2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
42 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
43 * as int64 which is wrong. MSVC doesn't define it at all, so just
48 # include <sys/param.h>
50 # define LITTLE_ENDIAN 1234
51 # define BIG_ENDIAN 4321
52 # define BYTE_ORDER LITTLE_ENDIAN
54 # define SSIZE_MAX INT_MAX
58 #define MDB_PID_T pid_t
59 #include <sys/param.h>
62 #ifdef HAVE_SYS_FILE_H
78 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
79 #include <netinet/in.h>
80 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
83 #if defined(__APPLE__) || defined (BSD)
84 # define MDB_USE_POSIX_SEM 1
85 # define MDB_FDATASYNC fsync
86 #elif defined(ANDROID)
87 # define MDB_FDATASYNC fsync
92 #ifdef MDB_USE_POSIX_SEM
93 # define MDB_USE_HASH 1
94 #include <semaphore.h>
99 #include <valgrind/memcheck.h>
100 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
101 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
102 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
103 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
104 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
106 #define VGMEMP_CREATE(h,r,z)
107 #define VGMEMP_ALLOC(h,a,s)
108 #define VGMEMP_FREE(h,a)
109 #define VGMEMP_DESTROY(h)
110 #define VGMEMP_DEFINED(a,s)
114 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
115 /* Solaris just defines one or the other */
116 # define LITTLE_ENDIAN 1234
117 # define BIG_ENDIAN 4321
118 # ifdef _LITTLE_ENDIAN
119 # define BYTE_ORDER LITTLE_ENDIAN
121 # define BYTE_ORDER BIG_ENDIAN
124 # define BYTE_ORDER __BYTE_ORDER
128 #ifndef LITTLE_ENDIAN
129 #define LITTLE_ENDIAN __LITTLE_ENDIAN
132 #define BIG_ENDIAN __BIG_ENDIAN
135 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
136 #define MISALIGNED_OK 1
142 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
143 # error "Unknown or unsupported endianness (BYTE_ORDER)"
144 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
145 # error "Two's complement, reasonably sized integer types, please"
148 /** @defgroup internal MDB Internals
151 /** @defgroup compat Compatibility Macros
152 * A bunch of macros to minimize the amount of platform-specific ifdefs
153 * needed throughout the rest of the code. When the features this library
154 * needs are similar enough to POSIX to be hidden in a one-or-two line
155 * replacement, this macro approach is used.
159 /** Wrapper around __func__, which is a C99 feature */
160 #if __STDC_VERSION__ >= 199901L
161 # define mdb_func_ __func__
162 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
163 # define mdb_func_ __FUNCTION__
165 /* If a debug message says <mdb_unknown>(), update the #if statements above */
166 # define mdb_func_ "<mdb_unknown>"
170 #define MDB_USE_HASH 1
171 #define MDB_PIDLOCK 0
172 #define pthread_t DWORD
173 #define pthread_mutex_t HANDLE
174 #define pthread_key_t DWORD
175 #define pthread_self() GetCurrentThreadId()
176 #define pthread_key_create(x,y) \
177 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
178 #define pthread_key_delete(x) TlsFree(x)
179 #define pthread_getspecific(x) TlsGetValue(x)
180 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
181 #define pthread_mutex_unlock(x) ReleaseMutex(x)
182 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
183 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
184 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
185 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
186 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
187 #define getpid() GetCurrentProcessId()
188 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
189 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
190 #define ErrCode() GetLastError()
191 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
192 #define close(fd) (CloseHandle(fd) ? 0 : -1)
193 #define munmap(ptr,len) UnmapViewOfFile(ptr)
194 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
195 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
197 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
202 #define Z "z" /**< printf format modifier for size_t */
204 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
205 #define MDB_PIDLOCK 1
207 #ifdef MDB_USE_POSIX_SEM
209 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
210 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
211 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
212 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
215 mdb_sem_wait(sem_t *sem)
218 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
223 /** Lock the reader mutex.
225 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
226 /** Unlock the reader mutex.
228 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
230 /** Lock the writer mutex.
231 * Only a single write transaction is allowed at a time. Other writers
232 * will block waiting for this mutex.
234 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
235 /** Unlock the writer mutex.
237 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
238 #endif /* MDB_USE_POSIX_SEM */
240 /** Get the error code for the last failed system function.
242 #define ErrCode() errno
244 /** An abstraction for a file handle.
245 * On POSIX systems file handles are small integers. On Windows
246 * they're opaque pointers.
250 /** A value for an invalid file handle.
251 * Mainly used to initialize file variables and signify that they are
254 #define INVALID_HANDLE_VALUE (-1)
256 /** Get the size of a memory page for the system.
257 * This is the basic size that the platform's memory manager uses, and is
258 * fundamental to the use of memory-mapped files.
260 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
263 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
266 #define MNAME_LEN (sizeof(pthread_mutex_t))
272 /** A flag for opening a file and requesting synchronous data writes.
273 * This is only used when writing a meta page. It's not strictly needed;
274 * we could just do a normal write and then immediately perform a flush.
275 * But if this flag is available it saves us an extra system call.
277 * @note If O_DSYNC is undefined but exists in /usr/include,
278 * preferably set some compiler flag to get the definition.
279 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
282 # define MDB_DSYNC O_DSYNC
286 /** Function for flushing the data of a file. Define this to fsync
287 * if fdatasync() is not supported.
289 #ifndef MDB_FDATASYNC
290 # define MDB_FDATASYNC fdatasync
294 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
305 /** A page number in the database.
306 * Note that 64 bit page numbers are overkill, since pages themselves
307 * already represent 12-13 bits of addressable memory, and the OS will
308 * always limit applications to a maximum of 63 bits of address space.
310 * @note In the #MDB_node structure, we only store 48 bits of this value,
311 * which thus limits us to only 60 bits of addressable data.
313 typedef MDB_ID pgno_t;
315 /** A transaction ID.
316 * See struct MDB_txn.mt_txnid for details.
318 typedef MDB_ID txnid_t;
320 /** @defgroup debug Debug Macros
324 /** Enable debug output. Needs variable argument macros (a C99 feature).
325 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
326 * read from and written to the database (used for free space management).
332 static int mdb_debug;
333 static txnid_t mdb_debug_start;
335 /** Print a debug message with printf formatting.
336 * Requires double parenthesis around 2 or more args.
338 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
339 # define DPRINTF0(fmt, ...) \
340 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
342 # define DPRINTF(args) ((void) 0)
344 /** Print a debug string.
345 * The string is printed literally, with no format processing.
347 #define DPUTS(arg) DPRINTF(("%s", arg))
348 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
350 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
353 /** @brief The maximum size of a database page.
355 * This is 32k, since it must fit in #MDB_page.%mp_upper.
357 * LMDB will use database pages < OS pages if needed.
358 * That causes more I/O in write transactions: The OS must
359 * know (read) the whole page before writing a partial page.
361 * Note that we don't currently support Huge pages. On Linux,
362 * regular data files cannot use Huge pages, and in general
363 * Huge pages aren't actually pageable. We rely on the OS
364 * demand-pager to read our data and page it out when memory
365 * pressure from other processes is high. So until OSs have
366 * actual paging support for Huge pages, they're not viable.
368 #define MAX_PAGESIZE 0x8000
370 /** The minimum number of keys required in a database page.
371 * Setting this to a larger value will place a smaller bound on the
372 * maximum size of a data item. Data items larger than this size will
373 * be pushed into overflow pages instead of being stored directly in
374 * the B-tree node. This value used to default to 4. With a page size
375 * of 4096 bytes that meant that any item larger than 1024 bytes would
376 * go into an overflow page. That also meant that on average 2-3KB of
377 * each overflow page was wasted space. The value cannot be lower than
378 * 2 because then there would no longer be a tree structure. With this
379 * value, items larger than 2KB will go into overflow pages, and on
380 * average only 1KB will be wasted.
382 #define MDB_MINKEYS 2
384 /** A stamp that identifies a file as an MDB file.
385 * There's nothing special about this value other than that it is easily
386 * recognizable, and it will reflect any byte order mismatches.
388 #define MDB_MAGIC 0xBEEFC0DE
390 /** The version number for a database's datafile format. */
391 #define MDB_DATA_VERSION 1
392 /** The version number for a database's lockfile format. */
393 #define MDB_LOCK_VERSION 1
395 /** @brief The max size of a key we can write, or 0 for dynamic max.
397 * Define this as 0 to compute the max from the page size. 511
398 * is default for backwards compat: liblmdb <= 0.9.10 can break
399 * when modifying a DB with keys/dupsort data bigger than its max.
401 * Data items in an #MDB_DUPSORT database are also limited to
402 * this size, since they're actually keys of a sub-DB. Keys and
403 * #MDB_DUPSORT data items must fit on a node in a regular page.
405 #ifndef MDB_MAXKEYSIZE
406 #define MDB_MAXKEYSIZE 511
409 /** The maximum size of a key we can write to the environment. */
411 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
413 #define ENV_MAXKEY(env) ((env)->me_maxkey)
416 /** @brief The maximum size of a data item.
418 * We only store a 32 bit value for node sizes.
420 #define MAXDATASIZE 0xffffffffUL
423 /** Key size which fits in a #DKBUF.
426 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
429 * This is used for printing a hex dump of a key's contents.
431 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
432 /** Display a key in hex.
434 * Invoke a function to display a key in hex.
436 #define DKEY(x) mdb_dkey(x, kbuf)
442 /** An invalid page number.
443 * Mainly used to denote an empty tree.
445 #define P_INVALID (~(pgno_t)0)
447 /** Test if the flags \b f are set in a flag word \b w. */
448 #define F_ISSET(w, f) (((w) & (f)) == (f))
450 /** Round \b n up to an even number. */
451 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
453 /** Used for offsets within a single page.
454 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
457 typedef uint16_t indx_t;
459 /** Default size of memory map.
460 * This is certainly too small for any actual applications. Apps should always set
461 * the size explicitly using #mdb_env_set_mapsize().
463 #define DEFAULT_MAPSIZE 1048576
465 /** @defgroup readers Reader Lock Table
466 * Readers don't acquire any locks for their data access. Instead, they
467 * simply record their transaction ID in the reader table. The reader
468 * mutex is needed just to find an empty slot in the reader table. The
469 * slot's address is saved in thread-specific data so that subsequent read
470 * transactions started by the same thread need no further locking to proceed.
472 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
474 * No reader table is used if the database is on a read-only filesystem, or
475 * if #MDB_NOLOCK is set.
477 * Since the database uses multi-version concurrency control, readers don't
478 * actually need any locking. This table is used to keep track of which
479 * readers are using data from which old transactions, so that we'll know
480 * when a particular old transaction is no longer in use. Old transactions
481 * that have discarded any data pages can then have those pages reclaimed
482 * for use by a later write transaction.
484 * The lock table is constructed such that reader slots are aligned with the
485 * processor's cache line size. Any slot is only ever used by one thread.
486 * This alignment guarantees that there will be no contention or cache
487 * thrashing as threads update their own slot info, and also eliminates
488 * any need for locking when accessing a slot.
490 * A writer thread will scan every slot in the table to determine the oldest
491 * outstanding reader transaction. Any freed pages older than this will be
492 * reclaimed by the writer. The writer doesn't use any locks when scanning
493 * this table. This means that there's no guarantee that the writer will
494 * see the most up-to-date reader info, but that's not required for correct
495 * operation - all we need is to know the upper bound on the oldest reader,
496 * we don't care at all about the newest reader. So the only consequence of
497 * reading stale information here is that old pages might hang around a
498 * while longer before being reclaimed. That's actually good anyway, because
499 * the longer we delay reclaiming old pages, the more likely it is that a
500 * string of contiguous pages can be found after coalescing old pages from
501 * many old transactions together.
504 /** Number of slots in the reader table.
505 * This value was chosen somewhat arbitrarily. 126 readers plus a
506 * couple mutexes fit exactly into 8KB on my development machine.
507 * Applications should set the table size using #mdb_env_set_maxreaders().
509 #define DEFAULT_READERS 126
511 /** The size of a CPU cache line in bytes. We want our lock structures
512 * aligned to this size to avoid false cache line sharing in the
514 * This value works for most CPUs. For Itanium this should be 128.
520 /** The information we store in a single slot of the reader table.
521 * In addition to a transaction ID, we also record the process and
522 * thread ID that owns a slot, so that we can detect stale information,
523 * e.g. threads or processes that went away without cleaning up.
524 * @note We currently don't check for stale records. We simply re-init
525 * the table when we know that we're the only process opening the
528 typedef struct MDB_rxbody {
529 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
530 * Multiple readers that start at the same time will probably have the
531 * same ID here. Again, it's not important to exclude them from
532 * anything; all we need to know is which version of the DB they
533 * started from so we can avoid overwriting any data used in that
534 * particular version.
537 /** The process ID of the process owning this reader txn. */
539 /** The thread ID of the thread owning this txn. */
543 /** The actual reader record, with cacheline padding. */
544 typedef struct MDB_reader {
547 /** shorthand for mrb_txnid */
548 #define mr_txnid mru.mrx.mrb_txnid
549 #define mr_pid mru.mrx.mrb_pid
550 #define mr_tid mru.mrx.mrb_tid
551 /** cache line alignment */
552 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
556 /** The header for the reader table.
557 * The table resides in a memory-mapped file. (This is a different file
558 * than is used for the main database.)
560 * For POSIX the actual mutexes reside in the shared memory of this
561 * mapped file. On Windows, mutexes are named objects allocated by the
562 * kernel; we store the mutex names in this mapped file so that other
563 * processes can grab them. This same approach is also used on
564 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
565 * process-shared POSIX mutexes. For these cases where a named object
566 * is used, the object name is derived from a 64 bit FNV hash of the
567 * environment pathname. As such, naming collisions are extremely
568 * unlikely. If a collision occurs, the results are unpredictable.
570 typedef struct MDB_txbody {
571 /** Stamp identifying this as an MDB file. It must be set
574 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
576 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
577 char mtb_rmname[MNAME_LEN];
579 /** Mutex protecting access to this table.
580 * This is the reader lock that #LOCK_MUTEX_R acquires.
582 pthread_mutex_t mtb_mutex;
584 /** The ID of the last transaction committed to the database.
585 * This is recorded here only for convenience; the value can always
586 * be determined by reading the main database meta pages.
589 /** The number of slots that have been used in the reader table.
590 * This always records the maximum count, it is not decremented
591 * when readers release their slots.
593 unsigned mtb_numreaders;
596 /** The actual reader table definition. */
597 typedef struct MDB_txninfo {
600 #define mti_magic mt1.mtb.mtb_magic
601 #define mti_format mt1.mtb.mtb_format
602 #define mti_mutex mt1.mtb.mtb_mutex
603 #define mti_rmname mt1.mtb.mtb_rmname
604 #define mti_txnid mt1.mtb.mtb_txnid
605 #define mti_numreaders mt1.mtb.mtb_numreaders
606 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
609 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
610 char mt2_wmname[MNAME_LEN];
611 #define mti_wmname mt2.mt2_wmname
613 pthread_mutex_t mt2_wmutex;
614 #define mti_wmutex mt2.mt2_wmutex
616 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
618 MDB_reader mti_readers[1];
621 /** Lockfile format signature: version, features and field layout */
622 #define MDB_LOCK_FORMAT \
624 ((MDB_LOCK_VERSION) \
625 /* Flags which describe functionality */ \
626 + (((MDB_PIDLOCK) != 0) << 16)))
629 /** Common header for all page types.
630 * Overflow records occupy a number of contiguous pages with no
631 * headers on any page after the first.
633 typedef struct MDB_page {
634 #define mp_pgno mp_p.p_pgno
635 #define mp_next mp_p.p_next
637 pgno_t p_pgno; /**< page number */
638 void * p_next; /**< for in-memory list of freed structs */
641 /** @defgroup mdb_page Page Flags
643 * Flags for the page headers.
646 #define P_BRANCH 0x01 /**< branch page */
647 #define P_LEAF 0x02 /**< leaf page */
648 #define P_OVERFLOW 0x04 /**< overflow page */
649 #define P_META 0x08 /**< meta page */
650 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
651 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
652 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
653 #define P_KEEP 0x8000 /**< leave this page alone during spill */
655 uint16_t mp_flags; /**< @ref mdb_page */
656 #define mp_lower mp_pb.pb.pb_lower
657 #define mp_upper mp_pb.pb.pb_upper
658 #define mp_pages mp_pb.pb_pages
661 indx_t pb_lower; /**< lower bound of free space */
662 indx_t pb_upper; /**< upper bound of free space */
664 uint32_t pb_pages; /**< number of overflow pages */
666 indx_t mp_ptrs[1]; /**< dynamic size */
669 /** Size of the page header, excluding dynamic data at the end */
670 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
672 /** Address of first usable data byte in a page, after the header */
673 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
675 /** Number of nodes on a page */
676 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
678 /** The amount of space remaining in the page */
679 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
681 /** The percentage of space used in the page, in tenths of a percent. */
682 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
683 ((env)->me_psize - PAGEHDRSZ))
684 /** The minimum page fill factor, in tenths of a percent.
685 * Pages emptier than this are candidates for merging.
687 #define FILL_THRESHOLD 250
689 /** Test if a page is a leaf page */
690 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
691 /** Test if a page is a LEAF2 page */
692 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
693 /** Test if a page is a branch page */
694 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
695 /** Test if a page is an overflow page */
696 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
697 /** Test if a page is a sub page */
698 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
700 /** The number of overflow pages needed to store the given size. */
701 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
703 /** Header for a single key/data pair within a page.
704 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
705 * We guarantee 2-byte alignment for 'MDB_node's.
707 typedef struct MDB_node {
708 /** lo and hi are used for data size on leaf nodes and for
709 * child pgno on branch nodes. On 64 bit platforms, flags
710 * is also used for pgno. (Branch nodes have no flags).
711 * They are in host byte order in case that lets some
712 * accesses be optimized into a 32-bit word access.
714 #if BYTE_ORDER == LITTLE_ENDIAN
715 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
717 unsigned short mn_hi, mn_lo;
719 /** @defgroup mdb_node Node Flags
721 * Flags for node headers.
724 #define F_BIGDATA 0x01 /**< data put on overflow page */
725 #define F_SUBDATA 0x02 /**< data is a sub-database */
726 #define F_DUPDATA 0x04 /**< data has duplicates */
728 /** valid flags for #mdb_node_add() */
729 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
732 unsigned short mn_flags; /**< @ref mdb_node */
733 unsigned short mn_ksize; /**< key size */
734 char mn_data[1]; /**< key and data are appended here */
737 /** Size of the node header, excluding dynamic data at the end */
738 #define NODESIZE offsetof(MDB_node, mn_data)
740 /** Bit position of top word in page number, for shifting mn_flags */
741 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
743 /** Size of a node in a branch page with a given key.
744 * This is just the node header plus the key, there is no data.
746 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
748 /** Size of a node in a leaf page with a given key and data.
749 * This is node header plus key plus data size.
751 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
753 /** Address of node \b i in page \b p */
754 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
756 /** Address of the key for the node */
757 #define NODEKEY(node) (void *)((node)->mn_data)
759 /** Address of the data for a node */
760 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
762 /** Get the page number pointed to by a branch node */
763 #define NODEPGNO(node) \
764 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
765 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
766 /** Set the page number in a branch node */
767 #define SETPGNO(node,pgno) do { \
768 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
769 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
771 /** Get the size of the data in a leaf node */
772 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
773 /** Set the size of the data for a leaf node */
774 #define SETDSZ(node,size) do { \
775 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
776 /** The size of a key in a node */
777 #define NODEKSZ(node) ((node)->mn_ksize)
779 /** Copy a page number from src to dst */
781 #define COPY_PGNO(dst,src) dst = src
783 #if SIZE_MAX > 4294967295UL
784 #define COPY_PGNO(dst,src) do { \
785 unsigned short *s, *d; \
786 s = (unsigned short *)&(src); \
787 d = (unsigned short *)&(dst); \
794 #define COPY_PGNO(dst,src) do { \
795 unsigned short *s, *d; \
796 s = (unsigned short *)&(src); \
797 d = (unsigned short *)&(dst); \
803 /** The address of a key in a LEAF2 page.
804 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
805 * There are no node headers, keys are stored contiguously.
807 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
809 /** Set the \b node's key into \b keyptr, if requested. */
810 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
811 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
813 /** Set the \b node's key into \b key. */
814 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
816 /** Information about a single database in the environment. */
817 typedef struct MDB_db {
818 uint32_t md_pad; /**< also ksize for LEAF2 pages */
819 uint16_t md_flags; /**< @ref mdb_dbi_open */
820 uint16_t md_depth; /**< depth of this tree */
821 pgno_t md_branch_pages; /**< number of internal pages */
822 pgno_t md_leaf_pages; /**< number of leaf pages */
823 pgno_t md_overflow_pages; /**< number of overflow pages */
824 size_t md_entries; /**< number of data items */
825 pgno_t md_root; /**< the root page of this tree */
828 /** mdb_dbi_open flags */
829 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
830 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
831 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
832 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
834 /** Handle for the DB used to track free pages. */
836 /** Handle for the default DB. */
839 /** Meta page content.
840 * A meta page is the start point for accessing a database snapshot.
841 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
843 typedef struct MDB_meta {
844 /** Stamp identifying this as an MDB file. It must be set
847 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
849 void *mm_address; /**< address for fixed mapping */
850 size_t mm_mapsize; /**< size of mmap region */
851 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
852 /** The size of pages used in this DB */
853 #define mm_psize mm_dbs[0].md_pad
854 /** Any persistent environment flags. @ref mdb_env */
855 #define mm_flags mm_dbs[0].md_flags
856 pgno_t mm_last_pg; /**< last used page in file */
857 txnid_t mm_txnid; /**< txnid that committed this page */
860 /** Buffer for a stack-allocated meta page.
861 * The members define size and alignment, and silence type
862 * aliasing warnings. They are not used directly; that could
863 * mean incorrectly using several union members in parallel.
865 typedef union MDB_metabuf {
868 char mm_pad[PAGEHDRSZ];
873 /** Auxiliary DB info.
874 * The information here is mostly static/read-only. There is
875 * only a single copy of this record in the environment.
877 typedef struct MDB_dbx {
878 MDB_val md_name; /**< name of the database */
879 MDB_cmp_func *md_cmp; /**< function for comparing keys */
880 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
881 MDB_rel_func *md_rel; /**< user relocate function */
882 void *md_relctx; /**< user-provided context for md_rel */
885 /** A database transaction.
886 * Every operation requires a transaction handle.
889 MDB_txn *mt_parent; /**< parent of a nested txn */
890 MDB_txn *mt_child; /**< nested txn under this txn */
891 pgno_t mt_next_pgno; /**< next unallocated page */
892 /** The ID of this transaction. IDs are integers incrementing from 1.
893 * Only committed write transactions increment the ID. If a transaction
894 * aborts, the ID may be re-used by the next writer.
897 MDB_env *mt_env; /**< the DB environment */
898 /** The list of pages that became unused during this transaction.
901 /** The sorted list of dirty pages we temporarily wrote to disk
902 * because the dirty list was full. page numbers in here are
903 * shifted left by 1, deleted slots have the LSB set.
