2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
42 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
43 * as int64 which is wrong. MSVC doesn't define it at all, so just
48 # include <sys/param.h>
50 # define LITTLE_ENDIAN 1234
51 # define BIG_ENDIAN 4321
52 # define BYTE_ORDER LITTLE_ENDIAN
54 # define SSIZE_MAX INT_MAX
58 #define MDB_PID_T pid_t
59 #include <sys/param.h>
62 #ifdef HAVE_SYS_FILE_H
78 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
79 #include <netinet/in.h>
80 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
83 #if defined(__APPLE__) || defined (BSD)
84 # define MDB_USE_POSIX_SEM 1
85 # define MDB_FDATASYNC fsync
86 #elif defined(ANDROID)
87 # define MDB_FDATASYNC fsync
92 #ifdef MDB_USE_POSIX_SEM
93 # define MDB_USE_HASH 1
94 #include <semaphore.h>
99 #include <valgrind/memcheck.h>
100 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
101 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
102 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
103 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
104 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
106 #define VGMEMP_CREATE(h,r,z)
107 #define VGMEMP_ALLOC(h,a,s)
108 #define VGMEMP_FREE(h,a)
109 #define VGMEMP_DESTROY(h)
110 #define VGMEMP_DEFINED(a,s)
114 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
115 /* Solaris just defines one or the other */
116 # define LITTLE_ENDIAN 1234
117 # define BIG_ENDIAN 4321
118 # ifdef _LITTLE_ENDIAN
119 # define BYTE_ORDER LITTLE_ENDIAN
121 # define BYTE_ORDER BIG_ENDIAN
124 # define BYTE_ORDER __BYTE_ORDER
128 #ifndef LITTLE_ENDIAN
129 #define LITTLE_ENDIAN __LITTLE_ENDIAN
132 #define BIG_ENDIAN __BIG_ENDIAN
135 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
136 #define MISALIGNED_OK 1
142 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
143 # error "Unknown or unsupported endianness (BYTE_ORDER)"
144 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
145 # error "Two's complement, reasonably sized integer types, please"
148 /** @defgroup internal MDB Internals
151 /** @defgroup compat Compatibility Macros
152 * A bunch of macros to minimize the amount of platform-specific ifdefs
153 * needed throughout the rest of the code. When the features this library
154 * needs are similar enough to POSIX to be hidden in a one-or-two line
155 * replacement, this macro approach is used.
159 /** Wrapper around __func__, which is a C99 feature */
160 #if __STDC_VERSION__ >= 199901L
161 # define mdb_func_ __func__
162 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
163 # define mdb_func_ __FUNCTION__
165 /* If a debug message says <mdb_unknown>(), update the #if statements above */
166 # define mdb_func_ "<mdb_unknown>"
170 #define MDB_USE_HASH 1
171 #define MDB_PIDLOCK 0
172 #define pthread_t DWORD
173 #define pthread_mutex_t HANDLE
174 #define pthread_key_t DWORD
175 #define pthread_self() GetCurrentThreadId()
176 #define pthread_key_create(x,y) \
177 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
178 #define pthread_key_delete(x) TlsFree(x)
179 #define pthread_getspecific(x) TlsGetValue(x)
180 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
181 #define pthread_mutex_unlock(x) ReleaseMutex(x)
182 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
183 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
184 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
185 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
186 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
187 #define getpid() GetCurrentProcessId()
188 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
189 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
190 #define ErrCode() GetLastError()
191 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
192 #define close(fd) (CloseHandle(fd) ? 0 : -1)
193 #define munmap(ptr,len) UnmapViewOfFile(ptr)
194 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
195 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
197 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
202 #define Z "z" /**< printf format modifier for size_t */
204 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
205 #define MDB_PIDLOCK 1
207 #ifdef MDB_USE_POSIX_SEM
209 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
210 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
211 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
212 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
215 mdb_sem_wait(sem_t *sem)
218 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
223 /** Lock the reader mutex.
225 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
226 /** Unlock the reader mutex.
228 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
230 /** Lock the writer mutex.
231 * Only a single write transaction is allowed at a time. Other writers
232 * will block waiting for this mutex.
234 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
235 /** Unlock the writer mutex.
237 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
238 #endif /* MDB_USE_POSIX_SEM */
240 /** Get the error code for the last failed system function.
242 #define ErrCode() errno
244 /** An abstraction for a file handle.
245 * On POSIX systems file handles are small integers. On Windows
246 * they're opaque pointers.
250 /** A value for an invalid file handle.
251 * Mainly used to initialize file variables and signify that they are
254 #define INVALID_HANDLE_VALUE (-1)
256 /** Get the size of a memory page for the system.
257 * This is the basic size that the platform's memory manager uses, and is
258 * fundamental to the use of memory-mapped files.
260 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
263 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
266 #define MNAME_LEN (sizeof(pthread_mutex_t))
272 /** A flag for opening a file and requesting synchronous data writes.
273 * This is only used when writing a meta page. It's not strictly needed;
274 * we could just do a normal write and then immediately perform a flush.
275 * But if this flag is available it saves us an extra system call.
277 * @note If O_DSYNC is undefined but exists in /usr/include,
278 * preferably set some compiler flag to get the definition.
279 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
282 # define MDB_DSYNC O_DSYNC
286 /** Function for flushing the data of a file. Define this to fsync
287 * if fdatasync() is not supported.
289 #ifndef MDB_FDATASYNC
290 # define MDB_FDATASYNC fdatasync
294 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
305 /** A page number in the database.
306 * Note that 64 bit page numbers are overkill, since pages themselves
307 * already represent 12-13 bits of addressable memory, and the OS will
308 * always limit applications to a maximum of 63 bits of address space.
310 * @note In the #MDB_node structure, we only store 48 bits of this value,
311 * which thus limits us to only 60 bits of addressable data.
313 typedef MDB_ID pgno_t;
315 /** A transaction ID.
316 * See struct MDB_txn.mt_txnid for details.
318 typedef MDB_ID txnid_t;
320 /** @defgroup debug Debug Macros
324 /** Enable debug output. Needs variable argument macros (a C99 feature).
325 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
326 * read from and written to the database (used for free space management).
332 static int mdb_debug;
333 static txnid_t mdb_debug_start;
335 /** Print a debug message with printf formatting.
336 * Requires double parenthesis around 2 or more args.
338 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
339 # define DPRINTF0(fmt, ...) \
340 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
342 # define DPRINTF(args) ((void) 0)
344 /** Print a debug string.
345 * The string is printed literally, with no format processing.
347 #define DPUTS(arg) DPRINTF(("%s", arg))
348 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
350 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
353 /** @brief The maximum size of a database page.
355 * This is 32k, since it must fit in #MDB_page.%mp_upper.
357 * LMDB will use database pages < OS pages if needed.
358 * That causes more I/O in write transactions: The OS must
359 * know (read) the whole page before writing a partial page.
361 * Note that we don't currently support Huge pages. On Linux,
362 * regular data files cannot use Huge pages, and in general
363 * Huge pages aren't actually pageable. We rely on the OS
364 * demand-pager to read our data and page it out when memory
365 * pressure from other processes is high. So until OSs have
366 * actual paging support for Huge pages, they're not viable.
368 #define MAX_PAGESIZE 0x8000
370 /** The minimum number of keys required in a database page.
371 * Setting this to a larger value will place a smaller bound on the
372 * maximum size of a data item. Data items larger than this size will
373 * be pushed into overflow pages instead of being stored directly in
374 * the B-tree node. This value used to default to 4. With a page size
375 * of 4096 bytes that meant that any item larger than 1024 bytes would
376 * go into an overflow page. That also meant that on average 2-3KB of
377 * each overflow page was wasted space. The value cannot be lower than
378 * 2 because then there would no longer be a tree structure. With this
379 * value, items larger than 2KB will go into overflow pages, and on
380 * average only 1KB will be wasted.
382 #define MDB_MINKEYS 2
384 /** A stamp that identifies a file as an MDB file.
385 * There's nothing special about this value other than that it is easily
386 * recognizable, and it will reflect any byte order mismatches.
388 #define MDB_MAGIC 0xBEEFC0DE
390 /** The version number for a database's datafile format. */
391 #define MDB_DATA_VERSION 1
392 /** The version number for a database's lockfile format. */
393 #define MDB_LOCK_VERSION 1
395 /** @brief The max size of a key we can write, or 0 for dynamic max.
397 * Define this as 0 to compute the max from the page size. 511
398 * is default for backwards compat: liblmdb <= 0.9.10 can break
399 * when modifying a DB with keys/dupsort data bigger than its max.
401 * Data items in an #MDB_DUPSORT database are also limited to
402 * this size, since they're actually keys of a sub-DB. Keys and
403 * #MDB_DUPSORT data items must fit on a node in a regular page.
405 #ifndef MDB_MAXKEYSIZE
406 #define MDB_MAXKEYSIZE 511
409 /** The maximum size of a key we can write to the environment. */
411 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
413 #define ENV_MAXKEY(env) ((env)->me_maxkey)
416 /** @brief The maximum size of a data item.
418 * We only store a 32 bit value for node sizes.
420 #define MAXDATASIZE 0xffffffffUL
423 /** Key size which fits in a #DKBUF.
426 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
429 * This is used for printing a hex dump of a key's contents.
431 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
432 /** Display a key in hex.
434 * Invoke a function to display a key in hex.
436 #define DKEY(x) mdb_dkey(x, kbuf)
442 /** An invalid page number.
443 * Mainly used to denote an empty tree.
445 #define P_INVALID (~(pgno_t)0)
447 /** Test if the flags \b f are set in a flag word \b w. */
448 #define F_ISSET(w, f) (((w) & (f)) == (f))
450 /** Round \b n up to an even number. */
451 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
453 /** Used for offsets within a single page.
454 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
457 typedef uint16_t indx_t;
459 /** Default size of memory map.
460 * This is certainly too small for any actual applications. Apps should always set
461 * the size explicitly using #mdb_env_set_mapsize().
463 #define DEFAULT_MAPSIZE 1048576
465 /** @defgroup readers Reader Lock Table
466 * Readers don't acquire any locks for their data access. Instead, they
467 * simply record their transaction ID in the reader table. The reader
468 * mutex is needed just to find an empty slot in the reader table. The
469 * slot's address is saved in thread-specific data so that subsequent read
470 * transactions started by the same thread need no further locking to proceed.
472 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
474 * No reader table is used if the database is on a read-only filesystem, or
475 * if #MDB_NOLOCK is set.
477 * Since the database uses multi-version concurrency control, readers don't
478 * actually need any locking. This table is used to keep track of which
479 * readers are using data from which old transactions, so that we'll know
480 * when a particular old transaction is no longer in use. Old transactions
481 * that have discarded any data pages can then have those pages reclaimed
482 * for use by a later write transaction.
484 * The lock table is constructed such that reader slots are aligned with the
485 * processor's cache line size. Any slot is only ever used by one thread.
486 * This alignment guarantees that there will be no contention or cache
487 * thrashing as threads update their own slot info, and also eliminates
488 * any need for locking when accessing a slot.
490 * A writer thread will scan every slot in the table to determine the oldest
491 * outstanding reader transaction. Any freed pages older than this will be
492 * reclaimed by the writer. The writer doesn't use any locks when scanning
493 * this table. This means that there's no guarantee that the writer will
494 * see the most up-to-date reader info, but that's not required for correct
495 * operation - all we need is to know the upper bound on the oldest reader,
496 * we don't care at all about the newest reader. So the only consequence of
497 * reading stale information here is that old pages might hang around a
498 * while longer before being reclaimed. That's actually good anyway, because
499 * the longer we delay reclaiming old pages, the more likely it is that a
500 * string of contiguous pages can be found after coalescing old pages from
501 * many old transactions together.
504 /** Number of slots in the reader table.
505 * This value was chosen somewhat arbitrarily. 126 readers plus a
506 * couple mutexes fit exactly into 8KB on my development machine.
507 * Applications should set the table size using #mdb_env_set_maxreaders().
509 #define DEFAULT_READERS 126
511 /** The size of a CPU cache line in bytes. We want our lock structures
512 * aligned to this size to avoid false cache line sharing in the
514 * This value works for most CPUs. For Itanium this should be 128.
520 /** The information we store in a single slot of the reader table.
521 * In addition to a transaction ID, we also record the process and
522 * thread ID that owns a slot, so that we can detect stale information,
523 * e.g. threads or processes that went away without cleaning up.
524 * @note We currently don't check for stale records. We simply re-init
525 * the table when we know that we're the only process opening the
528 typedef struct MDB_rxbody {
529 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
530 * Multiple readers that start at the same time will probably have the
531 * same ID here. Again, it's not important to exclude them from
532 * anything; all we need to know is which version of the DB they
533 * started from so we can avoid overwriting any data used in that
534 * particular version.
537 /** The process ID of the process owning this reader txn. */
539 /** The thread ID of the thread owning this txn. */
543 /** The actual reader record, with cacheline padding. */
544 typedef struct MDB_reader {
547 /** shorthand for mrb_txnid */
548 #define mr_txnid mru.mrx.mrb_txnid
549 #define mr_pid mru.mrx.mrb_pid
550 #define mr_tid mru.mrx.mrb_tid
551 /** cache line alignment */
552 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
556 /** The header for the reader table.
557 * The table resides in a memory-mapped file. (This is a different file
558 * than is used for the main database.)
560 * For POSIX the actual mutexes reside in the shared memory of this
561 * mapped file. On Windows, mutexes are named objects allocated by the
562 * kernel; we store the mutex names in this mapped file so that other
563 * processes can grab them. This same approach is also used on
564 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
565 * process-shared POSIX mutexes. For these cases where a named object
566 * is used, the object name is derived from a 64 bit FNV hash of the
567 * environment pathname. As such, naming collisions are extremely
568 * unlikely. If a collision occurs, the results are unpredictable.
570 typedef struct MDB_txbody {
571 /** Stamp identifying this as an MDB file. It must be set
574 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
576 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
577 char mtb_rmname[MNAME_LEN];
579 /** Mutex protecting access to this table.
580 * This is the reader lock that #LOCK_MUTEX_R acquires.
582 pthread_mutex_t mtb_mutex;
584 /** The ID of the last transaction committed to the database.
585 * This is recorded here only for convenience; the value can always
586 * be determined by reading the main database meta pages.
589 /** The number of slots that have been used in the reader table.
590 * This always records the maximum count, it is not decremented
591 * when readers release their slots.
593 unsigned mtb_numreaders;
596 /** The actual reader table definition. */
597 typedef struct MDB_txninfo {
600 #define mti_magic mt1.mtb.mtb_magic
601 #define mti_format mt1.mtb.mtb_format
602 #define mti_mutex mt1.mtb.mtb_mutex
603 #define mti_rmname mt1.mtb.mtb_rmname
604 #define mti_txnid mt1.mtb.mtb_txnid
605 #define mti_numreaders mt1.mtb.mtb_numreaders
606 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
609 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
610 char mt2_wmname[MNAME_LEN];
611 #define mti_wmname mt2.mt2_wmname
613 pthread_mutex_t mt2_wmutex;
614 #define mti_wmutex mt2.mt2_wmutex
616 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
618 MDB_reader mti_readers[1];
621 /** Lockfile format signature: version, features and field layout */
622 #define MDB_LOCK_FORMAT \
624 ((MDB_LOCK_VERSION) \
625 /* Flags which describe functionality */ \
626 + (((MDB_PIDLOCK) != 0) << 16)))
629 /** Common header for all page types.
630 * Overflow records occupy a number of contiguous pages with no
631 * headers on any page after the first.
633 typedef struct MDB_page {
634 #define mp_pgno mp_p.p_pgno
635 #define mp_next mp_p.p_next
637 pgno_t p_pgno; /**< page number */
638 void * p_next; /**< for in-memory list of freed structs */
641 /** @defgroup mdb_page Page Flags
643 * Flags for the page headers.
646 #define P_BRANCH 0x01 /**< branch page */
647 #define P_LEAF 0x02 /**< leaf page */
648 #define P_OVERFLOW 0x04 /**< overflow page */
649 #define P_META 0x08 /**< meta page */
650 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
651 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
652 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
653 #define P_KEEP 0x8000 /**< leave this page alone during spill */
655 uint16_t mp_flags; /**< @ref mdb_page */
656 #define mp_lower mp_pb.pb.pb_lower
657 #define mp_upper mp_pb.pb.pb_upper
658 #define mp_pages mp_pb.pb_pages
661 indx_t pb_lower; /**< lower bound of free space */
662 indx_t pb_upper; /**< upper bound of free space */
664 uint32_t pb_pages; /**< number of overflow pages */
666 indx_t mp_ptrs[1]; /**< dynamic size */
669 /** Size of the page header, excluding dynamic data at the end */
670 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
672 /** Address of first usable data byte in a page, after the header */
673 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
675 /** Number of nodes on a page */
676 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
678 /** The amount of space remaining in the page */
679 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
681 /** The percentage of space used in the page, in tenths of a percent. */
682 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
683 ((env)->me_psize - PAGEHDRSZ))
684 /** The minimum page fill factor, in tenths of a percent.
685 * Pages emptier than this are candidates for merging.
687 #define FILL_THRESHOLD 250
689 /** Test if a page is a leaf page */
690 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
691 /** Test if a page is a LEAF2 page */
692 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
693 /** Test if a page is a branch page */
694 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
695 /** Test if a page is an overflow page */
696 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
697 /** Test if a page is a sub page */
698 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
700 /** The number of overflow pages needed to store the given size. */
701 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
703 /** Header for a single key/data pair within a page.
704 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
705 * We guarantee 2-byte alignment for 'MDB_node's.
707 typedef struct MDB_node {
708 /** lo and hi are used for data size on leaf nodes and for
709 * child pgno on branch nodes. On 64 bit platforms, flags
710 * is also used for pgno. (Branch nodes have no flags).
711 * They are in host byte order in case that lets some
712 * accesses be optimized into a 32-bit word access.
714 #if BYTE_ORDER == LITTLE_ENDIAN
715 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
717 unsigned short mn_hi, mn_lo;
719 /** @defgroup mdb_node Node Flags
721 * Flags for node headers.
724 #define F_BIGDATA 0x01 /**< data put on overflow page */
725 #define F_SUBDATA 0x02 /**< data is a sub-database */
726 #define F_DUPDATA 0x04 /**< data has duplicates */
728 /** valid flags for #mdb_node_add() */
729 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
732 unsigned short mn_flags; /**< @ref mdb_node */
733 unsigned short mn_ksize; /**< key size */
734 char mn_data[1]; /**< key and data are appended here */
737 /** Size of the node header, excluding dynamic data at the end */
738 #define NODESIZE offsetof(MDB_node, mn_data)
740 /** Bit position of top word in page number, for shifting mn_flags */
741 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
743 /** Size of a node in a branch page with a given key.
744 * This is just the node header plus the key, there is no data.
746 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
748 /** Size of a node in a leaf page with a given key and data.
749 * This is node header plus key plus data size.
751 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
753 /** Address of node \b i in page \b p */
754 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
756 /** Address of the key for the node */
757 #define NODEKEY(node) (void *)((node)->mn_data)
759 /** Address of the data for a node */
760 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
762 /** Get the page number pointed to by a branch node */
763 #define NODEPGNO(node) \
764 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
765 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
766 /** Set the page number in a branch node */
767 #define SETPGNO(node,pgno) do { \
768 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
769 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
771 /** Get the size of the data in a leaf node */
772 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
773 /** Set the size of the data for a leaf node */
774 #define SETDSZ(node,size) do { \
775 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
776 /** The size of a key in a node */
777 #define NODEKSZ(node) ((node)->mn_ksize)
779 /** Copy a page number from src to dst */
781 #define COPY_PGNO(dst,src) dst = src
783 #if SIZE_MAX > 4294967295UL
784 #define COPY_PGNO(dst,src) do { \
785 unsigned short *s, *d; \
786 s = (unsigned short *)&(src); \
787 d = (unsigned short *)&(dst); \
794 #define COPY_PGNO(dst,src) do { \
795 unsigned short *s, *d; \
796 s = (unsigned short *)&(src); \
797 d = (unsigned short *)&(dst); \
803 /** The address of a key in a LEAF2 page.
804 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
805 * There are no node headers, keys are stored contiguously.
807 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
809 /** Set the \b node's key into \b keyptr, if requested. */
810 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
811 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
813 /** Set the \b node's key into \b key. */
814 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
816 /** Information about a single database in the environment. */
817 typedef struct MDB_db {
818 uint32_t md_pad; /**< also ksize for LEAF2 pages */
819 uint16_t md_flags; /**< @ref mdb_dbi_open */
820 uint16_t md_depth; /**< depth of this tree */
821 pgno_t md_branch_pages; /**< number of internal pages */
822 pgno_t md_leaf_pages; /**< number of leaf pages */
823 pgno_t md_overflow_pages; /**< number of overflow pages */
824 size_t md_entries; /**< number of data items */
825 pgno_t md_root; /**< the root page of this tree */
828 /** mdb_dbi_open flags */
829 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
830 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
831 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
832 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
834 /** Handle for the DB used to track free pages. */
836 /** Handle for the default DB. */
839 /** Meta page content.
840 * A meta page is the start point for accessing a database snapshot.
841 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
843 typedef struct MDB_meta {
844 /** Stamp identifying this as an MDB file. It must be set
847 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
849 void *mm_address; /**< address for fixed mapping */
850 size_t mm_mapsize; /**< size of mmap region */
851 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
852 /** The size of pages used in this DB */
853 #define mm_psize mm_dbs[0].md_pad
854 /** Any persistent environment flags. @ref mdb_env */
855 #define mm_flags mm_dbs[0].md_flags
856 pgno_t mm_last_pg; /**< last used page in file */
857 txnid_t mm_txnid; /**< txnid that committed this page */
860 /** Buffer for a stack-allocated meta page.
861 * The members define size and alignment, and silence type
862 * aliasing warnings. They are not used directly; that could
863 * mean incorrectly using several union members in parallel.
865 typedef union MDB_metabuf {
868 char mm_pad[PAGEHDRSZ];
873 /** Auxiliary DB info.
874 * The information here is mostly static/read-only. There is
875 * only a single copy of this record in the environment.
877 typedef struct MDB_dbx {
878 MDB_val md_name; /**< name of the database */
879 MDB_cmp_func *md_cmp; /**< function for comparing keys */
880 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
881 MDB_rel_func *md_rel; /**< user relocate function */
882 void *md_relctx; /**< user-provided context for md_rel */
885 /** A database transaction.
886 * Every operation requires a transaction handle.
889 MDB_txn *mt_parent; /**< parent of a nested txn */
890 MDB_txn *mt_child; /**< nested txn under this txn */
891 pgno_t mt_next_pgno; /**< next unallocated page */
892 /** The ID of this transaction. IDs are integers incrementing from 1.
893 * Only committed write transactions increment the ID. If a transaction
894 * aborts, the ID may be re-used by the next writer.
897 MDB_env *mt_env; /**< the DB environment */
898 /** The list of pages that became unused during this transaction.
