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 2011-2014 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_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
654 #define P_KEEP 0x8000 /**< leave this page alone during spill */
656 uint16_t mp_flags; /**< @ref mdb_page */
657 #define mp_lower mp_pb.pb.pb_lower
658 #define mp_upper mp_pb.pb.pb_upper
659 #define mp_pages mp_pb.pb_pages
662 indx_t pb_lower; /**< lower bound of free space */
663 indx_t pb_upper; /**< upper bound of free space */
665 uint32_t pb_pages; /**< number of overflow pages */
667 indx_t mp_ptrs[1]; /**< dynamic size */
670 /** Size of the page header, excluding dynamic data at the end */
671 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
673 /** Address of first usable data byte in a page, after the header */
674 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
676 /** Number of nodes on a page */
677 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
679 /** The amount of space remaining in the page */
680 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
682 /** The percentage of space used in the page, in tenths of a percent. */
683 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
684 ((env)->me_psize - PAGEHDRSZ))
685 /** The minimum page fill factor, in tenths of a percent.
686 * Pages emptier than this are candidates for merging.
688 #define FILL_THRESHOLD 250
690 /** Test if a page is a leaf page */
691 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
692 /** Test if a page is a LEAF2 page */
693 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
694 /** Test if a page is a branch page */
695 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
696 /** Test if a page is an overflow page */
697 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
698 /** Test if a page is a sub page */
699 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
701 /** The number of overflow pages needed to store the given size. */
702 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
704 /** Header for a single key/data pair within a page.
705 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
706 * We guarantee 2-byte alignment for 'MDB_node's.
708 typedef struct MDB_node {
709 /** lo and hi are used for data size on leaf nodes and for
710 * child pgno on branch nodes. On 64 bit platforms, flags
711 * is also used for pgno. (Branch nodes have no flags).
712 * They are in host byte order in case that lets some
713 * accesses be optimized into a 32-bit word access.
715 #if BYTE_ORDER == LITTLE_ENDIAN
716 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
718 unsigned short mn_hi, mn_lo;
720 /** @defgroup mdb_node Node Flags
722 * Flags for node headers.
725 #define F_BIGDATA 0x01 /**< data put on overflow page */
726 #define F_SUBDATA 0x02 /**< data is a sub-database */
727 #define F_DUPDATA 0x04 /**< data has duplicates */
729 /** valid flags for #mdb_node_add() */
730 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
733 unsigned short mn_flags; /**< @ref mdb_node */
734 unsigned short mn_ksize; /**< key size */
735 char mn_data[1]; /**< key and data are appended here */
738 /** Size of the node header, excluding dynamic data at the end */
739 #define NODESIZE offsetof(MDB_node, mn_data)
741 /** Bit position of top word in page number, for shifting mn_flags */
742 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
744 /** Size of a node in a branch page with a given key.
745 * This is just the node header plus the key, there is no data.
747 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
749 /** Size of a node in a leaf page with a given key and data.
750 * This is node header plus key plus data size.
752 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
754 /** Address of node \b i in page \b p */
755 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
757 /** Address of the key for the node */
758 #define NODEKEY(node) (void *)((node)->mn_data)
760 /** Address of the data for a node */
761 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
763 /** Get the page number pointed to by a branch node */
764 #define NODEPGNO(node) \
765 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
766 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
767 /** Set the page number in a branch node */
768 #define SETPGNO(node,pgno) do { \
769 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
770 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
772 /** Get the size of the data in a leaf node */
773 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
774 /** Set the size of the data for a leaf node */
775 #define SETDSZ(node,size) do { \
776 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
777 /** The size of a key in a node */
778 #define NODEKSZ(node) ((node)->mn_ksize)
780 /** Copy a page number from src to dst */
782 #define COPY_PGNO(dst,src) dst = src
784 #if SIZE_MAX > 4294967295UL
785 #define COPY_PGNO(dst,src) do { \
786 unsigned short *s, *d; \
787 s = (unsigned short *)&(src); \
788 d = (unsigned short *)&(dst); \
795 #define COPY_PGNO(dst,src) do { \
796 unsigned short *s, *d; \
797 s = (unsigned short *)&(src); \
798 d = (unsigned short *)&(dst); \
804 /** The address of a key in a LEAF2 page.
805 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
806 * There are no node headers, keys are stored contiguously.
808 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
810 /** Set the \b node's key into \b keyptr, if requested. */
811 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
812 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
814 /** Set the \b node's key into \b key. */
815 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
817 /** Information about a single database in the environment. */
818 typedef struct MDB_db {
819 uint32_t md_pad; /**< also ksize for LEAF2 pages */
820 uint16_t md_flags; /**< @ref mdb_dbi_open */
821 uint16_t md_depth; /**< depth of this tree */
822 pgno_t md_branch_pages; /**< number of internal pages */
823 pgno_t md_leaf_pages; /**< number of leaf pages */
824 pgno_t md_overflow_pages; /**< number of overflow pages */
825 size_t md_entries; /**< number of data items */
826 pgno_t md_root; /**< the root page of this tree */
829 /** mdb_dbi_open flags */
830 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
831 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
832 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
833 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
835 /** Handle for the DB used to track free pages. */
837 /** Handle for the default DB. */
840 /** Meta page content.
841 * A meta page is the start point for accessing a database snapshot.
842 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
844 typedef struct MDB_meta {
845 /** Stamp identifying this as an MDB file. It must be set
848 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
850 void *mm_address; /**< address for fixed mapping */
851 size_t mm_mapsize; /**< size of mmap region */
852 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
853 /** The size of pages used in this DB */
854 #define mm_psize mm_dbs[0].md_pad
855 /** Any persistent environment flags. @ref mdb_env */
856 #define mm_flags mm_dbs[0].md_flags
857 pgno_t mm_last_pg; /**< last used page in file */
858 txnid_t mm_txnid; /**< txnid that committed this page */
861 /** Buffer for a stack-allocated meta page.
862 * The members define size and alignment, and silence type
863 * aliasing warnings. They are not used directly; that could
864 * mean incorrectly using several union members in parallel.
866 typedef union MDB_metabuf {
869 char mm_pad[PAGEHDRSZ];
874 /** Auxiliary DB info.
875 * The information here is mostly static/read-only. There is
876 * only a single copy of this record in the environment.
878 typedef struct MDB_dbx {
879 MDB_val md_name; /**< name of the database */
880 MDB_cmp_func *md_cmp; /**< function for comparing keys */
881 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
882 MDB_rel_func *md_rel; /**< user relocate function */
883 void *md_relctx; /**< user-provided context for md_rel */
886 /** A database transaction.
887 * Every operation requires a transaction handle.
890 MDB_txn *mt_parent; /**< parent of a nested txn */
891 MDB_txn *mt_child; /**< nested txn under this txn */
892 pgno_t mt_next_pgno; /**< next unallocated page */
893 /** The ID of this transaction. IDs are integers incrementing from 1.
894 * Only committed write transactions increment the ID. If a transaction
895 * aborts, the ID may be re-used by the next writer.
898 MDB_env *mt_env; /**< the DB environment */
899 /** The list of pages that became unused during this transaction.
902 /** The sorted list of dirty pages we temporarily wrote to disk
903 * because the dirty list was full. page numbers in here are
904 * shifted left by 1, deleted slots have the LSB set.
906 MDB_IDL mt_spill_pgs;
908 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
910 /** For read txns: This thread/txn's reader table slot, or NULL. */
913 /** Array of records for each DB known in the environment. */
915 /** Array of MDB_db records for each known DB */
917 /** @defgroup mt_dbflag Transaction DB Flags
921 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
922 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
923 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
924 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
926 /** In write txns, array of cursors for each DB */
927 MDB_cursor **mt_cursors;
928 /** Array of flags for each DB */
929 unsigned char *mt_dbflags;
930 /** Number of DB records in use. This number only ever increments;
931 * we don't decrement it when individual DB handles are closed.
935 /** @defgroup mdb_txn Transaction Flags
939 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
940 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
941 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
942 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
944 unsigned int mt_flags; /**< @ref mdb_txn */
945 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
946 * Includes ancestor txns' dirty pages not hidden by other txns'
947 * dirty/spilled pages. Thus commit(nested txn) has room to merge
948 * dirty_list into mt_parent after freeing hidden mt_parent pages.
950 unsigned int mt_dirty_room;
953 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
954 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
955 * raise this on a 64 bit machine.
957 #define CURSOR_STACK 32
961 /** Cursors are used for all DB operations.
962 * A cursor holds a path of (page pointer, key index) from the DB
963 * root to a position in the DB, plus other state. #MDB_DUPSORT
964 * cursors include an xcursor to the current data item. Write txns
965 * track their cursors and keep them up to date when data moves.
966 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
967 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
970 /** Next cursor on this DB in this txn */
972 /** Backup of the original cursor if this cursor is a shadow */
973 MDB_cursor *mc_backup;
974 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
975 struct MDB_xcursor *mc_xcursor;
976 /** The transaction that owns this cursor */
978 /** The database handle this cursor operates on */
980 /** The database record for this cursor */
982 /** The database auxiliary record for this cursor */
984 /** The @ref mt_dbflag for this database */
985 unsigned char *mc_dbflag;
986 unsigned short mc_snum; /**< number of pushed pages */
987 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
988 /** @defgroup mdb_cursor Cursor Flags
990 * Cursor state flags.
993 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
994 #define C_EOF 0x02 /**< No more data */
995 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
996 #define C_DEL 0x08 /**< last op was a cursor_del */
997 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
998 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1000 unsigned int mc_flags; /**< @ref mdb_cursor */
1001 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1002 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1005 /** Context for sorted-dup records.
1006 * We could have gone to a fully recursive design, with arbitrarily
1007 * deep nesting of sub-databases. But for now we only handle these
1008 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1010 typedef struct MDB_xcursor {
1011 /** A sub-cursor for traversing the Dup DB */
1012 MDB_cursor mx_cursor;
1013 /** The database record for this Dup DB */
1015 /** The auxiliary DB record for this Dup DB */
1017 /** The @ref mt_dbflag for this Dup DB */
1018 unsigned char mx_dbflag;
1021 /** State of FreeDB old pages, stored in the MDB_env */
1022 typedef struct MDB_pgstate {
1023 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1024 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1025 MDB_page *mf_pgloose; /**< Dirty pages that can be reused */
1028 /** The database environment. */
1030 HANDLE me_fd; /**< The main data file */
1031 HANDLE me_lfd; /**< The lock file */
1032 HANDLE me_mfd; /**< just for writing the meta pages */
1033 /** Failed to update the meta page. Probably an I/O error. */
1034 #define MDB_FATAL_ERROR 0x80000000U
1035 /** Some fields are initialized. */
1036 #define MDB_ENV_ACTIVE 0x20000000U
1037 /** me_txkey is set */
1038 #define MDB_ENV_TXKEY 0x10000000U
1039 uint32_t me_flags; /**< @ref mdb_env */
1040 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1041 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1042 unsigned int me_maxreaders; /**< size of the reader table */
1043 unsigned int me_numreaders; /**< max numreaders set by this env */
1044 MDB_dbi me_numdbs; /**< number of DBs opened */
1045 MDB_dbi me_maxdbs; /**< size of the DB table */
1046 MDB_PID_T me_pid; /**< process ID of this env */
1047 char *me_path; /**< path to the DB files */
1048 char *me_map; /**< the memory map of the data file */
1049 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1050 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1051 void *me_pbuf; /**< scratch area for DUPSORT put() */
1052 MDB_txn *me_txn; /**< current write transaction */
1053 size_t me_mapsize; /**< size of the data memory map */
1054 off_t me_size; /**< current file size */
1055 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1056 MDB_dbx *me_dbxs; /**< array of static DB info */
1057 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1058 pthread_key_t me_txkey; /**< thread-key for readers */
1059 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1060 # define me_pglast me_pgstate.mf_pglast
1061 # define me_pghead me_pgstate.mf_pghead
1062 # define me_pgloose me_pgstate.mf_pgloose
1063 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1064 /** IDL of pages that became unused in a write txn */
1065 MDB_IDL me_free_pgs;
1066 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1067 MDB_ID2L me_dirty_list;
1068 /** Max number of freelist items that can fit in a single overflow page */
1070 /** Max size of a node on a page */
1071 unsigned int me_nodemax;
1072 #if !(MDB_MAXKEYSIZE)
1073 unsigned int me_maxkey; /**< max size of a key */
1075 int me_live_reader; /**< have liveness lock in reader table */
1077 int me_pidquery; /**< Used in OpenProcess */
1078 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1080 #elif defined(MDB_USE_POSIX_SEM)
1081 sem_t *me_rmutex; /* Shared mutexes are not supported */
1084 void *me_userctx; /**< User-settable context */
1085 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1088 /** Nested transaction */
1089 typedef struct MDB_ntxn {
1090 MDB_txn mnt_txn; /**< the transaction */
1091 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1094 /** max number of pages to commit in one writev() call */
1095 #define MDB_COMMIT_PAGES 64
1096 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1097 #undef MDB_COMMIT_PAGES
1098 #define MDB_COMMIT_PAGES IOV_MAX
1101 /** max bytes to write in one call */
1102 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1104 /** Check \b txn and \b dbi arguments to a function */
1105 #define TXN_DBI_EXIST(txn, dbi) \
1106 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1108 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1109 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1110 static int mdb_page_touch(MDB_cursor *mc);
1112 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1113 static int mdb_page_search_root(MDB_cursor *mc,
1114 MDB_val *key, int modify);
1115 #define MDB_PS_MODIFY 1
1116 #define MDB_PS_ROOTONLY 2
1117 #define MDB_PS_FIRST 4
1118 #define MDB_PS_LAST 8
1119 static int mdb_page_search(MDB_cursor *mc,
1120 MDB_val *key, int flags);
1121 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1123 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1124 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1125 pgno_t newpgno, unsigned int nflags);
1127 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1128 static int mdb_env_pick_meta(const MDB_env *env);
1129 static int mdb_env_write_meta(MDB_txn *txn);
1130 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1131 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1133 static void mdb_env_close0(MDB_env *env, int excl);
1135 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1136 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1137 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1138 static void mdb_node_del(MDB_cursor *mc, int ksize);
1139 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1140 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1141 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1142 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1143 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1145 static int mdb_rebalance(MDB_cursor *mc);
1146 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1148 static void mdb_cursor_pop(MDB_cursor *mc);
1149 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1151 static int mdb_cursor_del0(MDB_cursor *mc);
1152 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1153 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1154 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1155 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1156 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1158 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1159 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1161 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1162 static void mdb_xcursor_init0(MDB_cursor *mc);
1163 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1165 static int mdb_drop0(MDB_cursor *mc, int subs);
1166 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1169 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1173 static SECURITY_DESCRIPTOR mdb_null_sd;
1174 static SECURITY_ATTRIBUTES mdb_all_sa;
1175 static int mdb_sec_inited;
1178 /** Return the library version info. */
1180 mdb_version(int *major, int *minor, int *patch)
1182 if (major) *major = MDB_VERSION_MAJOR;
1183 if (minor) *minor = MDB_VERSION_MINOR;
1184 if (patch) *patch = MDB_VERSION_PATCH;
1185 return MDB_VERSION_STRING;
1188 /** Table of descriptions for MDB @ref errors */
1189 static char *const mdb_errstr[] = {
1190 "MDB_KEYEXIST: Key/data pair already exists",
1191 "MDB_NOTFOUND: No matching key/data pair found",
1192 "MDB_PAGE_NOTFOUND: Requested page not found",
1193 "MDB_CORRUPTED: Located page was wrong type",
1194 "MDB_PANIC: Update of meta page failed",
1195 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1196 "MDB_INVALID: File is not an MDB file",
1197 "MDB_MAP_FULL: Environment mapsize limit reached",
1198 "MDB_DBS_FULL: Environment maxdbs limit reached",
1199 "MDB_READERS_FULL: Environment maxreaders limit reached",
1200 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1201 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1202 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1203 "MDB_PAGE_FULL: Internal error - page has no more space",
1204 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1205 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1206 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1207 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1208 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1212 mdb_strerror(int err)
1216 return ("Successful return: 0");
1218 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1219 i = err - MDB_KEYEXIST;
1220 return mdb_errstr[i];
1223 return strerror(err);
1226 /** assert(3) variant in cursor context */
1227 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1228 /** assert(3) variant in transaction context */
1229 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1230 /** assert(3) variant in environment context */
1231 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1234 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1235 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1238 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1239 const char *func, const char *file, int line)
1242 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1243 file, line, expr_txt, func);
1244 if (env->me_assert_func)
1245 env->me_assert_func(env, buf);
1246 fprintf(stderr, "%s\n", buf);
1250 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1254 /** Return the page number of \b mp which may be sub-page, for debug output */
1256 mdb_dbg_pgno(MDB_page *mp)
1259 COPY_PGNO(ret, mp->mp_pgno);
1263 /** Display a key in hexadecimal and return the address of the result.
1264 * @param[in] key the key to display
1265 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1266 * @return The key in hexadecimal form.
1269 mdb_dkey(MDB_val *key, char *buf)
1272 unsigned char *c = key->mv_data;
1278 if (key->mv_size > DKBUF_MAXKEYSIZE)
1279 return "MDB_MAXKEYSIZE";
1280 /* may want to make this a dynamic check: if the key is mostly
1281 * printable characters, print it as-is instead of converting to hex.
1285 for (i=0; i<key->mv_size; i++)
1286 ptr += sprintf(ptr, "%02x", *c++);
1288 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1294 mdb_leafnode_type(MDB_node *n)
1296 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1297 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1298 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1301 /** Display all the keys in the page. */
1303 mdb_page_list(MDB_page *mp)
1305 pgno_t pgno = mdb_dbg_pgno(mp);
1306 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1308 unsigned int i, nkeys, nsize, total = 0;
1312 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1313 case P_BRANCH: type = "Branch page"; break;
1314 case P_LEAF: type = "Leaf page"; break;
1315 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1316 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1317 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1319 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1320 pgno, mp->mp_pages, state);
1323 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1324 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1327 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1331 nkeys = NUMKEYS(mp);
1332 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1334 for (i=0; i<nkeys; i++) {
1335 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1336 key.mv_size = nsize = mp->mp_pad;
1337 key.mv_data = LEAF2KEY(mp, i, nsize);
1339 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1342 node = NODEPTR(mp, i);
1343 key.mv_size = node->mn_ksize;
1344 key.mv_data = node->mn_data;
1345 nsize = NODESIZE + key.mv_size;
1346 if (IS_BRANCH(mp)) {
1347 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1351 if (F_ISSET(node->mn_flags, F_BIGDATA))
1352 nsize += sizeof(pgno_t);
1354 nsize += NODEDSZ(node);
1356 nsize += sizeof(indx_t);
1357 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1358 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1360 total = EVEN(total);
1362 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1363 IS_LEAF2(mp) ? PAGEHDRSZ : mp->mp_lower, total, SIZELEFT(mp));
1367 mdb_cursor_chk(MDB_cursor *mc)
1373 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1374 for (i=0; i<mc->mc_top; i++) {
1376 node = NODEPTR(mp, mc->mc_ki[i]);
1377 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1380 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1386 /** Count all the pages in each DB and in the freelist
1387 * and make sure it matches the actual number of pages
1389 * All named DBs must be open for a correct count.
1391 static void mdb_audit(MDB_txn *txn)
1395 MDB_ID freecount, count;
1400 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1401 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1402 freecount += *(MDB_ID *)data.mv_data;
1403 mdb_tassert(txn, rc == MDB_NOTFOUND);
1406 for (i = 0; i<txn->mt_numdbs; i++) {
1408 if (!(txn->mt_dbflags[i] & DB_VALID))
1410 mdb_cursor_init(&mc, txn, i, &mx);
1411 if (txn->mt_dbs[i].md_root == P_INVALID)
1413 count += txn->mt_dbs[i].md_branch_pages +
1414 txn->mt_dbs[i].md_leaf_pages +
1415 txn->mt_dbs[i].md_overflow_pages;
1416 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1417 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1418 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1421 mp = mc.mc_pg[mc.mc_top];
1422 for (j=0; j<NUMKEYS(mp); j++) {
1423 MDB_node *leaf = NODEPTR(mp, j);
1424 if (leaf->mn_flags & F_SUBDATA) {
1426 memcpy(&db, NODEDATA(leaf), sizeof(db));
1427 count += db.md_branch_pages + db.md_leaf_pages +
1428 db.md_overflow_pages;
1432 mdb_tassert(txn, rc == MDB_NOTFOUND);
1435 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1436 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1437 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1443 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1445 return txn->mt_dbxs[dbi].md_cmp(a, b);
1449 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1451 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1454 /** Allocate memory for a page.
1455 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1458 mdb_page_malloc(MDB_txn *txn, unsigned num)
1460 MDB_env *env = txn->mt_env;
1461 MDB_page *ret = env->me_dpages;
1462 size_t psize = env->me_psize, sz = psize, off;
1463 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1464 * For a single page alloc, we init everything after the page header.
1465 * For multi-page, we init the final page; if the caller needed that
1466 * many pages they will be filling in at least up to the last page.
1470 VGMEMP_ALLOC(env, ret, sz);
1471 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1472 env->me_dpages = ret->mp_next;
1475 psize -= off = PAGEHDRSZ;
1480 if ((ret = malloc(sz)) != NULL) {
1481 VGMEMP_ALLOC(env, ret, sz);
1482 if (!(env->me_flags & MDB_NOMEMINIT)) {
1483 memset((char *)ret + off, 0, psize);
1487 txn->mt_flags |= MDB_TXN_ERROR;
1491 /** Free a single page.
