2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
42 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
43 * as int64 which is wrong. MSVC doesn't define it at all, so just
48 # include <sys/param.h>
50 # define LITTLE_ENDIAN 1234
51 # define BIG_ENDIAN 4321
52 # define BYTE_ORDER LITTLE_ENDIAN
54 # define SSIZE_MAX INT_MAX
58 #define MDB_PID_T pid_t
59 #include <sys/param.h>
62 #ifdef HAVE_SYS_FILE_H
78 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
79 #include <netinet/in.h>
80 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
83 #if defined(__APPLE__) || defined (BSD)
84 # define MDB_USE_POSIX_SEM 1
85 # define MDB_FDATASYNC fsync
86 #elif defined(ANDROID)
87 # define MDB_FDATASYNC fsync
92 #ifdef MDB_USE_POSIX_SEM
93 # define MDB_USE_HASH 1
94 #include <semaphore.h>
99 #include <valgrind/memcheck.h>
100 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
101 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
102 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
103 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
104 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
106 #define VGMEMP_CREATE(h,r,z)
107 #define VGMEMP_ALLOC(h,a,s)
108 #define VGMEMP_FREE(h,a)
109 #define VGMEMP_DESTROY(h)
110 #define VGMEMP_DEFINED(a,s)
114 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
115 /* Solaris just defines one or the other */
116 # define LITTLE_ENDIAN 1234
117 # define BIG_ENDIAN 4321
118 # ifdef _LITTLE_ENDIAN
119 # define BYTE_ORDER LITTLE_ENDIAN
121 # define BYTE_ORDER BIG_ENDIAN
124 # define BYTE_ORDER __BYTE_ORDER
128 #ifndef LITTLE_ENDIAN
129 #define LITTLE_ENDIAN __LITTLE_ENDIAN
132 #define BIG_ENDIAN __BIG_ENDIAN
135 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
136 #define MISALIGNED_OK 1
142 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
143 # error "Unknown or unsupported endianness (BYTE_ORDER)"
144 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
145 # error "Two's complement, reasonably sized integer types, please"
148 /** @defgroup internal MDB Internals
151 /** @defgroup compat Compatibility Macros
152 * A bunch of macros to minimize the amount of platform-specific ifdefs
153 * needed throughout the rest of the code. When the features this library
154 * needs are similar enough to POSIX to be hidden in a one-or-two line
155 * replacement, this macro approach is used.
159 /** Wrapper around __func__, which is a C99 feature */
160 #if __STDC_VERSION__ >= 199901L
161 # define mdb_func_ __func__
162 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
163 # define mdb_func_ __FUNCTION__
165 /* If a debug message says <mdb_unknown>(), update the #if statements above */
166 # define mdb_func_ "<mdb_unknown>"
170 #define MDB_USE_HASH 1
171 #define MDB_PIDLOCK 0
172 #define pthread_t DWORD
173 #define pthread_mutex_t HANDLE
174 #define pthread_key_t DWORD
175 #define pthread_self() GetCurrentThreadId()
176 #define pthread_key_create(x,y) \
177 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
178 #define pthread_key_delete(x) TlsFree(x)
179 #define pthread_getspecific(x) TlsGetValue(x)
180 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
181 #define pthread_mutex_unlock(x) ReleaseMutex(x)
182 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
183 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
184 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
185 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
186 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
187 #define getpid() GetCurrentProcessId()
188 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
189 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
190 #define ErrCode() GetLastError()
191 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
192 #define close(fd) (CloseHandle(fd) ? 0 : -1)
193 #define munmap(ptr,len) UnmapViewOfFile(ptr)
194 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
195 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
197 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
202 #define Z "z" /**< printf format modifier for size_t */
204 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
205 #define MDB_PIDLOCK 1
207 #ifdef MDB_USE_POSIX_SEM
209 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
210 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
211 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
212 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
215 mdb_sem_wait(sem_t *sem)
218 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
223 /** Lock the reader mutex.
225 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
226 /** Unlock the reader mutex.
228 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
230 /** Lock the writer mutex.
231 * Only a single write transaction is allowed at a time. Other writers
232 * will block waiting for this mutex.
234 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
235 /** Unlock the writer mutex.
237 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
238 #endif /* MDB_USE_POSIX_SEM */
240 /** Get the error code for the last failed system function.
242 #define ErrCode() errno
244 /** An abstraction for a file handle.
245 * On POSIX systems file handles are small integers. On Windows
246 * they're opaque pointers.
250 /** A value for an invalid file handle.
251 * Mainly used to initialize file variables and signify that they are
254 #define INVALID_HANDLE_VALUE (-1)
256 /** Get the size of a memory page for the system.
257 * This is the basic size that the platform's memory manager uses, and is
258 * fundamental to the use of memory-mapped files.
260 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
263 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
266 #define MNAME_LEN (sizeof(pthread_mutex_t))
272 /** A flag for opening a file and requesting synchronous data writes.
273 * This is only used when writing a meta page. It's not strictly needed;
274 * we could just do a normal write and then immediately perform a flush.
275 * But if this flag is available it saves us an extra system call.
277 * @note If O_DSYNC is undefined but exists in /usr/include,
278 * preferably set some compiler flag to get the definition.
279 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
282 # define MDB_DSYNC O_DSYNC
286 /** Function for flushing the data of a file. Define this to fsync
287 * if fdatasync() is not supported.
289 #ifndef MDB_FDATASYNC
290 # define MDB_FDATASYNC fdatasync
294 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
305 /** A page number in the database.
306 * Note that 64 bit page numbers are overkill, since pages themselves
307 * already represent 12-13 bits of addressable memory, and the OS will
308 * always limit applications to a maximum of 63 bits of address space.
310 * @note In the #MDB_node structure, we only store 48 bits of this value,
311 * which thus limits us to only 60 bits of addressable data.
313 typedef MDB_ID pgno_t;
315 /** A transaction ID.
316 * See struct MDB_txn.mt_txnid for details.
318 typedef MDB_ID txnid_t;
320 /** @defgroup debug Debug Macros
324 /** Enable debug output. Needs variable argument macros (a C99 feature).
325 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
326 * read from and written to the database (used for free space management).
332 static int mdb_debug;
333 static txnid_t mdb_debug_start;
335 /** Print a debug message with printf formatting.
336 * Requires double parenthesis around 2 or more args.
338 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
339 # define DPRINTF0(fmt, ...) \
340 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
342 # define DPRINTF(args) ((void) 0)
344 /** Print a debug string.
345 * The string is printed literally, with no format processing.
347 #define DPUTS(arg) DPRINTF(("%s", arg))
348 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
350 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
353 /** @brief The maximum size of a database page.
355 * This is 32k, since it must fit in #MDB_page.%mp_upper.
357 * LMDB will use database pages < OS pages if needed.
358 * That causes more I/O in write transactions: The OS must
359 * know (read) the whole page before writing a partial page.
361 * Note that we don't currently support Huge pages. On Linux,
362 * regular data files cannot use Huge pages, and in general
363 * Huge pages aren't actually pageable. We rely on the OS
364 * demand-pager to read our data and page it out when memory
365 * pressure from other processes is high. So until OSs have
366 * actual paging support for Huge pages, they're not viable.
368 #define MAX_PAGESIZE 0x8000
370 /** The minimum number of keys required in a database page.
371 * Setting this to a larger value will place a smaller bound on the
372 * maximum size of a data item. Data items larger than this size will
373 * be pushed into overflow pages instead of being stored directly in
374 * the B-tree node. This value used to default to 4. With a page size
375 * of 4096 bytes that meant that any item larger than 1024 bytes would
376 * go into an overflow page. That also meant that on average 2-3KB of
377 * each overflow page was wasted space. The value cannot be lower than
378 * 2 because then there would no longer be a tree structure. With this
379 * value, items larger than 2KB will go into overflow pages, and on
380 * average only 1KB will be wasted.
382 #define MDB_MINKEYS 2
384 /** A stamp that identifies a file as an MDB file.
385 * There's nothing special about this value other than that it is easily
386 * recognizable, and it will reflect any byte order mismatches.
388 #define MDB_MAGIC 0xBEEFC0DE
390 /** The version number for a database's datafile format. */
391 #define MDB_DATA_VERSION 1
392 /** The version number for a database's lockfile format. */
393 #define MDB_LOCK_VERSION 1
395 /** @brief The max size of a key we can write, or 0 for dynamic max.
397 * Define this as 0 to compute the max from the page size. 511
398 * is default for backwards compat: liblmdb <= 0.9.10 can break
399 * when modifying a DB with keys/dupsort data bigger than its max.
401 * Data items in an #MDB_DUPSORT database are also limited to
402 * this size, since they're actually keys of a sub-DB. Keys and
403 * #MDB_DUPSORT data items must fit on a node in a regular page.
405 #ifndef MDB_MAXKEYSIZE
406 #define MDB_MAXKEYSIZE 511
409 /** The maximum size of a key we can write to the environment. */
411 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
413 #define ENV_MAXKEY(env) ((env)->me_maxkey)
416 /** @brief The maximum size of a data item.
418 * We only store a 32 bit value for node sizes.
420 #define MAXDATASIZE 0xffffffffUL
423 /** Key size which fits in a #DKBUF.
426 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
429 * This is used for printing a hex dump of a key's contents.
431 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
432 /** Display a key in hex.
434 * Invoke a function to display a key in hex.
436 #define DKEY(x) mdb_dkey(x, kbuf)
442 /** An invalid page number.
443 * Mainly used to denote an empty tree.
445 #define P_INVALID (~(pgno_t)0)
447 /** Test if the flags \b f are set in a flag word \b w. */
448 #define F_ISSET(w, f) (((w) & (f)) == (f))
450 /** Round \b n up to an even number. */
451 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
453 /** Used for offsets within a single page.
454 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
457 typedef uint16_t indx_t;
459 /** Default size of memory map.
460 * This is certainly too small for any actual applications. Apps should always set
461 * the size explicitly using #mdb_env_set_mapsize().
463 #define DEFAULT_MAPSIZE 1048576
465 /** @defgroup readers Reader Lock Table
466 * Readers don't acquire any locks for their data access. Instead, they
467 * simply record their transaction ID in the reader table. The reader
468 * mutex is needed just to find an empty slot in the reader table. The
469 * slot's address is saved in thread-specific data so that subsequent read
470 * transactions started by the same thread need no further locking to proceed.
472 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
474 * No reader table is used if the database is on a read-only filesystem, or
475 * if #MDB_NOLOCK is set.
477 * Since the database uses multi-version concurrency control, readers don't
478 * actually need any locking. This table is used to keep track of which
479 * readers are using data from which old transactions, so that we'll know
480 * when a particular old transaction is no longer in use. Old transactions
481 * that have discarded any data pages can then have those pages reclaimed
482 * for use by a later write transaction.
484 * The lock table is constructed such that reader slots are aligned with the
485 * processor's cache line size. Any slot is only ever used by one thread.
486 * This alignment guarantees that there will be no contention or cache
487 * thrashing as threads update their own slot info, and also eliminates
488 * any need for locking when accessing a slot.
490 * A writer thread will scan every slot in the table to determine the oldest
491 * outstanding reader transaction. Any freed pages older than this will be
492 * reclaimed by the writer. The writer doesn't use any locks when scanning
493 * this table. This means that there's no guarantee that the writer will
494 * see the most up-to-date reader info, but that's not required for correct
495 * operation - all we need is to know the upper bound on the oldest reader,
496 * we don't care at all about the newest reader. So the only consequence of
497 * reading stale information here is that old pages might hang around a
498 * while longer before being reclaimed. That's actually good anyway, because
499 * the longer we delay reclaiming old pages, the more likely it is that a
500 * string of contiguous pages can be found after coalescing old pages from
501 * many old transactions together.
504 /** Number of slots in the reader table.
505 * This value was chosen somewhat arbitrarily. 126 readers plus a
506 * couple mutexes fit exactly into 8KB on my development machine.
507 * Applications should set the table size using #mdb_env_set_maxreaders().
509 #define DEFAULT_READERS 126
511 /** The size of a CPU cache line in bytes. We want our lock structures
512 * aligned to this size to avoid false cache line sharing in the
514 * This value works for most CPUs. For Itanium this should be 128.
520 /** The information we store in a single slot of the reader table.
521 * In addition to a transaction ID, we also record the process and
522 * thread ID that owns a slot, so that we can detect stale information,
523 * e.g. threads or processes that went away without cleaning up.
524 * @note We currently don't check for stale records. We simply re-init
525 * the table when we know that we're the only process opening the
528 typedef struct MDB_rxbody {
529 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
530 * Multiple readers that start at the same time will probably have the
531 * same ID here. Again, it's not important to exclude them from
532 * anything; all we need to know is which version of the DB they
533 * started from so we can avoid overwriting any data used in that
534 * particular version.
537 /** The process ID of the process owning this reader txn. */
539 /** The thread ID of the thread owning this txn. */
543 /** The actual reader record, with cacheline padding. */
544 typedef struct MDB_reader {
547 /** shorthand for mrb_txnid */
548 #define mr_txnid mru.mrx.mrb_txnid
549 #define mr_pid mru.mrx.mrb_pid
550 #define mr_tid mru.mrx.mrb_tid
551 /** cache line alignment */
552 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
556 /** The header for the reader table.
557 * The table resides in a memory-mapped file. (This is a different file
558 * than is used for the main database.)
560 * For POSIX the actual mutexes reside in the shared memory of this
561 * mapped file. On Windows, mutexes are named objects allocated by the
562 * kernel; we store the mutex names in this mapped file so that other
563 * processes can grab them. This same approach is also used on
564 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
565 * process-shared POSIX mutexes. For these cases where a named object
566 * is used, the object name is derived from a 64 bit FNV hash of the
567 * environment pathname. As such, naming collisions are extremely
568 * unlikely. If a collision occurs, the results are unpredictable.
570 typedef struct MDB_txbody {
571 /** Stamp identifying this as an MDB file. It must be set
574 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
576 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
577 char mtb_rmname[MNAME_LEN];
579 /** Mutex protecting access to this table.
580 * This is the reader lock that #LOCK_MUTEX_R acquires.
582 pthread_mutex_t mtb_mutex;
584 /** The ID of the last transaction committed to the database.
585 * This is recorded here only for convenience; the value can always
586 * be determined by reading the main database meta pages.
589 /** The number of slots that have been used in the reader table.
590 * This always records the maximum count, it is not decremented
591 * when readers release their slots.
593 unsigned mtb_numreaders;
596 /** The actual reader table definition. */
597 typedef struct MDB_txninfo {
600 #define mti_magic mt1.mtb.mtb_magic
601 #define mti_format mt1.mtb.mtb_format
602 #define mti_mutex mt1.mtb.mtb_mutex
603 #define mti_rmname mt1.mtb.mtb_rmname
604 #define mti_txnid mt1.mtb.mtb_txnid
605 #define mti_numreaders mt1.mtb.mtb_numreaders
606 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
609 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
610 char mt2_wmname[MNAME_LEN];
611 #define mti_wmname mt2.mt2_wmname
613 pthread_mutex_t mt2_wmutex;
614 #define mti_wmutex mt2.mt2_wmutex
616 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
618 MDB_reader mti_readers[1];
621 /** Lockfile format signature: version, features and field layout */
622 #define MDB_LOCK_FORMAT \
624 ((MDB_LOCK_VERSION) \
625 /* Flags which describe functionality */ \
626 + (((MDB_PIDLOCK) != 0) << 16)))
629 /** Common header for all page types.
630 * Overflow records occupy a number of contiguous pages with no
631 * headers on any page after the first.
633 typedef struct MDB_page {
634 #define mp_pgno mp_p.p_pgno
635 #define mp_next mp_p.p_next
637 pgno_t p_pgno; /**< page number */
638 void * p_next; /**< for in-memory list of freed structs */
641 /** @defgroup mdb_page Page Flags
643 * Flags for the page headers.
646 #define P_BRANCH 0x01 /**< branch page */
647 #define P_LEAF 0x02 /**< leaf page */
648 #define P_OVERFLOW 0x04 /**< overflow page */
649 #define P_META 0x08 /**< meta page */
650 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
651 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
652 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
653 #define P_KEEP 0x8000 /**< leave this page alone during spill */
655 uint16_t mp_flags; /**< @ref mdb_page */
656 #define mp_lower mp_pb.pb.pb_lower
657 #define mp_upper mp_pb.pb.pb_upper
658 #define mp_pages mp_pb.pb_pages
661 indx_t pb_lower; /**< lower bound of free space */
662 indx_t pb_upper; /**< upper bound of free space */
664 uint32_t pb_pages; /**< number of overflow pages */
666 indx_t mp_ptrs[1]; /**< dynamic size */
669 /** Size of the page header, excluding dynamic data at the end */
670 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
672 /** Address of first usable data byte in a page, after the header */
673 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
675 /** Number of nodes on a page */
676 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
678 /** The amount of space remaining in the page */
679 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
681 /** The percentage of space used in the page, in tenths of a percent. */
682 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
683 ((env)->me_psize - PAGEHDRSZ))
684 /** The minimum page fill factor, in tenths of a percent.
685 * Pages emptier than this are candidates for merging.
687 #define FILL_THRESHOLD 250
689 /** Test if a page is a leaf page */
690 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
691 /** Test if a page is a LEAF2 page */
692 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
693 /** Test if a page is a branch page */
694 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
695 /** Test if a page is an overflow page */
696 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
697 /** Test if a page is a sub page */
698 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
700 /** The number of overflow pages needed to store the given size. */
701 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
703 /** Header for a single key/data pair within a page.
704 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
705 * We guarantee 2-byte alignment for 'MDB_node's.
707 typedef struct MDB_node {
708 /** lo and hi are used for data size on leaf nodes and for
709 * child pgno on branch nodes. On 64 bit platforms, flags
710 * is also used for pgno. (Branch nodes have no flags).
711 * They are in host byte order in case that lets some
712 * accesses be optimized into a 32-bit word access.
714 #if BYTE_ORDER == LITTLE_ENDIAN
715 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
717 unsigned short mn_hi, mn_lo;
719 /** @defgroup mdb_node Node Flags
721 * Flags for node headers.
724 #define F_BIGDATA 0x01 /**< data put on overflow page */
725 #define F_SUBDATA 0x02 /**< data is a sub-database */
726 #define F_DUPDATA 0x04 /**< data has duplicates */
728 /** valid flags for #mdb_node_add() */
729 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
732 unsigned short mn_flags; /**< @ref mdb_node */
733 unsigned short mn_ksize; /**< key size */
734 char mn_data[1]; /**< key and data are appended here */
737 /** Size of the node header, excluding dynamic data at the end */
738 #define NODESIZE offsetof(MDB_node, mn_data)
740 /** Bit position of top word in page number, for shifting mn_flags */
741 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
743 /** Size of a node in a branch page with a given key.
744 * This is just the node header plus the key, there is no data.
746 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
748 /** Size of a node in a leaf page with a given key and data.
749 * This is node header plus key plus data size.
751 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
753 /** Address of node \b i in page \b p */
754 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
756 /** Address of the key for the node */
757 #define NODEKEY(node) (void *)((node)->mn_data)
759 /** Address of the data for a node */
760 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
762 /** Get the page number pointed to by a branch node */
763 #define NODEPGNO(node) \
764 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
765 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
766 /** Set the page number in a branch node */
767 #define SETPGNO(node,pgno) do { \
768 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
769 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
771 /** Get the size of the data in a leaf node */
772 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
773 /** Set the size of the data for a leaf node */
774 #define SETDSZ(node,size) do { \
775 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
776 /** The size of a key in a node */
777 #define NODEKSZ(node) ((node)->mn_ksize)
779 /** Copy a page number from src to dst */
781 #define COPY_PGNO(dst,src) dst = src
783 #if SIZE_MAX > 4294967295UL
784 #define COPY_PGNO(dst,src) do { \
785 unsigned short *s, *d; \
786 s = (unsigned short *)&(src); \
787 d = (unsigned short *)&(dst); \
794 #define COPY_PGNO(dst,src) do { \
795 unsigned short *s, *d; \
796 s = (unsigned short *)&(src); \
797 d = (unsigned short *)&(dst); \
803 /** The address of a key in a LEAF2 page.
804 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
805 * There are no node headers, keys are stored contiguously.
807 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
809 /** Set the \b node's key into \b keyptr, if requested. */
810 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
811 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
813 /** Set the \b node's key into \b key. */
814 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
816 /** Information about a single database in the environment. */
817 typedef struct MDB_db {
818 uint32_t md_pad; /**< also ksize for LEAF2 pages */
819 uint16_t md_flags; /**< @ref mdb_dbi_open */
820 uint16_t md_depth; /**< depth of this tree */
821 pgno_t md_branch_pages; /**< number of internal pages */
822 pgno_t md_leaf_pages; /**< number of leaf pages */
823 pgno_t md_overflow_pages; /**< number of overflow pages */
824 size_t md_entries; /**< number of data items */
825 pgno_t md_root; /**< the root page of this tree */
828 /** mdb_dbi_open flags */
829 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
830 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
831 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
832 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
834 /** Handle for the DB used to track free pages. */
836 /** Handle for the default DB. */
839 /** Meta page content.
840 * A meta page is the start point for accessing a database snapshot.
841 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
843 typedef struct MDB_meta {
844 /** Stamp identifying this as an MDB file. It must be set
847 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
849 void *mm_address; /**< address for fixed mapping */
850 size_t mm_mapsize; /**< size of mmap region */
851 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
852 /** The size of pages used in this DB */
853 #define mm_psize mm_dbs[0].md_pad
854 /** Any persistent environment flags. @ref mdb_env */
855 #define mm_flags mm_dbs[0].md_flags
856 pgno_t mm_last_pg; /**< last used page in file */
857 txnid_t mm_txnid; /**< txnid that committed this page */
860 /** Buffer for a stack-allocated meta page.
861 * The members define size and alignment, and silence type
862 * aliasing warnings. They are not used directly; that could
863 * mean incorrectly using several union members in parallel.
865 typedef union MDB_metabuf {
868 char mm_pad[PAGEHDRSZ];
873 /** Auxiliary DB info.
874 * The information here is mostly static/read-only. There is
875 * only a single copy of this record in the environment.
877 typedef struct MDB_dbx {
878 MDB_val md_name; /**< name of the database */
879 MDB_cmp_func *md_cmp; /**< function for comparing keys */
880 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
881 MDB_rel_func *md_rel; /**< user relocate function */
882 void *md_relctx; /**< user-provided context for md_rel */
885 /** A database transaction.
886 * Every operation requires a transaction handle.
889 MDB_txn *mt_parent; /**< parent of a nested txn */
890 MDB_txn *mt_child; /**< nested txn under this txn */
891 pgno_t mt_next_pgno; /**< next unallocated page */
892 /** The ID of this transaction. IDs are integers incrementing from 1.
893 * Only committed write transactions increment the ID. If a transaction
894 * aborts, the ID may be re-used by the next writer.
897 MDB_env *mt_env; /**< the DB environment */
898 /** The list of pages that became unused during this transaction.
901 /** The sorted list of dirty pages we temporarily wrote to disk
902 * because the dirty list was full. page numbers in here are
903 * shifted left by 1, deleted slots have the LSB set.
