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
79 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
80 #include <netinet/in.h>
81 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
84 #if defined(__APPLE__) || defined (BSD)
85 # define MDB_USE_POSIX_SEM 1
86 # define MDB_FDATASYNC fsync
87 #elif defined(ANDROID)
88 # define MDB_FDATASYNC fsync
93 #ifdef MDB_USE_POSIX_SEM
94 # define MDB_USE_HASH 1
95 #include <semaphore.h>
100 #include <valgrind/memcheck.h>
101 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
102 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
103 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
104 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
105 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
107 #define VGMEMP_CREATE(h,r,z)
108 #define VGMEMP_ALLOC(h,a,s)
109 #define VGMEMP_FREE(h,a)
110 #define VGMEMP_DESTROY(h)
111 #define VGMEMP_DEFINED(a,s)
115 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
116 /* Solaris just defines one or the other */
117 # define LITTLE_ENDIAN 1234
118 # define BIG_ENDIAN 4321
119 # ifdef _LITTLE_ENDIAN
120 # define BYTE_ORDER LITTLE_ENDIAN
122 # define BYTE_ORDER BIG_ENDIAN
125 # define BYTE_ORDER __BYTE_ORDER
129 #ifndef LITTLE_ENDIAN
130 #define LITTLE_ENDIAN __LITTLE_ENDIAN
133 #define BIG_ENDIAN __BIG_ENDIAN
136 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
137 #define MISALIGNED_OK 1
143 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
144 # error "Unknown or unsupported endianness (BYTE_ORDER)"
145 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
146 # error "Two's complement, reasonably sized integer types, please"
149 /** @defgroup internal MDB Internals
152 /** @defgroup compat Windows Compatibility Macros
153 * A bunch of macros to minimize the amount of platform-specific ifdefs
154 * needed throughout the rest of the code. When the features this library
155 * needs are similar enough to POSIX to be hidden in a one-or-two line
156 * replacement, this macro approach is used.
160 #define MDB_USE_HASH 1
161 #define MDB_PIDLOCK 0
162 #define pthread_t DWORD
163 #define pthread_mutex_t HANDLE
164 #define pthread_key_t DWORD
165 #define pthread_self() GetCurrentThreadId()
166 #define pthread_key_create(x,y) \
167 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
168 #define pthread_key_delete(x) TlsFree(x)
169 #define pthread_getspecific(x) TlsGetValue(x)
170 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
171 #define pthread_mutex_unlock(x) ReleaseMutex(x)
172 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
173 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
174 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
175 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
176 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
177 #define getpid() GetCurrentProcessId()
178 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
179 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
180 #define ErrCode() GetLastError()
181 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
182 #define close(fd) (CloseHandle(fd) ? 0 : -1)
183 #define munmap(ptr,len) UnmapViewOfFile(ptr)
184 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
185 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
187 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
192 #define Z "z" /**< printf format modifier for size_t */
194 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
195 #define MDB_PIDLOCK 1
197 #ifdef MDB_USE_POSIX_SEM
199 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
200 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
201 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
202 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
205 mdb_sem_wait(sem_t *sem)
208 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
213 /** Lock the reader mutex.
215 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
216 /** Unlock the reader mutex.
218 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
220 /** Lock the writer mutex.
221 * Only a single write transaction is allowed at a time. Other writers
222 * will block waiting for this mutex.
224 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
225 /** Unlock the writer mutex.
227 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
228 #endif /* MDB_USE_POSIX_SEM */
230 /** Get the error code for the last failed system function.
232 #define ErrCode() errno
234 /** An abstraction for a file handle.
235 * On POSIX systems file handles are small integers. On Windows
236 * they're opaque pointers.
240 /** A value for an invalid file handle.
241 * Mainly used to initialize file variables and signify that they are
244 #define INVALID_HANDLE_VALUE (-1)
246 /** Get the size of a memory page for the system.
247 * This is the basic size that the platform's memory manager uses, and is
248 * fundamental to the use of memory-mapped files.
250 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
253 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
256 #define MNAME_LEN (sizeof(pthread_mutex_t))
262 /** A flag for opening a file and requesting synchronous data writes.
263 * This is only used when writing a meta page. It's not strictly needed;
264 * we could just do a normal write and then immediately perform a flush.