905 MDB_IDL mt_spill_pgs;
907 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
909 /** For read txns: This thread/txn's reader table slot, or NULL. */
912 /** Array of records for each DB known in the environment. */
914 /** Array of MDB_db records for each known DB */
916 /** @defgroup mt_dbflag Transaction DB Flags
920 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
921 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
922 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
923 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
925 /** In write txns, array of cursors for each DB */
926 MDB_cursor **mt_cursors;
927 /** Array of flags for each DB */
928 unsigned char *mt_dbflags;
929 /** Number of DB records in use. This number only ever increments;
930 * we don't decrement it when individual DB handles are closed.
934 /** @defgroup mdb_txn Transaction Flags
938 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
939 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
940 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
941 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
943 unsigned int mt_flags; /**< @ref mdb_txn */
944 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
945 * Includes ancestor txns' dirty pages not hidden by other txns'
946 * dirty/spilled pages. Thus commit(nested txn) has room to merge
947 * dirty_list into mt_parent after freeing hidden mt_parent pages.
949 unsigned int mt_dirty_room;
952 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
953 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
954 * raise this on a 64 bit machine.
956 #define CURSOR_STACK 32
960 /** Cursors are used for all DB operations.
961 * A cursor holds a path of (page pointer, key index) from the DB
962 * root to a position in the DB, plus other state. #MDB_DUPSORT
963 * cursors include an xcursor to the current data item. Write txns
964 * track their cursors and keep them up to date when data moves.
965 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
966 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
969 /** Next cursor on this DB in this txn */
971 /** Backup of the original cursor if this cursor is a shadow */
972 MDB_cursor *mc_backup;
973 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
974 struct MDB_xcursor *mc_xcursor;
975 /** The transaction that owns this cursor */
977 /** The database handle this cursor operates on */
979 /** The database record for this cursor */
981 /** The database auxiliary record for this cursor */
983 /** The @ref mt_dbflag for this database */
984 unsigned char *mc_dbflag;
985 unsigned short mc_snum; /**< number of pushed pages */
986 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
987 /** @defgroup mdb_cursor Cursor Flags
989 * Cursor state flags.
992 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
993 #define C_EOF 0x02 /**< No more data */
994 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
995 #define C_DEL 0x08 /**< last op was a cursor_del */
996 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
997 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
999 unsigned int mc_flags; /**< @ref mdb_cursor */
1000 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1001 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1004 /** Context for sorted-dup records.
1005 * We could have gone to a fully recursive design, with arbitrarily
1006 * deep nesting of sub-databases. But for now we only handle these
1007 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1009 typedef struct MDB_xcursor {
1010 /** A sub-cursor for traversing the Dup DB */
1011 MDB_cursor mx_cursor;
1012 /** The database record for this Dup DB */
1014 /** The auxiliary DB record for this Dup DB */
1016 /** The @ref mt_dbflag for this Dup DB */
1017 unsigned char mx_dbflag;
1020 /** State of FreeDB old pages, stored in the MDB_env */
1021 typedef struct MDB_pgstate {
1022 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1023 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1026 /** The database environment. */
1028 HANDLE me_fd; /**< The main data file */
1029 HANDLE me_lfd; /**< The lock file */
1030 HANDLE me_mfd; /**< just for writing the meta pages */
1031 /** Failed to update the meta page. Probably an I/O error. */
1032 #define MDB_FATAL_ERROR 0x80000000U
1033 /** Some fields are initialized. */
1034 #define MDB_ENV_ACTIVE 0x20000000U
1035 /** me_txkey is set */
1036 #define MDB_ENV_TXKEY 0x10000000U
1037 uint32_t me_flags; /**< @ref mdb_env */
1038 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1039 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1040 unsigned int me_maxreaders; /**< size of the reader table */
1041 unsigned int me_numreaders; /**< max numreaders set by this env */
1042 MDB_dbi me_numdbs; /**< number of DBs opened */
1043 MDB_dbi me_maxdbs; /**< size of the DB table */
1044 MDB_PID_T me_pid; /**< process ID of this env */
1045 char *me_path; /**< path to the DB files */
1046 char *me_map; /**< the memory map of the data file */
1047 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1048 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1049 void *me_pbuf; /**< scratch area for DUPSORT put() */
1050 MDB_txn *me_txn; /**< current write transaction */
1051 size_t me_mapsize; /**< size of the data memory map */
1052 off_t me_size; /**< current file size */
1053 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1054 MDB_dbx *me_dbxs; /**< array of static DB info */
1055 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1056 pthread_key_t me_txkey; /**< thread-key for readers */
1057 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1058 # define me_pglast me_pgstate.mf_pglast
1059 # define me_pghead me_pgstate.mf_pghead
1060 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1061 /** IDL of pages that became unused in a write txn */
1062 MDB_IDL me_free_pgs;
1063 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1064 MDB_ID2L me_dirty_list;
1065 /** Max number of freelist items that can fit in a single overflow page */
1067 /** Max size of a node on a page */
1068 unsigned int me_nodemax;
1069 #if !(MDB_MAXKEYSIZE)
1070 unsigned int me_maxkey; /**< max size of a key */
1072 int me_live_reader; /**< have liveness lock in reader table */
1074 int me_pidquery; /**< Used in OpenProcess */
1075 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1077 #elif defined(MDB_USE_POSIX_SEM)
1078 sem_t *me_rmutex; /* Shared mutexes are not supported */
1081 void *me_userctx; /**< User-settable context */
1082 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1085 /** Nested transaction */
1086 typedef struct MDB_ntxn {
1087 MDB_txn mnt_txn; /**< the transaction */
1088 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1091 /** max number of pages to commit in one writev() call */
1092 #define MDB_COMMIT_PAGES 64
1093 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1094 #undef MDB_COMMIT_PAGES
1095 #define MDB_COMMIT_PAGES IOV_MAX
1098 /** max bytes to write in one call */
1099 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1101 /** Check \b txn and \b dbi arguments to a function */
1102 #define TXN_DBI_EXIST(txn, dbi) \
1103 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1105 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1106 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1107 static int mdb_page_touch(MDB_cursor *mc);
1109 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1110 static int mdb_page_search_root(MDB_cursor *mc,
1111 MDB_val *key, int modify);
1112 #define MDB_PS_MODIFY 1
1113 #define MDB_PS_ROOTONLY 2
1114 #define MDB_PS_FIRST 4
1115 #define MDB_PS_LAST 8
1116 static int mdb_page_search(MDB_cursor *mc,
1117 MDB_val *key, int flags);
1118 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1120 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1121 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1122 pgno_t newpgno, unsigned int nflags);
1124 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1125 static int mdb_env_pick_meta(const MDB_env *env);
1126 static int mdb_env_write_meta(MDB_txn *txn);
1127 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1128 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1130 static void mdb_env_close0(MDB_env *env, int excl);
1132 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1133 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1134 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1135 static void mdb_node_del(MDB_cursor *mc, int ksize);
1136 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1137 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1138 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1139 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1140 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1142 static int mdb_rebalance(MDB_cursor *mc);
1143 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1145 static void mdb_cursor_pop(MDB_cursor *mc);
1146 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1148 static int mdb_cursor_del0(MDB_cursor *mc);
1149 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1150 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1151 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1152 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1153 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1155 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1156 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1158 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1159 static void mdb_xcursor_init0(MDB_cursor *mc);
1160 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1162 static int mdb_drop0(MDB_cursor *mc, int subs);
1163 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1166 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1170 static SECURITY_DESCRIPTOR mdb_null_sd;
1171 static SECURITY_ATTRIBUTES mdb_all_sa;
1172 static int mdb_sec_inited;
1175 /** Return the library version info. */
1177 mdb_version(int *major, int *minor, int *patch)
1179 if (major) *major = MDB_VERSION_MAJOR;
1180 if (minor) *minor = MDB_VERSION_MINOR;
1181 if (patch) *patch = MDB_VERSION_PATCH;
1182 return MDB_VERSION_STRING;
1185 /** Table of descriptions for MDB @ref errors */
1186 static char *const mdb_errstr[] = {
1187 "MDB_KEYEXIST: Key/data pair already exists",
1188 "MDB_NOTFOUND: No matching key/data pair found",
1189 "MDB_PAGE_NOTFOUND: Requested page not found",
1190 "MDB_CORRUPTED: Located page was wrong type",
1191 "MDB_PANIC: Update of meta page failed",
1192 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1193 "MDB_INVALID: File is not an MDB file",
1194 "MDB_MAP_FULL: Environment mapsize limit reached",
1195 "MDB_DBS_FULL: Environment maxdbs limit reached",
1196 "MDB_READERS_FULL: Environment maxreaders limit reached",
1197 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1198 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1199 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1200 "MDB_PAGE_FULL: Internal error - page has no more space",
1201 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1202 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1203 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1204 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1205 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1209 mdb_strerror(int err)
1213 return ("Successful return: 0");
1215 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1216 i = err - MDB_KEYEXIST;
1217 return mdb_errstr[i];
1220 return strerror(err);
1223 /** assert(3) variant in cursor context */
1224 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1225 /** assert(3) variant in transaction context */
1226 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1227 /** assert(3) variant in environment context */
1228 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1231 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1232 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1235 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1236 const char *func, const char *file, int line)
1239 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1240 file, line, expr_txt, func);
1241 if (env->me_assert_func)
1242 env->me_assert_func(env, buf);
1243 fprintf(stderr, "%s\n", buf);
1247 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1251 /** Return the page number of \b mp which may be sub-page, for debug output */
1253 mdb_dbg_pgno(MDB_page *mp)
1256 COPY_PGNO(ret, mp->mp_pgno);
1260 /** Display a key in hexadecimal and return the address of the result.
1261 * @param[in] key the key to display
1262 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1263 * @return The key in hexadecimal form.
1266 mdb_dkey(MDB_val *key, char *buf)
1269 unsigned char *c = key->mv_data;
1275 if (key->mv_size > DKBUF_MAXKEYSIZE)
1276 return "MDB_MAXKEYSIZE";
1277 /* may want to make this a dynamic check: if the key is mostly
1278 * printable characters, print it as-is instead of converting to hex.
1282 for (i=0; i<key->mv_size; i++)
1283 ptr += sprintf(ptr, "%02x", *c++);
1285 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1291 mdb_leafnode_type(MDB_node *n)
1293 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1294 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1295 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1298 /** Display all the keys in the page. */
1300 mdb_page_list(MDB_page *mp)
1302 pgno_t pgno = mdb_dbg_pgno(mp);
1303 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1305 unsigned int i, nkeys, nsize, total = 0;
1309 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1310 case P_BRANCH: type = "Branch page"; break;
1311 case P_LEAF: type = "Leaf page"; break;
1312 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1313 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1314 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1316 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1317 pgno, mp->mp_pages, state);
1320 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1321 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1324 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1328 nkeys = NUMKEYS(mp);
1329 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1331 for (i=0; i<nkeys; i++) {
1332 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1333 key.mv_size = nsize = mp->mp_pad;
1334 key.mv_data = LEAF2KEY(mp, i, nsize);
1336 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1339 node = NODEPTR(mp, i);
1340 key.mv_size = node->mn_ksize;
1341 key.mv_data = node->mn_data;
1342 nsize = NODESIZE + key.mv_size;
1343 if (IS_BRANCH(mp)) {
1344 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1348 if (F_ISSET(node->mn_flags, F_BIGDATA))
1349 nsize += sizeof(pgno_t);
1351 nsize += NODEDSZ(node);
1353 nsize += sizeof(indx_t);
1354 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1355 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1357 total = EVEN(total);
1359 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1360 IS_LEAF2(mp) ? PAGEHDRSZ : mp->mp_lower, total, SIZELEFT(mp));
1364 mdb_cursor_chk(MDB_cursor *mc)
1370 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1371 for (i=0; i<mc->mc_top; i++) {
1373 node = NODEPTR(mp, mc->mc_ki[i]);
1374 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1377 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1383 /** Count all the pages in each DB and in the freelist
1384 * and make sure it matches the actual number of pages
1386 * All named DBs must be open for a correct count.
1388 static void mdb_audit(MDB_txn *txn)
1392 MDB_ID freecount, count;
1397 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1398 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1399 freecount += *(MDB_ID *)data.mv_data;
1400 mdb_tassert(txn, rc == MDB_NOTFOUND);
1403 for (i = 0; i<txn->mt_numdbs; i++) {
1405 if (!(txn->mt_dbflags[i] & DB_VALID))
1407 mdb_cursor_init(&mc, txn, i, &mx);
1408 if (txn->mt_dbs[i].md_root == P_INVALID)
1410 count += txn->mt_dbs[i].md_branch_pages +
1411 txn->mt_dbs[i].md_leaf_pages +
1412 txn->mt_dbs[i].md_overflow_pages;
1413 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1414 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1415 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1418 mp = mc.mc_pg[mc.mc_top];
1419 for (j=0; j<NUMKEYS(mp); j++) {
1420 MDB_node *leaf = NODEPTR(mp, j);
1421 if (leaf->mn_flags & F_SUBDATA) {
1423 memcpy(&db, NODEDATA(leaf), sizeof(db));
1424 count += db.md_branch_pages + db.md_leaf_pages +
1425 db.md_overflow_pages;
1429 mdb_tassert(txn, rc == MDB_NOTFOUND);
1432 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1433 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1434 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1440 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1442 return txn->mt_dbxs[dbi].md_cmp(a, b);
1446 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1448 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1451 /** Allocate memory for a page.
1452 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1455 mdb_page_malloc(MDB_txn *txn, unsigned num)
1457 MDB_env *env = txn->mt_env;
1458 MDB_page *ret = env->me_dpages;
1459 size_t psize = env->me_psize, sz = psize, off;
1460 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1461 * For a single page alloc, we init everything after the page header.
1462 * For multi-page, we init the final page; if the caller needed that
1463 * many pages they will be filling in at least up to the last page.
1467 VGMEMP_ALLOC(env, ret, sz);
1468 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1469 env->me_dpages = ret->mp_next;
1472 psize -= off = PAGEHDRSZ;
1477 if ((ret = malloc(sz)) != NULL) {
1478 VGMEMP_ALLOC(env, ret, sz);
1479 if (!(env->me_flags & MDB_NOMEMINIT)) {
1480 memset((char *)ret + off, 0, psize);
1484 txn->mt_flags |= MDB_TXN_ERROR;
1489 /** Free a single page.
1490 * Saves single pages to a list, for future reuse.
1491 * (This is not used for multi-page overflow pages.)
1494 mdb_page_free(MDB_env *env, MDB_page *mp)
1496 mp->mp_next = env->me_dpages;
1497 VGMEMP_FREE(env, mp);
1498 env->me_dpages = mp;
1501 /** Free a dirty page */
1503 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1505 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1506 mdb_page_free(env, dp);
1508 /* large pages just get freed directly */
1509 VGMEMP_FREE(env, dp);
1514 /** Return all dirty pages to dpage list */
1516 mdb_dlist_free(MDB_txn *txn)
1518 MDB_env *env = txn->mt_env;
1519 MDB_ID2L dl = txn->mt_u.dirty_list;
1520 unsigned i, n = dl[0].mid;
1522 for (i = 1; i <= n; i++) {
1523 mdb_dpage_free(env, dl[i].mptr);
1528 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1529 * @param[in] mc A cursor handle for the current operation.
1530 * @param[in] pflags Flags of the pages to update:
1531 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1532 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1533 * @return 0 on success, non-zero on failure.
1536 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1538 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1539 MDB_txn *txn = mc->mc_txn;
1545 int rc = MDB_SUCCESS, level;
1547 /* Mark pages seen by cursors */
1548 if (mc->mc_flags & C_UNTRACK)
1549 mc = NULL; /* will find mc in mt_cursors */
1550 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1551 for (; mc; mc=mc->mc_next) {
1552 if (!(mc->mc_flags & C_INITIALIZED))
1554 for (m3 = mc;; m3 = &mx->mx_cursor) {
1556 for (j=0; j<m3->mc_snum; j++) {
1558 if ((mp->mp_flags & Mask) == pflags)
1559 mp->mp_flags ^= P_KEEP;
1561 mx = m3->mc_xcursor;
1562 /* Proceed to mx if it is at a sub-database */
1563 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1565 if (! (mp && (mp->mp_flags & P_LEAF)))
1567 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1568 if (!(leaf->mn_flags & F_SUBDATA))
1577 /* Mark dirty root pages */
1578 for (i=0; i<txn->mt_numdbs; i++) {
1579 if (txn->mt_dbflags[i] & DB_DIRTY) {
1580 pgno_t pgno = txn->mt_dbs[i].md_root;
1581 if (pgno == P_INVALID)
1583 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1585 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1586 dp->mp_flags ^= P_KEEP;
1594 static int mdb_page_flush(MDB_txn *txn, int keep);
1596 /** Spill pages from the dirty list back to disk.
1597 * This is intended to prevent running into #MDB_TXN_FULL situations,
1598 * but note that they may still occur in a few cases:
1599 * 1) our estimate of the txn size could be too small. Currently this
1600 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1601 * 2) child txns may run out of space if their parents dirtied a
1602 * lot of pages and never spilled them. TODO: we probably should do
1603 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1604 * the parent's dirty_room is below a given threshold.
1606 * Otherwise, if not using nested txns, it is expected that apps will
1607 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1608 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1609 * If the txn never references them again, they can be left alone.
1610 * If the txn only reads them, they can be used without any fuss.
1611 * If the txn writes them again, they can be dirtied immediately without
1612 * going thru all of the work of #mdb_page_touch(). Such references are
1613 * handled by #mdb_page_unspill().
1615 * Also note, we never spill DB root pages, nor pages of active cursors,
1616 * because we'll need these back again soon anyway. And in nested txns,
1617 * we can't spill a page in a child txn if it was already spilled in a
1618 * parent txn. That would alter the parent txns' data even though
1619 * the child hasn't committed yet, and we'd have no way to undo it if
1620 * the child aborted.
1622 * @param[in] m0 cursor A cursor handle identifying the transaction and
1623 * database for which we are checking space.
1624 * @param[in] key For a put operation, the key being stored.
1625 * @param[in] data For a put operation, the data being stored.
1626 * @return 0 on success, non-zero on failure.
1629 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1631 MDB_txn *txn = m0->mc_txn;
1633 MDB_ID2L dl = txn->mt_u.dirty_list;
1634 unsigned int i, j, need;
1637 if (m0->mc_flags & C_SUB)
1640 /* Estimate how much space this op will take */
1641 i = m0->mc_db->md_depth;
1642 /* Named DBs also dirty the main DB */
1643 if (m0->mc_dbi > MAIN_DBI)
1644 i += txn->mt_dbs[MAIN_DBI].md_depth;
1645 /* For puts, roughly factor in the key+data size */
1647 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1648 i += i; /* double it for good measure */
1651 if (txn->mt_dirty_room > i)
1654 if (!txn->mt_spill_pgs) {
1655 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1656 if (!txn->mt_spill_pgs)
1659 /* purge deleted slots */
1660 MDB_IDL sl = txn->mt_spill_pgs;
1661 unsigned int num = sl[0];
1663 for (i=1; i<=num; i++) {
1670 /* Preserve pages which may soon be dirtied again */
1671 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1674 /* Less aggressive spill - we originally spilled the entire dirty list,
1675 * with a few exceptions for cursor pages and DB root pages. But this
1676 * turns out to be a lot of wasted effort because in a large txn many
1677 * of those pages will need to be used again. So now we spill only 1/8th
1678 * of the dirty pages. Testing revealed this to be a good tradeoff,
1679 * better than 1/2, 1/4, or 1/10.
1681 if (need < MDB_IDL_UM_MAX / 8)
1682 need = MDB_IDL_UM_MAX / 8;
1684 /* Save the page IDs of all the pages we're flushing */
1685 /* flush from the tail forward, this saves a lot of shifting later on. */
1686 for (i=dl[0].mid; i && need; i--) {
1687 MDB_ID pn = dl[i].mid << 1;
1689 if (dp->mp_flags & P_KEEP)
1691 /* Can't spill twice, make sure it's not already in a parent's
1694 if (txn->mt_parent) {
1696 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1697 if (tx2->mt_spill_pgs) {
1698 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1699 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1700 dp->mp_flags |= P_KEEP;
1708 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1712 mdb_midl_sort(txn->mt_spill_pgs);
1714 /* Flush the spilled part of dirty list */
1715 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1718 /* Reset any dirty pages we kept that page_flush didn't see */
1719 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1722 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1726 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1728 mdb_find_oldest(MDB_txn *txn)
1731 txnid_t mr, oldest = txn->mt_txnid - 1;
1732 if (txn->mt_env->me_txns) {
1733 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1734 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1745 /** Add a page to the txn's dirty list */
1747 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1750 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1752 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1753 insert = mdb_mid2l_append;
1755 insert = mdb_mid2l_insert;
1757 mid.mid = mp->mp_pgno;
1759 rc = insert(txn->mt_u.dirty_list, &mid);
1760 mdb_tassert(txn, rc == 0);
1761 txn->mt_dirty_room--;
1764 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1765 * me_pghead and mt_next_pgno.
1767 * If there are free pages available from older transactions, they
1768 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1769 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1770 * and move me_pglast to say which records were consumed. Only this
1771 * function can create me_pghead and move me_pglast/mt_next_pgno.
1772 * @param[in] mc cursor A cursor handle identifying the transaction and
1773 * database for which we are allocating.
1774 * @param[in] num the number of pages to allocate.
1775 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1776 * will always be satisfied by a single contiguous chunk of memory.
1777 * @return 0 on success, non-zero on failure.
1780 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1782 /* use PARANOID for now, default infinite search slows down too much
1783 * when the freelist is large
1785 #define MDB_PARANOID
1786 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1787 /* Get at most <Max_retries> more freeDB records once me_pghead
1788 * has enough pages. If not enough, use new pages from the map.
1789 * If <Paranoid> and mc is updating the freeDB, only get new
1790 * records if me_pghead is empty. Then the freelist cannot play
1791 * catch-up with itself by growing while trying to save it.
1793 enum { Paranoid = 1, Max_retries = 500 };
1795 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1797 int rc, retry = num;
1798 MDB_txn *txn = mc->mc_txn;
1799 MDB_env *env = txn->mt_env;
1800 pgno_t pgno, *mop = env->me_pghead;
1801 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1803 txnid_t oldest = 0, last;
1809 /* If our dirty list is already full, we can't do anything */
1810 if (txn->mt_dirty_room == 0) {
1815 for (op = MDB_FIRST;; op = MDB_NEXT) {
1818 pgno_t *idl, old_id, new_id;
1820 /* Seek a big enough contiguous page range. Prefer
1821 * pages at the tail, just truncating the list.