905 MDB_IDL mt_spill_pgs;
907 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
909 /** For read txns: This thread/txn's reader table slot, or NULL. */
912 /** Array of records for each DB known in the environment. */
914 /** Array of MDB_db records for each known DB */
916 /** @defgroup mt_dbflag Transaction DB Flags
920 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
921 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
922 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
923 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
925 /** In write txns, array of cursors for each DB */
926 MDB_cursor **mt_cursors;
927 /** Array of flags for each DB */
928 unsigned char *mt_dbflags;
929 /** Number of DB records in use. This number only ever increments;
930 * we don't decrement it when individual DB handles are closed.
934 /** @defgroup mdb_txn Transaction Flags
938 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
939 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
940 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
941 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
943 unsigned int mt_flags; /**< @ref mdb_txn */
944 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
945 * Includes ancestor txns' dirty pages not hidden by other txns'
946 * dirty/spilled pages. Thus commit(nested txn) has room to merge
947 * dirty_list into mt_parent after freeing hidden mt_parent pages.
949 unsigned int mt_dirty_room;
952 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
953 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
954 * raise this on a 64 bit machine.
956 #define CURSOR_STACK 32
960 /** Cursors are used for all DB operations.
961 * A cursor holds a path of (page pointer, key index) from the DB
962 * root to a position in the DB, plus other state. #MDB_DUPSORT
963 * cursors include an xcursor to the current data item. Write txns
964 * track their cursors and keep them up to date when data moves.
965 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
966 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
969 /** Next cursor on this DB in this txn */
971 /** Backup of the original cursor if this cursor is a shadow */
972 MDB_cursor *mc_backup;
973 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
974 struct MDB_xcursor *mc_xcursor;
975 /** The transaction that owns this cursor */
977 /** The database handle this cursor operates on */
979 /** The database record for this cursor */
981 /** The database auxiliary record for this cursor */
983 /** The @ref mt_dbflag for this database */
984 unsigned char *mc_dbflag;
985 unsigned short mc_snum; /**< number of pushed pages */
986 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
987 /** @defgroup mdb_cursor Cursor Flags
989 * Cursor state flags.
992 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
993 #define C_EOF 0x02 /**< No more data */
994 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
995 #define C_DEL 0x08 /**< last op was a cursor_del */
996 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
997 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
999 unsigned int mc_flags; /**< @ref mdb_cursor */
1000 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1001 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1004 /** Context for sorted-dup records.
1005 * We could have gone to a fully recursive design, with arbitrarily
1006 * deep nesting of sub-databases. But for now we only handle these
1007 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1009 typedef struct MDB_xcursor {
1010 /** A sub-cursor for traversing the Dup DB */
1011 MDB_cursor mx_cursor;
1012 /** The database record for this Dup DB */
1014 /** The auxiliary DB record for this Dup DB */
1016 /** The @ref mt_dbflag for this Dup DB */
1017 unsigned char mx_dbflag;
1020 /** State of FreeDB old pages, stored in the MDB_env */
1021 typedef struct MDB_pgstate {
1022 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1023 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1026 /** The database environment. */
1028 HANDLE me_fd; /**< The main data file */
1029 HANDLE me_lfd; /**< The lock file */
1030 HANDLE me_mfd; /**< just for writing the meta pages */
1031 /** Failed to update the meta page. Probably an I/O error. */
1032 #define MDB_FATAL_ERROR 0x80000000U
1033 /** Some fields are initialized. */
1034 #define MDB_ENV_ACTIVE 0x20000000U
1035 /** me_txkey is set */
1036 #define MDB_ENV_TXKEY 0x10000000U
1037 uint32_t me_flags; /**< @ref mdb_env */
1038 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1039 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1040 unsigned int me_maxreaders; /**< size of the reader table */
1041 unsigned int me_numreaders; /**< max numreaders set by this env */
1042 MDB_dbi me_numdbs; /**< number of DBs opened */
1043 MDB_dbi me_maxdbs; /**< size of the DB table */
1044 MDB_PID_T me_pid; /**< process ID of this env */
1045 char *me_path; /**< path to the DB files */
1046 char *me_map; /**< the memory map of the data file */
1047 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1048 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1049 void *me_pbuf; /**< scratch area for DUPSORT put() */
1050 MDB_txn *me_txn; /**< current write transaction */
1051 size_t me_mapsize; /**< size of the data memory map */
1052 off_t me_size; /**< current file size */
1053 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1054 MDB_dbx *me_dbxs; /**< array of static DB info */
1055 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1056 pthread_key_t me_txkey; /**< thread-key for readers */
1057 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1058 # define me_pglast me_pgstate.mf_pglast
1059 # define me_pghead me_pgstate.mf_pghead
1060 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1061 /** IDL of pages that became unused in a write txn */
1062 MDB_IDL me_free_pgs;
1063 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1064 MDB_ID2L me_dirty_list;
1065 /** Max number of freelist items that can fit in a single overflow page */
1067 /** Max size of a node on a page */
1068 unsigned int me_nodemax;
1069 #if !(MDB_MAXKEYSIZE)
1070 unsigned int me_maxkey; /**< max size of a key */
1072 int me_live_reader; /**< have liveness lock in reader table */
1074 int me_pidquery; /**< Used in OpenProcess */
1075 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1077 #elif defined(MDB_USE_POSIX_SEM)
1078 sem_t *me_rmutex; /* Shared mutexes are not supported */
1081 void *me_userctx; /**< User-settable context */
1082 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1085 /** Nested transaction */
1086 typedef struct MDB_ntxn {
1087 MDB_txn mnt_txn; /**< the transaction */
1088 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1091 /** max number of pages to commit in one writev() call */
1092 #define MDB_COMMIT_PAGES 64
1093 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1094 #undef MDB_COMMIT_PAGES
1095 #define MDB_COMMIT_PAGES IOV_MAX
1098 /** max bytes to write in one call */
1099 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1101 /** Check \b txn and \b dbi arguments to a function */
1102 #define TXN_DBI_EXIST(txn, dbi) \
1103 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1105 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1106 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1107 static int mdb_page_touch(MDB_cursor *mc);
1109 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1110 static int mdb_page_search_root(MDB_cursor *mc,
1111 MDB_val *key, int modify);
1112 #define MDB_PS_MODIFY 1
1113 #define MDB_PS_ROOTONLY 2
1114 #define MDB_PS_FIRST 4
1115 #define MDB_PS_LAST 8
1116 static int mdb_page_search(MDB_cursor *mc,
1117 MDB_val *key, int flags);
1118 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1120 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1121 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1122 pgno_t newpgno, unsigned int nflags);
1124 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1125 static int mdb_env_pick_meta(const MDB_env *env);
1126 static int mdb_env_write_meta(MDB_txn *txn);
1127 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1128 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1130 static void mdb_env_close0(MDB_env *env, int excl);
1132 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1133 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1134 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1135 static void mdb_node_del(MDB_cursor *mc, int ksize);
1136 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1137 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1138 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1139 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1140 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1142 static int mdb_rebalance(MDB_cursor *mc);
1143 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1145 static void mdb_cursor_pop(MDB_cursor *mc);
1146 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1148 static int mdb_cursor_del0(MDB_cursor *mc);
1149 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1150 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1151 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1152 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1153 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1155 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1156 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1158 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1159 static void mdb_xcursor_init0(MDB_cursor *mc);
1160 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1162 static int mdb_drop0(MDB_cursor *mc, int subs);
1163 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1166 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1170 static SECURITY_DESCRIPTOR mdb_null_sd;
1171 static SECURITY_ATTRIBUTES mdb_all_sa;
1172 static int mdb_sec_inited;
1175 /** Return the library version info. */
1177 mdb_version(int *major, int *minor, int *patch)
1179 if (major) *major = MDB_VERSION_MAJOR;
1180 if (minor) *minor = MDB_VERSION_MINOR;
1181 if (patch) *patch = MDB_VERSION_PATCH;
1182 return MDB_VERSION_STRING;
1185 /** Table of descriptions for MDB @ref errors */
1186 static char *const mdb_errstr[] = {
1187 "MDB_KEYEXIST: Key/data pair already exists",
1188 "MDB_NOTFOUND: No matching key/data pair found",
1189 "MDB_PAGE_NOTFOUND: Requested page not found",
1190 "MDB_CORRUPTED: Located page was wrong type",
1191 "MDB_PANIC: Update of meta page failed",
1192 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1193 "MDB_INVALID: File is not an MDB file",
1194 "MDB_MAP_FULL: Environment mapsize limit reached",
1195 "MDB_DBS_FULL: Environment maxdbs limit reached",
1196 "MDB_READERS_FULL: Environment maxreaders limit reached",
1197 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1198 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1199 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1200 "MDB_PAGE_FULL: Internal error - page has no more space",
1201 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1202 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1203 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1204 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1205 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1209 mdb_strerror(int err)
1213 return ("Successful return: 0");
1215 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1216 i = err - MDB_KEYEXIST;
1217 return mdb_errstr[i];
1220 return strerror(err);
1223 /** assert(3) variant in cursor context */
1224 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1225 /** assert(3) variant in transaction context */
1226 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1227 /** assert(3) variant in environment context */
1228 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1231 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1232 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1235 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1236 const char *func, const char *file, int line)
1239 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1240 file, line, expr_txt, func);
1241 if (env->me_assert_func)
1242 env->me_assert_func(env, buf);
1243 fprintf(stderr, "%s\n", buf);
1247 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1251 /** Return the page number of \b mp which may be sub-page, for debug output */
1253 mdb_dbg_pgno(MDB_page *mp)
1256 COPY_PGNO(ret, mp->mp_pgno);
1260 /** Display a key in hexadecimal and return the address of the result.
1261 * @param[in] key the key to display
1262 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1263 * @return The key in hexadecimal form.
1266 mdb_dkey(MDB_val *key, char *buf)
1269 unsigned char *c = key->mv_data;
1275 if (key->mv_size > DKBUF_MAXKEYSIZE)
1276 return "MDB_MAXKEYSIZE";
1277 /* may want to make this a dynamic check: if the key is mostly
1278 * printable characters, print it as-is instead of converting to hex.
1282 for (i=0; i<key->mv_size; i++)
1283 ptr += sprintf(ptr, "%02x", *c++);
1285 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1291 mdb_leafnode_type(MDB_node *n)
1293 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1294 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1295 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1298 /** Display all the keys in the page. */
1300 mdb_page_list(MDB_page *mp)
1302 pgno_t pgno = mdb_dbg_pgno(mp);
1303 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1305 unsigned int i, nkeys, nsize, total = 0;
1309 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1310 case P_BRANCH: type = "Branch page"; break;
1311 case P_LEAF: type = "Leaf page"; break;
1312 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1313 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1314 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1316 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1317 pgno, mp->mp_pages, state);
1320 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1321 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1324 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1328 nkeys = NUMKEYS(mp);
1329 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1331 for (i=0; i<nkeys; i++) {
1332 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1333 key.mv_size = nsize = mp->mp_pad;
1334 key.mv_data = LEAF2KEY(mp, i, nsize);
1336 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1339 node = NODEPTR(mp, i);
1340 key.mv_size = node->mn_ksize;
1341 key.mv_data = node->mn_data;
1342 nsize = NODESIZE + key.mv_size;
1343 if (IS_BRANCH(mp)) {
1344 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1348 if (F_ISSET(node->mn_flags, F_BIGDATA))
1349 nsize += sizeof(pgno_t);
1351 nsize += NODEDSZ(node);
1353 nsize += sizeof(indx_t);
1354 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1355 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1357 total = EVEN(total);
1359 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1360 IS_LEAF2(mp) ? PAGEHDRSZ : mp->mp_lower, total, SIZELEFT(mp));
1364 mdb_cursor_chk(MDB_cursor *mc)
1370 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1371 for (i=0; i<mc->mc_top; i++) {
1373 node = NODEPTR(mp, mc->mc_ki[i]);
1374 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1377 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1383 /** Count all the pages in each DB and in the freelist
1384 * and make sure it matches the actual number of pages
1386 * All named DBs must be open for a correct count.
1388 static void mdb_audit(MDB_txn *txn)
1392 MDB_ID freecount, count;
1397 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1398 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1399 freecount += *(MDB_ID *)data.mv_data;
1400 mdb_tassert(txn, rc == MDB_NOTFOUND);
1403 for (i = 0; i<txn->mt_numdbs; i++) {
1405 if (!(txn->mt_dbflags[i] & DB_VALID))
1407 mdb_cursor_init(&mc, txn, i, &mx);
1408 if (txn->mt_dbs[i].md_root == P_INVALID)
1410 count += txn->mt_dbs[i].md_branch_pages +
1411 txn->mt_dbs[i].md_leaf_pages +
1412 txn->mt_dbs[i].md_overflow_pages;
1413 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1414 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1415 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1418 mp = mc.mc_pg[mc.mc_top];
1419 for (j=0; j<NUMKEYS(mp); j++) {
1420 MDB_node *leaf = NODEPTR(mp, j);
1421 if (leaf->mn_flags & F_SUBDATA) {
1423 memcpy(&db, NODEDATA(leaf), sizeof(db));
1424 count += db.md_branch_pages + db.md_leaf_pages +
1425 db.md_overflow_pages;
1429 mdb_tassert(txn, rc == MDB_NOTFOUND);
1432 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1433 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1434 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1440 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1442 return txn->mt_dbxs[dbi].md_cmp(a, b);
1446 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1448 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1451 /** Allocate memory for a page.
1452 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1455 mdb_page_malloc(MDB_txn *txn, unsigned num)
1457 MDB_env *env = txn->mt_env;
1458 MDB_page *ret = env->me_dpages;
1459 size_t psize = env->me_psize, sz = psize, off;
1460 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1461 * For a single page alloc, we init everything after the page header.
1462 * For multi-page, we init the final page; if the caller needed that
1463 * many pages they will be filling in at least up to the last page.
1467 VGMEMP_ALLOC(env, ret, sz);
1468 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1469 env->me_dpages = ret->mp_next;
1472 psize -= off = PAGEHDRSZ;
1477 if ((ret = malloc(sz)) != NULL) {
1478 VGMEMP_ALLOC(env, ret, sz);
1479 if (!(env->me_flags & MDB_NOMEMINIT)) {
1480 memset((char *)ret + off, 0, psize);
1484 txn->mt_flags |= MDB_TXN_ERROR;
1489 /** Free a single page.
1490 * Saves single pages to a list, for future reuse.
1491 * (This is not used for multi-page overflow pages.)
1494 mdb_page_free(MDB_env *env, MDB_page *mp)
1496 mp->mp_next = env->me_dpages;
1497 VGMEMP_FREE(env, mp);
1498 env->me_dpages = mp;
1501 /** Free a dirty page */
1503 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1505 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1506 mdb_page_free(env, dp);
1508 /* large pages just get freed directly */
1509 VGMEMP_FREE(env, dp);
1514 /** Return all dirty pages to dpage list */
1516 mdb_dlist_free(MDB_txn *txn)
1518 MDB_env *env = txn->mt_env;
1519 MDB_ID2L dl = txn->mt_u.dirty_list;
1520 unsigned i, n = dl[0].mid;
1522 for (i = 1; i <= n; i++) {
1523 mdb_dpage_free(env, dl[i].mptr);
1528 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1529 * @param[in] mc A cursor handle for the current operation.
1530 * @param[in] pflags Flags of the pages to update:
1531 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1532 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1533 * @return 0 on success, non-zero on failure.
1536 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1538 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1539 MDB_txn *txn = mc->mc_txn;
1545 int rc = MDB_SUCCESS, level;
1547 /* Mark pages seen by cursors */
1548 if (mc->mc_flags & C_UNTRACK)
1549 mc = NULL; /* will find mc in mt_cursors */
1550 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1551 for (; mc; mc=mc->mc_next) {
1552 if (!(mc->mc_flags & C_INITIALIZED))
1554 for (m3 = mc;; m3 = &mx->mx_cursor) {
1556 for (j=0; j<m3->mc_snum; j++) {
1558 if ((mp->mp_flags & Mask) == pflags)
1559 mp->mp_flags ^= P_KEEP;
1561 mx = m3->mc_xcursor;
1562 /* Proceed to mx if it is at a sub-database */
1563 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1565 if (! (mp && (mp->mp_flags & P_LEAF)))
1567 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1568 if (!(leaf->mn_flags & F_SUBDATA))
1577 /* Mark dirty root pages */
1578 for (i=0; i<txn->mt_numdbs; i++) {
1579 if (txn->mt_dbflags[i] & DB_DIRTY) {
1580 pgno_t pgno = txn->mt_dbs[i].md_root;
1581 if (pgno == P_INVALID)
1583 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1585 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1586 dp->mp_flags ^= P_KEEP;
1594 static int mdb_page_flush(MDB_txn *txn, int keep);
1596 /** Spill pages from the dirty list back to disk.
1597 * This is intended to prevent running into #MDB_TXN_FULL situations,
1598 * but note that they may still occur in a few cases:
1599 * 1) our estimate of the txn size could be too small. Currently this
1600 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1601 * 2) child txns may run out of space if their parents dirtied a
1602 * lot of pages and never spilled them. TODO: we probably should do
1603 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1604 * the parent's dirty_room is below a given threshold.
1606 * Otherwise, if not using nested txns, it is expected that apps will
1607 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1608 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1609 * If the txn never references them again, they can be left alone.
1610 * If the txn only reads them, they can be used without any fuss.
1611 * If the txn writes them again, they can be dirtied immediately without
1612 * going thru all of the work of #mdb_page_touch(). Such references are
1613 * handled by #mdb_page_unspill().
1615 * Also note, we never spill DB root pages, nor pages of active cursors,
1616 * because we'll need these back again soon anyway. And in nested txns,
1617 * we can't spill a page in a child txn if it was already spilled in a
1618 * parent txn. That would alter the parent txns' data even though
1619 * the child hasn't committed yet, and we'd have no way to undo it if
1620 * the child aborted.
1622 * @param[in] m0 cursor A cursor handle identifying the transaction and
1623 * database for which we are checking space.
1624 * @param[in] key For a put operation, the key being stored.
1625 * @param[in] data For a put operation, the data being stored.
1626 * @return 0 on success, non-zero on failure.
1629 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1631 MDB_txn *txn = m0->mc_txn;
1633 MDB_ID2L dl = txn->mt_u.dirty_list;
1634 unsigned int i, j, need;
1637 if (m0->mc_flags & C_SUB)
1640 /* Estimate how much space this op will take */
1641 i = m0->mc_db->md_depth;
1642 /* Named DBs also dirty the main DB */
1643 if (m0->mc_dbi > MAIN_DBI)
1644 i += txn->mt_dbs[MAIN_DBI].md_depth;
1645 /* For puts, roughly factor in the key+data size */
1647 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1648 i += i; /* double it for good measure */
1651 if (txn->mt_dirty_room > i)
1654 if (!txn->mt_spill_pgs) {
1655 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1656 if (!txn->mt_spill_pgs)
1659 /* purge deleted slots */
1660 MDB_IDL sl = txn->mt_spill_pgs;
1661 unsigned int num = sl[0];
1663 for (i=1; i<=num; i++) {
1670 /* Preserve pages which may soon be dirtied again */
1671 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1674 /* Less aggressive spill - we originally spilled the entire dirty list,
1675 * with a few exceptions for cursor pages and DB root pages. But this
1676 * turns out to be a lot of wasted effort because in a large txn many
1677 * of those pages will need to be used again. So now we spill only 1/8th
1678 * of the dirty pages. Testing revealed this to be a good tradeoff,
1679 * better than 1/2, 1/4, or 1/10.
1681 if (need < MDB_IDL_UM_MAX / 8)
1682 need = MDB_IDL_UM_MAX / 8;
1684 /* Save the page IDs of all the pages we're flushing */
1685 /* flush from the tail forward, this saves a lot of shifting later on. */
1686 for (i=dl[0].mid; i && need; i--) {
1687 MDB_ID pn = dl[i].mid << 1;
1689 if (dp->mp_flags & P_KEEP)
1691 /* Can't spill twice, make sure it's not already in a parent's
1694 if (txn->mt_parent) {
1696 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1697 if (tx2->mt_spill_pgs) {
1698 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1699 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1700 dp->mp_flags |= P_KEEP;
1708 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1712 mdb_midl_sort(txn->mt_spill_pgs);
1714 /* Flush the spilled part of dirty list */
1715 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1718 /* Reset any dirty pages we kept that page_flush didn't see */
1719 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1722 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1726 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1728 mdb_find_oldest(MDB_txn *txn)
1731 txnid_t mr, oldest = txn->mt_txnid - 1;
1732 if (txn->mt_env->me_txns) {
1733 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1734 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1745 /** Add a page to the txn's dirty list */
1747 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1750 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1752 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1753 insert = mdb_mid2l_append;
1755 insert = mdb_mid2l_insert;
1757 mid.mid = mp->mp_pgno;
1759 rc = insert(txn->mt_u.dirty_list, &mid);
1760 mdb_tassert(txn, rc == 0);
1761 txn->mt_dirty_room--;
1764 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1765 * me_pghead and mt_next_pgno.
1767 * If there are free pages available from older transactions, they
1768 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1769 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1770 * and move me_pglast to say which records were consumed. Only this
1771 * function can create me_pghead and move me_pglast/mt_next_pgno.
1772 * @param[in] mc cursor A cursor handle identifying the transaction and
1773 * database for which we are allocating.
1774 * @param[in] num the number of pages to allocate.
1775 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1776 * will always be satisfied by a single contiguous chunk of memory.
1777 * @return 0 on success, non-zero on failure.
1780 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1782 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1783 /* Get at most <Max_retries> more freeDB records once me_pghead
1784 * has enough pages. If not enough, use new pages from the map.
1785 * If <Paranoid> and mc is updating the freeDB, only get new
1786 * records if me_pghead is empty. Then the freelist cannot play
1787 * catch-up with itself by growing while trying to save it.
1789 enum { Paranoid = 1, Max_retries = 500 };
1791 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1793 int rc, retry = Max_retries;
1794 MDB_txn *txn = mc->mc_txn;
1795 MDB_env *env = txn->mt_env;
1796 pgno_t pgno, *mop = env->me_pghead;
1797 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1799 txnid_t oldest = 0, last;
1805 /* If our dirty list is already full, we can't do anything */
1806 if (txn->mt_dirty_room == 0) {
1811 for (op = MDB_FIRST;; op = MDB_NEXT) {
1814 pgno_t *idl, old_id, new_id;
1816 /* Seek a big enough contiguous page range. Prefer
1817 * pages at the tail, just truncating the list.