265 * But if this flag is available it saves us an extra system call.
267 * @note If O_DSYNC is undefined but exists in /usr/include,
268 * preferably set some compiler flag to get the definition.
269 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
272 # define MDB_DSYNC O_DSYNC
276 /** Function for flushing the data of a file. Define this to fsync
277 * if fdatasync() is not supported.
279 #ifndef MDB_FDATASYNC
280 # define MDB_FDATASYNC fdatasync
284 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
295 /** A page number in the database.
296 * Note that 64 bit page numbers are overkill, since pages themselves
297 * already represent 12-13 bits of addressable memory, and the OS will
298 * always limit applications to a maximum of 63 bits of address space.
300 * @note In the #MDB_node structure, we only store 48 bits of this value,
301 * which thus limits us to only 60 bits of addressable data.
303 typedef MDB_ID pgno_t;
305 /** A transaction ID.
306 * See struct MDB_txn.mt_txnid for details.
308 typedef MDB_ID txnid_t;
310 /** @defgroup debug Debug Macros
314 /** Enable debug output. Needs variable argument macros (a C99 feature).
315 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
316 * read from and written to the database (used for free space management).
322 static int mdb_debug;
323 static txnid_t mdb_debug_start;
325 /** Print a debug message with printf formatting.
326 * Requires double parenthesis around 2 or more args.
328 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
329 # define DPRINTF0(fmt, ...) \
330 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
332 # define DPRINTF(args) ((void) 0)
334 /** Print a debug string.
335 * The string is printed literally, with no format processing.
337 #define DPUTS(arg) DPRINTF(("%s", arg))
338 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
340 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
343 /** @brief The maximum size of a database page.
345 * This is 32k, since it must fit in #MDB_page.#mp_upper.
347 * LMDB will use database pages < OS pages if needed.
348 * That causes more I/O in write transactions: The OS must
349 * know (read) the whole page before writing a partial page.
351 * Note that we don't currently support Huge pages. On Linux,
352 * regular data files cannot use Huge pages, and in general
353 * Huge pages aren't actually pageable. We rely on the OS
354 * demand-pager to read our data and page it out when memory
355 * pressure from other processes is high. So until OSs have
356 * actual paging support for Huge pages, they're not viable.
358 #define MAX_PAGESIZE 0x8000
360 /** The minimum number of keys required in a database page.
361 * Setting this to a larger value will place a smaller bound on the
362 * maximum size of a data item. Data items larger than this size will
363 * be pushed into overflow pages instead of being stored directly in
364 * the B-tree node. This value used to default to 4. With a page size
365 * of 4096 bytes that meant that any item larger than 1024 bytes would
366 * go into an overflow page. That also meant that on average 2-3KB of
367 * each overflow page was wasted space. The value cannot be lower than
368 * 2 because then there would no longer be a tree structure. With this
369 * value, items larger than 2KB will go into overflow pages, and on
370 * average only 1KB will be wasted.
372 #define MDB_MINKEYS 2
374 /** A stamp that identifies a file as an MDB file.
375 * There's nothing special about this value other than that it is easily
376 * recognizable, and it will reflect any byte order mismatches.
378 #define MDB_MAGIC 0xBEEFC0DE
380 /** The version number for a database's datafile format. */
381 #define MDB_DATA_VERSION 1
382 /** The version number for a database's lockfile format. */
383 #define MDB_LOCK_VERSION 1
385 /** @brief The max size of a key we can write, or 0 for dynamic max.
387 * Define this as 0 to compute the max from the page size. 511
388 * is default for backwards compat: liblmdb <= 0.9.10 can break
389 * when modifying a DB with keys/dupsort data bigger than its max.
391 * Data items in an #MDB_DUPSORT database are also limited to
392 * this size, since they're actually keys of a sub-DB. Keys and
393 * #MDB_DUPSORT data items must fit on a node in a regular page.
395 #ifndef MDB_MAXKEYSIZE
396 #define MDB_MAXKEYSIZE 511
399 /** The maximum size of a key we can write to the environment. */
401 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
403 #define ENV_MAXKEY(env) ((env)->me_maxkey)
406 /** @brief The maximum size of a data item.
408 * We only store a 32 bit value for node sizes.