1492 * Saves single pages to a list, for future reuse.
1493 * (This is not used for multi-page overflow pages.)
1496 mdb_page_free(MDB_env *env, MDB_page *mp)
1498 mp->mp_next = env->me_dpages;
1499 VGMEMP_FREE(env, mp);
1500 env->me_dpages = mp;
1503 /** Free a dirty page */
1505 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1507 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1508 mdb_page_free(env, dp);
1510 /* large pages just get freed directly */
1511 VGMEMP_FREE(env, dp);
1516 /** Return all dirty pages to dpage list */
1518 mdb_dlist_free(MDB_txn *txn)
1520 MDB_env *env = txn->mt_env;
1521 MDB_ID2L dl = txn->mt_u.dirty_list;
1522 unsigned i, n = dl[0].mid;
1524 for (i = 1; i <= n; i++) {
1525 mdb_dpage_free(env, dl[i].mptr);
1530 /** Loosen a single page.
1531 * Saves single pages to a list for future reuse
1532 * in this same txn. It has been pulled from the freeDB
1533 * and already resides on the dirty list, but has been
1534 * deleted. Use these pages first before pulling again
1538 mdb_page_loose(MDB_env *env, MDB_page *mp)
1540 pgno_t *pp = (pgno_t *)mp->mp_ptrs;
1542 mp->mp_next = env->me_pgloose;
1543 env->me_pgloose = mp;
1544 mp->mp_flags |= P_LOOSE;
1547 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1548 * @param[in] mc A cursor handle for the current operation.
1549 * @param[in] pflags Flags of the pages to update:
1550 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1551 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1552 * @return 0 on success, non-zero on failure.
1555 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1557 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1558 MDB_txn *txn = mc->mc_txn;
1564 int rc = MDB_SUCCESS, level;
1566 /* Mark pages seen by cursors */
1567 if (mc->mc_flags & C_UNTRACK)
1568 mc = NULL; /* will find mc in mt_cursors */
1569 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1570 for (; mc; mc=mc->mc_next) {
1571 if (!(mc->mc_flags & C_INITIALIZED))
1573 for (m3 = mc;; m3 = &mx->mx_cursor) {
1575 for (j=0; j<m3->mc_snum; j++) {
1577 if ((mp->mp_flags & Mask) == pflags)
1578 mp->mp_flags ^= P_KEEP;
1580 mx = m3->mc_xcursor;
1581 /* Proceed to mx if it is at a sub-database */
1582 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1584 if (! (mp && (mp->mp_flags & P_LEAF)))
1586 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1587 if (!(leaf->mn_flags & F_SUBDATA))
1595 /* Loose pages shouldn't be spilled */
1596 for (dp = txn->mt_env->me_pgloose; dp; dp=dp->mp_next) {
1597 if ((dp->mp_flags & Mask) == pflags)
1598 dp->mp_flags ^= P_KEEP;
1602 /* Mark dirty root pages */
1603 for (i=0; i<txn->mt_numdbs; i++) {
1604 if (txn->mt_dbflags[i] & DB_DIRTY) {
1605 pgno_t pgno = txn->mt_dbs[i].md_root;
1606 if (pgno == P_INVALID)
1608 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1610 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1611 dp->mp_flags ^= P_KEEP;
1619 static int mdb_page_flush(MDB_txn *txn, int keep);
1621 /** Spill pages from the dirty list back to disk.
1622 * This is intended to prevent running into #MDB_TXN_FULL situations,
1623 * but note that they may still occur in a few cases:
1624 * 1) our estimate of the txn size could be too small. Currently this
1625 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1626 * 2) child txns may run out of space if their parents dirtied a
1627 * lot of pages and never spilled them. TODO: we probably should do
1628 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1629 * the parent's dirty_room is below a given threshold.
1631 * Otherwise, if not using nested txns, it is expected that apps will
1632 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1633 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1634 * If the txn never references them again, they can be left alone.
1635 * If the txn only reads them, they can be used without any fuss.
1636 * If the txn writes them again, they can be dirtied immediately without
1637 * going thru all of the work of #mdb_page_touch(). Such references are
1638 * handled by #mdb_page_unspill().
1640 * Also note, we never spill DB root pages, nor pages of active cursors,
1641 * because we'll need these back again soon anyway. And in nested txns,
1642 * we can't spill a page in a child txn if it was already spilled in a
1643 * parent txn. That would alter the parent txns' data even though
1644 * the child hasn't committed yet, and we'd have no way to undo it if
1645 * the child aborted.
1647 * @param[in] m0 cursor A cursor handle identifying the transaction and
1648 * database for which we are checking space.
1649 * @param[in] key For a put operation, the key being stored.
1650 * @param[in] data For a put operation, the data being stored.
1651 * @return 0 on success, non-zero on failure.
1654 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1656 MDB_txn *txn = m0->mc_txn;
1658 MDB_ID2L dl = txn->mt_u.dirty_list;
1659 unsigned int i, j, need;
1662 if (m0->mc_flags & C_SUB)
1665 /* Estimate how much space this op will take */
1666 i = m0->mc_db->md_depth;
1667 /* Named DBs also dirty the main DB */
1668 if (m0->mc_dbi > MAIN_DBI)
1669 i += txn->mt_dbs[MAIN_DBI].md_depth;
1670 /* For puts, roughly factor in the key+data size */
1672 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1673 i += i; /* double it for good measure */
1676 if (txn->mt_dirty_room > i)
1679 if (!txn->mt_spill_pgs) {
1680 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1681 if (!txn->mt_spill_pgs)
1684 /* purge deleted slots */
1685 MDB_IDL sl = txn->mt_spill_pgs;
1686 unsigned int num = sl[0];
1688 for (i=1; i<=num; i++) {
1695 /* Preserve pages which may soon be dirtied again */
1696 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1699 /* Less aggressive spill - we originally spilled the entire dirty list,
1700 * with a few exceptions for cursor pages and DB root pages. But this
1701 * turns out to be a lot of wasted effort because in a large txn many
1702 * of those pages will need to be used again. So now we spill only 1/8th
1703 * of the dirty pages. Testing revealed this to be a good tradeoff,
1704 * better than 1/2, 1/4, or 1/10.
1706 if (need < MDB_IDL_UM_MAX / 8)
1707 need = MDB_IDL_UM_MAX / 8;
1709 /* Save the page IDs of all the pages we're flushing */
1710 /* flush from the tail forward, this saves a lot of shifting later on. */
1711 for (i=dl[0].mid; i && need; i--) {
1712 MDB_ID pn = dl[i].mid << 1;
1714 if (dp->mp_flags & P_KEEP)
1716 /* Can't spill twice, make sure it's not already in a parent's
1719 if (txn->mt_parent) {
1721 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1722 if (tx2->mt_spill_pgs) {
1723 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1724 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1725 dp->mp_flags |= P_KEEP;
1733 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1737 mdb_midl_sort(txn->mt_spill_pgs);
1739 /* Flush the spilled part of dirty list */
1740 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1743 /* Reset any dirty pages we kept that page_flush didn't see */
1744 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1747 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1751 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1753 mdb_find_oldest(MDB_txn *txn)
1756 txnid_t mr, oldest = txn->mt_txnid - 1;
1757 if (txn->mt_env->me_txns) {
1758 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1759 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1770 /** Add a page to the txn's dirty list */
1772 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1775 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1777 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1778 insert = mdb_mid2l_append;
1780 insert = mdb_mid2l_insert;
1782 mid.mid = mp->mp_pgno;
1784 rc = insert(txn->mt_u.dirty_list, &mid);
1785 mdb_tassert(txn, rc == 0);
1786 txn->mt_dirty_room--;
1789 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1790 * me_pghead and mt_next_pgno.
1792 * If there are free pages available from older transactions, they
1793 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1794 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1795 * and move me_pglast to say which records were consumed. Only this
1796 * function can create me_pghead and move me_pglast/mt_next_pgno.
1797 * @param[in] mc cursor A cursor handle identifying the transaction and
1798 * database for which we are allocating.
1799 * @param[in] num the number of pages to allocate.
1800 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1801 * will always be satisfied by a single contiguous chunk of memory.
1802 * @return 0 on success, non-zero on failure.
1805 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1807 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1808 /* Get at most <Max_retries> more freeDB records once me_pghead
1809 * has enough pages. If not enough, use new pages from the map.
1810 * If <Paranoid> and mc is updating the freeDB, only get new
1811 * records if me_pghead is empty. Then the freelist cannot play
1812 * catch-up with itself by growing while trying to save it.
1814 enum { Paranoid = 1, Max_retries = 500 };
1816 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1818 int rc, retry = num * 20;
1819 MDB_txn *txn = mc->mc_txn;
1820 MDB_env *env = txn->mt_env;
1821 pgno_t pgno, *mop = env->me_pghead;
1822 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1824 txnid_t oldest = 0, last;
1828 /* If there are any loose pages, just use them */
1829 if (num == 1 && env->me_pgloose) {
1831 np = env->me_pgloose;
1832 env->me_pgloose = np->mp_next;
1833 pp = (pgno_t *)np->mp_ptrs;
1841 /* If our dirty list is already full, we can't do anything */
1842 if (txn->mt_dirty_room == 0) {
1847 for (op = MDB_FIRST;; op = MDB_NEXT) {
1850 pgno_t *idl, old_id, new_id;
1852 /* Seek a big enough contiguous page range. Prefer
1853 * pages at the tail, just truncating the list.
1859 if (mop[i-n2] == pgno+n2)
1866 if (op == MDB_FIRST) { /* 1st iteration */
1867 /* Prepare to fetch more and coalesce */
1868 oldest = mdb_find_oldest(txn);
1869 last = env->me_pglast;
1870 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1873 key.mv_data = &last; /* will look up last+1 */
1874 key.mv_size = sizeof(last);
1876 if (Paranoid && mc->mc_dbi == FREE_DBI)
1879 if (Paranoid && retry < 0 && mop_len)
1883 /* Do not fetch more if the record will be too recent */
1886 rc = mdb_cursor_get(&m2, &key, NULL, op);
1888 if (rc == MDB_NOTFOUND)
1892 last = *(txnid_t*)key.mv_data;
1895 np = m2.mc_pg[m2.mc_top];
1896 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1897 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1900 idl = (MDB_ID *) data.mv_data;
1903 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1908 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1910 mop = env->me_pghead;
1912 env->me_pglast = last;
1914 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1915 last, txn->mt_dbs[FREE_DBI].md_root, i));
1917 DPRINTF(("IDL %"Z"u", idl[k]));
1919 /* Merge in descending sorted order */
1922 mop[0] = (pgno_t)-1;
1926 for (; old_id < new_id; old_id = mop[--j])
1933 /* Use new pages from the map when nothing suitable in the freeDB */
1935 pgno = txn->mt_next_pgno;
1936 if (pgno + num >= env->me_maxpg) {
1937 DPUTS("DB size maxed out");
1943 if (env->me_flags & MDB_WRITEMAP) {
1944 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1946 if (!(np = mdb_page_malloc(txn, num))) {
1952 mop[0] = mop_len -= num;
1953 /* Move any stragglers down */
1954 for (j = i-num; j < mop_len; )
1955 mop[++j] = mop[++i];
1957 txn->mt_next_pgno = pgno + num;
1960 mdb_page_dirty(txn, np);
1966 txn->mt_flags |= MDB_TXN_ERROR;
1970 /** Copy the used portions of a non-overflow page.
1971 * @param[in] dst page to copy into
1972 * @param[in] src page to copy from
1973 * @param[in] psize size of a page
1976 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1978 enum { Align = sizeof(pgno_t) };
1979 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1981 /* If page isn't full, just copy the used portion. Adjust
1982 * alignment so memcpy may copy words instead of bytes.
1984 if ((unused &= -Align) && !IS_LEAF2(src)) {
1986 memcpy(dst, src, (lower + (Align-1)) & -Align);
1987 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1990 memcpy(dst, src, psize - unused);
1994 /** Pull a page off the txn's spill list, if present.
1995 * If a page being referenced was spilled to disk in this txn, bring
1996 * it back and make it dirty/writable again.
1997 * @param[in] txn the transaction handle.
1998 * @param[in] mp the page being referenced. It must not be dirty.
1999 * @param[out] ret the writable page, if any. ret is unchanged if
2000 * mp wasn't spilled.
2003 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2005 MDB_env *env = txn->mt_env;
2008 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2010 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2011 if (!tx2->mt_spill_pgs)
2013 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2014 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2017 if (txn->mt_dirty_room == 0)
2018 return MDB_TXN_FULL;
2019 if (IS_OVERFLOW(mp))
2023 if (env->me_flags & MDB_WRITEMAP) {
2026 np = mdb_page_malloc(txn, num);
2030 memcpy(np, mp, num * env->me_psize);
2032 mdb_page_copy(np, mp, env->me_psize);
2035 /* If in current txn, this page is no longer spilled.
2036 * If it happens to be the last page, truncate the spill list.
2037 * Otherwise mark it as deleted by setting the LSB.
2039 if (x == txn->mt_spill_pgs[0])
2040 txn->mt_spill_pgs[0]--;
2042 txn->mt_spill_pgs[x] |= 1;
2043 } /* otherwise, if belonging to a parent txn, the
2044 * page remains spilled until child commits
2047 mdb_page_dirty(txn, np);
2048 np->mp_flags |= P_DIRTY;
2056 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2057 * @param[in] mc cursor pointing to the page to be touched
2058 * @return 0 on success, non-zero on failure.
2061 mdb_page_touch(MDB_cursor *mc)
2063 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2064 MDB_txn *txn = mc->mc_txn;
2065 MDB_cursor *m2, *m3;
2069 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2070 if (txn->mt_flags & MDB_TXN_SPILLS) {
2072 rc = mdb_page_unspill(txn, mp, &np);
2078 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2079 (rc = mdb_page_alloc(mc, 1, &np)))
2082 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2083 mp->mp_pgno, pgno));
2084 mdb_cassert(mc, mp->mp_pgno != pgno);
2085 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2086 /* Update the parent page, if any, to point to the new page */
2088 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2089 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2090 SETPGNO(node, pgno);
2092 mc->mc_db->md_root = pgno;
2094 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2095 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2097 /* If txn has a parent, make sure the page is in our
2101 unsigned x = mdb_mid2l_search(dl, pgno);
2102 if (x <= dl[0].mid && dl[x].mid == pgno) {
2103 if (mp != dl[x].mptr) { /* bad cursor? */
2104 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2105 txn->mt_flags |= MDB_TXN_ERROR;
2106 return MDB_CORRUPTED;
2111 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2113 np = mdb_page_malloc(txn, 1);
2118 rc = mdb_mid2l_insert(dl, &mid);
2119 mdb_cassert(mc, rc == 0);
2124 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2126 np->mp_flags |= P_DIRTY;
2129 /* Adjust cursors pointing to mp */
2130 mc->mc_pg[mc->mc_top] = np;
2131 m2 = txn->mt_cursors[mc->mc_dbi];
2132 if (mc->mc_flags & C_SUB) {
2133 for (; m2; m2=m2->mc_next) {
2134 m3 = &m2->mc_xcursor->mx_cursor;
2135 if (m3->mc_snum < mc->mc_snum) continue;
2136 if (m3->mc_pg[mc->mc_top] == mp)
2137 m3->mc_pg[mc->mc_top] = np;
2140 for (; m2; m2=m2->mc_next) {
2141 if (m2->mc_snum < mc->mc_snum) continue;
2142 if (m2->mc_pg[mc->mc_top] == mp) {
2143 m2->mc_pg[mc->mc_top] = np;
2144 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2146 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2148 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2149 if (!(leaf->mn_flags & F_SUBDATA))
2150 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2158 txn->mt_flags |= MDB_TXN_ERROR;
2163 mdb_env_sync(MDB_env *env, int force)
2166 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2167 if (env->me_flags & MDB_WRITEMAP) {
2168 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2169 ? MS_ASYNC : MS_SYNC;
2170 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2173 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2177 if (MDB_FDATASYNC(env->me_fd))
2184 /** Back up parent txn's cursors, then grab the originals for tracking */
2186 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2188 MDB_cursor *mc, *bk;
2193 for (i = src->mt_numdbs; --i >= 0; ) {
2194 if ((mc = src->mt_cursors[i]) != NULL) {
2195 size = sizeof(MDB_cursor);
2197 size += sizeof(MDB_xcursor);
2198 for (; mc; mc = bk->mc_next) {
2204 mc->mc_db = &dst->mt_dbs[i];
2205 /* Kill pointers into src - and dst to reduce abuse: The
2206 * user may not use mc until dst ends. Otherwise we'd...
2208 mc->mc_txn = NULL; /* ...set this to dst */
2209 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2210 if ((mx = mc->mc_xcursor) != NULL) {
2211 *(MDB_xcursor *)(bk+1) = *mx;
2212 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2214 mc->mc_next = dst->mt_cursors[i];
2215 dst->mt_cursors[i] = mc;
2222 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2223 * @param[in] txn the transaction handle.
2224 * @param[in] merge true to keep changes to parent cursors, false to revert.
2225 * @return 0 on success, non-zero on failure.
2228 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2230 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2234 for (i = txn->mt_numdbs; --i >= 0; ) {
2235 for (mc = cursors[i]; mc; mc = next) {
2237 if ((bk = mc->mc_backup) != NULL) {
2239 /* Commit changes to parent txn */
2240 mc->mc_next = bk->mc_next;
2241 mc->mc_backup = bk->mc_backup;
2242 mc->mc_txn = bk->mc_txn;
2243 mc->mc_db = bk->mc_db;
2244 mc->mc_dbflag = bk->mc_dbflag;
2245 if ((mx = mc->mc_xcursor) != NULL)
2246 mx->mx_cursor.mc_txn = bk->mc_txn;
2248 /* Abort nested txn */
2250 if ((mx = mc->mc_xcursor) != NULL)
2251 *mx = *(MDB_xcursor *)(bk+1);
2255 /* Only malloced cursors are permanently tracked. */
2263 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2266 mdb_txn_reset0(MDB_txn *txn, const char *act);
2268 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2274 Pidset = F_SETLK, Pidcheck = F_GETLK
2278 /** Set or check a pid lock. Set returns 0 on success.
2279 * Check returns 0 if the process is certainly dead, nonzero if it may
2280 * be alive (the lock exists or an error happened so we do not know).
2282 * On Windows Pidset is a no-op, we merely check for the existence
2283 * of the process with the given pid. On POSIX we use a single byte
2284 * lock on the lockfile, set at an offset equal to the pid.
2287 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2289 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2292 if (op == Pidcheck) {
2293 h = OpenProcess(env->me_pidquery, FALSE, pid);
2294 /* No documented "no such process" code, but other program use this: */
2296 return ErrCode() != ERROR_INVALID_PARAMETER;
2297 /* A process exists until all handles to it close. Has it exited? */
2298 ret = WaitForSingleObject(h, 0) != 0;
2305 struct flock lock_info;
2306 memset(&lock_info, 0, sizeof(lock_info));
2307 lock_info.l_type = F_WRLCK;
2308 lock_info.l_whence = SEEK_SET;
2309 lock_info.l_start = pid;
2310 lock_info.l_len = 1;
2311 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2312 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2314 } else if ((rc = ErrCode()) == EINTR) {
2322 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2323 * @param[in] txn the transaction handle to initialize
2324 * @return 0 on success, non-zero on failure.