901 /** The sorted list of dirty pages we temporarily wrote to disk
902 * because the dirty list was full. page numbers in here are
903 * shifted left by 1, deleted slots have the LSB set.
905 MDB_IDL mt_spill_pgs;
907 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
909 /** For read txns: This thread/txn's reader table slot, or NULL. */
912 /** Array of records for each DB known in the environment. */
914 /** Array of MDB_db records for each known DB */
916 /** @defgroup mt_dbflag Transaction DB Flags
920 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
921 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
922 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
923 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
925 /** In write txns, array of cursors for each DB */
926 MDB_cursor **mt_cursors;
927 /** Array of flags for each DB */
928 unsigned char *mt_dbflags;
929 /** Number of DB records in use. This number only ever increments;
930 * we don't decrement it when individual DB handles are closed.
934 /** @defgroup mdb_txn Transaction Flags
938 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
939 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
940 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
941 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
943 unsigned int mt_flags; /**< @ref mdb_txn */
944 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
945 * Includes ancestor txns' dirty pages not hidden by other txns'
946 * dirty/spilled pages. Thus commit(nested txn) has room to merge
947 * dirty_list into mt_parent after freeing hidden mt_parent pages.
949 unsigned int mt_dirty_room;
952 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
953 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
954 * raise this on a 64 bit machine.
956 #define CURSOR_STACK 32
960 /** Cursors are used for all DB operations.
961 * A cursor holds a path of (page pointer, key index) from the DB
962 * root to a position in the DB, plus other state. #MDB_DUPSORT
963 * cursors include an xcursor to the current data item. Write txns
964 * track their cursors and keep them up to date when data moves.
965 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
966 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
969 /** Next cursor on this DB in this txn */
971 /** Backup of the original cursor if this cursor is a shadow */
972 MDB_cursor *mc_backup;
973 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
974 struct MDB_xcursor *mc_xcursor;
975 /** The transaction that owns this cursor */
977 /** The database handle this cursor operates on */
979 /** The database record for this cursor */
981 /** The database auxiliary record for this cursor */
983 /** The @ref mt_dbflag for this database */
984 unsigned char *mc_dbflag;
985 unsigned short mc_snum; /**< number of pushed pages */
986 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
987 /** @defgroup mdb_cursor Cursor Flags
989 * Cursor state flags.
992 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
993 #define C_EOF 0x02 /**< No more data */
994 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
995 #define C_DEL 0x08 /**< last op was a cursor_del */
996 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
997 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
999 unsigned int mc_flags; /**< @ref mdb_cursor */
1000 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1001 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1004 /** Context for sorted-dup records.
1005 * We could have gone to a fully recursive design, with arbitrarily
1006 * deep nesting of sub-databases. But for now we only handle these
1007 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1009 typedef struct MDB_xcursor {
1010 /** A sub-cursor for traversing the Dup DB */
1011 MDB_cursor mx_cursor;
1012 /** The database record for this Dup DB */
1014 /** The auxiliary DB record for this Dup DB */
1016 /** The @ref mt_dbflag for this Dup DB */
1017 unsigned char mx_dbflag;
1020 /** State of FreeDB old pages, stored in the MDB_env */
1021 typedef struct MDB_pgstate {
1022 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1023 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1026 /** The database environment. */
1028 HANDLE me_fd; /**< The main data file */
1029 HANDLE me_lfd; /**< The lock file */
1030 HANDLE me_mfd; /**< just for writing the meta pages */
1031 /** Failed to update the meta page. Probably an I/O error. */
1032 #define MDB_FATAL_ERROR 0x80000000U
1033 /** Some fields are initialized. */
1034 #define MDB_ENV_ACTIVE 0x20000000U
1035 /** me_txkey is set */
1036 #define MDB_ENV_TXKEY 0x10000000U
1037 /** Have liveness lock in reader table */
1038 #define MDB_LIVE_READER 0x08000000U
1039 uint32_t me_flags; /**< @ref mdb_env */
1040 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1041 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1042 unsigned int me_maxreaders; /**< size of the reader table */
1043 unsigned int me_numreaders; /**< max numreaders set by this env */
1044 MDB_dbi me_numdbs; /**< number of DBs opened */
1045 MDB_dbi me_maxdbs; /**< size of the DB table */
1046 MDB_PID_T me_pid; /**< process ID of this env */
1047 char *me_path; /**< path to the DB files */
1048 char *me_map; /**< the memory map of the data file */
1049 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1050 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1051 void *me_pbuf; /**< scratch area for DUPSORT put() */
1052 MDB_txn *me_txn; /**< current write transaction */
1053 size_t me_mapsize; /**< size of the data memory map */
1054 off_t me_size; /**< current file size */
1055 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1056 MDB_dbx *me_dbxs; /**< array of static DB info */
1057 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1058 pthread_key_t me_txkey; /**< thread-key for readers */
1059 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1060 # define me_pglast me_pgstate.mf_pglast
1061 # define me_pghead me_pgstate.mf_pghead
1062 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1063 /** IDL of pages that became unused in a write txn */
1064 MDB_IDL me_free_pgs;
1065 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1066 MDB_ID2L me_dirty_list;
1067 /** Max number of freelist items that can fit in a single overflow page */
1069 /** Max size of a node on a page */
1070 unsigned int me_nodemax;
1071 #if !(MDB_MAXKEYSIZE)
1072 unsigned int me_maxkey; /**< max size of a key */
1075 int me_pidquery; /**< Used in OpenProcess */
1076 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1078 #elif defined(MDB_USE_POSIX_SEM)
1079 sem_t *me_rmutex; /* Shared mutexes are not supported */
1082 void *me_userctx; /**< User-settable context */
1083 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1086 /** Nested transaction */
1087 typedef struct MDB_ntxn {
1088 MDB_txn mnt_txn; /**< the transaction */
1089 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1092 /** max number of pages to commit in one writev() call */
1093 #define MDB_COMMIT_PAGES 64
1094 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1095 #undef MDB_COMMIT_PAGES
1096 #define MDB_COMMIT_PAGES IOV_MAX
1099 /** max bytes to write in one call */
1100 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1102 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1103 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1104 static int mdb_page_touch(MDB_cursor *mc);
1106 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1107 static int mdb_page_search_root(MDB_cursor *mc,
1108 MDB_val *key, int modify);
1109 #define MDB_PS_MODIFY 1
1110 #define MDB_PS_ROOTONLY 2
1111 #define MDB_PS_FIRST 4
1112 #define MDB_PS_LAST 8
1113 static int mdb_page_search(MDB_cursor *mc,
1114 MDB_val *key, int flags);
1115 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1117 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1118 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1119 pgno_t newpgno, unsigned int nflags);
1121 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1122 static int mdb_env_pick_meta(const MDB_env *env);
1123 static int mdb_env_write_meta(MDB_txn *txn);
1124 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1125 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1127 static void mdb_env_close0(MDB_env *env, int excl);
1129 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1130 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1131 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1132 static void mdb_node_del(MDB_cursor *mc, int ksize);
1133 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1134 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1135 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1136 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1137 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1139 static int mdb_rebalance(MDB_cursor *mc);
1140 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1142 static void mdb_cursor_pop(MDB_cursor *mc);
1143 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1145 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1146 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1147 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1148 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1149 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1151 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1152 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1154 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1155 static void mdb_xcursor_init0(MDB_cursor *mc);
1156 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1158 static int mdb_drop0(MDB_cursor *mc, int subs);
1159 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1162 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1166 static SECURITY_DESCRIPTOR mdb_null_sd;
1167 static SECURITY_ATTRIBUTES mdb_all_sa;
1168 static int mdb_sec_inited;
1171 /** Return the library version info. */
1173 mdb_version(int *major, int *minor, int *patch)
1175 if (major) *major = MDB_VERSION_MAJOR;
1176 if (minor) *minor = MDB_VERSION_MINOR;
1177 if (patch) *patch = MDB_VERSION_PATCH;
1178 return MDB_VERSION_STRING;
1181 /** Table of descriptions for MDB @ref errors */
1182 static char *const mdb_errstr[] = {
1183 "MDB_KEYEXIST: Key/data pair already exists",
1184 "MDB_NOTFOUND: No matching key/data pair found",
1185 "MDB_PAGE_NOTFOUND: Requested page not found",
1186 "MDB_CORRUPTED: Located page was wrong type",
1187 "MDB_PANIC: Update of meta page failed",
1188 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1189 "MDB_INVALID: File is not an MDB file",
1190 "MDB_MAP_FULL: Environment mapsize limit reached",
1191 "MDB_DBS_FULL: Environment maxdbs limit reached",
1192 "MDB_READERS_FULL: Environment maxreaders limit reached",
1193 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1194 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1195 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1196 "MDB_PAGE_FULL: Internal error - page has no more space",
1197 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1198 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1199 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1200 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1201 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1205 mdb_strerror(int err)
1209 return ("Successful return: 0");
1211 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1212 i = err - MDB_KEYEXIST;
1213 return mdb_errstr[i];
1216 return strerror(err);
1219 /** assert(3) variant in cursor context */
1220 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1221 /** assert(3) variant in transaction context */
1222 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1223 /** assert(3) variant in environment context */
1224 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1227 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1228 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1231 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1232 const char *func, const char *file, int line)
1235 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1236 file, line, expr_txt, func);
1237 if (env->me_assert_func)
1238 env->me_assert_func(env, buf);
1239 fprintf(stderr, "%s\n", buf);
1243 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1247 /** Return the page number of \b mp which may be sub-page, for debug output */
1249 mdb_dbg_pgno(MDB_page *mp)
1252 COPY_PGNO(ret, mp->mp_pgno);
1256 /** Display a key in hexadecimal and return the address of the result.
1257 * @param[in] key the key to display
1258 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1259 * @return The key in hexadecimal form.
1262 mdb_dkey(MDB_val *key, char *buf)
1265 unsigned char *c = key->mv_data;
1271 if (key->mv_size > DKBUF_MAXKEYSIZE)
1272 return "MDB_MAXKEYSIZE";
1273 /* may want to make this a dynamic check: if the key is mostly
1274 * printable characters, print it as-is instead of converting to hex.
1278 for (i=0; i<key->mv_size; i++)
1279 ptr += sprintf(ptr, "%02x", *c++);
1281 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1286 /** Display all the keys in the page. */
1288 mdb_page_list(MDB_page *mp)
1291 unsigned int i, nkeys, nsize, total = 0;
1295 nkeys = NUMKEYS(mp);
1296 fprintf(stderr, "Page %"Z"u numkeys %d\n", mdb_dbg_pgno(mp), nkeys);
1297 for (i=0; i<nkeys; i++) {
1298 node = NODEPTR(mp, i);
1299 key.mv_size = node->mn_ksize;
1300 key.mv_data = node->mn_data;
1301 nsize = NODESIZE + key.mv_size;
1302 if (IS_BRANCH(mp)) {
1303 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1307 if (F_ISSET(node->mn_flags, F_BIGDATA))
1308 nsize += sizeof(pgno_t);
1310 nsize += NODEDSZ(node);
1312 nsize += sizeof(indx_t);
1313 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1315 total = EVEN(total);
1317 fprintf(stderr, "Total: %d\n", total);
1321 mdb_cursor_chk(MDB_cursor *mc)
1327 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1328 for (i=0; i<mc->mc_top; i++) {
1330 node = NODEPTR(mp, mc->mc_ki[i]);
1331 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1334 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1340 /** Count all the pages in each DB and in the freelist
1341 * and make sure it matches the actual number of pages
1344 static void mdb_audit(MDB_txn *txn)
1348 MDB_ID freecount, count;
1353 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1354 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1355 freecount += *(MDB_ID *)data.mv_data;
1358 for (i = 0; i<txn->mt_numdbs; i++) {
1360 mdb_cursor_init(&mc, txn, i, &mx);
1361 if (txn->mt_dbs[i].md_root == P_INVALID)
1363 count += txn->mt_dbs[i].md_branch_pages +
1364 txn->mt_dbs[i].md_leaf_pages +
1365 txn->mt_dbs[i].md_overflow_pages;
1366 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1367 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1371 mp = mc.mc_pg[mc.mc_top];
1372 for (j=0; j<NUMKEYS(mp); j++) {
1373 MDB_node *leaf = NODEPTR(mp, j);
1374 if (leaf->mn_flags & F_SUBDATA) {
1376 memcpy(&db, NODEDATA(leaf), sizeof(db));
1377 count += db.md_branch_pages + db.md_leaf_pages +
1378 db.md_overflow_pages;
1382 while (mdb_cursor_sibling(&mc, 1) == 0);
1385 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1386 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1387 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1393 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1395 return txn->mt_dbxs[dbi].md_cmp(a, b);
1399 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1401 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1404 /** Allocate memory for a page.
1405 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1408 mdb_page_malloc(MDB_txn *txn, unsigned num)
1410 MDB_env *env = txn->mt_env;
1411 MDB_page *ret = env->me_dpages;
1412 size_t psize = env->me_psize, sz = psize, off;
1413 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1414 * For a single page alloc, we init everything after the page header.
1415 * For multi-page, we init the final page; if the caller needed that
1416 * many pages they will be filling in at least up to the last page.
1420 VGMEMP_ALLOC(env, ret, sz);
1421 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1422 env->me_dpages = ret->mp_next;
1425 psize -= off = PAGEHDRSZ;
1430 if ((ret = malloc(sz)) != NULL) {
1431 VGMEMP_ALLOC(env, ret, sz);
1432 if (!(env->me_flags & MDB_NOMEMINIT)) {
1433 memset((char *)ret + off, 0, psize);
1437 txn->mt_flags |= MDB_TXN_ERROR;
1442 /** Free a single page.
1443 * Saves single pages to a list, for future reuse.
1444 * (This is not used for multi-page overflow pages.)
1447 mdb_page_free(MDB_env *env, MDB_page *mp)
1449 mp->mp_next = env->me_dpages;
1450 VGMEMP_FREE(env, mp);
1451 env->me_dpages = mp;
1454 /** Free a dirty page */
1456 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1458 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1459 mdb_page_free(env, dp);
1461 /* large pages just get freed directly */
1462 VGMEMP_FREE(env, dp);
1467 /** Return all dirty pages to dpage list */
1469 mdb_dlist_free(MDB_txn *txn)
1471 MDB_env *env = txn->mt_env;
1472 MDB_ID2L dl = txn->mt_u.dirty_list;
1473 unsigned i, n = dl[0].mid;
1475 for (i = 1; i <= n; i++) {
1476 mdb_dpage_free(env, dl[i].mptr);
1481 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1482 * @param[in] mc A cursor handle for the current operation.
1483 * @param[in] pflags Flags of the pages to update:
1484 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1485 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1486 * @return 0 on success, non-zero on failure.
1489 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1491 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1492 MDB_txn *txn = mc->mc_txn;
1498 int rc = MDB_SUCCESS, level;
1500 /* Mark pages seen by cursors */
1501 if (mc->mc_flags & C_UNTRACK)
1502 mc = NULL; /* will find mc in mt_cursors */
1503 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1504 for (; mc; mc=mc->mc_next) {
1505 if (!(mc->mc_flags & C_INITIALIZED))
1507 for (m3 = mc;; m3 = &mx->mx_cursor) {
1509 for (j=0; j<m3->mc_snum; j++) {
1511 if ((mp->mp_flags & Mask) == pflags)
1512 mp->mp_flags ^= P_KEEP;
1514 mx = m3->mc_xcursor;
1515 /* Proceed to mx if it is at a sub-database */
1516 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1518 if (! (mp && (mp->mp_flags & P_LEAF)))
1520 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1521 if (!(leaf->mn_flags & F_SUBDATA))
1530 /* Mark dirty root pages */
1531 for (i=0; i<txn->mt_numdbs; i++) {
1532 if (txn->mt_dbflags[i] & DB_DIRTY) {
1533 pgno_t pgno = txn->mt_dbs[i].md_root;
1534 if (pgno == P_INVALID)
1536 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1538 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1539 dp->mp_flags ^= P_KEEP;
1547 static int mdb_page_flush(MDB_txn *txn, int keep);
1549 /** Spill pages from the dirty list back to disk.
1550 * This is intended to prevent running into #MDB_TXN_FULL situations,
1551 * but note that they may still occur in a few cases:
1552 * 1) our estimate of the txn size could be too small. Currently this
1553 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1554 * 2) child txns may run out of space if their parents dirtied a
1555 * lot of pages and never spilled them. TODO: we probably should do
1556 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1557 * the parent's dirty_room is below a given threshold.
1559 * Otherwise, if not using nested txns, it is expected that apps will
1560 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1561 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1562 * If the txn never references them again, they can be left alone.
1563 * If the txn only reads them, they can be used without any fuss.
1564 * If the txn writes them again, they can be dirtied immediately without
1565 * going thru all of the work of #mdb_page_touch(). Such references are
1566 * handled by #mdb_page_unspill().
1568 * Also note, we never spill DB root pages, nor pages of active cursors,
1569 * because we'll need these back again soon anyway. And in nested txns,
1570 * we can't spill a page in a child txn if it was already spilled in a
1571 * parent txn. That would alter the parent txns' data even though
1572 * the child hasn't committed yet, and we'd have no way to undo it if
1573 * the child aborted.
1575 * @param[in] m0 cursor A cursor handle identifying the transaction and
1576 * database for which we are checking space.
1577 * @param[in] key For a put operation, the key being stored.
1578 * @param[in] data For a put operation, the data being stored.
1579 * @return 0 on success, non-zero on failure.
1582 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1584 MDB_txn *txn = m0->mc_txn;
1586 MDB_ID2L dl = txn->mt_u.dirty_list;
1587 unsigned int i, j, need;
1590 if (m0->mc_flags & C_SUB)
1593 /* Estimate how much space this op will take */
1594 i = m0->mc_db->md_depth;
1595 /* Named DBs also dirty the main DB */
1596 if (m0->mc_dbi > MAIN_DBI)
1597 i += txn->mt_dbs[MAIN_DBI].md_depth;
1598 /* For puts, roughly factor in the key+data size */
1600 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1601 i += i; /* double it for good measure */
1604 if (txn->mt_dirty_room > i)
1607 if (!txn->mt_spill_pgs) {
1608 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1609 if (!txn->mt_spill_pgs)
1612 /* purge deleted slots */
1613 MDB_IDL sl = txn->mt_spill_pgs;
1614 unsigned int num = sl[0];
1616 for (i=1; i<=num; i++) {
1623 /* Preserve pages which may soon be dirtied again */
1624 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1627 /* Less aggressive spill - we originally spilled the entire dirty list,
1628 * with a few exceptions for cursor pages and DB root pages. But this
1629 * turns out to be a lot of wasted effort because in a large txn many
1630 * of those pages will need to be used again. So now we spill only 1/8th
1631 * of the dirty pages. Testing revealed this to be a good tradeoff,
1632 * better than 1/2, 1/4, or 1/10.
1634 if (need < MDB_IDL_UM_MAX / 8)
1635 need = MDB_IDL_UM_MAX / 8;
1637 /* Save the page IDs of all the pages we're flushing */
1638 /* flush from the tail forward, this saves a lot of shifting later on. */
1639 for (i=dl[0].mid; i && need; i--) {
1640 MDB_ID pn = dl[i].mid << 1;
1642 if (dp->mp_flags & P_KEEP)
1644 /* Can't spill twice, make sure it's not already in a parent's
1647 if (txn->mt_parent) {
1649 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1650 if (tx2->mt_spill_pgs) {
1651 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1652 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1653 dp->mp_flags |= P_KEEP;
1661 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1665 mdb_midl_sort(txn->mt_spill_pgs);
1667 /* Flush the spilled part of dirty list */
1668 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1671 /* Reset any dirty pages we kept that page_flush didn't see */
1672 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1675 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1679 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1681 mdb_find_oldest(MDB_txn *txn)
1684 txnid_t mr, oldest = txn->mt_txnid - 1;
1685 if (txn->mt_env->me_txns) {
1686 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1687 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1698 /** Add a page to the txn's dirty list */
1700 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1703 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1705 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1706 insert = mdb_mid2l_append;
1708 insert = mdb_mid2l_insert;
1710 mid.mid = mp->mp_pgno;
1712 rc = insert(txn->mt_u.dirty_list, &mid);
1713 mdb_tassert(txn, rc == 0);
1714 txn->mt_dirty_room--;
1717 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1718 * me_pghead and mt_next_pgno.
1720 * If there are free pages available from older transactions, they
1721 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1722 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1723 * and move me_pglast to say which records were consumed. Only this
1724 * function can create me_pghead and move me_pglast/mt_next_pgno.
1725 * @param[in] mc cursor A cursor handle identifying the transaction and
1726 * database for which we are allocating.
1727 * @param[in] num the number of pages to allocate.
1728 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1729 * will always be satisfied by a single contiguous chunk of memory.
1730 * @return 0 on success, non-zero on failure.
1733 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1735 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1736 /* Get at most <Max_retries> more freeDB records once me_pghead
1737 * has enough pages. If not enough, use new pages from the map.
1738 * If <Paranoid> and mc is updating the freeDB, only get new
1739 * records if me_pghead is empty. Then the freelist cannot play
1740 * catch-up with itself by growing while trying to save it.
1742 enum { Paranoid = 1, Max_retries = 500 };
1744 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1746 int rc, retry = Max_retries;
1747 MDB_txn *txn = mc->mc_txn;
1748 MDB_env *env = txn->mt_env;
1749 pgno_t pgno, *mop = env->me_pghead;
1750 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1752 txnid_t oldest = 0, last;
1758 /* If our dirty list is already full, we can't do anything */
1759 if (txn->mt_dirty_room == 0) {
1764 for (op = MDB_FIRST;; op = MDB_NEXT) {
1767 pgno_t *idl, old_id, new_id;
1769 /* Seek a big enough contiguous page range. Prefer
1770 * pages at the tail, just truncating the list.
1776 if (mop[i-n2] == pgno+n2)
1779 if (Max_retries < INT_MAX && --retry < 0)
1783 if (op == MDB_FIRST) { /* 1st iteration */
1784 /* Prepare to fetch more and coalesce */
1785 oldest = mdb_find_oldest(txn);
1786 last = env->me_pglast;
1787 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1790 key.mv_data = &last; /* will look up last+1 */
1791 key.mv_size = sizeof(last);
1793 if (Paranoid && mc->mc_dbi == FREE_DBI)
1796 if (Paranoid && retry < 0 && mop_len)
1800 /* Do not fetch more if the record will be too recent */
1803 rc = mdb_cursor_get(&m2, &key, NULL, op);
1805 if (rc == MDB_NOTFOUND)
1809 last = *(txnid_t*)key.mv_data;
1812 np = m2.mc_pg[m2.mc_top];
1813 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1814 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1817 idl = (MDB_ID *) data.mv_data;
1820 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1825 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1827 mop = env->me_pghead;
1829 env->me_pglast = last;
1831 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1832 last, txn->mt_dbs[FREE_DBI].md_root, i));
1834 DPRINTF(("IDL %"Z"u", idl[k]));
1836 /* Merge in descending sorted order */
1839 mop[0] = (pgno_t)-1;
1843 for (; old_id < new_id; old_id = mop[--j])
1850 /* Use new pages from the map when nothing suitable in the freeDB */
1852 pgno = txn->mt_next_pgno;
1853 if (pgno + num >= env->me_maxpg) {
1854 DPUTS("DB size maxed out");
1860 if (env->me_flags & MDB_WRITEMAP) {
1861 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1863 if (!(np = mdb_page_malloc(txn, num))) {
1869 mop[0] = mop_len -= num;
1870 /* Move any stragglers down */
1871 for (j = i-num; j < mop_len; )
1872 mop[++j] = mop[++i];
1874 txn->mt_next_pgno = pgno + num;
1877 mdb_page_dirty(txn, np);
1883 txn->mt_flags |= MDB_TXN_ERROR;
1887 /** Copy the used portions of a non-overflow page.