410 #define MAXDATASIZE 0xffffffffUL
413 /** Key size which fits in a #DKBUF.
416 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
419 * This is used for printing a hex dump of a key's contents.
421 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
422 /** Display a key in hex.
424 * Invoke a function to display a key in hex.
426 #define DKEY(x) mdb_dkey(x, kbuf)
432 /** An invalid page number.
433 * Mainly used to denote an empty tree.
435 #define P_INVALID (~(pgno_t)0)
437 /** Test if the flags \b f are set in a flag word \b w. */
438 #define F_ISSET(w, f) (((w) & (f)) == (f))
440 /** Round \b n up to an even number. */
441 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
443 /** Used for offsets within a single page.
444 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
447 typedef uint16_t indx_t;
449 /** Default size of memory map.
450 * This is certainly too small for any actual applications. Apps should always set
451 * the size explicitly using #mdb_env_set_mapsize().
453 #define DEFAULT_MAPSIZE 1048576
455 /** @defgroup readers Reader Lock Table
456 * Readers don't acquire any locks for their data access. Instead, they
457 * simply record their transaction ID in the reader table. The reader
458 * mutex is needed just to find an empty slot in the reader table. The
459 * slot's address is saved in thread-specific data so that subsequent read
460 * transactions started by the same thread need no further locking to proceed.
462 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
464 * No reader table is used if the database is on a read-only filesystem, or
465 * if #MDB_NOLOCK is set.
467 * Since the database uses multi-version concurrency control, readers don't
468 * actually need any locking. This table is used to keep track of which
469 * readers are using data from which old transactions, so that we'll know
470 * when a particular old transaction is no longer in use. Old transactions
471 * that have discarded any data pages can then have those pages reclaimed
472 * for use by a later write transaction.
474 * The lock table is constructed such that reader slots are aligned with the
475 * processor's cache line size. Any slot is only ever used by one thread.
476 * This alignment guarantees that there will be no contention or cache
477 * thrashing as threads update their own slot info, and also eliminates
478 * any need for locking when accessing a slot.
480 * A writer thread will scan every slot in the table to determine the oldest
481 * outstanding reader transaction. Any freed pages older than this will be
482 * reclaimed by the writer. The writer doesn't use any locks when scanning
483 * this table. This means that there's no guarantee that the writer will
484 * see the most up-to-date reader info, but that's not required for correct
485 * operation - all we need is to know the upper bound on the oldest reader,
486 * we don't care at all about the newest reader. So the only consequence of
487 * reading stale information here is that old pages might hang around a
488 * while longer before being reclaimed. That's actually good anyway, because
489 * the longer we delay reclaiming old pages, the more likely it is that a
490 * string of contiguous pages can be found after coalescing old pages from
491 * many old transactions together.
494 /** Number of slots in the reader table.
495 * This value was chosen somewhat arbitrarily. 126 readers plus a
496 * couple mutexes fit exactly into 8KB on my development machine.
497 * Applications should set the table size using #mdb_env_set_maxreaders().
499 #define DEFAULT_READERS 126
501 /** The size of a CPU cache line in bytes. We want our lock structures
502 * aligned to this size to avoid false cache line sharing in the
504 * This value works for most CPUs. For Itanium this should be 128.
510 /** The information we store in a single slot of the reader table.
511 * In addition to a transaction ID, we also record the process and
512 * thread ID that owns a slot, so that we can detect stale information,
513 * e.g. threads or processes that went away without cleaning up.
514 * @note We currently don't check for stale records. We simply re-init
515 * the table when we know that we're the only process opening the
518 typedef struct MDB_rxbody {
519 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
520 * Multiple readers that start at the same time will probably have the
521 * same ID here. Again, it's not important to exclude them from
522 * anything; all we need to know is which version of the DB they
523 * started from so we can avoid overwriting any data used in that
524 * particular version.
527 /** The process ID of the process owning this reader txn. */
529 /** The thread ID of the thread owning this txn. */
533 /** The actual reader record, with cacheline padding. */
534 typedef struct MDB_reader {
537 /** shorthand for mrb_txnid */
538 #define mr_txnid mru.mrx.mrb_txnid
539 #define mr_pid mru.mrx.mrb_pid
540 #define mr_tid mru.mrx.mrb_tid
541 /** cache line alignment */
542 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
546 /** The header for the reader table.