1827 if (mop[i-n2] == pgno+n2)
1830 if (Max_retries < INT_MAX && --retry < 0)
1834 if (op == MDB_FIRST) { /* 1st iteration */
1835 /* Prepare to fetch more and coalesce */
1836 oldest = mdb_find_oldest(txn);
1837 last = env->me_pglast;
1838 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1841 key.mv_data = &last; /* will look up last+1 */
1842 key.mv_size = sizeof(last);
1844 if (Paranoid && mc->mc_dbi == FREE_DBI)
1847 if (Paranoid && retry < 0 && mop_len)
1851 /* Do not fetch more if the record will be too recent */
1854 rc = mdb_cursor_get(&m2, &key, NULL, op);
1856 if (rc == MDB_NOTFOUND)
1860 last = *(txnid_t*)key.mv_data;
1863 np = m2.mc_pg[m2.mc_top];
1864 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1865 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1868 idl = (MDB_ID *) data.mv_data;
1871 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1876 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1878 mop = env->me_pghead;
1880 env->me_pglast = last;
1882 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1883 last, txn->mt_dbs[FREE_DBI].md_root, i));
1885 DPRINTF(("IDL %"Z"u", idl[k]));
1887 /* Merge in descending sorted order */
1890 mop[0] = (pgno_t)-1;
1894 for (; old_id < new_id; old_id = mop[--j])
1901 /* Use new pages from the map when nothing suitable in the freeDB */
1903 pgno = txn->mt_next_pgno;
1904 if (pgno + num >= env->me_maxpg) {
1905 DPUTS("DB size maxed out");
1911 if (env->me_flags & MDB_WRITEMAP) {
1912 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1914 if (!(np = mdb_page_malloc(txn, num))) {
1920 mop[0] = mop_len -= num;
1921 /* Move any stragglers down */
1922 for (j = i-num; j < mop_len; )
1923 mop[++j] = mop[++i];
1925 txn->mt_next_pgno = pgno + num;
1928 mdb_page_dirty(txn, np);
1934 txn->mt_flags |= MDB_TXN_ERROR;
1938 /** Copy the used portions of a non-overflow page.
1939 * @param[in] dst page to copy into
1940 * @param[in] src page to copy from
1941 * @param[in] psize size of a page
1944 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1946 enum { Align = sizeof(pgno_t) };
1947 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1949 /* If page isn't full, just copy the used portion. Adjust
1950 * alignment so memcpy may copy words instead of bytes.
1952 if ((unused &= -Align) && !IS_LEAF2(src)) {
1954 memcpy(dst, src, (lower + (Align-1)) & -Align);
1955 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1958 memcpy(dst, src, psize - unused);
1962 /** Pull a page off the txn's spill list, if present.
1963 * If a page being referenced was spilled to disk in this txn, bring
1964 * it back and make it dirty/writable again.
1965 * @param[in] txn the transaction handle.
1966 * @param[in] mp the page being referenced. It must not be dirty.
1967 * @param[out] ret the writable page, if any. ret is unchanged if
1968 * mp wasn't spilled.
1971 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1973 MDB_env *env = txn->mt_env;
1976 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1978 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1979 if (!tx2->mt_spill_pgs)
1981 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1982 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1985 if (txn->mt_dirty_room == 0)
1986 return MDB_TXN_FULL;
1987 if (IS_OVERFLOW(mp))
1991 if (env->me_flags & MDB_WRITEMAP) {
1994 np = mdb_page_malloc(txn, num);
1998 memcpy(np, mp, num * env->me_psize);
2000 mdb_page_copy(np, mp, env->me_psize);
2003 /* If in current txn, this page is no longer spilled.
2004 * If it happens to be the last page, truncate the spill list.
2005 * Otherwise mark it as deleted by setting the LSB.
2007 if (x == txn->mt_spill_pgs[0])
2008 txn->mt_spill_pgs[0]--;
2010 txn->mt_spill_pgs[x] |= 1;
2011 } /* otherwise, if belonging to a parent txn, the
2012 * page remains spilled until child commits
2015 mdb_page_dirty(txn, np);
2016 np->mp_flags |= P_DIRTY;
2024 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2025 * @param[in] mc cursor pointing to the page to be touched
2026 * @return 0 on success, non-zero on failure.
2029 mdb_page_touch(MDB_cursor *mc)
2031 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2032 MDB_txn *txn = mc->mc_txn;
2033 MDB_cursor *m2, *m3;
2037 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2038 if (txn->mt_flags & MDB_TXN_SPILLS) {
2040 rc = mdb_page_unspill(txn, mp, &np);
2046 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2047 (rc = mdb_page_alloc(mc, 1, &np)))
2050 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2051 mp->mp_pgno, pgno));
2052 mdb_cassert(mc, mp->mp_pgno != pgno);
2053 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2054 /* Update the parent page, if any, to point to the new page */
2056 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2057 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2058 SETPGNO(node, pgno);
2060 mc->mc_db->md_root = pgno;
2062 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2063 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2065 /* If txn has a parent, make sure the page is in our
2069 unsigned x = mdb_mid2l_search(dl, pgno);
2070 if (x <= dl[0].mid && dl[x].mid == pgno) {
2071 if (mp != dl[x].mptr) { /* bad cursor? */
2072 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2073 txn->mt_flags |= MDB_TXN_ERROR;
2074 return MDB_CORRUPTED;
2079 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2081 np = mdb_page_malloc(txn, 1);
2086 rc = mdb_mid2l_insert(dl, &mid);
2087 mdb_cassert(mc, rc == 0);
2092 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2094 np->mp_flags |= P_DIRTY;
2097 /* Adjust cursors pointing to mp */
2098 mc->mc_pg[mc->mc_top] = np;
2099 m2 = txn->mt_cursors[mc->mc_dbi];
2100 if (mc->mc_flags & C_SUB) {
2101 for (; m2; m2=m2->mc_next) {
2102 m3 = &m2->mc_xcursor->mx_cursor;
2103 if (m3->mc_snum < mc->mc_snum) continue;
2104 if (m3->mc_pg[mc->mc_top] == mp)
2105 m3->mc_pg[mc->mc_top] = np;
2108 for (; m2; m2=m2->mc_next) {
2109 if (m2->mc_snum < mc->mc_snum) continue;
2110 if (m2->mc_pg[mc->mc_top] == mp) {
2111 m2->mc_pg[mc->mc_top] = np;
2112 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2114 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2116 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2117 if (!(leaf->mn_flags & F_SUBDATA))
2118 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2126 txn->mt_flags |= MDB_TXN_ERROR;
2131 mdb_env_sync(MDB_env *env, int force)
2134 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2135 if (env->me_flags & MDB_WRITEMAP) {
2136 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2137 ? MS_ASYNC : MS_SYNC;
2138 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2141 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2145 if (MDB_FDATASYNC(env->me_fd))
2152 /** Back up parent txn's cursors, then grab the originals for tracking */
2154 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2156 MDB_cursor *mc, *bk;
2161 for (i = src->mt_numdbs; --i >= 0; ) {
2162 if ((mc = src->mt_cursors[i]) != NULL) {
2163 size = sizeof(MDB_cursor);
2165 size += sizeof(MDB_xcursor);
2166 for (; mc; mc = bk->mc_next) {
2172 mc->mc_db = &dst->mt_dbs[i];
2173 /* Kill pointers into src - and dst to reduce abuse: The
2174 * user may not use mc until dst ends. Otherwise we'd...
2176 mc->mc_txn = NULL; /* ...set this to dst */
2177 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2178 if ((mx = mc->mc_xcursor) != NULL) {
2179 *(MDB_xcursor *)(bk+1) = *mx;
2180 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2182 mc->mc_next = dst->mt_cursors[i];
2183 dst->mt_cursors[i] = mc;
2190 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2191 * @param[in] txn the transaction handle.
2192 * @param[in] merge true to keep changes to parent cursors, false to revert.
2193 * @return 0 on success, non-zero on failure.
2196 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2198 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2202 for (i = txn->mt_numdbs; --i >= 0; ) {
2203 for (mc = cursors[i]; mc; mc = next) {
2205 if ((bk = mc->mc_backup) != NULL) {
2207 /* Commit changes to parent txn */
2208 mc->mc_next = bk->mc_next;
2209 mc->mc_backup = bk->mc_backup;
2210 mc->mc_txn = bk->mc_txn;
2211 mc->mc_db = bk->mc_db;
2212 mc->mc_dbflag = bk->mc_dbflag;
2213 if ((mx = mc->mc_xcursor) != NULL)
2214 mx->mx_cursor.mc_txn = bk->mc_txn;
2216 /* Abort nested txn */
2218 if ((mx = mc->mc_xcursor) != NULL)
2219 *mx = *(MDB_xcursor *)(bk+1);
2223 /* Only malloced cursors are permanently tracked. */
2231 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2234 mdb_txn_reset0(MDB_txn *txn, const char *act);
2236 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2242 Pidset = F_SETLK, Pidcheck = F_GETLK
2246 /** Set or check a pid lock. Set returns 0 on success.
2247 * Check returns 0 if the process is certainly dead, nonzero if it may
2248 * be alive (the lock exists or an error happened so we do not know).
2250 * On Windows Pidset is a no-op, we merely check for the existence
2251 * of the process with the given pid. On POSIX we use a single byte
2252 * lock on the lockfile, set at an offset equal to the pid.
2255 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2257 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2260 if (op == Pidcheck) {
2261 h = OpenProcess(env->me_pidquery, FALSE, pid);
2262 /* No documented "no such process" code, but other program use this: */
2264 return ErrCode() != ERROR_INVALID_PARAMETER;
2265 /* A process exists until all handles to it close. Has it exited? */
2266 ret = WaitForSingleObject(h, 0) != 0;
2273 struct flock lock_info;
2274 memset(&lock_info, 0, sizeof(lock_info));
2275 lock_info.l_type = F_WRLCK;
2276 lock_info.l_whence = SEEK_SET;
2277 lock_info.l_start = pid;
2278 lock_info.l_len = 1;
2279 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2280 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2282 } else if ((rc = ErrCode()) == EINTR) {
2290 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2291 * @param[in] txn the transaction handle to initialize
2292 * @return 0 on success, non-zero on failure.
2295 mdb_txn_renew0(MDB_txn *txn)
2297 MDB_env *env = txn->mt_env;
2298 MDB_txninfo *ti = env->me_txns;
2302 int rc, new_notls = 0;
2305 txn->mt_numdbs = env->me_numdbs;
2306 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2308 if (txn->mt_flags & MDB_TXN_RDONLY) {
2310 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2311 txn->mt_txnid = meta->mm_txnid;
2312 txn->mt_u.reader = NULL;
2314 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2315 pthread_getspecific(env->me_txkey);
2317 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2318 return MDB_BAD_RSLOT;
2320 MDB_PID_T pid = env->me_pid;
2321 pthread_t tid = pthread_self();
2323 if (!env->me_live_reader) {
2324 rc = mdb_reader_pid(env, Pidset, pid);
2327 env->me_live_reader = 1;
2331 nr = ti->mti_numreaders;
2332 for (i=0; i<nr; i++)
2333 if (ti->mti_readers[i].mr_pid == 0)
2335 if (i == env->me_maxreaders) {
2336 UNLOCK_MUTEX_R(env);
2337 return MDB_READERS_FULL;
2339 ti->mti_readers[i].mr_pid = pid;
2340 ti->mti_readers[i].mr_tid = tid;
2342 ti->mti_numreaders = ++nr;
2343 /* Save numreaders for un-mutexed mdb_env_close() */
2344 env->me_numreaders = nr;
2345 UNLOCK_MUTEX_R(env);
2347 r = &ti->mti_readers[i];
2348 new_notls = (env->me_flags & MDB_NOTLS);
2349 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2354 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2355 txn->mt_u.reader = r;
2356 meta = env->me_metas[txn->mt_txnid & 1];
2362 txn->mt_txnid = ti->mti_txnid;
2363 meta = env->me_metas[txn->mt_txnid & 1];
2365 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2366 txn->mt_txnid = meta->mm_txnid;
2370 if (txn->mt_txnid == mdb_debug_start)
2373 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2374 txn->mt_u.dirty_list = env->me_dirty_list;
2375 txn->mt_u.dirty_list[0].mid = 0;
2376 txn->mt_free_pgs = env->me_free_pgs;
2377 txn->mt_free_pgs[0] = 0;
2378 txn->mt_spill_pgs = NULL;
2382 /* Copy the DB info and flags */
2383 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2385 /* Moved to here to avoid a data race in read TXNs */
2386 txn->mt_next_pgno = meta->mm_last_pg+1;
2388 for (i=2; i<txn->mt_numdbs; i++) {
2389 x = env->me_dbflags[i];
2390 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2391 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2393 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2395 if (env->me_maxpg < txn->mt_next_pgno) {
2396 mdb_txn_reset0(txn, "renew0-mapfail");
2398 txn->mt_u.reader->mr_pid = 0;
2399 txn->mt_u.reader = NULL;
2401 return MDB_MAP_RESIZED;
2408 mdb_txn_renew(MDB_txn *txn)
2412 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2415 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2416 DPUTS("environment had fatal error, must shutdown!");
2420 rc = mdb_txn_renew0(txn);
2421 if (rc == MDB_SUCCESS) {
2422 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2423 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2424 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2430 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2434 int rc, size, tsize = sizeof(MDB_txn);
2436 if (env->me_flags & MDB_FATAL_ERROR) {
2437 DPUTS("environment had fatal error, must shutdown!");
2440 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2443 /* Nested transactions: Max 1 child, write txns only, no writemap */
2444 if (parent->mt_child ||
2445 (flags & MDB_RDONLY) ||
2446 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2447 (env->me_flags & MDB_WRITEMAP))
2449 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2451 tsize = sizeof(MDB_ntxn);
2453 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2454 if (!(flags & MDB_RDONLY))
2455 size += env->me_maxdbs * sizeof(MDB_cursor *);
2457 if ((txn = calloc(1, size)) == NULL) {
2458 DPRINTF(("calloc: %s", strerror(ErrCode())));
2461 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2462 if (flags & MDB_RDONLY) {
2463 txn->mt_flags |= MDB_TXN_RDONLY;
2464 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2466 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2467 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2473 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2474 if (!txn->mt_u.dirty_list ||
2475 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2477 free(txn->mt_u.dirty_list);
2481 txn->mt_txnid = parent->mt_txnid;
2482 txn->mt_dirty_room = parent->mt_dirty_room;
2483 txn->mt_u.dirty_list[0].mid = 0;
2484 txn->mt_spill_pgs = NULL;
2485 txn->mt_next_pgno = parent->mt_next_pgno;
2486 parent->mt_child = txn;
2487 txn->mt_parent = parent;
2488 txn->mt_numdbs = parent->mt_numdbs;
2489 txn->mt_flags = parent->mt_flags;
2490 txn->mt_dbxs = parent->mt_dbxs;
2491 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2492 /* Copy parent's mt_dbflags, but clear DB_NEW */
2493 for (i=0; i<txn->mt_numdbs; i++)
2494 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2496 ntxn = (MDB_ntxn *)txn;
2497 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2498 if (env->me_pghead) {
2499 size = MDB_IDL_SIZEOF(env->me_pghead);
2500 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2502 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2507 rc = mdb_cursor_shadow(parent, txn);
2509 mdb_txn_reset0(txn, "beginchild-fail");
2511 rc = mdb_txn_renew0(txn);
2517 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2518 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2519 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2526 mdb_txn_env(MDB_txn *txn)
2528 if(!txn) return NULL;
2532 /** Export or close DBI handles opened in this txn. */
2534 mdb_dbis_update(MDB_txn *txn, int keep)
2537 MDB_dbi n = txn->mt_numdbs;
2538 MDB_env *env = txn->mt_env;
2539 unsigned char *tdbflags = txn->mt_dbflags;
2541 for (i = n; --i >= 2;) {
2542 if (tdbflags[i] & DB_NEW) {
2544 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2546 char *ptr = env->me_dbxs[i].md_name.mv_data;
2547 env->me_dbxs[i].md_name.mv_data = NULL;
2548 env->me_dbxs[i].md_name.mv_size = 0;
2549 env->me_dbflags[i] = 0;
2554 if (keep && env->me_numdbs < n)
2558 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2559 * May be called twice for readonly txns: First reset it, then abort.
2560 * @param[in] txn the transaction handle to reset
2561 * @param[in] act why the transaction is being reset
2564 mdb_txn_reset0(MDB_txn *txn, const char *act)
2566 MDB_env *env = txn->mt_env;
2568 /* Close any DBI handles opened in this txn */
2569 mdb_dbis_update(txn, 0);
2571 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2572 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2573 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2575 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2576 if (txn->mt_u.reader) {
2577 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2578 if (!(env->me_flags & MDB_NOTLS))
2579 txn->mt_u.reader = NULL; /* txn does not own reader */
2581 txn->mt_numdbs = 0; /* close nothing if called again */
2582 txn->mt_dbxs = NULL; /* mark txn as reset */
2584 mdb_cursors_close(txn, 0);
2586 if (!(env->me_flags & MDB_WRITEMAP)) {
2587 mdb_dlist_free(txn);
2589 mdb_midl_free(env->me_pghead);
2591 if (txn->mt_parent) {
2592 txn->mt_parent->mt_child = NULL;
2593 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2594 mdb_midl_free(txn->mt_free_pgs);
2595 mdb_midl_free(txn->mt_spill_pgs);
2596 free(txn->mt_u.dirty_list);
2600 if (mdb_midl_shrink(&txn->mt_free_pgs))
2601 env->me_free_pgs = txn->mt_free_pgs;
2602 env->me_pghead = NULL;
2606 /* The writer mutex was locked in mdb_txn_begin. */
2608 UNLOCK_MUTEX_W(env);
2613 mdb_txn_reset(MDB_txn *txn)
2618 /* This call is only valid for read-only txns */
2619 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2622 mdb_txn_reset0(txn, "reset");
2626 mdb_txn_abort(MDB_txn *txn)
2632 mdb_txn_abort(txn->mt_child);
2634 mdb_txn_reset0(txn, "abort");
2635 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2636 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2637 txn->mt_u.reader->mr_pid = 0;
2642 /** Save the freelist as of this transaction to the freeDB.
2643 * This changes the freelist. Keep trying until it stabilizes.
2646 mdb_freelist_save(MDB_txn *txn)
2648 /* env->me_pghead[] can grow and shrink during this call.
2649 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2650 * Page numbers cannot disappear from txn->mt_free_pgs[].
2653 MDB_env *env = txn->mt_env;
2654 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2655 txnid_t pglast = 0, head_id = 0;
2656 pgno_t freecnt = 0, *free_pgs, *mop;
2657 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2659 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2661 if (env->me_pghead) {
2662 /* Make sure first page of freeDB is touched and on freelist */
2663 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2664 if (rc && rc != MDB_NOTFOUND)
2668 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2669 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2670 ? SSIZE_MAX : maxfree_1pg;
2673 /* Come back here after each Put() in case freelist changed */
2678 /* If using records from freeDB which we have not yet
2679 * deleted, delete them and any we reserved for me_pghead.
2681 while (pglast < env->me_pglast) {
2682 rc = mdb_cursor_first(&mc, &key, NULL);
2685 pglast = head_id = *(txnid_t *)key.mv_data;
2686 total_room = head_room = 0;
2687 mdb_tassert(txn, pglast <= env->me_pglast);
2688 rc = mdb_cursor_del(&mc, 0);
2693 /* Save the IDL of pages freed by this txn, to a single record */
2694 if (freecnt < txn->mt_free_pgs[0]) {
2696 /* Make sure last page of freeDB is touched and on freelist */
2697 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2698 if (rc && rc != MDB_NOTFOUND)
2701 free_pgs = txn->mt_free_pgs;
2702 /* Write to last page of freeDB */
2703 key.mv_size = sizeof(txn->mt_txnid);
2704 key.mv_data = &txn->mt_txnid;
2706 freecnt = free_pgs[0];
2707 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2708 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2711 /* Retry if mt_free_pgs[] grew during the Put() */
2712 free_pgs = txn->mt_free_pgs;
2713 } while (freecnt < free_pgs[0]);
2714 mdb_midl_sort(free_pgs);
2715 memcpy(data.mv_data, free_pgs, data.mv_size);
2718 unsigned int i = free_pgs[0];
2719 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2720 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2722 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2728 mop = env->me_pghead;
2729 mop_len = mop ? mop[0] : 0;
2731 /* Reserve records for me_pghead[]. Split it if multi-page,
2732 * to avoid searching freeDB for a page range. Use keys in
2733 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2735 if (total_room >= mop_len) {
2736 if (total_room == mop_len || --more < 0)
2738 } else if (head_room >= maxfree_1pg && head_id > 1) {
2739 /* Keep current record (overflow page), add a new one */
2743 /* (Re)write {key = head_id, IDL length = head_room} */
2744 total_room -= head_room;
2745 head_room = mop_len - total_room;
2746 if (head_room > maxfree_1pg && head_id > 1) {
2747 /* Overflow multi-page for part of me_pghead */
2748 head_room /= head_id; /* amortize page sizes */
2749 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2750 } else if (head_room < 0) {
2751 /* Rare case, not bothering to delete this record */
2754 key.mv_size = sizeof(head_id);
2755 key.mv_data = &head_id;
2756 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2757 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2760 /* IDL is initially empty, zero out at least the length */
2761 pgs = (pgno_t *)data.mv_data;
2762 j = head_room > clean_limit ? head_room : 0;
2766 total_room += head_room;
2769 /* Fill in the reserved me_pghead records */
2775 rc = mdb_cursor_first(&mc, &key, &data);
2776 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2777 txnid_t id = *(txnid_t *)key.mv_data;
2778 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2781 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2783 if (len > mop_len) {
2785 data.mv_size = (len + 1) * sizeof(MDB_ID);
2787 data.mv_data = mop -= len;
2790 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2792 if (rc || !(mop_len -= len))
2799 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2800 * @param[in] txn the transaction that's being committed
2801 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2802 * @return 0 on success, non-zero on failure.
2805 mdb_page_flush(MDB_txn *txn, int keep)
2807 MDB_env *env = txn->mt_env;
2808 MDB_ID2L dl = txn->mt_u.dirty_list;
2809 unsigned psize = env->me_psize, j;
2810 int i, pagecount = dl[0].mid, rc;
2811 size_t size = 0, pos = 0;
2813 MDB_page *dp = NULL;
2817 struct iovec iov[MDB_COMMIT_PAGES];
2818 ssize_t wpos = 0, wsize = 0, wres;
2819 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2825 if (env->me_flags & MDB_WRITEMAP) {
2826 /* Clear dirty flags */
2827 while (++i <= pagecount) {
2829 /* Don't flush this page yet */
2830 if (dp->mp_flags & P_KEEP) {
2831 dp->mp_flags ^= P_KEEP;
2835 dp->mp_flags &= ~P_DIRTY;
2840 /* Write the pages */
2842 if (++i <= pagecount) {
2844 /* Don't flush this page yet */
2845 if (dp->mp_flags & P_KEEP) {
2846 dp->mp_flags ^= P_KEEP;
2851 /* clear dirty flag */
2852 dp->mp_flags &= ~P_DIRTY;
2855 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2860 /* Windows actually supports scatter/gather I/O, but only on
2861 * unbuffered file handles. Since we're relying on the OS page
2862 * cache for all our data, that's self-defeating. So we just
2863 * write pages one at a time. We use the ov structure to set
2864 * the write offset, to at least save the overhead of a Seek
2867 DPRINTF(("committing page %"Z"u", pgno));
2868 memset(&ov, 0, sizeof(ov));
2869 ov.Offset = pos & 0xffffffff;
2870 ov.OffsetHigh = pos >> 16 >> 16;
2871 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2873 DPRINTF(("WriteFile: %d", rc));
2877 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2878 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2880 /* Write previous page(s) */
2881 #ifdef MDB_USE_PWRITEV
2882 wres = pwritev(env->me_fd, iov, n, wpos);
2885 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2887 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2889 DPRINTF(("lseek: %s", strerror(rc)));
2892 wres = writev(env->me_fd, iov, n);
2895 if (wres != wsize) {
2898 DPRINTF(("Write error: %s", strerror(rc)));
2900 rc = EIO; /* TODO: Use which error code? */
2901 DPUTS("short write, filesystem full?");
2912 DPRINTF(("committing page %"Z"u", pgno));
2913 next_pos = pos + size;
2914 iov[n].iov_len = size;
2915 iov[n].iov_base = (char *)dp;
2921 for (i = keep; ++i <= pagecount; ) {
2923 /* This is a page we skipped above */
2926 dl[j].mid = dp->mp_pgno;
2929 mdb_dpage_free(env, dp);
2934 txn->mt_dirty_room += i - j;
2940 mdb_txn_commit(MDB_txn *txn)
2946 if (txn == NULL || txn->mt_env == NULL)
2949 if (txn->mt_child) {
2950 rc = mdb_txn_commit(txn->mt_child);
2951 txn->mt_child = NULL;
2958 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2959 mdb_dbis_update(txn, 1);
2960 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2965 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2966 DPUTS("error flag is set, can't commit");
2968 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2973 if (txn->mt_parent) {
2974 MDB_txn *parent = txn->mt_parent;
2977 unsigned x, y, len, ps_len;
2979 /* Append our free list to parent's */
2980 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2983 mdb_midl_free(txn->mt_free_pgs);
2984 /* Failures after this must either undo the changes
2985 * to the parent or set MDB_TXN_ERROR in the parent.