1823 if (mop[i-n2] == pgno+n2)
1826 if (Max_retries < INT_MAX && --retry < 0)
1830 if (op == MDB_FIRST) { /* 1st iteration */
1831 /* Prepare to fetch more and coalesce */
1832 oldest = mdb_find_oldest(txn);
1833 last = env->me_pglast;
1834 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1837 key.mv_data = &last; /* will look up last+1 */
1838 key.mv_size = sizeof(last);
1840 if (Paranoid && mc->mc_dbi == FREE_DBI)
1843 if (Paranoid && retry < 0 && mop_len)
1847 /* Do not fetch more if the record will be too recent */
1850 rc = mdb_cursor_get(&m2, &key, NULL, op);
1852 if (rc == MDB_NOTFOUND)
1856 last = *(txnid_t*)key.mv_data;
1859 np = m2.mc_pg[m2.mc_top];
1860 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1861 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1864 idl = (MDB_ID *) data.mv_data;
1867 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1872 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1874 mop = env->me_pghead;
1876 env->me_pglast = last;
1878 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1879 last, txn->mt_dbs[FREE_DBI].md_root, i));
1881 DPRINTF(("IDL %"Z"u", idl[k]));
1883 /* Merge in descending sorted order */
1886 mop[0] = (pgno_t)-1;
1890 for (; old_id < new_id; old_id = mop[--j])
1897 /* Use new pages from the map when nothing suitable in the freeDB */
1899 pgno = txn->mt_next_pgno;
1900 if (pgno + num >= env->me_maxpg) {
1901 DPUTS("DB size maxed out");
1907 if (env->me_flags & MDB_WRITEMAP) {
1908 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1910 if (!(np = mdb_page_malloc(txn, num))) {
1916 mop[0] = mop_len -= num;
1917 /* Move any stragglers down */
1918 for (j = i-num; j < mop_len; )
1919 mop[++j] = mop[++i];
1921 txn->mt_next_pgno = pgno + num;
1924 mdb_page_dirty(txn, np);
1930 txn->mt_flags |= MDB_TXN_ERROR;
1934 /** Copy the used portions of a non-overflow page.
1935 * @param[in] dst page to copy into
1936 * @param[in] src page to copy from
1937 * @param[in] psize size of a page
1940 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1942 enum { Align = sizeof(pgno_t) };
1943 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1945 /* If page isn't full, just copy the used portion. Adjust
1946 * alignment so memcpy may copy words instead of bytes.
1948 if ((unused &= -Align) && !IS_LEAF2(src)) {
1950 memcpy(dst, src, (lower + (Align-1)) & -Align);
1951 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1954 memcpy(dst, src, psize - unused);
1958 /** Pull a page off the txn's spill list, if present.
1959 * If a page being referenced was spilled to disk in this txn, bring
1960 * it back and make it dirty/writable again.
1961 * @param[in] txn the transaction handle.
1962 * @param[in] mp the page being referenced. It must not be dirty.
1963 * @param[out] ret the writable page, if any. ret is unchanged if
1964 * mp wasn't spilled.
1967 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1969 MDB_env *env = txn->mt_env;
1972 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1974 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1975 if (!tx2->mt_spill_pgs)
1977 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1978 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1981 if (txn->mt_dirty_room == 0)
1982 return MDB_TXN_FULL;
1983 if (IS_OVERFLOW(mp))
1987 if (env->me_flags & MDB_WRITEMAP) {
1990 np = mdb_page_malloc(txn, num);
1994 memcpy(np, mp, num * env->me_psize);
1996 mdb_page_copy(np, mp, env->me_psize);
1999 /* If in current txn, this page is no longer spilled.
2000 * If it happens to be the last page, truncate the spill list.
2001 * Otherwise mark it as deleted by setting the LSB.
2003 if (x == txn->mt_spill_pgs[0])
2004 txn->mt_spill_pgs[0]--;
2006 txn->mt_spill_pgs[x] |= 1;
2007 } /* otherwise, if belonging to a parent txn, the
2008 * page remains spilled until child commits
2011 mdb_page_dirty(txn, np);
2012 np->mp_flags |= P_DIRTY;
2020 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2021 * @param[in] mc cursor pointing to the page to be touched
2022 * @return 0 on success, non-zero on failure.
2025 mdb_page_touch(MDB_cursor *mc)
2027 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2028 MDB_txn *txn = mc->mc_txn;
2029 MDB_cursor *m2, *m3;
2033 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2034 if (txn->mt_flags & MDB_TXN_SPILLS) {
2036 rc = mdb_page_unspill(txn, mp, &np);
2042 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2043 (rc = mdb_page_alloc(mc, 1, &np)))
2046 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2047 mp->mp_pgno, pgno));
2048 mdb_cassert(mc, mp->mp_pgno != pgno);
2049 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2050 /* Update the parent page, if any, to point to the new page */
2052 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2053 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2054 SETPGNO(node, pgno);
2056 mc->mc_db->md_root = pgno;
2058 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2059 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2061 /* If txn has a parent, make sure the page is in our
2065 unsigned x = mdb_mid2l_search(dl, pgno);
2066 if (x <= dl[0].mid && dl[x].mid == pgno) {
2067 if (mp != dl[x].mptr) { /* bad cursor? */
2068 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2069 txn->mt_flags |= MDB_TXN_ERROR;
2070 return MDB_CORRUPTED;
2075 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2077 np = mdb_page_malloc(txn, 1);
2082 rc = mdb_mid2l_insert(dl, &mid);
2083 mdb_cassert(mc, rc == 0);
2088 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2090 np->mp_flags |= P_DIRTY;
2093 /* Adjust cursors pointing to mp */
2094 mc->mc_pg[mc->mc_top] = np;
2095 m2 = txn->mt_cursors[mc->mc_dbi];
2096 if (mc->mc_flags & C_SUB) {
2097 for (; m2; m2=m2->mc_next) {
2098 m3 = &m2->mc_xcursor->mx_cursor;
2099 if (m3->mc_snum < mc->mc_snum) continue;
2100 if (m3->mc_pg[mc->mc_top] == mp)
2101 m3->mc_pg[mc->mc_top] = np;
2104 for (; m2; m2=m2->mc_next) {
2105 if (m2->mc_snum < mc->mc_snum) continue;
2106 if (m2->mc_pg[mc->mc_top] == mp) {
2107 m2->mc_pg[mc->mc_top] = np;
2108 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2110 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2112 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2113 if (!(leaf->mn_flags & F_SUBDATA))
2114 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2122 txn->mt_flags |= MDB_TXN_ERROR;
2127 mdb_env_sync(MDB_env *env, int force)
2130 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2131 if (env->me_flags & MDB_WRITEMAP) {
2132 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2133 ? MS_ASYNC : MS_SYNC;
2134 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2137 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2141 if (MDB_FDATASYNC(env->me_fd))
2148 /** Back up parent txn's cursors, then grab the originals for tracking */
2150 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2152 MDB_cursor *mc, *bk;
2157 for (i = src->mt_numdbs; --i >= 0; ) {
2158 if ((mc = src->mt_cursors[i]) != NULL) {
2159 size = sizeof(MDB_cursor);
2161 size += sizeof(MDB_xcursor);
2162 for (; mc; mc = bk->mc_next) {
2168 mc->mc_db = &dst->mt_dbs[i];
2169 /* Kill pointers into src - and dst to reduce abuse: The
2170 * user may not use mc until dst ends. Otherwise we'd...
2172 mc->mc_txn = NULL; /* ...set this to dst */
2173 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2174 if ((mx = mc->mc_xcursor) != NULL) {
2175 *(MDB_xcursor *)(bk+1) = *mx;
2176 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2178 mc->mc_next = dst->mt_cursors[i];
2179 dst->mt_cursors[i] = mc;
2186 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2187 * @param[in] txn the transaction handle.
2188 * @param[in] merge true to keep changes to parent cursors, false to revert.
2189 * @return 0 on success, non-zero on failure.
2192 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2194 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2198 for (i = txn->mt_numdbs; --i >= 0; ) {
2199 for (mc = cursors[i]; mc; mc = next) {
2201 if ((bk = mc->mc_backup) != NULL) {
2203 /* Commit changes to parent txn */
2204 mc->mc_next = bk->mc_next;
2205 mc->mc_backup = bk->mc_backup;
2206 mc->mc_txn = bk->mc_txn;
2207 mc->mc_db = bk->mc_db;
2208 mc->mc_dbflag = bk->mc_dbflag;
2209 if ((mx = mc->mc_xcursor) != NULL)
2210 mx->mx_cursor.mc_txn = bk->mc_txn;
2212 /* Abort nested txn */
2214 if ((mx = mc->mc_xcursor) != NULL)
2215 *mx = *(MDB_xcursor *)(bk+1);
2219 /* Only malloced cursors are permanently tracked. */
2227 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2230 mdb_txn_reset0(MDB_txn *txn, const char *act);
2232 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2238 Pidset = F_SETLK, Pidcheck = F_GETLK
2242 /** Set or check a pid lock. Set returns 0 on success.
2243 * Check returns 0 if the process is certainly dead, nonzero if it may
2244 * be alive (the lock exists or an error happened so we do not know).
2246 * On Windows Pidset is a no-op, we merely check for the existence
2247 * of the process with the given pid. On POSIX we use a single byte
2248 * lock on the lockfile, set at an offset equal to the pid.
2251 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2253 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2256 if (op == Pidcheck) {
2257 h = OpenProcess(env->me_pidquery, FALSE, pid);
2258 /* No documented "no such process" code, but other program use this: */
2260 return ErrCode() != ERROR_INVALID_PARAMETER;
2261 /* A process exists until all handles to it close. Has it exited? */
2262 ret = WaitForSingleObject(h, 0) != 0;
2269 struct flock lock_info;
2270 memset(&lock_info, 0, sizeof(lock_info));
2271 lock_info.l_type = F_WRLCK;
2272 lock_info.l_whence = SEEK_SET;
2273 lock_info.l_start = pid;
2274 lock_info.l_len = 1;
2275 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2276 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2278 } else if ((rc = ErrCode()) == EINTR) {
2286 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2287 * @param[in] txn the transaction handle to initialize
2288 * @return 0 on success, non-zero on failure.
2291 mdb_txn_renew0(MDB_txn *txn)
2293 MDB_env *env = txn->mt_env;
2294 MDB_txninfo *ti = env->me_txns;
2298 int rc, new_notls = 0;
2301 txn->mt_numdbs = env->me_numdbs;
2302 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2304 if (txn->mt_flags & MDB_TXN_RDONLY) {
2306 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2307 txn->mt_txnid = meta->mm_txnid;
2308 txn->mt_u.reader = NULL;
2310 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2311 pthread_getspecific(env->me_txkey);
2313 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2314 return MDB_BAD_RSLOT;
2316 MDB_PID_T pid = env->me_pid;
2317 pthread_t tid = pthread_self();
2319 if (!env->me_live_reader) {
2320 rc = mdb_reader_pid(env, Pidset, pid);
2323 env->me_live_reader = 1;
2327 nr = ti->mti_numreaders;
2328 for (i=0; i<nr; i++)
2329 if (ti->mti_readers[i].mr_pid == 0)
2331 if (i == env->me_maxreaders) {
2332 UNLOCK_MUTEX_R(env);
2333 return MDB_READERS_FULL;
2335 ti->mti_readers[i].mr_pid = pid;
2336 ti->mti_readers[i].mr_tid = tid;
2338 ti->mti_numreaders = ++nr;
2339 /* Save numreaders for un-mutexed mdb_env_close() */
2340 env->me_numreaders = nr;
2341 UNLOCK_MUTEX_R(env);
2343 r = &ti->mti_readers[i];
2344 new_notls = (env->me_flags & MDB_NOTLS);
2345 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2350 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2351 txn->mt_u.reader = r;
2352 meta = env->me_metas[txn->mt_txnid & 1];
2358 txn->mt_txnid = ti->mti_txnid;
2359 meta = env->me_metas[txn->mt_txnid & 1];
2361 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2362 txn->mt_txnid = meta->mm_txnid;
2366 if (txn->mt_txnid == mdb_debug_start)
2369 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2370 txn->mt_u.dirty_list = env->me_dirty_list;
2371 txn->mt_u.dirty_list[0].mid = 0;
2372 txn->mt_free_pgs = env->me_free_pgs;
2373 txn->mt_free_pgs[0] = 0;
2374 txn->mt_spill_pgs = NULL;
2378 /* Copy the DB info and flags */
2379 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2381 /* Moved to here to avoid a data race in read TXNs */
2382 txn->mt_next_pgno = meta->mm_last_pg+1;
2384 for (i=2; i<txn->mt_numdbs; i++) {
2385 x = env->me_dbflags[i];
2386 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2387 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2389 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2391 if (env->me_maxpg < txn->mt_next_pgno) {
2392 mdb_txn_reset0(txn, "renew0-mapfail");
2394 txn->mt_u.reader->mr_pid = 0;
2395 txn->mt_u.reader = NULL;
2397 return MDB_MAP_RESIZED;
2404 mdb_txn_renew(MDB_txn *txn)
2408 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2411 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2412 DPUTS("environment had fatal error, must shutdown!");
2416 rc = mdb_txn_renew0(txn);
2417 if (rc == MDB_SUCCESS) {
2418 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2419 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2420 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2426 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2430 int rc, size, tsize = sizeof(MDB_txn);
2432 if (env->me_flags & MDB_FATAL_ERROR) {
2433 DPUTS("environment had fatal error, must shutdown!");
2436 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2439 /* Nested transactions: Max 1 child, write txns only, no writemap */
2440 if (parent->mt_child ||
2441 (flags & MDB_RDONLY) ||
2442 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2443 (env->me_flags & MDB_WRITEMAP))
2445 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2447 tsize = sizeof(MDB_ntxn);
2449 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2450 if (!(flags & MDB_RDONLY))
2451 size += env->me_maxdbs * sizeof(MDB_cursor *);
2453 if ((txn = calloc(1, size)) == NULL) {
2454 DPRINTF(("calloc: %s", strerror(ErrCode())));
2457 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2458 if (flags & MDB_RDONLY) {
2459 txn->mt_flags |= MDB_TXN_RDONLY;
2460 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2462 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2463 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2469 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2470 if (!txn->mt_u.dirty_list ||
2471 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2473 free(txn->mt_u.dirty_list);
2477 txn->mt_txnid = parent->mt_txnid;
2478 txn->mt_dirty_room = parent->mt_dirty_room;
2479 txn->mt_u.dirty_list[0].mid = 0;
2480 txn->mt_spill_pgs = NULL;
2481 txn->mt_next_pgno = parent->mt_next_pgno;
2482 parent->mt_child = txn;
2483 txn->mt_parent = parent;
2484 txn->mt_numdbs = parent->mt_numdbs;
2485 txn->mt_flags = parent->mt_flags;
2486 txn->mt_dbxs = parent->mt_dbxs;
2487 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2488 /* Copy parent's mt_dbflags, but clear DB_NEW */
2489 for (i=0; i<txn->mt_numdbs; i++)
2490 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2492 ntxn = (MDB_ntxn *)txn;
2493 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2494 if (env->me_pghead) {
2495 size = MDB_IDL_SIZEOF(env->me_pghead);
2496 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2498 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2503 rc = mdb_cursor_shadow(parent, txn);
2505 mdb_txn_reset0(txn, "beginchild-fail");
2507 rc = mdb_txn_renew0(txn);
2513 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2514 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2515 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2522 mdb_txn_env(MDB_txn *txn)
2524 if(!txn) return NULL;
2528 /** Export or close DBI handles opened in this txn. */
2530 mdb_dbis_update(MDB_txn *txn, int keep)
2533 MDB_dbi n = txn->mt_numdbs;
2534 MDB_env *env = txn->mt_env;
2535 unsigned char *tdbflags = txn->mt_dbflags;
2537 for (i = n; --i >= 2;) {
2538 if (tdbflags[i] & DB_NEW) {
2540 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2542 char *ptr = env->me_dbxs[i].md_name.mv_data;
2543 env->me_dbxs[i].md_name.mv_data = NULL;
2544 env->me_dbxs[i].md_name.mv_size = 0;
2545 env->me_dbflags[i] = 0;
2550 if (keep && env->me_numdbs < n)
2554 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2555 * May be called twice for readonly txns: First reset it, then abort.
2556 * @param[in] txn the transaction handle to reset
2557 * @param[in] act why the transaction is being reset
2560 mdb_txn_reset0(MDB_txn *txn, const char *act)
2562 MDB_env *env = txn->mt_env;
2564 /* Close any DBI handles opened in this txn */
2565 mdb_dbis_update(txn, 0);
2567 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2568 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2569 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2571 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2572 if (txn->mt_u.reader) {
2573 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2574 if (!(env->me_flags & MDB_NOTLS))
2575 txn->mt_u.reader = NULL; /* txn does not own reader */
2577 txn->mt_numdbs = 0; /* close nothing if called again */
2578 txn->mt_dbxs = NULL; /* mark txn as reset */
2580 mdb_cursors_close(txn, 0);
2582 if (!(env->me_flags & MDB_WRITEMAP)) {
2583 mdb_dlist_free(txn);
2585 mdb_midl_free(env->me_pghead);
2587 if (txn->mt_parent) {
2588 txn->mt_parent->mt_child = NULL;
2589 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2590 mdb_midl_free(txn->mt_free_pgs);
2591 mdb_midl_free(txn->mt_spill_pgs);
2592 free(txn->mt_u.dirty_list);
2596 if (mdb_midl_shrink(&txn->mt_free_pgs))
2597 env->me_free_pgs = txn->mt_free_pgs;
2598 env->me_pghead = NULL;
2602 /* The writer mutex was locked in mdb_txn_begin. */
2604 UNLOCK_MUTEX_W(env);
2609 mdb_txn_reset(MDB_txn *txn)
2614 /* This call is only valid for read-only txns */
2615 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2618 mdb_txn_reset0(txn, "reset");
2622 mdb_txn_abort(MDB_txn *txn)
2628 mdb_txn_abort(txn->mt_child);
2630 mdb_txn_reset0(txn, "abort");
2631 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2632 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2633 txn->mt_u.reader->mr_pid = 0;
2638 /** Save the freelist as of this transaction to the freeDB.
2639 * This changes the freelist. Keep trying until it stabilizes.
2642 mdb_freelist_save(MDB_txn *txn)
2644 /* env->me_pghead[] can grow and shrink during this call.
2645 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2646 * Page numbers cannot disappear from txn->mt_free_pgs[].
2649 MDB_env *env = txn->mt_env;
2650 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2651 txnid_t pglast = 0, head_id = 0;
2652 pgno_t freecnt = 0, *free_pgs, *mop;
2653 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2655 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2657 if (env->me_pghead) {
2658 /* Make sure first page of freeDB is touched and on freelist */
2659 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2660 if (rc && rc != MDB_NOTFOUND)
2664 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2665 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2666 ? SSIZE_MAX : maxfree_1pg;
2669 /* Come back here after each Put() in case freelist changed */
2674 /* If using records from freeDB which we have not yet
2675 * deleted, delete them and any we reserved for me_pghead.
2677 while (pglast < env->me_pglast) {
2678 rc = mdb_cursor_first(&mc, &key, NULL);
2681 pglast = head_id = *(txnid_t *)key.mv_data;
2682 total_room = head_room = 0;
2683 mdb_tassert(txn, pglast <= env->me_pglast);
2684 rc = mdb_cursor_del(&mc, 0);
2689 /* Save the IDL of pages freed by this txn, to a single record */
2690 if (freecnt < txn->mt_free_pgs[0]) {
2692 /* Make sure last page of freeDB is touched and on freelist */
2693 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2694 if (rc && rc != MDB_NOTFOUND)
2697 free_pgs = txn->mt_free_pgs;
2698 /* Write to last page of freeDB */
2699 key.mv_size = sizeof(txn->mt_txnid);
2700 key.mv_data = &txn->mt_txnid;
2702 freecnt = free_pgs[0];
2703 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2704 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2707 /* Retry if mt_free_pgs[] grew during the Put() */
2708 free_pgs = txn->mt_free_pgs;
2709 } while (freecnt < free_pgs[0]);
2710 mdb_midl_sort(free_pgs);
2711 memcpy(data.mv_data, free_pgs, data.mv_size);
2714 unsigned int i = free_pgs[0];
2715 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2716 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2718 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2724 mop = env->me_pghead;
2725 mop_len = mop ? mop[0] : 0;
2727 /* Reserve records for me_pghead[]. Split it if multi-page,
2728 * to avoid searching freeDB for a page range. Use keys in
2729 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2731 if (total_room >= mop_len) {
2732 if (total_room == mop_len || --more < 0)
2734 } else if (head_room >= maxfree_1pg && head_id > 1) {
2735 /* Keep current record (overflow page), add a new one */
2739 /* (Re)write {key = head_id, IDL length = head_room} */
2740 total_room -= head_room;
2741 head_room = mop_len - total_room;
2742 if (head_room > maxfree_1pg && head_id > 1) {
2743 /* Overflow multi-page for part of me_pghead */
2744 head_room /= head_id; /* amortize page sizes */
2745 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2746 } else if (head_room < 0) {
2747 /* Rare case, not bothering to delete this record */
2750 key.mv_size = sizeof(head_id);
2751 key.mv_data = &head_id;
2752 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2753 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2756 /* IDL is initially empty, zero out at least the length */
2757 pgs = (pgno_t *)data.mv_data;
2758 j = head_room > clean_limit ? head_room : 0;
2762 total_room += head_room;
2765 /* Fill in the reserved me_pghead records */
2771 rc = mdb_cursor_first(&mc, &key, &data);
2772 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2773 txnid_t id = *(txnid_t *)key.mv_data;
2774 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2777 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2779 if (len > mop_len) {
2781 data.mv_size = (len + 1) * sizeof(MDB_ID);
2783 data.mv_data = mop -= len;
2786 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2788 if (rc || !(mop_len -= len))
2795 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2796 * @param[in] txn the transaction that's being committed
2797 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2798 * @return 0 on success, non-zero on failure.
2801 mdb_page_flush(MDB_txn *txn, int keep)
2803 MDB_env *env = txn->mt_env;
2804 MDB_ID2L dl = txn->mt_u.dirty_list;
2805 unsigned psize = env->me_psize, j;
2806 int i, pagecount = dl[0].mid, rc;
2807 size_t size = 0, pos = 0;
2809 MDB_page *dp = NULL;
2813 struct iovec iov[MDB_COMMIT_PAGES];
2814 ssize_t wpos = 0, wsize = 0, wres;
2815 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2821 if (env->me_flags & MDB_WRITEMAP) {
2822 /* Clear dirty flags */
2823 while (++i <= pagecount) {
2825 /* Don't flush this page yet */
2826 if (dp->mp_flags & P_KEEP) {
2827 dp->mp_flags ^= P_KEEP;
2831 dp->mp_flags &= ~P_DIRTY;
2836 /* Write the pages */
2838 if (++i <= pagecount) {
2840 /* Don't flush this page yet */
2841 if (dp->mp_flags & P_KEEP) {
2842 dp->mp_flags ^= P_KEEP;
2847 /* clear dirty flag */
2848 dp->mp_flags &= ~P_DIRTY;
2851 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2856 /* Windows actually supports scatter/gather I/O, but only on
2857 * unbuffered file handles. Since we're relying on the OS page
2858 * cache for all our data, that's self-defeating. So we just
2859 * write pages one at a time. We use the ov structure to set
2860 * the write offset, to at least save the overhead of a Seek
2863 DPRINTF(("committing page %"Z"u", pgno));
2864 memset(&ov, 0, sizeof(ov));
2865 ov.Offset = pos & 0xffffffff;
2866 ov.OffsetHigh = pos >> 16 >> 16;
2867 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2869 DPRINTF(("WriteFile: %d", rc));
2873 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2874 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2876 /* Write previous page(s) */
2877 #ifdef MDB_USE_PWRITEV
2878 wres = pwritev(env->me_fd, iov, n, wpos);
2881 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2883 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2885 DPRINTF(("lseek: %s", strerror(rc)));
2888 wres = writev(env->me_fd, iov, n);
2891 if (wres != wsize) {
2894 DPRINTF(("Write error: %s", strerror(rc)));
2896 rc = EIO; /* TODO: Use which error code? */
2897 DPUTS("short write, filesystem full?");
2908 DPRINTF(("committing page %"Z"u", pgno));
2909 next_pos = pos + size;
2910 iov[n].iov_len = size;
2911 iov[n].iov_base = (char *)dp;
2917 for (i = keep; ++i <= pagecount; ) {
2919 /* This is a page we skipped above */
2922 dl[j].mid = dp->mp_pgno;
2925 mdb_dpage_free(env, dp);
2930 txn->mt_dirty_room += i - j;
2936 mdb_txn_commit(MDB_txn *txn)
2942 if (txn == NULL || txn->mt_env == NULL)
2945 if (txn->mt_child) {
2946 rc = mdb_txn_commit(txn->mt_child);
2947 txn->mt_child = NULL;
2954 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2955 mdb_dbis_update(txn, 1);
2956 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2961 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2962 DPUTS("error flag is set, can't commit");
2964 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2969 if (txn->mt_parent) {
2970 MDB_txn *parent = txn->mt_parent;
2973 unsigned x, y, len, ps_len;
2975 /* Append our free list to parent's */
2976 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2979 mdb_midl_free(txn->mt_free_pgs);
2980 /* Failures after this must either undo the changes
2981 * to the parent or set MDB_TXN_ERROR in the parent.