547 * The table resides in a memory-mapped file. (This is a different file
548 * than is used for the main database.)
550 * For POSIX the actual mutexes reside in the shared memory of this
551 * mapped file. On Windows, mutexes are named objects allocated by the
552 * kernel; we store the mutex names in this mapped file so that other
553 * processes can grab them. This same approach is also used on
554 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
555 * process-shared POSIX mutexes. For these cases where a named object
556 * is used, the object name is derived from a 64 bit FNV hash of the
557 * environment pathname. As such, naming collisions are extremely
558 * unlikely. If a collision occurs, the results are unpredictable.
560 typedef struct MDB_txbody {
561 /** Stamp identifying this as an MDB file. It must be set
564 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
566 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
567 char mtb_rmname[MNAME_LEN];
569 /** Mutex protecting access to this table.
570 * This is the reader lock that #LOCK_MUTEX_R acquires.
572 pthread_mutex_t mtb_mutex;
574 /** The ID of the last transaction committed to the database.
575 * This is recorded here only for convenience; the value can always
576 * be determined by reading the main database meta pages.
579 /** The number of slots that have been used in the reader table.
580 * This always records the maximum count, it is not decremented
581 * when readers release their slots.
583 unsigned mtb_numreaders;
586 /** The actual reader table definition. */
587 typedef struct MDB_txninfo {
590 #define mti_magic mt1.mtb.mtb_magic
591 #define mti_format mt1.mtb.mtb_format
592 #define mti_mutex mt1.mtb.mtb_mutex
593 #define mti_rmname mt1.mtb.mtb_rmname
594 #define mti_txnid mt1.mtb.mtb_txnid
595 #define mti_numreaders mt1.mtb.mtb_numreaders
596 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
599 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
600 char mt2_wmname[MNAME_LEN];
601 #define mti_wmname mt2.mt2_wmname
603 pthread_mutex_t mt2_wmutex;
604 #define mti_wmutex mt2.mt2_wmutex
606 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
608 MDB_reader mti_readers[1];
611 /** Lockfile format signature: version, features and field layout */
612 #define MDB_LOCK_FORMAT \
614 ((MDB_LOCK_VERSION) \
615 /* Flags which describe functionality */ \
616 + (((MDB_PIDLOCK) != 0) << 16)))
619 /** Common header for all page types.
620 * Overflow records occupy a number of contiguous pages with no
621 * headers on any page after the first.
623 typedef struct MDB_page {
624 #define mp_pgno mp_p.p_pgno
625 #define mp_next mp_p.p_next
627 pgno_t p_pgno; /**< page number */
628 void * p_next; /**< for in-memory list of freed structs */
631 /** @defgroup mdb_page Page Flags
633 * Flags for the page headers.
636 #define P_BRANCH 0x01 /**< branch page */
637 #define P_LEAF 0x02 /**< leaf page */
638 #define P_OVERFLOW 0x04 /**< overflow page */
639 #define P_META 0x08 /**< meta page */
640 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
641 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
642 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
643 #define P_KEEP 0x8000 /**< leave this page alone during spill */
645 uint16_t mp_flags; /**< @ref mdb_page */
646 #define mp_lower mp_pb.pb.pb_lower
647 #define mp_upper mp_pb.pb.pb_upper
648 #define mp_pages mp_pb.pb_pages
651 indx_t pb_lower; /**< lower bound of free space */
652 indx_t pb_upper; /**< upper bound of free space */
654 uint32_t pb_pages; /**< number of overflow pages */
656 indx_t mp_ptrs[1]; /**< dynamic size */
659 /** Size of the page header, excluding dynamic data at the end */
660 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
662 /** Address of first usable data byte in a page, after the header */
663 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
665 /** Number of nodes on a page */
666 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
668 /** The amount of space remaining in the page */
669 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
671 /** The percentage of space used in the page, in tenths of a percent. */
672 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
673 ((env)->me_psize - PAGEHDRSZ))
674 /** The minimum page fill factor, in tenths of a percent.
675 * Pages emptier than this are candidates for merging.