2327 mdb_txn_renew0(MDB_txn *txn)
2329 MDB_env *env = txn->mt_env;
2330 MDB_txninfo *ti = env->me_txns;
2334 int rc, new_notls = 0;
2337 txn->mt_numdbs = env->me_numdbs;
2338 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2340 if (txn->mt_flags & MDB_TXN_RDONLY) {
2342 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2343 txn->mt_txnid = meta->mm_txnid;
2344 txn->mt_u.reader = NULL;
2346 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2347 pthread_getspecific(env->me_txkey);
2349 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2350 return MDB_BAD_RSLOT;
2352 MDB_PID_T pid = env->me_pid;
2353 pthread_t tid = pthread_self();
2355 if (!env->me_live_reader) {
2356 rc = mdb_reader_pid(env, Pidset, pid);
2359 env->me_live_reader = 1;
2363 nr = ti->mti_numreaders;
2364 for (i=0; i<nr; i++)
2365 if (ti->mti_readers[i].mr_pid == 0)
2367 if (i == env->me_maxreaders) {
2368 UNLOCK_MUTEX_R(env);
2369 return MDB_READERS_FULL;
2371 ti->mti_readers[i].mr_pid = pid;
2372 ti->mti_readers[i].mr_tid = tid;
2374 ti->mti_numreaders = ++nr;
2375 /* Save numreaders for un-mutexed mdb_env_close() */
2376 env->me_numreaders = nr;
2377 UNLOCK_MUTEX_R(env);
2379 r = &ti->mti_readers[i];
2380 new_notls = (env->me_flags & MDB_NOTLS);
2381 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2386 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2387 txn->mt_u.reader = r;
2388 meta = env->me_metas[txn->mt_txnid & 1];
2394 txn->mt_txnid = ti->mti_txnid;
2395 meta = env->me_metas[txn->mt_txnid & 1];
2397 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2398 txn->mt_txnid = meta->mm_txnid;
2402 if (txn->mt_txnid == mdb_debug_start)
2405 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2406 txn->mt_u.dirty_list = env->me_dirty_list;
2407 txn->mt_u.dirty_list[0].mid = 0;
2408 txn->mt_free_pgs = env->me_free_pgs;
2409 txn->mt_free_pgs[0] = 0;
2410 txn->mt_spill_pgs = NULL;
2414 /* Copy the DB info and flags */
2415 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2417 /* Moved to here to avoid a data race in read TXNs */
2418 txn->mt_next_pgno = meta->mm_last_pg+1;
2420 for (i=2; i<txn->mt_numdbs; i++) {
2421 x = env->me_dbflags[i];
2422 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2423 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2425 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2427 if (env->me_maxpg < txn->mt_next_pgno) {
2428 mdb_txn_reset0(txn, "renew0-mapfail");
2430 txn->mt_u.reader->mr_pid = 0;
2431 txn->mt_u.reader = NULL;
2433 return MDB_MAP_RESIZED;
2440 mdb_txn_renew(MDB_txn *txn)
2444 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2447 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2448 DPUTS("environment had fatal error, must shutdown!");
2452 rc = mdb_txn_renew0(txn);
2453 if (rc == MDB_SUCCESS) {
2454 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2455 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2456 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2462 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2466 int rc, size, tsize = sizeof(MDB_txn);
2468 if (env->me_flags & MDB_FATAL_ERROR) {
2469 DPUTS("environment had fatal error, must shutdown!");
2472 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2475 /* Nested transactions: Max 1 child, write txns only, no writemap */
2476 if (parent->mt_child ||
2477 (flags & MDB_RDONLY) ||
2478 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2479 (env->me_flags & MDB_WRITEMAP))
2481 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2483 tsize = sizeof(MDB_ntxn);
2485 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2486 if (!(flags & MDB_RDONLY))
2487 size += env->me_maxdbs * sizeof(MDB_cursor *);
2489 if ((txn = calloc(1, size)) == NULL) {
2490 DPRINTF(("calloc: %s", strerror(ErrCode())));
2493 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2494 if (flags & MDB_RDONLY) {
2495 txn->mt_flags |= MDB_TXN_RDONLY;
2496 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2498 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2499 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2505 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2506 if (!txn->mt_u.dirty_list ||
2507 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2509 free(txn->mt_u.dirty_list);
2513 txn->mt_txnid = parent->mt_txnid;
2514 txn->mt_dirty_room = parent->mt_dirty_room;
2515 txn->mt_u.dirty_list[0].mid = 0;
2516 txn->mt_spill_pgs = NULL;
2517 txn->mt_next_pgno = parent->mt_next_pgno;
2518 parent->mt_child = txn;
2519 txn->mt_parent = parent;
2520 txn->mt_numdbs = parent->mt_numdbs;
2521 txn->mt_flags = parent->mt_flags;
2522 txn->mt_dbxs = parent->mt_dbxs;
2523 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2524 /* Copy parent's mt_dbflags, but clear DB_NEW */
2525 for (i=0; i<txn->mt_numdbs; i++)
2526 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2528 ntxn = (MDB_ntxn *)txn;
2529 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2530 if (env->me_pghead) {
2531 size = MDB_IDL_SIZEOF(env->me_pghead);
2532 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2534 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2539 rc = mdb_cursor_shadow(parent, txn);
2541 mdb_txn_reset0(txn, "beginchild-fail");
2543 rc = mdb_txn_renew0(txn);
2549 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2550 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2551 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2558 mdb_txn_env(MDB_txn *txn)
2560 if(!txn) return NULL;
2564 /** Export or close DBI handles opened in this txn. */
2566 mdb_dbis_update(MDB_txn *txn, int keep)
2569 MDB_dbi n = txn->mt_numdbs;
2570 MDB_env *env = txn->mt_env;
2571 unsigned char *tdbflags = txn->mt_dbflags;
2573 for (i = n; --i >= 2;) {
2574 if (tdbflags[i] & DB_NEW) {
2576 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2578 char *ptr = env->me_dbxs[i].md_name.mv_data;
2579 env->me_dbxs[i].md_name.mv_data = NULL;
2580 env->me_dbxs[i].md_name.mv_size = 0;
2581 env->me_dbflags[i] = 0;
2586 if (keep && env->me_numdbs < n)
2590 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2591 * May be called twice for readonly txns: First reset it, then abort.
2592 * @param[in] txn the transaction handle to reset
2593 * @param[in] act why the transaction is being reset
2596 mdb_txn_reset0(MDB_txn *txn, const char *act)
2598 MDB_env *env = txn->mt_env;
2600 /* Close any DBI handles opened in this txn */
2601 mdb_dbis_update(txn, 0);
2603 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2604 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2605 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2607 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2608 if (txn->mt_u.reader) {
2609 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2610 if (!(env->me_flags & MDB_NOTLS))
2611 txn->mt_u.reader = NULL; /* txn does not own reader */
2613 txn->mt_numdbs = 0; /* close nothing if called again */
2614 txn->mt_dbxs = NULL; /* mark txn as reset */
2616 mdb_cursors_close(txn, 0);
2618 if (!(env->me_flags & MDB_WRITEMAP)) {
2619 mdb_dlist_free(txn);
2621 mdb_midl_free(env->me_pghead);
2623 if (txn->mt_parent) {
2624 txn->mt_parent->mt_child = NULL;
2625 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2626 mdb_midl_free(txn->mt_free_pgs);
2627 mdb_midl_free(txn->mt_spill_pgs);
2628 free(txn->mt_u.dirty_list);
2632 if (mdb_midl_shrink(&txn->mt_free_pgs))
2633 env->me_free_pgs = txn->mt_free_pgs;
2634 env->me_pghead = NULL;
2638 /* The writer mutex was locked in mdb_txn_begin. */
2640 UNLOCK_MUTEX_W(env);
2645 mdb_txn_reset(MDB_txn *txn)
2650 /* This call is only valid for read-only txns */
2651 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2654 mdb_txn_reset0(txn, "reset");
2658 mdb_txn_abort(MDB_txn *txn)
2664 mdb_txn_abort(txn->mt_child);
2666 mdb_txn_reset0(txn, "abort");
2667 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2668 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2669 txn->mt_u.reader->mr_pid = 0;
2674 /** Save the freelist as of this transaction to the freeDB.
2675 * This changes the freelist. Keep trying until it stabilizes.
2678 mdb_freelist_save(MDB_txn *txn)
2680 /* env->me_pghead[] can grow and shrink during this call.
2681 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2682 * Page numbers cannot disappear from txn->mt_free_pgs[].
2685 MDB_env *env = txn->mt_env;
2686 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2687 txnid_t pglast = 0, head_id = 0;
2688 pgno_t freecnt = 0, *free_pgs, *mop;
2689 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2691 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2693 if (env->me_pghead) {
2694 /* Make sure first page of freeDB is touched and on freelist */
2695 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2696 if (rc && rc != MDB_NOTFOUND)
2700 /* Dispose of loose pages. Usually they will have all
2701 * been used up by the time we get here.
2703 if (env->me_pgloose) {
2704 MDB_page *mp = env->me_pgloose;
2706 /* Just return them to freeDB */
2707 if (env->me_pghead) {
2709 mop = env->me_pghead;
2712 pp = (pgno_t *)mp->mp_ptrs;
2715 for (i = mop[0]; i && mop[i] < pg; i--)
2722 /* Oh well, they were wasted. Put on freelist */
2724 pp = (pgno_t *)mp->mp_ptrs;
2725 mdb_midl_append(&txn->mt_free_pgs, *pp);
2729 env->me_pgloose = NULL;
2732 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2733 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2734 ? SSIZE_MAX : maxfree_1pg;
2737 /* Come back here after each Put() in case freelist changed */
2742 /* If using records from freeDB which we have not yet
2743 * deleted, delete them and any we reserved for me_pghead.
2745 while (pglast < env->me_pglast) {
2746 rc = mdb_cursor_first(&mc, &key, NULL);
2749 pglast = head_id = *(txnid_t *)key.mv_data;
2750 total_room = head_room = 0;
2751 mdb_tassert(txn, pglast <= env->me_pglast);
2752 rc = mdb_cursor_del(&mc, 0);
2757 /* Save the IDL of pages freed by this txn, to a single record */
2758 if (freecnt < txn->mt_free_pgs[0]) {
2760 /* Make sure last page of freeDB is touched and on freelist */
2761 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2762 if (rc && rc != MDB_NOTFOUND)
2765 free_pgs = txn->mt_free_pgs;
2766 /* Write to last page of freeDB */
2767 key.mv_size = sizeof(txn->mt_txnid);
2768 key.mv_data = &txn->mt_txnid;
2770 freecnt = free_pgs[0];
2771 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2772 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2775 /* Retry if mt_free_pgs[] grew during the Put() */
2776 free_pgs = txn->mt_free_pgs;
2777 } while (freecnt < free_pgs[0]);
2778 mdb_midl_sort(free_pgs);
2779 memcpy(data.mv_data, free_pgs, data.mv_size);
2782 unsigned int i = free_pgs[0];
2783 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2784 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2786 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2792 mop = env->me_pghead;
2793 mop_len = mop ? mop[0] : 0;
2795 /* Reserve records for me_pghead[]. Split it if multi-page,
2796 * to avoid searching freeDB for a page range. Use keys in
2797 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2799 if (total_room >= mop_len) {
2800 if (total_room == mop_len || --more < 0)
2802 } else if (head_room >= maxfree_1pg && head_id > 1) {
2803 /* Keep current record (overflow page), add a new one */
2807 /* (Re)write {key = head_id, IDL length = head_room} */
2808 total_room -= head_room;
2809 head_room = mop_len - total_room;
2810 if (head_room > maxfree_1pg && head_id > 1) {
2811 /* Overflow multi-page for part of me_pghead */
2812 head_room /= head_id; /* amortize page sizes */
2813 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2814 } else if (head_room < 0) {
2815 /* Rare case, not bothering to delete this record */
2818 key.mv_size = sizeof(head_id);
2819 key.mv_data = &head_id;
2820 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2821 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2824 /* IDL is initially empty, zero out at least the length */
2825 pgs = (pgno_t *)data.mv_data;
2826 j = head_room > clean_limit ? head_room : 0;
2830 total_room += head_room;
2833 /* Fill in the reserved me_pghead records */
2839 rc = mdb_cursor_first(&mc, &key, &data);
2840 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2841 txnid_t id = *(txnid_t *)key.mv_data;
2842 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2845 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2847 if (len > mop_len) {
2849 data.mv_size = (len + 1) * sizeof(MDB_ID);
2851 data.mv_data = mop -= len;
2854 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2856 if (rc || !(mop_len -= len))
2863 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2864 * @param[in] txn the transaction that's being committed
2865 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2866 * @return 0 on success, non-zero on failure.
2869 mdb_page_flush(MDB_txn *txn, int keep)
2871 MDB_env *env = txn->mt_env;
2872 MDB_ID2L dl = txn->mt_u.dirty_list;
2873 unsigned psize = env->me_psize, j;
2874 int i, pagecount = dl[0].mid, rc;
2875 size_t size = 0, pos = 0;
2877 MDB_page *dp = NULL;
2881 struct iovec iov[MDB_COMMIT_PAGES];
2882 ssize_t wpos = 0, wsize = 0, wres;
2883 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2889 if (env->me_flags & MDB_WRITEMAP) {
2890 /* Clear dirty flags */
2891 while (++i <= pagecount) {
2893 /* Don't flush this page yet */
2894 if (dp->mp_flags & P_KEEP) {
2895 dp->mp_flags ^= P_KEEP;
2899 dp->mp_flags &= ~P_DIRTY;
2904 /* Write the pages */
2906 if (++i <= pagecount) {
2908 /* Don't flush this page yet */
2909 if (dp->mp_flags & P_KEEP) {
2910 dp->mp_flags ^= P_KEEP;
2915 /* clear dirty flag */
2916 dp->mp_flags &= ~P_DIRTY;
2919 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2924 /* Windows actually supports scatter/gather I/O, but only on
2925 * unbuffered file handles. Since we're relying on the OS page
2926 * cache for all our data, that's self-defeating. So we just
2927 * write pages one at a time. We use the ov structure to set
2928 * the write offset, to at least save the overhead of a Seek
2931 DPRINTF(("committing page %"Z"u", pgno));
2932 memset(&ov, 0, sizeof(ov));
2933 ov.Offset = pos & 0xffffffff;
2934 ov.OffsetHigh = pos >> 16 >> 16;
2935 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2937 DPRINTF(("WriteFile: %d", rc));
2941 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2942 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2944 /* Write previous page(s) */
2945 #ifdef MDB_USE_PWRITEV
2946 wres = pwritev(env->me_fd, iov, n, wpos);
2949 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2951 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2953 DPRINTF(("lseek: %s", strerror(rc)));
2956 wres = writev(env->me_fd, iov, n);
2959 if (wres != wsize) {
2962 DPRINTF(("Write error: %s", strerror(rc)));
2964 rc = EIO; /* TODO: Use which error code? */
2965 DPUTS("short write, filesystem full?");
2976 DPRINTF(("committing page %"Z"u", pgno));
2977 next_pos = pos + size;
2978 iov[n].iov_len = size;
2979 iov[n].iov_base = (char *)dp;
2985 for (i = keep; ++i <= pagecount; ) {
2987 /* This is a page we skipped above */
2990 dl[j].mid = dp->mp_pgno;
2993 mdb_dpage_free(env, dp);
2998 txn->mt_dirty_room += i - j;
3004 mdb_txn_commit(MDB_txn *txn)
3010 if (txn == NULL || txn->mt_env == NULL)
3013 if (txn->mt_child) {
3014 rc = mdb_txn_commit(txn->mt_child);
3015 txn->mt_child = NULL;
3022 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3023 mdb_dbis_update(txn, 1);
3024 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3029 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3030 DPUTS("error flag is set, can't commit");
3032 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3037 if (txn->mt_parent) {
3038 MDB_txn *parent = txn->mt_parent;
3041 unsigned x, y, len, ps_len;
3043 /* Append our free list to parent's */
3044 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3047 mdb_midl_free(txn->mt_free_pgs);
3048 /* Failures after this must either undo the changes
3049 * to the parent or set MDB_TXN_ERROR in the parent.
3052 parent->mt_next_pgno = txn->mt_next_pgno;
3053 parent->mt_flags = txn->mt_flags;
3055 /* Merge our cursors into parent's and close them */
3056 mdb_cursors_close(txn, 1);
3058 /* Update parent's DB table. */
3059 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3060 parent->mt_numdbs = txn->mt_numdbs;
3061 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3062 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3063 for (i=2; i<txn->mt_numdbs; i++) {
3064 /* preserve parent's DB_NEW status */
3065 x = parent->mt_dbflags[i] & DB_NEW;
3066 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3069 dst = parent->mt_u.dirty_list;
3070 src = txn->mt_u.dirty_list;
3071 /* Remove anything in our dirty list from parent's spill list */
3072 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3074 pspill[0] = (pgno_t)-1;
3075 /* Mark our dirty pages as deleted in parent spill list */
3076 for (i=0, len=src[0].mid; ++i <= len; ) {
3077 MDB_ID pn = src[i].mid << 1;
3078 while (pn > pspill[x])
3080 if (pn == pspill[x]) {
3085 /* Squash deleted pagenums if we deleted any */
3086 for (x=y; ++x <= ps_len; )
3087 if (!(pspill[x] & 1))
3088 pspill[++y] = pspill[x];
3092 /* Find len = length of merging our dirty list with parent's */
3094 dst[0].mid = 0; /* simplify loops */
3095 if (parent->mt_parent) {
3096 len = x + src[0].mid;
3097 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3098 for (i = x; y && i; y--) {
3099 pgno_t yp = src[y].mid;
3100 while (yp < dst[i].mid)
3102 if (yp == dst[i].mid) {
3107 } else { /* Simplify the above for single-ancestor case */
3108 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3110 /* Merge our dirty list with parent's */
3112 for (i = len; y; dst[i--] = src[y--]) {
3113 pgno_t yp = src[y].mid;
3114 while (yp < dst[x].mid)
3115 dst[i--] = dst[x--];
3116 if (yp == dst[x].mid)
3117 free(dst[x--].mptr);
3119 mdb_tassert(txn, i == x);
3121 free(txn->mt_u.dirty_list);
3122 parent->mt_dirty_room = txn->mt_dirty_room;
3123 if (txn->mt_spill_pgs) {
3124 if (parent->mt_spill_pgs) {
3125 /* TODO: Prevent failure here, so parent does not fail */
3126 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3128 parent->mt_flags |= MDB_TXN_ERROR;
3129 mdb_midl_free(txn->mt_spill_pgs);
3130 mdb_midl_sort(parent->mt_spill_pgs);
3132 parent->mt_spill_pgs = txn->mt_spill_pgs;
3136 parent->mt_child = NULL;
3137 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3142 if (txn != env->me_txn) {
3143 DPUTS("attempt to commit unknown transaction");
3148 mdb_cursors_close(txn, 0);
3150 if (!txn->mt_u.dirty_list[0].mid &&
3151 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3154 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3155 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3157 /* Update DB root pointers */
3158 if (txn->mt_numdbs > 2) {
3162 data.mv_size = sizeof(MDB_db);
3164 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3165 for (i = 2; i < txn->mt_numdbs; i++) {
3166 if (txn->mt_dbflags[i] & DB_DIRTY) {
3167 data.mv_data = &txn->mt_dbs[i];
3168 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3175 rc = mdb_freelist_save(txn);
3179 mdb_midl_free(env->me_pghead);
3180 env->me_pghead = NULL;
3181 if (mdb_midl_shrink(&txn->mt_free_pgs))
3182 env->me_free_pgs = txn->mt_free_pgs;
3188 if ((rc = mdb_page_flush(txn, 0)) ||
3189 (rc = mdb_env_sync(env, 0)) ||
3190 (rc = mdb_env_write_meta(txn)))
3196 mdb_dbis_update(txn, 1);
3199 UNLOCK_MUTEX_W(env);
3209 /** Read the environment parameters of a DB environment before
3210 * mapping it into memory.
3211 * @param[in] env the environment handle
3212 * @param[out] meta address of where to store the meta information
3213 * @return 0 on success, non-zero on failure.
3216 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3222 enum { Size = sizeof(pbuf) };
3224 /* We don't know the page size yet, so use a minimum value.
3225 * Read both meta pages so we can use the latest one.
3228 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3232 memset(&ov, 0, sizeof(ov));
3234 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3235 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3238 rc = pread(env->me_fd, &pbuf, Size, off);
3241 if (rc == 0 && off == 0)
3243 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3244 DPRINTF(("read: %s", mdb_strerror(rc)));
3248 p = (MDB_page *)&pbuf;
3250 if (!F_ISSET(p->mp_flags, P_META)) {
3251 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3256 if (m->mm_magic != MDB_MAGIC) {
3257 DPUTS("meta has invalid magic");
3261 if (m->mm_version != MDB_DATA_VERSION) {
3262 DPRINTF(("database is version %u, expected version %u",
3263 m->mm_version, MDB_DATA_VERSION));
3264 return MDB_VERSION_MISMATCH;
3267 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3273 /** Write the environment parameters of a freshly created DB environment.
3274 * @param[in] env the environment handle
3275 * @param[out] meta address of where to store the meta information
3276 * @return 0 on success, non-zero on failure.
3279 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3287 memset(&ov, 0, sizeof(ov));
3288 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3290 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3293 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3294 len = pwrite(fd, ptr, size, pos); \
3295 rc = (len >= 0); } while(0)
3298 DPUTS("writing new meta page");
3300 psize = env->me_psize;
3302 meta->mm_magic = MDB_MAGIC;
3303 meta->mm_version = MDB_DATA_VERSION;
3304 meta->mm_mapsize = env->me_mapsize;
3305 meta->mm_psize = psize;
3306 meta->mm_last_pg = 1;
3307 meta->mm_flags = env->me_flags & 0xffff;
3308 meta->mm_flags |= MDB_INTEGERKEY;
3309 meta->mm_dbs[0].md_root = P_INVALID;
3310 meta->mm_dbs[1].md_root = P_INVALID;
3312 p = calloc(2, psize);
3314 p->mp_flags = P_META;
3315 *(MDB_meta *)METADATA(p) = *meta;
3317 q = (MDB_page *)((char *)p + psize);
3319 q->mp_flags = P_META;
3320 *(MDB_meta *)METADATA(q) = *meta;
3322 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3325 else if ((unsigned) len == psize * 2)
3333 /** Update the environment info to commit a transaction.
3334 * @param[in] txn the transaction that's being committed
3335 * @return 0 on success, non-zero on failure.
3338 mdb_env_write_meta(MDB_txn *txn)
3341 MDB_meta meta, metab, *mp;
3343 int rc, len, toggle;
3352 toggle = txn->mt_txnid & 1;
3353 DPRINTF(("writing meta page %d for root page %"Z"u",
3354 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3357 mp = env->me_metas[toggle];
3359 if (env->me_flags & MDB_WRITEMAP) {
3360 /* Persist any increases of mapsize config */
3361 if (env->me_mapsize > mp->mm_mapsize)
3362 mp->mm_mapsize = env->me_mapsize;
3363 mp->mm_dbs[0] = txn->mt_dbs[0];
3364 mp->mm_dbs[1] = txn->mt_dbs[1];
3365 mp->mm_last_pg = txn->mt_next_pgno - 1;
3366 mp->mm_txnid = txn->mt_txnid;
3367 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3368 unsigned meta_size = env->me_psize;
3369 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3372 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3373 if (meta_size < env->me_os_psize)
3374 meta_size += meta_size;
3379 if (MDB_MSYNC(ptr, meta_size, rc)) {
3386 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3387 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3389 ptr = (char *)&meta;
3390 if (env->me_mapsize > mp->mm_mapsize) {
3391 /* Persist any increases of mapsize config */
3392 meta.mm_mapsize = env->me_mapsize;
3393 off = offsetof(MDB_meta, mm_mapsize);
3395 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3397 len = sizeof(MDB_meta) - off;
3400 meta.mm_dbs[0] = txn->mt_dbs[0];
3401 meta.mm_dbs[1] = txn->mt_dbs[1];
3402 meta.mm_last_pg = txn->mt_next_pgno - 1;
3403 meta.mm_txnid = txn->mt_txnid;
3406 off += env->me_psize;
3409 /* Write to the SYNC fd */
3410 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3411 env->me_fd : env->me_mfd;
3414 memset(&ov, 0, sizeof(ov));
3416 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3420 rc = pwrite(mfd, ptr, len, off);
3423 rc = rc < 0 ? ErrCode() : EIO;
3424 DPUTS("write failed, disk error?");
3425 /* On a failure, the pagecache still contains the new data.