2988 parent->mt_next_pgno = txn->mt_next_pgno;
2989 parent->mt_flags = txn->mt_flags;
2991 /* Merge our cursors into parent's and close them */
2992 mdb_cursors_close(txn, 1);
2994 /* Update parent's DB table. */
2995 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2996 parent->mt_numdbs = txn->mt_numdbs;
2997 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2998 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2999 for (i=2; i<txn->mt_numdbs; i++) {
3000 /* preserve parent's DB_NEW status */
3001 x = parent->mt_dbflags[i] & DB_NEW;
3002 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3005 dst = parent->mt_u.dirty_list;
3006 src = txn->mt_u.dirty_list;
3007 /* Remove anything in our dirty list from parent's spill list */
3008 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3010 pspill[0] = (pgno_t)-1;
3011 /* Mark our dirty pages as deleted in parent spill list */
3012 for (i=0, len=src[0].mid; ++i <= len; ) {
3013 MDB_ID pn = src[i].mid << 1;
3014 while (pn > pspill[x])
3016 if (pn == pspill[x]) {
3021 /* Squash deleted pagenums if we deleted any */
3022 for (x=y; ++x <= ps_len; )
3023 if (!(pspill[x] & 1))
3024 pspill[++y] = pspill[x];
3028 /* Find len = length of merging our dirty list with parent's */
3030 dst[0].mid = 0; /* simplify loops */
3031 if (parent->mt_parent) {
3032 len = x + src[0].mid;
3033 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3034 for (i = x; y && i; y--) {
3035 pgno_t yp = src[y].mid;
3036 while (yp < dst[i].mid)
3038 if (yp == dst[i].mid) {
3043 } else { /* Simplify the above for single-ancestor case */
3044 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3046 /* Merge our dirty list with parent's */
3048 for (i = len; y; dst[i--] = src[y--]) {
3049 pgno_t yp = src[y].mid;
3050 while (yp < dst[x].mid)
3051 dst[i--] = dst[x--];
3052 if (yp == dst[x].mid)
3053 free(dst[x--].mptr);
3055 mdb_tassert(txn, i == x);
3057 free(txn->mt_u.dirty_list);
3058 parent->mt_dirty_room = txn->mt_dirty_room;
3059 if (txn->mt_spill_pgs) {
3060 if (parent->mt_spill_pgs) {
3061 /* TODO: Prevent failure here, so parent does not fail */
3062 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3064 parent->mt_flags |= MDB_TXN_ERROR;
3065 mdb_midl_free(txn->mt_spill_pgs);
3066 mdb_midl_sort(parent->mt_spill_pgs);
3068 parent->mt_spill_pgs = txn->mt_spill_pgs;
3072 parent->mt_child = NULL;
3073 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3078 if (txn != env->me_txn) {
3079 DPUTS("attempt to commit unknown transaction");
3084 mdb_cursors_close(txn, 0);
3086 if (!txn->mt_u.dirty_list[0].mid &&
3087 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3090 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3091 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3093 /* Update DB root pointers */
3094 if (txn->mt_numdbs > 2) {
3098 data.mv_size = sizeof(MDB_db);
3100 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3101 for (i = 2; i < txn->mt_numdbs; i++) {
3102 if (txn->mt_dbflags[i] & DB_DIRTY) {
3103 data.mv_data = &txn->mt_dbs[i];
3104 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3111 rc = mdb_freelist_save(txn);
3115 mdb_midl_free(env->me_pghead);
3116 env->me_pghead = NULL;
3117 if (mdb_midl_shrink(&txn->mt_free_pgs))
3118 env->me_free_pgs = txn->mt_free_pgs;
3124 if ((rc = mdb_page_flush(txn, 0)) ||
3125 (rc = mdb_env_sync(env, 0)) ||
3126 (rc = mdb_env_write_meta(txn)))
3132 mdb_dbis_update(txn, 1);
3135 UNLOCK_MUTEX_W(env);
3145 /** Read the environment parameters of a DB environment before
3146 * mapping it into memory.
3147 * @param[in] env the environment handle
3148 * @param[out] meta address of where to store the meta information
3149 * @return 0 on success, non-zero on failure.
3152 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3158 enum { Size = sizeof(pbuf) };
3160 /* We don't know the page size yet, so use a minimum value.
3161 * Read both meta pages so we can use the latest one.
3164 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3168 memset(&ov, 0, sizeof(ov));
3170 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3171 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3174 rc = pread(env->me_fd, &pbuf, Size, off);
3177 if (rc == 0 && off == 0)
3179 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3180 DPRINTF(("read: %s", mdb_strerror(rc)));
3184 p = (MDB_page *)&pbuf;
3186 if (!F_ISSET(p->mp_flags, P_META)) {
3187 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3192 if (m->mm_magic != MDB_MAGIC) {
3193 DPUTS("meta has invalid magic");
3197 if (m->mm_version != MDB_DATA_VERSION) {
3198 DPRINTF(("database is version %u, expected version %u",
3199 m->mm_version, MDB_DATA_VERSION));
3200 return MDB_VERSION_MISMATCH;
3203 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3209 /** Write the environment parameters of a freshly created DB environment.
3210 * @param[in] env the environment handle
3211 * @param[out] meta address of where to store the meta information
3212 * @return 0 on success, non-zero on failure.
3215 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3223 memset(&ov, 0, sizeof(ov));
3224 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3226 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3229 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3230 len = pwrite(fd, ptr, size, pos); \
3231 rc = (len >= 0); } while(0)
3234 DPUTS("writing new meta page");
3236 psize = env->me_psize;
3238 meta->mm_magic = MDB_MAGIC;
3239 meta->mm_version = MDB_DATA_VERSION;
3240 meta->mm_mapsize = env->me_mapsize;
3241 meta->mm_psize = psize;
3242 meta->mm_last_pg = 1;
3243 meta->mm_flags = env->me_flags & 0xffff;
3244 meta->mm_flags |= MDB_INTEGERKEY;
3245 meta->mm_dbs[0].md_root = P_INVALID;
3246 meta->mm_dbs[1].md_root = P_INVALID;
3248 p = calloc(2, psize);
3250 p->mp_flags = P_META;
3251 *(MDB_meta *)METADATA(p) = *meta;
3253 q = (MDB_page *)((char *)p + psize);
3255 q->mp_flags = P_META;
3256 *(MDB_meta *)METADATA(q) = *meta;
3258 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3261 else if ((unsigned) len == psize * 2)
3269 /** Update the environment info to commit a transaction.
3270 * @param[in] txn the transaction that's being committed
3271 * @return 0 on success, non-zero on failure.
3274 mdb_env_write_meta(MDB_txn *txn)
3277 MDB_meta meta, metab, *mp;
3279 int rc, len, toggle;
3288 toggle = txn->mt_txnid & 1;
3289 DPRINTF(("writing meta page %d for root page %"Z"u",
3290 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3293 mp = env->me_metas[toggle];
3295 if (env->me_flags & MDB_WRITEMAP) {
3296 /* Persist any increases of mapsize config */
3297 if (env->me_mapsize > mp->mm_mapsize)
3298 mp->mm_mapsize = env->me_mapsize;
3299 mp->mm_dbs[0] = txn->mt_dbs[0];
3300 mp->mm_dbs[1] = txn->mt_dbs[1];
3301 mp->mm_last_pg = txn->mt_next_pgno - 1;
3302 mp->mm_txnid = txn->mt_txnid;
3303 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3304 unsigned meta_size = env->me_psize;
3305 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3308 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3309 if (meta_size < env->me_os_psize)
3310 meta_size += meta_size;
3315 if (MDB_MSYNC(ptr, meta_size, rc)) {
3322 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3323 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3325 ptr = (char *)&meta;
3326 if (env->me_mapsize > mp->mm_mapsize) {
3327 /* Persist any increases of mapsize config */
3328 meta.mm_mapsize = env->me_mapsize;
3329 off = offsetof(MDB_meta, mm_mapsize);
3331 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3333 len = sizeof(MDB_meta) - off;
3336 meta.mm_dbs[0] = txn->mt_dbs[0];
3337 meta.mm_dbs[1] = txn->mt_dbs[1];
3338 meta.mm_last_pg = txn->mt_next_pgno - 1;
3339 meta.mm_txnid = txn->mt_txnid;
3342 off += env->me_psize;
3345 /* Write to the SYNC fd */
3346 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3347 env->me_fd : env->me_mfd;
3350 memset(&ov, 0, sizeof(ov));
3352 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3356 rc = pwrite(mfd, ptr, len, off);
3359 rc = rc < 0 ? ErrCode() : EIO;
3360 DPUTS("write failed, disk error?");
3361 /* On a failure, the pagecache still contains the new data.
3362 * Write some old data back, to prevent it from being used.
3363 * Use the non-SYNC fd; we know it will fail anyway.
3365 meta.mm_last_pg = metab.mm_last_pg;
3366 meta.mm_txnid = metab.mm_txnid;
3368 memset(&ov, 0, sizeof(ov));
3370 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3372 r2 = pwrite(env->me_fd, ptr, len, off);
3373 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3376 env->me_flags |= MDB_FATAL_ERROR;
3380 /* Memory ordering issues are irrelevant; since the entire writer
3381 * is wrapped by wmutex, all of these changes will become visible
3382 * after the wmutex is unlocked. Since the DB is multi-version,
3383 * readers will get consistent data regardless of how fresh or
3384 * how stale their view of these values is.
3387 env->me_txns->mti_txnid = txn->mt_txnid;
3392 /** Check both meta pages to see which one is newer.
3393 * @param[in] env the environment handle
3394 * @return meta toggle (0 or 1).
3397 mdb_env_pick_meta(const MDB_env *env)
3399 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3403 mdb_env_create(MDB_env **env)
3407 e = calloc(1, sizeof(MDB_env));
3411 e->me_maxreaders = DEFAULT_READERS;
3412 e->me_maxdbs = e->me_numdbs = 2;
3413 e->me_fd = INVALID_HANDLE_VALUE;
3414 e->me_lfd = INVALID_HANDLE_VALUE;
3415 e->me_mfd = INVALID_HANDLE_VALUE;
3416 #ifdef MDB_USE_POSIX_SEM
3417 e->me_rmutex = SEM_FAILED;
3418 e->me_wmutex = SEM_FAILED;
3420 e->me_pid = getpid();
3421 GET_PAGESIZE(e->me_os_psize);
3422 VGMEMP_CREATE(e,0,0);
3428 mdb_env_map(MDB_env *env, void *addr, int newsize)
3431 unsigned int flags = env->me_flags;
3435 LONG sizelo, sizehi;
3436 sizelo = env->me_mapsize & 0xffffffff;
3437 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3439 /* Windows won't create mappings for zero length files.
3440 * Just allocate the maxsize right now.
3443 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3444 || !SetEndOfFile(env->me_fd)
3445 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3448 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3449 PAGE_READWRITE : PAGE_READONLY,
3450 sizehi, sizelo, NULL);
3453 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3454 FILE_MAP_WRITE : FILE_MAP_READ,
3455 0, 0, env->me_mapsize, addr);
3456 rc = env->me_map ? 0 : ErrCode();
3461 int prot = PROT_READ;
3462 if (flags & MDB_WRITEMAP) {
3464 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3467 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3469 if (env->me_map == MAP_FAILED) {
3474 if (flags & MDB_NORDAHEAD) {
3475 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3477 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3479 #ifdef POSIX_MADV_RANDOM
3480 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3481 #endif /* POSIX_MADV_RANDOM */
3482 #endif /* MADV_RANDOM */
3486 /* Can happen because the address argument to mmap() is just a
3487 * hint. mmap() can pick another, e.g. if the range is in use.
3488 * The MAP_FIXED flag would prevent that, but then mmap could
3489 * instead unmap existing pages to make room for the new map.
3491 if (addr && env->me_map != addr)
3492 return EBUSY; /* TODO: Make a new MDB_* error code? */
3494 p = (MDB_page *)env->me_map;
3495 env->me_metas[0] = METADATA(p);
3496 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3502 mdb_env_set_mapsize(MDB_env *env, size_t size)
3504 /* If env is already open, caller is responsible for making
3505 * sure there are no active txns.
3513 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3514 else if (size < env->me_mapsize) {
3515 /* If the configured size is smaller, make sure it's
3516 * still big enough. Silently round up to minimum if not.
3518 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3522 munmap(env->me_map, env->me_mapsize);
3523 env->me_mapsize = size;
3524 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3525 rc = mdb_env_map(env, old, 1);
3529 env->me_mapsize = size;
3531 env->me_maxpg = env->me_mapsize / env->me_psize;
3536 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3540 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3545 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3547 if (env->me_map || readers < 1)
3549 env->me_maxreaders = readers;
3554 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3556 if (!env || !readers)
3558 *readers = env->me_maxreaders;
3562 /** Further setup required for opening an MDB environment
3565 mdb_env_open2(MDB_env *env)
3567 unsigned int flags = env->me_flags;
3568 int i, newenv = 0, rc;
3572 /* See if we should use QueryLimited */
3574 if ((rc & 0xff) > 5)
3575 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3577 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3580 memset(&meta, 0, sizeof(meta));
3582 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3585 DPUTS("new mdbenv");
3587 env->me_psize = env->me_os_psize;
3588 if (env->me_psize > MAX_PAGESIZE)
3589 env->me_psize = MAX_PAGESIZE;
3591 env->me_psize = meta.mm_psize;
3594 /* Was a mapsize configured? */
3595 if (!env->me_mapsize) {
3596 /* If this is a new environment, take the default,
3597 * else use the size recorded in the existing env.
3599 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3600 } else if (env->me_mapsize < meta.mm_mapsize) {
3601 /* If the configured size is smaller, make sure it's
3602 * still big enough. Silently round up to minimum if not.
3604 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3605 if (env->me_mapsize < minsize)
3606 env->me_mapsize = minsize;
3609 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3614 if (flags & MDB_FIXEDMAP)
3615 meta.mm_address = env->me_map;
3616 i = mdb_env_init_meta(env, &meta);
3617 if (i != MDB_SUCCESS) {
3622 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3623 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3625 #if !(MDB_MAXKEYSIZE)
3626 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3628 env->me_maxpg = env->me_mapsize / env->me_psize;
3632 int toggle = mdb_env_pick_meta(env);
3633 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3635 DPRINTF(("opened database version %u, pagesize %u",
3636 env->me_metas[0]->mm_version, env->me_psize));
3637 DPRINTF(("using meta page %d", toggle));
3638 DPRINTF(("depth: %u", db->md_depth));
3639 DPRINTF(("entries: %"Z"u", db->md_entries));
3640 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3641 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3642 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3643 DPRINTF(("root: %"Z"u", db->md_root));
3651 /** Release a reader thread's slot in the reader lock table.
3652 * This function is called automatically when a thread exits.
3653 * @param[in] ptr This points to the slot in the reader lock table.
3656 mdb_env_reader_dest(void *ptr)
3658 MDB_reader *reader = ptr;
3664 /** Junk for arranging thread-specific callbacks on Windows. This is
3665 * necessarily platform and compiler-specific. Windows supports up
3666 * to 1088 keys. Let's assume nobody opens more than 64 environments
3667 * in a single process, for now. They can override this if needed.
3669 #ifndef MAX_TLS_KEYS
3670 #define MAX_TLS_KEYS 64
3672 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3673 static int mdb_tls_nkeys;
3675 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3679 case DLL_PROCESS_ATTACH: break;
3680 case DLL_THREAD_ATTACH: break;
3681 case DLL_THREAD_DETACH:
3682 for (i=0; i<mdb_tls_nkeys; i++) {
3683 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3685 mdb_env_reader_dest(r);
3689 case DLL_PROCESS_DETACH: break;
3694 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3696 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3700 /* Force some symbol references.
3701 * _tls_used forces the linker to create the TLS directory if not already done
3702 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3704 #pragma comment(linker, "/INCLUDE:_tls_used")
3705 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3706 #pragma const_seg(".CRT$XLB")
3707 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3708 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3711 #pragma comment(linker, "/INCLUDE:__tls_used")
3712 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3713 #pragma data_seg(".CRT$XLB")
3714 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3716 #endif /* WIN 32/64 */
3717 #endif /* !__GNUC__ */
3720 /** Downgrade the exclusive lock on the region back to shared */
3722 mdb_env_share_locks(MDB_env *env, int *excl)
3724 int rc = 0, toggle = mdb_env_pick_meta(env);
3726 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3731 /* First acquire a shared lock. The Unlock will
3732 * then release the existing exclusive lock.
3734 memset(&ov, 0, sizeof(ov));
3735 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3738 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3744 struct flock lock_info;
3745 /* The shared lock replaces the existing lock */
3746 memset((void *)&lock_info, 0, sizeof(lock_info));
3747 lock_info.l_type = F_RDLCK;
3748 lock_info.l_whence = SEEK_SET;
3749 lock_info.l_start = 0;
3750 lock_info.l_len = 1;
3751 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3752 (rc = ErrCode()) == EINTR) ;
3753 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3760 /** Try to get exlusive lock, otherwise shared.
3761 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3764 mdb_env_excl_lock(MDB_env *env, int *excl)
3768 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3772 memset(&ov, 0, sizeof(ov));
3773 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3780 struct flock lock_info;
3781 memset((void *)&lock_info, 0, sizeof(lock_info));
3782 lock_info.l_type = F_WRLCK;
3783 lock_info.l_whence = SEEK_SET;
3784 lock_info.l_start = 0;
3785 lock_info.l_len = 1;
3786 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3787 (rc = ErrCode()) == EINTR) ;
3791 # ifdef MDB_USE_POSIX_SEM
3792 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3795 lock_info.l_type = F_RDLCK;
3796 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3797 (rc = ErrCode()) == EINTR) ;
3807 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3809 * @(#) $Revision: 5.1 $
3810 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3811 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3813 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3817 * Please do not copyright this code. This code is in the public domain.
3819 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3820 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3821 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3822 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3823 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3824 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3825 * PERFORMANCE OF THIS SOFTWARE.
3828 * chongo <Landon Curt Noll> /\oo/\
3829 * http://www.isthe.com/chongo/
3831 * Share and Enjoy! :-)
3834 typedef unsigned long long mdb_hash_t;
3835 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3837 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3838 * @param[in] val value to hash
3839 * @param[in] hval initial value for hash
3840 * @return 64 bit hash
3842 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3843 * hval arg on the first call.
3846 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3848 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3849 unsigned char *end = s + val->mv_size;
3851 * FNV-1a hash each octet of the string
3854 /* xor the bottom with the current octet */
3855 hval ^= (mdb_hash_t)*s++;
3857 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3858 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3859 (hval << 7) + (hval << 8) + (hval << 40);
3861 /* return our new hash value */
3865 /** Hash the string and output the encoded hash.
3866 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3867 * very short name limits. We don't care about the encoding being reversible,
3868 * we just want to preserve as many bits of the input as possible in a
3869 * small printable string.
3870 * @param[in] str string to hash
3871 * @param[out] encbuf an array of 11 chars to hold the hash
3873 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3876 mdb_pack85(unsigned long l, char *out)
3880 for (i=0; i<5; i++) {
3881 *out++ = mdb_a85[l % 85];
3887 mdb_hash_enc(MDB_val *val, char *encbuf)
3889 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3891 mdb_pack85(h, encbuf);
3892 mdb_pack85(h>>32, encbuf+5);
3897 /** Open and/or initialize the lock region for the environment.
3898 * @param[in] env The MDB environment.
3899 * @param[in] lpath The pathname of the file used for the lock region.
3900 * @param[in] mode The Unix permissions for the file, if we create it.
3901 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3902 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3903 * @return 0 on success, non-zero on failure.
3906 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3909 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3911 # define MDB_ERRCODE_ROFS EROFS
3912 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3913 # define MDB_CLOEXEC O_CLOEXEC
3916 # define MDB_CLOEXEC 0
3923 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3924 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3925 FILE_ATTRIBUTE_NORMAL, NULL);
3927 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3929 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3931 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3936 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3937 /* Lose record locks when exec*() */
3938 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3939 fcntl(env->me_lfd, F_SETFD, fdflags);
3942 if (!(env->me_flags & MDB_NOTLS)) {
3943 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3946 env->me_flags |= MDB_ENV_TXKEY;
3948 /* Windows TLS callbacks need help finding their TLS info. */
3949 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3953 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3957 /* Try to get exclusive lock. If we succeed, then
3958 * nobody is using the lock region and we should initialize it.
3960 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3963 size = GetFileSize(env->me_lfd, NULL);
3965 size = lseek(env->me_lfd, 0, SEEK_END);
3966 if (size == -1) goto fail_errno;
3968 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3969 if (size < rsize && *excl > 0) {
3971 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3972 || !SetEndOfFile(env->me_lfd))
3975 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3979 size = rsize - sizeof(MDB_txninfo);
3980 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3985 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3987 if (!mh) goto fail_errno;
3988 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3990 if (!env->me_txns) goto fail_errno;
3992 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3994 if (m == MAP_FAILED) goto fail_errno;
4000 BY_HANDLE_FILE_INFORMATION stbuf;
4009 if (!mdb_sec_inited) {
4010 InitializeSecurityDescriptor(&mdb_null_sd,
4011 SECURITY_DESCRIPTOR_REVISION);
4012 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4013 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4014 mdb_all_sa.bInheritHandle = FALSE;
4015 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4018 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4019 idbuf.volume = stbuf.dwVolumeSerialNumber;
4020 idbuf.nhigh = stbuf.nFileIndexHigh;
4021 idbuf.nlow = stbuf.nFileIndexLow;
4022 val.mv_data = &idbuf;
4023 val.mv_size = sizeof(idbuf);
4024 mdb_hash_enc(&val, encbuf);
4025 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4026 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4027 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4028 if (!env->me_rmutex) goto fail_errno;
4029 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4030 if (!env->me_wmutex) goto fail_errno;
4031 #elif defined(MDB_USE_POSIX_SEM)
4040 #if defined(__NetBSD__)
4041 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4043 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4044 idbuf.dev = stbuf.st_dev;
4045 idbuf.ino = stbuf.st_ino;
4046 val.mv_data = &idbuf;
4047 val.mv_size = sizeof(idbuf);
4048 mdb_hash_enc(&val, encbuf);
4049 #ifdef MDB_SHORT_SEMNAMES
4050 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4052 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4053 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4054 /* Clean up after a previous run, if needed: Try to
4055 * remove both semaphores before doing anything else.