677 #define FILL_THRESHOLD 250
679 /** Test if a page is a leaf page */
680 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
681 /** Test if a page is a LEAF2 page */
682 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
683 /** Test if a page is a branch page */
684 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
685 /** Test if a page is an overflow page */
686 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
687 /** Test if a page is a sub page */
688 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
690 /** The number of overflow pages needed to store the given size. */
691 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
693 /** Header for a single key/data pair within a page.
694 * We guarantee 2-byte alignment for nodes.
696 typedef struct MDB_node {
697 /** lo and hi are used for data size on leaf nodes and for
698 * child pgno on branch nodes. On 64 bit platforms, flags
699 * is also used for pgno. (Branch nodes have no flags).
700 * They are in host byte order in case that lets some
701 * accesses be optimized into a 32-bit word access.
703 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
704 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
705 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
706 /** @defgroup mdb_node Node Flags
708 * Flags for node headers.
711 #define F_BIGDATA 0x01 /**< data put on overflow page */
712 #define F_SUBDATA 0x02 /**< data is a sub-database */
713 #define F_DUPDATA 0x04 /**< data has duplicates */
715 /** valid flags for #mdb_node_add() */
716 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
719 unsigned short mn_flags; /**< @ref mdb_node */
720 unsigned short mn_ksize; /**< key size */
721 char mn_data[1]; /**< key and data are appended here */
724 /** Size of the node header, excluding dynamic data at the end */
725 #define NODESIZE offsetof(MDB_node, mn_data)
727 /** Bit position of top word in page number, for shifting mn_flags */
728 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
730 /** Size of a node in a branch page with a given key.
731 * This is just the node header plus the key, there is no data.
733 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
735 /** Size of a node in a leaf page with a given key and data.
736 * This is node header plus key plus data size.
738 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
740 /** Address of node \b i in page \b p */
741 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
743 /** Address of the key for the node */
744 #define NODEKEY(node) (void *)((node)->mn_data)
746 /** Address of the data for a node */
747 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
749 /** Get the page number pointed to by a branch node */
750 #define NODEPGNO(node) \
751 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
752 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
753 /** Set the page number in a branch node */
754 #define SETPGNO(node,pgno) do { \
755 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
756 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
758 /** Get the size of the data in a leaf node */
759 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
760 /** Set the size of the data for a leaf node */
761 #define SETDSZ(node,size) do { \
762 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
763 /** The size of a key in a node */
764 #define NODEKSZ(node) ((node)->mn_ksize)
766 /** Copy a page number from src to dst */
768 #define COPY_PGNO(dst,src) dst = src
770 #if SIZE_MAX > 4294967295UL
771 #define COPY_PGNO(dst,src) do { \
772 unsigned short *s, *d; \
773 s = (unsigned short *)&(src); \
774 d = (unsigned short *)&(dst); \
781 #define COPY_PGNO(dst,src) do { \
782 unsigned short *s, *d; \
783 s = (unsigned short *)&(src); \
784 d = (unsigned short *)&(dst); \
790 /** The address of a key in a LEAF2 page.
791 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
792 * There are no node headers, keys are stored contiguously.
794 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
796 /** Set the \b node's key into \b keyptr, if requested. */
797 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
798 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
800 /** Set the \b node's key into \b key. */
801 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
803 /** Information about a single database in the environment. */
804 typedef struct MDB_db {
805 uint32_t md_pad; /**< also ksize for LEAF2 pages */
806 uint16_t md_flags; /**< @ref mdb_dbi_open */
807 uint16_t md_depth; /**< depth of this tree */
808 pgno_t md_branch_pages; /**< number of internal pages */
809 pgno_t md_leaf_pages; /**< number of leaf pages */
810 pgno_t md_overflow_pages; /**< number of overflow pages */
811 size_t md_entries; /**< number of data items */
812 pgno_t md_root; /**< the root page of this tree */
815 /** mdb_dbi_open flags */
816 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
817 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
818 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
819 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
821 /** Handle for the DB used to track free pages. */
823 /** Handle for the default DB. */
826 /** Meta page content.
827 * A meta page is the start point for accessing a database snapshot.