3426 * Write some old data back, to prevent it from being used.
3427 * Use the non-SYNC fd; we know it will fail anyway.
3429 meta.mm_last_pg = metab.mm_last_pg;
3430 meta.mm_txnid = metab.mm_txnid;
3432 memset(&ov, 0, sizeof(ov));
3434 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3436 r2 = pwrite(env->me_fd, ptr, len, off);
3437 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3440 env->me_flags |= MDB_FATAL_ERROR;
3444 /* Memory ordering issues are irrelevant; since the entire writer
3445 * is wrapped by wmutex, all of these changes will become visible
3446 * after the wmutex is unlocked. Since the DB is multi-version,
3447 * readers will get consistent data regardless of how fresh or
3448 * how stale their view of these values is.
3451 env->me_txns->mti_txnid = txn->mt_txnid;
3456 /** Check both meta pages to see which one is newer.
3457 * @param[in] env the environment handle
3458 * @return meta toggle (0 or 1).
3461 mdb_env_pick_meta(const MDB_env *env)
3463 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3467 mdb_env_create(MDB_env **env)
3471 e = calloc(1, sizeof(MDB_env));
3475 e->me_maxreaders = DEFAULT_READERS;
3476 e->me_maxdbs = e->me_numdbs = 2;
3477 e->me_fd = INVALID_HANDLE_VALUE;
3478 e->me_lfd = INVALID_HANDLE_VALUE;
3479 e->me_mfd = INVALID_HANDLE_VALUE;
3480 #ifdef MDB_USE_POSIX_SEM
3481 e->me_rmutex = SEM_FAILED;
3482 e->me_wmutex = SEM_FAILED;
3484 e->me_pid = getpid();
3485 GET_PAGESIZE(e->me_os_psize);
3486 VGMEMP_CREATE(e,0,0);
3492 mdb_env_map(MDB_env *env, void *addr, int newsize)
3495 unsigned int flags = env->me_flags;
3499 LONG sizelo, sizehi;
3500 sizelo = env->me_mapsize & 0xffffffff;
3501 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3503 /* Windows won't create mappings for zero length files.
3504 * Just allocate the maxsize right now.
3507 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3508 || !SetEndOfFile(env->me_fd)
3509 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3512 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3513 PAGE_READWRITE : PAGE_READONLY,
3514 sizehi, sizelo, NULL);
3517 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3518 FILE_MAP_WRITE : FILE_MAP_READ,
3519 0, 0, env->me_mapsize, addr);
3520 rc = env->me_map ? 0 : ErrCode();
3525 int prot = PROT_READ;
3526 if (flags & MDB_WRITEMAP) {
3528 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3531 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3533 if (env->me_map == MAP_FAILED) {
3538 if (flags & MDB_NORDAHEAD) {
3539 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3541 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3543 #ifdef POSIX_MADV_RANDOM
3544 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3545 #endif /* POSIX_MADV_RANDOM */
3546 #endif /* MADV_RANDOM */
3550 /* Can happen because the address argument to mmap() is just a
3551 * hint. mmap() can pick another, e.g. if the range is in use.
3552 * The MAP_FIXED flag would prevent that, but then mmap could
3553 * instead unmap existing pages to make room for the new map.
3555 if (addr && env->me_map != addr)
3556 return EBUSY; /* TODO: Make a new MDB_* error code? */
3558 p = (MDB_page *)env->me_map;
3559 env->me_metas[0] = METADATA(p);
3560 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3566 mdb_env_set_mapsize(MDB_env *env, size_t size)
3568 /* If env is already open, caller is responsible for making
3569 * sure there are no active txns.
3577 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3578 else if (size < env->me_mapsize) {
3579 /* If the configured size is smaller, make sure it's
3580 * still big enough. Silently round up to minimum if not.
3582 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3586 munmap(env->me_map, env->me_mapsize);
3587 env->me_mapsize = size;
3588 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3589 rc = mdb_env_map(env, old, 1);
3593 env->me_mapsize = size;
3595 env->me_maxpg = env->me_mapsize / env->me_psize;
3600 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3604 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3609 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3611 if (env->me_map || readers < 1)
3613 env->me_maxreaders = readers;
3618 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3620 if (!env || !readers)
3622 *readers = env->me_maxreaders;
3626 /** Further setup required for opening an MDB environment
3629 mdb_env_open2(MDB_env *env)
3631 unsigned int flags = env->me_flags;
3632 int i, newenv = 0, rc;
3636 /* See if we should use QueryLimited */
3638 if ((rc & 0xff) > 5)
3639 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3641 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3644 memset(&meta, 0, sizeof(meta));
3646 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3649 DPUTS("new mdbenv");
3651 env->me_psize = env->me_os_psize;
3652 if (env->me_psize > MAX_PAGESIZE)
3653 env->me_psize = MAX_PAGESIZE;
3655 env->me_psize = meta.mm_psize;
3658 /* Was a mapsize configured? */
3659 if (!env->me_mapsize) {
3660 /* If this is a new environment, take the default,
3661 * else use the size recorded in the existing env.
3663 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3664 } else if (env->me_mapsize < meta.mm_mapsize) {
3665 /* If the configured size is smaller, make sure it's
3666 * still big enough. Silently round up to minimum if not.
3668 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3669 if (env->me_mapsize < minsize)
3670 env->me_mapsize = minsize;
3673 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3678 if (flags & MDB_FIXEDMAP)
3679 meta.mm_address = env->me_map;
3680 i = mdb_env_init_meta(env, &meta);
3681 if (i != MDB_SUCCESS) {
3686 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3687 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3689 #if !(MDB_MAXKEYSIZE)
3690 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3692 env->me_maxpg = env->me_mapsize / env->me_psize;
3696 int toggle = mdb_env_pick_meta(env);
3697 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3699 DPRINTF(("opened database version %u, pagesize %u",
3700 env->me_metas[0]->mm_version, env->me_psize));
3701 DPRINTF(("using meta page %d", toggle));
3702 DPRINTF(("depth: %u", db->md_depth));
3703 DPRINTF(("entries: %"Z"u", db->md_entries));
3704 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3705 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3706 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3707 DPRINTF(("root: %"Z"u", db->md_root));
3715 /** Release a reader thread's slot in the reader lock table.
3716 * This function is called automatically when a thread exits.
3717 * @param[in] ptr This points to the slot in the reader lock table.
3720 mdb_env_reader_dest(void *ptr)
3722 MDB_reader *reader = ptr;
3728 /** Junk for arranging thread-specific callbacks on Windows. This is
3729 * necessarily platform and compiler-specific. Windows supports up
3730 * to 1088 keys. Let's assume nobody opens more than 64 environments
3731 * in a single process, for now. They can override this if needed.
3733 #ifndef MAX_TLS_KEYS
3734 #define MAX_TLS_KEYS 64
3736 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3737 static int mdb_tls_nkeys;
3739 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3743 case DLL_PROCESS_ATTACH: break;
3744 case DLL_THREAD_ATTACH: break;
3745 case DLL_THREAD_DETACH:
3746 for (i=0; i<mdb_tls_nkeys; i++) {
3747 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3749 mdb_env_reader_dest(r);
3753 case DLL_PROCESS_DETACH: break;
3758 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3760 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3764 /* Force some symbol references.
3765 * _tls_used forces the linker to create the TLS directory if not already done
3766 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3768 #pragma comment(linker, "/INCLUDE:_tls_used")
3769 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3770 #pragma const_seg(".CRT$XLB")
3771 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3772 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3775 #pragma comment(linker, "/INCLUDE:__tls_used")
3776 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3777 #pragma data_seg(".CRT$XLB")
3778 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3780 #endif /* WIN 32/64 */
3781 #endif /* !__GNUC__ */
3784 /** Downgrade the exclusive lock on the region back to shared */
3786 mdb_env_share_locks(MDB_env *env, int *excl)
3788 int rc = 0, toggle = mdb_env_pick_meta(env);
3790 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3795 /* First acquire a shared lock. The Unlock will
3796 * then release the existing exclusive lock.
3798 memset(&ov, 0, sizeof(ov));
3799 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3802 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3808 struct flock lock_info;
3809 /* The shared lock replaces the existing lock */
3810 memset((void *)&lock_info, 0, sizeof(lock_info));
3811 lock_info.l_type = F_RDLCK;
3812 lock_info.l_whence = SEEK_SET;
3813 lock_info.l_start = 0;
3814 lock_info.l_len = 1;
3815 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3816 (rc = ErrCode()) == EINTR) ;
3817 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3824 /** Try to get exlusive lock, otherwise shared.
3825 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3828 mdb_env_excl_lock(MDB_env *env, int *excl)
3832 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3836 memset(&ov, 0, sizeof(ov));
3837 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3844 struct flock lock_info;
3845 memset((void *)&lock_info, 0, sizeof(lock_info));
3846 lock_info.l_type = F_WRLCK;
3847 lock_info.l_whence = SEEK_SET;
3848 lock_info.l_start = 0;
3849 lock_info.l_len = 1;
3850 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3851 (rc = ErrCode()) == EINTR) ;
3855 # ifdef MDB_USE_POSIX_SEM
3856 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3859 lock_info.l_type = F_RDLCK;
3860 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3861 (rc = ErrCode()) == EINTR) ;
3871 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3873 * @(#) $Revision: 5.1 $
3874 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3875 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3877 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3881 * Please do not copyright this code. This code is in the public domain.
3883 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3884 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3885 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3886 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3887 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3888 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3889 * PERFORMANCE OF THIS SOFTWARE.
3892 * chongo <Landon Curt Noll> /\oo/\
3893 * http://www.isthe.com/chongo/
3895 * Share and Enjoy! :-)
3898 typedef unsigned long long mdb_hash_t;
3899 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3901 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3902 * @param[in] val value to hash
3903 * @param[in] hval initial value for hash
3904 * @return 64 bit hash
3906 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3907 * hval arg on the first call.
3910 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3912 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3913 unsigned char *end = s + val->mv_size;
3915 * FNV-1a hash each octet of the string
3918 /* xor the bottom with the current octet */
3919 hval ^= (mdb_hash_t)*s++;
3921 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3922 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3923 (hval << 7) + (hval << 8) + (hval << 40);
3925 /* return our new hash value */
3929 /** Hash the string and output the encoded hash.
3930 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3931 * very short name limits. We don't care about the encoding being reversible,
3932 * we just want to preserve as many bits of the input as possible in a
3933 * small printable string.
3934 * @param[in] str string to hash
3935 * @param[out] encbuf an array of 11 chars to hold the hash
3937 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3940 mdb_pack85(unsigned long l, char *out)
3944 for (i=0; i<5; i++) {
3945 *out++ = mdb_a85[l % 85];
3951 mdb_hash_enc(MDB_val *val, char *encbuf)
3953 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3955 mdb_pack85(h, encbuf);
3956 mdb_pack85(h>>32, encbuf+5);
3961 /** Open and/or initialize the lock region for the environment.
3962 * @param[in] env The MDB environment.
3963 * @param[in] lpath The pathname of the file used for the lock region.
3964 * @param[in] mode The Unix permissions for the file, if we create it.
3965 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3966 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3967 * @return 0 on success, non-zero on failure.
3970 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3973 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3975 # define MDB_ERRCODE_ROFS EROFS
3976 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3977 # define MDB_CLOEXEC O_CLOEXEC
3980 # define MDB_CLOEXEC 0
3987 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3988 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3989 FILE_ATTRIBUTE_NORMAL, NULL);
3991 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3993 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3995 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4000 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4001 /* Lose record locks when exec*() */
4002 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4003 fcntl(env->me_lfd, F_SETFD, fdflags);
4006 if (!(env->me_flags & MDB_NOTLS)) {
4007 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4010 env->me_flags |= MDB_ENV_TXKEY;
4012 /* Windows TLS callbacks need help finding their TLS info. */
4013 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4017 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4021 /* Try to get exclusive lock. If we succeed, then
4022 * nobody is using the lock region and we should initialize it.
4024 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4027 size = GetFileSize(env->me_lfd, NULL);
4029 size = lseek(env->me_lfd, 0, SEEK_END);
4030 if (size == -1) goto fail_errno;
4032 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4033 if (size < rsize && *excl > 0) {
4035 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4036 || !SetEndOfFile(env->me_lfd))
4039 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4043 size = rsize - sizeof(MDB_txninfo);
4044 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4049 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4051 if (!mh) goto fail_errno;
4052 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4054 if (!env->me_txns) goto fail_errno;
4056 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4058 if (m == MAP_FAILED) goto fail_errno;
4064 BY_HANDLE_FILE_INFORMATION stbuf;
4073 if (!mdb_sec_inited) {
4074 InitializeSecurityDescriptor(&mdb_null_sd,
4075 SECURITY_DESCRIPTOR_REVISION);
4076 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4077 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4078 mdb_all_sa.bInheritHandle = FALSE;
4079 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4082 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4083 idbuf.volume = stbuf.dwVolumeSerialNumber;
4084 idbuf.nhigh = stbuf.nFileIndexHigh;
4085 idbuf.nlow = stbuf.nFileIndexLow;
4086 val.mv_data = &idbuf;
4087 val.mv_size = sizeof(idbuf);
4088 mdb_hash_enc(&val, encbuf);
4089 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4090 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4091 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4092 if (!env->me_rmutex) goto fail_errno;
4093 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4094 if (!env->me_wmutex) goto fail_errno;
4095 #elif defined(MDB_USE_POSIX_SEM)
4104 #if defined(__NetBSD__)
4105 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4107 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4108 idbuf.dev = stbuf.st_dev;
4109 idbuf.ino = stbuf.st_ino;
4110 val.mv_data = &idbuf;
4111 val.mv_size = sizeof(idbuf);
4112 mdb_hash_enc(&val, encbuf);
4113 #ifdef MDB_SHORT_SEMNAMES
4114 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4116 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4117 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4118 /* Clean up after a previous run, if needed: Try to
4119 * remove both semaphores before doing anything else.
4121 sem_unlink(env->me_txns->mti_rmname);
4122 sem_unlink(env->me_txns->mti_wmname);
4123 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4124 O_CREAT|O_EXCL, mode, 1);
4125 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4126 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4127 O_CREAT|O_EXCL, mode, 1);
4128 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4129 #else /* MDB_USE_POSIX_SEM */
4130 pthread_mutexattr_t mattr;
4132 if ((rc = pthread_mutexattr_init(&mattr))
4133 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4134 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4135 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4137 pthread_mutexattr_destroy(&mattr);
4138 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4140 env->me_txns->mti_magic = MDB_MAGIC;
4141 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4142 env->me_txns->mti_txnid = 0;
4143 env->me_txns->mti_numreaders = 0;
4146 if (env->me_txns->mti_magic != MDB_MAGIC) {
4147 DPUTS("lock region has invalid magic");
4151 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4152 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4153 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4154 rc = MDB_VERSION_MISMATCH;
4158 if (rc && rc != EACCES && rc != EAGAIN) {
4162 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4163 if (!env->me_rmutex) goto fail_errno;
4164 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4165 if (!env->me_wmutex) goto fail_errno;
4166 #elif defined(MDB_USE_POSIX_SEM)
4167 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4168 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4169 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4170 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4181 /** The name of the lock file in the DB environment */
4182 #define LOCKNAME "/lock.mdb"
4183 /** The name of the data file in the DB environment */
4184 #define DATANAME "/data.mdb"
4185 /** The suffix of the lock file when no subdir is used */
4186 #define LOCKSUFF "-lock"
4187 /** Only a subset of the @ref mdb_env flags can be changed
4188 * at runtime. Changing other flags requires closing the
4189 * environment and re-opening it with the new flags.
4191 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4192 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4193 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4195 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4196 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4200 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4202 int oflags, rc, len, excl = -1;
4203 char *lpath, *dpath;
4205 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4209 if (flags & MDB_NOSUBDIR) {
4210 rc = len + sizeof(LOCKSUFF) + len + 1;
4212 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4217 if (flags & MDB_NOSUBDIR) {
4218 dpath = lpath + len + sizeof(LOCKSUFF);
4219 sprintf(lpath, "%s" LOCKSUFF, path);
4220 strcpy(dpath, path);
4222 dpath = lpath + len + sizeof(LOCKNAME);
4223 sprintf(lpath, "%s" LOCKNAME, path);
4224 sprintf(dpath, "%s" DATANAME, path);
4228 flags |= env->me_flags;
4229 if (flags & MDB_RDONLY) {
4230 /* silently ignore WRITEMAP when we're only getting read access */
4231 flags &= ~MDB_WRITEMAP;
4233 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4234 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4237 env->me_flags = flags |= MDB_ENV_ACTIVE;
4241 env->me_path = strdup(path);
4242 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4243 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4244 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4249 /* For RDONLY, get lockfile after we know datafile exists */
4250 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4251 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4257 if (F_ISSET(flags, MDB_RDONLY)) {
4258 oflags = GENERIC_READ;
4259 len = OPEN_EXISTING;
4261 oflags = GENERIC_READ|GENERIC_WRITE;
4264 mode = FILE_ATTRIBUTE_NORMAL;
4265 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4266 NULL, len, mode, NULL);
4268 if (F_ISSET(flags, MDB_RDONLY))
4271 oflags = O_RDWR | O_CREAT;
4273 env->me_fd = open(dpath, oflags, mode);
4275 if (env->me_fd == INVALID_HANDLE_VALUE) {
4280 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4281 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4286 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4287 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4288 env->me_mfd = env->me_fd;
4290 /* Synchronous fd for meta writes. Needed even with
4291 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4294 len = OPEN_EXISTING;
4295 env->me_mfd = CreateFile(dpath, oflags,
4296 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4297 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4300 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4302 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4307 DPRINTF(("opened dbenv %p", (void *) env));
4309 rc = mdb_env_share_locks(env, &excl);
4313 if (!((flags & MDB_RDONLY) ||
4314 (env->me_pbuf = calloc(1, env->me_psize))))
4320 mdb_env_close0(env, excl);
4326 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4328 mdb_env_close0(MDB_env *env, int excl)
4332 if (!(env->me_flags & MDB_ENV_ACTIVE))
4335 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4336 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4337 free(env->me_dbxs[i].md_name.mv_data);
4340 free(env->me_dbflags);
4343 free(env->me_dirty_list);
4344 mdb_midl_free(env->me_free_pgs);
4346 if (env->me_flags & MDB_ENV_TXKEY) {
4347 pthread_key_delete(env->me_txkey);
4349 /* Delete our key from the global list */
4350 for (i=0; i<mdb_tls_nkeys; i++)
4351 if (mdb_tls_keys[i] == env->me_txkey) {
4352 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4360 munmap(env->me_map, env->me_mapsize);
4362 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4363 (void) close(env->me_mfd);
4364 if (env->me_fd != INVALID_HANDLE_VALUE)
4365 (void) close(env->me_fd);
4367 MDB_PID_T pid = env->me_pid;
4368 /* Clearing readers is done in this function because
4369 * me_txkey with its destructor must be disabled first.
4371 for (i = env->me_numreaders; --i >= 0; )
4372 if (env->me_txns->mti_readers[i].mr_pid == pid)
4373 env->me_txns->mti_readers[i].mr_pid = 0;
4375 if (env->me_rmutex) {
4376 CloseHandle(env->me_rmutex);
4377 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4379 /* Windows automatically destroys the mutexes when
4380 * the last handle closes.
4382 #elif defined(MDB_USE_POSIX_SEM)
4383 if (env->me_rmutex != SEM_FAILED) {
4384 sem_close(env->me_rmutex);
4385 if (env->me_wmutex != SEM_FAILED)
4386 sem_close(env->me_wmutex);
4387 /* If we have the filelock: If we are the
4388 * only remaining user, clean up semaphores.
4391 mdb_env_excl_lock(env, &excl);
4393 sem_unlink(env->me_txns->mti_rmname);
4394 sem_unlink(env->me_txns->mti_wmname);
4398 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4400 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4403 /* Unlock the lockfile. Windows would have unlocked it
4404 * after closing anyway, but not necessarily at once.
4406 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4409 (void) close(env->me_lfd);
4412 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4416 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4418 MDB_txn *txn = NULL;
4424 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4428 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4431 /* Do the lock/unlock of the reader mutex before starting the
4432 * write txn. Otherwise other read txns could block writers.