2984 parent->mt_next_pgno = txn->mt_next_pgno;
2985 parent->mt_flags = txn->mt_flags;
2987 /* Merge our cursors into parent's and close them */
2988 mdb_cursors_close(txn, 1);
2990 /* Update parent's DB table. */
2991 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2992 parent->mt_numdbs = txn->mt_numdbs;
2993 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2994 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2995 for (i=2; i<txn->mt_numdbs; i++) {
2996 /* preserve parent's DB_NEW status */
2997 x = parent->mt_dbflags[i] & DB_NEW;
2998 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3001 dst = parent->mt_u.dirty_list;
3002 src = txn->mt_u.dirty_list;
3003 /* Remove anything in our dirty list from parent's spill list */
3004 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3006 pspill[0] = (pgno_t)-1;
3007 /* Mark our dirty pages as deleted in parent spill list */
3008 for (i=0, len=src[0].mid; ++i <= len; ) {
3009 MDB_ID pn = src[i].mid << 1;
3010 while (pn > pspill[x])
3012 if (pn == pspill[x]) {
3017 /* Squash deleted pagenums if we deleted any */
3018 for (x=y; ++x <= ps_len; )
3019 if (!(pspill[x] & 1))
3020 pspill[++y] = pspill[x];
3024 /* Find len = length of merging our dirty list with parent's */
3026 dst[0].mid = 0; /* simplify loops */
3027 if (parent->mt_parent) {
3028 len = x + src[0].mid;
3029 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3030 for (i = x; y && i; y--) {
3031 pgno_t yp = src[y].mid;
3032 while (yp < dst[i].mid)
3034 if (yp == dst[i].mid) {
3039 } else { /* Simplify the above for single-ancestor case */
3040 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3042 /* Merge our dirty list with parent's */
3044 for (i = len; y; dst[i--] = src[y--]) {
3045 pgno_t yp = src[y].mid;
3046 while (yp < dst[x].mid)
3047 dst[i--] = dst[x--];
3048 if (yp == dst[x].mid)
3049 free(dst[x--].mptr);
3051 mdb_tassert(txn, i == x);
3053 free(txn->mt_u.dirty_list);
3054 parent->mt_dirty_room = txn->mt_dirty_room;
3055 if (txn->mt_spill_pgs) {
3056 if (parent->mt_spill_pgs) {
3057 /* TODO: Prevent failure here, so parent does not fail */
3058 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3060 parent->mt_flags |= MDB_TXN_ERROR;
3061 mdb_midl_free(txn->mt_spill_pgs);
3062 mdb_midl_sort(parent->mt_spill_pgs);
3064 parent->mt_spill_pgs = txn->mt_spill_pgs;
3068 parent->mt_child = NULL;
3069 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3074 if (txn != env->me_txn) {
3075 DPUTS("attempt to commit unknown transaction");
3080 mdb_cursors_close(txn, 0);
3082 if (!txn->mt_u.dirty_list[0].mid &&
3083 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3086 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3087 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3089 /* Update DB root pointers */
3090 if (txn->mt_numdbs > 2) {
3094 data.mv_size = sizeof(MDB_db);
3096 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3097 for (i = 2; i < txn->mt_numdbs; i++) {
3098 if (txn->mt_dbflags[i] & DB_DIRTY) {
3099 data.mv_data = &txn->mt_dbs[i];
3100 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3107 rc = mdb_freelist_save(txn);
3111 mdb_midl_free(env->me_pghead);
3112 env->me_pghead = NULL;
3113 if (mdb_midl_shrink(&txn->mt_free_pgs))
3114 env->me_free_pgs = txn->mt_free_pgs;
3120 if ((rc = mdb_page_flush(txn, 0)) ||
3121 (rc = mdb_env_sync(env, 0)) ||
3122 (rc = mdb_env_write_meta(txn)))
3128 mdb_dbis_update(txn, 1);
3131 UNLOCK_MUTEX_W(env);
3141 /** Read the environment parameters of a DB environment before
3142 * mapping it into memory.
3143 * @param[in] env the environment handle
3144 * @param[out] meta address of where to store the meta information
3145 * @return 0 on success, non-zero on failure.
3148 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3154 enum { Size = sizeof(pbuf) };
3156 /* We don't know the page size yet, so use a minimum value.
3157 * Read both meta pages so we can use the latest one.
3160 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3164 memset(&ov, 0, sizeof(ov));
3166 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3167 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3170 rc = pread(env->me_fd, &pbuf, Size, off);
3173 if (rc == 0 && off == 0)
3175 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3176 DPRINTF(("read: %s", mdb_strerror(rc)));
3180 p = (MDB_page *)&pbuf;
3182 if (!F_ISSET(p->mp_flags, P_META)) {
3183 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3188 if (m->mm_magic != MDB_MAGIC) {
3189 DPUTS("meta has invalid magic");
3193 if (m->mm_version != MDB_DATA_VERSION) {
3194 DPRINTF(("database is version %u, expected version %u",
3195 m->mm_version, MDB_DATA_VERSION));
3196 return MDB_VERSION_MISMATCH;
3199 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3205 /** Write the environment parameters of a freshly created DB environment.
3206 * @param[in] env the environment handle
3207 * @param[out] meta address of where to store the meta information
3208 * @return 0 on success, non-zero on failure.
3211 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3219 memset(&ov, 0, sizeof(ov));
3220 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3222 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3225 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3226 len = pwrite(fd, ptr, size, pos); \
3227 rc = (len >= 0); } while(0)
3230 DPUTS("writing new meta page");
3232 psize = env->me_psize;
3234 meta->mm_magic = MDB_MAGIC;
3235 meta->mm_version = MDB_DATA_VERSION;
3236 meta->mm_mapsize = env->me_mapsize;
3237 meta->mm_psize = psize;
3238 meta->mm_last_pg = 1;
3239 meta->mm_flags = env->me_flags & 0xffff;
3240 meta->mm_flags |= MDB_INTEGERKEY;
3241 meta->mm_dbs[0].md_root = P_INVALID;
3242 meta->mm_dbs[1].md_root = P_INVALID;
3244 p = calloc(2, psize);
3246 p->mp_flags = P_META;
3247 *(MDB_meta *)METADATA(p) = *meta;
3249 q = (MDB_page *)((char *)p + psize);
3251 q->mp_flags = P_META;
3252 *(MDB_meta *)METADATA(q) = *meta;
3254 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3257 else if ((unsigned) len == psize * 2)
3265 /** Update the environment info to commit a transaction.
3266 * @param[in] txn the transaction that's being committed
3267 * @return 0 on success, non-zero on failure.
3270 mdb_env_write_meta(MDB_txn *txn)
3273 MDB_meta meta, metab, *mp;
3275 int rc, len, toggle;
3284 toggle = txn->mt_txnid & 1;
3285 DPRINTF(("writing meta page %d for root page %"Z"u",
3286 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3289 mp = env->me_metas[toggle];
3291 if (env->me_flags & MDB_WRITEMAP) {
3292 /* Persist any increases of mapsize config */
3293 if (env->me_mapsize > mp->mm_mapsize)
3294 mp->mm_mapsize = env->me_mapsize;
3295 mp->mm_dbs[0] = txn->mt_dbs[0];
3296 mp->mm_dbs[1] = txn->mt_dbs[1];
3297 mp->mm_last_pg = txn->mt_next_pgno - 1;
3298 mp->mm_txnid = txn->mt_txnid;
3299 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3300 unsigned meta_size = env->me_psize;
3301 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3304 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3305 if (meta_size < env->me_os_psize)
3306 meta_size += meta_size;
3311 if (MDB_MSYNC(ptr, meta_size, rc)) {
3318 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3319 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3321 ptr = (char *)&meta;
3322 if (env->me_mapsize > mp->mm_mapsize) {
3323 /* Persist any increases of mapsize config */
3324 meta.mm_mapsize = env->me_mapsize;
3325 off = offsetof(MDB_meta, mm_mapsize);
3327 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3329 len = sizeof(MDB_meta) - off;
3332 meta.mm_dbs[0] = txn->mt_dbs[0];
3333 meta.mm_dbs[1] = txn->mt_dbs[1];
3334 meta.mm_last_pg = txn->mt_next_pgno - 1;
3335 meta.mm_txnid = txn->mt_txnid;
3338 off += env->me_psize;
3341 /* Write to the SYNC fd */
3342 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3343 env->me_fd : env->me_mfd;
3346 memset(&ov, 0, sizeof(ov));
3348 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3352 rc = pwrite(mfd, ptr, len, off);
3355 rc = rc < 0 ? ErrCode() : EIO;
3356 DPUTS("write failed, disk error?");
3357 /* On a failure, the pagecache still contains the new data.
3358 * Write some old data back, to prevent it from being used.
3359 * Use the non-SYNC fd; we know it will fail anyway.
3361 meta.mm_last_pg = metab.mm_last_pg;
3362 meta.mm_txnid = metab.mm_txnid;
3364 memset(&ov, 0, sizeof(ov));
3366 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3368 r2 = pwrite(env->me_fd, ptr, len, off);
3369 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3372 env->me_flags |= MDB_FATAL_ERROR;
3376 /* Memory ordering issues are irrelevant; since the entire writer
3377 * is wrapped by wmutex, all of these changes will become visible
3378 * after the wmutex is unlocked. Since the DB is multi-version,
3379 * readers will get consistent data regardless of how fresh or
3380 * how stale their view of these values is.
3383 env->me_txns->mti_txnid = txn->mt_txnid;
3388 /** Check both meta pages to see which one is newer.
3389 * @param[in] env the environment handle
3390 * @return meta toggle (0 or 1).
3393 mdb_env_pick_meta(const MDB_env *env)
3395 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3399 mdb_env_create(MDB_env **env)
3403 e = calloc(1, sizeof(MDB_env));
3407 e->me_maxreaders = DEFAULT_READERS;
3408 e->me_maxdbs = e->me_numdbs = 2;
3409 e->me_fd = INVALID_HANDLE_VALUE;
3410 e->me_lfd = INVALID_HANDLE_VALUE;
3411 e->me_mfd = INVALID_HANDLE_VALUE;
3412 #ifdef MDB_USE_POSIX_SEM
3413 e->me_rmutex = SEM_FAILED;
3414 e->me_wmutex = SEM_FAILED;
3416 e->me_pid = getpid();
3417 GET_PAGESIZE(e->me_os_psize);
3418 VGMEMP_CREATE(e,0,0);
3424 mdb_env_map(MDB_env *env, void *addr, int newsize)
3427 unsigned int flags = env->me_flags;
3431 LONG sizelo, sizehi;
3432 sizelo = env->me_mapsize & 0xffffffff;
3433 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3435 /* Windows won't create mappings for zero length files.
3436 * Just allocate the maxsize right now.
3439 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3440 || !SetEndOfFile(env->me_fd)
3441 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3444 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3445 PAGE_READWRITE : PAGE_READONLY,
3446 sizehi, sizelo, NULL);
3449 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3450 FILE_MAP_WRITE : FILE_MAP_READ,
3451 0, 0, env->me_mapsize, addr);
3452 rc = env->me_map ? 0 : ErrCode();
3457 int prot = PROT_READ;
3458 if (flags & MDB_WRITEMAP) {
3460 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3463 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3465 if (env->me_map == MAP_FAILED) {
3470 if (flags & MDB_NORDAHEAD) {
3471 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3473 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3475 #ifdef POSIX_MADV_RANDOM
3476 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3477 #endif /* POSIX_MADV_RANDOM */
3478 #endif /* MADV_RANDOM */
3482 /* Can happen because the address argument to mmap() is just a
3483 * hint. mmap() can pick another, e.g. if the range is in use.
3484 * The MAP_FIXED flag would prevent that, but then mmap could
3485 * instead unmap existing pages to make room for the new map.
3487 if (addr && env->me_map != addr)
3488 return EBUSY; /* TODO: Make a new MDB_* error code? */
3490 p = (MDB_page *)env->me_map;
3491 env->me_metas[0] = METADATA(p);
3492 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3498 mdb_env_set_mapsize(MDB_env *env, size_t size)
3500 /* If env is already open, caller is responsible for making
3501 * sure there are no active txns.
3509 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3510 else if (size < env->me_mapsize) {
3511 /* If the configured size is smaller, make sure it's
3512 * still big enough. Silently round up to minimum if not.
3514 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3518 munmap(env->me_map, env->me_mapsize);
3519 env->me_mapsize = size;
3520 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3521 rc = mdb_env_map(env, old, 1);
3525 env->me_mapsize = size;
3527 env->me_maxpg = env->me_mapsize / env->me_psize;
3532 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3536 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3541 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3543 if (env->me_map || readers < 1)
3545 env->me_maxreaders = readers;
3550 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3552 if (!env || !readers)
3554 *readers = env->me_maxreaders;
3558 /** Further setup required for opening an MDB environment
3561 mdb_env_open2(MDB_env *env)
3563 unsigned int flags = env->me_flags;
3564 int i, newenv = 0, rc;
3568 /* See if we should use QueryLimited */
3570 if ((rc & 0xff) > 5)
3571 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3573 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3576 memset(&meta, 0, sizeof(meta));
3578 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3581 DPUTS("new mdbenv");
3583 env->me_psize = env->me_os_psize;
3584 if (env->me_psize > MAX_PAGESIZE)
3585 env->me_psize = MAX_PAGESIZE;
3587 env->me_psize = meta.mm_psize;
3590 /* Was a mapsize configured? */
3591 if (!env->me_mapsize) {
3592 /* If this is a new environment, take the default,
3593 * else use the size recorded in the existing env.
3595 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3596 } else if (env->me_mapsize < meta.mm_mapsize) {
3597 /* If the configured size is smaller, make sure it's
3598 * still big enough. Silently round up to minimum if not.
3600 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3601 if (env->me_mapsize < minsize)
3602 env->me_mapsize = minsize;
3605 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3610 if (flags & MDB_FIXEDMAP)
3611 meta.mm_address = env->me_map;
3612 i = mdb_env_init_meta(env, &meta);
3613 if (i != MDB_SUCCESS) {
3618 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3619 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3621 #if !(MDB_MAXKEYSIZE)
3622 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3624 env->me_maxpg = env->me_mapsize / env->me_psize;
3628 int toggle = mdb_env_pick_meta(env);
3629 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3631 DPRINTF(("opened database version %u, pagesize %u",
3632 env->me_metas[0]->mm_version, env->me_psize));
3633 DPRINTF(("using meta page %d", toggle));
3634 DPRINTF(("depth: %u", db->md_depth));
3635 DPRINTF(("entries: %"Z"u", db->md_entries));
3636 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3637 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3638 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3639 DPRINTF(("root: %"Z"u", db->md_root));
3647 /** Release a reader thread's slot in the reader lock table.
3648 * This function is called automatically when a thread exits.
3649 * @param[in] ptr This points to the slot in the reader lock table.
3652 mdb_env_reader_dest(void *ptr)
3654 MDB_reader *reader = ptr;
3660 /** Junk for arranging thread-specific callbacks on Windows. This is
3661 * necessarily platform and compiler-specific. Windows supports up
3662 * to 1088 keys. Let's assume nobody opens more than 64 environments
3663 * in a single process, for now. They can override this if needed.
3665 #ifndef MAX_TLS_KEYS
3666 #define MAX_TLS_KEYS 64
3668 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3669 static int mdb_tls_nkeys;
3671 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3675 case DLL_PROCESS_ATTACH: break;
3676 case DLL_THREAD_ATTACH: break;
3677 case DLL_THREAD_DETACH:
3678 for (i=0; i<mdb_tls_nkeys; i++) {
3679 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3680 mdb_env_reader_dest(r);
3683 case DLL_PROCESS_DETACH: break;
3688 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3690 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3694 /* Force some symbol references.
3695 * _tls_used forces the linker to create the TLS directory if not already done
3696 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3698 #pragma comment(linker, "/INCLUDE:_tls_used")
3699 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3700 #pragma const_seg(".CRT$XLB")
3701 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3702 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3705 #pragma comment(linker, "/INCLUDE:__tls_used")
3706 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3707 #pragma data_seg(".CRT$XLB")
3708 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3710 #endif /* WIN 32/64 */
3711 #endif /* !__GNUC__ */
3714 /** Downgrade the exclusive lock on the region back to shared */
3716 mdb_env_share_locks(MDB_env *env, int *excl)
3718 int rc = 0, toggle = mdb_env_pick_meta(env);
3720 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3725 /* First acquire a shared lock. The Unlock will
3726 * then release the existing exclusive lock.
3728 memset(&ov, 0, sizeof(ov));
3729 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3732 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3738 struct flock lock_info;
3739 /* The shared lock replaces the existing lock */
3740 memset((void *)&lock_info, 0, sizeof(lock_info));
3741 lock_info.l_type = F_RDLCK;
3742 lock_info.l_whence = SEEK_SET;
3743 lock_info.l_start = 0;
3744 lock_info.l_len = 1;
3745 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3746 (rc = ErrCode()) == EINTR) ;
3747 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3754 /** Try to get exlusive lock, otherwise shared.
3755 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3758 mdb_env_excl_lock(MDB_env *env, int *excl)
3762 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3766 memset(&ov, 0, sizeof(ov));
3767 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3774 struct flock lock_info;
3775 memset((void *)&lock_info, 0, sizeof(lock_info));
3776 lock_info.l_type = F_WRLCK;
3777 lock_info.l_whence = SEEK_SET;
3778 lock_info.l_start = 0;
3779 lock_info.l_len = 1;
3780 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3781 (rc = ErrCode()) == EINTR) ;
3785 # ifdef MDB_USE_POSIX_SEM
3786 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3789 lock_info.l_type = F_RDLCK;
3790 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3791 (rc = ErrCode()) == EINTR) ;
3801 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3803 * @(#) $Revision: 5.1 $
3804 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3805 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3807 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3811 * Please do not copyright this code. This code is in the public domain.
3813 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3814 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3815 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3816 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3817 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3818 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3819 * PERFORMANCE OF THIS SOFTWARE.
3822 * chongo <Landon Curt Noll> /\oo/\
3823 * http://www.isthe.com/chongo/
3825 * Share and Enjoy! :-)
3828 typedef unsigned long long mdb_hash_t;
3829 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3831 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3832 * @param[in] val value to hash
3833 * @param[in] hval initial value for hash
3834 * @return 64 bit hash
3836 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3837 * hval arg on the first call.
3840 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3842 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3843 unsigned char *end = s + val->mv_size;
3845 * FNV-1a hash each octet of the string
3848 /* xor the bottom with the current octet */
3849 hval ^= (mdb_hash_t)*s++;
3851 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3852 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3853 (hval << 7) + (hval << 8) + (hval << 40);
3855 /* return our new hash value */
3859 /** Hash the string and output the encoded hash.
3860 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3861 * very short name limits. We don't care about the encoding being reversible,
3862 * we just want to preserve as many bits of the input as possible in a
3863 * small printable string.
3864 * @param[in] str string to hash
3865 * @param[out] encbuf an array of 11 chars to hold the hash
3867 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3870 mdb_pack85(unsigned long l, char *out)
3874 for (i=0; i<5; i++) {
3875 *out++ = mdb_a85[l % 85];
3881 mdb_hash_enc(MDB_val *val, char *encbuf)
3883 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3885 mdb_pack85(h, encbuf);
3886 mdb_pack85(h>>32, encbuf+5);
3891 /** Open and/or initialize the lock region for the environment.
3892 * @param[in] env The MDB environment.
3893 * @param[in] lpath The pathname of the file used for the lock region.
3894 * @param[in] mode The Unix permissions for the file, if we create it.
3895 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3896 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3897 * @return 0 on success, non-zero on failure.
3900 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3903 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3905 # define MDB_ERRCODE_ROFS EROFS
3906 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3907 # define MDB_CLOEXEC O_CLOEXEC
3910 # define MDB_CLOEXEC 0
3917 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3918 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3919 FILE_ATTRIBUTE_NORMAL, NULL);
3921 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3923 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3925 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3930 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3931 /* Lose record locks when exec*() */
3932 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3933 fcntl(env->me_lfd, F_SETFD, fdflags);
3936 if (!(env->me_flags & MDB_NOTLS)) {
3937 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3940 env->me_flags |= MDB_ENV_TXKEY;
3942 /* Windows TLS callbacks need help finding their TLS info. */
3943 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3947 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3951 /* Try to get exclusive lock. If we succeed, then
3952 * nobody is using the lock region and we should initialize it.
3954 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3957 size = GetFileSize(env->me_lfd, NULL);
3959 size = lseek(env->me_lfd, 0, SEEK_END);
3960 if (size == -1) goto fail_errno;
3962 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3963 if (size < rsize && *excl > 0) {
3965 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3966 || !SetEndOfFile(env->me_lfd))
3969 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3973 size = rsize - sizeof(MDB_txninfo);
3974 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3979 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3981 if (!mh) goto fail_errno;
3982 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3984 if (!env->me_txns) goto fail_errno;
3986 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3988 if (m == MAP_FAILED) goto fail_errno;
3994 BY_HANDLE_FILE_INFORMATION stbuf;
4003 if (!mdb_sec_inited) {
4004 InitializeSecurityDescriptor(&mdb_null_sd,
4005 SECURITY_DESCRIPTOR_REVISION);
4006 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4007 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4008 mdb_all_sa.bInheritHandle = FALSE;
4009 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4012 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4013 idbuf.volume = stbuf.dwVolumeSerialNumber;
4014 idbuf.nhigh = stbuf.nFileIndexHigh;
4015 idbuf.nlow = stbuf.nFileIndexLow;
4016 val.mv_data = &idbuf;
4017 val.mv_size = sizeof(idbuf);
4018 mdb_hash_enc(&val, encbuf);
4019 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4020 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4021 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4022 if (!env->me_rmutex) goto fail_errno;
4023 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4024 if (!env->me_wmutex) goto fail_errno;
4025 #elif defined(MDB_USE_POSIX_SEM)
4034 #if defined(__NetBSD__)
4035 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4037 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4038 idbuf.dev = stbuf.st_dev;
4039 idbuf.ino = stbuf.st_ino;
4040 val.mv_data = &idbuf;
4041 val.mv_size = sizeof(idbuf);
4042 mdb_hash_enc(&val, encbuf);
4043 #ifdef MDB_SHORT_SEMNAMES
4044 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4046 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4047 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4048 /* Clean up after a previous run, if needed: Try to
4049 * remove both semaphores before doing anything else.