828 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
830 typedef struct MDB_meta {
831 /** Stamp identifying this as an MDB file. It must be set
834 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
836 void *mm_address; /**< address for fixed mapping */
837 size_t mm_mapsize; /**< size of mmap region */
838 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
839 /** The size of pages used in this DB */
840 #define mm_psize mm_dbs[0].md_pad
841 /** Any persistent environment flags. @ref mdb_env */
842 #define mm_flags mm_dbs[0].md_flags
843 pgno_t mm_last_pg; /**< last used page in file */
844 txnid_t mm_txnid; /**< txnid that committed this page */
847 /** Buffer for a stack-allocated meta page.
848 * The members define size and alignment, and silence type
849 * aliasing warnings. They are not used directly; that could
850 * mean incorrectly using several union members in parallel.
852 typedef union MDB_metabuf {
855 char mm_pad[PAGEHDRSZ];
860 /** Auxiliary DB info.
861 * The information here is mostly static/read-only. There is
862 * only a single copy of this record in the environment.
864 typedef struct MDB_dbx {
865 MDB_val md_name; /**< name of the database */
866 MDB_cmp_func *md_cmp; /**< function for comparing keys */
867 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
868 MDB_rel_func *md_rel; /**< user relocate function */
869 void *md_relctx; /**< user-provided context for md_rel */
872 /** A database transaction.
873 * Every operation requires a transaction handle.
876 MDB_txn *mt_parent; /**< parent of a nested txn */
877 MDB_txn *mt_child; /**< nested txn under this txn */
878 pgno_t mt_next_pgno; /**< next unallocated page */
879 /** The ID of this transaction. IDs are integers incrementing from 1.
880 * Only committed write transactions increment the ID. If a transaction
881 * aborts, the ID may be re-used by the next writer.
884 MDB_env *mt_env; /**< the DB environment */
885 /** The list of pages that became unused during this transaction.
888 /** The sorted list of dirty pages we temporarily wrote to disk
889 * because the dirty list was full. page numbers in here are
890 * shifted left by 1, deleted slots have the LSB set.
892 MDB_IDL mt_spill_pgs;
894 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
896 /** For read txns: This thread/txn's reader table slot, or NULL. */
899 /** Array of records for each DB known in the environment. */
901 /** Array of MDB_db records for each known DB */
903 /** @defgroup mt_dbflag Transaction DB Flags
907 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
908 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
909 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
910 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
912 /** In write txns, array of cursors for each DB */
913 MDB_cursor **mt_cursors;
914 /** Array of flags for each DB */
915 unsigned char *mt_dbflags;
916 /** Number of DB records in use. This number only ever increments;
917 * we don't decrement it when individual DB handles are closed.
921 /** @defgroup mdb_txn Transaction Flags
925 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
926 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
927 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
928 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
930 unsigned int mt_flags; /**< @ref mdb_txn */
931 /** dirty_list room: Array size - #dirty pages visible to this txn.
932 * Includes ancestor txns' dirty pages not hidden by other txns'
933 * dirty/spilled pages. Thus commit(nested txn) has room to merge
934 * dirty_list into mt_parent after freeing hidden mt_parent pages.
936 unsigned int mt_dirty_room;
939 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
940 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
941 * raise this on a 64 bit machine.
943 #define CURSOR_STACK 32
947 /** Cursors are used for all DB operations.
948 * A cursor holds a path of (page pointer, key index) from the DB
949 * root to a position in the DB, plus other state. #MDB_DUPSORT
950 * cursors include an xcursor to the current data item. Write txns
951 * track their cursors and keep them up to date when data moves.
952 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
953 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
956 /** Next cursor on this DB in this txn */
958 /** Backup of the original cursor if this cursor is a shadow */
959 MDB_cursor *mc_backup;
960 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
961 struct MDB_xcursor *mc_xcursor;
962 /** The transaction that owns this cursor */
964 /** The database handle this cursor operates on */
966 /** The database record for this cursor */
968 /** The database auxiliary record for this cursor */
970 /** The @ref mt_dbflag for this database */
971 unsigned char *mc_dbflag;
972 unsigned short mc_snum; /**< number of pushed pages */
973 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
974 /** @defgroup mdb_cursor Cursor Flags
976 * Cursor state flags.
979 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
980 #define C_EOF 0x02 /**< No more data */
981 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
982 #define C_DEL 0x08 /**< last op was a cursor_del */
983 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
984 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
986 unsigned int mc_flags; /**< @ref mdb_cursor */
987 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
988 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
991 /** Context for sorted-dup records.