4434 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4439 /* We must start the actual read txn after blocking writers */
4440 mdb_txn_reset0(txn, "reset-stage1");
4442 /* Temporarily block writers until we snapshot the meta pages */
4445 rc = mdb_txn_renew0(txn);
4447 UNLOCK_MUTEX_W(env);
4452 wsize = env->me_psize * 2;
4456 DO_WRITE(rc, fd, ptr, w2, len);
4460 } else if (len > 0) {
4466 /* Non-blocking or async handles are not supported */
4472 UNLOCK_MUTEX_W(env);
4477 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4479 if (wsize > MAX_WRITE)
4483 DO_WRITE(rc, fd, ptr, w2, len);
4487 } else if (len > 0) {
4504 mdb_env_copy(MDB_env *env, const char *path)
4508 HANDLE newfd = INVALID_HANDLE_VALUE;
4510 if (env->me_flags & MDB_NOSUBDIR) {
4511 lpath = (char *)path;
4514 len += sizeof(DATANAME);
4515 lpath = malloc(len);
4518 sprintf(lpath, "%s" DATANAME, path);
4521 /* The destination path must exist, but the destination file must not.
4522 * We don't want the OS to cache the writes, since the source data is
4523 * already in the OS cache.
4526 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4527 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4529 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4531 if (newfd == INVALID_HANDLE_VALUE) {
4537 /* Set O_DIRECT if the file system supports it */
4538 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4539 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4541 #ifdef F_NOCACHE /* __APPLE__ */
4542 rc = fcntl(newfd, F_NOCACHE, 1);
4549 rc = mdb_env_copyfd(env, newfd);
4552 if (!(env->me_flags & MDB_NOSUBDIR))
4554 if (newfd != INVALID_HANDLE_VALUE)
4555 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4562 mdb_env_close(MDB_env *env)
4569 VGMEMP_DESTROY(env);
4570 while ((dp = env->me_dpages) != NULL) {
4571 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4572 env->me_dpages = dp->mp_next;
4576 mdb_env_close0(env, 0);
4580 /** Compare two items pointing at aligned size_t's */
4582 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4584 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4585 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4588 /** Compare two items pointing at aligned unsigned int's */
4590 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4592 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4593 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4596 /** Compare two items pointing at unsigned ints of unknown alignment.
4597 * Nodes and keys are guaranteed to be 2-byte aligned.
4600 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4602 #if BYTE_ORDER == LITTLE_ENDIAN
4603 unsigned short *u, *c;
4606 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4607 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4610 } while(!x && u > (unsigned short *)a->mv_data);
4613 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4617 /** Compare two items pointing at size_t's of unknown alignment. */
4618 #ifdef MISALIGNED_OK
4619 # define mdb_cmp_clong mdb_cmp_long
4621 # define mdb_cmp_clong mdb_cmp_cint
4624 /** Compare two items lexically */
4626 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4633 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4639 diff = memcmp(a->mv_data, b->mv_data, len);
4640 return diff ? diff : len_diff<0 ? -1 : len_diff;
4643 /** Compare two items in reverse byte order */
4645 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4647 const unsigned char *p1, *p2, *p1_lim;
4651 p1_lim = (const unsigned char *)a->mv_data;
4652 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4653 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4655 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4661 while (p1 > p1_lim) {
4662 diff = *--p1 - *--p2;
4666 return len_diff<0 ? -1 : len_diff;
4669 /** Search for key within a page, using binary search.
4670 * Returns the smallest entry larger or equal to the key.
4671 * If exactp is non-null, stores whether the found entry was an exact match
4672 * in *exactp (1 or 0).
4673 * Updates the cursor index with the index of the found entry.
4674 * If no entry larger or equal to the key is found, returns NULL.
4677 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4679 unsigned int i = 0, nkeys;
4682 MDB_page *mp = mc->mc_pg[mc->mc_top];
4683 MDB_node *node = NULL;
4688 nkeys = NUMKEYS(mp);
4690 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4691 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4694 low = IS_LEAF(mp) ? 0 : 1;
4696 cmp = mc->mc_dbx->md_cmp;
4698 /* Branch pages have no data, so if using integer keys,
4699 * alignment is guaranteed. Use faster mdb_cmp_int.
4701 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4702 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4709 nodekey.mv_size = mc->mc_db->md_pad;
4710 node = NODEPTR(mp, 0); /* fake */
4711 while (low <= high) {
4712 i = (low + high) >> 1;
4713 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4714 rc = cmp(key, &nodekey);
4715 DPRINTF(("found leaf index %u [%s], rc = %i",
4716 i, DKEY(&nodekey), rc));
4725 while (low <= high) {
4726 i = (low + high) >> 1;
4728 node = NODEPTR(mp, i);
4729 nodekey.mv_size = NODEKSZ(node);
4730 nodekey.mv_data = NODEKEY(node);
4732 rc = cmp(key, &nodekey);
4735 DPRINTF(("found leaf index %u [%s], rc = %i",
4736 i, DKEY(&nodekey), rc));
4738 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4739 i, DKEY(&nodekey), NODEPGNO(node), rc));
4750 if (rc > 0) { /* Found entry is less than the key. */
4751 i++; /* Skip to get the smallest entry larger than key. */
4753 node = NODEPTR(mp, i);
4756 *exactp = (rc == 0 && nkeys > 0);
4757 /* store the key index */
4758 mc->mc_ki[mc->mc_top] = i;
4760 /* There is no entry larger or equal to the key. */
4763 /* nodeptr is fake for LEAF2 */
4769 mdb_cursor_adjust(MDB_cursor *mc, func)
4773 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4774 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4781 /** Pop a page off the top of the cursor's stack. */
4783 mdb_cursor_pop(MDB_cursor *mc)
4787 MDB_page *top = mc->mc_pg[mc->mc_top];
4793 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4794 DDBI(mc), (void *) mc));
4798 /** Push a page onto the top of the cursor's stack. */
4800 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4802 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4803 DDBI(mc), (void *) mc));
4805 if (mc->mc_snum >= CURSOR_STACK) {
4806 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4807 return MDB_CURSOR_FULL;
4810 mc->mc_top = mc->mc_snum++;
4811 mc->mc_pg[mc->mc_top] = mp;
4812 mc->mc_ki[mc->mc_top] = 0;
4817 /** Find the address of the page corresponding to a given page number.
4818 * @param[in] txn the transaction for this access.
4819 * @param[in] pgno the page number for the page to retrieve.
4820 * @param[out] ret address of a pointer where the page's address will be stored.
4821 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4822 * @return 0 on success, non-zero on failure.
4825 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4827 MDB_env *env = txn->mt_env;
4831 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4835 MDB_ID2L dl = tx2->mt_u.dirty_list;
4837 /* Spilled pages were dirtied in this txn and flushed
4838 * because the dirty list got full. Bring this page
4839 * back in from the map (but don't unspill it here,
4840 * leave that unless page_touch happens again).
4842 if (tx2->mt_spill_pgs) {
4843 MDB_ID pn = pgno << 1;
4844 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4845 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4846 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4851 unsigned x = mdb_mid2l_search(dl, pgno);
4852 if (x <= dl[0].mid && dl[x].mid == pgno) {
4858 } while ((tx2 = tx2->mt_parent) != NULL);
4861 if (pgno < txn->mt_next_pgno) {
4863 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4865 DPRINTF(("page %"Z"u not found", pgno));
4866 txn->mt_flags |= MDB_TXN_ERROR;
4867 return MDB_PAGE_NOTFOUND;
4877 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4878 * The cursor is at the root page, set up the rest of it.
4881 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4883 MDB_page *mp = mc->mc_pg[mc->mc_top];
4887 while (IS_BRANCH(mp)) {
4891 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4892 mdb_cassert(mc, NUMKEYS(mp) > 1);
4893 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4895 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4897 if (flags & MDB_PS_LAST)
4898 i = NUMKEYS(mp) - 1;
4901 node = mdb_node_search(mc, key, &exact);
4903 i = NUMKEYS(mp) - 1;
4905 i = mc->mc_ki[mc->mc_top];
4907 mdb_cassert(mc, i > 0);
4911 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4914 mdb_cassert(mc, i < NUMKEYS(mp));
4915 node = NODEPTR(mp, i);
4917 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4920 mc->mc_ki[mc->mc_top] = i;
4921 if ((rc = mdb_cursor_push(mc, mp)))
4924 if (flags & MDB_PS_MODIFY) {
4925 if ((rc = mdb_page_touch(mc)) != 0)
4927 mp = mc->mc_pg[mc->mc_top];
4932 DPRINTF(("internal error, index points to a %02X page!?",
4934 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4935 return MDB_CORRUPTED;
4938 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4939 key ? DKEY(key) : "null"));
4940 mc->mc_flags |= C_INITIALIZED;
4941 mc->mc_flags &= ~C_EOF;
4946 /** Search for the lowest key under the current branch page.
4947 * This just bypasses a NUMKEYS check in the current page
4948 * before calling mdb_page_search_root(), because the callers
4949 * are all in situations where the current page is known to
4953 mdb_page_search_lowest(MDB_cursor *mc)
4955 MDB_page *mp = mc->mc_pg[mc->mc_top];
4956 MDB_node *node = NODEPTR(mp, 0);
4959 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4962 mc->mc_ki[mc->mc_top] = 0;
4963 if ((rc = mdb_cursor_push(mc, mp)))
4965 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4968 /** Search for the page a given key should be in.
4969 * Push it and its parent pages on the cursor stack.
4970 * @param[in,out] mc the cursor for this operation.
4971 * @param[in] key the key to search for, or NULL for first/last page.
4972 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4973 * are touched (updated with new page numbers).
4974 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4975 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4976 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4977 * @return 0 on success, non-zero on failure.
4980 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4985 /* Make sure the txn is still viable, then find the root from
4986 * the txn's db table and set it as the root of the cursor's stack.
4988 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4989 DPUTS("transaction has failed, must abort");
4992 /* Make sure we're using an up-to-date root */
4993 if (*mc->mc_dbflag & DB_STALE) {
4995 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4996 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5003 MDB_node *leaf = mdb_node_search(&mc2,
5004 &mc->mc_dbx->md_name, &exact);
5006 return MDB_NOTFOUND;
5007 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5010 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5012 /* The txn may not know this DBI, or another process may
5013 * have dropped and recreated the DB with other flags.
5015 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5016 return MDB_INCOMPATIBLE;
5017 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5019 *mc->mc_dbflag &= ~DB_STALE;
5021 root = mc->mc_db->md_root;
5023 if (root == P_INVALID) { /* Tree is empty. */
5024 DPUTS("tree is empty");
5025 return MDB_NOTFOUND;
5029 mdb_cassert(mc, root > 1);
5030 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5031 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5037 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5038 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5040 if (flags & MDB_PS_MODIFY) {
5041 if ((rc = mdb_page_touch(mc)))
5045 if (flags & MDB_PS_ROOTONLY)
5048 return mdb_page_search_root(mc, key, flags);
5052 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5054 MDB_txn *txn = mc->mc_txn;
5055 pgno_t pg = mp->mp_pgno;
5056 unsigned x = 0, ovpages = mp->mp_pages;
5057 MDB_env *env = txn->mt_env;
5058 MDB_IDL sl = txn->mt_spill_pgs;
5059 MDB_ID pn = pg << 1;
5062 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5063 /* If the page is dirty or on the spill list we just acquired it,
5064 * so we should give it back to our current free list, if any.
5065 * Otherwise put it onto the list of pages we freed in this txn.
5067 * Won't create me_pghead: me_pglast must be inited along with it.
5068 * Unsupported in nested txns: They would need to hide the page
5069 * range in ancestor txns' dirty and spilled lists.
5071 if (env->me_pghead &&
5073 ((mp->mp_flags & P_DIRTY) ||
5074 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5078 MDB_ID2 *dl, ix, iy;
5079 rc = mdb_midl_need(&env->me_pghead, ovpages);
5082 if (!(mp->mp_flags & P_DIRTY)) {
5083 /* This page is no longer spilled */
5090 /* Remove from dirty list */
5091 dl = txn->mt_u.dirty_list;
5093 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5099 mdb_cassert(mc, x > 1);
5101 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5102 txn->mt_flags |= MDB_TXN_ERROR;
5103 return MDB_CORRUPTED;
5106 if (!(env->me_flags & MDB_WRITEMAP))
5107 mdb_dpage_free(env, mp);
5109 /* Insert in me_pghead */
5110 mop = env->me_pghead;
5111 j = mop[0] + ovpages;
5112 for (i = mop[0]; i && mop[i] < pg; i--)
5118 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5122 mc->mc_db->md_overflow_pages -= ovpages;
5126 /** Return the data associated with a given node.
5127 * @param[in] txn The transaction for this operation.
5128 * @param[in] leaf The node being read.
5129 * @param[out] data Updated to point to the node's data.
5130 * @return 0 on success, non-zero on failure.
5133 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5135 MDB_page *omp; /* overflow page */
5139 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5140 data->mv_size = NODEDSZ(leaf);
5141 data->mv_data = NODEDATA(leaf);
5145 /* Read overflow data.
5147 data->mv_size = NODEDSZ(leaf);
5148 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5149 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5150 DPRINTF(("read overflow page %"Z"u failed", pgno));
5153 data->mv_data = METADATA(omp);
5159 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5160 MDB_val *key, MDB_val *data)
5167 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5169 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5172 if (txn->mt_flags & MDB_TXN_ERROR)
5175 mdb_cursor_init(&mc, txn, dbi, &mx);
5176 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5179 /** Find a sibling for a page.
5180 * Replaces the page at the top of the cursor's stack with the
5181 * specified sibling, if one exists.
5182 * @param[in] mc The cursor for this operation.
5183 * @param[in] move_right Non-zero if the right sibling is requested,
5184 * otherwise the left sibling.
5185 * @return 0 on success, non-zero on failure.
5188 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5194 if (mc->mc_snum < 2) {
5195 return MDB_NOTFOUND; /* root has no siblings */
5199 DPRINTF(("parent page is page %"Z"u, index %u",
5200 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5202 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5203 : (mc->mc_ki[mc->mc_top] == 0)) {
5204 DPRINTF(("no more keys left, moving to %s sibling",
5205 move_right ? "right" : "left"));
5206 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5207 /* undo cursor_pop before returning */
5214 mc->mc_ki[mc->mc_top]++;
5216 mc->mc_ki[mc->mc_top]--;
5217 DPRINTF(("just moving to %s index key %u",
5218 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5220 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5222 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5223 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5224 /* mc will be inconsistent if caller does mc_snum++ as above */
5225 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5229 mdb_cursor_push(mc, mp);
5231 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5236 /** Move the cursor to the next data item. */
5238 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5244 if (mc->mc_flags & C_EOF) {
5245 return MDB_NOTFOUND;
5248 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5250 mp = mc->mc_pg[mc->mc_top];
5252 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5253 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5254 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5255 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5256 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5257 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5258 if (rc == MDB_SUCCESS)
5259 MDB_GET_KEY(leaf, key);
5264 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5265 if (op == MDB_NEXT_DUP)
5266 return MDB_NOTFOUND;
5270 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5271 mdb_dbg_pgno(mp), (void *) mc));
5272 if (mc->mc_flags & C_DEL)
5275 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5276 DPUTS("=====> move to next sibling page");
5277 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5278 mc->mc_flags |= C_EOF;
5281 mp = mc->mc_pg[mc->mc_top];
5282 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5284 mc->mc_ki[mc->mc_top]++;
5287 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5288 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5291 key->mv_size = mc->mc_db->md_pad;
5292 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5296 mdb_cassert(mc, IS_LEAF(mp));
5297 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5299 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5300 mdb_xcursor_init1(mc, leaf);
5303 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5306 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5307 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5308 if (rc != MDB_SUCCESS)
5313 MDB_GET_KEY(leaf, key);
5317 /** Move the cursor to the previous data item. */
5319 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5325 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5327 mp = mc->mc_pg[mc->mc_top];
5329 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5330 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5331 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5332 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5333 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5334 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5335 if (rc == MDB_SUCCESS)
5336 MDB_GET_KEY(leaf, key);
5340 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5341 if (op == MDB_PREV_DUP)
5342 return MDB_NOTFOUND;
5347 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5348 mdb_dbg_pgno(mp), (void *) mc));
5350 if (mc->mc_ki[mc->mc_top] == 0) {
5351 DPUTS("=====> move to prev sibling page");
5352 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5355 mp = mc->mc_pg[mc->mc_top];
5356 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5357 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5359 mc->mc_ki[mc->mc_top]--;
5361 mc->mc_flags &= ~C_EOF;
5363 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5364 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5367 key->mv_size = mc->mc_db->md_pad;
5368 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5372 mdb_cassert(mc, IS_LEAF(mp));
5373 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5375 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5376 mdb_xcursor_init1(mc, leaf);
5379 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5382 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5383 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5384 if (rc != MDB_SUCCESS)
5389 MDB_GET_KEY(leaf, key);
5393 /** Set the cursor on a specific data item. */
5395 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5396 MDB_cursor_op op, int *exactp)
5400 MDB_node *leaf = NULL;
5403 if (key->mv_size == 0)
5404 return MDB_BAD_VALSIZE;
5407 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5409 /* See if we're already on the right page */
5410 if (mc->mc_flags & C_INITIALIZED) {
5413 mp = mc->mc_pg[mc->mc_top];
5415 mc->mc_ki[mc->mc_top] = 0;
5416 return MDB_NOTFOUND;
5418 if (mp->mp_flags & P_LEAF2) {
5419 nodekey.mv_size = mc->mc_db->md_pad;
5420 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5422 leaf = NODEPTR(mp, 0);
5423 MDB_GET_KEY2(leaf, nodekey);
5425 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5427 /* Probably happens rarely, but first node on the page
5428 * was the one we wanted.
5430 mc->mc_ki[mc->mc_top] = 0;
5437 unsigned int nkeys = NUMKEYS(mp);
5439 if (mp->mp_flags & P_LEAF2) {
5440 nodekey.mv_data = LEAF2KEY(mp,
5441 nkeys-1, nodekey.mv_size);
5443 leaf = NODEPTR(mp, nkeys-1);
5444 MDB_GET_KEY2(leaf, nodekey);
5446 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5448 /* last node was the one we wanted */
5449 mc->mc_ki[mc->mc_top] = nkeys-1;
5455 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5456 /* This is definitely the right page, skip search_page */
5457 if (mp->mp_flags & P_LEAF2) {
5458 nodekey.mv_data = LEAF2KEY(mp,
5459 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5461 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5462 MDB_GET_KEY2(leaf, nodekey);
5464 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5466 /* current node was the one we wanted */
5476 /* If any parents have right-sibs, search.
5477 * Otherwise, there's nothing further.
5479 for (i=0; i<mc->mc_top; i++)
5481 NUMKEYS(mc->mc_pg[i])-1)
5483 if (i == mc->mc_top) {
5484 /* There are no other pages */
5485 mc->mc_ki[mc->mc_top] = nkeys;
5486 return MDB_NOTFOUND;
5490 /* There are no other pages */
5491 mc->mc_ki[mc->mc_top] = 0;
5492 if (op == MDB_SET_RANGE && !exactp) {
5496 return MDB_NOTFOUND;
5500 rc = mdb_page_search(mc, key, 0);
5501 if (rc != MDB_SUCCESS)
5504 mp = mc->mc_pg[mc->mc_top];
5505 mdb_cassert(mc, IS_LEAF(mp));
5508 leaf = mdb_node_search(mc, key, exactp);
5509 if (exactp != NULL && !*exactp) {
5510 /* MDB_SET specified and not an exact match. */
5511 return MDB_NOTFOUND;
5515 DPUTS("===> inexact leaf not found, goto sibling");
5516 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5517 return rc; /* no entries matched */
5518 mp = mc->mc_pg[mc->mc_top];
5519 mdb_cassert(mc, IS_LEAF(mp));
5520 leaf = NODEPTR(mp, 0);
5524 mc->mc_flags |= C_INITIALIZED;
5525 mc->mc_flags &= ~C_EOF;
5528 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5529 key->mv_size = mc->mc_db->md_pad;
5530 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5535 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5536 mdb_xcursor_init1(mc, leaf);
5539 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5540 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5541 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5544 if (op == MDB_GET_BOTH) {
5550 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5551 if (rc != MDB_SUCCESS)
5554 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5556 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5558 rc = mc->mc_dbx->md_dcmp(data, &d2);
5560 if (op == MDB_GET_BOTH || rc > 0)
5561 return MDB_NOTFOUND;
5568 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5569 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5574 /* The key already matches in all other cases */
5575 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5576 MDB_GET_KEY(leaf, key);
5577 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5582 /** Move the cursor to the first item in the database. */
5584 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5590 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5592 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5593 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5594 if (rc != MDB_SUCCESS)
5597 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5599 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5600 mc->mc_flags |= C_INITIALIZED;
5601 mc->mc_flags &= ~C_EOF;
5603 mc->mc_ki[mc->mc_top] = 0;
5605 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5606 key->mv_size = mc->mc_db->md_pad;
5607 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5612 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5613 mdb_xcursor_init1(mc, leaf);
5614 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5618 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5622 MDB_GET_KEY(leaf, key);
5626 /** Move the cursor to the last item in the database. */
5628 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5634 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5636 if (!(mc->mc_flags & C_EOF)) {
5638 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5639 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5640 if (rc != MDB_SUCCESS)
5643 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5646 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5647 mc->mc_flags |= C_INITIALIZED|C_EOF;
5648 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5650 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5651 key->mv_size = mc->mc_db->md_pad;
5652 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5657 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5658 mdb_xcursor_init1(mc, leaf);
5659 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5663 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5668 MDB_GET_KEY(leaf, key);
5673 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5678 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5683 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5687 case MDB_GET_CURRENT:
5688 if (!(mc->mc_flags & C_INITIALIZED)) {
5691 MDB_page *mp = mc->mc_pg[mc->mc_top];
5692 int nkeys = NUMKEYS(mp);
5693 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5694 mc->mc_ki[mc->mc_top] = nkeys;
5700 key->mv_size = mc->mc_db->md_pad;
5701 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5703 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5704 MDB_GET_KEY(leaf, key);
5706 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5707 if (mc->mc_flags & C_DEL)
5708 mdb_xcursor_init1(mc, leaf);
5709 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5711 rc = mdb_node_read(mc->mc_txn, leaf, data);
5718 case MDB_GET_BOTH_RANGE:
5723 if (mc->mc_xcursor == NULL) {
5724 rc = MDB_INCOMPATIBLE;
5734 rc = mdb_cursor_set(mc, key, data, op,
5735 op == MDB_SET_RANGE ? NULL : &exact);
5738 case MDB_GET_MULTIPLE:
5739 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5743 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5744 rc = MDB_INCOMPATIBLE;
5748 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5749 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5752 case MDB_NEXT_MULTIPLE:
5757 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5758 rc = MDB_INCOMPATIBLE;
5761 if (!(mc->mc_flags & C_INITIALIZED))
5762 rc = mdb_cursor_first(mc, key, data);
5764 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5765 if (rc == MDB_SUCCESS) {
5766 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5769 mx = &mc->mc_xcursor->mx_cursor;
5770 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5772 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5773 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5781 case MDB_NEXT_NODUP:
5782 if (!(mc->mc_flags & C_INITIALIZED))
5783 rc = mdb_cursor_first(mc, key, data);
5785 rc = mdb_cursor_next(mc, key, data, op);
5789 case MDB_PREV_NODUP:
5790 if (!(mc->mc_flags & C_INITIALIZED)) {
5791 rc = mdb_cursor_last(mc, key, data);
5794 mc->mc_flags |= C_INITIALIZED;
5795 mc->mc_ki[mc->mc_top]++;
5797 rc = mdb_cursor_prev(mc, key, data, op);
5800 rc = mdb_cursor_first(mc, key, data);
5803 mfunc = mdb_cursor_first;
5805 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5809 if (mc->mc_xcursor == NULL) {
5810 rc = MDB_INCOMPATIBLE;
5813 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5817 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5820 rc = mdb_cursor_last(mc, key, data);
5823 mfunc = mdb_cursor_last;
5826 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5831 if (mc->mc_flags & C_DEL)
5832 mc->mc_flags ^= C_DEL;
5837 /** Touch all the pages in the cursor stack. Set mc_top.