4051 sem_unlink(env->me_txns->mti_rmname);
4052 sem_unlink(env->me_txns->mti_wmname);
4053 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4054 O_CREAT|O_EXCL, mode, 1);
4055 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4056 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4057 O_CREAT|O_EXCL, mode, 1);
4058 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4059 #else /* MDB_USE_POSIX_SEM */
4060 pthread_mutexattr_t mattr;
4062 if ((rc = pthread_mutexattr_init(&mattr))
4063 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4064 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4065 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4067 pthread_mutexattr_destroy(&mattr);
4068 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4070 env->me_txns->mti_magic = MDB_MAGIC;
4071 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4072 env->me_txns->mti_txnid = 0;
4073 env->me_txns->mti_numreaders = 0;
4076 if (env->me_txns->mti_magic != MDB_MAGIC) {
4077 DPUTS("lock region has invalid magic");
4081 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4082 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4083 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4084 rc = MDB_VERSION_MISMATCH;
4088 if (rc && rc != EACCES && rc != EAGAIN) {
4092 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4093 if (!env->me_rmutex) goto fail_errno;
4094 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4095 if (!env->me_wmutex) goto fail_errno;
4096 #elif defined(MDB_USE_POSIX_SEM)
4097 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4098 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4099 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4100 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4111 /** The name of the lock file in the DB environment */
4112 #define LOCKNAME "/lock.mdb"
4113 /** The name of the data file in the DB environment */
4114 #define DATANAME "/data.mdb"
4115 /** The suffix of the lock file when no subdir is used */
4116 #define LOCKSUFF "-lock"
4117 /** Only a subset of the @ref mdb_env flags can be changed
4118 * at runtime. Changing other flags requires closing the
4119 * environment and re-opening it with the new flags.
4121 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4122 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4123 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4125 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4126 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4130 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4132 int oflags, rc, len, excl = -1;
4133 char *lpath, *dpath;
4135 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4139 if (flags & MDB_NOSUBDIR) {
4140 rc = len + sizeof(LOCKSUFF) + len + 1;
4142 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4147 if (flags & MDB_NOSUBDIR) {
4148 dpath = lpath + len + sizeof(LOCKSUFF);
4149 sprintf(lpath, "%s" LOCKSUFF, path);
4150 strcpy(dpath, path);
4152 dpath = lpath + len + sizeof(LOCKNAME);
4153 sprintf(lpath, "%s" LOCKNAME, path);
4154 sprintf(dpath, "%s" DATANAME, path);
4158 flags |= env->me_flags;
4159 if (flags & MDB_RDONLY) {
4160 /* silently ignore WRITEMAP when we're only getting read access */
4161 flags &= ~MDB_WRITEMAP;
4163 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4164 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4167 env->me_flags = flags |= MDB_ENV_ACTIVE;
4171 env->me_path = strdup(path);
4172 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4173 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4174 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4179 /* For RDONLY, get lockfile after we know datafile exists */
4180 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4181 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4187 if (F_ISSET(flags, MDB_RDONLY)) {
4188 oflags = GENERIC_READ;
4189 len = OPEN_EXISTING;
4191 oflags = GENERIC_READ|GENERIC_WRITE;
4194 mode = FILE_ATTRIBUTE_NORMAL;
4195 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4196 NULL, len, mode, NULL);
4198 if (F_ISSET(flags, MDB_RDONLY))
4201 oflags = O_RDWR | O_CREAT;
4203 env->me_fd = open(dpath, oflags, mode);
4205 if (env->me_fd == INVALID_HANDLE_VALUE) {
4210 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4211 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4216 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4217 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4218 env->me_mfd = env->me_fd;
4220 /* Synchronous fd for meta writes. Needed even with
4221 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4224 len = OPEN_EXISTING;
4225 env->me_mfd = CreateFile(dpath, oflags,
4226 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4227 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4230 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4232 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4237 DPRINTF(("opened dbenv %p", (void *) env));
4239 rc = mdb_env_share_locks(env, &excl);
4243 if (!((flags & MDB_RDONLY) ||
4244 (env->me_pbuf = calloc(1, env->me_psize))))
4250 mdb_env_close0(env, excl);
4256 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4258 mdb_env_close0(MDB_env *env, int excl)
4262 if (!(env->me_flags & MDB_ENV_ACTIVE))
4265 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4266 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4267 free(env->me_dbxs[i].md_name.mv_data);
4270 free(env->me_dbflags);
4273 free(env->me_dirty_list);
4274 mdb_midl_free(env->me_free_pgs);
4276 if (env->me_flags & MDB_ENV_TXKEY) {
4277 pthread_key_delete(env->me_txkey);
4279 /* Delete our key from the global list */
4280 for (i=0; i<mdb_tls_nkeys; i++)
4281 if (mdb_tls_keys[i] == env->me_txkey) {
4282 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4290 munmap(env->me_map, env->me_mapsize);
4292 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4293 (void) close(env->me_mfd);
4294 if (env->me_fd != INVALID_HANDLE_VALUE)
4295 (void) close(env->me_fd);
4297 MDB_PID_T pid = env->me_pid;
4298 /* Clearing readers is done in this function because
4299 * me_txkey with its destructor must be disabled first.
4301 for (i = env->me_numreaders; --i >= 0; )
4302 if (env->me_txns->mti_readers[i].mr_pid == pid)
4303 env->me_txns->mti_readers[i].mr_pid = 0;
4305 if (env->me_rmutex) {
4306 CloseHandle(env->me_rmutex);
4307 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4309 /* Windows automatically destroys the mutexes when
4310 * the last handle closes.
4312 #elif defined(MDB_USE_POSIX_SEM)
4313 if (env->me_rmutex != SEM_FAILED) {
4314 sem_close(env->me_rmutex);
4315 if (env->me_wmutex != SEM_FAILED)
4316 sem_close(env->me_wmutex);
4317 /* If we have the filelock: If we are the
4318 * only remaining user, clean up semaphores.
4321 mdb_env_excl_lock(env, &excl);
4323 sem_unlink(env->me_txns->mti_rmname);
4324 sem_unlink(env->me_txns->mti_wmname);
4328 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4330 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4333 /* Unlock the lockfile. Windows would have unlocked it
4334 * after closing anyway, but not necessarily at once.
4336 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4339 (void) close(env->me_lfd);
4342 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4346 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4348 MDB_txn *txn = NULL;
4354 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4358 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4361 /* Do the lock/unlock of the reader mutex before starting the
4362 * write txn. Otherwise other read txns could block writers.
4364 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4369 /* We must start the actual read txn after blocking writers */
4370 mdb_txn_reset0(txn, "reset-stage1");
4372 /* Temporarily block writers until we snapshot the meta pages */
4375 rc = mdb_txn_renew0(txn);
4377 UNLOCK_MUTEX_W(env);
4382 wsize = env->me_psize * 2;
4386 DO_WRITE(rc, fd, ptr, w2, len);
4390 } else if (len > 0) {
4396 /* Non-blocking or async handles are not supported */
4402 UNLOCK_MUTEX_W(env);
4407 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4409 if (wsize > MAX_WRITE)
4413 DO_WRITE(rc, fd, ptr, w2, len);
4417 } else if (len > 0) {
4434 mdb_env_copy(MDB_env *env, const char *path)
4438 HANDLE newfd = INVALID_HANDLE_VALUE;
4440 if (env->me_flags & MDB_NOSUBDIR) {
4441 lpath = (char *)path;
4444 len += sizeof(DATANAME);
4445 lpath = malloc(len);
4448 sprintf(lpath, "%s" DATANAME, path);
4451 /* The destination path must exist, but the destination file must not.
4452 * We don't want the OS to cache the writes, since the source data is
4453 * already in the OS cache.
4456 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4457 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4459 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4461 if (newfd == INVALID_HANDLE_VALUE) {
4467 /* Set O_DIRECT if the file system supports it */
4468 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4469 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4471 #ifdef F_NOCACHE /* __APPLE__ */
4472 rc = fcntl(newfd, F_NOCACHE, 1);
4479 rc = mdb_env_copyfd(env, newfd);
4482 if (!(env->me_flags & MDB_NOSUBDIR))
4484 if (newfd != INVALID_HANDLE_VALUE)
4485 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4492 mdb_env_close(MDB_env *env)
4499 VGMEMP_DESTROY(env);
4500 while ((dp = env->me_dpages) != NULL) {
4501 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4502 env->me_dpages = dp->mp_next;
4506 mdb_env_close0(env, 0);
4510 /** Compare two items pointing at aligned size_t's */
4512 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4514 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4515 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4518 /** Compare two items pointing at aligned unsigned int's */
4520 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4522 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4523 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4526 /** Compare two items pointing at unsigned ints of unknown alignment.
4527 * Nodes and keys are guaranteed to be 2-byte aligned.
4530 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4532 #if BYTE_ORDER == LITTLE_ENDIAN
4533 unsigned short *u, *c;
4536 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4537 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4540 } while(!x && u > (unsigned short *)a->mv_data);
4543 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4547 /** Compare two items lexically */
4549 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4556 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4562 diff = memcmp(a->mv_data, b->mv_data, len);
4563 return diff ? diff : len_diff<0 ? -1 : len_diff;
4566 /** Compare two items in reverse byte order */
4568 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4570 const unsigned char *p1, *p2, *p1_lim;
4574 p1_lim = (const unsigned char *)a->mv_data;
4575 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4576 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4578 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4584 while (p1 > p1_lim) {
4585 diff = *--p1 - *--p2;
4589 return len_diff<0 ? -1 : len_diff;
4592 /** Search for key within a page, using binary search.
4593 * Returns the smallest entry larger or equal to the key.
4594 * If exactp is non-null, stores whether the found entry was an exact match
4595 * in *exactp (1 or 0).
4596 * Updates the cursor index with the index of the found entry.
4597 * If no entry larger or equal to the key is found, returns NULL.
4600 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4602 unsigned int i = 0, nkeys;
4605 MDB_page *mp = mc->mc_pg[mc->mc_top];
4606 MDB_node *node = NULL;
4611 nkeys = NUMKEYS(mp);
4613 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4614 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4617 low = IS_LEAF(mp) ? 0 : 1;
4619 cmp = mc->mc_dbx->md_cmp;
4621 /* Branch pages have no data, so if using integer keys,
4622 * alignment is guaranteed. Use faster mdb_cmp_int.
4624 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4625 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4632 nodekey.mv_size = mc->mc_db->md_pad;
4633 node = NODEPTR(mp, 0); /* fake */
4634 while (low <= high) {
4635 i = (low + high) >> 1;
4636 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4637 rc = cmp(key, &nodekey);
4638 DPRINTF(("found leaf index %u [%s], rc = %i",
4639 i, DKEY(&nodekey), rc));
4648 while (low <= high) {
4649 i = (low + high) >> 1;
4651 node = NODEPTR(mp, i);
4652 nodekey.mv_size = NODEKSZ(node);
4653 nodekey.mv_data = NODEKEY(node);
4655 rc = cmp(key, &nodekey);
4658 DPRINTF(("found leaf index %u [%s], rc = %i",
4659 i, DKEY(&nodekey), rc));
4661 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4662 i, DKEY(&nodekey), NODEPGNO(node), rc));
4673 if (rc > 0) { /* Found entry is less than the key. */
4674 i++; /* Skip to get the smallest entry larger than key. */
4676 node = NODEPTR(mp, i);
4679 *exactp = (rc == 0 && nkeys > 0);
4680 /* store the key index */
4681 mc->mc_ki[mc->mc_top] = i;
4683 /* There is no entry larger or equal to the key. */
4686 /* nodeptr is fake for LEAF2 */
4692 mdb_cursor_adjust(MDB_cursor *mc, func)
4696 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4697 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4704 /** Pop a page off the top of the cursor's stack. */
4706 mdb_cursor_pop(MDB_cursor *mc)
4710 MDB_page *top = mc->mc_pg[mc->mc_top];
4716 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4717 DDBI(mc), (void *) mc));
4721 /** Push a page onto the top of the cursor's stack. */
4723 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4725 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4726 DDBI(mc), (void *) mc));
4728 if (mc->mc_snum >= CURSOR_STACK) {
4729 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4730 return MDB_CURSOR_FULL;
4733 mc->mc_top = mc->mc_snum++;
4734 mc->mc_pg[mc->mc_top] = mp;
4735 mc->mc_ki[mc->mc_top] = 0;
4740 /** Find the address of the page corresponding to a given page number.
4741 * @param[in] txn the transaction for this access.
4742 * @param[in] pgno the page number for the page to retrieve.
4743 * @param[out] ret address of a pointer where the page's address will be stored.
4744 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4745 * @return 0 on success, non-zero on failure.
4748 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4750 MDB_env *env = txn->mt_env;
4754 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4758 MDB_ID2L dl = tx2->mt_u.dirty_list;
4760 /* Spilled pages were dirtied in this txn and flushed
4761 * because the dirty list got full. Bring this page
4762 * back in from the map (but don't unspill it here,
4763 * leave that unless page_touch happens again).
4765 if (tx2->mt_spill_pgs) {
4766 MDB_ID pn = pgno << 1;
4767 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4768 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4769 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4774 unsigned x = mdb_mid2l_search(dl, pgno);
4775 if (x <= dl[0].mid && dl[x].mid == pgno) {
4781 } while ((tx2 = tx2->mt_parent) != NULL);
4784 if (pgno < txn->mt_next_pgno) {
4786 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4788 DPRINTF(("page %"Z"u not found", pgno));
4789 txn->mt_flags |= MDB_TXN_ERROR;
4790 return MDB_PAGE_NOTFOUND;
4800 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4801 * The cursor is at the root page, set up the rest of it.
4804 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4806 MDB_page *mp = mc->mc_pg[mc->mc_top];
4810 while (IS_BRANCH(mp)) {
4814 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4815 mdb_cassert(mc, NUMKEYS(mp) > 1);
4816 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4818 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4820 if (flags & MDB_PS_LAST)
4821 i = NUMKEYS(mp) - 1;
4824 node = mdb_node_search(mc, key, &exact);
4826 i = NUMKEYS(mp) - 1;
4828 i = mc->mc_ki[mc->mc_top];
4830 mdb_cassert(mc, i > 0);
4834 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4837 mdb_cassert(mc, i < NUMKEYS(mp));
4838 node = NODEPTR(mp, i);
4840 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4843 mc->mc_ki[mc->mc_top] = i;
4844 if ((rc = mdb_cursor_push(mc, mp)))
4847 if (flags & MDB_PS_MODIFY) {
4848 if ((rc = mdb_page_touch(mc)) != 0)
4850 mp = mc->mc_pg[mc->mc_top];
4855 DPRINTF(("internal error, index points to a %02X page!?",
4857 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4858 return MDB_CORRUPTED;
4861 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4862 key ? DKEY(key) : "null"));
4863 mc->mc_flags |= C_INITIALIZED;
4864 mc->mc_flags &= ~C_EOF;
4869 /** Search for the lowest key under the current branch page.
4870 * This just bypasses a NUMKEYS check in the current page
4871 * before calling mdb_page_search_root(), because the callers
4872 * are all in situations where the current page is known to
4876 mdb_page_search_lowest(MDB_cursor *mc)
4878 MDB_page *mp = mc->mc_pg[mc->mc_top];
4879 MDB_node *node = NODEPTR(mp, 0);
4882 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4885 mc->mc_ki[mc->mc_top] = 0;
4886 if ((rc = mdb_cursor_push(mc, mp)))
4888 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4891 /** Search for the page a given key should be in.
4892 * Push it and its parent pages on the cursor stack.
4893 * @param[in,out] mc the cursor for this operation.
4894 * @param[in] key the key to search for, or NULL for first/last page.
4895 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4896 * are touched (updated with new page numbers).
4897 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4898 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4899 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4900 * @return 0 on success, non-zero on failure.
4903 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4908 /* Make sure the txn is still viable, then find the root from
4909 * the txn's db table and set it as the root of the cursor's stack.
4911 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4912 DPUTS("transaction has failed, must abort");
4915 /* Make sure we're using an up-to-date root */
4916 if (*mc->mc_dbflag & DB_STALE) {
4918 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4919 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4926 MDB_node *leaf = mdb_node_search(&mc2,
4927 &mc->mc_dbx->md_name, &exact);
4929 return MDB_NOTFOUND;
4930 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4933 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4935 /* The txn may not know this DBI, or another process may
4936 * have dropped and recreated the DB with other flags.
4938 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4939 return MDB_INCOMPATIBLE;
4940 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4942 *mc->mc_dbflag &= ~DB_STALE;
4944 root = mc->mc_db->md_root;
4946 if (root == P_INVALID) { /* Tree is empty. */
4947 DPUTS("tree is empty");
4948 return MDB_NOTFOUND;
4952 mdb_cassert(mc, root > 1);
4953 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4954 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4960 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4961 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4963 if (flags & MDB_PS_MODIFY) {
4964 if ((rc = mdb_page_touch(mc)))
4968 if (flags & MDB_PS_ROOTONLY)
4971 return mdb_page_search_root(mc, key, flags);
4975 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4977 MDB_txn *txn = mc->mc_txn;
4978 pgno_t pg = mp->mp_pgno;
4979 unsigned x = 0, ovpages = mp->mp_pages;
4980 MDB_env *env = txn->mt_env;
4981 MDB_IDL sl = txn->mt_spill_pgs;
4982 MDB_ID pn = pg << 1;
4985 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4986 /* If the page is dirty or on the spill list we just acquired it,
4987 * so we should give it back to our current free list, if any.
4988 * Otherwise put it onto the list of pages we freed in this txn.
4990 * Won't create me_pghead: me_pglast must be inited along with it.
4991 * Unsupported in nested txns: They would need to hide the page
4992 * range in ancestor txns' dirty and spilled lists.
4994 if (env->me_pghead &&
4996 ((mp->mp_flags & P_DIRTY) ||
4997 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5001 MDB_ID2 *dl, ix, iy;
5002 rc = mdb_midl_need(&env->me_pghead, ovpages);
5005 if (!(mp->mp_flags & P_DIRTY)) {
5006 /* This page is no longer spilled */
5013 /* Remove from dirty list */
5014 dl = txn->mt_u.dirty_list;
5016 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5022 mdb_cassert(mc, x > 1);
5024 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5025 txn->mt_flags |= MDB_TXN_ERROR;
5026 return MDB_CORRUPTED;
5029 if (!(env->me_flags & MDB_WRITEMAP))
5030 mdb_dpage_free(env, mp);
5032 /* Insert in me_pghead */
5033 mop = env->me_pghead;
5034 j = mop[0] + ovpages;
5035 for (i = mop[0]; i && mop[i] < pg; i--)
5041 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5045 mc->mc_db->md_overflow_pages -= ovpages;
5049 /** Return the data associated with a given node.
5050 * @param[in] txn The transaction for this operation.
5051 * @param[in] leaf The node being read.
5052 * @param[out] data Updated to point to the node's data.
5053 * @return 0 on success, non-zero on failure.
5056 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5058 MDB_page *omp; /* overflow page */
5062 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5063 data->mv_size = NODEDSZ(leaf);
5064 data->mv_data = NODEDATA(leaf);
5068 /* Read overflow data.
5070 data->mv_size = NODEDSZ(leaf);
5071 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5072 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5073 DPRINTF(("read overflow page %"Z"u failed", pgno));
5076 data->mv_data = METADATA(omp);
5082 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5083 MDB_val *key, MDB_val *data)
5090 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5092 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5095 if (txn->mt_flags & MDB_TXN_ERROR)
5098 mdb_cursor_init(&mc, txn, dbi, &mx);
5099 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5102 /** Find a sibling for a page.
5103 * Replaces the page at the top of the cursor's stack with the
5104 * specified sibling, if one exists.
5105 * @param[in] mc The cursor for this operation.
5106 * @param[in] move_right Non-zero if the right sibling is requested,
5107 * otherwise the left sibling.
5108 * @return 0 on success, non-zero on failure.
5111 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5117 if (mc->mc_snum < 2) {
5118 return MDB_NOTFOUND; /* root has no siblings */
5122 DPRINTF(("parent page is page %"Z"u, index %u",
5123 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5125 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5126 : (mc->mc_ki[mc->mc_top] == 0)) {
5127 DPRINTF(("no more keys left, moving to %s sibling",
5128 move_right ? "right" : "left"));
5129 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5130 /* undo cursor_pop before returning */
5137 mc->mc_ki[mc->mc_top]++;
5139 mc->mc_ki[mc->mc_top]--;
5140 DPRINTF(("just moving to %s index key %u",
5141 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5143 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5145 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5146 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5147 /* mc will be inconsistent if caller does mc_snum++ as above */
5148 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5152 mdb_cursor_push(mc, mp);
5154 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5159 /** Move the cursor to the next data item. */
5161 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5167 if (mc->mc_flags & C_EOF) {
5168 return MDB_NOTFOUND;
5171 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5173 mp = mc->mc_pg[mc->mc_top];
5175 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5176 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5177 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5178 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5179 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5180 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5181 if (rc == MDB_SUCCESS)
5182 MDB_GET_KEY(leaf, key);
5187 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5188 if (op == MDB_NEXT_DUP)
5189 return MDB_NOTFOUND;
5193 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5194 mdb_dbg_pgno(mp), (void *) mc));
5195 if (mc->mc_flags & C_DEL)
5198 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5199 DPUTS("=====> move to next sibling page");
5200 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5201 mc->mc_flags |= C_EOF;
5204 mp = mc->mc_pg[mc->mc_top];
5205 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5207 mc->mc_ki[mc->mc_top]++;
5210 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5211 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5214 key->mv_size = mc->mc_db->md_pad;
5215 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5219 mdb_cassert(mc, IS_LEAF(mp));
5220 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5222 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5223 mdb_xcursor_init1(mc, leaf);
5226 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5229 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5230 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5231 if (rc != MDB_SUCCESS)
5236 MDB_GET_KEY(leaf, key);
5240 /** Move the cursor to the previous data item. */
5242 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5248 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5250 mp = mc->mc_pg[mc->mc_top];
5252 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5253 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5254 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5255 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5256 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5257 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5258 if (rc == MDB_SUCCESS)
5259 MDB_GET_KEY(leaf, key);
5263 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5264 if (op == MDB_PREV_DUP)
5265 return MDB_NOTFOUND;
5270 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5271 mdb_dbg_pgno(mp), (void *) mc));
5273 if (mc->mc_ki[mc->mc_top] == 0) {
5274 DPUTS("=====> move to prev sibling page");
5275 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5278 mp = mc->mc_pg[mc->mc_top];
5279 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5280 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5282 mc->mc_ki[mc->mc_top]--;
5284 mc->mc_flags &= ~C_EOF;
5286 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5287 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5290 key->mv_size = mc->mc_db->md_pad;
5291 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5295 mdb_cassert(mc, IS_LEAF(mp));
5296 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5298 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5299 mdb_xcursor_init1(mc, leaf);
5302 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5305 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5306 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5307 if (rc != MDB_SUCCESS)
5312 MDB_GET_KEY(leaf, key);
5316 /** Set the cursor on a specific data item. */
5318 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5319 MDB_cursor_op op, int *exactp)
5323 MDB_node *leaf = NULL;
5326 if (key->mv_size == 0)
5327 return MDB_BAD_VALSIZE;
5330 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5332 /* See if we're already on the right page */
5333 if (mc->mc_flags & C_INITIALIZED) {
5336 mp = mc->mc_pg[mc->mc_top];
5338 mc->mc_ki[mc->mc_top] = 0;
5339 return MDB_NOTFOUND;
5341 if (mp->mp_flags & P_LEAF2) {
5342 nodekey.mv_size = mc->mc_db->md_pad;
5343 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5345 leaf = NODEPTR(mp, 0);
5346 MDB_GET_KEY2(leaf, nodekey);
5348 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5350 /* Probably happens rarely, but first node on the page
5351 * was the one we wanted.