1888 * @param[in] dst page to copy into
1889 * @param[in] src page to copy from
1890 * @param[in] psize size of a page
1893 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1895 enum { Align = sizeof(pgno_t) };
1896 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1898 /* If page isn't full, just copy the used portion. Adjust
1899 * alignment so memcpy may copy words instead of bytes.
1901 if ((unused &= -Align) && !IS_LEAF2(src)) {
1903 memcpy(dst, src, (lower + (Align-1)) & -Align);
1904 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1907 memcpy(dst, src, psize - unused);
1911 /** Pull a page off the txn's spill list, if present.
1912 * If a page being referenced was spilled to disk in this txn, bring
1913 * it back and make it dirty/writable again.
1914 * @param[in] txn the transaction handle.
1915 * @param[in] mp the page being referenced. It must not be dirty.
1916 * @param[out] ret the writable page, if any. ret is unchanged if
1917 * mp wasn't spilled.
1920 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1922 MDB_env *env = txn->mt_env;
1925 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1927 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1928 if (!tx2->mt_spill_pgs)
1930 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1931 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1934 if (txn->mt_dirty_room == 0)
1935 return MDB_TXN_FULL;
1936 if (IS_OVERFLOW(mp))
1940 if (env->me_flags & MDB_WRITEMAP) {
1943 np = mdb_page_malloc(txn, num);
1947 memcpy(np, mp, num * env->me_psize);
1949 mdb_page_copy(np, mp, env->me_psize);
1952 /* If in current txn, this page is no longer spilled.
1953 * If it happens to be the last page, truncate the spill list.
1954 * Otherwise mark it as deleted by setting the LSB.
1956 if (x == txn->mt_spill_pgs[0])
1957 txn->mt_spill_pgs[0]--;
1959 txn->mt_spill_pgs[x] |= 1;
1960 } /* otherwise, if belonging to a parent txn, the
1961 * page remains spilled until child commits
1964 mdb_page_dirty(txn, np);
1965 np->mp_flags |= P_DIRTY;
1973 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1974 * @param[in] mc cursor pointing to the page to be touched
1975 * @return 0 on success, non-zero on failure.
1978 mdb_page_touch(MDB_cursor *mc)
1980 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1981 MDB_txn *txn = mc->mc_txn;
1982 MDB_cursor *m2, *m3;
1986 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1987 if (txn->mt_flags & MDB_TXN_SPILLS) {
1989 rc = mdb_page_unspill(txn, mp, &np);
1995 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1996 (rc = mdb_page_alloc(mc, 1, &np)))
1999 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2000 mp->mp_pgno, pgno));
2001 mdb_cassert(mc, mp->mp_pgno != pgno);
2002 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2003 /* Update the parent page, if any, to point to the new page */
2005 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2006 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2007 SETPGNO(node, pgno);
2009 mc->mc_db->md_root = pgno;
2011 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2012 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2014 /* If txn has a parent, make sure the page is in our
2018 unsigned x = mdb_mid2l_search(dl, pgno);
2019 if (x <= dl[0].mid && dl[x].mid == pgno) {
2020 if (mp != dl[x].mptr) { /* bad cursor? */
2021 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2022 txn->mt_flags |= MDB_TXN_ERROR;
2023 return MDB_CORRUPTED;
2028 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2030 np = mdb_page_malloc(txn, 1);
2035 rc = mdb_mid2l_insert(dl, &mid);
2036 mdb_cassert(mc, rc == 0);
2041 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2043 np->mp_flags |= P_DIRTY;
2046 /* Adjust cursors pointing to mp */
2047 mc->mc_pg[mc->mc_top] = np;
2048 m2 = txn->mt_cursors[mc->mc_dbi];
2049 if (mc->mc_flags & C_SUB) {
2050 for (; m2; m2=m2->mc_next) {
2051 m3 = &m2->mc_xcursor->mx_cursor;
2052 if (m3->mc_snum < mc->mc_snum) continue;
2053 if (m3->mc_pg[mc->mc_top] == mp)
2054 m3->mc_pg[mc->mc_top] = np;
2057 for (; m2; m2=m2->mc_next) {
2058 if (m2->mc_snum < mc->mc_snum) continue;
2059 if (m2->mc_pg[mc->mc_top] == mp) {
2060 m2->mc_pg[mc->mc_top] = np;
2061 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2062 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2064 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2065 if (!(leaf->mn_flags & F_SUBDATA))
2066 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2074 txn->mt_flags |= MDB_TXN_ERROR;
2079 mdb_env_sync(MDB_env *env, int force)
2082 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2083 if (env->me_flags & MDB_WRITEMAP) {
2084 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2085 ? MS_ASYNC : MS_SYNC;
2086 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2089 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2093 if (MDB_FDATASYNC(env->me_fd))
2100 /** Back up parent txn's cursors, then grab the originals for tracking */
2102 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2104 MDB_cursor *mc, *bk;
2109 for (i = src->mt_numdbs; --i >= 0; ) {
2110 if ((mc = src->mt_cursors[i]) != NULL) {
2111 size = sizeof(MDB_cursor);
2113 size += sizeof(MDB_xcursor);
2114 for (; mc; mc = bk->mc_next) {
2120 mc->mc_db = &dst->mt_dbs[i];
2121 /* Kill pointers into src - and dst to reduce abuse: The
2122 * user may not use mc until dst ends. Otherwise we'd...
2124 mc->mc_txn = NULL; /* ...set this to dst */
2125 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2126 if ((mx = mc->mc_xcursor) != NULL) {
2127 *(MDB_xcursor *)(bk+1) = *mx;
2128 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2130 mc->mc_next = dst->mt_cursors[i];
2131 dst->mt_cursors[i] = mc;
2138 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2139 * @param[in] txn the transaction handle.
2140 * @param[in] merge true to keep changes to parent cursors, false to revert.
2141 * @return 0 on success, non-zero on failure.
2144 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2146 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2150 for (i = txn->mt_numdbs; --i >= 0; ) {
2151 for (mc = cursors[i]; mc; mc = next) {
2153 if ((bk = mc->mc_backup) != NULL) {
2155 /* Commit changes to parent txn */
2156 mc->mc_next = bk->mc_next;
2157 mc->mc_backup = bk->mc_backup;
2158 mc->mc_txn = bk->mc_txn;
2159 mc->mc_db = bk->mc_db;
2160 mc->mc_dbflag = bk->mc_dbflag;
2161 if ((mx = mc->mc_xcursor) != NULL)
2162 mx->mx_cursor.mc_txn = bk->mc_txn;
2164 /* Abort nested txn */
2166 if ((mx = mc->mc_xcursor) != NULL)
2167 *mx = *(MDB_xcursor *)(bk+1);
2171 /* Only malloced cursors are permanently tracked. */
2179 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2182 mdb_txn_reset0(MDB_txn *txn, const char *act);
2184 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2190 Pidset = F_SETLK, Pidcheck = F_GETLK
2194 /** Set or check a pid lock. Set returns 0 on success.
2195 * Check returns 0 if the process is certainly dead, nonzero if it may
2196 * be alive (the lock exists or an error happened so we do not know).
2198 * On Windows Pidset is a no-op, we merely check for the existence
2199 * of the process with the given pid. On POSIX we use a single byte
2200 * lock on the lockfile, set at an offset equal to the pid.
2203 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2205 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2208 if (op == Pidcheck) {
2209 h = OpenProcess(env->me_pidquery, FALSE, pid);
2210 /* No documented "no such process" code, but other program use this: */
2212 return ErrCode() != ERROR_INVALID_PARAMETER;
2213 /* A process exists until all handles to it close. Has it exited? */
2214 ret = WaitForSingleObject(h, 0) != 0;
2221 struct flock lock_info;
2222 memset(&lock_info, 0, sizeof(lock_info));
2223 lock_info.l_type = F_WRLCK;
2224 lock_info.l_whence = SEEK_SET;
2225 lock_info.l_start = pid;
2226 lock_info.l_len = 1;
2227 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2228 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2230 } else if ((rc = ErrCode()) == EINTR) {
2238 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2239 * @param[in] txn the transaction handle to initialize
2240 * @return 0 on success, non-zero on failure.
2243 mdb_txn_renew0(MDB_txn *txn)
2245 MDB_env *env = txn->mt_env;
2246 MDB_txninfo *ti = env->me_txns;
2250 int rc, new_notls = 0;
2253 txn->mt_numdbs = env->me_numdbs;
2254 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2256 if (txn->mt_flags & MDB_TXN_RDONLY) {
2258 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2259 txn->mt_txnid = meta->mm_txnid;
2260 txn->mt_u.reader = NULL;
2262 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2263 pthread_getspecific(env->me_txkey);
2265 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2266 return MDB_BAD_RSLOT;
2268 MDB_PID_T pid = env->me_pid;
2269 pthread_t tid = pthread_self();
2271 if (!(env->me_flags & MDB_LIVE_READER)) {
2272 rc = mdb_reader_pid(env, Pidset, pid);
2275 env->me_flags |= MDB_LIVE_READER;
2279 nr = ti->mti_numreaders;
2280 for (i=0; i<nr; i++)
2281 if (ti->mti_readers[i].mr_pid == 0)
2283 if (i == env->me_maxreaders) {
2284 UNLOCK_MUTEX_R(env);
2285 return MDB_READERS_FULL;
2287 ti->mti_readers[i].mr_pid = pid;
2288 ti->mti_readers[i].mr_tid = tid;
2290 ti->mti_numreaders = ++nr;
2291 /* Save numreaders for un-mutexed mdb_env_close() */
2292 env->me_numreaders = nr;
2293 UNLOCK_MUTEX_R(env);
2295 r = &ti->mti_readers[i];
2296 new_notls = (env->me_flags & MDB_NOTLS);
2297 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2302 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2303 txn->mt_u.reader = r;
2304 meta = env->me_metas[txn->mt_txnid & 1];
2310 txn->mt_txnid = ti->mti_txnid;
2311 meta = env->me_metas[txn->mt_txnid & 1];
2313 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2314 txn->mt_txnid = meta->mm_txnid;
2318 if (txn->mt_txnid == mdb_debug_start)
2321 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2322 txn->mt_u.dirty_list = env->me_dirty_list;
2323 txn->mt_u.dirty_list[0].mid = 0;
2324 txn->mt_free_pgs = env->me_free_pgs;
2325 txn->mt_free_pgs[0] = 0;
2326 txn->mt_spill_pgs = NULL;
2330 /* Copy the DB info and flags */
2331 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2333 /* Moved to here to avoid a data race in read TXNs */
2334 txn->mt_next_pgno = meta->mm_last_pg+1;
2336 for (i=2; i<txn->mt_numdbs; i++) {
2337 x = env->me_dbflags[i];
2338 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2339 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2341 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2343 if (env->me_maxpg < txn->mt_next_pgno) {
2344 mdb_txn_reset0(txn, "renew0-mapfail");
2346 txn->mt_u.reader->mr_pid = 0;
2347 txn->mt_u.reader = NULL;
2349 return MDB_MAP_RESIZED;
2356 mdb_txn_renew(MDB_txn *txn)
2360 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2363 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2364 DPUTS("environment had fatal error, must shutdown!");
2368 rc = mdb_txn_renew0(txn);
2369 if (rc == MDB_SUCCESS) {
2370 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2371 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2372 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2378 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2382 int rc, size, tsize = sizeof(MDB_txn);
2384 if (env->me_flags & MDB_FATAL_ERROR) {
2385 DPUTS("environment had fatal error, must shutdown!");
2388 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2391 /* Nested transactions: Max 1 child, write txns only, no writemap */
2392 if (parent->mt_child ||
2393 (flags & MDB_RDONLY) ||
2394 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2395 (env->me_flags & MDB_WRITEMAP))
2397 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2399 tsize = sizeof(MDB_ntxn);
2401 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2402 if (!(flags & MDB_RDONLY))
2403 size += env->me_maxdbs * sizeof(MDB_cursor *);
2405 if ((txn = calloc(1, size)) == NULL) {
2406 DPRINTF(("calloc: %s", strerror(ErrCode())));
2409 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2410 if (flags & MDB_RDONLY) {
2411 txn->mt_flags |= MDB_TXN_RDONLY;
2412 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2414 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2415 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2421 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2422 if (!txn->mt_u.dirty_list ||
2423 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2425 free(txn->mt_u.dirty_list);
2429 txn->mt_txnid = parent->mt_txnid;
2430 txn->mt_dirty_room = parent->mt_dirty_room;
2431 txn->mt_u.dirty_list[0].mid = 0;
2432 txn->mt_spill_pgs = NULL;
2433 txn->mt_next_pgno = parent->mt_next_pgno;
2434 parent->mt_child = txn;
2435 txn->mt_parent = parent;
2436 txn->mt_numdbs = parent->mt_numdbs;
2437 txn->mt_flags = parent->mt_flags;
2438 txn->mt_dbxs = parent->mt_dbxs;
2439 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2440 /* Copy parent's mt_dbflags, but clear DB_NEW */
2441 for (i=0; i<txn->mt_numdbs; i++)
2442 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2444 ntxn = (MDB_ntxn *)txn;
2445 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2446 if (env->me_pghead) {
2447 size = MDB_IDL_SIZEOF(env->me_pghead);
2448 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2450 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2455 rc = mdb_cursor_shadow(parent, txn);
2457 mdb_txn_reset0(txn, "beginchild-fail");
2459 rc = mdb_txn_renew0(txn);
2465 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2466 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2467 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2474 mdb_txn_env(MDB_txn *txn)
2476 if(!txn) return NULL;
2480 /** Export or close DBI handles opened in this txn. */
2482 mdb_dbis_update(MDB_txn *txn, int keep)
2485 MDB_dbi n = txn->mt_numdbs;
2486 MDB_env *env = txn->mt_env;
2487 unsigned char *tdbflags = txn->mt_dbflags;
2489 for (i = n; --i >= 2;) {
2490 if (tdbflags[i] & DB_NEW) {
2492 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2494 char *ptr = env->me_dbxs[i].md_name.mv_data;
2495 env->me_dbxs[i].md_name.mv_data = NULL;
2496 env->me_dbxs[i].md_name.mv_size = 0;
2497 env->me_dbflags[i] = 0;
2502 if (keep && env->me_numdbs < n)
2506 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2507 * May be called twice for readonly txns: First reset it, then abort.
2508 * @param[in] txn the transaction handle to reset
2509 * @param[in] act why the transaction is being reset
2512 mdb_txn_reset0(MDB_txn *txn, const char *act)
2514 MDB_env *env = txn->mt_env;
2516 /* Close any DBI handles opened in this txn */
2517 mdb_dbis_update(txn, 0);
2519 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2520 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2521 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2523 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2524 if (txn->mt_u.reader) {
2525 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2526 if (!(env->me_flags & MDB_NOTLS))
2527 txn->mt_u.reader = NULL; /* txn does not own reader */
2529 txn->mt_numdbs = 0; /* close nothing if called again */
2530 txn->mt_dbxs = NULL; /* mark txn as reset */
2532 mdb_cursors_close(txn, 0);
2534 if (!(env->me_flags & MDB_WRITEMAP)) {
2535 mdb_dlist_free(txn);
2537 mdb_midl_free(env->me_pghead);
2539 if (txn->mt_parent) {
2540 txn->mt_parent->mt_child = NULL;
2541 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2542 mdb_midl_free(txn->mt_free_pgs);
2543 mdb_midl_free(txn->mt_spill_pgs);
2544 free(txn->mt_u.dirty_list);
2548 if (mdb_midl_shrink(&txn->mt_free_pgs))
2549 env->me_free_pgs = txn->mt_free_pgs;
2550 env->me_pghead = NULL;
2554 /* The writer mutex was locked in mdb_txn_begin. */
2556 UNLOCK_MUTEX_W(env);
2561 mdb_txn_reset(MDB_txn *txn)
2566 /* This call is only valid for read-only txns */
2567 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2570 mdb_txn_reset0(txn, "reset");
2574 mdb_txn_abort(MDB_txn *txn)
2580 mdb_txn_abort(txn->mt_child);
2582 mdb_txn_reset0(txn, "abort");
2583 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2584 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2585 txn->mt_u.reader->mr_pid = 0;
2590 /** Save the freelist as of this transaction to the freeDB.
2591 * This changes the freelist. Keep trying until it stabilizes.
2594 mdb_freelist_save(MDB_txn *txn)
2596 /* env->me_pghead[] can grow and shrink during this call.
2597 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2598 * Page numbers cannot disappear from txn->mt_free_pgs[].
2601 MDB_env *env = txn->mt_env;
2602 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2603 txnid_t pglast = 0, head_id = 0;
2604 pgno_t freecnt = 0, *free_pgs, *mop;
2605 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2607 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2609 if (env->me_pghead) {
2610 /* Make sure first page of freeDB is touched and on freelist */
2611 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2612 if (rc && rc != MDB_NOTFOUND)
2616 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2617 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2618 ? SSIZE_MAX : maxfree_1pg;
2621 /* Come back here after each Put() in case freelist changed */
2626 /* If using records from freeDB which we have not yet
2627 * deleted, delete them and any we reserved for me_pghead.
2629 while (pglast < env->me_pglast) {
2630 rc = mdb_cursor_first(&mc, &key, NULL);
2633 pglast = head_id = *(txnid_t *)key.mv_data;
2634 total_room = head_room = 0;
2635 mdb_tassert(txn, pglast <= env->me_pglast);
2636 rc = mdb_cursor_del(&mc, 0);
2641 /* Save the IDL of pages freed by this txn, to a single record */
2642 if (freecnt < txn->mt_free_pgs[0]) {
2644 /* Make sure last page of freeDB is touched and on freelist */
2645 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2646 if (rc && rc != MDB_NOTFOUND)
2649 free_pgs = txn->mt_free_pgs;
2650 /* Write to last page of freeDB */
2651 key.mv_size = sizeof(txn->mt_txnid);
2652 key.mv_data = &txn->mt_txnid;
2654 freecnt = free_pgs[0];
2655 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2656 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2659 /* Retry if mt_free_pgs[] grew during the Put() */
2660 free_pgs = txn->mt_free_pgs;
2661 } while (freecnt < free_pgs[0]);
2662 mdb_midl_sort(free_pgs);
2663 memcpy(data.mv_data, free_pgs, data.mv_size);
2666 unsigned int i = free_pgs[0];
2667 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2668 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2670 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2676 mop = env->me_pghead;
2677 mop_len = mop ? mop[0] : 0;
2679 /* Reserve records for me_pghead[]. Split it if multi-page,
2680 * to avoid searching freeDB for a page range. Use keys in
2681 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2683 if (total_room >= mop_len) {
2684 if (total_room == mop_len || --more < 0)
2686 } else if (head_room >= maxfree_1pg && head_id > 1) {
2687 /* Keep current record (overflow page), add a new one */
2691 /* (Re)write {key = head_id, IDL length = head_room} */
2692 total_room -= head_room;
2693 head_room = mop_len - total_room;
2694 if (head_room > maxfree_1pg && head_id > 1) {
2695 /* Overflow multi-page for part of me_pghead */
2696 head_room /= head_id; /* amortize page sizes */
2697 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2698 } else if (head_room < 0) {
2699 /* Rare case, not bothering to delete this record */
2702 key.mv_size = sizeof(head_id);
2703 key.mv_data = &head_id;
2704 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2705 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2708 /* IDL is initially empty, zero out at least the length */
2709 pgs = (pgno_t *)data.mv_data;
2710 j = head_room > clean_limit ? head_room : 0;
2714 total_room += head_room;
2717 /* Fill in the reserved me_pghead records */
2723 rc = mdb_cursor_first(&mc, &key, &data);
2724 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2725 unsigned flags = MDB_CURRENT;
2726 txnid_t id = *(txnid_t *)key.mv_data;
2727 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2730 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2731 if (len > mop_len) {
2733 data.mv_size = (len + 1) * sizeof(MDB_ID);
2734 /* Drop MDB_CURRENT when changing the data size */
2738 data.mv_data = mop -= len;
2741 rc = mdb_cursor_put(&mc, &key, &data, flags);
2743 if (rc || !(mop_len -= len))
2750 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2751 * @param[in] txn the transaction that's being committed
2752 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2753 * @return 0 on success, non-zero on failure.
2756 mdb_page_flush(MDB_txn *txn, int keep)
2758 MDB_env *env = txn->mt_env;
2759 MDB_ID2L dl = txn->mt_u.dirty_list;
2760 unsigned psize = env->me_psize, j;
2761 int i, pagecount = dl[0].mid, rc;
2762 size_t size = 0, pos = 0;
2764 MDB_page *dp = NULL;
2768 struct iovec iov[MDB_COMMIT_PAGES];
2769 ssize_t wpos = 0, wsize = 0, wres;
2770 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2776 if (env->me_flags & MDB_WRITEMAP) {
2777 /* Clear dirty flags */
2778 while (++i <= pagecount) {
2780 /* Don't flush this page yet */
2781 if (dp->mp_flags & P_KEEP) {
2782 dp->mp_flags ^= P_KEEP;
2786 dp->mp_flags &= ~P_DIRTY;
2791 /* Write the pages */
2793 if (++i <= pagecount) {
2795 /* Don't flush this page yet */
2796 if (dp->mp_flags & P_KEEP) {
2797 dp->mp_flags ^= P_KEEP;
2802 /* clear dirty flag */
2803 dp->mp_flags &= ~P_DIRTY;
2806 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2811 /* Windows actually supports scatter/gather I/O, but only on
2812 * unbuffered file handles. Since we're relying on the OS page
2813 * cache for all our data, that's self-defeating. So we just
2814 * write pages one at a time. We use the ov structure to set
2815 * the write offset, to at least save the overhead of a Seek
2818 DPRINTF(("committing page %"Z"u", pgno));
2819 memset(&ov, 0, sizeof(ov));
2820 ov.Offset = pos & 0xffffffff;
2821 ov.OffsetHigh = pos >> 16 >> 16;
2822 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2824 DPRINTF(("WriteFile: %d", rc));
2828 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2829 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2831 /* Write previous page(s) */
2832 #ifdef MDB_USE_PWRITEV
2833 wres = pwritev(env->me_fd, iov, n, wpos);
2836 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2838 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2840 DPRINTF(("lseek: %s", strerror(rc)));
2843 wres = writev(env->me_fd, iov, n);
2846 if (wres != wsize) {
2849 DPRINTF(("Write error: %s", strerror(rc)));
2851 rc = EIO; /* TODO: Use which error code? */
2852 DPUTS("short write, filesystem full?");
2863 DPRINTF(("committing page %"Z"u", pgno));
2864 next_pos = pos + size;
2865 iov[n].iov_len = size;
2866 iov[n].iov_base = (char *)dp;
2872 for (i = keep; ++i <= pagecount; ) {
2874 /* This is a page we skipped above */
2877 dl[j].mid = dp->mp_pgno;
2880 mdb_dpage_free(env, dp);
2885 txn->mt_dirty_room += i - j;
2891 mdb_txn_commit(MDB_txn *txn)
2897 if (txn == NULL || txn->mt_env == NULL)
2900 if (txn->mt_child) {
2901 rc = mdb_txn_commit(txn->mt_child);
2902 txn->mt_child = NULL;
2909 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2910 mdb_dbis_update(txn, 1);
2911 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2916 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2917 DPUTS("error flag is set, can't commit");
2919 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2924 if (txn->mt_parent) {
2925 MDB_txn *parent = txn->mt_parent;
2928 unsigned x, y, len, ps_len;
2930 /* Append our free list to parent's */
2931 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2934 mdb_midl_free(txn->mt_free_pgs);
2935 /* Failures after this must either undo the changes
2936 * to the parent or set MDB_TXN_ERROR in the parent.