5838 * Makes sure all the pages are writable, before attempting a write operation.
5839 * @param[in] mc The cursor to operate on.
5842 mdb_cursor_touch(MDB_cursor *mc)
5844 int rc = MDB_SUCCESS;
5846 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5849 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5850 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5853 *mc->mc_dbflag |= DB_DIRTY;
5858 rc = mdb_page_touch(mc);
5859 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5860 mc->mc_top = mc->mc_snum-1;
5865 /** Do not spill pages to disk if txn is getting full, may fail instead */
5866 #define MDB_NOSPILL 0x8000
5869 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5872 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5874 MDB_node *leaf = NULL;
5877 MDB_val xdata, *rdata, dkey, olddata;
5879 int do_sub = 0, insert_key, insert_data;
5880 unsigned int mcount = 0, dcount = 0, nospill;
5883 unsigned int nflags;
5886 if (mc == NULL || key == NULL)
5889 env = mc->mc_txn->mt_env;
5891 /* Check this first so counter will always be zero on any
5894 if (flags & MDB_MULTIPLE) {
5895 dcount = data[1].mv_size;
5896 data[1].mv_size = 0;
5897 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5898 return MDB_INCOMPATIBLE;
5901 nospill = flags & MDB_NOSPILL;
5902 flags &= ~MDB_NOSPILL;
5904 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5905 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5907 if (key->mv_size-1 >= ENV_MAXKEY(env))
5908 return MDB_BAD_VALSIZE;
5910 #if SIZE_MAX > MAXDATASIZE
5911 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5912 return MDB_BAD_VALSIZE;
5914 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5915 return MDB_BAD_VALSIZE;
5918 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5919 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5923 if (flags == MDB_CURRENT) {
5924 if (!(mc->mc_flags & C_INITIALIZED))
5927 } else if (mc->mc_db->md_root == P_INVALID) {
5928 /* new database, cursor has nothing to point to */
5931 mc->mc_flags &= ~C_INITIALIZED;
5936 if (flags & MDB_APPEND) {
5938 rc = mdb_cursor_last(mc, &k2, &d2);
5940 rc = mc->mc_dbx->md_cmp(key, &k2);
5943 mc->mc_ki[mc->mc_top]++;
5945 /* new key is <= last key */
5950 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5952 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5953 DPRINTF(("duplicate key [%s]", DKEY(key)));
5955 return MDB_KEYEXIST;
5957 if (rc && rc != MDB_NOTFOUND)
5961 if (mc->mc_flags & C_DEL)
5962 mc->mc_flags ^= C_DEL;
5964 /* Cursor is positioned, check for room in the dirty list */
5966 if (flags & MDB_MULTIPLE) {
5968 xdata.mv_size = data->mv_size * dcount;
5972 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5976 if (rc == MDB_NO_ROOT) {
5978 /* new database, write a root leaf page */
5979 DPUTS("allocating new root leaf page");
5980 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5983 mdb_cursor_push(mc, np);
5984 mc->mc_db->md_root = np->mp_pgno;
5985 mc->mc_db->md_depth++;
5986 *mc->mc_dbflag |= DB_DIRTY;
5987 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5989 np->mp_flags |= P_LEAF2;
5990 mc->mc_flags |= C_INITIALIZED;
5992 /* make sure all cursor pages are writable */
5993 rc2 = mdb_cursor_touch(mc);
5998 insert_key = insert_data = rc;
6000 /* The key does not exist */
6001 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6002 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6003 LEAFSIZE(key, data) > env->me_nodemax)
6005 /* Too big for a node, insert in sub-DB */
6006 fp_flags = P_LEAF|P_DIRTY;
6008 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6009 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
6013 /* there's only a key anyway, so this is a no-op */
6014 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6016 unsigned int ksize = mc->mc_db->md_pad;
6017 if (key->mv_size != ksize)
6018 return MDB_BAD_VALSIZE;
6019 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6020 memcpy(ptr, key->mv_data, ksize);
6025 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6026 olddata.mv_size = NODEDSZ(leaf);
6027 olddata.mv_data = NODEDATA(leaf);
6030 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6031 /* Prepare (sub-)page/sub-DB to accept the new item,
6032 * if needed. fp: old sub-page or a header faking
6033 * it. mp: new (sub-)page. offset: growth in page
6034 * size. xdata: node data with new page or DB.
6036 unsigned i, offset = 0;
6037 mp = fp = xdata.mv_data = env->me_pbuf;
6038 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6040 /* Was a single item before, must convert now */
6041 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6042 /* Just overwrite the current item */
6043 if (flags == MDB_CURRENT)
6046 #if UINT_MAX < SIZE_MAX
6047 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6048 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6050 /* does data match? */
6051 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6052 if (flags & MDB_NODUPDATA)
6053 return MDB_KEYEXIST;
6058 /* Back up original data item */
6059 dkey.mv_size = olddata.mv_size;
6060 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6062 /* Make sub-page header for the dup items, with dummy body */
6063 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6064 fp->mp_lower = PAGEHDRSZ;
6065 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6066 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6067 fp->mp_flags |= P_LEAF2;
6068 fp->mp_pad = data->mv_size;
6069 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6071 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6072 (dkey.mv_size & 1) + (data->mv_size & 1);
6074 fp->mp_upper = xdata.mv_size;
6075 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
6076 } else if (leaf->mn_flags & F_SUBDATA) {
6077 /* Data is on sub-DB, just store it */
6078 flags |= F_DUPDATA|F_SUBDATA;
6081 /* Data is on sub-page */
6082 fp = olddata.mv_data;
6085 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6086 offset = EVEN(NODESIZE + sizeof(indx_t) +
6090 offset = fp->mp_pad;
6091 if (SIZELEFT(fp) < offset) {
6092 offset *= 4; /* space for 4 more */
6095 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6097 fp->mp_flags |= P_DIRTY;
6098 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6099 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6103 xdata.mv_size = olddata.mv_size + offset;
6106 fp_flags = fp->mp_flags;
6107 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6108 /* Too big for a sub-page, convert to sub-DB */
6109 fp_flags &= ~P_SUBP;
6111 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6112 fp_flags |= P_LEAF2;
6113 dummy.md_pad = fp->mp_pad;
6114 dummy.md_flags = MDB_DUPFIXED;
6115 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6116 dummy.md_flags |= MDB_INTEGERKEY;
6122 dummy.md_branch_pages = 0;
6123 dummy.md_leaf_pages = 1;
6124 dummy.md_overflow_pages = 0;
6125 dummy.md_entries = NUMKEYS(fp);
6126 xdata.mv_size = sizeof(MDB_db);
6127 xdata.mv_data = &dummy;
6128 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6130 offset = env->me_psize - olddata.mv_size;
6131 flags |= F_DUPDATA|F_SUBDATA;
6132 dummy.md_root = mp->mp_pgno;
6135 mp->mp_flags = fp_flags | P_DIRTY;
6136 mp->mp_pad = fp->mp_pad;
6137 mp->mp_lower = fp->mp_lower;
6138 mp->mp_upper = fp->mp_upper + offset;
6139 if (fp_flags & P_LEAF2) {
6140 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6142 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
6143 olddata.mv_size - fp->mp_upper);
6144 for (i=0; i<NUMKEYS(fp); i++)
6145 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6153 mdb_node_del(mc, 0);
6157 /* overflow page overwrites need special handling */
6158 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6161 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6163 memcpy(&pg, olddata.mv_data, sizeof(pg));
6164 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6166 ovpages = omp->mp_pages;
6168 /* Is the ov page large enough? */
6169 if (ovpages >= dpages) {
6170 if (!(omp->mp_flags & P_DIRTY) &&
6171 (level || (env->me_flags & MDB_WRITEMAP)))
6173 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6176 level = 0; /* dirty in this txn or clean */
6179 if (omp->mp_flags & P_DIRTY) {
6180 /* yes, overwrite it. Note in this case we don't
6181 * bother to try shrinking the page if the new data
6182 * is smaller than the overflow threshold.
6185 /* It is writable only in a parent txn */
6186 size_t sz = (size_t) env->me_psize * ovpages, off;
6187 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6193 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6194 mdb_cassert(mc, rc2 == 0);
6195 if (!(flags & MDB_RESERVE)) {
6196 /* Copy end of page, adjusting alignment so
6197 * compiler may copy words instead of bytes.
6199 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6200 memcpy((size_t *)((char *)np + off),
6201 (size_t *)((char *)omp + off), sz - off);
6204 memcpy(np, omp, sz); /* Copy beginning of page */
6207 SETDSZ(leaf, data->mv_size);
6208 if (F_ISSET(flags, MDB_RESERVE))
6209 data->mv_data = METADATA(omp);
6211 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6215 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6217 } else if (data->mv_size == olddata.mv_size) {
6218 /* same size, just replace it. Note that we could
6219 * also reuse this node if the new data is smaller,
6220 * but instead we opt to shrink the node in that case.
6222 if (F_ISSET(flags, MDB_RESERVE))
6223 data->mv_data = olddata.mv_data;
6224 else if (!(mc->mc_flags & C_SUB))
6225 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6227 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6230 mdb_node_del(mc, 0);
6236 nflags = flags & NODE_ADD_FLAGS;
6237 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6238 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6239 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6240 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6242 nflags |= MDB_SPLIT_REPLACE;
6243 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6245 /* There is room already in this leaf page. */
6246 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6247 if (rc == 0 && insert_key) {
6248 /* Adjust other cursors pointing to mp */
6249 MDB_cursor *m2, *m3;
6250 MDB_dbi dbi = mc->mc_dbi;
6251 unsigned i = mc->mc_top;
6252 MDB_page *mp = mc->mc_pg[i];
6254 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6255 if (mc->mc_flags & C_SUB)
6256 m3 = &m2->mc_xcursor->mx_cursor;
6259 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6260 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6267 if (rc == MDB_SUCCESS) {
6268 /* Now store the actual data in the child DB. Note that we're
6269 * storing the user data in the keys field, so there are strict
6270 * size limits on dupdata. The actual data fields of the child
6271 * DB are all zero size.
6279 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6280 if (flags & MDB_CURRENT) {
6281 xflags = MDB_CURRENT|MDB_NOSPILL;
6283 mdb_xcursor_init1(mc, leaf);
6284 xflags = (flags & MDB_NODUPDATA) ?
6285 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6287 /* converted, write the original data first */
6289 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6293 /* Adjust other cursors pointing to mp */
6295 unsigned i = mc->mc_top;
6296 MDB_page *mp = mc->mc_pg[i];
6298 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6299 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6300 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6301 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6302 mdb_xcursor_init1(m2, leaf);
6306 /* we've done our job */
6309 ecount = mc->mc_xcursor->mx_db.md_entries;
6310 if (flags & MDB_APPENDDUP)
6311 xflags |= MDB_APPEND;
6312 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6313 if (flags & F_SUBDATA) {
6314 void *db = NODEDATA(leaf);
6315 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6317 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6319 /* Increment count unless we just replaced an existing item. */
6321 mc->mc_db->md_entries++;
6323 /* Invalidate txn if we created an empty sub-DB */
6326 /* If we succeeded and the key didn't exist before,
6327 * make sure the cursor is marked valid.
6329 mc->mc_flags |= C_INITIALIZED;
6331 if (flags & MDB_MULTIPLE) {
6334 /* let caller know how many succeeded, if any */
6335 data[1].mv_size = mcount;
6336 if (mcount < dcount) {
6337 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6338 insert_key = insert_data = 0;
6345 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6348 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6353 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6359 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6360 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6362 if (!(mc->mc_flags & C_INITIALIZED))
6365 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6366 return MDB_NOTFOUND;
6368 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6371 rc = mdb_cursor_touch(mc);
6375 mp = mc->mc_pg[mc->mc_top];
6378 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6380 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6381 if (flags & MDB_NODUPDATA) {
6382 /* mdb_cursor_del0() will subtract the final entry */
6383 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6385 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6386 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6388 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6391 /* If sub-DB still has entries, we're done */
6392 if (mc->mc_xcursor->mx_db.md_entries) {
6393 if (leaf->mn_flags & F_SUBDATA) {
6394 /* update subDB info */
6395 void *db = NODEDATA(leaf);
6396 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6399 /* shrink fake page */
6400 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6401 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6402 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6403 /* fix other sub-DB cursors pointed at this fake page */
6404 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6405 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6406 if (m2->mc_pg[mc->mc_top] == mp &&
6407 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6408 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6411 mc->mc_db->md_entries--;
6412 mc->mc_flags |= C_DEL;
6415 /* otherwise fall thru and delete the sub-DB */
6418 if (leaf->mn_flags & F_SUBDATA) {
6419 /* add all the child DB's pages to the free list */
6420 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6426 /* add overflow pages to free list */
6427 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6431 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6432 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6433 (rc = mdb_ovpage_free(mc, omp)))
6438 return mdb_cursor_del0(mc);
6441 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6445 /** Allocate and initialize new pages for a database.
6446 * @param[in] mc a cursor on the database being added to.
6447 * @param[in] flags flags defining what type of page is being allocated.
6448 * @param[in] num the number of pages to allocate. This is usually 1,
6449 * unless allocating overflow pages for a large record.
6450 * @param[out] mp Address of a page, or NULL on failure.
6451 * @return 0 on success, non-zero on failure.
6454 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6459 if ((rc = mdb_page_alloc(mc, num, &np)))
6461 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6462 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6463 np->mp_flags = flags | P_DIRTY;
6464 np->mp_lower = PAGEHDRSZ;
6465 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6468 mc->mc_db->md_branch_pages++;
6469 else if (IS_LEAF(np))
6470 mc->mc_db->md_leaf_pages++;
6471 else if (IS_OVERFLOW(np)) {
6472 mc->mc_db->md_overflow_pages += num;
6480 /** Calculate the size of a leaf node.
6481 * The size depends on the environment's page size; if a data item
6482 * is too large it will be put onto an overflow page and the node
6483 * size will only include the key and not the data. Sizes are always
6484 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6485 * of the #MDB_node headers.
6486 * @param[in] env The environment handle.
6487 * @param[in] key The key for the node.
6488 * @param[in] data The data for the node.
6489 * @return The number of bytes needed to store the node.
6492 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6496 sz = LEAFSIZE(key, data);
6497 if (sz > env->me_nodemax) {
6498 /* put on overflow page */
6499 sz -= data->mv_size - sizeof(pgno_t);
6502 return EVEN(sz + sizeof(indx_t));
6505 /** Calculate the size of a branch node.
6506 * The size should depend on the environment's page size but since
6507 * we currently don't support spilling large keys onto overflow
6508 * pages, it's simply the size of the #MDB_node header plus the
6509 * size of the key. Sizes are always rounded up to an even number
6510 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6511 * @param[in] env The environment handle.
6512 * @param[in] key The key for the node.
6513 * @return The number of bytes needed to store the node.
6516 mdb_branch_size(MDB_env *env, MDB_val *key)
6521 if (sz > env->me_nodemax) {
6522 /* put on overflow page */
6523 /* not implemented */
6524 /* sz -= key->size - sizeof(pgno_t); */
6527 return sz + sizeof(indx_t);
6530 /** Add a node to the page pointed to by the cursor.
6531 * @param[in] mc The cursor for this operation.
6532 * @param[in] indx The index on the page where the new node should be added.
6533 * @param[in] key The key for the new node.
6534 * @param[in] data The data for the new node, if any.
6535 * @param[in] pgno The page number, if adding a branch node.
6536 * @param[in] flags Flags for the node.
6537 * @return 0 on success, non-zero on failure. Possible errors are:
6539 * <li>ENOMEM - failed to allocate overflow pages for the node.
6540 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6541 * should never happen since all callers already calculate the
6542 * page's free space before calling this function.
6546 mdb_node_add(MDB_cursor *mc, indx_t indx,
6547 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6550 size_t node_size = NODESIZE;
6554 MDB_page *mp = mc->mc_pg[mc->mc_top];
6555 MDB_page *ofp = NULL; /* overflow page */
6558 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6560 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6561 IS_LEAF(mp) ? "leaf" : "branch",
6562 IS_SUBP(mp) ? "sub-" : "",
6563 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6564 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6567 /* Move higher keys up one slot. */
6568 int ksize = mc->mc_db->md_pad, dif;
6569 char *ptr = LEAF2KEY(mp, indx, ksize);
6570 dif = NUMKEYS(mp) - indx;
6572 memmove(ptr+ksize, ptr, dif*ksize);
6573 /* insert new key */
6574 memcpy(ptr, key->mv_data, ksize);
6576 /* Just using these for counting */
6577 mp->mp_lower += sizeof(indx_t);
6578 mp->mp_upper -= ksize - sizeof(indx_t);
6582 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6584 node_size += key->mv_size;
6586 mdb_cassert(mc, data);
6587 if (F_ISSET(flags, F_BIGDATA)) {
6588 /* Data already on overflow page. */
6589 node_size += sizeof(pgno_t);
6590 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6591 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6593 /* Put data on overflow page. */
6594 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6595 data->mv_size, node_size+data->mv_size));
6596 node_size = EVEN(node_size + sizeof(pgno_t));
6597 if ((ssize_t)node_size > room)
6599 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6601 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6605 node_size += data->mv_size;
6608 node_size = EVEN(node_size);
6609 if ((ssize_t)node_size > room)
6613 /* Move higher pointers up one slot. */
6614 for (i = NUMKEYS(mp); i > indx; i--)
6615 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6617 /* Adjust free space offsets. */
6618 ofs = mp->mp_upper - node_size;
6619 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6620 mp->mp_ptrs[indx] = ofs;
6622 mp->mp_lower += sizeof(indx_t);
6624 /* Write the node data. */
6625 node = NODEPTR(mp, indx);
6626 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6627 node->mn_flags = flags;
6629 SETDSZ(node,data->mv_size);
6634 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6637 mdb_cassert(mc, key);
6639 if (F_ISSET(flags, F_BIGDATA))
6640 memcpy(node->mn_data + key->mv_size, data->mv_data,
6642 else if (F_ISSET(flags, MDB_RESERVE))
6643 data->mv_data = node->mn_data + key->mv_size;
6645 memcpy(node->mn_data + key->mv_size, data->mv_data,
6648 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6650 if (F_ISSET(flags, MDB_RESERVE))
6651 data->mv_data = METADATA(ofp);
6653 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6660 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6661 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6662 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6663 DPRINTF(("node size = %"Z"u", node_size));
6664 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6665 return MDB_PAGE_FULL;
6668 /** Delete the specified node from a page.
6669 * @param[in] mc Cursor pointing to the node to delete.
6670 * @param[in] ksize The size of a node. Only used if the page is
6671 * part of a #MDB_DUPFIXED database.