4057 sem_unlink(env->me_txns->mti_rmname);
4058 sem_unlink(env->me_txns->mti_wmname);
4059 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4060 O_CREAT|O_EXCL, mode, 1);
4061 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4062 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4063 O_CREAT|O_EXCL, mode, 1);
4064 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4065 #else /* MDB_USE_POSIX_SEM */
4066 pthread_mutexattr_t mattr;
4068 if ((rc = pthread_mutexattr_init(&mattr))
4069 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4070 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4071 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4073 pthread_mutexattr_destroy(&mattr);
4074 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4076 env->me_txns->mti_magic = MDB_MAGIC;
4077 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4078 env->me_txns->mti_txnid = 0;
4079 env->me_txns->mti_numreaders = 0;
4082 if (env->me_txns->mti_magic != MDB_MAGIC) {
4083 DPUTS("lock region has invalid magic");
4087 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4088 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4089 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4090 rc = MDB_VERSION_MISMATCH;
4094 if (rc && rc != EACCES && rc != EAGAIN) {
4098 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4099 if (!env->me_rmutex) goto fail_errno;
4100 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4101 if (!env->me_wmutex) goto fail_errno;
4102 #elif defined(MDB_USE_POSIX_SEM)
4103 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4104 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4105 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4106 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4117 /** The name of the lock file in the DB environment */
4118 #define LOCKNAME "/lock.mdb"
4119 /** The name of the data file in the DB environment */
4120 #define DATANAME "/data.mdb"
4121 /** The suffix of the lock file when no subdir is used */
4122 #define LOCKSUFF "-lock"
4123 /** Only a subset of the @ref mdb_env flags can be changed
4124 * at runtime. Changing other flags requires closing the
4125 * environment and re-opening it with the new flags.
4127 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4128 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4129 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4131 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4132 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4136 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4138 int oflags, rc, len, excl = -1;
4139 char *lpath, *dpath;
4141 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4145 if (flags & MDB_NOSUBDIR) {
4146 rc = len + sizeof(LOCKSUFF) + len + 1;
4148 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4153 if (flags & MDB_NOSUBDIR) {
4154 dpath = lpath + len + sizeof(LOCKSUFF);
4155 sprintf(lpath, "%s" LOCKSUFF, path);
4156 strcpy(dpath, path);
4158 dpath = lpath + len + sizeof(LOCKNAME);
4159 sprintf(lpath, "%s" LOCKNAME, path);
4160 sprintf(dpath, "%s" DATANAME, path);
4164 flags |= env->me_flags;
4165 if (flags & MDB_RDONLY) {
4166 /* silently ignore WRITEMAP when we're only getting read access */
4167 flags &= ~MDB_WRITEMAP;
4169 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4170 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4173 env->me_flags = flags |= MDB_ENV_ACTIVE;
4177 env->me_path = strdup(path);
4178 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4179 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4180 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4185 /* For RDONLY, get lockfile after we know datafile exists */
4186 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4187 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4193 if (F_ISSET(flags, MDB_RDONLY)) {
4194 oflags = GENERIC_READ;
4195 len = OPEN_EXISTING;
4197 oflags = GENERIC_READ|GENERIC_WRITE;
4200 mode = FILE_ATTRIBUTE_NORMAL;
4201 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4202 NULL, len, mode, NULL);
4204 if (F_ISSET(flags, MDB_RDONLY))
4207 oflags = O_RDWR | O_CREAT;
4209 env->me_fd = open(dpath, oflags, mode);
4211 if (env->me_fd == INVALID_HANDLE_VALUE) {
4216 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4217 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4222 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4223 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4224 env->me_mfd = env->me_fd;
4226 /* Synchronous fd for meta writes. Needed even with
4227 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4230 len = OPEN_EXISTING;
4231 env->me_mfd = CreateFile(dpath, oflags,
4232 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4233 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4236 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4238 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4243 DPRINTF(("opened dbenv %p", (void *) env));
4245 rc = mdb_env_share_locks(env, &excl);
4249 if (!((flags & MDB_RDONLY) ||
4250 (env->me_pbuf = calloc(1, env->me_psize))))
4256 mdb_env_close0(env, excl);
4262 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4264 mdb_env_close0(MDB_env *env, int excl)
4268 if (!(env->me_flags & MDB_ENV_ACTIVE))
4271 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4272 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4273 free(env->me_dbxs[i].md_name.mv_data);
4276 free(env->me_dbflags);
4279 free(env->me_dirty_list);
4280 mdb_midl_free(env->me_free_pgs);
4282 if (env->me_flags & MDB_ENV_TXKEY) {
4283 pthread_key_delete(env->me_txkey);
4285 /* Delete our key from the global list */
4286 for (i=0; i<mdb_tls_nkeys; i++)
4287 if (mdb_tls_keys[i] == env->me_txkey) {
4288 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4296 munmap(env->me_map, env->me_mapsize);
4298 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4299 (void) close(env->me_mfd);
4300 if (env->me_fd != INVALID_HANDLE_VALUE)
4301 (void) close(env->me_fd);
4303 MDB_PID_T pid = env->me_pid;
4304 /* Clearing readers is done in this function because
4305 * me_txkey with its destructor must be disabled first.
4307 for (i = env->me_numreaders; --i >= 0; )
4308 if (env->me_txns->mti_readers[i].mr_pid == pid)
4309 env->me_txns->mti_readers[i].mr_pid = 0;
4311 if (env->me_rmutex) {
4312 CloseHandle(env->me_rmutex);
4313 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4315 /* Windows automatically destroys the mutexes when
4316 * the last handle closes.
4318 #elif defined(MDB_USE_POSIX_SEM)
4319 if (env->me_rmutex != SEM_FAILED) {
4320 sem_close(env->me_rmutex);
4321 if (env->me_wmutex != SEM_FAILED)
4322 sem_close(env->me_wmutex);
4323 /* If we have the filelock: If we are the
4324 * only remaining user, clean up semaphores.
4327 mdb_env_excl_lock(env, &excl);
4329 sem_unlink(env->me_txns->mti_rmname);
4330 sem_unlink(env->me_txns->mti_wmname);
4334 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4336 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4339 /* Unlock the lockfile. Windows would have unlocked it
4340 * after closing anyway, but not necessarily at once.
4342 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4345 (void) close(env->me_lfd);
4348 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4352 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4354 MDB_txn *txn = NULL;
4360 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4364 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4367 /* Do the lock/unlock of the reader mutex before starting the
4368 * write txn. Otherwise other read txns could block writers.
4370 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4375 /* We must start the actual read txn after blocking writers */
4376 mdb_txn_reset0(txn, "reset-stage1");
4378 /* Temporarily block writers until we snapshot the meta pages */
4381 rc = mdb_txn_renew0(txn);
4383 UNLOCK_MUTEX_W(env);
4388 wsize = env->me_psize * 2;
4392 DO_WRITE(rc, fd, ptr, w2, len);
4396 } else if (len > 0) {
4402 /* Non-blocking or async handles are not supported */
4408 UNLOCK_MUTEX_W(env);
4413 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4415 if (wsize > MAX_WRITE)
4419 DO_WRITE(rc, fd, ptr, w2, len);
4423 } else if (len > 0) {
4440 mdb_env_copy(MDB_env *env, const char *path)
4444 HANDLE newfd = INVALID_HANDLE_VALUE;
4446 if (env->me_flags & MDB_NOSUBDIR) {
4447 lpath = (char *)path;
4450 len += sizeof(DATANAME);
4451 lpath = malloc(len);
4454 sprintf(lpath, "%s" DATANAME, path);
4457 /* The destination path must exist, but the destination file must not.
4458 * We don't want the OS to cache the writes, since the source data is
4459 * already in the OS cache.
4462 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4463 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4465 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4467 if (newfd == INVALID_HANDLE_VALUE) {
4473 /* Set O_DIRECT if the file system supports it */
4474 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4475 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4477 #ifdef F_NOCACHE /* __APPLE__ */
4478 rc = fcntl(newfd, F_NOCACHE, 1);
4485 rc = mdb_env_copyfd(env, newfd);
4488 if (!(env->me_flags & MDB_NOSUBDIR))
4490 if (newfd != INVALID_HANDLE_VALUE)
4491 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4498 mdb_env_close(MDB_env *env)
4505 VGMEMP_DESTROY(env);
4506 while ((dp = env->me_dpages) != NULL) {
4507 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4508 env->me_dpages = dp->mp_next;
4512 mdb_env_close0(env, 0);
4516 /** Compare two items pointing at aligned size_t's */
4518 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4520 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4521 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4524 /** Compare two items pointing at aligned unsigned int's */
4526 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4528 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4529 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4532 /** Compare two items pointing at unsigned ints of unknown alignment.
4533 * Nodes and keys are guaranteed to be 2-byte aligned.
4536 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4538 #if BYTE_ORDER == LITTLE_ENDIAN
4539 unsigned short *u, *c;
4542 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4543 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4546 } while(!x && u > (unsigned short *)a->mv_data);
4549 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4553 /** Compare two items pointing at size_t's of unknown alignment. */
4554 #ifdef MISALIGNED_OK
4555 # define mdb_cmp_clong mdb_cmp_long
4557 # define mdb_cmp_clong mdb_cmp_cint
4560 /** Compare two items lexically */
4562 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4569 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4575 diff = memcmp(a->mv_data, b->mv_data, len);
4576 return diff ? diff : len_diff<0 ? -1 : len_diff;
4579 /** Compare two items in reverse byte order */
4581 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4583 const unsigned char *p1, *p2, *p1_lim;
4587 p1_lim = (const unsigned char *)a->mv_data;
4588 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4589 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4591 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4597 while (p1 > p1_lim) {
4598 diff = *--p1 - *--p2;
4602 return len_diff<0 ? -1 : len_diff;
4605 /** Search for key within a page, using binary search.
4606 * Returns the smallest entry larger or equal to the key.
4607 * If exactp is non-null, stores whether the found entry was an exact match
4608 * in *exactp (1 or 0).
4609 * Updates the cursor index with the index of the found entry.
4610 * If no entry larger or equal to the key is found, returns NULL.
4613 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4615 unsigned int i = 0, nkeys;
4618 MDB_page *mp = mc->mc_pg[mc->mc_top];
4619 MDB_node *node = NULL;
4624 nkeys = NUMKEYS(mp);
4626 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4627 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4630 low = IS_LEAF(mp) ? 0 : 1;
4632 cmp = mc->mc_dbx->md_cmp;
4634 /* Branch pages have no data, so if using integer keys,
4635 * alignment is guaranteed. Use faster mdb_cmp_int.
4637 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4638 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4645 nodekey.mv_size = mc->mc_db->md_pad;
4646 node = NODEPTR(mp, 0); /* fake */
4647 while (low <= high) {
4648 i = (low + high) >> 1;
4649 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4650 rc = cmp(key, &nodekey);
4651 DPRINTF(("found leaf index %u [%s], rc = %i",
4652 i, DKEY(&nodekey), rc));
4661 while (low <= high) {
4662 i = (low + high) >> 1;
4664 node = NODEPTR(mp, i);
4665 nodekey.mv_size = NODEKSZ(node);
4666 nodekey.mv_data = NODEKEY(node);
4668 rc = cmp(key, &nodekey);
4671 DPRINTF(("found leaf index %u [%s], rc = %i",
4672 i, DKEY(&nodekey), rc));
4674 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4675 i, DKEY(&nodekey), NODEPGNO(node), rc));
4686 if (rc > 0) { /* Found entry is less than the key. */
4687 i++; /* Skip to get the smallest entry larger than key. */
4689 node = NODEPTR(mp, i);
4692 *exactp = (rc == 0 && nkeys > 0);
4693 /* store the key index */
4694 mc->mc_ki[mc->mc_top] = i;
4696 /* There is no entry larger or equal to the key. */
4699 /* nodeptr is fake for LEAF2 */
4705 mdb_cursor_adjust(MDB_cursor *mc, func)
4709 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4710 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4717 /** Pop a page off the top of the cursor's stack. */
4719 mdb_cursor_pop(MDB_cursor *mc)
4723 MDB_page *top = mc->mc_pg[mc->mc_top];
4729 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4730 DDBI(mc), (void *) mc));
4734 /** Push a page onto the top of the cursor's stack. */
4736 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4738 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4739 DDBI(mc), (void *) mc));
4741 if (mc->mc_snum >= CURSOR_STACK) {
4742 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4743 return MDB_CURSOR_FULL;
4746 mc->mc_top = mc->mc_snum++;
4747 mc->mc_pg[mc->mc_top] = mp;
4748 mc->mc_ki[mc->mc_top] = 0;
4753 /** Find the address of the page corresponding to a given page number.
4754 * @param[in] txn the transaction for this access.
4755 * @param[in] pgno the page number for the page to retrieve.
4756 * @param[out] ret address of a pointer where the page's address will be stored.
4757 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4758 * @return 0 on success, non-zero on failure.
4761 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4763 MDB_env *env = txn->mt_env;
4767 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4771 MDB_ID2L dl = tx2->mt_u.dirty_list;
4773 /* Spilled pages were dirtied in this txn and flushed
4774 * because the dirty list got full. Bring this page
4775 * back in from the map (but don't unspill it here,
4776 * leave that unless page_touch happens again).
4778 if (tx2->mt_spill_pgs) {
4779 MDB_ID pn = pgno << 1;
4780 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4781 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4782 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4787 unsigned x = mdb_mid2l_search(dl, pgno);
4788 if (x <= dl[0].mid && dl[x].mid == pgno) {
4794 } while ((tx2 = tx2->mt_parent) != NULL);
4797 if (pgno < txn->mt_next_pgno) {
4799 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4801 DPRINTF(("page %"Z"u not found", pgno));
4802 txn->mt_flags |= MDB_TXN_ERROR;
4803 return MDB_PAGE_NOTFOUND;
4813 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4814 * The cursor is at the root page, set up the rest of it.
4817 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4819 MDB_page *mp = mc->mc_pg[mc->mc_top];
4823 while (IS_BRANCH(mp)) {
4827 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4828 mdb_cassert(mc, NUMKEYS(mp) > 1);
4829 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4831 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4833 if (flags & MDB_PS_LAST)
4834 i = NUMKEYS(mp) - 1;
4837 node = mdb_node_search(mc, key, &exact);
4839 i = NUMKEYS(mp) - 1;
4841 i = mc->mc_ki[mc->mc_top];
4843 mdb_cassert(mc, i > 0);
4847 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4850 mdb_cassert(mc, i < NUMKEYS(mp));
4851 node = NODEPTR(mp, i);
4853 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4856 mc->mc_ki[mc->mc_top] = i;
4857 if ((rc = mdb_cursor_push(mc, mp)))
4860 if (flags & MDB_PS_MODIFY) {
4861 if ((rc = mdb_page_touch(mc)) != 0)
4863 mp = mc->mc_pg[mc->mc_top];
4868 DPRINTF(("internal error, index points to a %02X page!?",
4870 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4871 return MDB_CORRUPTED;
4874 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4875 key ? DKEY(key) : "null"));
4876 mc->mc_flags |= C_INITIALIZED;
4877 mc->mc_flags &= ~C_EOF;
4882 /** Search for the lowest key under the current branch page.
4883 * This just bypasses a NUMKEYS check in the current page
4884 * before calling mdb_page_search_root(), because the callers
4885 * are all in situations where the current page is known to
4889 mdb_page_search_lowest(MDB_cursor *mc)
4891 MDB_page *mp = mc->mc_pg[mc->mc_top];
4892 MDB_node *node = NODEPTR(mp, 0);
4895 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4898 mc->mc_ki[mc->mc_top] = 0;
4899 if ((rc = mdb_cursor_push(mc, mp)))
4901 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4904 /** Search for the page a given key should be in.
4905 * Push it and its parent pages on the cursor stack.
4906 * @param[in,out] mc the cursor for this operation.
4907 * @param[in] key the key to search for, or NULL for first/last page.
4908 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4909 * are touched (updated with new page numbers).
4910 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4911 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4912 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4913 * @return 0 on success, non-zero on failure.
4916 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4921 /* Make sure the txn is still viable, then find the root from
4922 * the txn's db table and set it as the root of the cursor's stack.
4924 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4925 DPUTS("transaction has failed, must abort");
4928 /* Make sure we're using an up-to-date root */
4929 if (*mc->mc_dbflag & DB_STALE) {
4931 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4932 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4939 MDB_node *leaf = mdb_node_search(&mc2,
4940 &mc->mc_dbx->md_name, &exact);
4942 return MDB_NOTFOUND;
4943 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4946 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4948 /* The txn may not know this DBI, or another process may
4949 * have dropped and recreated the DB with other flags.
4951 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4952 return MDB_INCOMPATIBLE;
4953 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4955 *mc->mc_dbflag &= ~DB_STALE;
4957 root = mc->mc_db->md_root;
4959 if (root == P_INVALID) { /* Tree is empty. */
4960 DPUTS("tree is empty");
4961 return MDB_NOTFOUND;
4965 mdb_cassert(mc, root > 1);
4966 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4967 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4973 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4974 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4976 if (flags & MDB_PS_MODIFY) {
4977 if ((rc = mdb_page_touch(mc)))
4981 if (flags & MDB_PS_ROOTONLY)
4984 return mdb_page_search_root(mc, key, flags);
4988 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4990 MDB_txn *txn = mc->mc_txn;
4991 pgno_t pg = mp->mp_pgno;
4992 unsigned x = 0, ovpages = mp->mp_pages;
4993 MDB_env *env = txn->mt_env;
4994 MDB_IDL sl = txn->mt_spill_pgs;
4995 MDB_ID pn = pg << 1;
4998 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4999 /* If the page is dirty or on the spill list we just acquired it,
5000 * so we should give it back to our current free list, if any.
5001 * Otherwise put it onto the list of pages we freed in this txn.
5003 * Won't create me_pghead: me_pglast must be inited along with it.
5004 * Unsupported in nested txns: They would need to hide the page
5005 * range in ancestor txns' dirty and spilled lists.
5007 if (env->me_pghead &&
5009 ((mp->mp_flags & P_DIRTY) ||
5010 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5014 MDB_ID2 *dl, ix, iy;
5015 rc = mdb_midl_need(&env->me_pghead, ovpages);
5018 if (!(mp->mp_flags & P_DIRTY)) {
5019 /* This page is no longer spilled */
5026 /* Remove from dirty list */
5027 dl = txn->mt_u.dirty_list;
5029 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5035 mdb_cassert(mc, x > 1);
5037 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5038 txn->mt_flags |= MDB_TXN_ERROR;
5039 return MDB_CORRUPTED;
5042 if (!(env->me_flags & MDB_WRITEMAP))
5043 mdb_dpage_free(env, mp);
5045 /* Insert in me_pghead */
5046 mop = env->me_pghead;
5047 j = mop[0] + ovpages;
5048 for (i = mop[0]; i && mop[i] < pg; i--)
5054 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5058 mc->mc_db->md_overflow_pages -= ovpages;
5062 /** Return the data associated with a given node.
5063 * @param[in] txn The transaction for this operation.
5064 * @param[in] leaf The node being read.
5065 * @param[out] data Updated to point to the node's data.
5066 * @return 0 on success, non-zero on failure.
5069 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5071 MDB_page *omp; /* overflow page */
5075 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5076 data->mv_size = NODEDSZ(leaf);
5077 data->mv_data = NODEDATA(leaf);
5081 /* Read overflow data.
5083 data->mv_size = NODEDSZ(leaf);
5084 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5085 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5086 DPRINTF(("read overflow page %"Z"u failed", pgno));
5089 data->mv_data = METADATA(omp);
5095 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5096 MDB_val *key, MDB_val *data)
5103 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5105 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5108 if (txn->mt_flags & MDB_TXN_ERROR)
5111 mdb_cursor_init(&mc, txn, dbi, &mx);
5112 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5115 /** Find a sibling for a page.
5116 * Replaces the page at the top of the cursor's stack with the
5117 * specified sibling, if one exists.
5118 * @param[in] mc The cursor for this operation.
5119 * @param[in] move_right Non-zero if the right sibling is requested,
5120 * otherwise the left sibling.
5121 * @return 0 on success, non-zero on failure.
5124 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5130 if (mc->mc_snum < 2) {
5131 return MDB_NOTFOUND; /* root has no siblings */
5135 DPRINTF(("parent page is page %"Z"u, index %u",
5136 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5138 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5139 : (mc->mc_ki[mc->mc_top] == 0)) {
5140 DPRINTF(("no more keys left, moving to %s sibling",
5141 move_right ? "right" : "left"));
5142 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5143 /* undo cursor_pop before returning */
5150 mc->mc_ki[mc->mc_top]++;
5152 mc->mc_ki[mc->mc_top]--;
5153 DPRINTF(("just moving to %s index key %u",
5154 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5156 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5158 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5159 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5160 /* mc will be inconsistent if caller does mc_snum++ as above */
5161 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5165 mdb_cursor_push(mc, mp);
5167 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5172 /** Move the cursor to the next data item. */
5174 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5180 if (mc->mc_flags & C_EOF) {
5181 return MDB_NOTFOUND;
5184 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5186 mp = mc->mc_pg[mc->mc_top];
5188 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5189 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5190 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5191 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5192 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5193 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5194 if (rc == MDB_SUCCESS)
5195 MDB_GET_KEY(leaf, key);
5200 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5201 if (op == MDB_NEXT_DUP)
5202 return MDB_NOTFOUND;
5206 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5207 mdb_dbg_pgno(mp), (void *) mc));
5208 if (mc->mc_flags & C_DEL)
5211 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5212 DPUTS("=====> move to next sibling page");
5213 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5214 mc->mc_flags |= C_EOF;
5217 mp = mc->mc_pg[mc->mc_top];
5218 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5220 mc->mc_ki[mc->mc_top]++;
5223 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5224 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5227 key->mv_size = mc->mc_db->md_pad;
5228 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5232 mdb_cassert(mc, IS_LEAF(mp));
5233 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5235 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5236 mdb_xcursor_init1(mc, leaf);
5239 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5242 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5243 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5244 if (rc != MDB_SUCCESS)
5249 MDB_GET_KEY(leaf, key);
5253 /** Move the cursor to the previous data item. */
5255 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5261 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5263 mp = mc->mc_pg[mc->mc_top];
5265 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5266 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5267 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5268 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5269 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5270 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5271 if (rc == MDB_SUCCESS)
5272 MDB_GET_KEY(leaf, key);
5276 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5277 if (op == MDB_PREV_DUP)
5278 return MDB_NOTFOUND;
5283 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5284 mdb_dbg_pgno(mp), (void *) mc));
5286 if (mc->mc_ki[mc->mc_top] == 0) {
5287 DPUTS("=====> move to prev sibling page");
5288 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5291 mp = mc->mc_pg[mc->mc_top];
5292 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5293 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5295 mc->mc_ki[mc->mc_top]--;
5297 mc->mc_flags &= ~C_EOF;
5299 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5300 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5303 key->mv_size = mc->mc_db->md_pad;
5304 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5308 mdb_cassert(mc, IS_LEAF(mp));
5309 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5311 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5312 mdb_xcursor_init1(mc, leaf);
5315 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5318 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5319 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5320 if (rc != MDB_SUCCESS)
5325 MDB_GET_KEY(leaf, key);
5329 /** Set the cursor on a specific data item. */
5331 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5332 MDB_cursor_op op, int *exactp)
5336 MDB_node *leaf = NULL;
5339 if (key->mv_size == 0)
5340 return MDB_BAD_VALSIZE;
5343 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5345 /* See if we're already on the right page */
5346 if (mc->mc_flags & C_INITIALIZED) {
5349 mp = mc->mc_pg[mc->mc_top];
5351 mc->mc_ki[mc->mc_top] = 0;
5352 return MDB_NOTFOUND;
5354 if (mp->mp_flags & P_LEAF2) {
5355 nodekey.mv_size = mc->mc_db->md_pad;
5356 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5358 leaf = NODEPTR(mp, 0);
5359 MDB_GET_KEY2(leaf, nodekey);
5361 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5363 /* Probably happens rarely, but first node on the page
5364 * was the one we wanted.