2939 parent->mt_next_pgno = txn->mt_next_pgno;
2940 parent->mt_flags = txn->mt_flags;
2942 /* Merge our cursors into parent's and close them */
2943 mdb_cursors_close(txn, 1);
2945 /* Update parent's DB table. */
2946 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2947 parent->mt_numdbs = txn->mt_numdbs;
2948 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2949 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2950 for (i=2; i<txn->mt_numdbs; i++) {
2951 /* preserve parent's DB_NEW status */
2952 x = parent->mt_dbflags[i] & DB_NEW;
2953 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2956 dst = parent->mt_u.dirty_list;
2957 src = txn->mt_u.dirty_list;
2958 /* Remove anything in our dirty list from parent's spill list */
2959 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2961 pspill[0] = (pgno_t)-1;
2962 /* Mark our dirty pages as deleted in parent spill list */
2963 for (i=0, len=src[0].mid; ++i <= len; ) {
2964 MDB_ID pn = src[i].mid << 1;
2965 while (pn > pspill[x])
2967 if (pn == pspill[x]) {
2972 /* Squash deleted pagenums if we deleted any */
2973 for (x=y; ++x <= ps_len; )
2974 if (!(pspill[x] & 1))
2975 pspill[++y] = pspill[x];
2979 /* Find len = length of merging our dirty list with parent's */
2981 dst[0].mid = 0; /* simplify loops */
2982 if (parent->mt_parent) {
2983 len = x + src[0].mid;
2984 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2985 for (i = x; y && i; y--) {
2986 pgno_t yp = src[y].mid;
2987 while (yp < dst[i].mid)
2989 if (yp == dst[i].mid) {
2994 } else { /* Simplify the above for single-ancestor case */
2995 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2997 /* Merge our dirty list with parent's */
2999 for (i = len; y; dst[i--] = src[y--]) {
3000 pgno_t yp = src[y].mid;
3001 while (yp < dst[x].mid)
3002 dst[i--] = dst[x--];
3003 if (yp == dst[x].mid)
3004 free(dst[x--].mptr);
3006 mdb_tassert(txn, i == x);
3008 free(txn->mt_u.dirty_list);
3009 parent->mt_dirty_room = txn->mt_dirty_room;
3010 if (txn->mt_spill_pgs) {
3011 if (parent->mt_spill_pgs) {
3012 /* TODO: Prevent failure here, so parent does not fail */
3013 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3015 parent->mt_flags |= MDB_TXN_ERROR;
3016 mdb_midl_free(txn->mt_spill_pgs);
3017 mdb_midl_sort(parent->mt_spill_pgs);
3019 parent->mt_spill_pgs = txn->mt_spill_pgs;
3023 parent->mt_child = NULL;
3024 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3029 if (txn != env->me_txn) {
3030 DPUTS("attempt to commit unknown transaction");
3035 mdb_cursors_close(txn, 0);
3037 if (!txn->mt_u.dirty_list[0].mid &&
3038 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3041 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3042 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3044 /* Update DB root pointers */
3045 if (txn->mt_numdbs > 2) {
3049 data.mv_size = sizeof(MDB_db);
3051 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3052 for (i = 2; i < txn->mt_numdbs; i++) {
3053 if (txn->mt_dbflags[i] & DB_DIRTY) {
3054 data.mv_data = &txn->mt_dbs[i];
3055 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3062 rc = mdb_freelist_save(txn);
3066 mdb_midl_free(env->me_pghead);
3067 env->me_pghead = NULL;
3068 if (mdb_midl_shrink(&txn->mt_free_pgs))
3069 env->me_free_pgs = txn->mt_free_pgs;
3075 if ((rc = mdb_page_flush(txn, 0)) ||
3076 (rc = mdb_env_sync(env, 0)) ||
3077 (rc = mdb_env_write_meta(txn)))
3083 mdb_dbis_update(txn, 1);
3086 UNLOCK_MUTEX_W(env);
3096 /** Read the environment parameters of a DB environment before
3097 * mapping it into memory.
3098 * @param[in] env the environment handle
3099 * @param[out] meta address of where to store the meta information
3100 * @return 0 on success, non-zero on failure.
3103 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3109 enum { Size = sizeof(pbuf) };
3111 /* We don't know the page size yet, so use a minimum value.
3112 * Read both meta pages so we can use the latest one.
3115 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3119 memset(&ov, 0, sizeof(ov));
3121 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3122 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3125 rc = pread(env->me_fd, &pbuf, Size, off);
3128 if (rc == 0 && off == 0)
3130 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3131 DPRINTF(("read: %s", mdb_strerror(rc)));
3135 p = (MDB_page *)&pbuf;
3137 if (!F_ISSET(p->mp_flags, P_META)) {
3138 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3143 if (m->mm_magic != MDB_MAGIC) {
3144 DPUTS("meta has invalid magic");
3148 if (m->mm_version != MDB_DATA_VERSION) {
3149 DPRINTF(("database is version %u, expected version %u",
3150 m->mm_version, MDB_DATA_VERSION));
3151 return MDB_VERSION_MISMATCH;
3154 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3160 /** Write the environment parameters of a freshly created DB environment.
3161 * @param[in] env the environment handle
3162 * @param[out] meta address of where to store the meta information
3163 * @return 0 on success, non-zero on failure.
3166 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3174 memset(&ov, 0, sizeof(ov));
3175 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3177 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3180 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3181 len = pwrite(fd, ptr, size, pos); \
3182 rc = (len >= 0); } while(0)
3185 DPUTS("writing new meta page");
3187 psize = env->me_psize;
3189 meta->mm_magic = MDB_MAGIC;
3190 meta->mm_version = MDB_DATA_VERSION;
3191 meta->mm_mapsize = env->me_mapsize;
3192 meta->mm_psize = psize;
3193 meta->mm_last_pg = 1;
3194 meta->mm_flags = env->me_flags & 0xffff;
3195 meta->mm_flags |= MDB_INTEGERKEY;
3196 meta->mm_dbs[0].md_root = P_INVALID;
3197 meta->mm_dbs[1].md_root = P_INVALID;
3199 p = calloc(2, psize);
3201 p->mp_flags = P_META;
3202 *(MDB_meta *)METADATA(p) = *meta;
3204 q = (MDB_page *)((char *)p + psize);
3206 q->mp_flags = P_META;
3207 *(MDB_meta *)METADATA(q) = *meta;
3209 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3212 else if ((unsigned) len == psize * 2)
3220 /** Update the environment info to commit a transaction.
3221 * @param[in] txn the transaction that's being committed
3222 * @return 0 on success, non-zero on failure.
3225 mdb_env_write_meta(MDB_txn *txn)
3228 MDB_meta meta, metab, *mp;
3230 int rc, len, toggle;
3239 toggle = txn->mt_txnid & 1;
3240 DPRINTF(("writing meta page %d for root page %"Z"u",
3241 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3244 mp = env->me_metas[toggle];
3246 if (env->me_flags & MDB_WRITEMAP) {
3247 /* Persist any increases of mapsize config */
3248 if (env->me_mapsize > mp->mm_mapsize)
3249 mp->mm_mapsize = env->me_mapsize;
3250 mp->mm_dbs[0] = txn->mt_dbs[0];
3251 mp->mm_dbs[1] = txn->mt_dbs[1];
3252 mp->mm_last_pg = txn->mt_next_pgno - 1;
3253 mp->mm_txnid = txn->mt_txnid;
3254 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3255 unsigned meta_size = env->me_psize;
3256 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3259 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3260 if (meta_size < env->me_os_psize)
3261 meta_size += meta_size;
3266 if (MDB_MSYNC(ptr, meta_size, rc)) {
3273 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3274 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3276 ptr = (char *)&meta;
3277 if (env->me_mapsize > mp->mm_mapsize) {
3278 /* Persist any increases of mapsize config */
3279 meta.mm_mapsize = env->me_mapsize;
3280 off = offsetof(MDB_meta, mm_mapsize);
3282 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3284 len = sizeof(MDB_meta) - off;
3287 meta.mm_dbs[0] = txn->mt_dbs[0];
3288 meta.mm_dbs[1] = txn->mt_dbs[1];
3289 meta.mm_last_pg = txn->mt_next_pgno - 1;
3290 meta.mm_txnid = txn->mt_txnid;
3293 off += env->me_psize;
3296 /* Write to the SYNC fd */
3297 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3298 env->me_fd : env->me_mfd;
3301 memset(&ov, 0, sizeof(ov));
3303 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3307 rc = pwrite(mfd, ptr, len, off);
3310 rc = rc < 0 ? ErrCode() : EIO;
3311 DPUTS("write failed, disk error?");
3312 /* On a failure, the pagecache still contains the new data.
3313 * Write some old data back, to prevent it from being used.
3314 * Use the non-SYNC fd; we know it will fail anyway.
3316 meta.mm_last_pg = metab.mm_last_pg;
3317 meta.mm_txnid = metab.mm_txnid;
3319 memset(&ov, 0, sizeof(ov));
3321 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3323 r2 = pwrite(env->me_fd, ptr, len, off);
3324 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3327 env->me_flags |= MDB_FATAL_ERROR;
3331 /* Memory ordering issues are irrelevant; since the entire writer
3332 * is wrapped by wmutex, all of these changes will become visible
3333 * after the wmutex is unlocked. Since the DB is multi-version,
3334 * readers will get consistent data regardless of how fresh or
3335 * how stale their view of these values is.
3338 env->me_txns->mti_txnid = txn->mt_txnid;
3343 /** Check both meta pages to see which one is newer.
3344 * @param[in] env the environment handle
3345 * @return meta toggle (0 or 1).
3348 mdb_env_pick_meta(const MDB_env *env)
3350 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3354 mdb_env_create(MDB_env **env)
3358 e = calloc(1, sizeof(MDB_env));
3362 e->me_maxreaders = DEFAULT_READERS;
3363 e->me_maxdbs = e->me_numdbs = 2;
3364 e->me_fd = INVALID_HANDLE_VALUE;
3365 e->me_lfd = INVALID_HANDLE_VALUE;
3366 e->me_mfd = INVALID_HANDLE_VALUE;
3367 #ifdef MDB_USE_POSIX_SEM
3368 e->me_rmutex = SEM_FAILED;
3369 e->me_wmutex = SEM_FAILED;
3371 e->me_pid = getpid();
3372 GET_PAGESIZE(e->me_os_psize);
3373 VGMEMP_CREATE(e,0,0);
3379 mdb_env_map(MDB_env *env, void *addr, int newsize)
3382 unsigned int flags = env->me_flags;
3386 LONG sizelo, sizehi;
3387 sizelo = env->me_mapsize & 0xffffffff;
3388 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3390 /* Windows won't create mappings for zero length files.
3391 * Just allocate the maxsize right now.
3394 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3395 || !SetEndOfFile(env->me_fd)
3396 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3399 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3400 PAGE_READWRITE : PAGE_READONLY,
3401 sizehi, sizelo, NULL);
3404 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3405 FILE_MAP_WRITE : FILE_MAP_READ,
3406 0, 0, env->me_mapsize, addr);
3407 rc = env->me_map ? 0 : ErrCode();
3412 int prot = PROT_READ;
3413 if (flags & MDB_WRITEMAP) {
3415 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3418 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3420 if (env->me_map == MAP_FAILED) {
3425 if (flags & MDB_NORDAHEAD) {
3426 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3428 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3430 #ifdef POSIX_MADV_RANDOM
3431 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3432 #endif /* POSIX_MADV_RANDOM */
3433 #endif /* MADV_RANDOM */
3437 /* Can happen because the address argument to mmap() is just a
3438 * hint. mmap() can pick another, e.g. if the range is in use.
3439 * The MAP_FIXED flag would prevent that, but then mmap could
3440 * instead unmap existing pages to make room for the new map.
3442 if (addr && env->me_map != addr)
3443 return EBUSY; /* TODO: Make a new MDB_* error code? */
3445 p = (MDB_page *)env->me_map;
3446 env->me_metas[0] = METADATA(p);
3447 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3453 mdb_env_set_mapsize(MDB_env *env, size_t size)
3455 /* If env is already open, caller is responsible for making
3456 * sure there are no active txns.
3464 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3465 else if (size < env->me_mapsize) {
3466 /* If the configured size is smaller, make sure it's
3467 * still big enough. Silently round up to minimum if not.
3469 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3473 munmap(env->me_map, env->me_mapsize);
3474 env->me_mapsize = size;
3475 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3476 rc = mdb_env_map(env, old, 1);
3480 env->me_mapsize = size;
3482 env->me_maxpg = env->me_mapsize / env->me_psize;
3487 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3491 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3496 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3498 if (env->me_map || readers < 1)
3500 env->me_maxreaders = readers;
3505 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3507 if (!env || !readers)
3509 *readers = env->me_maxreaders;
3513 /** Further setup required for opening an MDB environment
3516 mdb_env_open2(MDB_env *env)
3518 unsigned int flags = env->me_flags;
3519 int i, newenv = 0, rc;
3523 /* See if we should use QueryLimited */
3525 if ((rc & 0xff) > 5)
3526 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3528 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3531 memset(&meta, 0, sizeof(meta));
3533 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3536 DPUTS("new mdbenv");
3538 env->me_psize = env->me_os_psize;
3539 if (env->me_psize > MAX_PAGESIZE)
3540 env->me_psize = MAX_PAGESIZE;
3542 env->me_psize = meta.mm_psize;
3545 /* Was a mapsize configured? */
3546 if (!env->me_mapsize) {
3547 /* If this is a new environment, take the default,
3548 * else use the size recorded in the existing env.
3550 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3551 } else if (env->me_mapsize < meta.mm_mapsize) {
3552 /* If the configured size is smaller, make sure it's
3553 * still big enough. Silently round up to minimum if not.
3555 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3556 if (env->me_mapsize < minsize)
3557 env->me_mapsize = minsize;
3560 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3565 if (flags & MDB_FIXEDMAP)
3566 meta.mm_address = env->me_map;
3567 i = mdb_env_init_meta(env, &meta);
3568 if (i != MDB_SUCCESS) {
3573 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3574 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3576 #if !(MDB_MAXKEYSIZE)
3577 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3579 env->me_maxpg = env->me_mapsize / env->me_psize;
3583 int toggle = mdb_env_pick_meta(env);
3584 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3586 DPRINTF(("opened database version %u, pagesize %u",
3587 env->me_metas[0]->mm_version, env->me_psize));
3588 DPRINTF(("using meta page %d", toggle));
3589 DPRINTF(("depth: %u", db->md_depth));
3590 DPRINTF(("entries: %"Z"u", db->md_entries));
3591 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3592 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3593 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3594 DPRINTF(("root: %"Z"u", db->md_root));
3602 /** Release a reader thread's slot in the reader lock table.
3603 * This function is called automatically when a thread exits.
3604 * @param[in] ptr This points to the slot in the reader lock table.
3607 mdb_env_reader_dest(void *ptr)
3609 MDB_reader *reader = ptr;
3615 /** Junk for arranging thread-specific callbacks on Windows. This is
3616 * necessarily platform and compiler-specific. Windows supports up
3617 * to 1088 keys. Let's assume nobody opens more than 64 environments
3618 * in a single process, for now. They can override this if needed.
3620 #ifndef MAX_TLS_KEYS
3621 #define MAX_TLS_KEYS 64
3623 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3624 static int mdb_tls_nkeys;
3626 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3630 case DLL_PROCESS_ATTACH: break;
3631 case DLL_THREAD_ATTACH: break;
3632 case DLL_THREAD_DETACH:
3633 for (i=0; i<mdb_tls_nkeys; i++) {
3634 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3635 mdb_env_reader_dest(r);
3638 case DLL_PROCESS_DETACH: break;
3643 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3645 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3649 /* Force some symbol references.
3650 * _tls_used forces the linker to create the TLS directory if not already done
3651 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3653 #pragma comment(linker, "/INCLUDE:_tls_used")
3654 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3655 #pragma const_seg(".CRT$XLB")
3656 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3657 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3660 #pragma comment(linker, "/INCLUDE:__tls_used")
3661 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3662 #pragma data_seg(".CRT$XLB")
3663 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3665 #endif /* WIN 32/64 */
3666 #endif /* !__GNUC__ */
3669 /** Downgrade the exclusive lock on the region back to shared */
3671 mdb_env_share_locks(MDB_env *env, int *excl)
3673 int rc = 0, toggle = mdb_env_pick_meta(env);
3675 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3680 /* First acquire a shared lock. The Unlock will
3681 * then release the existing exclusive lock.
3683 memset(&ov, 0, sizeof(ov));
3684 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3687 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3693 struct flock lock_info;
3694 /* The shared lock replaces the existing lock */
3695 memset((void *)&lock_info, 0, sizeof(lock_info));
3696 lock_info.l_type = F_RDLCK;
3697 lock_info.l_whence = SEEK_SET;
3698 lock_info.l_start = 0;
3699 lock_info.l_len = 1;
3700 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3701 (rc = ErrCode()) == EINTR) ;
3702 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3709 /** Try to get exlusive lock, otherwise shared.
3710 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3713 mdb_env_excl_lock(MDB_env *env, int *excl)
3717 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3721 memset(&ov, 0, sizeof(ov));
3722 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3729 struct flock lock_info;
3730 memset((void *)&lock_info, 0, sizeof(lock_info));
3731 lock_info.l_type = F_WRLCK;
3732 lock_info.l_whence = SEEK_SET;
3733 lock_info.l_start = 0;
3734 lock_info.l_len = 1;
3735 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3736 (rc = ErrCode()) == EINTR) ;
3740 # ifdef MDB_USE_POSIX_SEM
3741 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3744 lock_info.l_type = F_RDLCK;
3745 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3746 (rc = ErrCode()) == EINTR) ;
3756 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3758 * @(#) $Revision: 5.1 $
3759 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3760 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3762 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3766 * Please do not copyright this code. This code is in the public domain.
3768 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3769 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3770 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3771 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3772 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3773 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3774 * PERFORMANCE OF THIS SOFTWARE.
3777 * chongo <Landon Curt Noll> /\oo/\
3778 * http://www.isthe.com/chongo/
3780 * Share and Enjoy! :-)
3783 typedef unsigned long long mdb_hash_t;
3784 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3786 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3787 * @param[in] val value to hash
3788 * @param[in] hval initial value for hash
3789 * @return 64 bit hash
3791 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3792 * hval arg on the first call.
3795 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3797 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3798 unsigned char *end = s + val->mv_size;
3800 * FNV-1a hash each octet of the string
3803 /* xor the bottom with the current octet */
3804 hval ^= (mdb_hash_t)*s++;
3806 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3807 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3808 (hval << 7) + (hval << 8) + (hval << 40);
3810 /* return our new hash value */
3814 /** Hash the string and output the encoded hash.
3815 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3816 * very short name limits. We don't care about the encoding being reversible,
3817 * we just want to preserve as many bits of the input as possible in a
3818 * small printable string.
3819 * @param[in] str string to hash
3820 * @param[out] encbuf an array of 11 chars to hold the hash
3822 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3825 mdb_pack85(unsigned long l, char *out)
3829 for (i=0; i<5; i++) {
3830 *out++ = mdb_a85[l % 85];
3836 mdb_hash_enc(MDB_val *val, char *encbuf)
3838 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3840 mdb_pack85(h, encbuf);
3841 mdb_pack85(h>>32, encbuf+5);
3846 /** Open and/or initialize the lock region for the environment.
3847 * @param[in] env The MDB environment.
3848 * @param[in] lpath The pathname of the file used for the lock region.
3849 * @param[in] mode The Unix permissions for the file, if we create it.
3850 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3851 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3852 * @return 0 on success, non-zero on failure.
3855 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3858 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3860 # define MDB_ERRCODE_ROFS EROFS
3861 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3862 # define MDB_CLOEXEC O_CLOEXEC
3865 # define MDB_CLOEXEC 0
3872 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3873 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3874 FILE_ATTRIBUTE_NORMAL, NULL);
3876 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3878 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3880 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3885 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3886 /* Lose record locks when exec*() */
3887 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3888 fcntl(env->me_lfd, F_SETFD, fdflags);
3891 if (!(env->me_flags & MDB_NOTLS)) {
3892 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3895 env->me_flags |= MDB_ENV_TXKEY;
3897 /* Windows TLS callbacks need help finding their TLS info. */
3898 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3902 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3906 /* Try to get exclusive lock. If we succeed, then
3907 * nobody is using the lock region and we should initialize it.
3909 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3912 size = GetFileSize(env->me_lfd, NULL);
3914 size = lseek(env->me_lfd, 0, SEEK_END);
3915 if (size == -1) goto fail_errno;
3917 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3918 if (size < rsize && *excl > 0) {
3920 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3921 || !SetEndOfFile(env->me_lfd))
3924 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3928 size = rsize - sizeof(MDB_txninfo);
3929 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3934 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3936 if (!mh) goto fail_errno;
3937 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3939 if (!env->me_txns) goto fail_errno;
3941 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3943 if (m == MAP_FAILED) goto fail_errno;
3949 BY_HANDLE_FILE_INFORMATION stbuf;
3958 if (!mdb_sec_inited) {
3959 InitializeSecurityDescriptor(&mdb_null_sd,
3960 SECURITY_DESCRIPTOR_REVISION);
3961 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3962 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3963 mdb_all_sa.bInheritHandle = FALSE;
3964 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3967 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3968 idbuf.volume = stbuf.dwVolumeSerialNumber;
3969 idbuf.nhigh = stbuf.nFileIndexHigh;
3970 idbuf.nlow = stbuf.nFileIndexLow;
3971 val.mv_data = &idbuf;
3972 val.mv_size = sizeof(idbuf);
3973 mdb_hash_enc(&val, encbuf);
3974 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3975 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3976 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3977 if (!env->me_rmutex) goto fail_errno;
3978 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3979 if (!env->me_wmutex) goto fail_errno;
3980 #elif defined(MDB_USE_POSIX_SEM)
3989 #if defined(__NetBSD__)
3990 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3992 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3993 idbuf.dev = stbuf.st_dev;
3994 idbuf.ino = stbuf.st_ino;
3995 val.mv_data = &idbuf;
3996 val.mv_size = sizeof(idbuf);
3997 mdb_hash_enc(&val, encbuf);
3998 #ifdef MDB_SHORT_SEMNAMES
3999 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4001 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4002 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4003 /* Clean up after a previous run, if needed: Try to
4004 * remove both semaphores before doing anything else.