5353 mc->mc_ki[mc->mc_top] = 0;
5360 unsigned int nkeys = NUMKEYS(mp);
5362 if (mp->mp_flags & P_LEAF2) {
5363 nodekey.mv_data = LEAF2KEY(mp,
5364 nkeys-1, nodekey.mv_size);
5366 leaf = NODEPTR(mp, nkeys-1);
5367 MDB_GET_KEY2(leaf, nodekey);
5369 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5371 /* last node was the one we wanted */
5372 mc->mc_ki[mc->mc_top] = nkeys-1;
5378 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5379 /* This is definitely the right page, skip search_page */
5380 if (mp->mp_flags & P_LEAF2) {
5381 nodekey.mv_data = LEAF2KEY(mp,
5382 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5384 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5385 MDB_GET_KEY2(leaf, nodekey);
5387 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5389 /* current node was the one we wanted */
5399 /* If any parents have right-sibs, search.
5400 * Otherwise, there's nothing further.
5402 for (i=0; i<mc->mc_top; i++)
5404 NUMKEYS(mc->mc_pg[i])-1)
5406 if (i == mc->mc_top) {
5407 /* There are no other pages */
5408 mc->mc_ki[mc->mc_top] = nkeys;
5409 return MDB_NOTFOUND;
5413 /* There are no other pages */
5414 mc->mc_ki[mc->mc_top] = 0;
5415 if (op == MDB_SET_RANGE) {
5419 return MDB_NOTFOUND;
5423 rc = mdb_page_search(mc, key, 0);
5424 if (rc != MDB_SUCCESS)
5427 mp = mc->mc_pg[mc->mc_top];
5428 mdb_cassert(mc, IS_LEAF(mp));
5431 leaf = mdb_node_search(mc, key, exactp);
5432 if (exactp != NULL && !*exactp) {
5433 /* MDB_SET specified and not an exact match. */
5434 return MDB_NOTFOUND;
5438 DPUTS("===> inexact leaf not found, goto sibling");
5439 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5440 return rc; /* no entries matched */
5441 mp = mc->mc_pg[mc->mc_top];
5442 mdb_cassert(mc, IS_LEAF(mp));
5443 leaf = NODEPTR(mp, 0);
5447 mc->mc_flags |= C_INITIALIZED;
5448 mc->mc_flags &= ~C_EOF;
5451 key->mv_size = mc->mc_db->md_pad;
5452 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5456 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5457 mdb_xcursor_init1(mc, leaf);
5460 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5461 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5462 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5465 if (op == MDB_GET_BOTH) {
5471 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5472 if (rc != MDB_SUCCESS)
5475 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5477 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5479 rc = mc->mc_dbx->md_dcmp(data, &d2);
5481 if (op == MDB_GET_BOTH || rc > 0)
5482 return MDB_NOTFOUND;
5489 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5490 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5495 /* The key already matches in all other cases */
5496 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5497 MDB_GET_KEY(leaf, key);
5498 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5503 /** Move the cursor to the first item in the database. */
5505 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5511 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5513 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5514 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5515 if (rc != MDB_SUCCESS)
5518 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5520 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5521 mc->mc_flags |= C_INITIALIZED;
5522 mc->mc_flags &= ~C_EOF;
5524 mc->mc_ki[mc->mc_top] = 0;
5526 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5527 key->mv_size = mc->mc_db->md_pad;
5528 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5533 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5534 mdb_xcursor_init1(mc, leaf);
5535 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5539 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5543 MDB_GET_KEY(leaf, key);
5547 /** Move the cursor to the last item in the database. */
5549 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5555 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5557 if (!(mc->mc_flags & C_EOF)) {
5559 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5560 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5561 if (rc != MDB_SUCCESS)
5564 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5567 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5568 mc->mc_flags |= C_INITIALIZED|C_EOF;
5569 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5571 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5572 key->mv_size = mc->mc_db->md_pad;
5573 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5578 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5579 mdb_xcursor_init1(mc, leaf);
5580 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5584 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5589 MDB_GET_KEY(leaf, key);
5594 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5599 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5604 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5608 case MDB_GET_CURRENT:
5609 if (!(mc->mc_flags & C_INITIALIZED)) {
5612 MDB_page *mp = mc->mc_pg[mc->mc_top];
5613 int nkeys = NUMKEYS(mp);
5614 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5615 mc->mc_ki[mc->mc_top] = nkeys;
5621 key->mv_size = mc->mc_db->md_pad;
5622 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5624 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5625 MDB_GET_KEY(leaf, key);
5627 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5628 if (mc->mc_flags & C_DEL)
5629 mdb_xcursor_init1(mc, leaf);
5630 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5632 rc = mdb_node_read(mc->mc_txn, leaf, data);
5639 case MDB_GET_BOTH_RANGE:
5644 if (mc->mc_xcursor == NULL) {
5645 rc = MDB_INCOMPATIBLE;
5655 rc = mdb_cursor_set(mc, key, data, op,
5656 op == MDB_SET_RANGE ? NULL : &exact);
5659 case MDB_GET_MULTIPLE:
5660 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5664 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5665 rc = MDB_INCOMPATIBLE;
5669 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5670 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5673 case MDB_NEXT_MULTIPLE:
5678 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5679 rc = MDB_INCOMPATIBLE;
5682 if (!(mc->mc_flags & C_INITIALIZED))
5683 rc = mdb_cursor_first(mc, key, data);
5685 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5686 if (rc == MDB_SUCCESS) {
5687 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5690 mx = &mc->mc_xcursor->mx_cursor;
5691 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5693 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5694 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5702 case MDB_NEXT_NODUP:
5703 if (!(mc->mc_flags & C_INITIALIZED))
5704 rc = mdb_cursor_first(mc, key, data);
5706 rc = mdb_cursor_next(mc, key, data, op);
5710 case MDB_PREV_NODUP:
5711 if (!(mc->mc_flags & C_INITIALIZED)) {
5712 rc = mdb_cursor_last(mc, key, data);
5715 mc->mc_flags |= C_INITIALIZED;
5716 mc->mc_ki[mc->mc_top]++;
5718 rc = mdb_cursor_prev(mc, key, data, op);
5721 rc = mdb_cursor_first(mc, key, data);
5724 mfunc = mdb_cursor_first;
5726 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5730 if (mc->mc_xcursor == NULL) {
5731 rc = MDB_INCOMPATIBLE;
5734 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5738 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5741 rc = mdb_cursor_last(mc, key, data);
5744 mfunc = mdb_cursor_last;
5747 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5752 if (mc->mc_flags & C_DEL)
5753 mc->mc_flags ^= C_DEL;
5758 /** Touch all the pages in the cursor stack. Set mc_top.
5759 * Makes sure all the pages are writable, before attempting a write operation.
5760 * @param[in] mc The cursor to operate on.
5763 mdb_cursor_touch(MDB_cursor *mc)
5765 int rc = MDB_SUCCESS;
5767 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5770 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5771 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5774 *mc->mc_dbflag |= DB_DIRTY;
5779 rc = mdb_page_touch(mc);
5780 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5781 mc->mc_top = mc->mc_snum-1;
5786 /** Do not spill pages to disk if txn is getting full, may fail instead */
5787 #define MDB_NOSPILL 0x8000
5790 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5793 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5795 MDB_node *leaf = NULL;
5798 MDB_val xdata, *rdata, dkey, olddata;
5800 int do_sub = 0, insert;
5801 unsigned int mcount = 0, dcount = 0, nospill;
5804 unsigned int nflags;
5807 if (mc == NULL || key == NULL)
5810 env = mc->mc_txn->mt_env;
5812 /* Check this first so counter will always be zero on any
5815 if (flags & MDB_MULTIPLE) {
5816 dcount = data[1].mv_size;
5817 data[1].mv_size = 0;
5818 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5819 return MDB_INCOMPATIBLE;
5822 nospill = flags & MDB_NOSPILL;
5823 flags &= ~MDB_NOSPILL;
5825 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5826 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5828 if (flags != MDB_CURRENT && key->mv_size-1 >= ENV_MAXKEY(env))
5829 return MDB_BAD_VALSIZE;
5831 #if SIZE_MAX > MAXDATASIZE
5832 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5833 return MDB_BAD_VALSIZE;
5835 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5836 return MDB_BAD_VALSIZE;
5839 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5840 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5844 if (flags == MDB_CURRENT) {
5845 if (!(mc->mc_flags & C_INITIALIZED))
5848 } else if (mc->mc_db->md_root == P_INVALID) {
5849 /* new database, cursor has nothing to point to */
5852 mc->mc_flags &= ~C_INITIALIZED;
5857 if (flags & MDB_APPEND) {
5859 rc = mdb_cursor_last(mc, &k2, &d2);
5861 rc = mc->mc_dbx->md_cmp(key, &k2);
5864 mc->mc_ki[mc->mc_top]++;
5866 /* new key is <= last key */
5871 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5873 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5874 DPRINTF(("duplicate key [%s]", DKEY(key)));
5876 return MDB_KEYEXIST;
5878 if (rc && rc != MDB_NOTFOUND)
5882 if (mc->mc_flags & C_DEL)
5883 mc->mc_flags ^= C_DEL;
5885 /* Cursor is positioned, check for room in the dirty list */
5887 if (flags & MDB_MULTIPLE) {
5889 xdata.mv_size = data->mv_size * dcount;
5893 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5897 if (rc == MDB_NO_ROOT) {
5899 /* new database, write a root leaf page */
5900 DPUTS("allocating new root leaf page");
5901 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5904 mdb_cursor_push(mc, np);
5905 mc->mc_db->md_root = np->mp_pgno;
5906 mc->mc_db->md_depth++;
5907 *mc->mc_dbflag |= DB_DIRTY;
5908 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5910 np->mp_flags |= P_LEAF2;
5911 mc->mc_flags |= C_INITIALIZED;
5913 /* make sure all cursor pages are writable */
5914 rc2 = mdb_cursor_touch(mc);
5921 /* The key does not exist */
5922 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5923 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5924 LEAFSIZE(key, data) > env->me_nodemax)
5926 /* Too big for a node, insert in sub-DB */
5927 fp_flags = P_LEAF|P_DIRTY;
5929 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5930 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5934 /* there's only a key anyway, so this is a no-op */
5935 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5936 unsigned int ksize = mc->mc_db->md_pad;
5937 if (key->mv_size != ksize)
5938 return MDB_BAD_VALSIZE;
5939 if (flags == MDB_CURRENT) {
5940 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5941 memcpy(ptr, key->mv_data, ksize);
5947 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5948 olddata.mv_size = NODEDSZ(leaf);
5949 olddata.mv_data = NODEDATA(leaf);
5952 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5953 /* Prepare (sub-)page/sub-DB to accept the new item,
5954 * if needed. fp: old sub-page or a header faking
5955 * it. mp: new (sub-)page. offset: growth in page
5956 * size. xdata: node data with new page or DB.
5958 unsigned i, offset = 0;
5959 mp = fp = xdata.mv_data = env->me_pbuf;
5960 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5962 /* Was a single item before, must convert now */
5963 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5964 /* Just overwrite the current item */
5965 if (flags == MDB_CURRENT)
5968 #if UINT_MAX < SIZE_MAX
5969 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5970 #ifdef MISALIGNED_OK
5971 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5973 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5976 /* if data matches, skip it */
5977 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5978 if (flags & MDB_NODUPDATA)
5980 else if (flags & MDB_MULTIPLE)
5987 /* Back up original data item */
5988 dkey.mv_size = olddata.mv_size;
5989 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5991 /* Make sub-page header for the dup items, with dummy body */
5992 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5993 fp->mp_lower = PAGEHDRSZ;
5994 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5995 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5996 fp->mp_flags |= P_LEAF2;
5997 fp->mp_pad = data->mv_size;
5998 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6000 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6001 (dkey.mv_size & 1) + (data->mv_size & 1);
6003 fp->mp_upper = xdata.mv_size;
6004 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
6005 } else if (leaf->mn_flags & F_SUBDATA) {
6006 /* Data is on sub-DB, just store it */
6007 flags |= F_DUPDATA|F_SUBDATA;
6010 /* Data is on sub-page */
6011 fp = olddata.mv_data;
6014 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6015 offset = EVEN(NODESIZE + sizeof(indx_t) +
6019 offset = fp->mp_pad;
6020 if (SIZELEFT(fp) < offset) {
6021 offset *= 4; /* space for 4 more */
6024 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6026 fp->mp_flags |= P_DIRTY;
6027 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6028 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6032 xdata.mv_size = olddata.mv_size + offset;
6035 fp_flags = fp->mp_flags;
6036 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6037 /* Too big for a sub-page, convert to sub-DB */
6038 fp_flags &= ~P_SUBP;
6040 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6041 fp_flags |= P_LEAF2;
6042 dummy.md_pad = fp->mp_pad;
6043 dummy.md_flags = MDB_DUPFIXED;
6044 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6045 dummy.md_flags |= MDB_INTEGERKEY;
6051 dummy.md_branch_pages = 0;
6052 dummy.md_leaf_pages = 1;
6053 dummy.md_overflow_pages = 0;
6054 dummy.md_entries = NUMKEYS(fp);
6055 xdata.mv_size = sizeof(MDB_db);
6056 xdata.mv_data = &dummy;
6057 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6059 offset = env->me_psize - olddata.mv_size;
6060 flags |= F_DUPDATA|F_SUBDATA;
6061 dummy.md_root = mp->mp_pgno;
6064 mp->mp_flags = fp_flags | P_DIRTY;
6065 mp->mp_pad = fp->mp_pad;
6066 mp->mp_lower = fp->mp_lower;
6067 mp->mp_upper = fp->mp_upper + offset;
6068 if (fp_flags & P_LEAF2) {
6069 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6071 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
6072 olddata.mv_size - fp->mp_upper);
6073 for (i=0; i<NUMKEYS(fp); i++)
6074 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6082 mdb_node_del(mc, 0);
6086 /* overflow page overwrites need special handling */
6087 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6090 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6092 memcpy(&pg, olddata.mv_data, sizeof(pg));
6093 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6095 ovpages = omp->mp_pages;
6097 /* Is the ov page large enough? */
6098 if (ovpages >= dpages) {
6099 if (!(omp->mp_flags & P_DIRTY) &&
6100 (level || (env->me_flags & MDB_WRITEMAP)))
6102 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6105 level = 0; /* dirty in this txn or clean */
6108 if (omp->mp_flags & P_DIRTY) {
6109 /* yes, overwrite it. Note in this case we don't
6110 * bother to try shrinking the page if the new data
6111 * is smaller than the overflow threshold.
6114 /* It is writable only in a parent txn */
6115 size_t sz = (size_t) env->me_psize * ovpages, off;
6116 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6122 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6123 mdb_cassert(mc, rc2 == 0);
6124 if (!(flags & MDB_RESERVE)) {
6125 /* Copy end of page, adjusting alignment so
6126 * compiler may copy words instead of bytes.
6128 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6129 memcpy((size_t *)((char *)np + off),
6130 (size_t *)((char *)omp + off), sz - off);
6133 memcpy(np, omp, sz); /* Copy beginning of page */
6136 SETDSZ(leaf, data->mv_size);
6137 if (F_ISSET(flags, MDB_RESERVE))
6138 data->mv_data = METADATA(omp);
6140 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6144 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6146 } else if (data->mv_size == olddata.mv_size) {
6147 /* same size, just replace it. Note that we could
6148 * also reuse this node if the new data is smaller,
6149 * but instead we opt to shrink the node in that case.
6151 if (F_ISSET(flags, MDB_RESERVE))
6152 data->mv_data = olddata.mv_data;
6153 else if (!(mc->mc_flags & C_SUB))
6154 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6156 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6159 mdb_node_del(mc, 0);
6165 nflags = flags & NODE_ADD_FLAGS;
6166 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6167 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6168 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6169 nflags &= ~MDB_APPEND;
6171 nflags |= MDB_SPLIT_REPLACE;
6172 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6174 /* There is room already in this leaf page. */
6175 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6176 if (rc == 0 && !do_sub && insert) {
6177 /* Adjust other cursors pointing to mp */
6178 MDB_cursor *m2, *m3;
6179 MDB_dbi dbi = mc->mc_dbi;
6180 unsigned i = mc->mc_top;
6181 MDB_page *mp = mc->mc_pg[i];
6183 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6184 if (mc->mc_flags & C_SUB)
6185 m3 = &m2->mc_xcursor->mx_cursor;
6188 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6189 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6196 if (rc != MDB_SUCCESS)
6197 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6199 /* Now store the actual data in the child DB. Note that we're
6200 * storing the user data in the keys field, so there are strict
6201 * size limits on dupdata. The actual data fields of the child
6202 * DB are all zero size.
6209 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6210 if (flags & MDB_CURRENT) {
6211 xflags = MDB_CURRENT|MDB_NOSPILL;
6213 mdb_xcursor_init1(mc, leaf);
6214 xflags = (flags & MDB_NODUPDATA) ?
6215 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6217 /* converted, write the original data first */
6219 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6221 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6222 return rc == MDB_KEYEXIST ? MDB_CORRUPTED : rc;
6225 /* Adjust other cursors pointing to mp */
6227 unsigned i = mc->mc_top;
6228 MDB_page *mp = mc->mc_pg[i];
6230 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6231 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6232 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6233 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6234 mdb_xcursor_init1(m2, leaf);
6238 /* we've done our job */
6241 if (flags & MDB_APPENDDUP)
6242 xflags |= MDB_APPEND;
6243 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6244 if (flags & F_SUBDATA) {
6245 void *db = NODEDATA(leaf);
6246 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6249 /* sub-writes might have failed so check rc again.
6250 * Don't increment count if we just replaced an existing item.
6253 mc->mc_db->md_entries++;
6254 if (flags & MDB_MULTIPLE) {
6258 /* let caller know how many succeeded, if any */
6259 data[1].mv_size = mcount;
6260 if (mcount < dcount) {
6261 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6268 /* If we succeeded and the key didn't exist before, make sure
6269 * the cursor is marked valid.
6272 mc->mc_flags |= C_INITIALIZED;
6277 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6283 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6284 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6286 if (!(mc->mc_flags & C_INITIALIZED))
6289 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6290 return MDB_NOTFOUND;
6292 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6295 rc = mdb_cursor_touch(mc);
6299 mp = mc->mc_pg[mc->mc_top];
6302 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6304 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6305 if (!(flags & MDB_NODUPDATA)) {
6306 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6307 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6309 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6312 /* If sub-DB still has entries, we're done */
6313 if (mc->mc_xcursor->mx_db.md_entries) {
6314 if (leaf->mn_flags & F_SUBDATA) {
6315 /* update subDB info */
6316 void *db = NODEDATA(leaf);
6317 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6320 /* shrink fake page */
6321 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6322 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6323 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6324 /* fix other sub-DB cursors pointed at this fake page */
6325 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6326 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6327 if (m2->mc_pg[mc->mc_top] == mp &&
6328 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6329 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6332 mc->mc_db->md_entries--;
6333 mc->mc_flags |= C_DEL;
6336 /* otherwise fall thru and delete the sub-DB */
6339 if (leaf->mn_flags & F_SUBDATA) {
6340 /* add all the child DB's pages to the free list */
6341 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6344 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries;
6348 /* add overflow pages to free list */
6349 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6353 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6354 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6355 (rc = mdb_ovpage_free(mc, omp)))
6360 return mdb_cursor_del0(mc);
6363 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6367 /** Allocate and initialize new pages for a database.
6368 * @param[in] mc a cursor on the database being added to.
6369 * @param[in] flags flags defining what type of page is being allocated.
6370 * @param[in] num the number of pages to allocate. This is usually 1,
6371 * unless allocating overflow pages for a large record.
6372 * @param[out] mp Address of a page, or NULL on failure.
6373 * @return 0 on success, non-zero on failure.
6376 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6381 if ((rc = mdb_page_alloc(mc, num, &np)))
6383 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6384 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6385 np->mp_flags = flags | P_DIRTY;
6386 np->mp_lower = PAGEHDRSZ;
6387 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6390 mc->mc_db->md_branch_pages++;
6391 else if (IS_LEAF(np))
6392 mc->mc_db->md_leaf_pages++;
6393 else if (IS_OVERFLOW(np)) {
6394 mc->mc_db->md_overflow_pages += num;
6402 /** Calculate the size of a leaf node.
6403 * The size depends on the environment's page size; if a data item
6404 * is too large it will be put onto an overflow page and the node
6405 * size will only include the key and not the data. Sizes are always
6406 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6407 * of the #MDB_node headers.
6408 * @param[in] env The environment handle.
6409 * @param[in] key The key for the node.
6410 * @param[in] data The data for the node.
6411 * @return The number of bytes needed to store the node.
6414 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6418 sz = LEAFSIZE(key, data);
6419 if (sz > env->me_nodemax) {
6420 /* put on overflow page */
6421 sz -= data->mv_size - sizeof(pgno_t);
6424 return EVEN(sz + sizeof(indx_t));
6427 /** Calculate the size of a branch node.
6428 * The size should depend on the environment's page size but since
6429 * we currently don't support spilling large keys onto overflow
6430 * pages, it's simply the size of the #MDB_node header plus the
6431 * size of the key. Sizes are always rounded up to an even number
6432 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6433 * @param[in] env The environment handle.
6434 * @param[in] key The key for the node.
6435 * @return The number of bytes needed to store the node.
6438 mdb_branch_size(MDB_env *env, MDB_val *key)
6443 if (sz > env->me_nodemax) {
6444 /* put on overflow page */
6445 /* not implemented */
6446 /* sz -= key->size - sizeof(pgno_t); */
6449 return sz + sizeof(indx_t);
6452 /** Add a node to the page pointed to by the cursor.
6453 * @param[in] mc The cursor for this operation.