992 * We could have gone to a fully recursive design, with arbitrarily
993 * deep nesting of sub-databases. But for now we only handle these
994 * levels - main DB, optional sub-DB, sorted-duplicate DB.
996 typedef struct MDB_xcursor {
997 /** A sub-cursor for traversing the Dup DB */
998 MDB_cursor mx_cursor;
999 /** The database record for this Dup DB */
1001 /** The auxiliary DB record for this Dup DB */
1003 /** The @ref mt_dbflag for this Dup DB */
1004 unsigned char mx_dbflag;
1007 /** State of FreeDB old pages, stored in the MDB_env */
1008 typedef struct MDB_pgstate {
1009 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1010 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1013 /** The database environment. */
1015 HANDLE me_fd; /**< The main data file */
1016 HANDLE me_lfd; /**< The lock file */
1017 HANDLE me_mfd; /**< just for writing the meta pages */
1018 /** Failed to update the meta page. Probably an I/O error. */
1019 #define MDB_FATAL_ERROR 0x80000000U
1020 /** Some fields are initialized. */
1021 #define MDB_ENV_ACTIVE 0x20000000U
1022 /** me_txkey is set */
1023 #define MDB_ENV_TXKEY 0x10000000U
1024 /** Have liveness lock in reader table */
1025 #define MDB_LIVE_READER 0x08000000U
1026 uint32_t me_flags; /**< @ref mdb_env */
1027 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1028 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1029 unsigned int me_maxreaders; /**< size of the reader table */
1030 unsigned int me_numreaders; /**< max numreaders set by this env */
1031 MDB_dbi me_numdbs; /**< number of DBs opened */
1032 MDB_dbi me_maxdbs; /**< size of the DB table */
1033 MDB_PID_T me_pid; /**< process ID of this env */
1034 char *me_path; /**< path to the DB files */
1035 char *me_map; /**< the memory map of the data file */
1036 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1037 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1038 void *me_pbuf; /**< scratch area for DUPSORT put() */
1039 MDB_txn *me_txn; /**< current write transaction */
1040 size_t me_mapsize; /**< size of the data memory map */
1041 off_t me_size; /**< current file size */
1042 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1043 MDB_dbx *me_dbxs; /**< array of static DB info */
1044 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1045 pthread_key_t me_txkey; /**< thread-key for readers */
1046 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1047 # define me_pglast me_pgstate.mf_pglast
1048 # define me_pghead me_pgstate.mf_pghead
1049 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1050 /** IDL of pages that became unused in a write txn */
1051 MDB_IDL me_free_pgs;
1052 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1053 MDB_ID2L me_dirty_list;
1054 /** Max number of freelist items that can fit in a single overflow page */
1056 /** Max size of a node on a page */
1057 unsigned int me_nodemax;
1058 #if !(MDB_MAXKEYSIZE)
1059 unsigned int me_maxkey; /**< max size of a key */
1062 int me_pidquery; /**< Used in OpenProcess */
1063 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1065 #elif defined(MDB_USE_POSIX_SEM)
1066 sem_t *me_rmutex; /* Shared mutexes are not supported */
1071 /** Nested transaction */
1072 typedef struct MDB_ntxn {
1073 MDB_txn mnt_txn; /**< the transaction */
1074 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1077 /** max number of pages to commit in one writev() call */
1078 #define MDB_COMMIT_PAGES 64
1079 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1080 #undef MDB_COMMIT_PAGES
1081 #define MDB_COMMIT_PAGES IOV_MAX
1084 /* max bytes to write in one call */
1085 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1087 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1088 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1089 static int mdb_page_touch(MDB_cursor *mc);
1091 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1092 static int mdb_page_search_root(MDB_cursor *mc,
1093 MDB_val *key, int modify);
1094 #define MDB_PS_MODIFY 1
1095 #define MDB_PS_ROOTONLY 2
1096 #define MDB_PS_FIRST 4
1097 #define MDB_PS_LAST 8
1098 static int mdb_page_search(MDB_cursor *mc,
1099 MDB_val *key, int flags);
1100 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1102 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1103 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1104 pgno_t newpgno, unsigned int nflags);
1106 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1107 static int mdb_env_pick_meta(const MDB_env *env);
1108 static int mdb_env_write_meta(MDB_txn *txn);
1109 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1110 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1112 static void mdb_env_close0(MDB_env *env, int excl);