4006 sem_unlink(env->me_txns->mti_rmname);
4007 sem_unlink(env->me_txns->mti_wmname);
4008 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4009 O_CREAT|O_EXCL, mode, 1);
4010 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4011 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4012 O_CREAT|O_EXCL, mode, 1);
4013 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4014 #else /* MDB_USE_POSIX_SEM */
4015 pthread_mutexattr_t mattr;
4017 if ((rc = pthread_mutexattr_init(&mattr))
4018 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4019 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4020 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4022 pthread_mutexattr_destroy(&mattr);
4023 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4025 env->me_txns->mti_magic = MDB_MAGIC;
4026 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4027 env->me_txns->mti_txnid = 0;
4028 env->me_txns->mti_numreaders = 0;
4031 if (env->me_txns->mti_magic != MDB_MAGIC) {
4032 DPUTS("lock region has invalid magic");
4036 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4037 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4038 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4039 rc = MDB_VERSION_MISMATCH;
4043 if (rc && rc != EACCES && rc != EAGAIN) {
4047 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4048 if (!env->me_rmutex) goto fail_errno;
4049 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4050 if (!env->me_wmutex) goto fail_errno;
4051 #elif defined(MDB_USE_POSIX_SEM)
4052 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4053 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4054 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4055 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4066 /** The name of the lock file in the DB environment */
4067 #define LOCKNAME "/lock.mdb"
4068 /** The name of the data file in the DB environment */
4069 #define DATANAME "/data.mdb"
4070 /** The suffix of the lock file when no subdir is used */
4071 #define LOCKSUFF "-lock"
4072 /** Only a subset of the @ref mdb_env flags can be changed
4073 * at runtime. Changing other flags requires closing the
4074 * environment and re-opening it with the new flags.
4076 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4077 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4078 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4080 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4081 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4085 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4087 int oflags, rc, len, excl = -1;
4088 char *lpath, *dpath;
4090 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4094 if (flags & MDB_NOSUBDIR) {
4095 rc = len + sizeof(LOCKSUFF) + len + 1;
4097 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4102 if (flags & MDB_NOSUBDIR) {
4103 dpath = lpath + len + sizeof(LOCKSUFF);
4104 sprintf(lpath, "%s" LOCKSUFF, path);
4105 strcpy(dpath, path);
4107 dpath = lpath + len + sizeof(LOCKNAME);
4108 sprintf(lpath, "%s" LOCKNAME, path);
4109 sprintf(dpath, "%s" DATANAME, path);
4113 flags |= env->me_flags;
4114 if (flags & MDB_RDONLY) {
4115 /* silently ignore WRITEMAP when we're only getting read access */
4116 flags &= ~MDB_WRITEMAP;
4118 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4119 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4122 env->me_flags = flags |= MDB_ENV_ACTIVE;
4126 env->me_path = strdup(path);
4127 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4128 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4129 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4134 /* For RDONLY, get lockfile after we know datafile exists */
4135 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4136 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4142 if (F_ISSET(flags, MDB_RDONLY)) {
4143 oflags = GENERIC_READ;
4144 len = OPEN_EXISTING;
4146 oflags = GENERIC_READ|GENERIC_WRITE;
4149 mode = FILE_ATTRIBUTE_NORMAL;
4150 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4151 NULL, len, mode, NULL);
4153 if (F_ISSET(flags, MDB_RDONLY))
4156 oflags = O_RDWR | O_CREAT;
4158 env->me_fd = open(dpath, oflags, mode);
4160 if (env->me_fd == INVALID_HANDLE_VALUE) {
4165 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4166 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4171 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4172 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4173 env->me_mfd = env->me_fd;
4175 /* Synchronous fd for meta writes. Needed even with
4176 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4179 len = OPEN_EXISTING;
4180 env->me_mfd = CreateFile(dpath, oflags,
4181 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4182 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4185 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4187 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4192 DPRINTF(("opened dbenv %p", (void *) env));
4194 rc = mdb_env_share_locks(env, &excl);
4198 if (!((flags & MDB_RDONLY) ||
4199 (env->me_pbuf = calloc(1, env->me_psize))))
4205 mdb_env_close0(env, excl);
4211 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4213 mdb_env_close0(MDB_env *env, int excl)
4217 if (!(env->me_flags & MDB_ENV_ACTIVE))
4220 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4221 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4222 free(env->me_dbxs[i].md_name.mv_data);
4225 free(env->me_dbflags);
4228 free(env->me_dirty_list);
4229 mdb_midl_free(env->me_free_pgs);
4231 if (env->me_flags & MDB_ENV_TXKEY) {
4232 pthread_key_delete(env->me_txkey);
4234 /* Delete our key from the global list */
4235 for (i=0; i<mdb_tls_nkeys; i++)
4236 if (mdb_tls_keys[i] == env->me_txkey) {
4237 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4245 munmap(env->me_map, env->me_mapsize);
4247 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4248 (void) close(env->me_mfd);
4249 if (env->me_fd != INVALID_HANDLE_VALUE)
4250 (void) close(env->me_fd);
4252 MDB_PID_T pid = env->me_pid;
4253 /* Clearing readers is done in this function because
4254 * me_txkey with its destructor must be disabled first.
4256 for (i = env->me_numreaders; --i >= 0; )
4257 if (env->me_txns->mti_readers[i].mr_pid == pid)
4258 env->me_txns->mti_readers[i].mr_pid = 0;
4260 if (env->me_rmutex) {
4261 CloseHandle(env->me_rmutex);
4262 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4264 /* Windows automatically destroys the mutexes when
4265 * the last handle closes.
4267 #elif defined(MDB_USE_POSIX_SEM)
4268 if (env->me_rmutex != SEM_FAILED) {
4269 sem_close(env->me_rmutex);
4270 if (env->me_wmutex != SEM_FAILED)
4271 sem_close(env->me_wmutex);
4272 /* If we have the filelock: If we are the
4273 * only remaining user, clean up semaphores.
4276 mdb_env_excl_lock(env, &excl);
4278 sem_unlink(env->me_txns->mti_rmname);
4279 sem_unlink(env->me_txns->mti_wmname);
4283 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4285 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4288 /* Unlock the lockfile. Windows would have unlocked it
4289 * after closing anyway, but not necessarily at once.
4291 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4294 (void) close(env->me_lfd);
4297 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4301 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4303 MDB_txn *txn = NULL;
4309 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4313 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4316 /* Do the lock/unlock of the reader mutex before starting the
4317 * write txn. Otherwise other read txns could block writers.
4319 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4324 /* We must start the actual read txn after blocking writers */
4325 mdb_txn_reset0(txn, "reset-stage1");
4327 /* Temporarily block writers until we snapshot the meta pages */
4330 rc = mdb_txn_renew0(txn);
4332 UNLOCK_MUTEX_W(env);
4337 wsize = env->me_psize * 2;
4341 DO_WRITE(rc, fd, ptr, w2, len);
4345 } else if (len > 0) {
4351 /* Non-blocking or async handles are not supported */
4357 UNLOCK_MUTEX_W(env);
4362 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4364 if (wsize > MAX_WRITE)
4368 DO_WRITE(rc, fd, ptr, w2, len);
4372 } else if (len > 0) {
4389 mdb_env_copy(MDB_env *env, const char *path)
4393 HANDLE newfd = INVALID_HANDLE_VALUE;
4395 if (env->me_flags & MDB_NOSUBDIR) {
4396 lpath = (char *)path;
4399 len += sizeof(DATANAME);
4400 lpath = malloc(len);
4403 sprintf(lpath, "%s" DATANAME, path);
4406 /* The destination path must exist, but the destination file must not.
4407 * We don't want the OS to cache the writes, since the source data is
4408 * already in the OS cache.
4411 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4412 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4414 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4416 if (newfd == INVALID_HANDLE_VALUE) {
4422 /* Set O_DIRECT if the file system supports it */
4423 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4424 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4426 #ifdef F_NOCACHE /* __APPLE__ */
4427 rc = fcntl(newfd, F_NOCACHE, 1);
4434 rc = mdb_env_copyfd(env, newfd);
4437 if (!(env->me_flags & MDB_NOSUBDIR))
4439 if (newfd != INVALID_HANDLE_VALUE)
4440 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4447 mdb_env_close(MDB_env *env)
4454 VGMEMP_DESTROY(env);
4455 while ((dp = env->me_dpages) != NULL) {
4456 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4457 env->me_dpages = dp->mp_next;
4461 mdb_env_close0(env, 0);
4465 /** Compare two items pointing at aligned size_t's */
4467 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4469 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4470 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4473 /** Compare two items pointing at aligned unsigned int's */
4475 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4477 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4478 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4481 /** Compare two items pointing at unsigned ints of unknown alignment.
4482 * Nodes and keys are guaranteed to be 2-byte aligned.
4485 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4487 #if BYTE_ORDER == LITTLE_ENDIAN
4488 unsigned short *u, *c;
4491 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4492 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4495 } while(!x && u > (unsigned short *)a->mv_data);
4498 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4502 /** Compare two items lexically */
4504 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4511 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4517 diff = memcmp(a->mv_data, b->mv_data, len);
4518 return diff ? diff : len_diff<0 ? -1 : len_diff;
4521 /** Compare two items in reverse byte order */
4523 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4525 const unsigned char *p1, *p2, *p1_lim;
4529 p1_lim = (const unsigned char *)a->mv_data;
4530 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4531 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4533 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4539 while (p1 > p1_lim) {
4540 diff = *--p1 - *--p2;
4544 return len_diff<0 ? -1 : len_diff;
4547 /** Search for key within a page, using binary search.
4548 * Returns the smallest entry larger or equal to the key.
4549 * If exactp is non-null, stores whether the found entry was an exact match
4550 * in *exactp (1 or 0).
4551 * Updates the cursor index with the index of the found entry.
4552 * If no entry larger or equal to the key is found, returns NULL.
4555 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4557 unsigned int i = 0, nkeys;
4560 MDB_page *mp = mc->mc_pg[mc->mc_top];
4561 MDB_node *node = NULL;
4566 nkeys = NUMKEYS(mp);
4568 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4569 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4572 low = IS_LEAF(mp) ? 0 : 1;
4574 cmp = mc->mc_dbx->md_cmp;
4576 /* Branch pages have no data, so if using integer keys,
4577 * alignment is guaranteed. Use faster mdb_cmp_int.
4579 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4580 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4587 nodekey.mv_size = mc->mc_db->md_pad;
4588 node = NODEPTR(mp, 0); /* fake */
4589 while (low <= high) {
4590 i = (low + high) >> 1;
4591 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4592 rc = cmp(key, &nodekey);
4593 DPRINTF(("found leaf index %u [%s], rc = %i",
4594 i, DKEY(&nodekey), rc));
4603 while (low <= high) {
4604 i = (low + high) >> 1;
4606 node = NODEPTR(mp, i);
4607 nodekey.mv_size = NODEKSZ(node);
4608 nodekey.mv_data = NODEKEY(node);
4610 rc = cmp(key, &nodekey);
4613 DPRINTF(("found leaf index %u [%s], rc = %i",
4614 i, DKEY(&nodekey), rc));
4616 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4617 i, DKEY(&nodekey), NODEPGNO(node), rc));
4628 if (rc > 0) { /* Found entry is less than the key. */
4629 i++; /* Skip to get the smallest entry larger than key. */
4631 node = NODEPTR(mp, i);
4634 *exactp = (rc == 0 && nkeys > 0);
4635 /* store the key index */
4636 mc->mc_ki[mc->mc_top] = i;
4638 /* There is no entry larger or equal to the key. */
4641 /* nodeptr is fake for LEAF2 */
4647 mdb_cursor_adjust(MDB_cursor *mc, func)
4651 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4652 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4659 /** Pop a page off the top of the cursor's stack. */
4661 mdb_cursor_pop(MDB_cursor *mc)
4665 MDB_page *top = mc->mc_pg[mc->mc_top];
4671 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4672 DDBI(mc), (void *) mc));
4676 /** Push a page onto the top of the cursor's stack. */
4678 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4680 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4681 DDBI(mc), (void *) mc));
4683 if (mc->mc_snum >= CURSOR_STACK) {
4684 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4685 return MDB_CURSOR_FULL;
4688 mc->mc_top = mc->mc_snum++;
4689 mc->mc_pg[mc->mc_top] = mp;
4690 mc->mc_ki[mc->mc_top] = 0;
4695 /** Find the address of the page corresponding to a given page number.
4696 * @param[in] txn the transaction for this access.
4697 * @param[in] pgno the page number for the page to retrieve.
4698 * @param[out] ret address of a pointer where the page's address will be stored.
4699 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4700 * @return 0 on success, non-zero on failure.
4703 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4705 MDB_env *env = txn->mt_env;
4709 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4713 MDB_ID2L dl = tx2->mt_u.dirty_list;
4715 /* Spilled pages were dirtied in this txn and flushed
4716 * because the dirty list got full. Bring this page
4717 * back in from the map (but don't unspill it here,
4718 * leave that unless page_touch happens again).
4720 if (tx2->mt_spill_pgs) {
4721 MDB_ID pn = pgno << 1;
4722 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4723 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4724 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4729 unsigned x = mdb_mid2l_search(dl, pgno);
4730 if (x <= dl[0].mid && dl[x].mid == pgno) {
4736 } while ((tx2 = tx2->mt_parent) != NULL);
4739 if (pgno < txn->mt_next_pgno) {
4741 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4743 DPRINTF(("page %"Z"u not found", pgno));
4744 txn->mt_flags |= MDB_TXN_ERROR;
4745 return MDB_PAGE_NOTFOUND;
4755 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4756 * The cursor is at the root page, set up the rest of it.
4759 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4761 MDB_page *mp = mc->mc_pg[mc->mc_top];
4765 while (IS_BRANCH(mp)) {
4769 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4770 mdb_cassert(mc, NUMKEYS(mp) > 1);
4771 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4773 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4775 if (flags & MDB_PS_LAST)
4776 i = NUMKEYS(mp) - 1;
4779 node = mdb_node_search(mc, key, &exact);
4781 i = NUMKEYS(mp) - 1;
4783 i = mc->mc_ki[mc->mc_top];
4785 mdb_cassert(mc, i > 0);
4789 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4792 mdb_cassert(mc, i < NUMKEYS(mp));
4793 node = NODEPTR(mp, i);
4795 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4798 mc->mc_ki[mc->mc_top] = i;
4799 if ((rc = mdb_cursor_push(mc, mp)))
4802 if (flags & MDB_PS_MODIFY) {
4803 if ((rc = mdb_page_touch(mc)) != 0)
4805 mp = mc->mc_pg[mc->mc_top];
4810 DPRINTF(("internal error, index points to a %02X page!?",
4812 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4813 return MDB_CORRUPTED;
4816 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4817 key ? DKEY(key) : "null"));
4818 mc->mc_flags |= C_INITIALIZED;
4819 mc->mc_flags &= ~C_EOF;
4824 /** Search for the lowest key under the current branch page.
4825 * This just bypasses a NUMKEYS check in the current page
4826 * before calling mdb_page_search_root(), because the callers
4827 * are all in situations where the current page is known to
4831 mdb_page_search_lowest(MDB_cursor *mc)
4833 MDB_page *mp = mc->mc_pg[mc->mc_top];
4834 MDB_node *node = NODEPTR(mp, 0);
4837 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4840 mc->mc_ki[mc->mc_top] = 0;
4841 if ((rc = mdb_cursor_push(mc, mp)))
4843 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4846 /** Search for the page a given key should be in.
4847 * Push it and its parent pages on the cursor stack.
4848 * @param[in,out] mc the cursor for this operation.
4849 * @param[in] key the key to search for, or NULL for first/last page.
4850 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4851 * are touched (updated with new page numbers).
4852 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4853 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4854 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4855 * @return 0 on success, non-zero on failure.
4858 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4863 /* Make sure the txn is still viable, then find the root from
4864 * the txn's db table and set it as the root of the cursor's stack.
4866 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4867 DPUTS("transaction has failed, must abort");
4870 /* Make sure we're using an up-to-date root */
4871 if (*mc->mc_dbflag & DB_STALE) {
4873 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4874 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4881 MDB_node *leaf = mdb_node_search(&mc2,
4882 &mc->mc_dbx->md_name, &exact);
4884 return MDB_NOTFOUND;
4885 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4888 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4890 /* The txn may not know this DBI, or another process may
4891 * have dropped and recreated the DB with other flags.
4893 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4894 return MDB_INCOMPATIBLE;
4895 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4897 *mc->mc_dbflag &= ~DB_STALE;
4899 root = mc->mc_db->md_root;
4901 if (root == P_INVALID) { /* Tree is empty. */
4902 DPUTS("tree is empty");
4903 return MDB_NOTFOUND;
4907 mdb_cassert(mc, root > 1);
4908 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4909 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4915 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4916 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4918 if (flags & MDB_PS_MODIFY) {
4919 if ((rc = mdb_page_touch(mc)))
4923 if (flags & MDB_PS_ROOTONLY)
4926 return mdb_page_search_root(mc, key, flags);
4930 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4932 MDB_txn *txn = mc->mc_txn;
4933 pgno_t pg = mp->mp_pgno;
4934 unsigned x = 0, ovpages = mp->mp_pages;
4935 MDB_env *env = txn->mt_env;
4936 MDB_IDL sl = txn->mt_spill_pgs;
4937 MDB_ID pn = pg << 1;
4940 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4941 /* If the page is dirty or on the spill list we just acquired it,
4942 * so we should give it back to our current free list, if any.
4943 * Otherwise put it onto the list of pages we freed in this txn.
4945 * Won't create me_pghead: me_pglast must be inited along with it.
4946 * Unsupported in nested txns: They would need to hide the page
4947 * range in ancestor txns' dirty and spilled lists.
4949 if (env->me_pghead &&
4951 ((mp->mp_flags & P_DIRTY) ||
4952 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4956 MDB_ID2 *dl, ix, iy;
4957 rc = mdb_midl_need(&env->me_pghead, ovpages);
4960 if (!(mp->mp_flags & P_DIRTY)) {
4961 /* This page is no longer spilled */
4968 /* Remove from dirty list */
4969 dl = txn->mt_u.dirty_list;
4971 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4977 mdb_cassert(mc, x > 1);
4979 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4980 txn->mt_flags |= MDB_TXN_ERROR;
4981 return MDB_CORRUPTED;
4984 if (!(env->me_flags & MDB_WRITEMAP))
4985 mdb_dpage_free(env, mp);
4987 /* Insert in me_pghead */
4988 mop = env->me_pghead;
4989 j = mop[0] + ovpages;
4990 for (i = mop[0]; i && mop[i] < pg; i--)
4996 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5000 mc->mc_db->md_overflow_pages -= ovpages;
5004 /** Return the data associated with a given node.
5005 * @param[in] txn The transaction for this operation.
5006 * @param[in] leaf The node being read.
5007 * @param[out] data Updated to point to the node's data.
5008 * @return 0 on success, non-zero on failure.
5011 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5013 MDB_page *omp; /* overflow page */
5017 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5018 data->mv_size = NODEDSZ(leaf);
5019 data->mv_data = NODEDATA(leaf);
5023 /* Read overflow data.
5025 data->mv_size = NODEDSZ(leaf);
5026 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5027 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5028 DPRINTF(("read overflow page %"Z"u failed", pgno));
5031 data->mv_data = METADATA(omp);
5037 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5038 MDB_val *key, MDB_val *data)
5045 if (key == NULL || data == NULL)
5048 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5050 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5053 if (txn->mt_flags & MDB_TXN_ERROR)
5056 mdb_cursor_init(&mc, txn, dbi, &mx);
5057 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5060 /** Find a sibling for a page.
5061 * Replaces the page at the top of the cursor's stack with the
5062 * specified sibling, if one exists.
5063 * @param[in] mc The cursor for this operation.
5064 * @param[in] move_right Non-zero if the right sibling is requested,
5065 * otherwise the left sibling.
5066 * @return 0 on success, non-zero on failure.
5069 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5075 if (mc->mc_snum < 2) {
5076 return MDB_NOTFOUND; /* root has no siblings */
5080 DPRINTF(("parent page is page %"Z"u, index %u",
5081 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5083 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5084 : (mc->mc_ki[mc->mc_top] == 0)) {
5085 DPRINTF(("no more keys left, moving to %s sibling",
5086 move_right ? "right" : "left"));
5087 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5088 /* undo cursor_pop before returning */
5095 mc->mc_ki[mc->mc_top]++;
5097 mc->mc_ki[mc->mc_top]--;
5098 DPRINTF(("just moving to %s index key %u",
5099 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5101 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5103 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5104 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5105 /* mc will be inconsistent if caller does mc_snum++ as above */
5106 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5110 mdb_cursor_push(mc, mp);
5112 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5117 /** Move the cursor to the next data item. */
5119 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5125 if (mc->mc_flags & C_EOF) {
5126 return MDB_NOTFOUND;
5129 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5131 mp = mc->mc_pg[mc->mc_top];
5133 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5134 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5135 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5136 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5137 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5138 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5139 if (rc == MDB_SUCCESS)
5140 MDB_GET_KEY(leaf, key);
5145 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5146 if (op == MDB_NEXT_DUP)
5147 return MDB_NOTFOUND;
5151 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5152 mdb_dbg_pgno(mp), (void *) mc));
5153 if (mc->mc_flags & C_DEL)
5156 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5157 DPUTS("=====> move to next sibling page");
5158 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5159 mc->mc_flags |= C_EOF;
5162 mp = mc->mc_pg[mc->mc_top];
5163 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5165 mc->mc_ki[mc->mc_top]++;
5168 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5169 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5172 key->mv_size = mc->mc_db->md_pad;
5173 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5177 mdb_cassert(mc, IS_LEAF(mp));
5178 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5180 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5181 mdb_xcursor_init1(mc, leaf);
5184 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5187 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5188 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5189 if (rc != MDB_SUCCESS)
5194 MDB_GET_KEY(leaf, key);
5198 /** Move the cursor to the previous data item. */
5200 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5206 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5208 mp = mc->mc_pg[mc->mc_top];
5210 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5211 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5212 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5213 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5214 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5215 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5216 if (rc == MDB_SUCCESS)
5217 MDB_GET_KEY(leaf, key);
5221 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5222 if (op == MDB_PREV_DUP)
5223 return MDB_NOTFOUND;
5228 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5229 mdb_dbg_pgno(mp), (void *) mc));
5231 if (mc->mc_ki[mc->mc_top] == 0) {
5232 DPUTS("=====> move to prev sibling page");
5233 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5236 mp = mc->mc_pg[mc->mc_top];
5237 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5238 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5240 mc->mc_ki[mc->mc_top]--;
5242 mc->mc_flags &= ~C_EOF;
5244 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5245 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5248 key->mv_size = mc->mc_db->md_pad;
5249 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5253 mdb_cassert(mc, IS_LEAF(mp));
5254 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5256 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5257 mdb_xcursor_init1(mc, leaf);
5260 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5263 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5264 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5265 if (rc != MDB_SUCCESS)
5270 MDB_GET_KEY(leaf, key);
5274 /** Set the cursor on a specific data item. */
5276 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5277 MDB_cursor_op op, int *exactp)
5281 MDB_node *leaf = NULL;
5284 if (key->mv_size == 0)
5285 return MDB_BAD_VALSIZE;
5288 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5290 /* See if we're already on the right page */
5291 if (mc->mc_flags & C_INITIALIZED) {
5294 mp = mc->mc_pg[mc->mc_top];
5296 mc->mc_ki[mc->mc_top] = 0;
5297 return MDB_NOTFOUND;
5299 if (mp->mp_flags & P_LEAF2) {
5300 nodekey.mv_size = mc->mc_db->md_pad;
5301 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5303 leaf = NODEPTR(mp, 0);
5304 MDB_GET_KEY2(leaf, nodekey);
5306 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5308 /* Probably happens rarely, but first node on the page
5309 * was the one we wanted.