6674 mdb_node_del(MDB_cursor *mc, int ksize)
6676 MDB_page *mp = mc->mc_pg[mc->mc_top];
6677 indx_t indx = mc->mc_ki[mc->mc_top];
6679 indx_t i, j, numkeys, ptr;
6683 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6684 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6685 numkeys = NUMKEYS(mp);
6686 mdb_cassert(mc, indx < numkeys);
6689 int x = numkeys - 1 - indx;
6690 base = LEAF2KEY(mp, indx, ksize);
6692 memmove(base, base + ksize, x * ksize);
6693 mp->mp_lower -= sizeof(indx_t);
6694 mp->mp_upper += ksize - sizeof(indx_t);
6698 node = NODEPTR(mp, indx);
6699 sz = NODESIZE + node->mn_ksize;
6701 if (F_ISSET(node->mn_flags, F_BIGDATA))
6702 sz += sizeof(pgno_t);
6704 sz += NODEDSZ(node);
6708 ptr = mp->mp_ptrs[indx];
6709 for (i = j = 0; i < numkeys; i++) {
6711 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6712 if (mp->mp_ptrs[i] < ptr)
6713 mp->mp_ptrs[j] += sz;
6718 base = (char *)mp + mp->mp_upper;
6719 memmove(base + sz, base, ptr - mp->mp_upper);
6721 mp->mp_lower -= sizeof(indx_t);
6725 /** Compact the main page after deleting a node on a subpage.
6726 * @param[in] mp The main page to operate on.
6727 * @param[in] indx The index of the subpage on the main page.
6730 mdb_node_shrink(MDB_page *mp, indx_t indx)
6736 indx_t i, numkeys, ptr;
6738 node = NODEPTR(mp, indx);
6739 sp = (MDB_page *)NODEDATA(node);
6740 delta = SIZELEFT(sp);
6741 xp = (MDB_page *)((char *)sp + delta);
6743 /* shift subpage upward */
6745 nsize = NUMKEYS(sp) * sp->mp_pad;
6747 return; /* do not make the node uneven-sized */
6748 memmove(METADATA(xp), METADATA(sp), nsize);
6751 numkeys = NUMKEYS(sp);
6752 for (i=numkeys-1; i>=0; i--)
6753 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6755 xp->mp_upper = sp->mp_lower;
6756 xp->mp_lower = sp->mp_lower;
6757 xp->mp_flags = sp->mp_flags;
6758 xp->mp_pad = sp->mp_pad;
6759 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6761 nsize = NODEDSZ(node) - delta;
6762 SETDSZ(node, nsize);
6764 /* shift lower nodes upward */
6765 ptr = mp->mp_ptrs[indx];
6766 numkeys = NUMKEYS(mp);
6767 for (i = 0; i < numkeys; i++) {
6768 if (mp->mp_ptrs[i] <= ptr)
6769 mp->mp_ptrs[i] += delta;
6772 base = (char *)mp + mp->mp_upper;
6773 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6774 mp->mp_upper += delta;
6777 /** Initial setup of a sorted-dups cursor.
6778 * Sorted duplicates are implemented as a sub-database for the given key.
6779 * The duplicate data items are actually keys of the sub-database.
6780 * Operations on the duplicate data items are performed using a sub-cursor
6781 * initialized when the sub-database is first accessed. This function does
6782 * the preliminary setup of the sub-cursor, filling in the fields that
6783 * depend only on the parent DB.
6784 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6787 mdb_xcursor_init0(MDB_cursor *mc)
6789 MDB_xcursor *mx = mc->mc_xcursor;
6791 mx->mx_cursor.mc_xcursor = NULL;
6792 mx->mx_cursor.mc_txn = mc->mc_txn;
6793 mx->mx_cursor.mc_db = &mx->mx_db;
6794 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6795 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6796 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6797 mx->mx_cursor.mc_snum = 0;
6798 mx->mx_cursor.mc_top = 0;
6799 mx->mx_cursor.mc_flags = C_SUB;
6800 mx->mx_dbx.md_name.mv_size = 0;
6801 mx->mx_dbx.md_name.mv_data = NULL;
6802 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6803 mx->mx_dbx.md_dcmp = NULL;
6804 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6807 /** Final setup of a sorted-dups cursor.
6808 * Sets up the fields that depend on the data from the main cursor.
6809 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6810 * @param[in] node The data containing the #MDB_db record for the
6811 * sorted-dup database.
6814 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6816 MDB_xcursor *mx = mc->mc_xcursor;
6818 if (node->mn_flags & F_SUBDATA) {
6819 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6820 mx->mx_cursor.mc_pg[0] = 0;
6821 mx->mx_cursor.mc_snum = 0;
6822 mx->mx_cursor.mc_top = 0;
6823 mx->mx_cursor.mc_flags = C_SUB;
6825 MDB_page *fp = NODEDATA(node);
6826 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6827 mx->mx_db.md_flags = 0;
6828 mx->mx_db.md_depth = 1;
6829 mx->mx_db.md_branch_pages = 0;
6830 mx->mx_db.md_leaf_pages = 1;
6831 mx->mx_db.md_overflow_pages = 0;
6832 mx->mx_db.md_entries = NUMKEYS(fp);
6833 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6834 mx->mx_cursor.mc_snum = 1;
6835 mx->mx_cursor.mc_top = 0;
6836 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6837 mx->mx_cursor.mc_pg[0] = fp;
6838 mx->mx_cursor.mc_ki[0] = 0;
6839 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6840 mx->mx_db.md_flags = MDB_DUPFIXED;
6841 mx->mx_db.md_pad = fp->mp_pad;
6842 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6843 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6846 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6847 mx->mx_db.md_root));
6848 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6849 #if UINT_MAX < SIZE_MAX
6850 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6851 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6855 /** Initialize a cursor for a given transaction and database. */
6857 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6860 mc->mc_backup = NULL;
6863 mc->mc_db = &txn->mt_dbs[dbi];
6864 mc->mc_dbx = &txn->mt_dbxs[dbi];
6865 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6870 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6871 mdb_tassert(txn, mx != NULL);
6872 mc->mc_xcursor = mx;
6873 mdb_xcursor_init0(mc);
6875 mc->mc_xcursor = NULL;
6877 if (*mc->mc_dbflag & DB_STALE) {
6878 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6883 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6886 size_t size = sizeof(MDB_cursor);
6888 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6891 if (txn->mt_flags & MDB_TXN_ERROR)
6894 /* Allow read access to the freelist */
6895 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6898 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6899 size += sizeof(MDB_xcursor);
6901 if ((mc = malloc(size)) != NULL) {
6902 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6903 if (txn->mt_cursors) {
6904 mc->mc_next = txn->mt_cursors[dbi];
6905 txn->mt_cursors[dbi] = mc;
6906 mc->mc_flags |= C_UNTRACK;
6918 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6920 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6923 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6926 if (txn->mt_flags & MDB_TXN_ERROR)
6929 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6933 /* Return the count of duplicate data items for the current key */
6935 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6939 if (mc == NULL || countp == NULL)
6942 if (mc->mc_xcursor == NULL)
6943 return MDB_INCOMPATIBLE;
6945 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6948 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6949 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6952 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6955 *countp = mc->mc_xcursor->mx_db.md_entries;
6961 mdb_cursor_close(MDB_cursor *mc)
6963 if (mc && !mc->mc_backup) {
6964 /* remove from txn, if tracked */
6965 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6966 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6967 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6969 *prev = mc->mc_next;
6976 mdb_cursor_txn(MDB_cursor *mc)
6978 if (!mc) return NULL;
6983 mdb_cursor_dbi(MDB_cursor *mc)
6988 /** Replace the key for a branch node with a new key.
6989 * @param[in] mc Cursor pointing to the node to operate on.
6990 * @param[in] key The new key to use.
6991 * @return 0 on success, non-zero on failure.
6994 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7000 int delta, ksize, oksize;
7001 indx_t ptr, i, numkeys, indx;
7004 indx = mc->mc_ki[mc->mc_top];
7005 mp = mc->mc_pg[mc->mc_top];
7006 node = NODEPTR(mp, indx);
7007 ptr = mp->mp_ptrs[indx];
7011 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7012 k2.mv_data = NODEKEY(node);
7013 k2.mv_size = node->mn_ksize;
7014 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7016 mdb_dkey(&k2, kbuf2),
7022 /* Sizes must be 2-byte aligned. */
7023 ksize = EVEN(key->mv_size);
7024 oksize = EVEN(node->mn_ksize);
7025 delta = ksize - oksize;
7027 /* Shift node contents if EVEN(key length) changed. */
7029 if (delta > 0 && SIZELEFT(mp) < delta) {
7031 /* not enough space left, do a delete and split */
7032 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7033 pgno = NODEPGNO(node);
7034 mdb_node_del(mc, 0);
7035 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7038 numkeys = NUMKEYS(mp);
7039 for (i = 0; i < numkeys; i++) {
7040 if (mp->mp_ptrs[i] <= ptr)
7041 mp->mp_ptrs[i] -= delta;
7044 base = (char *)mp + mp->mp_upper;
7045 len = ptr - mp->mp_upper + NODESIZE;
7046 memmove(base - delta, base, len);
7047 mp->mp_upper -= delta;
7049 node = NODEPTR(mp, indx);
7052 /* But even if no shift was needed, update ksize */
7053 if (node->mn_ksize != key->mv_size)
7054 node->mn_ksize = key->mv_size;
7057 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7063 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7065 /** Move a node from csrc to cdst.
7068 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7075 unsigned short flags;
7079 /* Mark src and dst as dirty. */
7080 if ((rc = mdb_page_touch(csrc)) ||
7081 (rc = mdb_page_touch(cdst)))
7084 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7085 key.mv_size = csrc->mc_db->md_pad;
7086 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7088 data.mv_data = NULL;
7092 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7093 mdb_cassert(csrc, !((size_t)srcnode & 1));
7094 srcpg = NODEPGNO(srcnode);
7095 flags = srcnode->mn_flags;
7096 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7097 unsigned int snum = csrc->mc_snum;
7099 /* must find the lowest key below src */
7100 rc = mdb_page_search_lowest(csrc);
7103 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7104 key.mv_size = csrc->mc_db->md_pad;
7105 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7107 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7108 key.mv_size = NODEKSZ(s2);
7109 key.mv_data = NODEKEY(s2);
7111 csrc->mc_snum = snum--;
7112 csrc->mc_top = snum;
7114 key.mv_size = NODEKSZ(srcnode);
7115 key.mv_data = NODEKEY(srcnode);
7117 data.mv_size = NODEDSZ(srcnode);
7118 data.mv_data = NODEDATA(srcnode);
7120 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7121 unsigned int snum = cdst->mc_snum;
7124 /* must find the lowest key below dst */
7125 mdb_cursor_copy(cdst, &mn);
7126 rc = mdb_page_search_lowest(&mn);
7129 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7130 bkey.mv_size = mn.mc_db->md_pad;
7131 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7133 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7134 bkey.mv_size = NODEKSZ(s2);
7135 bkey.mv_data = NODEKEY(s2);
7137 mn.mc_snum = snum--;
7140 rc = mdb_update_key(&mn, &bkey);
7145 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7146 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7147 csrc->mc_ki[csrc->mc_top],
7149 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7150 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7152 /* Add the node to the destination page.
7154 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7155 if (rc != MDB_SUCCESS)
7158 /* Delete the node from the source page.
7160 mdb_node_del(csrc, key.mv_size);
7163 /* Adjust other cursors pointing to mp */
7164 MDB_cursor *m2, *m3;
7165 MDB_dbi dbi = csrc->mc_dbi;
7166 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7168 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7169 if (csrc->mc_flags & C_SUB)
7170 m3 = &m2->mc_xcursor->mx_cursor;
7173 if (m3 == csrc) continue;
7174 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7175 csrc->mc_ki[csrc->mc_top]) {
7176 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7177 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7182 /* Update the parent separators.
7184 if (csrc->mc_ki[csrc->mc_top] == 0) {
7185 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7186 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7187 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7189 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7190 key.mv_size = NODEKSZ(srcnode);
7191 key.mv_data = NODEKEY(srcnode);
7193 DPRINTF(("update separator for source page %"Z"u to [%s]",
7194 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7195 mdb_cursor_copy(csrc, &mn);
7198 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7201 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7203 indx_t ix = csrc->mc_ki[csrc->mc_top];
7204 nullkey.mv_size = 0;
7205 csrc->mc_ki[csrc->mc_top] = 0;
7206 rc = mdb_update_key(csrc, &nullkey);
7207 csrc->mc_ki[csrc->mc_top] = ix;
7208 mdb_cassert(csrc, rc == MDB_SUCCESS);
7212 if (cdst->mc_ki[cdst->mc_top] == 0) {
7213 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7214 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7215 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7217 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7218 key.mv_size = NODEKSZ(srcnode);
7219 key.mv_data = NODEKEY(srcnode);
7221 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7222 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7223 mdb_cursor_copy(cdst, &mn);
7226 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7229 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7231 indx_t ix = cdst->mc_ki[cdst->mc_top];
7232 nullkey.mv_size = 0;
7233 cdst->mc_ki[cdst->mc_top] = 0;
7234 rc = mdb_update_key(cdst, &nullkey);
7235 cdst->mc_ki[cdst->mc_top] = ix;
7236 mdb_cassert(csrc, rc == MDB_SUCCESS);
7243 /** Merge one page into another.
7244 * The nodes from the page pointed to by \b csrc will
7245 * be copied to the page pointed to by \b cdst and then
7246 * the \b csrc page will be freed.
7247 * @param[in] csrc Cursor pointing to the source page.
7248 * @param[in] cdst Cursor pointing to the destination page.
7249 * @return 0 on success, non-zero on failure.
7252 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7260 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7261 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7263 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7264 mdb_cassert(csrc, cdst->mc_snum > 1);
7266 /* Mark dst as dirty. */
7267 if ((rc = mdb_page_touch(cdst)))
7270 /* Move all nodes from src to dst.
7272 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7273 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7274 key.mv_size = csrc->mc_db->md_pad;
7275 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7276 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7277 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7278 if (rc != MDB_SUCCESS)
7280 key.mv_data = (char *)key.mv_data + key.mv_size;
7283 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7284 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7285 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7288 mdb_cursor_copy(csrc, &mn);
7289 /* must find the lowest key below src */
7290 rc = mdb_page_search_lowest(&mn);
7293 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7294 key.mv_size = mn.mc_db->md_pad;
7295 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7297 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7298 key.mv_size = NODEKSZ(s2);
7299 key.mv_data = NODEKEY(s2);
7302 key.mv_size = srcnode->mn_ksize;
7303 key.mv_data = NODEKEY(srcnode);
7306 data.mv_size = NODEDSZ(srcnode);
7307 data.mv_data = NODEDATA(srcnode);
7308 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7309 if (rc != MDB_SUCCESS)
7314 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7315 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7316 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7318 /* Unlink the src page from parent and add to free list.
7321 mdb_node_del(csrc, 0);
7322 if (csrc->mc_ki[csrc->mc_top] == 0) {
7324 rc = mdb_update_key(csrc, &key);
7332 mdb_page_loose(csrc->mc_txn->mt_env, csrc->mc_pg[csrc->mc_top]);
7333 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7334 csrc->mc_db->md_leaf_pages--;
7336 csrc->mc_db->md_branch_pages--;
7338 /* Adjust other cursors pointing to mp */
7339 MDB_cursor *m2, *m3;
7340 MDB_dbi dbi = csrc->mc_dbi;
7341 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7343 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7344 if (csrc->mc_flags & C_SUB)
7345 m3 = &m2->mc_xcursor->mx_cursor;
7348 if (m3 == csrc) continue;
7349 if (m3->mc_snum < csrc->mc_snum) continue;
7350 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7351 m3->mc_pg[csrc->mc_top] = mp;
7352 m3->mc_ki[csrc->mc_top] += nkeys;
7357 unsigned int snum = cdst->mc_snum;
7358 uint16_t depth = cdst->mc_db->md_depth;
7359 mdb_cursor_pop(cdst);
7360 rc = mdb_rebalance(cdst);
7361 /* Did the tree shrink? */
7362 if (depth > cdst->mc_db->md_depth)
7364 cdst->mc_snum = snum;
7365 cdst->mc_top = snum-1;
7370 /** Copy the contents of a cursor.
7371 * @param[in] csrc The cursor to copy from.
7372 * @param[out] cdst The cursor to copy to.
7375 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7379 cdst->mc_txn = csrc->mc_txn;
7380 cdst->mc_dbi = csrc->mc_dbi;
7381 cdst->mc_db = csrc->mc_db;
7382 cdst->mc_dbx = csrc->mc_dbx;
7383 cdst->mc_snum = csrc->mc_snum;
7384 cdst->mc_top = csrc->mc_top;
7385 cdst->mc_flags = csrc->mc_flags;
7387 for (i=0; i<csrc->mc_snum; i++) {
7388 cdst->mc_pg[i] = csrc->mc_pg[i];
7389 cdst->mc_ki[i] = csrc->mc_ki[i];
7393 /** Rebalance the tree after a delete operation.
7394 * @param[in] mc Cursor pointing to the page where rebalancing
7396 * @return 0 on success, non-zero on failure.
7399 mdb_rebalance(MDB_cursor *mc)
7403 unsigned int ptop, minkeys;
7407 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7408 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7409 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7410 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7411 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7413 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7414 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7415 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7416 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7420 if (mc->mc_snum < 2) {
7421 MDB_page *mp = mc->mc_pg[0];
7423 DPUTS("Can't rebalance a subpage, ignoring");
7426 if (NUMKEYS(mp) == 0) {
7427 DPUTS("tree is completely empty");
7428 mc->mc_db->md_root = P_INVALID;
7429 mc->mc_db->md_depth = 0;
7430 mc->mc_db->md_leaf_pages = 0;
7431 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7434 /* Adjust cursors pointing to mp */
7437 mc->mc_flags &= ~C_INITIALIZED;
7439 MDB_cursor *m2, *m3;
7440 MDB_dbi dbi = mc->mc_dbi;
7442 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7443 if (mc->mc_flags & C_SUB)
7444 m3 = &m2->mc_xcursor->mx_cursor;
7447 if (m3->mc_snum < mc->mc_snum) continue;
7448 if (m3->mc_pg[0] == mp) {
7451 m3->mc_flags &= ~C_INITIALIZED;
7455 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7457 DPUTS("collapsing root page!");
7458 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7461 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7462 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7465 mc->mc_db->md_depth--;
7466 mc->mc_db->md_branch_pages--;
7467 mc->mc_ki[0] = mc->mc_ki[1];
7468 for (i = 1; i<mc->mc_db->md_depth; i++) {
7469 mc->mc_pg[i] = mc->mc_pg[i+1];
7470 mc->mc_ki[i] = mc->mc_ki[i+1];
7473 /* Adjust other cursors pointing to mp */
7474 MDB_cursor *m2, *m3;
7475 MDB_dbi dbi = mc->mc_dbi;
7477 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7478 if (mc->mc_flags & C_SUB)
7479 m3 = &m2->mc_xcursor->mx_cursor;
7482 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7483 if (m3->mc_pg[0] == mp) {
7486 for (i=0; i<m3->mc_snum; i++) {
7487 m3->mc_pg[i] = m3->mc_pg[i+1];
7488 m3->mc_ki[i] = m3->mc_ki[i+1];
7494 DPUTS("root page doesn't need rebalancing");
7498 /* The parent (branch page) must have at least 2 pointers,
7499 * otherwise the tree is invalid.
7501 ptop = mc->mc_top-1;
7502 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7504 /* Leaf page fill factor is below the threshold.
7505 * Try to move keys from left or right neighbor, or
7506 * merge with a neighbor page.
7511 mdb_cursor_copy(mc, &mn);
7512 mn.mc_xcursor = NULL;
7514 oldki = mc->mc_ki[mc->mc_top];
7515 if (mc->mc_ki[ptop] == 0) {
7516 /* We're the leftmost leaf in our parent.
7518 DPUTS("reading right neighbor");
7520 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7521 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7524 mn.mc_ki[mn.mc_top] = 0;
7525 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7527 /* There is at least one neighbor to the left.
7529 DPUTS("reading left neighbor");
7531 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7532 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7535 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7536 mc->mc_ki[mc->mc_top] = 0;
7539 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7540 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7541 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7543 /* If the neighbor page is above threshold and has enough keys,
7544 * move one key from it. Otherwise we should try to merge them.
7545 * (A branch page must never have less than 2 keys.)
7547 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7548 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7549 rc = mdb_node_move(&mn, mc);
7550 if (mc->mc_ki[ptop]) {
7554 if (mc->mc_ki[ptop] == 0) {
7555 rc = mdb_page_merge(&mn, mc);
7557 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7558 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7559 rc = mdb_page_merge(mc, &mn);
7560 mdb_cursor_copy(&mn, mc);
7562 mc->mc_flags &= ~C_EOF;
7564 mc->mc_ki[mc->mc_top] = oldki;
7568 /** Complete a delete operation started by #mdb_cursor_del(). */
7570 mdb_cursor_del0(MDB_cursor *mc)
7577 ki = mc->mc_ki[mc->mc_top];
7578 mdb_node_del(mc, mc->mc_db->md_pad);
7579 mc->mc_db->md_entries--;
7580 rc = mdb_rebalance(mc);
7582 if (rc == MDB_SUCCESS) {
7583 MDB_cursor *m2, *m3;
7584 MDB_dbi dbi = mc->mc_dbi;
7586 mp = mc->mc_pg[mc->mc_top];
7587 nkeys = NUMKEYS(mp);
7589 /* if mc points past last node in page, find next sibling */
7590 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7591 rc = mdb_cursor_sibling(mc, 1);
7592 if (rc == MDB_NOTFOUND)
7596 /* Adjust other cursors pointing to mp */
7597 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7598 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7599 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7601 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7603 if (m3->mc_pg[mc->mc_top] == mp) {
7604 if (m3->mc_ki[mc->mc_top] >= ki) {
7605 m3->mc_flags |= C_DEL;
7606 if (m3->mc_ki[mc->mc_top] > ki)
7607 m3->mc_ki[mc->mc_top]--;
7609 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7610 rc = mdb_cursor_sibling(m3, 1);
7611 if (rc == MDB_NOTFOUND)
7616 mc->mc_flags |= C_DEL;
7620 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7625 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7626 MDB_val *key, MDB_val *data)
7628 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7631 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7632 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7634 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7635 /* must ignore any data */
7639 return mdb_del0(txn, dbi, key, data, 0);
7643 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7644 MDB_val *key, MDB_val *data, unsigned flags)
7649 MDB_val rdata, *xdata;
7653 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7655 mdb_cursor_init(&mc, txn, dbi, &mx);
7664 flags |= MDB_NODUPDATA;
7666 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7668 /* let mdb_page_split know about this cursor if needed:
7669 * delete will trigger a rebalance; if it needs to move
7670 * a node from one page to another, it will have to
7671 * update the parent's separator key(s). If the new sepkey
7672 * is larger than the current one, the parent page may
7673 * run out of space, triggering a split. We need this
7674 * cursor to be consistent until the end of the rebalance.