6454 * @param[in] indx The index on the page where the new node should be added.
6455 * @param[in] key The key for the new node.
6456 * @param[in] data The data for the new node, if any.
6457 * @param[in] pgno The page number, if adding a branch node.
6458 * @param[in] flags Flags for the node.
6459 * @return 0 on success, non-zero on failure. Possible errors are:
6461 * <li>ENOMEM - failed to allocate overflow pages for the node.
6462 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6463 * should never happen since all callers already calculate the
6464 * page's free space before calling this function.
6468 mdb_node_add(MDB_cursor *mc, indx_t indx,
6469 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6472 size_t node_size = NODESIZE;
6476 MDB_page *mp = mc->mc_pg[mc->mc_top];
6477 MDB_page *ofp = NULL; /* overflow page */
6480 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6482 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6483 IS_LEAF(mp) ? "leaf" : "branch",
6484 IS_SUBP(mp) ? "sub-" : "",
6485 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6486 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6489 /* Move higher keys up one slot. */
6490 int ksize = mc->mc_db->md_pad, dif;
6491 char *ptr = LEAF2KEY(mp, indx, ksize);
6492 dif = NUMKEYS(mp) - indx;
6494 memmove(ptr+ksize, ptr, dif*ksize);
6495 /* insert new key */
6496 memcpy(ptr, key->mv_data, ksize);
6498 /* Just using these for counting */
6499 mp->mp_lower += sizeof(indx_t);
6500 mp->mp_upper -= ksize - sizeof(indx_t);
6504 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6506 node_size += key->mv_size;
6508 mdb_cassert(mc, data);
6509 if (F_ISSET(flags, F_BIGDATA)) {
6510 /* Data already on overflow page. */
6511 node_size += sizeof(pgno_t);
6512 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6513 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6515 /* Put data on overflow page. */
6516 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6517 data->mv_size, node_size+data->mv_size));
6518 node_size = EVEN(node_size + sizeof(pgno_t));
6519 if ((ssize_t)node_size > room)
6521 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6523 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6527 node_size += data->mv_size;
6530 node_size = EVEN(node_size);
6531 if ((ssize_t)node_size > room)
6535 /* Move higher pointers up one slot. */
6536 for (i = NUMKEYS(mp); i > indx; i--)
6537 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6539 /* Adjust free space offsets. */
6540 ofs = mp->mp_upper - node_size;
6541 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6542 mp->mp_ptrs[indx] = ofs;
6544 mp->mp_lower += sizeof(indx_t);
6546 /* Write the node data. */
6547 node = NODEPTR(mp, indx);
6548 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6549 node->mn_flags = flags;
6551 SETDSZ(node,data->mv_size);
6556 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6559 mdb_cassert(mc, key);
6561 if (F_ISSET(flags, F_BIGDATA))
6562 memcpy(node->mn_data + key->mv_size, data->mv_data,
6564 else if (F_ISSET(flags, MDB_RESERVE))
6565 data->mv_data = node->mn_data + key->mv_size;
6567 memcpy(node->mn_data + key->mv_size, data->mv_data,
6570 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6572 if (F_ISSET(flags, MDB_RESERVE))
6573 data->mv_data = METADATA(ofp);
6575 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6582 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6583 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6584 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6585 DPRINTF(("node size = %"Z"u", node_size));
6586 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6587 return MDB_PAGE_FULL;
6590 /** Delete the specified node from a page.
6591 * @param[in] mc Cursor pointing to the node to delete.
6592 * @param[in] ksize The size of a node. Only used if the page is
6593 * part of a #MDB_DUPFIXED database.
6596 mdb_node_del(MDB_cursor *mc, int ksize)
6598 MDB_page *mp = mc->mc_pg[mc->mc_top];
6599 indx_t indx = mc->mc_ki[mc->mc_top];
6601 indx_t i, j, numkeys, ptr;
6605 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6606 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6607 numkeys = NUMKEYS(mp);
6608 mdb_cassert(mc, indx < numkeys);
6611 int x = numkeys - 1 - indx;
6612 base = LEAF2KEY(mp, indx, ksize);
6614 memmove(base, base + ksize, x * ksize);
6615 mp->mp_lower -= sizeof(indx_t);
6616 mp->mp_upper += ksize - sizeof(indx_t);
6620 node = NODEPTR(mp, indx);
6621 sz = NODESIZE + node->mn_ksize;
6623 if (F_ISSET(node->mn_flags, F_BIGDATA))
6624 sz += sizeof(pgno_t);
6626 sz += NODEDSZ(node);
6630 ptr = mp->mp_ptrs[indx];
6631 for (i = j = 0; i < numkeys; i++) {
6633 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6634 if (mp->mp_ptrs[i] < ptr)
6635 mp->mp_ptrs[j] += sz;
6640 base = (char *)mp + mp->mp_upper;
6641 memmove(base + sz, base, ptr - mp->mp_upper);
6643 mp->mp_lower -= sizeof(indx_t);
6647 /** Compact the main page after deleting a node on a subpage.
6648 * @param[in] mp The main page to operate on.
6649 * @param[in] indx The index of the subpage on the main page.
6652 mdb_node_shrink(MDB_page *mp, indx_t indx)
6658 indx_t i, numkeys, ptr;
6660 node = NODEPTR(mp, indx);
6661 sp = (MDB_page *)NODEDATA(node);
6662 delta = SIZELEFT(sp);
6663 xp = (MDB_page *)((char *)sp + delta);
6665 /* shift subpage upward */
6667 nsize = NUMKEYS(sp) * sp->mp_pad;
6669 return; /* do not make the node uneven-sized */
6670 memmove(METADATA(xp), METADATA(sp), nsize);
6673 numkeys = NUMKEYS(sp);
6674 for (i=numkeys-1; i>=0; i--)
6675 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6677 xp->mp_upper = sp->mp_lower;
6678 xp->mp_lower = sp->mp_lower;
6679 xp->mp_flags = sp->mp_flags;
6680 xp->mp_pad = sp->mp_pad;
6681 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6683 nsize = NODEDSZ(node) - delta;
6684 SETDSZ(node, nsize);
6686 /* shift lower nodes upward */
6687 ptr = mp->mp_ptrs[indx];
6688 numkeys = NUMKEYS(mp);
6689 for (i = 0; i < numkeys; i++) {
6690 if (mp->mp_ptrs[i] <= ptr)
6691 mp->mp_ptrs[i] += delta;
6694 base = (char *)mp + mp->mp_upper;
6695 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6696 mp->mp_upper += delta;
6699 /** Initial setup of a sorted-dups cursor.
6700 * Sorted duplicates are implemented as a sub-database for the given key.
6701 * The duplicate data items are actually keys of the sub-database.
6702 * Operations on the duplicate data items are performed using a sub-cursor
6703 * initialized when the sub-database is first accessed. This function does
6704 * the preliminary setup of the sub-cursor, filling in the fields that
6705 * depend only on the parent DB.
6706 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6709 mdb_xcursor_init0(MDB_cursor *mc)
6711 MDB_xcursor *mx = mc->mc_xcursor;
6713 mx->mx_cursor.mc_xcursor = NULL;
6714 mx->mx_cursor.mc_txn = mc->mc_txn;
6715 mx->mx_cursor.mc_db = &mx->mx_db;
6716 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6717 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6718 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6719 mx->mx_cursor.mc_snum = 0;
6720 mx->mx_cursor.mc_top = 0;
6721 mx->mx_cursor.mc_flags = C_SUB;
6722 mx->mx_dbx.md_name.mv_size = 0;
6723 mx->mx_dbx.md_name.mv_data = NULL;
6724 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6725 mx->mx_dbx.md_dcmp = NULL;
6726 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6729 /** Final setup of a sorted-dups cursor.
6730 * Sets up the fields that depend on the data from the main cursor.
6731 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6732 * @param[in] node The data containing the #MDB_db record for the
6733 * sorted-dup database.
6736 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6738 MDB_xcursor *mx = mc->mc_xcursor;
6740 if (node->mn_flags & F_SUBDATA) {
6741 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6742 mx->mx_cursor.mc_pg[0] = 0;
6743 mx->mx_cursor.mc_snum = 0;
6744 mx->mx_cursor.mc_top = 0;
6745 mx->mx_cursor.mc_flags = C_SUB;
6747 MDB_page *fp = NODEDATA(node);
6748 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6749 mx->mx_db.md_flags = 0;
6750 mx->mx_db.md_depth = 1;
6751 mx->mx_db.md_branch_pages = 0;
6752 mx->mx_db.md_leaf_pages = 1;
6753 mx->mx_db.md_overflow_pages = 0;
6754 mx->mx_db.md_entries = NUMKEYS(fp);
6755 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6756 mx->mx_cursor.mc_snum = 1;
6757 mx->mx_cursor.mc_top = 0;
6758 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6759 mx->mx_cursor.mc_pg[0] = fp;
6760 mx->mx_cursor.mc_ki[0] = 0;
6761 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6762 mx->mx_db.md_flags = MDB_DUPFIXED;
6763 mx->mx_db.md_pad = fp->mp_pad;
6764 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6765 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6768 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6769 mx->mx_db.md_root));
6770 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6771 #if UINT_MAX < SIZE_MAX
6772 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6773 #ifdef MISALIGNED_OK
6774 mx->mx_dbx.md_cmp = mdb_cmp_long;
6776 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6781 /** Initialize a cursor for a given transaction and database. */
6783 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6786 mc->mc_backup = NULL;
6789 mc->mc_db = &txn->mt_dbs[dbi];
6790 mc->mc_dbx = &txn->mt_dbxs[dbi];
6791 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6796 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6797 mdb_tassert(txn, mx != NULL);
6798 mc->mc_xcursor = mx;
6799 mdb_xcursor_init0(mc);
6801 mc->mc_xcursor = NULL;
6803 if (*mc->mc_dbflag & DB_STALE) {
6804 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6809 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6812 size_t size = sizeof(MDB_cursor);
6814 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6817 if (txn->mt_flags & MDB_TXN_ERROR)
6820 /* Allow read access to the freelist */
6821 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6824 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6825 size += sizeof(MDB_xcursor);
6827 if ((mc = malloc(size)) != NULL) {
6828 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6829 if (txn->mt_cursors) {
6830 mc->mc_next = txn->mt_cursors[dbi];
6831 txn->mt_cursors[dbi] = mc;
6832 mc->mc_flags |= C_UNTRACK;
6844 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6846 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6849 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6852 if (txn->mt_flags & MDB_TXN_ERROR)
6855 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6859 /* Return the count of duplicate data items for the current key */
6861 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6865 if (mc == NULL || countp == NULL)
6868 if (mc->mc_xcursor == NULL)
6869 return MDB_INCOMPATIBLE;
6871 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6874 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6875 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6878 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6881 *countp = mc->mc_xcursor->mx_db.md_entries;
6887 mdb_cursor_close(MDB_cursor *mc)
6889 if (mc && !mc->mc_backup) {
6890 /* remove from txn, if tracked */
6891 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6892 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6893 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6895 *prev = mc->mc_next;
6902 mdb_cursor_txn(MDB_cursor *mc)
6904 if (!mc) return NULL;
6909 mdb_cursor_dbi(MDB_cursor *mc)
6914 /** Replace the key for a branch node with a new key.
6915 * @param[in] mc Cursor pointing to the node to operate on.
6916 * @param[in] key The new key to use.
6917 * @return 0 on success, non-zero on failure.
6920 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6926 int delta, ksize, oksize;
6927 indx_t ptr, i, numkeys, indx;
6930 indx = mc->mc_ki[mc->mc_top];
6931 mp = mc->mc_pg[mc->mc_top];
6932 node = NODEPTR(mp, indx);
6933 ptr = mp->mp_ptrs[indx];
6937 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6938 k2.mv_data = NODEKEY(node);
6939 k2.mv_size = node->mn_ksize;
6940 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6942 mdb_dkey(&k2, kbuf2),
6948 /* Sizes must be 2-byte aligned. */
6949 ksize = EVEN(key->mv_size);
6950 oksize = EVEN(node->mn_ksize);
6951 delta = ksize - oksize;
6953 /* Shift node contents if EVEN(key length) changed. */
6955 if (delta > 0 && SIZELEFT(mp) < delta) {
6957 /* not enough space left, do a delete and split */
6958 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6959 pgno = NODEPGNO(node);
6960 mdb_node_del(mc, 0);
6961 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6964 numkeys = NUMKEYS(mp);
6965 for (i = 0; i < numkeys; i++) {
6966 if (mp->mp_ptrs[i] <= ptr)
6967 mp->mp_ptrs[i] -= delta;
6970 base = (char *)mp + mp->mp_upper;
6971 len = ptr - mp->mp_upper + NODESIZE;
6972 memmove(base - delta, base, len);
6973 mp->mp_upper -= delta;
6975 node = NODEPTR(mp, indx);
6978 /* But even if no shift was needed, update ksize */
6979 if (node->mn_ksize != key->mv_size)
6980 node->mn_ksize = key->mv_size;
6983 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6989 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6991 /** Move a node from csrc to cdst.
6994 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7001 unsigned short flags;
7005 /* Mark src and dst as dirty. */
7006 if ((rc = mdb_page_touch(csrc)) ||
7007 (rc = mdb_page_touch(cdst)))
7010 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7011 key.mv_size = csrc->mc_db->md_pad;
7012 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7014 data.mv_data = NULL;
7018 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7019 mdb_cassert(csrc, !((size_t)srcnode & 1));
7020 srcpg = NODEPGNO(srcnode);
7021 flags = srcnode->mn_flags;
7022 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7023 unsigned int snum = csrc->mc_snum;
7025 /* must find the lowest key below src */
7026 rc = mdb_page_search_lowest(csrc);
7029 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7030 key.mv_size = csrc->mc_db->md_pad;
7031 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7033 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7034 key.mv_size = NODEKSZ(s2);
7035 key.mv_data = NODEKEY(s2);
7037 csrc->mc_snum = snum--;
7038 csrc->mc_top = snum;
7040 key.mv_size = NODEKSZ(srcnode);
7041 key.mv_data = NODEKEY(srcnode);
7043 data.mv_size = NODEDSZ(srcnode);
7044 data.mv_data = NODEDATA(srcnode);
7046 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7047 unsigned int snum = cdst->mc_snum;
7050 /* must find the lowest key below dst */
7051 rc = mdb_page_search_lowest(cdst);
7054 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
7055 bkey.mv_size = cdst->mc_db->md_pad;
7056 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
7058 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7059 bkey.mv_size = NODEKSZ(s2);
7060 bkey.mv_data = NODEKEY(s2);
7062 cdst->mc_snum = snum--;
7063 cdst->mc_top = snum;
7064 mdb_cursor_copy(cdst, &mn);
7066 rc = mdb_update_key(&mn, &bkey);
7071 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7072 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7073 csrc->mc_ki[csrc->mc_top],
7075 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7076 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7078 /* Add the node to the destination page.
7080 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7081 if (rc != MDB_SUCCESS)
7084 /* Delete the node from the source page.
7086 mdb_node_del(csrc, key.mv_size);
7089 /* Adjust other cursors pointing to mp */
7090 MDB_cursor *m2, *m3;
7091 MDB_dbi dbi = csrc->mc_dbi;
7092 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7094 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7095 if (csrc->mc_flags & C_SUB)
7096 m3 = &m2->mc_xcursor->mx_cursor;
7099 if (m3 == csrc) continue;
7100 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7101 csrc->mc_ki[csrc->mc_top]) {
7102 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7103 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7108 /* Update the parent separators.
7110 if (csrc->mc_ki[csrc->mc_top] == 0) {
7111 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7112 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7113 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7115 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7116 key.mv_size = NODEKSZ(srcnode);
7117 key.mv_data = NODEKEY(srcnode);
7119 DPRINTF(("update separator for source page %"Z"u to [%s]",
7120 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7121 mdb_cursor_copy(csrc, &mn);
7124 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7127 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7129 indx_t ix = csrc->mc_ki[csrc->mc_top];
7130 nullkey.mv_size = 0;
7131 csrc->mc_ki[csrc->mc_top] = 0;
7132 rc = mdb_update_key(csrc, &nullkey);
7133 csrc->mc_ki[csrc->mc_top] = ix;
7134 mdb_cassert(csrc, rc == MDB_SUCCESS);
7138 if (cdst->mc_ki[cdst->mc_top] == 0) {
7139 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7140 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7141 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7143 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7144 key.mv_size = NODEKSZ(srcnode);
7145 key.mv_data = NODEKEY(srcnode);
7147 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7148 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7149 mdb_cursor_copy(cdst, &mn);
7152 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7155 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7157 indx_t ix = cdst->mc_ki[cdst->mc_top];
7158 nullkey.mv_size = 0;
7159 cdst->mc_ki[cdst->mc_top] = 0;
7160 rc = mdb_update_key(cdst, &nullkey);
7161 cdst->mc_ki[cdst->mc_top] = ix;
7162 mdb_cassert(csrc, rc == MDB_SUCCESS);
7169 /** Merge one page into another.
7170 * The nodes from the page pointed to by \b csrc will
7171 * be copied to the page pointed to by \b cdst and then
7172 * the \b csrc page will be freed.
7173 * @param[in] csrc Cursor pointing to the source page.
7174 * @param[in] cdst Cursor pointing to the destination page.
7175 * @return 0 on success, non-zero on failure.
7178 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7186 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7187 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7189 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7190 mdb_cassert(csrc, cdst->mc_snum > 1);
7192 /* Mark dst as dirty. */
7193 if ((rc = mdb_page_touch(cdst)))
7196 /* Move all nodes from src to dst.
7198 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7199 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7200 key.mv_size = csrc->mc_db->md_pad;
7201 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7202 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7203 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7204 if (rc != MDB_SUCCESS)
7206 key.mv_data = (char *)key.mv_data + key.mv_size;
7209 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7210 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7211 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7212 unsigned int snum = csrc->mc_snum;
7214 /* must find the lowest key below src */
7215 rc = mdb_page_search_lowest(csrc);
7218 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7219 key.mv_size = csrc->mc_db->md_pad;
7220 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7222 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7223 key.mv_size = NODEKSZ(s2);
7224 key.mv_data = NODEKEY(s2);
7226 csrc->mc_snum = snum--;
7227 csrc->mc_top = snum;
7229 key.mv_size = srcnode->mn_ksize;
7230 key.mv_data = NODEKEY(srcnode);
7233 data.mv_size = NODEDSZ(srcnode);
7234 data.mv_data = NODEDATA(srcnode);
7235 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7236 if (rc != MDB_SUCCESS)
7241 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7242 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7243 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7245 /* Unlink the src page from parent and add to free list.
7248 mdb_node_del(csrc, 0);
7249 if (csrc->mc_ki[csrc->mc_top] == 0) {
7251 rc = mdb_update_key(csrc, &key);
7259 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7260 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7263 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7264 csrc->mc_db->md_leaf_pages--;
7266 csrc->mc_db->md_branch_pages--;
7268 /* Adjust other cursors pointing to mp */
7269 MDB_cursor *m2, *m3;
7270 MDB_dbi dbi = csrc->mc_dbi;
7271 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7273 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7274 if (csrc->mc_flags & C_SUB)
7275 m3 = &m2->mc_xcursor->mx_cursor;
7278 if (m3 == csrc) continue;
7279 if (m3->mc_snum < csrc->mc_snum) continue;
7280 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7281 m3->mc_pg[csrc->mc_top] = mp;
7282 m3->mc_ki[csrc->mc_top] += nkeys;
7286 mdb_cursor_pop(csrc);
7288 return mdb_rebalance(csrc);
7291 /** Copy the contents of a cursor.
7292 * @param[in] csrc The cursor to copy from.
7293 * @param[out] cdst The cursor to copy to.
7296 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7300 cdst->mc_txn = csrc->mc_txn;
7301 cdst->mc_dbi = csrc->mc_dbi;
7302 cdst->mc_db = csrc->mc_db;
7303 cdst->mc_dbx = csrc->mc_dbx;
7304 cdst->mc_snum = csrc->mc_snum;
7305 cdst->mc_top = csrc->mc_top;
7306 cdst->mc_flags = csrc->mc_flags;
7308 for (i=0; i<csrc->mc_snum; i++) {
7309 cdst->mc_pg[i] = csrc->mc_pg[i];
7310 cdst->mc_ki[i] = csrc->mc_ki[i];
7314 /** Rebalance the tree after a delete operation.
7315 * @param[in] mc Cursor pointing to the page where rebalancing
7317 * @return 0 on success, non-zero on failure.
7320 mdb_rebalance(MDB_cursor *mc)
7324 unsigned int ptop, minkeys;
7327 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7328 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7329 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7330 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7331 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7333 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7334 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7335 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7336 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7340 if (mc->mc_snum < 2) {
7341 MDB_page *mp = mc->mc_pg[0];
7343 DPUTS("Can't rebalance a subpage, ignoring");
7346 if (NUMKEYS(mp) == 0) {
7347 DPUTS("tree is completely empty");
7348 mc->mc_db->md_root = P_INVALID;
7349 mc->mc_db->md_depth = 0;
7350 mc->mc_db->md_leaf_pages = 0;
7351 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7354 /* Adjust cursors pointing to mp */
7357 mc->mc_flags &= ~C_INITIALIZED;
7359 MDB_cursor *m2, *m3;
7360 MDB_dbi dbi = mc->mc_dbi;
7362 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7363 if (mc->mc_flags & C_SUB)
7364 m3 = &m2->mc_xcursor->mx_cursor;
7367 if (m3->mc_snum < mc->mc_snum) continue;
7368 if (m3->mc_pg[0] == mp) {
7371 m3->mc_flags &= ~C_INITIALIZED;
7375 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7376 DPUTS("collapsing root page!");
7377 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7380 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7381 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7384 mc->mc_db->md_depth--;
7385 mc->mc_db->md_branch_pages--;
7386 mc->mc_ki[0] = mc->mc_ki[1];
7388 /* Adjust other cursors pointing to mp */
7389 MDB_cursor *m2, *m3;
7390 MDB_dbi dbi = mc->mc_dbi;
7392 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7393 if (mc->mc_flags & C_SUB)
7394 m3 = &m2->mc_xcursor->mx_cursor;
7397 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7398 if (m3->mc_pg[0] == mp) {
7402 for (i=0; i<m3->mc_snum; i++) {
7403 m3->mc_pg[i] = m3->mc_pg[i+1];
7404 m3->mc_ki[i] = m3->mc_ki[i+1];
7410 DPUTS("root page doesn't need rebalancing");
7414 /* The parent (branch page) must have at least 2 pointers,
7415 * otherwise the tree is invalid.
7417 ptop = mc->mc_top-1;
7418 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7420 /* Leaf page fill factor is below the threshold.
7421 * Try to move keys from left or right neighbor, or
7422 * merge with a neighbor page.
7427 mdb_cursor_copy(mc, &mn);
7428 mn.mc_xcursor = NULL;
7430 if (mc->mc_ki[ptop] == 0) {
7431 /* We're the leftmost leaf in our parent.
7433 DPUTS("reading right neighbor");
7435 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7436 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7439 mn.mc_ki[mn.mc_top] = 0;
7440 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7442 /* There is at least one neighbor to the left.