5311 mc->mc_ki[mc->mc_top] = 0;
5318 unsigned int nkeys = NUMKEYS(mp);
5320 if (mp->mp_flags & P_LEAF2) {
5321 nodekey.mv_data = LEAF2KEY(mp,
5322 nkeys-1, nodekey.mv_size);
5324 leaf = NODEPTR(mp, nkeys-1);
5325 MDB_GET_KEY2(leaf, nodekey);
5327 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5329 /* last node was the one we wanted */
5330 mc->mc_ki[mc->mc_top] = nkeys-1;
5336 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5337 /* This is definitely the right page, skip search_page */
5338 if (mp->mp_flags & P_LEAF2) {
5339 nodekey.mv_data = LEAF2KEY(mp,
5340 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5342 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5343 MDB_GET_KEY2(leaf, nodekey);
5345 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5347 /* current node was the one we wanted */
5357 /* If any parents have right-sibs, search.
5358 * Otherwise, there's nothing further.
5360 for (i=0; i<mc->mc_top; i++)
5362 NUMKEYS(mc->mc_pg[i])-1)
5364 if (i == mc->mc_top) {
5365 /* There are no other pages */
5366 mc->mc_ki[mc->mc_top] = nkeys;
5367 return MDB_NOTFOUND;
5371 /* There are no other pages */
5372 mc->mc_ki[mc->mc_top] = 0;
5373 if (op == MDB_SET_RANGE) {
5377 return MDB_NOTFOUND;
5381 rc = mdb_page_search(mc, key, 0);
5382 if (rc != MDB_SUCCESS)
5385 mp = mc->mc_pg[mc->mc_top];
5386 mdb_cassert(mc, IS_LEAF(mp));
5389 leaf = mdb_node_search(mc, key, exactp);
5390 if (exactp != NULL && !*exactp) {
5391 /* MDB_SET specified and not an exact match. */
5392 return MDB_NOTFOUND;
5396 DPUTS("===> inexact leaf not found, goto sibling");
5397 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5398 return rc; /* no entries matched */
5399 mp = mc->mc_pg[mc->mc_top];
5400 mdb_cassert(mc, IS_LEAF(mp));
5401 leaf = NODEPTR(mp, 0);
5405 mc->mc_flags |= C_INITIALIZED;
5406 mc->mc_flags &= ~C_EOF;
5409 key->mv_size = mc->mc_db->md_pad;
5410 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5414 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5415 mdb_xcursor_init1(mc, leaf);
5418 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5419 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5420 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5423 if (op == MDB_GET_BOTH) {
5429 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5430 if (rc != MDB_SUCCESS)
5433 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5435 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5437 rc = mc->mc_dbx->md_dcmp(data, &d2);
5439 if (op == MDB_GET_BOTH || rc > 0)
5440 return MDB_NOTFOUND;
5447 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5448 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5453 /* The key already matches in all other cases */
5454 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5455 MDB_GET_KEY(leaf, key);
5456 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5461 /** Move the cursor to the first item in the database. */
5463 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5469 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5471 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5472 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5473 if (rc != MDB_SUCCESS)
5476 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5478 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5479 mc->mc_flags |= C_INITIALIZED;
5480 mc->mc_flags &= ~C_EOF;
5482 mc->mc_ki[mc->mc_top] = 0;
5484 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5485 key->mv_size = mc->mc_db->md_pad;
5486 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5491 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5492 mdb_xcursor_init1(mc, leaf);
5493 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5497 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5501 MDB_GET_KEY(leaf, key);
5505 /** Move the cursor to the last item in the database. */
5507 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5513 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5515 if (!(mc->mc_flags & C_EOF)) {
5517 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5518 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5519 if (rc != MDB_SUCCESS)
5522 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5525 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5526 mc->mc_flags |= C_INITIALIZED|C_EOF;
5527 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5529 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5530 key->mv_size = mc->mc_db->md_pad;
5531 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5536 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5537 mdb_xcursor_init1(mc, leaf);
5538 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5542 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5547 MDB_GET_KEY(leaf, key);
5552 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5557 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5562 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5566 case MDB_GET_CURRENT:
5567 if (!(mc->mc_flags & C_INITIALIZED)) {
5570 MDB_page *mp = mc->mc_pg[mc->mc_top];
5571 int nkeys = NUMKEYS(mp);
5572 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5573 mc->mc_ki[mc->mc_top] = nkeys;
5579 key->mv_size = mc->mc_db->md_pad;
5580 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5582 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5583 MDB_GET_KEY(leaf, key);
5585 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5586 if (mc->mc_flags & C_DEL)
5587 mdb_xcursor_init1(mc, leaf);
5588 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5590 rc = mdb_node_read(mc->mc_txn, leaf, data);
5597 case MDB_GET_BOTH_RANGE:
5602 if (mc->mc_xcursor == NULL) {
5603 rc = MDB_INCOMPATIBLE;
5613 rc = mdb_cursor_set(mc, key, data, op,
5614 op == MDB_SET_RANGE ? NULL : &exact);
5617 case MDB_GET_MULTIPLE:
5618 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5622 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5623 rc = MDB_INCOMPATIBLE;
5627 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5628 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5631 case MDB_NEXT_MULTIPLE:
5636 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5637 rc = MDB_INCOMPATIBLE;
5640 if (!(mc->mc_flags & C_INITIALIZED))
5641 rc = mdb_cursor_first(mc, key, data);
5643 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5644 if (rc == MDB_SUCCESS) {
5645 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5648 mx = &mc->mc_xcursor->mx_cursor;
5649 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5651 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5652 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5660 case MDB_NEXT_NODUP:
5661 if (!(mc->mc_flags & C_INITIALIZED))
5662 rc = mdb_cursor_first(mc, key, data);
5664 rc = mdb_cursor_next(mc, key, data, op);
5668 case MDB_PREV_NODUP:
5669 if (!(mc->mc_flags & C_INITIALIZED)) {
5670 rc = mdb_cursor_last(mc, key, data);
5673 mc->mc_flags |= C_INITIALIZED;
5674 mc->mc_ki[mc->mc_top]++;
5676 rc = mdb_cursor_prev(mc, key, data, op);
5679 rc = mdb_cursor_first(mc, key, data);
5682 mfunc = mdb_cursor_first;
5684 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5688 if (mc->mc_xcursor == NULL) {
5689 rc = MDB_INCOMPATIBLE;
5692 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5696 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5699 rc = mdb_cursor_last(mc, key, data);
5702 mfunc = mdb_cursor_last;
5705 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5710 if (mc->mc_flags & C_DEL)
5711 mc->mc_flags ^= C_DEL;
5716 /** Touch all the pages in the cursor stack. Set mc_top.
5717 * Makes sure all the pages are writable, before attempting a write operation.
5718 * @param[in] mc The cursor to operate on.
5721 mdb_cursor_touch(MDB_cursor *mc)
5723 int rc = MDB_SUCCESS;
5725 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5728 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5729 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5732 *mc->mc_dbflag |= DB_DIRTY;
5737 rc = mdb_page_touch(mc);
5738 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5739 mc->mc_top = mc->mc_snum-1;
5744 /** Do not spill pages to disk if txn is getting full, may fail instead */
5745 #define MDB_NOSPILL 0x8000
5748 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5751 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5753 MDB_node *leaf = NULL;
5756 MDB_val xdata, *rdata, dkey, olddata;
5758 int do_sub = 0, insert;
5759 unsigned int mcount = 0, dcount = 0, nospill;
5762 unsigned int nflags;
5768 env = mc->mc_txn->mt_env;
5770 /* Check this first so counter will always be zero on any
5773 if (flags & MDB_MULTIPLE) {
5774 dcount = data[1].mv_size;
5775 data[1].mv_size = 0;
5776 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5777 return MDB_INCOMPATIBLE;
5780 nospill = flags & MDB_NOSPILL;
5781 flags &= ~MDB_NOSPILL;
5783 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5784 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5786 if (flags != MDB_CURRENT) {
5789 if (key->mv_size-1 >= ENV_MAXKEY(env))
5790 return MDB_BAD_VALSIZE;
5792 /* Ignore key except in sub-cursor, where key holds the data */
5793 if (!(mc->mc_flags & C_SUB))
5797 #if SIZE_MAX > MAXDATASIZE
5798 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5799 return MDB_BAD_VALSIZE;
5801 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5802 return MDB_BAD_VALSIZE;
5805 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5806 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5810 if (flags == MDB_CURRENT) {
5811 if (!(mc->mc_flags & C_INITIALIZED))
5814 } else if (mc->mc_db->md_root == P_INVALID) {
5815 /* new database, cursor has nothing to point to */
5818 mc->mc_flags &= ~C_INITIALIZED;
5823 if (flags & MDB_APPEND) {
5825 rc = mdb_cursor_last(mc, &k2, &d2);
5827 rc = mc->mc_dbx->md_cmp(key, &k2);
5830 mc->mc_ki[mc->mc_top]++;
5832 /* new key is <= last key */
5837 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5839 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5840 DPRINTF(("duplicate key [%s]", DKEY(key)));
5842 return MDB_KEYEXIST;
5844 if (rc && rc != MDB_NOTFOUND)
5848 if (mc->mc_flags & C_DEL)
5849 mc->mc_flags ^= C_DEL;
5851 /* Cursor is positioned, check for room in the dirty list */
5853 if (flags & MDB_MULTIPLE) {
5855 xdata.mv_size = data->mv_size * dcount;
5859 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5863 if (rc == MDB_NO_ROOT) {
5865 /* new database, write a root leaf page */
5866 DPUTS("allocating new root leaf page");
5867 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5870 mdb_cursor_push(mc, np);
5871 mc->mc_db->md_root = np->mp_pgno;
5872 mc->mc_db->md_depth++;
5873 *mc->mc_dbflag |= DB_DIRTY;
5874 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5876 np->mp_flags |= P_LEAF2;
5877 mc->mc_flags |= C_INITIALIZED;
5879 /* make sure all cursor pages are writable */
5880 rc2 = mdb_cursor_touch(mc);
5887 /* The key does not exist */
5888 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5889 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5890 LEAFSIZE(key, data) > env->me_nodemax)
5892 /* Too big for a node, insert in sub-DB */
5893 fp_flags = P_LEAF|P_DIRTY;
5895 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5896 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5900 /* there's only a key anyway, so this is a no-op */
5901 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5902 unsigned int ksize = mc->mc_db->md_pad;
5903 if (key->mv_size != ksize)
5904 return MDB_BAD_VALSIZE;
5905 if (flags == MDB_CURRENT) {
5906 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5907 memcpy(ptr, key->mv_data, ksize);
5913 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5914 olddata.mv_size = NODEDSZ(leaf);
5915 olddata.mv_data = NODEDATA(leaf);
5918 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5919 /* Prepare (sub-)page/sub-DB to accept the new item,
5920 * if needed. fp: old sub-page or a header faking
5921 * it. mp: new (sub-)page. offset: growth in page
5922 * size. xdata: node data with new page or DB.
5924 unsigned i, offset = 0;
5925 mp = fp = xdata.mv_data = env->me_pbuf;
5926 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5928 /* Was a single item before, must convert now */
5929 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5930 /* Just overwrite the current item */
5931 if (flags == MDB_CURRENT)
5934 #if UINT_MAX < SIZE_MAX
5935 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5936 #ifdef MISALIGNED_OK
5937 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5939 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5942 /* if data matches, skip it */
5943 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5944 if (flags & MDB_NODUPDATA)
5946 else if (flags & MDB_MULTIPLE)
5953 /* Back up original data item */
5954 dkey.mv_size = olddata.mv_size;
5955 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5957 /* Make sub-page header for the dup items, with dummy body */
5958 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5959 fp->mp_lower = PAGEHDRSZ;
5960 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5961 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5962 fp->mp_flags |= P_LEAF2;
5963 fp->mp_pad = data->mv_size;
5964 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
5966 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
5967 (dkey.mv_size & 1) + (data->mv_size & 1);
5969 fp->mp_upper = xdata.mv_size;
5970 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
5971 } else if (leaf->mn_flags & F_SUBDATA) {
5972 /* Data is on sub-DB, just store it */
5973 flags |= F_DUPDATA|F_SUBDATA;
5976 /* Data is on sub-page */
5977 fp = olddata.mv_data;
5980 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5981 offset = EVEN(NODESIZE + sizeof(indx_t) +
5985 offset = fp->mp_pad;
5986 if (SIZELEFT(fp) < offset) {
5987 offset *= 4; /* space for 4 more */
5990 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
5992 fp->mp_flags |= P_DIRTY;
5993 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
5994 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5998 xdata.mv_size = olddata.mv_size + offset;
6001 fp_flags = fp->mp_flags;
6002 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6003 /* Too big for a sub-page, convert to sub-DB */
6004 fp_flags &= ~P_SUBP;
6006 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6007 fp_flags |= P_LEAF2;
6008 dummy.md_pad = fp->mp_pad;
6009 dummy.md_flags = MDB_DUPFIXED;
6010 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6011 dummy.md_flags |= MDB_INTEGERKEY;
6017 dummy.md_branch_pages = 0;
6018 dummy.md_leaf_pages = 1;
6019 dummy.md_overflow_pages = 0;
6020 dummy.md_entries = NUMKEYS(fp);
6021 xdata.mv_size = sizeof(MDB_db);
6022 xdata.mv_data = &dummy;
6023 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6025 offset = env->me_psize - olddata.mv_size;
6026 flags |= F_DUPDATA|F_SUBDATA;
6027 dummy.md_root = mp->mp_pgno;
6030 mp->mp_flags = fp_flags | P_DIRTY;
6031 mp->mp_pad = fp->mp_pad;
6032 mp->mp_lower = fp->mp_lower;
6033 mp->mp_upper = fp->mp_upper + offset;
6034 if (fp_flags & P_LEAF2) {
6035 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6037 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
6038 olddata.mv_size - fp->mp_upper);
6039 for (i=0; i<NUMKEYS(fp); i++)
6040 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6048 mdb_node_del(mc, 0);
6052 /* overflow page overwrites need special handling */
6053 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6056 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6058 memcpy(&pg, olddata.mv_data, sizeof(pg));
6059 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6061 ovpages = omp->mp_pages;
6063 /* Is the ov page large enough? */
6064 if (ovpages >= dpages) {
6065 if (!(omp->mp_flags & P_DIRTY) &&
6066 (level || (env->me_flags & MDB_WRITEMAP)))
6068 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6071 level = 0; /* dirty in this txn or clean */
6074 if (omp->mp_flags & P_DIRTY) {
6075 /* yes, overwrite it. Note in this case we don't
6076 * bother to try shrinking the page if the new data
6077 * is smaller than the overflow threshold.
6080 /* It is writable only in a parent txn */
6081 size_t sz = (size_t) env->me_psize * ovpages, off;
6082 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6088 rc = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6089 mdb_cassert(mc, rc == 0);
6090 if (!(flags & MDB_RESERVE)) {
6091 /* Copy end of page, adjusting alignment so
6092 * compiler may copy words instead of bytes.
6094 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6095 memcpy((size_t *)((char *)np + off),
6096 (size_t *)((char *)omp + off), sz - off);
6099 memcpy(np, omp, sz); /* Copy beginning of page */
6102 SETDSZ(leaf, data->mv_size);
6103 if (F_ISSET(flags, MDB_RESERVE))
6104 data->mv_data = METADATA(omp);
6106 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6110 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6112 } else if (data->mv_size == olddata.mv_size) {
6113 /* same size, just replace it. Note that we could
6114 * also reuse this node if the new data is smaller,
6115 * but instead we opt to shrink the node in that case.
6117 if (F_ISSET(flags, MDB_RESERVE))
6118 data->mv_data = olddata.mv_data;
6119 else if (!(mc->mc_flags & C_SUB))
6120 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6122 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6125 mdb_node_del(mc, 0);
6126 mc->mc_db->md_entries--;
6132 nflags = flags & NODE_ADD_FLAGS;
6133 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6134 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6135 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6136 nflags &= ~MDB_APPEND;
6138 nflags |= MDB_SPLIT_REPLACE;
6139 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6141 /* There is room already in this leaf page. */
6142 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6143 if (rc == 0 && !do_sub && insert) {
6144 /* Adjust other cursors pointing to mp */
6145 MDB_cursor *m2, *m3;
6146 MDB_dbi dbi = mc->mc_dbi;
6147 unsigned i = mc->mc_top;
6148 MDB_page *mp = mc->mc_pg[i];
6150 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6151 if (mc->mc_flags & C_SUB)
6152 m3 = &m2->mc_xcursor->mx_cursor;
6155 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6156 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6163 if (rc != MDB_SUCCESS)
6164 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6166 /* Now store the actual data in the child DB. Note that we're
6167 * storing the user data in the keys field, so there are strict
6168 * size limits on dupdata. The actual data fields of the child
6169 * DB are all zero size.
6176 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6177 if (flags & MDB_CURRENT) {
6178 xflags = MDB_CURRENT|MDB_NOSPILL;
6180 mdb_xcursor_init1(mc, leaf);
6181 xflags = (flags & MDB_NODUPDATA) ?
6182 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6184 /* converted, write the original data first */
6186 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6190 /* Adjust other cursors pointing to mp */
6192 unsigned i = mc->mc_top;
6193 MDB_page *mp = mc->mc_pg[i];
6195 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6196 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6197 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6198 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6199 mdb_xcursor_init1(m2, leaf);
6203 /* we've done our job */
6206 if (flags & MDB_APPENDDUP)
6207 xflags |= MDB_APPEND;
6208 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6209 if (flags & F_SUBDATA) {
6210 void *db = NODEDATA(leaf);
6211 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6214 /* sub-writes might have failed so check rc again.
6215 * Don't increment count if we just replaced an existing item.
6217 if (!rc && !(flags & MDB_CURRENT))
6218 mc->mc_db->md_entries++;
6219 if (flags & MDB_MULTIPLE) {
6223 /* let caller know how many succeeded, if any */
6224 data[1].mv_size = mcount;
6225 if (mcount < dcount) {
6226 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6233 /* If we succeeded and the key didn't exist before, make sure
6234 * the cursor is marked valid.
6237 mc->mc_flags |= C_INITIALIZED;
6242 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6248 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6249 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6251 if (!(mc->mc_flags & C_INITIALIZED))
6254 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6255 return MDB_NOTFOUND;
6257 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6260 rc = mdb_cursor_touch(mc);
6264 mp = mc->mc_pg[mc->mc_top];
6265 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6267 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6268 if (!(flags & MDB_NODUPDATA)) {
6269 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6270 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6272 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6273 /* If sub-DB still has entries, we're done */
6274 if (mc->mc_xcursor->mx_db.md_entries) {
6275 if (leaf->mn_flags & F_SUBDATA) {
6276 /* update subDB info */
6277 void *db = NODEDATA(leaf);
6278 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6281 /* shrink fake page */
6282 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6283 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6284 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6285 /* fix other sub-DB cursors pointed at this fake page */
6286 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6287 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6288 if (m2->mc_pg[mc->mc_top] == mp &&
6289 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6290 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6293 mc->mc_db->md_entries--;
6294 mc->mc_flags |= C_DEL;
6297 /* otherwise fall thru and delete the sub-DB */
6300 if (leaf->mn_flags & F_SUBDATA) {
6301 /* add all the child DB's pages to the free list */
6302 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6303 if (rc == MDB_SUCCESS) {
6304 mc->mc_db->md_entries -=
6305 mc->mc_xcursor->mx_db.md_entries;
6310 return mdb_cursor_del0(mc, leaf);
6313 /** Allocate and initialize new pages for a database.
6314 * @param[in] mc a cursor on the database being added to.
6315 * @param[in] flags flags defining what type of page is being allocated.
6316 * @param[in] num the number of pages to allocate. This is usually 1,
6317 * unless allocating overflow pages for a large record.
6318 * @param[out] mp Address of a page, or NULL on failure.
6319 * @return 0 on success, non-zero on failure.
6322 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6327 if ((rc = mdb_page_alloc(mc, num, &np)))
6329 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6330 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6331 np->mp_flags = flags | P_DIRTY;
6332 np->mp_lower = PAGEHDRSZ;
6333 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6336 mc->mc_db->md_branch_pages++;
6337 else if (IS_LEAF(np))
6338 mc->mc_db->md_leaf_pages++;
6339 else if (IS_OVERFLOW(np)) {
6340 mc->mc_db->md_overflow_pages += num;
6348 /** Calculate the size of a leaf node.
6349 * The size depends on the environment's page size; if a data item
6350 * is too large it will be put onto an overflow page and the node
6351 * size will only include the key and not the data. Sizes are always
6352 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6353 * of the #MDB_node headers.
6354 * @param[in] env The environment handle.
6355 * @param[in] key The key for the node.
6356 * @param[in] data The data for the node.
6357 * @return The number of bytes needed to store the node.
6360 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6364 sz = LEAFSIZE(key, data);
6365 if (sz > env->me_nodemax) {
6366 /* put on overflow page */
6367 sz -= data->mv_size - sizeof(pgno_t);
6370 return EVEN(sz + sizeof(indx_t));
6373 /** Calculate the size of a branch node.
6374 * The size should depend on the environment's page size but since
6375 * we currently don't support spilling large keys onto overflow
6376 * pages, it's simply the size of the #MDB_node header plus the
6377 * size of the key. Sizes are always rounded up to an even number
6378 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6379 * @param[in] env The environment handle.
6380 * @param[in] key The key for the node.
6381 * @return The number of bytes needed to store the node.
6384 mdb_branch_size(MDB_env *env, MDB_val *key)
6389 if (sz > env->me_nodemax) {
6390 /* put on overflow page */
6391 /* not implemented */
6392 /* sz -= key->size - sizeof(pgno_t); */
6395 return sz + sizeof(indx_t);
6398 /** Add a node to the page pointed to by the cursor.