7676 mc.mc_flags |= C_UNTRACK;
7677 mc.mc_next = txn->mt_cursors[dbi];
7678 txn->mt_cursors[dbi] = &mc;
7679 rc = mdb_cursor_del(&mc, flags);
7680 txn->mt_cursors[dbi] = mc.mc_next;
7685 /** Split a page and insert a new node.
7686 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7687 * The cursor will be updated to point to the actual page and index where
7688 * the node got inserted after the split.
7689 * @param[in] newkey The key for the newly inserted node.
7690 * @param[in] newdata The data for the newly inserted node.
7691 * @param[in] newpgno The page number, if the new node is a branch node.
7692 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7693 * @return 0 on success, non-zero on failure.
7696 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7697 unsigned int nflags)
7700 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7703 int i, j, split_indx, nkeys, pmax;
7704 MDB_env *env = mc->mc_txn->mt_env;
7706 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7707 MDB_page *copy = NULL;
7708 MDB_page *mp, *rp, *pp;
7713 mp = mc->mc_pg[mc->mc_top];
7714 newindx = mc->mc_ki[mc->mc_top];
7715 nkeys = NUMKEYS(mp);
7717 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7718 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7719 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7721 /* Create a right sibling. */
7722 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7724 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7726 if (mc->mc_snum < 2) {
7727 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7729 /* shift current top to make room for new parent */
7730 mc->mc_pg[1] = mc->mc_pg[0];
7731 mc->mc_ki[1] = mc->mc_ki[0];
7734 mc->mc_db->md_root = pp->mp_pgno;
7735 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7736 mc->mc_db->md_depth++;
7739 /* Add left (implicit) pointer. */
7740 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7741 /* undo the pre-push */
7742 mc->mc_pg[0] = mc->mc_pg[1];
7743 mc->mc_ki[0] = mc->mc_ki[1];
7744 mc->mc_db->md_root = mp->mp_pgno;
7745 mc->mc_db->md_depth--;
7752 ptop = mc->mc_top-1;
7753 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7756 mc->mc_flags |= C_SPLITTING;
7757 mdb_cursor_copy(mc, &mn);
7758 mn.mc_pg[mn.mc_top] = rp;
7759 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7761 if (nflags & MDB_APPEND) {
7762 mn.mc_ki[mn.mc_top] = 0;
7764 split_indx = newindx;
7768 split_indx = (nkeys+1) / 2;
7773 unsigned int lsize, rsize, ksize;
7774 /* Move half of the keys to the right sibling */
7775 x = mc->mc_ki[mc->mc_top] - split_indx;
7776 ksize = mc->mc_db->md_pad;
7777 split = LEAF2KEY(mp, split_indx, ksize);
7778 rsize = (nkeys - split_indx) * ksize;
7779 lsize = (nkeys - split_indx) * sizeof(indx_t);
7780 mp->mp_lower -= lsize;
7781 rp->mp_lower += lsize;
7782 mp->mp_upper += rsize - lsize;
7783 rp->mp_upper -= rsize - lsize;
7784 sepkey.mv_size = ksize;
7785 if (newindx == split_indx) {
7786 sepkey.mv_data = newkey->mv_data;
7788 sepkey.mv_data = split;
7791 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7792 memcpy(rp->mp_ptrs, split, rsize);
7793 sepkey.mv_data = rp->mp_ptrs;
7794 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7795 memcpy(ins, newkey->mv_data, ksize);
7796 mp->mp_lower += sizeof(indx_t);
7797 mp->mp_upper -= ksize - sizeof(indx_t);
7800 memcpy(rp->mp_ptrs, split, x * ksize);
7801 ins = LEAF2KEY(rp, x, ksize);
7802 memcpy(ins, newkey->mv_data, ksize);
7803 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7804 rp->mp_lower += sizeof(indx_t);
7805 rp->mp_upper -= ksize - sizeof(indx_t);
7806 mc->mc_ki[mc->mc_top] = x;
7807 mc->mc_pg[mc->mc_top] = rp;
7810 int psize, nsize, k;
7811 /* Maximum free space in an empty page */
7812 pmax = env->me_psize - PAGEHDRSZ;
7814 nsize = mdb_leaf_size(env, newkey, newdata);
7816 nsize = mdb_branch_size(env, newkey);
7817 nsize = EVEN(nsize);
7819 /* grab a page to hold a temporary copy */
7820 copy = mdb_page_malloc(mc->mc_txn, 1);
7825 copy->mp_pgno = mp->mp_pgno;
7826 copy->mp_flags = mp->mp_flags;
7827 copy->mp_lower = PAGEHDRSZ;
7828 copy->mp_upper = env->me_psize;
7830 /* prepare to insert */
7831 for (i=0, j=0; i<nkeys; i++) {
7833 copy->mp_ptrs[j++] = 0;
7835 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7838 /* When items are relatively large the split point needs
7839 * to be checked, because being off-by-one will make the
7840 * difference between success or failure in mdb_node_add.
7842 * It's also relevant if a page happens to be laid out
7843 * such that one half of its nodes are all "small" and
7844 * the other half of its nodes are "large." If the new
7845 * item is also "large" and falls on the half with
7846 * "large" nodes, it also may not fit.
7848 * As a final tweak, if the new item goes on the last
7849 * spot on the page (and thus, onto the new page), bias
7850 * the split so the new page is emptier than the old page.
7851 * This yields better packing during sequential inserts.
7853 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7854 /* Find split point */
7856 if (newindx <= split_indx || newindx >= nkeys) {
7858 k = newindx >= nkeys ? nkeys : split_indx+2;
7863 for (; i!=k; i+=j) {
7868 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7869 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7871 if (F_ISSET(node->mn_flags, F_BIGDATA))
7872 psize += sizeof(pgno_t);
7874 psize += NODEDSZ(node);
7876 psize = EVEN(psize);
7878 if (psize > pmax || i == k-j) {
7879 split_indx = i + (j<0);
7884 if (split_indx == newindx) {
7885 sepkey.mv_size = newkey->mv_size;
7886 sepkey.mv_data = newkey->mv_data;
7888 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7889 sepkey.mv_size = node->mn_ksize;
7890 sepkey.mv_data = NODEKEY(node);
7895 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7897 /* Copy separator key to the parent.
7899 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7903 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7908 if (mn.mc_snum == mc->mc_snum) {
7909 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7910 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7911 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7912 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7917 /* Right page might now have changed parent.
7918 * Check if left page also changed parent.
7920 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7921 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7922 for (i=0; i<ptop; i++) {
7923 mc->mc_pg[i] = mn.mc_pg[i];
7924 mc->mc_ki[i] = mn.mc_ki[i];
7926 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7927 if (mn.mc_ki[ptop]) {
7928 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7930 /* find right page's left sibling */
7931 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7932 mdb_cursor_sibling(mc, 0);
7937 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7940 mc->mc_flags ^= C_SPLITTING;
7941 if (rc != MDB_SUCCESS) {
7944 if (nflags & MDB_APPEND) {
7945 mc->mc_pg[mc->mc_top] = rp;
7946 mc->mc_ki[mc->mc_top] = 0;
7947 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7950 for (i=0; i<mc->mc_top; i++)
7951 mc->mc_ki[i] = mn.mc_ki[i];
7952 } else if (!IS_LEAF2(mp)) {
7954 mc->mc_pg[mc->mc_top] = rp;
7959 rkey.mv_data = newkey->mv_data;
7960 rkey.mv_size = newkey->mv_size;
7966 /* Update index for the new key. */
7967 mc->mc_ki[mc->mc_top] = j;
7969 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7970 rkey.mv_data = NODEKEY(node);
7971 rkey.mv_size = node->mn_ksize;
7973 xdata.mv_data = NODEDATA(node);
7974 xdata.mv_size = NODEDSZ(node);
7977 pgno = NODEPGNO(node);
7978 flags = node->mn_flags;
7981 if (!IS_LEAF(mp) && j == 0) {
7982 /* First branch index doesn't need key data. */
7986 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7992 mc->mc_pg[mc->mc_top] = copy;
7997 } while (i != split_indx);
7999 nkeys = NUMKEYS(copy);
8000 for (i=0; i<nkeys; i++)
8001 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8002 mp->mp_lower = copy->mp_lower;
8003 mp->mp_upper = copy->mp_upper;
8004 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8005 env->me_psize - copy->mp_upper);
8007 /* reset back to original page */
8008 if (newindx < split_indx) {
8009 mc->mc_pg[mc->mc_top] = mp;
8010 if (nflags & MDB_RESERVE) {
8011 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8012 if (!(node->mn_flags & F_BIGDATA))
8013 newdata->mv_data = NODEDATA(node);
8016 mc->mc_pg[mc->mc_top] = rp;
8018 /* Make sure mc_ki is still valid.
8020 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8021 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8022 for (i=0; i<=ptop; i++) {
8023 mc->mc_pg[i] = mn.mc_pg[i];
8024 mc->mc_ki[i] = mn.mc_ki[i];
8031 /* Adjust other cursors pointing to mp */
8032 MDB_cursor *m2, *m3;
8033 MDB_dbi dbi = mc->mc_dbi;
8034 int fixup = NUMKEYS(mp);
8036 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8037 if (mc->mc_flags & C_SUB)
8038 m3 = &m2->mc_xcursor->mx_cursor;
8043 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8045 if (m3->mc_flags & C_SPLITTING)
8050 for (k=m3->mc_top; k>=0; k--) {
8051 m3->mc_ki[k+1] = m3->mc_ki[k];
8052 m3->mc_pg[k+1] = m3->mc_pg[k];
8054 if (m3->mc_ki[0] >= split_indx) {
8059 m3->mc_pg[0] = mc->mc_pg[0];
8063 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8064 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8065 m3->mc_ki[mc->mc_top]++;
8066 if (m3->mc_ki[mc->mc_top] >= fixup) {
8067 m3->mc_pg[mc->mc_top] = rp;
8068 m3->mc_ki[mc->mc_top] -= fixup;
8069 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8071 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8072 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8077 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8080 if (copy) /* tmp page */
8081 mdb_page_free(env, copy);
8083 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8088 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8089 MDB_val *key, MDB_val *data, unsigned int flags)
8094 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8097 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8100 mdb_cursor_init(&mc, txn, dbi, &mx);
8101 return mdb_cursor_put(&mc, key, data, flags);
8105 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8107 if ((flag & CHANGEABLE) != flag)
8110 env->me_flags |= flag;
8112 env->me_flags &= ~flag;
8117 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8122 *arg = env->me_flags;
8127 mdb_env_set_userctx(MDB_env *env, void *ctx)
8131 env->me_userctx = ctx;
8136 mdb_env_get_userctx(MDB_env *env)
8138 return env ? env->me_userctx : NULL;
8142 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8147 env->me_assert_func = func;
8153 mdb_env_get_path(MDB_env *env, const char **arg)
8158 *arg = env->me_path;
8163 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8172 /** Common code for #mdb_stat() and #mdb_env_stat().
8173 * @param[in] env the environment to operate in.
8174 * @param[in] db the #MDB_db record containing the stats to return.
8175 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8176 * @return 0, this function always succeeds.
8179 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8181 arg->ms_psize = env->me_psize;
8182 arg->ms_depth = db->md_depth;
8183 arg->ms_branch_pages = db->md_branch_pages;
8184 arg->ms_leaf_pages = db->md_leaf_pages;
8185 arg->ms_overflow_pages = db->md_overflow_pages;
8186 arg->ms_entries = db->md_entries;
8191 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8195 if (env == NULL || arg == NULL)
8198 toggle = mdb_env_pick_meta(env);
8200 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8204 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8208 if (env == NULL || arg == NULL)
8211 toggle = mdb_env_pick_meta(env);
8212 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
8213 arg->me_mapsize = env->me_mapsize;
8214 arg->me_maxreaders = env->me_maxreaders;
8216 /* me_numreaders may be zero if this process never used any readers. Use
8217 * the shared numreader count if it exists.
8219 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8221 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8222 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8226 /** Set the default comparison functions for a database.
8227 * Called immediately after a database is opened to set the defaults.
8228 * The user can then override them with #mdb_set_compare() or
8229 * #mdb_set_dupsort().
8230 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8231 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8234 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8236 uint16_t f = txn->mt_dbs[dbi].md_flags;
8238 txn->mt_dbxs[dbi].md_cmp =
8239 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8240 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8242 txn->mt_dbxs[dbi].md_dcmp =
8243 !(f & MDB_DUPSORT) ? 0 :
8244 ((f & MDB_INTEGERDUP)
8245 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8246 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8249 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8254 int rc, dbflag, exact;
8255 unsigned int unused = 0;
8258 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8259 mdb_default_cmp(txn, FREE_DBI);
8262 if ((flags & VALID_FLAGS) != flags)
8264 if (txn->mt_flags & MDB_TXN_ERROR)
8270 if (flags & PERSISTENT_FLAGS) {
8271 uint16_t f2 = flags & PERSISTENT_FLAGS;
8272 /* make sure flag changes get committed */
8273 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8274 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8275 txn->mt_flags |= MDB_TXN_DIRTY;
8278 mdb_default_cmp(txn, MAIN_DBI);
8282 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8283 mdb_default_cmp(txn, MAIN_DBI);
8286 /* Is the DB already open? */
8288 for (i=2; i<txn->mt_numdbs; i++) {
8289 if (!txn->mt_dbxs[i].md_name.mv_size) {
8290 /* Remember this free slot */
8291 if (!unused) unused = i;
8294 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8295 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8301 /* If no free slot and max hit, fail */
8302 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8303 return MDB_DBS_FULL;
8305 /* Cannot mix named databases with some mainDB flags */
8306 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8307 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8309 /* Find the DB info */
8310 dbflag = DB_NEW|DB_VALID;
8313 key.mv_data = (void *)name;
8314 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8315 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8316 if (rc == MDB_SUCCESS) {
8317 /* make sure this is actually a DB */
8318 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8319 if (!(node->mn_flags & F_SUBDATA))
8320 return MDB_INCOMPATIBLE;
8321 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8322 /* Create if requested */
8324 data.mv_size = sizeof(MDB_db);
8325 data.mv_data = &dummy;
8326 memset(&dummy, 0, sizeof(dummy));
8327 dummy.md_root = P_INVALID;
8328 dummy.md_flags = flags & PERSISTENT_FLAGS;
8329 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8333 /* OK, got info, add to table */
8334 if (rc == MDB_SUCCESS) {
8335 unsigned int slot = unused ? unused : txn->mt_numdbs;
8336 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8337 txn->mt_dbxs[slot].md_name.mv_size = len;
8338 txn->mt_dbxs[slot].md_rel = NULL;
8339 txn->mt_dbflags[slot] = dbflag;
8340 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8342 mdb_default_cmp(txn, slot);
8351 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8353 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8356 if (txn->mt_flags & MDB_TXN_ERROR)
8359 if (txn->mt_dbflags[dbi] & DB_STALE) {
8362 /* Stale, must read the DB's root. cursor_init does it for us. */
8363 mdb_cursor_init(&mc, txn, dbi, &mx);
8365 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8368 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8371 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8373 ptr = env->me_dbxs[dbi].md_name.mv_data;
8374 env->me_dbxs[dbi].md_name.mv_data = NULL;
8375 env->me_dbxs[dbi].md_name.mv_size = 0;
8376 env->me_dbflags[dbi] = 0;
8380 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8382 /* We could return the flags for the FREE_DBI too but what's the point? */
8383 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8385 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8389 /** Add all the DB's pages to the free list.
8390 * @param[in] mc Cursor on the DB to free.
8391 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8392 * @return 0 on success, non-zero on failure.
8395 mdb_drop0(MDB_cursor *mc, int subs)
8399 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8400 if (rc == MDB_SUCCESS) {
8401 MDB_txn *txn = mc->mc_txn;
8406 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8407 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8410 mdb_cursor_copy(mc, &mx);
8411 while (mc->mc_snum > 0) {
8412 MDB_page *mp = mc->mc_pg[mc->mc_top];
8413 unsigned n = NUMKEYS(mp);
8415 for (i=0; i<n; i++) {
8416 ni = NODEPTR(mp, i);
8417 if (ni->mn_flags & F_BIGDATA) {
8420 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8421 rc = mdb_page_get(txn, pg, &omp, NULL);
8424 mdb_cassert(mc, IS_OVERFLOW(omp));
8425 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8429 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8430 mdb_xcursor_init1(mc, ni);
8431 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8437 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8439 for (i=0; i<n; i++) {
8441 ni = NODEPTR(mp, i);
8444 mdb_midl_xappend(txn->mt_free_pgs, pg);
8449 mc->mc_ki[mc->mc_top] = i;
8450 rc = mdb_cursor_sibling(mc, 1);
8452 if (rc != MDB_NOTFOUND)
8454 /* no more siblings, go back to beginning
8455 * of previous level.
8459 for (i=1; i<mc->mc_snum; i++) {
8461 mc->mc_pg[i] = mx.mc_pg[i];
8466 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8469 txn->mt_flags |= MDB_TXN_ERROR;
8470 } else if (rc == MDB_NOTFOUND) {
8476 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8478 MDB_cursor *mc, *m2;
8481 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8484 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8487 rc = mdb_cursor_open(txn, dbi, &mc);
8491 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8492 /* Invalidate the dropped DB's cursors */
8493 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8494 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8498 /* Can't delete the main DB */
8499 if (del && dbi > MAIN_DBI) {
8500 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
8502 txn->mt_dbflags[dbi] = DB_STALE;
8503 mdb_dbi_close(txn->mt_env, dbi);
8505 txn->mt_flags |= MDB_TXN_ERROR;
8508 /* reset the DB record, mark it dirty */
8509 txn->mt_dbflags[dbi] |= DB_DIRTY;
8510 txn->mt_dbs[dbi].md_depth = 0;
8511 txn->mt_dbs[dbi].md_branch_pages = 0;
8512 txn->mt_dbs[dbi].md_leaf_pages = 0;
8513 txn->mt_dbs[dbi].md_overflow_pages = 0;
8514 txn->mt_dbs[dbi].md_entries = 0;
8515 txn->mt_dbs[dbi].md_root = P_INVALID;
8517 txn->mt_flags |= MDB_TXN_DIRTY;
8520 mdb_cursor_close(mc);
8524 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8526 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8529 txn->mt_dbxs[dbi].md_cmp = cmp;
8533 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8535 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8538 txn->mt_dbxs[dbi].md_dcmp = cmp;
8542 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8544 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8547 txn->mt_dbxs[dbi].md_rel = rel;
8551 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8553 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8556 txn->mt_dbxs[dbi].md_relctx = ctx;
8560 int mdb_env_get_maxkeysize(MDB_env *env)
8562 return ENV_MAXKEY(env);
8565 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8567 unsigned int i, rdrs;
8570 int rc = 0, first = 1;
8574 if (!env->me_txns) {
8575 return func("(no reader locks)\n", ctx);
8577 rdrs = env->me_txns->mti_numreaders;
8578 mr = env->me_txns->mti_readers;
8579 for (i=0; i<rdrs; i++) {
8581 txnid_t txnid = mr[i].mr_txnid;
8582 sprintf(buf, txnid == (txnid_t)-1 ?
8583 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
8584 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
8587 rc = func(" pid thread txnid\n", ctx);
8591 rc = func(buf, ctx);
8597 rc = func("(no active readers)\n", ctx);
8602 /** Insert pid into list if not already present.
8603 * return -1 if already present.
8605 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8607 /* binary search of pid in list */
8609 unsigned cursor = 1;
8611 unsigned n = ids[0];
8614 unsigned pivot = n >> 1;
8615 cursor = base + pivot + 1;
8616 val = pid - ids[cursor];
8621 } else if ( val > 0 ) {
8626 /* found, so it's a duplicate */
8635 for (n = ids[0]; n > cursor; n--)
8641 int mdb_reader_check(MDB_env *env, int *dead)
8643 unsigned int i, j, rdrs;
8645 MDB_PID_T *pids, pid;
8654 rdrs = env->me_txns->mti_numreaders;
8655 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8659 mr = env->me_txns->mti_readers;
8660 for (i=0; i<rdrs; i++) {
8661 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8663 if (mdb_pid_insert(pids, pid) == 0) {
8664 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8666 /* Recheck, a new process may have reused pid */
8667 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8668 for (j=i; j<rdrs; j++)
8669 if (mr[j].mr_pid == pid) {
8670 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8671 (unsigned) pid, mr[j].mr_txnid));
8676 UNLOCK_MUTEX_R(env);