7444 DPUTS("reading left neighbor");
7446 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7447 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7450 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7451 mc->mc_ki[mc->mc_top] = 0;
7454 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7455 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7456 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7458 /* If the neighbor page is above threshold and has enough keys,
7459 * move one key from it. Otherwise we should try to merge them.
7460 * (A branch page must never have less than 2 keys.)
7462 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7463 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7464 return mdb_node_move(&mn, mc);
7466 if (mc->mc_ki[ptop] == 0)
7467 rc = mdb_page_merge(&mn, mc);
7469 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7470 rc = mdb_page_merge(mc, &mn);
7471 mdb_cursor_copy(&mn, mc);
7473 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7478 /** Complete a delete operation started by #mdb_cursor_del(). */
7480 mdb_cursor_del0(MDB_cursor *mc)
7487 ki = mc->mc_ki[mc->mc_top];
7488 mdb_node_del(mc, mc->mc_db->md_pad);
7489 mc->mc_db->md_entries--;
7490 rc = mdb_rebalance(mc);
7492 if (rc == MDB_SUCCESS) {
7493 MDB_cursor *m2, *m3;
7494 MDB_dbi dbi = mc->mc_dbi;
7496 mp = mc->mc_pg[mc->mc_top];
7497 nkeys = NUMKEYS(mp);
7499 /* if mc points past last node in page, find next sibling */
7500 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7501 rc = mdb_cursor_sibling(mc, 1);
7502 if (rc == MDB_NOTFOUND)
7506 /* Adjust other cursors pointing to mp */
7507 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7508 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7509 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7511 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7513 if (m3->mc_pg[mc->mc_top] == mp) {
7514 if (m3->mc_ki[mc->mc_top] >= ki) {
7515 m3->mc_flags |= C_DEL;
7516 if (m3->mc_ki[mc->mc_top] > ki)
7517 m3->mc_ki[mc->mc_top]--;
7519 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7520 rc = mdb_cursor_sibling(m3, 1);
7521 if (rc == MDB_NOTFOUND)
7526 mc->mc_flags |= C_DEL;
7530 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7535 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7536 MDB_val *key, MDB_val *data)
7538 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7541 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7542 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7544 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7545 /* must ignore any data */
7549 return mdb_del0(txn, dbi, key, data, 0);
7553 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7554 MDB_val *key, MDB_val *data, unsigned flags)
7559 MDB_val rdata, *xdata;
7563 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7565 mdb_cursor_init(&mc, txn, dbi, &mx);
7574 flags |= MDB_NODUPDATA;
7576 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7578 /* let mdb_page_split know about this cursor if needed:
7579 * delete will trigger a rebalance; if it needs to move
7580 * a node from one page to another, it will have to
7581 * update the parent's separator key(s). If the new sepkey
7582 * is larger than the current one, the parent page may
7583 * run out of space, triggering a split. We need this
7584 * cursor to be consistent until the end of the rebalance.
7586 mc.mc_flags |= C_UNTRACK;
7587 mc.mc_next = txn->mt_cursors[dbi];
7588 txn->mt_cursors[dbi] = &mc;
7589 rc = mdb_cursor_del(&mc, flags);
7590 txn->mt_cursors[dbi] = mc.mc_next;
7595 /** Split a page and insert a new node.
7596 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7597 * The cursor will be updated to point to the actual page and index where
7598 * the node got inserted after the split.
7599 * @param[in] newkey The key for the newly inserted node.
7600 * @param[in] newdata The data for the newly inserted node.
7601 * @param[in] newpgno The page number, if the new node is a branch node.
7602 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7603 * @return 0 on success, non-zero on failure.
7606 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7607 unsigned int nflags)
7610 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7613 int i, j, split_indx, nkeys, pmax;
7614 MDB_env *env = mc->mc_txn->mt_env;
7616 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7617 MDB_page *copy = NULL;
7618 MDB_page *mp, *rp, *pp;
7623 mp = mc->mc_pg[mc->mc_top];
7624 newindx = mc->mc_ki[mc->mc_top];
7625 nkeys = NUMKEYS(mp);
7627 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7628 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7629 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7631 /* Create a right sibling. */
7632 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7634 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7636 if (mc->mc_snum < 2) {
7637 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7639 /* shift current top to make room for new parent */
7640 mc->mc_pg[1] = mc->mc_pg[0];
7641 mc->mc_ki[1] = mc->mc_ki[0];
7644 mc->mc_db->md_root = pp->mp_pgno;
7645 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7646 mc->mc_db->md_depth++;
7649 /* Add left (implicit) pointer. */
7650 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7651 /* undo the pre-push */
7652 mc->mc_pg[0] = mc->mc_pg[1];
7653 mc->mc_ki[0] = mc->mc_ki[1];
7654 mc->mc_db->md_root = mp->mp_pgno;
7655 mc->mc_db->md_depth--;
7662 ptop = mc->mc_top-1;
7663 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7666 mc->mc_flags |= C_SPLITTING;
7667 mdb_cursor_copy(mc, &mn);
7668 mn.mc_pg[mn.mc_top] = rp;
7669 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7671 if (nflags & MDB_APPEND) {
7672 mn.mc_ki[mn.mc_top] = 0;
7674 split_indx = newindx;
7678 split_indx = (nkeys+1) / 2;
7683 unsigned int lsize, rsize, ksize;
7684 /* Move half of the keys to the right sibling */
7685 x = mc->mc_ki[mc->mc_top] - split_indx;
7686 ksize = mc->mc_db->md_pad;
7687 split = LEAF2KEY(mp, split_indx, ksize);
7688 rsize = (nkeys - split_indx) * ksize;
7689 lsize = (nkeys - split_indx) * sizeof(indx_t);
7690 mp->mp_lower -= lsize;
7691 rp->mp_lower += lsize;
7692 mp->mp_upper += rsize - lsize;
7693 rp->mp_upper -= rsize - lsize;
7694 sepkey.mv_size = ksize;
7695 if (newindx == split_indx) {
7696 sepkey.mv_data = newkey->mv_data;
7698 sepkey.mv_data = split;
7701 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7702 memcpy(rp->mp_ptrs, split, rsize);
7703 sepkey.mv_data = rp->mp_ptrs;
7704 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7705 memcpy(ins, newkey->mv_data, ksize);
7706 mp->mp_lower += sizeof(indx_t);
7707 mp->mp_upper -= ksize - sizeof(indx_t);
7710 memcpy(rp->mp_ptrs, split, x * ksize);
7711 ins = LEAF2KEY(rp, x, ksize);
7712 memcpy(ins, newkey->mv_data, ksize);
7713 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7714 rp->mp_lower += sizeof(indx_t);
7715 rp->mp_upper -= ksize - sizeof(indx_t);
7716 mc->mc_ki[mc->mc_top] = x;
7717 mc->mc_pg[mc->mc_top] = rp;
7720 int psize, nsize, k;
7721 /* Maximum free space in an empty page */
7722 pmax = env->me_psize - PAGEHDRSZ;
7724 nsize = mdb_leaf_size(env, newkey, newdata);
7726 nsize = mdb_branch_size(env, newkey);
7727 nsize = EVEN(nsize);
7729 /* grab a page to hold a temporary copy */
7730 copy = mdb_page_malloc(mc->mc_txn, 1);
7735 copy->mp_pgno = mp->mp_pgno;
7736 copy->mp_flags = mp->mp_flags;
7737 copy->mp_lower = PAGEHDRSZ;
7738 copy->mp_upper = env->me_psize;
7740 /* prepare to insert */
7741 for (i=0, j=0; i<nkeys; i++) {
7743 copy->mp_ptrs[j++] = 0;
7745 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7748 /* When items are relatively large the split point needs
7749 * to be checked, because being off-by-one will make the
7750 * difference between success or failure in mdb_node_add.
7752 * It's also relevant if a page happens to be laid out
7753 * such that one half of its nodes are all "small" and
7754 * the other half of its nodes are "large." If the new
7755 * item is also "large" and falls on the half with
7756 * "large" nodes, it also may not fit.
7758 * As a final tweak, if the new item goes on the last
7759 * spot on the page (and thus, onto the new page), bias
7760 * the split so the new page is emptier than the old page.
7761 * This yields better packing during sequential inserts.
7763 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7764 /* Find split point */
7766 if (newindx <= split_indx || newindx >= nkeys) {
7768 k = newindx >= nkeys ? nkeys : split_indx+2;
7773 for (; i!=k; i+=j) {
7778 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7779 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7781 if (F_ISSET(node->mn_flags, F_BIGDATA))
7782 psize += sizeof(pgno_t);
7784 psize += NODEDSZ(node);
7786 psize = EVEN(psize);
7788 if (psize > pmax || i == k-j) {
7789 split_indx = i + (j<0);
7794 if (split_indx == newindx) {
7795 sepkey.mv_size = newkey->mv_size;
7796 sepkey.mv_data = newkey->mv_data;
7798 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7799 sepkey.mv_size = node->mn_ksize;
7800 sepkey.mv_data = NODEKEY(node);
7805 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7807 /* Copy separator key to the parent.
7809 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7813 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7818 if (mn.mc_snum == mc->mc_snum) {
7819 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7820 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7821 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7822 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7827 /* Right page might now have changed parent.
7828 * Check if left page also changed parent.
7830 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7831 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7832 for (i=0; i<ptop; i++) {
7833 mc->mc_pg[i] = mn.mc_pg[i];
7834 mc->mc_ki[i] = mn.mc_ki[i];
7836 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7837 if (mn.mc_ki[ptop]) {
7838 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7840 /* find right page's left sibling */
7841 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7842 mdb_cursor_sibling(mc, 0);
7847 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7850 mc->mc_flags ^= C_SPLITTING;
7851 if (rc != MDB_SUCCESS) {
7854 if (nflags & MDB_APPEND) {
7855 mc->mc_pg[mc->mc_top] = rp;
7856 mc->mc_ki[mc->mc_top] = 0;
7857 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7860 for (i=0; i<mc->mc_top; i++)
7861 mc->mc_ki[i] = mn.mc_ki[i];
7862 } else if (!IS_LEAF2(mp)) {
7864 mc->mc_pg[mc->mc_top] = rp;
7869 rkey.mv_data = newkey->mv_data;
7870 rkey.mv_size = newkey->mv_size;
7876 /* Update index for the new key. */
7877 mc->mc_ki[mc->mc_top] = j;
7879 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7880 rkey.mv_data = NODEKEY(node);
7881 rkey.mv_size = node->mn_ksize;
7883 xdata.mv_data = NODEDATA(node);
7884 xdata.mv_size = NODEDSZ(node);
7887 pgno = NODEPGNO(node);
7888 flags = node->mn_flags;
7891 if (!IS_LEAF(mp) && j == 0) {
7892 /* First branch index doesn't need key data. */
7896 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7902 mc->mc_pg[mc->mc_top] = copy;
7907 } while (i != split_indx);
7909 nkeys = NUMKEYS(copy);
7910 for (i=0; i<nkeys; i++)
7911 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7912 mp->mp_lower = copy->mp_lower;
7913 mp->mp_upper = copy->mp_upper;
7914 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7915 env->me_psize - copy->mp_upper);
7917 /* reset back to original page */
7918 if (newindx < split_indx) {
7919 mc->mc_pg[mc->mc_top] = mp;
7920 if (nflags & MDB_RESERVE) {
7921 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7922 if (!(node->mn_flags & F_BIGDATA))
7923 newdata->mv_data = NODEDATA(node);
7926 mc->mc_pg[mc->mc_top] = rp;
7928 /* Make sure mc_ki is still valid.
7930 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7931 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7932 for (i=0; i<=ptop; i++) {
7933 mc->mc_pg[i] = mn.mc_pg[i];
7934 mc->mc_ki[i] = mn.mc_ki[i];
7941 /* Adjust other cursors pointing to mp */
7942 MDB_cursor *m2, *m3;
7943 MDB_dbi dbi = mc->mc_dbi;
7944 int fixup = NUMKEYS(mp);
7946 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7947 if (mc->mc_flags & C_SUB)
7948 m3 = &m2->mc_xcursor->mx_cursor;
7953 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7955 if (m3->mc_flags & C_SPLITTING)
7960 for (k=m3->mc_top; k>=0; k--) {
7961 m3->mc_ki[k+1] = m3->mc_ki[k];
7962 m3->mc_pg[k+1] = m3->mc_pg[k];
7964 if (m3->mc_ki[0] >= split_indx) {
7969 m3->mc_pg[0] = mc->mc_pg[0];
7973 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7974 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7975 m3->mc_ki[mc->mc_top]++;
7976 if (m3->mc_ki[mc->mc_top] >= fixup) {
7977 m3->mc_pg[mc->mc_top] = rp;
7978 m3->mc_ki[mc->mc_top] -= fixup;
7979 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7981 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7982 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7987 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
7990 if (copy) /* tmp page */
7991 mdb_page_free(env, copy);
7993 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7998 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7999 MDB_val *key, MDB_val *data, unsigned int flags)
8004 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8007 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8010 mdb_cursor_init(&mc, txn, dbi, &mx);
8011 return mdb_cursor_put(&mc, key, data, flags);
8015 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8017 if ((flag & CHANGEABLE) != flag)
8020 env->me_flags |= flag;
8022 env->me_flags &= ~flag;
8027 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8032 *arg = env->me_flags;
8037 mdb_env_set_userctx(MDB_env *env, void *ctx)
8041 env->me_userctx = ctx;
8046 mdb_env_get_userctx(MDB_env *env)
8048 return env ? env->me_userctx : NULL;
8052 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8057 env->me_assert_func = func;
8063 mdb_env_get_path(MDB_env *env, const char **arg)
8068 *arg = env->me_path;
8073 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8082 /** Common code for #mdb_stat() and #mdb_env_stat().
8083 * @param[in] env the environment to operate in.
8084 * @param[in] db the #MDB_db record containing the stats to return.
8085 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8086 * @return 0, this function always succeeds.
8089 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8091 arg->ms_psize = env->me_psize;
8092 arg->ms_depth = db->md_depth;
8093 arg->ms_branch_pages = db->md_branch_pages;
8094 arg->ms_leaf_pages = db->md_leaf_pages;
8095 arg->ms_overflow_pages = db->md_overflow_pages;
8096 arg->ms_entries = db->md_entries;
8101 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8105 if (env == NULL || arg == NULL)
8108 toggle = mdb_env_pick_meta(env);
8110 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8114 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8118 if (env == NULL || arg == NULL)
8121 toggle = mdb_env_pick_meta(env);
8122 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
8123 arg->me_mapsize = env->me_mapsize;
8124 arg->me_maxreaders = env->me_maxreaders;
8126 /* me_numreaders may be zero if this process never used any readers. Use
8127 * the shared numreader count if it exists.
8129 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8131 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8132 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8136 /** Set the default comparison functions for a database.
8137 * Called immediately after a database is opened to set the defaults.
8138 * The user can then override them with #mdb_set_compare() or
8139 * #mdb_set_dupsort().
8140 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8141 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8144 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8146 uint16_t f = txn->mt_dbs[dbi].md_flags;
8148 txn->mt_dbxs[dbi].md_cmp =
8149 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8150 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8152 txn->mt_dbxs[dbi].md_dcmp =
8153 !(f & MDB_DUPSORT) ? 0 :
8154 ((f & MDB_INTEGERDUP)
8155 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8156 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8159 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8164 int rc, dbflag, exact;
8165 unsigned int unused = 0;
8168 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8169 mdb_default_cmp(txn, FREE_DBI);
8172 if ((flags & VALID_FLAGS) != flags)
8174 if (txn->mt_flags & MDB_TXN_ERROR)
8180 if (flags & PERSISTENT_FLAGS) {
8181 uint16_t f2 = flags & PERSISTENT_FLAGS;
8182 /* make sure flag changes get committed */
8183 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8184 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8185 txn->mt_flags |= MDB_TXN_DIRTY;
8188 mdb_default_cmp(txn, MAIN_DBI);
8192 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8193 mdb_default_cmp(txn, MAIN_DBI);
8196 /* Is the DB already open? */
8198 for (i=2; i<txn->mt_numdbs; i++) {
8199 if (!txn->mt_dbxs[i].md_name.mv_size) {
8200 /* Remember this free slot */
8201 if (!unused) unused = i;
8204 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8205 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8211 /* If no free slot and max hit, fail */
8212 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8213 return MDB_DBS_FULL;
8215 /* Cannot mix named databases with some mainDB flags */
8216 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8217 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8219 /* Find the DB info */
8220 dbflag = DB_NEW|DB_VALID;
8223 key.mv_data = (void *)name;
8224 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8225 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8226 if (rc == MDB_SUCCESS) {
8227 /* make sure this is actually a DB */
8228 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8229 if (!(node->mn_flags & F_SUBDATA))
8230 return MDB_INCOMPATIBLE;
8231 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8232 /* Create if requested */
8234 data.mv_size = sizeof(MDB_db);
8235 data.mv_data = &dummy;
8236 memset(&dummy, 0, sizeof(dummy));
8237 dummy.md_root = P_INVALID;
8238 dummy.md_flags = flags & PERSISTENT_FLAGS;
8239 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8243 /* OK, got info, add to table */
8244 if (rc == MDB_SUCCESS) {
8245 unsigned int slot = unused ? unused : txn->mt_numdbs;
8246 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8247 txn->mt_dbxs[slot].md_name.mv_size = len;
8248 txn->mt_dbxs[slot].md_rel = NULL;
8249 txn->mt_dbflags[slot] = dbflag;
8250 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8252 mdb_default_cmp(txn, slot);
8261 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8263 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8266 if (txn->mt_flags & MDB_TXN_ERROR)
8269 if (txn->mt_dbflags[dbi] & DB_STALE) {
8272 /* Stale, must read the DB's root. cursor_init does it for us. */
8273 mdb_cursor_init(&mc, txn, dbi, &mx);
8275 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8278 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8281 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8283 ptr = env->me_dbxs[dbi].md_name.mv_data;
8284 env->me_dbxs[dbi].md_name.mv_data = NULL;
8285 env->me_dbxs[dbi].md_name.mv_size = 0;
8286 env->me_dbflags[dbi] = 0;
8290 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8292 /* We could return the flags for the FREE_DBI too but what's the point? */
8293 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8295 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8299 /** Add all the DB's pages to the free list.
8300 * @param[in] mc Cursor on the DB to free.
8301 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8302 * @return 0 on success, non-zero on failure.
8305 mdb_drop0(MDB_cursor *mc, int subs)
8309 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8310 if (rc == MDB_SUCCESS) {
8311 MDB_txn *txn = mc->mc_txn;
8316 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8317 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8320 mdb_cursor_copy(mc, &mx);
8321 while (mc->mc_snum > 0) {
8322 MDB_page *mp = mc->mc_pg[mc->mc_top];
8323 unsigned n = NUMKEYS(mp);
8325 for (i=0; i<n; i++) {
8326 ni = NODEPTR(mp, i);
8327 if (ni->mn_flags & F_BIGDATA) {
8330 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8331 rc = mdb_page_get(txn, pg, &omp, NULL);
8334 mdb_cassert(mc, IS_OVERFLOW(omp));
8335 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8339 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8340 mdb_xcursor_init1(mc, ni);
8341 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8347 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8349 for (i=0; i<n; i++) {
8351 ni = NODEPTR(mp, i);
8354 mdb_midl_xappend(txn->mt_free_pgs, pg);
8359 mc->mc_ki[mc->mc_top] = i;
8360 rc = mdb_cursor_sibling(mc, 1);
8362 if (rc != MDB_NOTFOUND)
8364 /* no more siblings, go back to beginning
8365 * of previous level.
8369 for (i=1; i<mc->mc_snum; i++) {
8371 mc->mc_pg[i] = mx.mc_pg[i];
8376 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8379 txn->mt_flags |= MDB_TXN_ERROR;
8380 } else if (rc == MDB_NOTFOUND) {
8386 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8388 MDB_cursor *mc, *m2;
8391 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8394 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8397 rc = mdb_cursor_open(txn, dbi, &mc);
8401 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8402 /* Invalidate the dropped DB's cursors */
8403 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8404 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8408 /* Can't delete the main DB */
8409 if (del && dbi > MAIN_DBI) {
8410 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
8412 txn->mt_dbflags[dbi] = DB_STALE;
8413 mdb_dbi_close(txn->mt_env, dbi);
8415 txn->mt_flags |= MDB_TXN_ERROR;
8418 /* reset the DB record, mark it dirty */
8419 txn->mt_dbflags[dbi] |= DB_DIRTY;
8420 txn->mt_dbs[dbi].md_depth = 0;
8421 txn->mt_dbs[dbi].md_branch_pages = 0;
8422 txn->mt_dbs[dbi].md_leaf_pages = 0;
8423 txn->mt_dbs[dbi].md_overflow_pages = 0;
8424 txn->mt_dbs[dbi].md_entries = 0;
8425 txn->mt_dbs[dbi].md_root = P_INVALID;
8427 txn->mt_flags |= MDB_TXN_DIRTY;
8430 mdb_cursor_close(mc);
8434 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8436 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8439 txn->mt_dbxs[dbi].md_cmp = cmp;
8443 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8445 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8448 txn->mt_dbxs[dbi].md_dcmp = cmp;
8452 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8454 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8457 txn->mt_dbxs[dbi].md_rel = rel;
8461 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8463 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8466 txn->mt_dbxs[dbi].md_relctx = ctx;
8470 int mdb_env_get_maxkeysize(MDB_env *env)
8472 return ENV_MAXKEY(env);
8475 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8477 unsigned int i, rdrs;
8480 int rc = 0, first = 1;
8484 if (!env->me_txns) {
8485 return func("(no reader locks)\n", ctx);
8487 rdrs = env->me_txns->mti_numreaders;
8488 mr = env->me_txns->mti_readers;
8489 for (i=0; i<rdrs; i++) {
8491 txnid_t txnid = mr[i].mr_txnid;
8492 sprintf(buf, txnid == (txnid_t)-1 ?
8493 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
8494 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
8497 rc = func(" pid thread txnid\n", ctx);
8501 rc = func(buf, ctx);
8507 rc = func("(no active readers)\n", ctx);
8512 /** Insert pid into list if not already present.
8513 * return -1 if already present.
8515 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8517 /* binary search of pid in list */
8519 unsigned cursor = 1;
8521 unsigned n = ids[0];
8524 unsigned pivot = n >> 1;
8525 cursor = base + pivot + 1;
8526 val = pid - ids[cursor];
8531 } else if ( val > 0 ) {
8536 /* found, so it's a duplicate */
8545 for (n = ids[0]; n > cursor; n--)
8551 int mdb_reader_check(MDB_env *env, int *dead)
8553 unsigned int i, j, rdrs;
8555 MDB_PID_T *pids, pid;
8564 rdrs = env->me_txns->mti_numreaders;
8565 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8569 mr = env->me_txns->mti_readers;
8570 for (i=0; i<rdrs; i++) {
8571 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8573 if (mdb_pid_insert(pids, pid) == 0) {
8574 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8576 /* Recheck, a new process may have reused pid */
8577 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8578 for (j=i; j<rdrs; j++)
8579 if (mr[j].mr_pid == pid) {
8580 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8581 (unsigned) pid, mr[j].mr_txnid));
8586 UNLOCK_MUTEX_R(env);