6399 * @param[in] mc The cursor for this operation.
6400 * @param[in] indx The index on the page where the new node should be added.
6401 * @param[in] key The key for the new node.
6402 * @param[in] data The data for the new node, if any.
6403 * @param[in] pgno The page number, if adding a branch node.
6404 * @param[in] flags Flags for the node.
6405 * @return 0 on success, non-zero on failure. Possible errors are:
6407 * <li>ENOMEM - failed to allocate overflow pages for the node.
6408 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6409 * should never happen since all callers already calculate the
6410 * page's free space before calling this function.
6414 mdb_node_add(MDB_cursor *mc, indx_t indx,
6415 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6418 size_t node_size = NODESIZE;
6422 MDB_page *mp = mc->mc_pg[mc->mc_top];
6423 MDB_page *ofp = NULL; /* overflow page */
6426 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6428 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6429 IS_LEAF(mp) ? "leaf" : "branch",
6430 IS_SUBP(mp) ? "sub-" : "",
6431 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6432 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6435 /* Move higher keys up one slot. */
6436 int ksize = mc->mc_db->md_pad, dif;
6437 char *ptr = LEAF2KEY(mp, indx, ksize);
6438 dif = NUMKEYS(mp) - indx;
6440 memmove(ptr+ksize, ptr, dif*ksize);
6441 /* insert new key */
6442 memcpy(ptr, key->mv_data, ksize);
6444 /* Just using these for counting */
6445 mp->mp_lower += sizeof(indx_t);
6446 mp->mp_upper -= ksize - sizeof(indx_t);
6450 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6452 node_size += key->mv_size;
6454 mdb_cassert(mc, data);
6455 if (F_ISSET(flags, F_BIGDATA)) {
6456 /* Data already on overflow page. */
6457 node_size += sizeof(pgno_t);
6458 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6459 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6461 /* Put data on overflow page. */
6462 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6463 data->mv_size, node_size+data->mv_size));
6464 node_size = EVEN(node_size + sizeof(pgno_t));
6465 if ((ssize_t)node_size > room)
6467 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6469 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6473 node_size += data->mv_size;
6476 node_size = EVEN(node_size);
6477 if ((ssize_t)node_size > room)
6481 /* Move higher pointers up one slot. */
6482 for (i = NUMKEYS(mp); i > indx; i--)
6483 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6485 /* Adjust free space offsets. */
6486 ofs = mp->mp_upper - node_size;
6487 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6488 mp->mp_ptrs[indx] = ofs;
6490 mp->mp_lower += sizeof(indx_t);
6492 /* Write the node data. */
6493 node = NODEPTR(mp, indx);
6494 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6495 node->mn_flags = flags;
6497 SETDSZ(node,data->mv_size);
6502 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6505 mdb_cassert(mc, key);
6507 if (F_ISSET(flags, F_BIGDATA))
6508 memcpy(node->mn_data + key->mv_size, data->mv_data,
6510 else if (F_ISSET(flags, MDB_RESERVE))
6511 data->mv_data = node->mn_data + key->mv_size;
6513 memcpy(node->mn_data + key->mv_size, data->mv_data,
6516 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6518 if (F_ISSET(flags, MDB_RESERVE))
6519 data->mv_data = METADATA(ofp);
6521 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6528 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6529 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6530 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6531 DPRINTF(("node size = %"Z"u", node_size));
6532 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6533 return MDB_PAGE_FULL;
6536 /** Delete the specified node from a page.
6537 * @param[in] mc Cursor pointing to the node to delete.
6538 * @param[in] ksize The size of a node. Only used if the page is
6539 * part of a #MDB_DUPFIXED database.
6542 mdb_node_del(MDB_cursor *mc, int ksize)
6544 MDB_page *mp = mc->mc_pg[mc->mc_top];
6545 indx_t indx = mc->mc_ki[mc->mc_top];
6547 indx_t i, j, numkeys, ptr;
6551 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6552 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6553 numkeys = NUMKEYS(mp);
6554 mdb_cassert(mc, indx < numkeys);
6557 int x = numkeys - 1 - indx;
6558 base = LEAF2KEY(mp, indx, ksize);
6560 memmove(base, base + ksize, x * ksize);
6561 mp->mp_lower -= sizeof(indx_t);
6562 mp->mp_upper += ksize - sizeof(indx_t);
6566 node = NODEPTR(mp, indx);
6567 sz = NODESIZE + node->mn_ksize;
6569 if (F_ISSET(node->mn_flags, F_BIGDATA))
6570 sz += sizeof(pgno_t);
6572 sz += NODEDSZ(node);
6576 ptr = mp->mp_ptrs[indx];
6577 for (i = j = 0; i < numkeys; i++) {
6579 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6580 if (mp->mp_ptrs[i] < ptr)
6581 mp->mp_ptrs[j] += sz;
6586 base = (char *)mp + mp->mp_upper;
6587 memmove(base + sz, base, ptr - mp->mp_upper);
6589 mp->mp_lower -= sizeof(indx_t);
6593 /** Compact the main page after deleting a node on a subpage.
6594 * @param[in] mp The main page to operate on.
6595 * @param[in] indx The index of the subpage on the main page.
6598 mdb_node_shrink(MDB_page *mp, indx_t indx)
6604 indx_t i, numkeys, ptr;
6606 node = NODEPTR(mp, indx);
6607 sp = (MDB_page *)NODEDATA(node);
6608 delta = SIZELEFT(sp);
6609 xp = (MDB_page *)((char *)sp + delta);
6611 /* shift subpage upward */
6613 nsize = NUMKEYS(sp) * sp->mp_pad;
6615 return; /* do not make the node uneven-sized */
6616 memmove(METADATA(xp), METADATA(sp), nsize);
6619 numkeys = NUMKEYS(sp);
6620 for (i=numkeys-1; i>=0; i--)
6621 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6623 xp->mp_upper = sp->mp_lower;
6624 xp->mp_lower = sp->mp_lower;
6625 xp->mp_flags = sp->mp_flags;
6626 xp->mp_pad = sp->mp_pad;
6627 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6629 nsize = NODEDSZ(node) - delta;
6630 SETDSZ(node, nsize);
6632 /* shift lower nodes upward */
6633 ptr = mp->mp_ptrs[indx];
6634 numkeys = NUMKEYS(mp);
6635 for (i = 0; i < numkeys; i++) {
6636 if (mp->mp_ptrs[i] <= ptr)
6637 mp->mp_ptrs[i] += delta;
6640 base = (char *)mp + mp->mp_upper;
6641 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6642 mp->mp_upper += delta;
6645 /** Initial setup of a sorted-dups cursor.
6646 * Sorted duplicates are implemented as a sub-database for the given key.
6647 * The duplicate data items are actually keys of the sub-database.
6648 * Operations on the duplicate data items are performed using a sub-cursor
6649 * initialized when the sub-database is first accessed. This function does
6650 * the preliminary setup of the sub-cursor, filling in the fields that
6651 * depend only on the parent DB.
6652 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6655 mdb_xcursor_init0(MDB_cursor *mc)
6657 MDB_xcursor *mx = mc->mc_xcursor;
6659 mx->mx_cursor.mc_xcursor = NULL;
6660 mx->mx_cursor.mc_txn = mc->mc_txn;
6661 mx->mx_cursor.mc_db = &mx->mx_db;
6662 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6663 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6664 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6665 mx->mx_cursor.mc_snum = 0;
6666 mx->mx_cursor.mc_top = 0;
6667 mx->mx_cursor.mc_flags = C_SUB;
6668 mx->mx_dbx.md_name.mv_size = 0;
6669 mx->mx_dbx.md_name.mv_data = NULL;
6670 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6671 mx->mx_dbx.md_dcmp = NULL;
6672 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6675 /** Final setup of a sorted-dups cursor.
6676 * Sets up the fields that depend on the data from the main cursor.
6677 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6678 * @param[in] node The data containing the #MDB_db record for the
6679 * sorted-dup database.
6682 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6684 MDB_xcursor *mx = mc->mc_xcursor;
6686 if (node->mn_flags & F_SUBDATA) {
6687 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6688 mx->mx_cursor.mc_pg[0] = 0;
6689 mx->mx_cursor.mc_snum = 0;
6690 mx->mx_cursor.mc_top = 0;
6691 mx->mx_cursor.mc_flags = C_SUB;
6693 MDB_page *fp = NODEDATA(node);
6694 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6695 mx->mx_db.md_flags = 0;
6696 mx->mx_db.md_depth = 1;
6697 mx->mx_db.md_branch_pages = 0;
6698 mx->mx_db.md_leaf_pages = 1;
6699 mx->mx_db.md_overflow_pages = 0;
6700 mx->mx_db.md_entries = NUMKEYS(fp);
6701 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6702 mx->mx_cursor.mc_snum = 1;
6703 mx->mx_cursor.mc_top = 0;
6704 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6705 mx->mx_cursor.mc_pg[0] = fp;
6706 mx->mx_cursor.mc_ki[0] = 0;
6707 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6708 mx->mx_db.md_flags = MDB_DUPFIXED;
6709 mx->mx_db.md_pad = fp->mp_pad;
6710 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6711 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6714 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6715 mx->mx_db.md_root));
6716 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6717 #if UINT_MAX < SIZE_MAX
6718 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6719 #ifdef MISALIGNED_OK
6720 mx->mx_dbx.md_cmp = mdb_cmp_long;
6722 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6727 /** Initialize a cursor for a given transaction and database. */
6729 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6732 mc->mc_backup = NULL;
6735 mc->mc_db = &txn->mt_dbs[dbi];
6736 mc->mc_dbx = &txn->mt_dbxs[dbi];
6737 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6742 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6743 mdb_tassert(txn, mx != NULL);
6744 mc->mc_xcursor = mx;
6745 mdb_xcursor_init0(mc);
6747 mc->mc_xcursor = NULL;
6749 if (*mc->mc_dbflag & DB_STALE) {
6750 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6755 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6758 size_t size = sizeof(MDB_cursor);
6760 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6763 if (txn->mt_flags & MDB_TXN_ERROR)
6766 /* Allow read access to the freelist */
6767 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6770 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6771 size += sizeof(MDB_xcursor);
6773 if ((mc = malloc(size)) != NULL) {
6774 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6775 if (txn->mt_cursors) {
6776 mc->mc_next = txn->mt_cursors[dbi];
6777 txn->mt_cursors[dbi] = mc;
6778 mc->mc_flags |= C_UNTRACK;
6790 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6792 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6795 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6798 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6802 /* Return the count of duplicate data items for the current key */
6804 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6808 if (mc == NULL || countp == NULL)
6811 if (mc->mc_xcursor == NULL)
6812 return MDB_INCOMPATIBLE;
6814 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6815 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6818 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6821 *countp = mc->mc_xcursor->mx_db.md_entries;
6827 mdb_cursor_close(MDB_cursor *mc)
6829 if (mc && !mc->mc_backup) {
6830 /* remove from txn, if tracked */
6831 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6832 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6833 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6835 *prev = mc->mc_next;
6842 mdb_cursor_txn(MDB_cursor *mc)
6844 if (!mc) return NULL;
6849 mdb_cursor_dbi(MDB_cursor *mc)
6854 /** Replace the key for a branch node with a new key.
6855 * @param[in] mc Cursor pointing to the node to operate on.
6856 * @param[in] key The new key to use.
6857 * @return 0 on success, non-zero on failure.
6860 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6866 int delta, ksize, oksize;
6867 indx_t ptr, i, numkeys, indx;
6870 indx = mc->mc_ki[mc->mc_top];
6871 mp = mc->mc_pg[mc->mc_top];
6872 node = NODEPTR(mp, indx);
6873 ptr = mp->mp_ptrs[indx];
6877 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6878 k2.mv_data = NODEKEY(node);
6879 k2.mv_size = node->mn_ksize;
6880 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6882 mdb_dkey(&k2, kbuf2),
6888 /* Sizes must be 2-byte aligned. */
6889 ksize = EVEN(key->mv_size);
6890 oksize = EVEN(node->mn_ksize);
6891 delta = ksize - oksize;
6893 /* Shift node contents if EVEN(key length) changed. */
6895 if (delta > 0 && SIZELEFT(mp) < delta) {
6897 /* not enough space left, do a delete and split */
6898 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6899 pgno = NODEPGNO(node);
6900 mdb_node_del(mc, 0);
6901 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6904 numkeys = NUMKEYS(mp);
6905 for (i = 0; i < numkeys; i++) {
6906 if (mp->mp_ptrs[i] <= ptr)
6907 mp->mp_ptrs[i] -= delta;
6910 base = (char *)mp + mp->mp_upper;
6911 len = ptr - mp->mp_upper + NODESIZE;
6912 memmove(base - delta, base, len);
6913 mp->mp_upper -= delta;
6915 node = NODEPTR(mp, indx);
6918 /* But even if no shift was needed, update ksize */
6919 if (node->mn_ksize != key->mv_size)
6920 node->mn_ksize = key->mv_size;
6923 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6929 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6931 /** Move a node from csrc to cdst.
6934 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6941 unsigned short flags;
6945 /* Mark src and dst as dirty. */
6946 if ((rc = mdb_page_touch(csrc)) ||
6947 (rc = mdb_page_touch(cdst)))
6950 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6951 key.mv_size = csrc->mc_db->md_pad;
6952 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6954 data.mv_data = NULL;
6958 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6959 mdb_cassert(csrc, !((size_t)srcnode & 1));
6960 srcpg = NODEPGNO(srcnode);
6961 flags = srcnode->mn_flags;
6962 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6963 unsigned int snum = csrc->mc_snum;
6965 /* must find the lowest key below src */
6966 mdb_page_search_lowest(csrc);
6967 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6968 key.mv_size = csrc->mc_db->md_pad;
6969 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6971 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6972 key.mv_size = NODEKSZ(s2);
6973 key.mv_data = NODEKEY(s2);
6975 csrc->mc_snum = snum--;
6976 csrc->mc_top = snum;
6978 key.mv_size = NODEKSZ(srcnode);
6979 key.mv_data = NODEKEY(srcnode);
6981 data.mv_size = NODEDSZ(srcnode);
6982 data.mv_data = NODEDATA(srcnode);
6984 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6985 unsigned int snum = cdst->mc_snum;
6988 /* must find the lowest key below dst */
6989 mdb_page_search_lowest(cdst);
6990 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6991 bkey.mv_size = cdst->mc_db->md_pad;
6992 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6994 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6995 bkey.mv_size = NODEKSZ(s2);
6996 bkey.mv_data = NODEKEY(s2);
6998 cdst->mc_snum = snum--;
6999 cdst->mc_top = snum;
7000 mdb_cursor_copy(cdst, &mn);
7002 rc = mdb_update_key(&mn, &bkey);
7007 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7008 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7009 csrc->mc_ki[csrc->mc_top],
7011 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7012 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7014 /* Add the node to the destination page.
7016 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7017 if (rc != MDB_SUCCESS)
7020 /* Delete the node from the source page.
7022 mdb_node_del(csrc, key.mv_size);
7025 /* Adjust other cursors pointing to mp */
7026 MDB_cursor *m2, *m3;
7027 MDB_dbi dbi = csrc->mc_dbi;
7028 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7030 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7031 if (csrc->mc_flags & C_SUB)
7032 m3 = &m2->mc_xcursor->mx_cursor;
7035 if (m3 == csrc) continue;
7036 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7037 csrc->mc_ki[csrc->mc_top]) {
7038 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7039 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7044 /* Update the parent separators.
7046 if (csrc->mc_ki[csrc->mc_top] == 0) {
7047 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7048 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7049 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7051 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7052 key.mv_size = NODEKSZ(srcnode);
7053 key.mv_data = NODEKEY(srcnode);
7055 DPRINTF(("update separator for source page %"Z"u to [%s]",
7056 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7057 mdb_cursor_copy(csrc, &mn);
7060 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7063 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7065 indx_t ix = csrc->mc_ki[csrc->mc_top];
7066 nullkey.mv_size = 0;
7067 csrc->mc_ki[csrc->mc_top] = 0;
7068 rc = mdb_update_key(csrc, &nullkey);
7069 csrc->mc_ki[csrc->mc_top] = ix;
7070 mdb_cassert(csrc, rc == MDB_SUCCESS);
7074 if (cdst->mc_ki[cdst->mc_top] == 0) {
7075 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7076 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7077 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7079 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7080 key.mv_size = NODEKSZ(srcnode);
7081 key.mv_data = NODEKEY(srcnode);
7083 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7084 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7085 mdb_cursor_copy(cdst, &mn);
7088 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7091 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7093 indx_t ix = cdst->mc_ki[cdst->mc_top];
7094 nullkey.mv_size = 0;
7095 cdst->mc_ki[cdst->mc_top] = 0;
7096 rc = mdb_update_key(cdst, &nullkey);
7097 cdst->mc_ki[cdst->mc_top] = ix;
7098 mdb_cassert(csrc, rc == MDB_SUCCESS);
7105 /** Merge one page into another.
7106 * The nodes from the page pointed to by \b csrc will
7107 * be copied to the page pointed to by \b cdst and then
7108 * the \b csrc page will be freed.
7109 * @param[in] csrc Cursor pointing to the source page.
7110 * @param[in] cdst Cursor pointing to the destination page.
7111 * @return 0 on success, non-zero on failure.
7114 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7122 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7123 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7125 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7126 mdb_cassert(csrc, cdst->mc_snum > 1);
7128 /* Mark dst as dirty. */
7129 if ((rc = mdb_page_touch(cdst)))
7132 /* Move all nodes from src to dst.
7134 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7135 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7136 key.mv_size = csrc->mc_db->md_pad;
7137 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7138 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7139 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7140 if (rc != MDB_SUCCESS)
7142 key.mv_data = (char *)key.mv_data + key.mv_size;
7145 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7146 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7147 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7148 unsigned int snum = csrc->mc_snum;
7150 /* must find the lowest key below src */
7151 mdb_page_search_lowest(csrc);
7152 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7153 key.mv_size = csrc->mc_db->md_pad;
7154 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7156 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7157 key.mv_size = NODEKSZ(s2);
7158 key.mv_data = NODEKEY(s2);
7160 csrc->mc_snum = snum--;
7161 csrc->mc_top = snum;
7163 key.mv_size = srcnode->mn_ksize;
7164 key.mv_data = NODEKEY(srcnode);
7167 data.mv_size = NODEDSZ(srcnode);
7168 data.mv_data = NODEDATA(srcnode);
7169 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7170 if (rc != MDB_SUCCESS)
7175 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7176 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7177 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7179 /* Unlink the src page from parent and add to free list.
7182 mdb_node_del(csrc, 0);
7183 if (csrc->mc_ki[csrc->mc_top] == 0) {
7185 rc = mdb_update_key(csrc, &key);
7193 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7194 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7197 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7198 csrc->mc_db->md_leaf_pages--;
7200 csrc->mc_db->md_branch_pages--;
7202 /* Adjust other cursors pointing to mp */
7203 MDB_cursor *m2, *m3;
7204 MDB_dbi dbi = csrc->mc_dbi;
7205 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7207 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7208 if (csrc->mc_flags & C_SUB)
7209 m3 = &m2->mc_xcursor->mx_cursor;
7212 if (m3 == csrc) continue;
7213 if (m3->mc_snum < csrc->mc_snum) continue;
7214 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7215 m3->mc_pg[csrc->mc_top] = mp;
7216 m3->mc_ki[csrc->mc_top] += nkeys;
7220 mdb_cursor_pop(csrc);
7222 return mdb_rebalance(csrc);
7225 /** Copy the contents of a cursor.
7226 * @param[in] csrc The cursor to copy from.
7227 * @param[out] cdst The cursor to copy to.
7230 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7234 cdst->mc_txn = csrc->mc_txn;
7235 cdst->mc_dbi = csrc->mc_dbi;
7236 cdst->mc_db = csrc->mc_db;
7237 cdst->mc_dbx = csrc->mc_dbx;
7238 cdst->mc_snum = csrc->mc_snum;
7239 cdst->mc_top = csrc->mc_top;
7240 cdst->mc_flags = csrc->mc_flags;
7242 for (i=0; i<csrc->mc_snum; i++) {
7243 cdst->mc_pg[i] = csrc->mc_pg[i];
7244 cdst->mc_ki[i] = csrc->mc_ki[i];
7248 /** Rebalance the tree after a delete operation.
7249 * @param[in] mc Cursor pointing to the page where rebalancing
7251 * @return 0 on success, non-zero on failure.
7254 mdb_rebalance(MDB_cursor *mc)
7258 unsigned int ptop, minkeys;
7261 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7262 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7263 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7264 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7265 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7267 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7268 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7269 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7270 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7274 if (mc->mc_snum < 2) {
7275 MDB_page *mp = mc->mc_pg[0];
7277 DPUTS("Can't rebalance a subpage, ignoring");
7280 if (NUMKEYS(mp) == 0) {
7281 DPUTS("tree is completely empty");
7282 mc->mc_db->md_root = P_INVALID;
7283 mc->mc_db->md_depth = 0;
7284 mc->mc_db->md_leaf_pages = 0;
7285 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7288 /* Adjust cursors pointing to mp */
7291 mc->mc_flags &= ~C_INITIALIZED;
7293 MDB_cursor *m2, *m3;
7294 MDB_dbi dbi = mc->mc_dbi;
7296 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7297 if (mc->mc_flags & C_SUB)
7298 m3 = &m2->mc_xcursor->mx_cursor;
7301 if (m3->mc_snum < mc->mc_snum) continue;
7302 if (m3->mc_pg[0] == mp) {
7305 m3->mc_flags &= ~C_INITIALIZED;
7309 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7310 DPUTS("collapsing root page!");
7311 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7314 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7315 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7318 mc->mc_db->md_depth--;
7319 mc->mc_db->md_branch_pages--;
7320 mc->mc_ki[0] = mc->mc_ki[1];
7322 /* Adjust other cursors pointing to mp */
7323 MDB_cursor *m2, *m3;
7324 MDB_dbi dbi = mc->mc_dbi;
7326 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7327 if (mc->mc_flags & C_SUB)
7328 m3 = &m2->mc_xcursor->mx_cursor;
7331 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7332 if (m3->mc_pg[0] == mp) {
7336 for (i=0; i<m3->mc_snum; i++) {
7337 m3->mc_pg[i] = m3->mc_pg[i+1];
7338 m3->mc_ki[i] = m3->mc_ki[i+1];
7344 DPUTS("root page doesn't need rebalancing");
7348 /* The parent (branch page) must have at least 2 pointers,
7349 * otherwise the tree is invalid.
7351 ptop = mc->mc_top-1;
7352 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7354 /* Leaf page fill factor is below the threshold.
7355 * Try to move keys from left or right neighbor, or
7356 * merge with a neighbor page.
7361 mdb_cursor_copy(mc, &mn);
7362 mn.mc_xcursor = NULL;
7364 if (mc->mc_ki[ptop] == 0) {
7365 /* We're the leftmost leaf in our parent.
7367 DPUTS("reading right neighbor");
7369 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7370 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7373 mn.mc_ki[mn.mc_top] = 0;
7374 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7376 /* There is at least one neighbor to the left.