5366 mc->mc_ki[mc->mc_top] = 0;
5373 unsigned int nkeys = NUMKEYS(mp);
5375 if (mp->mp_flags & P_LEAF2) {
5376 nodekey.mv_data = LEAF2KEY(mp,
5377 nkeys-1, nodekey.mv_size);
5379 leaf = NODEPTR(mp, nkeys-1);
5380 MDB_GET_KEY2(leaf, nodekey);
5382 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5384 /* last node was the one we wanted */
5385 mc->mc_ki[mc->mc_top] = nkeys-1;
5391 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5392 /* This is definitely the right page, skip search_page */
5393 if (mp->mp_flags & P_LEAF2) {
5394 nodekey.mv_data = LEAF2KEY(mp,
5395 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5397 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5398 MDB_GET_KEY2(leaf, nodekey);
5400 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5402 /* current node was the one we wanted */
5412 /* If any parents have right-sibs, search.
5413 * Otherwise, there's nothing further.
5415 for (i=0; i<mc->mc_top; i++)
5417 NUMKEYS(mc->mc_pg[i])-1)
5419 if (i == mc->mc_top) {
5420 /* There are no other pages */
5421 mc->mc_ki[mc->mc_top] = nkeys;
5422 return MDB_NOTFOUND;
5426 /* There are no other pages */
5427 mc->mc_ki[mc->mc_top] = 0;
5428 if (op == MDB_SET_RANGE && !exactp) {
5432 return MDB_NOTFOUND;
5436 rc = mdb_page_search(mc, key, 0);
5437 if (rc != MDB_SUCCESS)
5440 mp = mc->mc_pg[mc->mc_top];
5441 mdb_cassert(mc, IS_LEAF(mp));
5444 leaf = mdb_node_search(mc, key, exactp);
5445 if (exactp != NULL && !*exactp) {
5446 /* MDB_SET specified and not an exact match. */
5447 return MDB_NOTFOUND;
5451 DPUTS("===> inexact leaf not found, goto sibling");
5452 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5453 return rc; /* no entries matched */
5454 mp = mc->mc_pg[mc->mc_top];
5455 mdb_cassert(mc, IS_LEAF(mp));
5456 leaf = NODEPTR(mp, 0);
5460 mc->mc_flags |= C_INITIALIZED;
5461 mc->mc_flags &= ~C_EOF;
5464 key->mv_size = mc->mc_db->md_pad;
5465 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5469 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5470 mdb_xcursor_init1(mc, leaf);
5473 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5474 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5475 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5478 if (op == MDB_GET_BOTH) {
5484 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5485 if (rc != MDB_SUCCESS)
5488 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5490 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5492 rc = mc->mc_dbx->md_dcmp(data, &d2);
5494 if (op == MDB_GET_BOTH || rc > 0)
5495 return MDB_NOTFOUND;
5502 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5503 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5508 /* The key already matches in all other cases */
5509 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5510 MDB_GET_KEY(leaf, key);
5511 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5516 /** Move the cursor to the first item in the database. */
5518 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5524 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5526 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5527 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5528 if (rc != MDB_SUCCESS)
5531 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5533 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5534 mc->mc_flags |= C_INITIALIZED;
5535 mc->mc_flags &= ~C_EOF;
5537 mc->mc_ki[mc->mc_top] = 0;
5539 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5540 key->mv_size = mc->mc_db->md_pad;
5541 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5546 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5547 mdb_xcursor_init1(mc, leaf);
5548 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5552 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5556 MDB_GET_KEY(leaf, key);
5560 /** Move the cursor to the last item in the database. */
5562 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5568 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5570 if (!(mc->mc_flags & C_EOF)) {
5572 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5573 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5574 if (rc != MDB_SUCCESS)
5577 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5580 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5581 mc->mc_flags |= C_INITIALIZED|C_EOF;
5582 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5584 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5585 key->mv_size = mc->mc_db->md_pad;
5586 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5591 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5592 mdb_xcursor_init1(mc, leaf);
5593 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5597 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5602 MDB_GET_KEY(leaf, key);
5607 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5612 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5617 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5621 case MDB_GET_CURRENT:
5622 if (!(mc->mc_flags & C_INITIALIZED)) {
5625 MDB_page *mp = mc->mc_pg[mc->mc_top];
5626 int nkeys = NUMKEYS(mp);
5627 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5628 mc->mc_ki[mc->mc_top] = nkeys;
5634 key->mv_size = mc->mc_db->md_pad;
5635 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5637 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5638 MDB_GET_KEY(leaf, key);
5640 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5641 if (mc->mc_flags & C_DEL)
5642 mdb_xcursor_init1(mc, leaf);
5643 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5645 rc = mdb_node_read(mc->mc_txn, leaf, data);
5652 case MDB_GET_BOTH_RANGE:
5657 if (mc->mc_xcursor == NULL) {
5658 rc = MDB_INCOMPATIBLE;
5668 rc = mdb_cursor_set(mc, key, data, op,
5669 op == MDB_SET_RANGE ? NULL : &exact);
5672 case MDB_GET_MULTIPLE:
5673 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5677 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5678 rc = MDB_INCOMPATIBLE;
5682 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5683 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5686 case MDB_NEXT_MULTIPLE:
5691 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5692 rc = MDB_INCOMPATIBLE;
5695 if (!(mc->mc_flags & C_INITIALIZED))
5696 rc = mdb_cursor_first(mc, key, data);
5698 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5699 if (rc == MDB_SUCCESS) {
5700 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5703 mx = &mc->mc_xcursor->mx_cursor;
5704 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5706 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5707 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5715 case MDB_NEXT_NODUP:
5716 if (!(mc->mc_flags & C_INITIALIZED))
5717 rc = mdb_cursor_first(mc, key, data);
5719 rc = mdb_cursor_next(mc, key, data, op);
5723 case MDB_PREV_NODUP:
5724 if (!(mc->mc_flags & C_INITIALIZED)) {
5725 rc = mdb_cursor_last(mc, key, data);
5728 mc->mc_flags |= C_INITIALIZED;
5729 mc->mc_ki[mc->mc_top]++;
5731 rc = mdb_cursor_prev(mc, key, data, op);
5734 rc = mdb_cursor_first(mc, key, data);
5737 mfunc = mdb_cursor_first;
5739 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5743 if (mc->mc_xcursor == NULL) {
5744 rc = MDB_INCOMPATIBLE;
5747 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5751 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5754 rc = mdb_cursor_last(mc, key, data);
5757 mfunc = mdb_cursor_last;
5760 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5765 if (mc->mc_flags & C_DEL)
5766 mc->mc_flags ^= C_DEL;
5771 /** Touch all the pages in the cursor stack. Set mc_top.
5772 * Makes sure all the pages are writable, before attempting a write operation.
5773 * @param[in] mc The cursor to operate on.
5776 mdb_cursor_touch(MDB_cursor *mc)
5778 int rc = MDB_SUCCESS;
5780 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5783 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5784 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5787 *mc->mc_dbflag |= DB_DIRTY;
5792 rc = mdb_page_touch(mc);
5793 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5794 mc->mc_top = mc->mc_snum-1;
5799 /** Do not spill pages to disk if txn is getting full, may fail instead */
5800 #define MDB_NOSPILL 0x8000
5803 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5806 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5808 MDB_node *leaf = NULL;
5811 MDB_val xdata, *rdata, dkey, olddata;
5813 int do_sub = 0, insert_key, insert_data;
5814 unsigned int mcount = 0, dcount = 0, nospill;
5817 unsigned int nflags;
5820 if (mc == NULL || key == NULL)
5823 env = mc->mc_txn->mt_env;
5825 /* Check this first so counter will always be zero on any
5828 if (flags & MDB_MULTIPLE) {
5829 dcount = data[1].mv_size;
5830 data[1].mv_size = 0;
5831 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5832 return MDB_INCOMPATIBLE;
5835 nospill = flags & MDB_NOSPILL;
5836 flags &= ~MDB_NOSPILL;
5838 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5839 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5841 if (key->mv_size-1 >= ENV_MAXKEY(env))
5842 return MDB_BAD_VALSIZE;
5844 #if SIZE_MAX > MAXDATASIZE
5845 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5846 return MDB_BAD_VALSIZE;
5848 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5849 return MDB_BAD_VALSIZE;
5852 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5853 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5857 if (flags == MDB_CURRENT) {
5858 if (!(mc->mc_flags & C_INITIALIZED))
5861 } else if (mc->mc_db->md_root == P_INVALID) {
5862 /* new database, cursor has nothing to point to */
5865 mc->mc_flags &= ~C_INITIALIZED;
5870 if (flags & MDB_APPEND) {
5872 rc = mdb_cursor_last(mc, &k2, &d2);
5874 rc = mc->mc_dbx->md_cmp(key, &k2);
5877 mc->mc_ki[mc->mc_top]++;
5879 /* new key is <= last key */
5884 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5886 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5887 DPRINTF(("duplicate key [%s]", DKEY(key)));
5889 return MDB_KEYEXIST;
5891 if (rc && rc != MDB_NOTFOUND)
5895 if (mc->mc_flags & C_DEL)
5896 mc->mc_flags ^= C_DEL;
5898 /* Cursor is positioned, check for room in the dirty list */
5900 if (flags & MDB_MULTIPLE) {
5902 xdata.mv_size = data->mv_size * dcount;
5906 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5910 if (rc == MDB_NO_ROOT) {
5912 /* new database, write a root leaf page */
5913 DPUTS("allocating new root leaf page");
5914 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5917 mdb_cursor_push(mc, np);
5918 mc->mc_db->md_root = np->mp_pgno;
5919 mc->mc_db->md_depth++;
5920 *mc->mc_dbflag |= DB_DIRTY;
5921 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5923 np->mp_flags |= P_LEAF2;
5924 mc->mc_flags |= C_INITIALIZED;
5926 /* make sure all cursor pages are writable */
5927 rc2 = mdb_cursor_touch(mc);
5932 insert_key = insert_data = rc;
5934 /* The key does not exist */
5935 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5936 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5937 LEAFSIZE(key, data) > env->me_nodemax)
5939 /* Too big for a node, insert in sub-DB */
5940 fp_flags = P_LEAF|P_DIRTY;
5942 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5943 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5947 /* there's only a key anyway, so this is a no-op */
5948 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5949 unsigned int ksize = mc->mc_db->md_pad;
5950 if (key->mv_size != ksize)
5951 return MDB_BAD_VALSIZE;
5952 if (flags == MDB_CURRENT) {
5953 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5954 memcpy(ptr, key->mv_data, ksize);
5960 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5961 olddata.mv_size = NODEDSZ(leaf);
5962 olddata.mv_data = NODEDATA(leaf);
5965 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5966 /* Prepare (sub-)page/sub-DB to accept the new item,
5967 * if needed. fp: old sub-page or a header faking
5968 * it. mp: new (sub-)page. offset: growth in page
5969 * size. xdata: node data with new page or DB.
5971 unsigned i, offset = 0;
5972 mp = fp = xdata.mv_data = env->me_pbuf;
5973 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5975 /* Was a single item before, must convert now */
5976 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5977 /* Just overwrite the current item */
5978 if (flags == MDB_CURRENT)
5981 #if UINT_MAX < SIZE_MAX
5982 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5983 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
5985 /* if data matches, skip it */
5986 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5987 if (flags & MDB_NODUPDATA)
5988 return MDB_KEYEXIST;
5993 /* Back up original data item */
5994 dkey.mv_size = olddata.mv_size;
5995 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5997 /* Make sub-page header for the dup items, with dummy body */
5998 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5999 fp->mp_lower = PAGEHDRSZ;
6000 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6001 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6002 fp->mp_flags |= P_LEAF2;
6003 fp->mp_pad = data->mv_size;
6004 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6006 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6007 (dkey.mv_size & 1) + (data->mv_size & 1);
6009 fp->mp_upper = xdata.mv_size;
6010 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
6011 } else if (leaf->mn_flags & F_SUBDATA) {
6012 /* Data is on sub-DB, just store it */
6013 flags |= F_DUPDATA|F_SUBDATA;
6016 /* Data is on sub-page */
6017 fp = olddata.mv_data;
6020 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6021 offset = EVEN(NODESIZE + sizeof(indx_t) +
6025 offset = fp->mp_pad;
6026 if (SIZELEFT(fp) < offset) {
6027 offset *= 4; /* space for 4 more */
6030 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6032 fp->mp_flags |= P_DIRTY;
6033 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6034 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6038 xdata.mv_size = olddata.mv_size + offset;
6041 fp_flags = fp->mp_flags;
6042 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6043 /* Too big for a sub-page, convert to sub-DB */
6044 fp_flags &= ~P_SUBP;
6046 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6047 fp_flags |= P_LEAF2;
6048 dummy.md_pad = fp->mp_pad;
6049 dummy.md_flags = MDB_DUPFIXED;
6050 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6051 dummy.md_flags |= MDB_INTEGERKEY;
6057 dummy.md_branch_pages = 0;
6058 dummy.md_leaf_pages = 1;
6059 dummy.md_overflow_pages = 0;
6060 dummy.md_entries = NUMKEYS(fp);
6061 xdata.mv_size = sizeof(MDB_db);
6062 xdata.mv_data = &dummy;
6063 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6065 offset = env->me_psize - olddata.mv_size;
6066 flags |= F_DUPDATA|F_SUBDATA;
6067 dummy.md_root = mp->mp_pgno;
6070 mp->mp_flags = fp_flags | P_DIRTY;
6071 mp->mp_pad = fp->mp_pad;
6072 mp->mp_lower = fp->mp_lower;
6073 mp->mp_upper = fp->mp_upper + offset;
6074 if (fp_flags & P_LEAF2) {
6075 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6077 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
6078 olddata.mv_size - fp->mp_upper);
6079 for (i=0; i<NUMKEYS(fp); i++)
6080 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6088 mdb_node_del(mc, 0);
6092 /* overflow page overwrites need special handling */
6093 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6096 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6098 memcpy(&pg, olddata.mv_data, sizeof(pg));
6099 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6101 ovpages = omp->mp_pages;
6103 /* Is the ov page large enough? */
6104 if (ovpages >= dpages) {
6105 if (!(omp->mp_flags & P_DIRTY) &&
6106 (level || (env->me_flags & MDB_WRITEMAP)))
6108 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6111 level = 0; /* dirty in this txn or clean */
6114 if (omp->mp_flags & P_DIRTY) {
6115 /* yes, overwrite it. Note in this case we don't
6116 * bother to try shrinking the page if the new data
6117 * is smaller than the overflow threshold.
6120 /* It is writable only in a parent txn */
6121 size_t sz = (size_t) env->me_psize * ovpages, off;
6122 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6128 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6129 mdb_cassert(mc, rc2 == 0);
6130 if (!(flags & MDB_RESERVE)) {
6131 /* Copy end of page, adjusting alignment so
6132 * compiler may copy words instead of bytes.
6134 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6135 memcpy((size_t *)((char *)np + off),
6136 (size_t *)((char *)omp + off), sz - off);
6139 memcpy(np, omp, sz); /* Copy beginning of page */
6142 SETDSZ(leaf, data->mv_size);
6143 if (F_ISSET(flags, MDB_RESERVE))
6144 data->mv_data = METADATA(omp);
6146 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6150 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6152 } else if (data->mv_size == olddata.mv_size) {
6153 /* same size, just replace it. Note that we could
6154 * also reuse this node if the new data is smaller,
6155 * but instead we opt to shrink the node in that case.
6157 if (F_ISSET(flags, MDB_RESERVE))
6158 data->mv_data = olddata.mv_data;
6159 else if (!(mc->mc_flags & C_SUB))
6160 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6162 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6165 mdb_node_del(mc, 0);
6171 nflags = flags & NODE_ADD_FLAGS;
6172 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6173 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6174 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6175 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6177 nflags |= MDB_SPLIT_REPLACE;
6178 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6180 /* There is room already in this leaf page. */
6181 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6182 if (rc == 0 && insert_key) {
6183 /* Adjust other cursors pointing to mp */
6184 MDB_cursor *m2, *m3;
6185 MDB_dbi dbi = mc->mc_dbi;
6186 unsigned i = mc->mc_top;
6187 MDB_page *mp = mc->mc_pg[i];
6189 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6190 if (mc->mc_flags & C_SUB)
6191 m3 = &m2->mc_xcursor->mx_cursor;
6194 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6195 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6202 if (rc == MDB_SUCCESS) {
6203 /* Now store the actual data in the child DB. Note that we're
6204 * storing the user data in the keys field, so there are strict
6205 * size limits on dupdata. The actual data fields of the child
6206 * DB are all zero size.
6214 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6215 if (flags & MDB_CURRENT) {
6216 xflags = MDB_CURRENT|MDB_NOSPILL;
6218 mdb_xcursor_init1(mc, leaf);
6219 xflags = (flags & MDB_NODUPDATA) ?
6220 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6222 /* converted, write the original data first */
6224 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6228 /* Adjust other cursors pointing to mp */
6230 unsigned i = mc->mc_top;
6231 MDB_page *mp = mc->mc_pg[i];
6233 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6234 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6235 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6236 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6237 mdb_xcursor_init1(m2, leaf);
6241 /* we've done our job */
6244 ecount = mc->mc_xcursor->mx_db.md_entries;
6245 if (flags & MDB_APPENDDUP)
6246 xflags |= MDB_APPEND;
6247 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6248 if (flags & F_SUBDATA) {
6249 void *db = NODEDATA(leaf);
6250 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6252 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6254 /* Increment count unless we just replaced an existing item. */
6256 mc->mc_db->md_entries++;
6258 /* Invalidate txn if we created an empty sub-DB */
6261 /* If we succeeded and the key didn't exist before,
6262 * make sure the cursor is marked valid.
6264 mc->mc_flags |= C_INITIALIZED;
6267 if (flags & MDB_MULTIPLE) {
6270 /* let caller know how many succeeded, if any */
6271 data[1].mv_size = mcount;
6272 if (mcount < dcount) {
6273 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6274 insert_key = insert_data = 0;
6281 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6284 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6289 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6295 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6296 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6298 if (!(mc->mc_flags & C_INITIALIZED))
6301 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6302 return MDB_NOTFOUND;
6304 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6307 rc = mdb_cursor_touch(mc);
6311 mp = mc->mc_pg[mc->mc_top];
6314 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6316 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6317 if (flags & MDB_NODUPDATA) {
6318 /* mdb_cursor_del0() will subtract the final entry */
6319 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6321 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6322 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6324 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6327 /* If sub-DB still has entries, we're done */
6328 if (mc->mc_xcursor->mx_db.md_entries) {
6329 if (leaf->mn_flags & F_SUBDATA) {
6330 /* update subDB info */
6331 void *db = NODEDATA(leaf);
6332 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6335 /* shrink fake page */
6336 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6337 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6338 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6339 /* fix other sub-DB cursors pointed at this fake page */
6340 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6341 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6342 if (m2->mc_pg[mc->mc_top] == mp &&
6343 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6344 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6347 mc->mc_db->md_entries--;
6348 mc->mc_flags |= C_DEL;
6351 /* otherwise fall thru and delete the sub-DB */
6354 if (leaf->mn_flags & F_SUBDATA) {
6355 /* add all the child DB's pages to the free list */
6356 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6362 /* add overflow pages to free list */
6363 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6367 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6368 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6369 (rc = mdb_ovpage_free(mc, omp)))
6374 return mdb_cursor_del0(mc);
6377 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6381 /** Allocate and initialize new pages for a database.
6382 * @param[in] mc a cursor on the database being added to.
6383 * @param[in] flags flags defining what type of page is being allocated.
6384 * @param[in] num the number of pages to allocate. This is usually 1,
6385 * unless allocating overflow pages for a large record.
6386 * @param[out] mp Address of a page, or NULL on failure.
6387 * @return 0 on success, non-zero on failure.
6390 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6395 if ((rc = mdb_page_alloc(mc, num, &np)))
6397 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6398 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6399 np->mp_flags = flags | P_DIRTY;
6400 np->mp_lower = PAGEHDRSZ;
6401 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6404 mc->mc_db->md_branch_pages++;
6405 else if (IS_LEAF(np))
6406 mc->mc_db->md_leaf_pages++;
6407 else if (IS_OVERFLOW(np)) {
6408 mc->mc_db->md_overflow_pages += num;
6416 /** Calculate the size of a leaf node.
6417 * The size depends on the environment's page size; if a data item
6418 * is too large it will be put onto an overflow page and the node
6419 * size will only include the key and not the data. Sizes are always
6420 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6421 * of the #MDB_node headers.
6422 * @param[in] env The environment handle.
6423 * @param[in] key The key for the node.
6424 * @param[in] data The data for the node.
6425 * @return The number of bytes needed to store the node.
6428 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6432 sz = LEAFSIZE(key, data);
6433 if (sz > env->me_nodemax) {
6434 /* put on overflow page */
6435 sz -= data->mv_size - sizeof(pgno_t);
6438 return EVEN(sz + sizeof(indx_t));
6441 /** Calculate the size of a branch node.
6442 * The size should depend on the environment's page size but since
6443 * we currently don't support spilling large keys onto overflow
6444 * pages, it's simply the size of the #MDB_node header plus the
6445 * size of the key. Sizes are always rounded up to an even number
6446 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6447 * @param[in] env The environment handle.
6448 * @param[in] key The key for the node.
6449 * @return The number of bytes needed to store the node.
6452 mdb_branch_size(MDB_env *env, MDB_val *key)
6457 if (sz > env->me_nodemax) {
6458 /* put on overflow page */
6459 /* not implemented */
6460 /* sz -= key->size - sizeof(pgno_t); */
6463 return sz + sizeof(indx_t);
6466 /** Add a node to the page pointed to by the cursor.
6467 * @param[in] mc The cursor for this operation.
6468 * @param[in] indx The index on the page where the new node should be added.
6469 * @param[in] key The key for the new node.
6470 * @param[in] data The data for the new node, if any.
6471 * @param[in] pgno The page number, if adding a branch node.
6472 * @param[in] flags Flags for the node.
6473 * @return 0 on success, non-zero on failure. Possible errors are:
6475 * <li>ENOMEM - failed to allocate overflow pages for the node.
6476 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6477 * should never happen since all callers already calculate the
6478 * page's free space before calling this function.
6482 mdb_node_add(MDB_cursor *mc, indx_t indx,
6483 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6486 size_t node_size = NODESIZE;
6490 MDB_page *mp = mc->mc_pg[mc->mc_top];
6491 MDB_page *ofp = NULL; /* overflow page */
6494 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6496 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6497 IS_LEAF(mp) ? "leaf" : "branch",
6498 IS_SUBP(mp) ? "sub-" : "",
6499 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6500 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6503 /* Move higher keys up one slot. */
6504 int ksize = mc->mc_db->md_pad, dif;
6505 char *ptr = LEAF2KEY(mp, indx, ksize);
6506 dif = NUMKEYS(mp) - indx;
6508 memmove(ptr+ksize, ptr, dif*ksize);
6509 /* insert new key */
6510 memcpy(ptr, key->mv_data, ksize);
6512 /* Just using these for counting */
6513 mp->mp_lower += sizeof(indx_t);
6514 mp->mp_upper -= ksize - sizeof(indx_t);
6518 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6520 node_size += key->mv_size;
6522 mdb_cassert(mc, data);
6523 if (F_ISSET(flags, F_BIGDATA)) {
6524 /* Data already on overflow page. */
6525 node_size += sizeof(pgno_t);
6526 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6527 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6529 /* Put data on overflow page. */
6530 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6531 data->mv_size, node_size+data->mv_size));
6532 node_size = EVEN(node_size + sizeof(pgno_t));
6533 if ((ssize_t)node_size > room)
6535 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6537 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6541 node_size += data->mv_size;
6544 node_size = EVEN(node_size);
6545 if ((ssize_t)node_size > room)
6549 /* Move higher pointers up one slot. */
6550 for (i = NUMKEYS(mp); i > indx; i--)
6551 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6553 /* Adjust free space offsets. */
6554 ofs = mp->mp_upper - node_size;
6555 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6556 mp->mp_ptrs[indx] = ofs;
6558 mp->mp_lower += sizeof(indx_t);
6560 /* Write the node data. */
6561 node = NODEPTR(mp, indx);
6562 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6563 node->mn_flags = flags;
6565 SETDSZ(node,data->mv_size);
6570 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6573 mdb_cassert(mc, key);
6575 if (F_ISSET(flags, F_BIGDATA))
6576 memcpy(node->mn_data + key->mv_size, data->mv_data,
6578 else if (F_ISSET(flags, MDB_RESERVE))
6579 data->mv_data = node->mn_data + key->mv_size;
6581 memcpy(node->mn_data + key->mv_size, data->mv_data,
6584 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6586 if (F_ISSET(flags, MDB_RESERVE))
6587 data->mv_data = METADATA(ofp);
6589 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6596 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6597 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6598 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6599 DPRINTF(("node size = %"Z"u", node_size));
6600 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6601 return MDB_PAGE_FULL;
6604 /** Delete the specified node from a page.
6605 * @param[in] mc Cursor pointing to the node to delete.
6606 * @param[in] ksize The size of a node. Only used if the page is
6607 * part of a #MDB_DUPFIXED database.
6610 mdb_node_del(MDB_cursor *mc, int ksize)
6612 MDB_page *mp = mc->mc_pg[mc->mc_top];
6613 indx_t indx = mc->mc_ki[mc->mc_top];
6615 indx_t i, j, numkeys, ptr;
6619 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6620 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6621 numkeys = NUMKEYS(mp);
6622 mdb_cassert(mc, indx < numkeys);
6625 int x = numkeys - 1 - indx;
6626 base = LEAF2KEY(mp, indx, ksize);
6628 memmove(base, base + ksize, x * ksize);
6629 mp->mp_lower -= sizeof(indx_t);
6630 mp->mp_upper += ksize - sizeof(indx_t);
6634 node = NODEPTR(mp, indx);
6635 sz = NODESIZE + node->mn_ksize;
6637 if (F_ISSET(node->mn_flags, F_BIGDATA))
6638 sz += sizeof(pgno_t);
6640 sz += NODEDSZ(node);
6644 ptr = mp->mp_ptrs[indx];
6645 for (i = j = 0; i < numkeys; i++) {
6647 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6648 if (mp->mp_ptrs[i] < ptr)
6649 mp->mp_ptrs[j] += sz;
6654 base = (char *)mp + mp->mp_upper;
6655 memmove(base + sz, base, ptr - mp->mp_upper);
6657 mp->mp_lower -= sizeof(indx_t);
6661 /** Compact the main page after deleting a node on a subpage.
6662 * @param[in] mp The main page to operate on.
6663 * @param[in] indx The index of the subpage on the main page.
6666 mdb_node_shrink(MDB_page *mp, indx_t indx)
6672 indx_t i, numkeys, ptr;
6674 node = NODEPTR(mp, indx);
6675 sp = (MDB_page *)NODEDATA(node);
6676 delta = SIZELEFT(sp);
6677 xp = (MDB_page *)((char *)sp + delta);
6679 /* shift subpage upward */
6681 nsize = NUMKEYS(sp) * sp->mp_pad;
6683 return; /* do not make the node uneven-sized */
6684 memmove(METADATA(xp), METADATA(sp), nsize);
6687 numkeys = NUMKEYS(sp);
6688 for (i=numkeys-1; i>=0; i--)
6689 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6691 xp->mp_upper = sp->mp_lower;
6692 xp->mp_lower = sp->mp_lower;
6693 xp->mp_flags = sp->mp_flags;
6694 xp->mp_pad = sp->mp_pad;
6695 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6697 nsize = NODEDSZ(node) - delta;
6698 SETDSZ(node, nsize);
6700 /* shift lower nodes upward */
6701 ptr = mp->mp_ptrs[indx];
6702 numkeys = NUMKEYS(mp);
6703 for (i = 0; i < numkeys; i++) {
6704 if (mp->mp_ptrs[i] <= ptr)
6705 mp->mp_ptrs[i] += delta;
6708 base = (char *)mp + mp->mp_upper;
6709 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6710 mp->mp_upper += delta;
6713 /** Initial setup of a sorted-dups cursor.
6714 * Sorted duplicates are implemented as a sub-database for the given key.
6715 * The duplicate data items are actually keys of the sub-database.
6716 * Operations on the duplicate data items are performed using a sub-cursor
6717 * initialized when the sub-database is first accessed. This function does
6718 * the preliminary setup of the sub-cursor, filling in the fields that
6719 * depend only on the parent DB.
6720 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6723 mdb_xcursor_init0(MDB_cursor *mc)
6725 MDB_xcursor *mx = mc->mc_xcursor;
6727 mx->mx_cursor.mc_xcursor = NULL;
6728 mx->mx_cursor.mc_txn = mc->mc_txn;
6729 mx->mx_cursor.mc_db = &mx->mx_db;
6730 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6731 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6732 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6733 mx->mx_cursor.mc_snum = 0;
6734 mx->mx_cursor.mc_top = 0;
6735 mx->mx_cursor.mc_flags = C_SUB;
6736 mx->mx_dbx.md_name.mv_size = 0;
6737 mx->mx_dbx.md_name.mv_data = NULL;
6738 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6739 mx->mx_dbx.md_dcmp = NULL;
6740 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6743 /** Final setup of a sorted-dups cursor.
6744 * Sets up the fields that depend on the data from the main cursor.
6745 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6746 * @param[in] node The data containing the #MDB_db record for the
6747 * sorted-dup database.
6750 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6752 MDB_xcursor *mx = mc->mc_xcursor;
6754 if (node->mn_flags & F_SUBDATA) {
6755 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6756 mx->mx_cursor.mc_pg[0] = 0;
6757 mx->mx_cursor.mc_snum = 0;
6758 mx->mx_cursor.mc_top = 0;
6759 mx->mx_cursor.mc_flags = C_SUB;
6761 MDB_page *fp = NODEDATA(node);
6762 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6763 mx->mx_db.md_flags = 0;
6764 mx->mx_db.md_depth = 1;
6765 mx->mx_db.md_branch_pages = 0;
6766 mx->mx_db.md_leaf_pages = 1;
6767 mx->mx_db.md_overflow_pages = 0;
6768 mx->mx_db.md_entries = NUMKEYS(fp);
6769 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6770 mx->mx_cursor.mc_snum = 1;
6771 mx->mx_cursor.mc_top = 0;
6772 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6773 mx->mx_cursor.mc_pg[0] = fp;
6774 mx->mx_cursor.mc_ki[0] = 0;
6775 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6776 mx->mx_db.md_flags = MDB_DUPFIXED;
6777 mx->mx_db.md_pad = fp->mp_pad;
6778 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6779 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6782 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6783 mx->mx_db.md_root));
6784 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6785 #if UINT_MAX < SIZE_MAX
6786 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6787 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6791 /** Initialize a cursor for a given transaction and database. */
6793 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6796 mc->mc_backup = NULL;
6799 mc->mc_db = &txn->mt_dbs[dbi];
6800 mc->mc_dbx = &txn->mt_dbxs[dbi];
6801 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6806 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6807 mdb_tassert(txn, mx != NULL);
6808 mc->mc_xcursor = mx;
6809 mdb_xcursor_init0(mc);
6811 mc->mc_xcursor = NULL;
6813 if (*mc->mc_dbflag & DB_STALE) {
6814 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6819 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6822 size_t size = sizeof(MDB_cursor);
6824 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6827 if (txn->mt_flags & MDB_TXN_ERROR)
6830 /* Allow read access to the freelist */
6831 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6834 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6835 size += sizeof(MDB_xcursor);
6837 if ((mc = malloc(size)) != NULL) {
6838 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6839 if (txn->mt_cursors) {
6840 mc->mc_next = txn->mt_cursors[dbi];
6841 txn->mt_cursors[dbi] = mc;
6842 mc->mc_flags |= C_UNTRACK;
6854 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6856 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6859 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6862 if (txn->mt_flags & MDB_TXN_ERROR)
6865 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6869 /* Return the count of duplicate data items for the current key */
6871 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6875 if (mc == NULL || countp == NULL)
6878 if (mc->mc_xcursor == NULL)
6879 return MDB_INCOMPATIBLE;
6881 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6884 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6885 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6888 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6891 *countp = mc->mc_xcursor->mx_db.md_entries;
6897 mdb_cursor_close(MDB_cursor *mc)
6899 if (mc && !mc->mc_backup) {
6900 /* remove from txn, if tracked */
6901 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6902 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6903 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6905 *prev = mc->mc_next;
6912 mdb_cursor_txn(MDB_cursor *mc)
6914 if (!mc) return NULL;
6919 mdb_cursor_dbi(MDB_cursor *mc)
6924 /** Replace the key for a branch node with a new key.
6925 * @param[in] mc Cursor pointing to the node to operate on.
6926 * @param[in] key The new key to use.
6927 * @return 0 on success, non-zero on failure.
6930 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6936 int delta, ksize, oksize;
6937 indx_t ptr, i, numkeys, indx;
6940 indx = mc->mc_ki[mc->mc_top];
6941 mp = mc->mc_pg[mc->mc_top];
6942 node = NODEPTR(mp, indx);
6943 ptr = mp->mp_ptrs[indx];
6947 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6948 k2.mv_data = NODEKEY(node);
6949 k2.mv_size = node->mn_ksize;
6950 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6952 mdb_dkey(&k2, kbuf2),
6958 /* Sizes must be 2-byte aligned. */
6959 ksize = EVEN(key->mv_size);
6960 oksize = EVEN(node->mn_ksize);
6961 delta = ksize - oksize;
6963 /* Shift node contents if EVEN(key length) changed. */
6965 if (delta > 0 && SIZELEFT(mp) < delta) {
6967 /* not enough space left, do a delete and split */
6968 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6969 pgno = NODEPGNO(node);
6970 mdb_node_del(mc, 0);
6971 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6974 numkeys = NUMKEYS(mp);
6975 for (i = 0; i < numkeys; i++) {
6976 if (mp->mp_ptrs[i] <= ptr)
6977 mp->mp_ptrs[i] -= delta;
6980 base = (char *)mp + mp->mp_upper;
6981 len = ptr - mp->mp_upper + NODESIZE;
6982 memmove(base - delta, base, len);
6983 mp->mp_upper -= delta;
6985 node = NODEPTR(mp, indx);
6988 /* But even if no shift was needed, update ksize */
6989 if (node->mn_ksize != key->mv_size)
6990 node->mn_ksize = key->mv_size;
6993 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6999 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7001 /** Move a node from csrc to cdst.
7004 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7011 unsigned short flags;
7015 /* Mark src and dst as dirty. */
7016 if ((rc = mdb_page_touch(csrc)) ||
7017 (rc = mdb_page_touch(cdst)))
7020 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7021 key.mv_size = csrc->mc_db->md_pad;
7022 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7024 data.mv_data = NULL;
7028 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7029 mdb_cassert(csrc, !((size_t)srcnode & 1));
7030 srcpg = NODEPGNO(srcnode);
7031 flags = srcnode->mn_flags;
7032 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7033 unsigned int snum = csrc->mc_snum;
7035 /* must find the lowest key below src */
7036 rc = mdb_page_search_lowest(csrc);
7039 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7040 key.mv_size = csrc->mc_db->md_pad;
7041 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7043 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7044 key.mv_size = NODEKSZ(s2);
7045 key.mv_data = NODEKEY(s2);
7047 csrc->mc_snum = snum--;
7048 csrc->mc_top = snum;
7050 key.mv_size = NODEKSZ(srcnode);
7051 key.mv_data = NODEKEY(srcnode);
7053 data.mv_size = NODEDSZ(srcnode);
7054 data.mv_data = NODEDATA(srcnode);
7056 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7057 unsigned int snum = cdst->mc_snum;
7060 /* must find the lowest key below dst */
7061 rc = mdb_page_search_lowest(cdst);
7064 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
7065 bkey.mv_size = cdst->mc_db->md_pad;
7066 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
7068 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7069 bkey.mv_size = NODEKSZ(s2);
7070 bkey.mv_data = NODEKEY(s2);
7072 cdst->mc_snum = snum--;
7073 cdst->mc_top = snum;
7074 mdb_cursor_copy(cdst, &mn);
7076 rc = mdb_update_key(&mn, &bkey);
7081 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7082 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7083 csrc->mc_ki[csrc->mc_top],
7085 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7086 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7088 /* Add the node to the destination page.
7090 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7091 if (rc != MDB_SUCCESS)
7094 /* Delete the node from the source page.
7096 mdb_node_del(csrc, key.mv_size);
7099 /* Adjust other cursors pointing to mp */
7100 MDB_cursor *m2, *m3;
7101 MDB_dbi dbi = csrc->mc_dbi;
7102 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7104 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7105 if (csrc->mc_flags & C_SUB)
7106 m3 = &m2->mc_xcursor->mx_cursor;
7109 if (m3 == csrc) continue;
7110 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7111 csrc->mc_ki[csrc->mc_top]) {
7112 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7113 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7118 /* Update the parent separators.
7120 if (csrc->mc_ki[csrc->mc_top] == 0) {
7121 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7122 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7123 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7125 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7126 key.mv_size = NODEKSZ(srcnode);
7127 key.mv_data = NODEKEY(srcnode);
7129 DPRINTF(("update separator for source page %"Z"u to [%s]",
7130 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7131 mdb_cursor_copy(csrc, &mn);
7134 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7137 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7139 indx_t ix = csrc->mc_ki[csrc->mc_top];
7140 nullkey.mv_size = 0;
7141 csrc->mc_ki[csrc->mc_top] = 0;
7142 rc = mdb_update_key(csrc, &nullkey);
7143 csrc->mc_ki[csrc->mc_top] = ix;
7144 mdb_cassert(csrc, rc == MDB_SUCCESS);
7148 if (cdst->mc_ki[cdst->mc_top] == 0) {
7149 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7150 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7151 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7153 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7154 key.mv_size = NODEKSZ(srcnode);
7155 key.mv_data = NODEKEY(srcnode);
7157 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7158 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7159 mdb_cursor_copy(cdst, &mn);
7162 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7165 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7167 indx_t ix = cdst->mc_ki[cdst->mc_top];
7168 nullkey.mv_size = 0;
7169 cdst->mc_ki[cdst->mc_top] = 0;
7170 rc = mdb_update_key(cdst, &nullkey);
7171 cdst->mc_ki[cdst->mc_top] = ix;
7172 mdb_cassert(csrc, rc == MDB_SUCCESS);
7179 /** Merge one page into another.
7180 * The nodes from the page pointed to by \b csrc will
7181 * be copied to the page pointed to by \b cdst and then
7182 * the \b csrc page will be freed.
7183 * @param[in] csrc Cursor pointing to the source page.
7184 * @param[in] cdst Cursor pointing to the destination page.
7185 * @return 0 on success, non-zero on failure.
7188 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7196 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7197 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7199 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7200 mdb_cassert(csrc, cdst->mc_snum > 1);
7202 /* Mark dst as dirty. */
7203 if ((rc = mdb_page_touch(cdst)))
7206 /* Move all nodes from src to dst.
7208 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7209 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7210 key.mv_size = csrc->mc_db->md_pad;
7211 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7212 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7213 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7214 if (rc != MDB_SUCCESS)
7216 key.mv_data = (char *)key.mv_data + key.mv_size;
7219 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7220 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7221 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7222 unsigned int snum = csrc->mc_snum;
7224 /* must find the lowest key below src */
7225 rc = mdb_page_search_lowest(csrc);
7228 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7229 key.mv_size = csrc->mc_db->md_pad;
7230 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7232 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7233 key.mv_size = NODEKSZ(s2);
7234 key.mv_data = NODEKEY(s2);
7236 csrc->mc_snum = snum--;
7237 csrc->mc_top = snum;
7239 key.mv_size = srcnode->mn_ksize;
7240 key.mv_data = NODEKEY(srcnode);
7243 data.mv_size = NODEDSZ(srcnode);
7244 data.mv_data = NODEDATA(srcnode);
7245 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7246 if (rc != MDB_SUCCESS)
7251 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7252 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7253 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7255 /* Unlink the src page from parent and add to free list.
7258 mdb_node_del(csrc, 0);
7259 if (csrc->mc_ki[csrc->mc_top] == 0) {
7261 rc = mdb_update_key(csrc, &key);
7269 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7270 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7273 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7274 csrc->mc_db->md_leaf_pages--;
7276 csrc->mc_db->md_branch_pages--;
7278 /* Adjust other cursors pointing to mp */
7279 MDB_cursor *m2, *m3;
7280 MDB_dbi dbi = csrc->mc_dbi;
7281 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7283 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7284 if (csrc->mc_flags & C_SUB)
7285 m3 = &m2->mc_xcursor->mx_cursor;
7288 if (m3 == csrc) continue;
7289 if (m3->mc_snum < csrc->mc_snum) continue;
7290 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7291 m3->mc_pg[csrc->mc_top] = mp;
7292 m3->mc_ki[csrc->mc_top] += nkeys;
7297 unsigned int snum = cdst->mc_snum;
7298 uint16_t depth = cdst->mc_db->md_depth;
7299 mdb_cursor_pop(cdst);
7300 rc = mdb_rebalance(cdst);
7301 /* Did the tree shrink? */
7302 if (depth > cdst->mc_db->md_depth)
7304 cdst->mc_snum = snum;
7305 cdst->mc_top = snum-1;
7310 /** Copy the contents of a cursor.
7311 * @param[in] csrc The cursor to copy from.
7312 * @param[out] cdst The cursor to copy to.
7315 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7319 cdst->mc_txn = csrc->mc_txn;
7320 cdst->mc_dbi = csrc->mc_dbi;
7321 cdst->mc_db = csrc->mc_db;
7322 cdst->mc_dbx = csrc->mc_dbx;
7323 cdst->mc_snum = csrc->mc_snum;
7324 cdst->mc_top = csrc->mc_top;
7325 cdst->mc_flags = csrc->mc_flags;
7327 for (i=0; i<csrc->mc_snum; i++) {
7328 cdst->mc_pg[i] = csrc->mc_pg[i];
7329 cdst->mc_ki[i] = csrc->mc_ki[i];
7333 /** Rebalance the tree after a delete operation.
7334 * @param[in] mc Cursor pointing to the page where rebalancing
7336 * @return 0 on success, non-zero on failure.
7339 mdb_rebalance(MDB_cursor *mc)
7343 unsigned int ptop, minkeys;
7347 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7348 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7349 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7350 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7351 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7353 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7354 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7355 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7356 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7360 if (mc->mc_snum < 2) {
7361 MDB_page *mp = mc->mc_pg[0];
7363 DPUTS("Can't rebalance a subpage, ignoring");
7366 if (NUMKEYS(mp) == 0) {
7367 DPUTS("tree is completely empty");
7368 mc->mc_db->md_root = P_INVALID;
7369 mc->mc_db->md_depth = 0;
7370 mc->mc_db->md_leaf_pages = 0;
7371 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7374 /* Adjust cursors pointing to mp */
7377 mc->mc_flags &= ~C_INITIALIZED;
7379 MDB_cursor *m2, *m3;
7380 MDB_dbi dbi = mc->mc_dbi;
7382 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7383 if (mc->mc_flags & C_SUB)
7384 m3 = &m2->mc_xcursor->mx_cursor;
7387 if (m3->mc_snum < mc->mc_snum) continue;
7388 if (m3->mc_pg[0] == mp) {
7391 m3->mc_flags &= ~C_INITIALIZED;
7395 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7397 DPUTS("collapsing root page!");
7398 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7401 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7402 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7405 mc->mc_db->md_depth--;
7406 mc->mc_db->md_branch_pages--;
7407 mc->mc_ki[0] = mc->mc_ki[1];
7408 for (i = 1; i<mc->mc_db->md_depth; i++) {
7409 mc->mc_pg[i] = mc->mc_pg[i+1];
7410 mc->mc_ki[i] = mc->mc_ki[i+1];
7413 /* Adjust other cursors pointing to mp */
7414 MDB_cursor *m2, *m3;
7415 MDB_dbi dbi = mc->mc_dbi;
7417 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7418 if (mc->mc_flags & C_SUB)
7419 m3 = &m2->mc_xcursor->mx_cursor;
7422 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7423 if (m3->mc_pg[0] == mp) {
7426 for (i=0; i<m3->mc_snum; i++) {
7427 m3->mc_pg[i] = m3->mc_pg[i+1];
7428 m3->mc_ki[i] = m3->mc_ki[i+1];
7434 DPUTS("root page doesn't need rebalancing");
7438 /* The parent (branch page) must have at least 2 pointers,
7439 * otherwise the tree is invalid.
7441 ptop = mc->mc_top-1;
7442 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7444 /* Leaf page fill factor is below the threshold.
7445 * Try to move keys from left or right neighbor, or
7446 * merge with a neighbor page.
7451 mdb_cursor_copy(mc, &mn);
7452 mn.mc_xcursor = NULL;
7454 oldki = mc->mc_ki[mc->mc_top];
7455 if (mc->mc_ki[ptop] == 0) {
7456 /* We're the leftmost leaf in our parent.
7458 DPUTS("reading right neighbor");
7460 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7461 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7464 mn.mc_ki[mn.mc_top] = 0;
7465 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7467 /* There is at least one neighbor to the left.