7378 DPUTS("reading left neighbor");
7380 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7381 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7384 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7385 mc->mc_ki[mc->mc_top] = 0;
7388 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7389 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7390 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7392 /* If the neighbor page is above threshold and has enough keys,
7393 * move one key from it. Otherwise we should try to merge them.
7394 * (A branch page must never have less than 2 keys.)
7396 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7397 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7398 return mdb_node_move(&mn, mc);
7400 if (mc->mc_ki[ptop] == 0)
7401 rc = mdb_page_merge(&mn, mc);
7403 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7404 rc = mdb_page_merge(mc, &mn);
7405 mdb_cursor_copy(&mn, mc);
7407 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7412 /** Complete a delete operation started by #mdb_cursor_del(). */
7414 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7421 mp = mc->mc_pg[mc->mc_top];
7422 ki = mc->mc_ki[mc->mc_top];
7424 /* add overflow pages to free list */
7425 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7429 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7430 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7431 (rc = mdb_ovpage_free(mc, omp)))
7434 mdb_node_del(mc, mc->mc_db->md_pad);
7435 mc->mc_db->md_entries--;
7436 rc = mdb_rebalance(mc);
7437 if (rc != MDB_SUCCESS)
7438 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7440 MDB_cursor *m2, *m3;
7441 MDB_dbi dbi = mc->mc_dbi;
7443 mp = mc->mc_pg[mc->mc_top];
7444 nkeys = NUMKEYS(mp);
7446 /* if mc points past last node in page, find next sibling */
7447 if (mc->mc_ki[mc->mc_top] >= nkeys)
7448 mdb_cursor_sibling(mc, 1);
7450 /* Adjust other cursors pointing to mp */
7451 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7452 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7453 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7455 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7457 if (m3->mc_pg[mc->mc_top] == mp) {
7458 if (m3->mc_ki[mc->mc_top] >= ki) {
7459 m3->mc_flags |= C_DEL;
7460 if (m3->mc_ki[mc->mc_top] > ki)
7461 m3->mc_ki[mc->mc_top]--;
7463 if (m3->mc_ki[mc->mc_top] >= nkeys)
7464 mdb_cursor_sibling(m3, 1);
7467 mc->mc_flags |= C_DEL;
7474 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7475 MDB_val *key, MDB_val *data)
7480 MDB_val rdata, *xdata;
7487 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7489 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7492 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7493 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7495 mdb_cursor_init(&mc, txn, dbi, &mx);
7498 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7499 /* must ignore any data */
7510 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7512 /* let mdb_page_split know about this cursor if needed:
7513 * delete will trigger a rebalance; if it needs to move
7514 * a node from one page to another, it will have to
7515 * update the parent's separator key(s). If the new sepkey
7516 * is larger than the current one, the parent page may
7517 * run out of space, triggering a split. We need this
7518 * cursor to be consistent until the end of the rebalance.
7520 mc.mc_flags |= C_UNTRACK;
7521 mc.mc_next = txn->mt_cursors[dbi];
7522 txn->mt_cursors[dbi] = &mc;
7523 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7524 txn->mt_cursors[dbi] = mc.mc_next;
7529 /** Split a page and insert a new node.
7530 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7531 * The cursor will be updated to point to the actual page and index where
7532 * the node got inserted after the split.
7533 * @param[in] newkey The key for the newly inserted node.
7534 * @param[in] newdata The data for the newly inserted node.
7535 * @param[in] newpgno The page number, if the new node is a branch node.
7536 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7537 * @return 0 on success, non-zero on failure.
7540 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7541 unsigned int nflags)
7544 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7547 int i, j, split_indx, nkeys, pmax;
7548 MDB_env *env = mc->mc_txn->mt_env;
7550 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7551 MDB_page *copy = NULL;
7552 MDB_page *mp, *rp, *pp;
7557 mp = mc->mc_pg[mc->mc_top];
7558 newindx = mc->mc_ki[mc->mc_top];
7559 nkeys = NUMKEYS(mp);
7561 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7562 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7563 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7565 /* Create a right sibling. */
7566 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7568 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7570 if (mc->mc_snum < 2) {
7571 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7573 /* shift current top to make room for new parent */
7574 mc->mc_pg[1] = mc->mc_pg[0];
7575 mc->mc_ki[1] = mc->mc_ki[0];
7578 mc->mc_db->md_root = pp->mp_pgno;
7579 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7580 mc->mc_db->md_depth++;
7583 /* Add left (implicit) pointer. */
7584 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7585 /* undo the pre-push */
7586 mc->mc_pg[0] = mc->mc_pg[1];
7587 mc->mc_ki[0] = mc->mc_ki[1];
7588 mc->mc_db->md_root = mp->mp_pgno;
7589 mc->mc_db->md_depth--;
7596 ptop = mc->mc_top-1;
7597 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7600 mc->mc_flags |= C_SPLITTING;
7601 mdb_cursor_copy(mc, &mn);
7602 mn.mc_pg[mn.mc_top] = rp;
7603 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7605 if (nflags & MDB_APPEND) {
7606 mn.mc_ki[mn.mc_top] = 0;
7608 split_indx = newindx;
7612 split_indx = (nkeys+1) / 2;
7617 unsigned int lsize, rsize, ksize;
7618 /* Move half of the keys to the right sibling */
7620 x = mc->mc_ki[mc->mc_top] - split_indx;
7621 ksize = mc->mc_db->md_pad;
7622 split = LEAF2KEY(mp, split_indx, ksize);
7623 rsize = (nkeys - split_indx) * ksize;
7624 lsize = (nkeys - split_indx) * sizeof(indx_t);
7625 mp->mp_lower -= lsize;
7626 rp->mp_lower += lsize;
7627 mp->mp_upper += rsize - lsize;
7628 rp->mp_upper -= rsize - lsize;
7629 sepkey.mv_size = ksize;
7630 if (newindx == split_indx) {
7631 sepkey.mv_data = newkey->mv_data;
7633 sepkey.mv_data = split;
7636 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7637 memcpy(rp->mp_ptrs, split, rsize);
7638 sepkey.mv_data = rp->mp_ptrs;
7639 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7640 memcpy(ins, newkey->mv_data, ksize);
7641 mp->mp_lower += sizeof(indx_t);
7642 mp->mp_upper -= ksize - sizeof(indx_t);
7645 memcpy(rp->mp_ptrs, split, x * ksize);
7646 ins = LEAF2KEY(rp, x, ksize);
7647 memcpy(ins, newkey->mv_data, ksize);
7648 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7649 rp->mp_lower += sizeof(indx_t);
7650 rp->mp_upper -= ksize - sizeof(indx_t);
7651 mc->mc_ki[mc->mc_top] = x;
7652 mc->mc_pg[mc->mc_top] = rp;
7655 int psize, nsize, k;
7656 /* Maximum free space in an empty page */
7657 pmax = env->me_psize - PAGEHDRSZ;
7659 nsize = mdb_leaf_size(env, newkey, newdata);
7661 nsize = mdb_branch_size(env, newkey);
7662 nsize = EVEN(nsize);
7664 /* grab a page to hold a temporary copy */
7665 copy = mdb_page_malloc(mc->mc_txn, 1);
7668 copy->mp_pgno = mp->mp_pgno;
7669 copy->mp_flags = mp->mp_flags;
7670 copy->mp_lower = PAGEHDRSZ;
7671 copy->mp_upper = env->me_psize;
7673 /* prepare to insert */
7674 for (i=0, j=0; i<nkeys; i++) {
7676 copy->mp_ptrs[j++] = 0;
7678 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7681 /* When items are relatively large the split point needs
7682 * to be checked, because being off-by-one will make the
7683 * difference between success or failure in mdb_node_add.
7685 * It's also relevant if a page happens to be laid out
7686 * such that one half of its nodes are all "small" and
7687 * the other half of its nodes are "large." If the new
7688 * item is also "large" and falls on the half with
7689 * "large" nodes, it also may not fit.
7691 * As a final tweak, if the new item goes on the last
7692 * spot on the page (and thus, onto the new page), bias
7693 * the split so the new page is emptier than the old page.
7694 * This yields better packing during sequential inserts.
7696 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7697 /* Find split point */
7699 if (newindx <= split_indx || newindx >= nkeys) {
7701 k = newindx >= nkeys ? nkeys : split_indx+2;
7706 for (; i!=k; i+=j) {
7711 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7712 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7714 if (F_ISSET(node->mn_flags, F_BIGDATA))
7715 psize += sizeof(pgno_t);
7717 psize += NODEDSZ(node);
7719 psize = EVEN(psize);
7721 if (psize > pmax || i == k-j) {
7722 split_indx = i + (j<0);
7727 if (split_indx == newindx) {
7728 sepkey.mv_size = newkey->mv_size;
7729 sepkey.mv_data = newkey->mv_data;
7731 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7732 sepkey.mv_size = node->mn_ksize;
7733 sepkey.mv_data = NODEKEY(node);
7738 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7740 /* Copy separator key to the parent.
7742 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7746 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7749 if (mn.mc_snum == mc->mc_snum) {
7750 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7751 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7752 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7753 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7758 /* Right page might now have changed parent.
7759 * Check if left page also changed parent.
7761 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7762 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7763 for (i=0; i<ptop; i++) {
7764 mc->mc_pg[i] = mn.mc_pg[i];
7765 mc->mc_ki[i] = mn.mc_ki[i];
7767 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7768 if (mn.mc_ki[ptop]) {
7769 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7771 /* find right page's left sibling */
7772 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7773 mdb_cursor_sibling(mc, 0);
7778 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7781 mc->mc_flags ^= C_SPLITTING;
7782 if (rc != MDB_SUCCESS) {
7785 if (nflags & MDB_APPEND) {
7786 mc->mc_pg[mc->mc_top] = rp;
7787 mc->mc_ki[mc->mc_top] = 0;
7788 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7791 for (i=0; i<mc->mc_top; i++)
7792 mc->mc_ki[i] = mn.mc_ki[i];
7793 } else if (!IS_LEAF2(mp)) {
7795 mc->mc_pg[mc->mc_top] = rp;
7800 rkey.mv_data = newkey->mv_data;
7801 rkey.mv_size = newkey->mv_size;
7807 /* Update index for the new key. */
7808 mc->mc_ki[mc->mc_top] = j;
7810 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7811 rkey.mv_data = NODEKEY(node);
7812 rkey.mv_size = node->mn_ksize;
7814 xdata.mv_data = NODEDATA(node);
7815 xdata.mv_size = NODEDSZ(node);
7818 pgno = NODEPGNO(node);
7819 flags = node->mn_flags;
7822 if (!IS_LEAF(mp) && j == 0) {
7823 /* First branch index doesn't need key data. */
7827 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7829 /* return tmp page to freelist */
7830 mdb_page_free(env, copy);
7836 mc->mc_pg[mc->mc_top] = copy;
7841 } while (i != split_indx);
7843 nkeys = NUMKEYS(copy);
7844 for (i=0; i<nkeys; i++)
7845 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7846 mp->mp_lower = copy->mp_lower;
7847 mp->mp_upper = copy->mp_upper;
7848 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7849 env->me_psize - copy->mp_upper);
7851 /* reset back to original page */
7852 if (newindx < split_indx) {
7853 mc->mc_pg[mc->mc_top] = mp;
7854 if (nflags & MDB_RESERVE) {
7855 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7856 if (!(node->mn_flags & F_BIGDATA))
7857 newdata->mv_data = NODEDATA(node);
7860 mc->mc_pg[mc->mc_top] = rp;
7862 /* Make sure mc_ki is still valid.
7864 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7865 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7866 for (i=0; i<=ptop; i++) {
7867 mc->mc_pg[i] = mn.mc_pg[i];
7868 mc->mc_ki[i] = mn.mc_ki[i];
7872 /* return tmp page to freelist */
7873 mdb_page_free(env, copy);
7877 /* Adjust other cursors pointing to mp */
7878 MDB_cursor *m2, *m3;
7879 MDB_dbi dbi = mc->mc_dbi;
7880 int fixup = NUMKEYS(mp);
7882 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7883 if (mc->mc_flags & C_SUB)
7884 m3 = &m2->mc_xcursor->mx_cursor;
7889 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7891 if (m3->mc_flags & C_SPLITTING)
7896 for (k=m3->mc_top; k>=0; k--) {
7897 m3->mc_ki[k+1] = m3->mc_ki[k];
7898 m3->mc_pg[k+1] = m3->mc_pg[k];
7900 if (m3->mc_ki[0] >= split_indx) {
7905 m3->mc_pg[0] = mc->mc_pg[0];
7909 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7910 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7911 m3->mc_ki[mc->mc_top]++;
7912 if (m3->mc_ki[mc->mc_top] >= fixup) {
7913 m3->mc_pg[mc->mc_top] = rp;
7914 m3->mc_ki[mc->mc_top] -= fixup;
7915 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7917 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7918 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7923 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
7928 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7929 MDB_val *key, MDB_val *data, unsigned int flags)
7934 if (key == NULL || data == NULL)
7937 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7940 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7943 mdb_cursor_init(&mc, txn, dbi, &mx);
7944 return mdb_cursor_put(&mc, key, data, flags);
7948 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7950 if ((flag & CHANGEABLE) != flag)
7953 env->me_flags |= flag;
7955 env->me_flags &= ~flag;
7960 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7965 *arg = env->me_flags;
7970 mdb_env_set_userctx(MDB_env *env, void *ctx)
7974 env->me_userctx = ctx;
7979 mdb_env_get_userctx(MDB_env *env)
7981 return env ? env->me_userctx : NULL;
7985 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
7990 env->me_assert_func = func;
7996 mdb_env_get_path(MDB_env *env, const char **arg)
8001 *arg = env->me_path;
8006 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8015 /** Common code for #mdb_stat() and #mdb_env_stat().
8016 * @param[in] env the environment to operate in.
8017 * @param[in] db the #MDB_db record containing the stats to return.
8018 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8019 * @return 0, this function always succeeds.
8022 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8024 arg->ms_psize = env->me_psize;
8025 arg->ms_depth = db->md_depth;
8026 arg->ms_branch_pages = db->md_branch_pages;
8027 arg->ms_leaf_pages = db->md_leaf_pages;
8028 arg->ms_overflow_pages = db->md_overflow_pages;
8029 arg->ms_entries = db->md_entries;
8034 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8038 if (env == NULL || arg == NULL)
8041 toggle = mdb_env_pick_meta(env);
8043 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8047 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8051 if (env == NULL || arg == NULL)
8054 toggle = mdb_env_pick_meta(env);
8055 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
8056 arg->me_mapsize = env->me_mapsize;
8057 arg->me_maxreaders = env->me_maxreaders;
8059 /* me_numreaders may be zero if this process never used any readers. Use
8060 * the shared numreader count if it exists.
8062 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8064 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8065 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8069 /** Set the default comparison functions for a database.
8070 * Called immediately after a database is opened to set the defaults.
8071 * The user can then override them with #mdb_set_compare() or
8072 * #mdb_set_dupsort().
8073 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8074 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8077 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8079 uint16_t f = txn->mt_dbs[dbi].md_flags;
8081 txn->mt_dbxs[dbi].md_cmp =
8082 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8083 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8085 txn->mt_dbxs[dbi].md_dcmp =
8086 !(f & MDB_DUPSORT) ? 0 :
8087 ((f & MDB_INTEGERDUP)
8088 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8089 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8092 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8097 int rc, dbflag, exact;
8098 unsigned int unused = 0;
8101 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8102 mdb_default_cmp(txn, FREE_DBI);
8105 if ((flags & VALID_FLAGS) != flags)
8107 if (txn->mt_flags & MDB_TXN_ERROR)
8113 if (flags & PERSISTENT_FLAGS) {
8114 uint16_t f2 = flags & PERSISTENT_FLAGS;
8115 /* make sure flag changes get committed */
8116 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8117 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8118 txn->mt_flags |= MDB_TXN_DIRTY;
8121 mdb_default_cmp(txn, MAIN_DBI);
8125 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8126 mdb_default_cmp(txn, MAIN_DBI);
8129 /* Is the DB already open? */
8131 for (i=2; i<txn->mt_numdbs; i++) {
8132 if (!txn->mt_dbxs[i].md_name.mv_size) {
8133 /* Remember this free slot */
8134 if (!unused) unused = i;
8137 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8138 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8144 /* If no free slot and max hit, fail */
8145 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8146 return MDB_DBS_FULL;
8148 /* Cannot mix named databases with some mainDB flags */
8149 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8150 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8152 /* Find the DB info */
8153 dbflag = DB_NEW|DB_VALID;
8156 key.mv_data = (void *)name;
8157 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8158 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8159 if (rc == MDB_SUCCESS) {
8160 /* make sure this is actually a DB */
8161 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8162 if (!(node->mn_flags & F_SUBDATA))
8163 return MDB_INCOMPATIBLE;
8164 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8165 /* Create if requested */
8167 data.mv_size = sizeof(MDB_db);
8168 data.mv_data = &dummy;
8169 memset(&dummy, 0, sizeof(dummy));
8170 dummy.md_root = P_INVALID;
8171 dummy.md_flags = flags & PERSISTENT_FLAGS;
8172 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8176 /* OK, got info, add to table */
8177 if (rc == MDB_SUCCESS) {
8178 unsigned int slot = unused ? unused : txn->mt_numdbs;
8179 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8180 txn->mt_dbxs[slot].md_name.mv_size = len;
8181 txn->mt_dbxs[slot].md_rel = NULL;
8182 txn->mt_dbflags[slot] = dbflag;
8183 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8185 mdb_default_cmp(txn, slot);
8194 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8196 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8199 if (txn->mt_dbflags[dbi] & DB_STALE) {
8202 /* Stale, must read the DB's root. cursor_init does it for us. */
8203 mdb_cursor_init(&mc, txn, dbi, &mx);
8205 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8208 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8211 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8213 ptr = env->me_dbxs[dbi].md_name.mv_data;
8214 env->me_dbxs[dbi].md_name.mv_data = NULL;
8215 env->me_dbxs[dbi].md_name.mv_size = 0;
8216 env->me_dbflags[dbi] = 0;
8220 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8222 /* We could return the flags for the FREE_DBI too but what's the point? */
8223 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8225 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8229 /** Add all the DB's pages to the free list.
8230 * @param[in] mc Cursor on the DB to free.
8231 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8232 * @return 0 on success, non-zero on failure.
8235 mdb_drop0(MDB_cursor *mc, int subs)
8239 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8240 if (rc == MDB_SUCCESS) {
8241 MDB_txn *txn = mc->mc_txn;
8246 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8247 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8250 mdb_cursor_copy(mc, &mx);
8251 while (mc->mc_snum > 0) {
8252 MDB_page *mp = mc->mc_pg[mc->mc_top];
8253 unsigned n = NUMKEYS(mp);
8255 for (i=0; i<n; i++) {
8256 ni = NODEPTR(mp, i);
8257 if (ni->mn_flags & F_BIGDATA) {
8260 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8261 rc = mdb_page_get(txn, pg, &omp, NULL);
8264 mdb_cassert(mc, IS_OVERFLOW(omp));
8265 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8269 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8270 mdb_xcursor_init1(mc, ni);
8271 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8277 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8279 for (i=0; i<n; i++) {
8281 ni = NODEPTR(mp, i);
8284 mdb_midl_xappend(txn->mt_free_pgs, pg);
8289 mc->mc_ki[mc->mc_top] = i;
8290 rc = mdb_cursor_sibling(mc, 1);
8292 /* no more siblings, go back to beginning
8293 * of previous level.
8297 for (i=1; i<mc->mc_snum; i++) {
8299 mc->mc_pg[i] = mx.mc_pg[i];
8304 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8305 } else if (rc == MDB_NOTFOUND) {
8311 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8313 MDB_cursor *mc, *m2;
8316 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8319 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8322 rc = mdb_cursor_open(txn, dbi, &mc);
8326 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8327 /* Invalidate the dropped DB's cursors */
8328 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8329 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8333 /* Can't delete the main DB */
8334 if (del && dbi > MAIN_DBI) {
8335 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8337 txn->mt_dbflags[dbi] = DB_STALE;
8338 mdb_dbi_close(txn->mt_env, dbi);
8341 /* reset the DB record, mark it dirty */
8342 txn->mt_dbflags[dbi] |= DB_DIRTY;
8343 txn->mt_dbs[dbi].md_depth = 0;
8344 txn->mt_dbs[dbi].md_branch_pages = 0;
8345 txn->mt_dbs[dbi].md_leaf_pages = 0;
8346 txn->mt_dbs[dbi].md_overflow_pages = 0;
8347 txn->mt_dbs[dbi].md_entries = 0;
8348 txn->mt_dbs[dbi].md_root = P_INVALID;
8350 txn->mt_flags |= MDB_TXN_DIRTY;
8353 mdb_cursor_close(mc);
8357 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8359 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8362 txn->mt_dbxs[dbi].md_cmp = cmp;
8366 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8368 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8371 txn->mt_dbxs[dbi].md_dcmp = cmp;
8375 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8377 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8380 txn->mt_dbxs[dbi].md_rel = rel;
8384 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8386 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8389 txn->mt_dbxs[dbi].md_relctx = ctx;
8393 int mdb_env_get_maxkeysize(MDB_env *env)
8395 return ENV_MAXKEY(env);
8398 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8400 unsigned int i, rdrs;
8403 int rc = 0, first = 1;
8407 if (!env->me_txns) {
8408 return func("(no reader locks)\n", ctx);
8410 rdrs = env->me_txns->mti_numreaders;
8411 mr = env->me_txns->mti_readers;
8412 for (i=0; i<rdrs; i++) {
8414 txnid_t txnid = mr[i].mr_txnid;
8415 sprintf(buf, txnid == (txnid_t)-1 ?
8416 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
8417 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
8420 rc = func(" pid thread txnid\n", ctx);
8424 rc = func(buf, ctx);
8430 rc = func("(no active readers)\n", ctx);
8435 /** Insert pid into list if not already present.
8436 * return -1 if already present.
8438 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8440 /* binary search of pid in list */
8442 unsigned cursor = 1;
8444 unsigned n = ids[0];
8447 unsigned pivot = n >> 1;
8448 cursor = base + pivot + 1;
8449 val = pid - ids[cursor];
8454 } else if ( val > 0 ) {
8459 /* found, so it's a duplicate */
8468 for (n = ids[0]; n > cursor; n--)
8474 int mdb_reader_check(MDB_env *env, int *dead)
8476 unsigned int i, j, rdrs;
8478 MDB_PID_T *pids, pid;
8487 rdrs = env->me_txns->mti_numreaders;
8488 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8492 mr = env->me_txns->mti_readers;
8493 for (i=0; i<rdrs; i++) {
8494 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8496 if (mdb_pid_insert(pids, pid) == 0) {
8497 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8499 /* Recheck, a new process may have reused pid */
8500 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8501 for (j=i; j<rdrs; j++)
8502 if (mr[j].mr_pid == pid) {
8503 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8504 (unsigned) pid, mr[j].mr_txnid));
8509 UNLOCK_MUTEX_R(env);
projects / openldap / blob