7469 DPUTS("reading left neighbor");
7471 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7472 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7475 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7476 mc->mc_ki[mc->mc_top] = 0;
7479 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7480 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7481 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7483 /* If the neighbor page is above threshold and has enough keys,
7484 * move one key from it. Otherwise we should try to merge them.
7485 * (A branch page must never have less than 2 keys.)
7487 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7488 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7489 rc = mdb_node_move(&mn, mc);
7490 if (mc->mc_ki[ptop]) {
7494 if (mc->mc_ki[ptop] == 0) {
7495 rc = mdb_page_merge(&mn, mc);
7497 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7498 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7499 rc = mdb_page_merge(mc, &mn);
7500 mdb_cursor_copy(&mn, mc);
7502 mc->mc_flags &= ~C_EOF;
7504 mc->mc_ki[mc->mc_top] = oldki;
7508 /** Complete a delete operation started by #mdb_cursor_del(). */
7510 mdb_cursor_del0(MDB_cursor *mc)
7517 ki = mc->mc_ki[mc->mc_top];
7518 mdb_node_del(mc, mc->mc_db->md_pad);
7519 mc->mc_db->md_entries--;
7520 rc = mdb_rebalance(mc);
7522 if (rc == MDB_SUCCESS) {
7523 MDB_cursor *m2, *m3;
7524 MDB_dbi dbi = mc->mc_dbi;
7526 mp = mc->mc_pg[mc->mc_top];
7527 nkeys = NUMKEYS(mp);
7529 /* if mc points past last node in page, find next sibling */
7530 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7531 rc = mdb_cursor_sibling(mc, 1);
7532 if (rc == MDB_NOTFOUND)
7536 /* Adjust other cursors pointing to mp */
7537 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7538 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7539 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7541 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7543 if (m3->mc_pg[mc->mc_top] == mp) {
7544 if (m3->mc_ki[mc->mc_top] >= ki) {
7545 m3->mc_flags |= C_DEL;
7546 if (m3->mc_ki[mc->mc_top] > ki)
7547 m3->mc_ki[mc->mc_top]--;
7549 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7550 rc = mdb_cursor_sibling(m3, 1);
7551 if (rc == MDB_NOTFOUND)
7556 mc->mc_flags |= C_DEL;
7560 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7565 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7566 MDB_val *key, MDB_val *data)
7568 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7571 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7572 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7574 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7575 /* must ignore any data */
7579 return mdb_del0(txn, dbi, key, data, 0);
7583 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7584 MDB_val *key, MDB_val *data, unsigned flags)
7589 MDB_val rdata, *xdata;
7593 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7595 mdb_cursor_init(&mc, txn, dbi, &mx);
7604 flags |= MDB_NODUPDATA;
7606 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7608 /* let mdb_page_split know about this cursor if needed:
7609 * delete will trigger a rebalance; if it needs to move
7610 * a node from one page to another, it will have to
7611 * update the parent's separator key(s). If the new sepkey
7612 * is larger than the current one, the parent page may
7613 * run out of space, triggering a split. We need this
7614 * cursor to be consistent until the end of the rebalance.
7616 mc.mc_flags |= C_UNTRACK;
7617 mc.mc_next = txn->mt_cursors[dbi];
7618 txn->mt_cursors[dbi] = &mc;
7619 rc = mdb_cursor_del(&mc, flags);
7620 txn->mt_cursors[dbi] = mc.mc_next;
7625 /** Split a page and insert a new node.
7626 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7627 * The cursor will be updated to point to the actual page and index where
7628 * the node got inserted after the split.
7629 * @param[in] newkey The key for the newly inserted node.
7630 * @param[in] newdata The data for the newly inserted node.
7631 * @param[in] newpgno The page number, if the new node is a branch node.
7632 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7633 * @return 0 on success, non-zero on failure.
7636 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7637 unsigned int nflags)
7640 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7643 int i, j, split_indx, nkeys, pmax;
7644 MDB_env *env = mc->mc_txn->mt_env;
7646 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7647 MDB_page *copy = NULL;
7648 MDB_page *mp, *rp, *pp;
7653 mp = mc->mc_pg[mc->mc_top];
7654 newindx = mc->mc_ki[mc->mc_top];
7655 nkeys = NUMKEYS(mp);
7657 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7658 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7659 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7661 /* Create a right sibling. */
7662 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7664 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7666 if (mc->mc_snum < 2) {
7667 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7669 /* shift current top to make room for new parent */
7670 mc->mc_pg[1] = mc->mc_pg[0];
7671 mc->mc_ki[1] = mc->mc_ki[0];
7674 mc->mc_db->md_root = pp->mp_pgno;
7675 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7676 mc->mc_db->md_depth++;
7679 /* Add left (implicit) pointer. */
7680 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7681 /* undo the pre-push */
7682 mc->mc_pg[0] = mc->mc_pg[1];
7683 mc->mc_ki[0] = mc->mc_ki[1];
7684 mc->mc_db->md_root = mp->mp_pgno;
7685 mc->mc_db->md_depth--;
7692 ptop = mc->mc_top-1;
7693 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7696 mc->mc_flags |= C_SPLITTING;
7697 mdb_cursor_copy(mc, &mn);
7698 mn.mc_pg[mn.mc_top] = rp;
7699 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7701 if (nflags & MDB_APPEND) {
7702 mn.mc_ki[mn.mc_top] = 0;
7704 split_indx = newindx;
7708 split_indx = (nkeys+1) / 2;
7713 unsigned int lsize, rsize, ksize;
7714 /* Move half of the keys to the right sibling */
7715 x = mc->mc_ki[mc->mc_top] - split_indx;
7716 ksize = mc->mc_db->md_pad;
7717 split = LEAF2KEY(mp, split_indx, ksize);
7718 rsize = (nkeys - split_indx) * ksize;
7719 lsize = (nkeys - split_indx) * sizeof(indx_t);
7720 mp->mp_lower -= lsize;
7721 rp->mp_lower += lsize;
7722 mp->mp_upper += rsize - lsize;
7723 rp->mp_upper -= rsize - lsize;
7724 sepkey.mv_size = ksize;
7725 if (newindx == split_indx) {
7726 sepkey.mv_data = newkey->mv_data;
7728 sepkey.mv_data = split;
7731 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7732 memcpy(rp->mp_ptrs, split, rsize);
7733 sepkey.mv_data = rp->mp_ptrs;
7734 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7735 memcpy(ins, newkey->mv_data, ksize);
7736 mp->mp_lower += sizeof(indx_t);
7737 mp->mp_upper -= ksize - sizeof(indx_t);
7740 memcpy(rp->mp_ptrs, split, x * ksize);
7741 ins = LEAF2KEY(rp, x, ksize);
7742 memcpy(ins, newkey->mv_data, ksize);
7743 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7744 rp->mp_lower += sizeof(indx_t);
7745 rp->mp_upper -= ksize - sizeof(indx_t);
7746 mc->mc_ki[mc->mc_top] = x;
7747 mc->mc_pg[mc->mc_top] = rp;
7750 int psize, nsize, k;
7751 /* Maximum free space in an empty page */
7752 pmax = env->me_psize - PAGEHDRSZ;
7754 nsize = mdb_leaf_size(env, newkey, newdata);
7756 nsize = mdb_branch_size(env, newkey);
7757 nsize = EVEN(nsize);
7759 /* grab a page to hold a temporary copy */
7760 copy = mdb_page_malloc(mc->mc_txn, 1);
7765 copy->mp_pgno = mp->mp_pgno;
7766 copy->mp_flags = mp->mp_flags;
7767 copy->mp_lower = PAGEHDRSZ;
7768 copy->mp_upper = env->me_psize;
7770 /* prepare to insert */
7771 for (i=0, j=0; i<nkeys; i++) {
7773 copy->mp_ptrs[j++] = 0;
7775 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7778 /* When items are relatively large the split point needs
7779 * to be checked, because being off-by-one will make the
7780 * difference between success or failure in mdb_node_add.
7782 * It's also relevant if a page happens to be laid out
7783 * such that one half of its nodes are all "small" and
7784 * the other half of its nodes are "large." If the new
7785 * item is also "large" and falls on the half with
7786 * "large" nodes, it also may not fit.
7788 * As a final tweak, if the new item goes on the last
7789 * spot on the page (and thus, onto the new page), bias
7790 * the split so the new page is emptier than the old page.
7791 * This yields better packing during sequential inserts.
7793 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7794 /* Find split point */
7796 if (newindx <= split_indx || newindx >= nkeys) {
7798 k = newindx >= nkeys ? nkeys : split_indx+2;
7803 for (; i!=k; i+=j) {
7808 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7809 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7811 if (F_ISSET(node->mn_flags, F_BIGDATA))
7812 psize += sizeof(pgno_t);
7814 psize += NODEDSZ(node);
7816 psize = EVEN(psize);
7818 if (psize > pmax || i == k-j) {
7819 split_indx = i + (j<0);
7824 if (split_indx == newindx) {
7825 sepkey.mv_size = newkey->mv_size;
7826 sepkey.mv_data = newkey->mv_data;
7828 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7829 sepkey.mv_size = node->mn_ksize;
7830 sepkey.mv_data = NODEKEY(node);
7835 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7837 /* Copy separator key to the parent.
7839 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7843 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7848 if (mn.mc_snum == mc->mc_snum) {
7849 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7850 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7851 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7852 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7857 /* Right page might now have changed parent.
7858 * Check if left page also changed parent.
7860 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7861 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7862 for (i=0; i<ptop; i++) {
7863 mc->mc_pg[i] = mn.mc_pg[i];
7864 mc->mc_ki[i] = mn.mc_ki[i];
7866 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7867 if (mn.mc_ki[ptop]) {
7868 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7870 /* find right page's left sibling */
7871 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7872 mdb_cursor_sibling(mc, 0);
7877 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7880 mc->mc_flags ^= C_SPLITTING;
7881 if (rc != MDB_SUCCESS) {
7884 if (nflags & MDB_APPEND) {
7885 mc->mc_pg[mc->mc_top] = rp;
7886 mc->mc_ki[mc->mc_top] = 0;
7887 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7890 for (i=0; i<mc->mc_top; i++)
7891 mc->mc_ki[i] = mn.mc_ki[i];
7892 } else if (!IS_LEAF2(mp)) {
7894 mc->mc_pg[mc->mc_top] = rp;
7899 rkey.mv_data = newkey->mv_data;
7900 rkey.mv_size = newkey->mv_size;
7906 /* Update index for the new key. */
7907 mc->mc_ki[mc->mc_top] = j;
7909 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7910 rkey.mv_data = NODEKEY(node);
7911 rkey.mv_size = node->mn_ksize;
7913 xdata.mv_data = NODEDATA(node);
7914 xdata.mv_size = NODEDSZ(node);
7917 pgno = NODEPGNO(node);
7918 flags = node->mn_flags;
7921 if (!IS_LEAF(mp) && j == 0) {
7922 /* First branch index doesn't need key data. */
7926 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7932 mc->mc_pg[mc->mc_top] = copy;
7937 } while (i != split_indx);
7939 nkeys = NUMKEYS(copy);
7940 for (i=0; i<nkeys; i++)
7941 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7942 mp->mp_lower = copy->mp_lower;
7943 mp->mp_upper = copy->mp_upper;
7944 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7945 env->me_psize - copy->mp_upper);
7947 /* reset back to original page */
7948 if (newindx < split_indx) {
7949 mc->mc_pg[mc->mc_top] = mp;
7950 if (nflags & MDB_RESERVE) {
7951 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7952 if (!(node->mn_flags & F_BIGDATA))
7953 newdata->mv_data = NODEDATA(node);
7956 mc->mc_pg[mc->mc_top] = rp;
7958 /* Make sure mc_ki is still valid.
7960 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7961 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7962 for (i=0; i<=ptop; i++) {
7963 mc->mc_pg[i] = mn.mc_pg[i];
7964 mc->mc_ki[i] = mn.mc_ki[i];
7971 /* Adjust other cursors pointing to mp */
7972 MDB_cursor *m2, *m3;
7973 MDB_dbi dbi = mc->mc_dbi;
7974 int fixup = NUMKEYS(mp);
7976 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7977 if (mc->mc_flags & C_SUB)
7978 m3 = &m2->mc_xcursor->mx_cursor;
7983 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7985 if (m3->mc_flags & C_SPLITTING)
7990 for (k=m3->mc_top; k>=0; k--) {
7991 m3->mc_ki[k+1] = m3->mc_ki[k];
7992 m3->mc_pg[k+1] = m3->mc_pg[k];
7994 if (m3->mc_ki[0] >= split_indx) {
7999 m3->mc_pg[0] = mc->mc_pg[0];
8003 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8004 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8005 m3->mc_ki[mc->mc_top]++;
8006 if (m3->mc_ki[mc->mc_top] >= fixup) {
8007 m3->mc_pg[mc->mc_top] = rp;
8008 m3->mc_ki[mc->mc_top] -= fixup;
8009 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8011 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8012 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8017 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8020 if (copy) /* tmp page */
8021 mdb_page_free(env, copy);
8023 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8028 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8029 MDB_val *key, MDB_val *data, unsigned int flags)
8034 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8037 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8040 mdb_cursor_init(&mc, txn, dbi, &mx);
8041 return mdb_cursor_put(&mc, key, data, flags);
8045 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8047 if ((flag & CHANGEABLE) != flag)
8050 env->me_flags |= flag;
8052 env->me_flags &= ~flag;
8057 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8062 *arg = env->me_flags;
8067 mdb_env_set_userctx(MDB_env *env, void *ctx)
8071 env->me_userctx = ctx;
8076 mdb_env_get_userctx(MDB_env *env)
8078 return env ? env->me_userctx : NULL;
8082 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8087 env->me_assert_func = func;
8093 mdb_env_get_path(MDB_env *env, const char **arg)
8098 *arg = env->me_path;
8103 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8112 /** Common code for #mdb_stat() and #mdb_env_stat().
8113 * @param[in] env the environment to operate in.
8114 * @param[in] db the #MDB_db record containing the stats to return.
8115 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8116 * @return 0, this function always succeeds.
8119 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8121 arg->ms_psize = env->me_psize;
8122 arg->ms_depth = db->md_depth;
8123 arg->ms_branch_pages = db->md_branch_pages;
8124 arg->ms_leaf_pages = db->md_leaf_pages;
8125 arg->ms_overflow_pages = db->md_overflow_pages;
8126 arg->ms_entries = db->md_entries;
8131 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8135 if (env == NULL || arg == NULL)
8138 toggle = mdb_env_pick_meta(env);
8140 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8144 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8148 if (env == NULL || arg == NULL)
8151 toggle = mdb_env_pick_meta(env);
8152 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
8153 arg->me_mapsize = env->me_mapsize;
8154 arg->me_maxreaders = env->me_maxreaders;
8156 /* me_numreaders may be zero if this process never used any readers. Use
8157 * the shared numreader count if it exists.
8159 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8161 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8162 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8166 /** Set the default comparison functions for a database.
8167 * Called immediately after a database is opened to set the defaults.
8168 * The user can then override them with #mdb_set_compare() or
8169 * #mdb_set_dupsort().
8170 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8171 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8174 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8176 uint16_t f = txn->mt_dbs[dbi].md_flags;
8178 txn->mt_dbxs[dbi].md_cmp =
8179 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8180 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8182 txn->mt_dbxs[dbi].md_dcmp =
8183 !(f & MDB_DUPSORT) ? 0 :
8184 ((f & MDB_INTEGERDUP)
8185 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8186 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8189 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8194 int rc, dbflag, exact;
8195 unsigned int unused = 0;
8198 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8199 mdb_default_cmp(txn, FREE_DBI);
8202 if ((flags & VALID_FLAGS) != flags)
8204 if (txn->mt_flags & MDB_TXN_ERROR)
8210 if (flags & PERSISTENT_FLAGS) {
8211 uint16_t f2 = flags & PERSISTENT_FLAGS;
8212 /* make sure flag changes get committed */
8213 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8214 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8215 txn->mt_flags |= MDB_TXN_DIRTY;
8218 mdb_default_cmp(txn, MAIN_DBI);
8222 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8223 mdb_default_cmp(txn, MAIN_DBI);
8226 /* Is the DB already open? */
8228 for (i=2; i<txn->mt_numdbs; i++) {
8229 if (!txn->mt_dbxs[i].md_name.mv_size) {
8230 /* Remember this free slot */
8231 if (!unused) unused = i;
8234 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8235 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8241 /* If no free slot and max hit, fail */
8242 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8243 return MDB_DBS_FULL;
8245 /* Cannot mix named databases with some mainDB flags */
8246 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8247 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8249 /* Find the DB info */
8250 dbflag = DB_NEW|DB_VALID;
8253 key.mv_data = (void *)name;
8254 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8255 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8256 if (rc == MDB_SUCCESS) {
8257 /* make sure this is actually a DB */
8258 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8259 if (!(node->mn_flags & F_SUBDATA))
8260 return MDB_INCOMPATIBLE;
8261 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8262 /* Create if requested */
8264 data.mv_size = sizeof(MDB_db);
8265 data.mv_data = &dummy;
8266 memset(&dummy, 0, sizeof(dummy));
8267 dummy.md_root = P_INVALID;
8268 dummy.md_flags = flags & PERSISTENT_FLAGS;
8269 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8273 /* OK, got info, add to table */
8274 if (rc == MDB_SUCCESS) {
8275 unsigned int slot = unused ? unused : txn->mt_numdbs;
8276 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8277 txn->mt_dbxs[slot].md_name.mv_size = len;
8278 txn->mt_dbxs[slot].md_rel = NULL;
8279 txn->mt_dbflags[slot] = dbflag;
8280 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8282 mdb_default_cmp(txn, slot);
8291 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8293 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8296 if (txn->mt_flags & MDB_TXN_ERROR)
8299 if (txn->mt_dbflags[dbi] & DB_STALE) {
8302 /* Stale, must read the DB's root. cursor_init does it for us. */
8303 mdb_cursor_init(&mc, txn, dbi, &mx);
8305 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8308 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8311 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8313 ptr = env->me_dbxs[dbi].md_name.mv_data;
8314 env->me_dbxs[dbi].md_name.mv_data = NULL;
8315 env->me_dbxs[dbi].md_name.mv_size = 0;
8316 env->me_dbflags[dbi] = 0;
8320 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8322 /* We could return the flags for the FREE_DBI too but what's the point? */
8323 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8325 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8329 /** Add all the DB's pages to the free list.
8330 * @param[in] mc Cursor on the DB to free.
8331 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8332 * @return 0 on success, non-zero on failure.
8335 mdb_drop0(MDB_cursor *mc, int subs)
8339 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8340 if (rc == MDB_SUCCESS) {
8341 MDB_txn *txn = mc->mc_txn;
8346 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8347 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8350 mdb_cursor_copy(mc, &mx);
8351 while (mc->mc_snum > 0) {
8352 MDB_page *mp = mc->mc_pg[mc->mc_top];
8353 unsigned n = NUMKEYS(mp);
8355 for (i=0; i<n; i++) {
8356 ni = NODEPTR(mp, i);
8357 if (ni->mn_flags & F_BIGDATA) {
8360 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8361 rc = mdb_page_get(txn, pg, &omp, NULL);
8364 mdb_cassert(mc, IS_OVERFLOW(omp));
8365 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8369 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8370 mdb_xcursor_init1(mc, ni);
8371 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8377 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8379 for (i=0; i<n; i++) {
8381 ni = NODEPTR(mp, i);
8384 mdb_midl_xappend(txn->mt_free_pgs, pg);
8389 mc->mc_ki[mc->mc_top] = i;
8390 rc = mdb_cursor_sibling(mc, 1);
8392 if (rc != MDB_NOTFOUND)
8394 /* no more siblings, go back to beginning
8395 * of previous level.
8399 for (i=1; i<mc->mc_snum; i++) {
8401 mc->mc_pg[i] = mx.mc_pg[i];
8406 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8409 txn->mt_flags |= MDB_TXN_ERROR;
8410 } else if (rc == MDB_NOTFOUND) {
8416 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8418 MDB_cursor *mc, *m2;
8421 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8424 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8427 rc = mdb_cursor_open(txn, dbi, &mc);
8431 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8432 /* Invalidate the dropped DB's cursors */
8433 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8434 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8438 /* Can't delete the main DB */
8439 if (del && dbi > MAIN_DBI) {
8440 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
8442 txn->mt_dbflags[dbi] = DB_STALE;
8443 mdb_dbi_close(txn->mt_env, dbi);
8445 txn->mt_flags |= MDB_TXN_ERROR;
8448 /* reset the DB record, mark it dirty */
8449 txn->mt_dbflags[dbi] |= DB_DIRTY;
8450 txn->mt_dbs[dbi].md_depth = 0;
8451 txn->mt_dbs[dbi].md_branch_pages = 0;
8452 txn->mt_dbs[dbi].md_leaf_pages = 0;
8453 txn->mt_dbs[dbi].md_overflow_pages = 0;
8454 txn->mt_dbs[dbi].md_entries = 0;
8455 txn->mt_dbs[dbi].md_root = P_INVALID;
8457 txn->mt_flags |= MDB_TXN_DIRTY;
8460 mdb_cursor_close(mc);
8464 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8466 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8469 txn->mt_dbxs[dbi].md_cmp = cmp;
8473 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8475 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8478 txn->mt_dbxs[dbi].md_dcmp = cmp;
8482 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8484 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8487 txn->mt_dbxs[dbi].md_rel = rel;
8491 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8493 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8496 txn->mt_dbxs[dbi].md_relctx = ctx;
8500 int mdb_env_get_maxkeysize(MDB_env *env)
8502 return ENV_MAXKEY(env);
8505 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8507 unsigned int i, rdrs;
8510 int rc = 0, first = 1;
8514 if (!env->me_txns) {
8515 return func("(no reader locks)\n", ctx);
8517 rdrs = env->me_txns->mti_numreaders;
8518 mr = env->me_txns->mti_readers;
8519 for (i=0; i<rdrs; i++) {
8521 txnid_t txnid = mr[i].mr_txnid;
8522 sprintf(buf, txnid == (txnid_t)-1 ?
8523 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
8524 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
8527 rc = func(" pid thread txnid\n", ctx);
8531 rc = func(buf, ctx);
8537 rc = func("(no active readers)\n", ctx);
8542 /** Insert pid into list if not already present.
8543 * return -1 if already present.
8545 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8547 /* binary search of pid in list */
8549 unsigned cursor = 1;
8551 unsigned n = ids[0];
8554 unsigned pivot = n >> 1;
8555 cursor = base + pivot + 1;
8556 val = pid - ids[cursor];
8561 } else if ( val > 0 ) {
8566 /* found, so it's a duplicate */
8575 for (n = ids[0]; n > cursor; n--)
8581 int mdb_reader_check(MDB_env *env, int *dead)
8583 unsigned int i, j, rdrs;
8585 MDB_PID_T *pids, pid;
8594 rdrs = env->me_txns->mti_numreaders;
8595 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8599 mr = env->me_txns->mti_readers;
8600 for (i=0; i<rdrs; i++) {
8601 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8603 if (mdb_pid_insert(pids, pid) == 0) {
8604 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8606 /* Recheck, a new process may have reused pid */
8607 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8608 for (j=i; j<rdrs; j++)
8609 if (mr[j].mr_pid == pid) {
8610 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8611 (unsigned) pid, mr[j].mr_txnid));
8616 UNLOCK_MUTEX_R(env);