1114 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1115 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1116 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1117 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1118 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1119 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1120 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1121 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1122 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1124 static int mdb_rebalance(MDB_cursor *mc);
1125 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1127 static void mdb_cursor_pop(MDB_cursor *mc);
1128 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1130 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1131 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1132 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1133 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1134 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1136 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1137 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1139 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1140 static void mdb_xcursor_init0(MDB_cursor *mc);
1141 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1143 static int mdb_drop0(MDB_cursor *mc, int subs);
1144 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1147 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1151 static SECURITY_DESCRIPTOR mdb_null_sd;
1152 static SECURITY_ATTRIBUTES mdb_all_sa;
1153 static int mdb_sec_inited;
1156 /** Return the library version info. */
1158 mdb_version(int *major, int *minor, int *patch)
1160 if (major) *major = MDB_VERSION_MAJOR;
1161 if (minor) *minor = MDB_VERSION_MINOR;
1162 if (patch) *patch = MDB_VERSION_PATCH;
1163 return MDB_VERSION_STRING;
1166 /** Table of descriptions for MDB @ref errors */
1167 static char *const mdb_errstr[] = {
1168 "MDB_KEYEXIST: Key/data pair already exists",
1169 "MDB_NOTFOUND: No matching key/data pair found",
1170 "MDB_PAGE_NOTFOUND: Requested page not found",
1171 "MDB_CORRUPTED: Located page was wrong type",
1172 "MDB_PANIC: Update of meta page failed",
1173 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1174 "MDB_INVALID: File is not an MDB file",
1175 "MDB_MAP_FULL: Environment mapsize limit reached",
1176 "MDB_DBS_FULL: Environment maxdbs limit reached",
1177 "MDB_READERS_FULL: Environment maxreaders limit reached",
1178 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1179 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1180 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1181 "MDB_PAGE_FULL: Internal error - page has no more space",
1182 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1183 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1184 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1185 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1186 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1190 mdb_strerror(int err)
1194 return ("Successful return: 0");
1196 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1197 i = err - MDB_KEYEXIST;
1198 return mdb_errstr[i];
1201 return strerror(err);
1205 /** Display a key in hexadecimal and return the address of the result.
1206 * @param[in] key the key to display
1207 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1208 * @return The key in hexadecimal form.
1211 mdb_dkey(MDB_val *key, char *buf)
1214 unsigned char *c = key->mv_data;
1220 if (key->mv_size > DKBUF_MAXKEYSIZE)
1221 return "MDB_MAXKEYSIZE";
1222 /* may want to make this a dynamic check: if the key is mostly
1223 * printable characters, print it as-is instead of converting to hex.
1227 for (i=0; i<key->mv_size; i++)
1228 ptr += sprintf(ptr, "%02x", *c++);
1230 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1235 /** Display all the keys in the page. */
1237 mdb_page_list(MDB_page *mp)
1240 unsigned int i, nkeys, nsize, total = 0;
1244 nkeys = NUMKEYS(mp);
1245 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1246 for (i=0; i<nkeys; i++) {
1247 node = NODEPTR(mp, i);
1248 key.mv_size = node->mn_ksize;
1249 key.mv_data = node->mn_data;
1250 nsize = NODESIZE + key.mv_size;
1251 if (IS_BRANCH(mp)) {
1252 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1256 if (F_ISSET(node->mn_flags, F_BIGDATA))
1257 nsize += sizeof(pgno_t);
1259 nsize += NODEDSZ(node);
1261 nsize += sizeof(indx_t);
1262 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1264 total = EVEN(total);
1266 fprintf(stderr, "Total: %d\n", total);
1270 mdb_cursor_chk(MDB_cursor *mc)
1276 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1277 for (i=0; i<mc->mc_top; i++) {
1279 node = NODEPTR(mp, mc->mc_ki[i]);
1280 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1283 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1289 /** Count all the pages in each DB and in the freelist
1290 * and make sure it matches the actual number of pages
1293 static void mdb_audit(MDB_txn *txn)
1297 MDB_ID freecount, count;
1302 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1303 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1304 freecount += *(MDB_ID *)data.mv_data;
1307 for (i = 0; i<txn->mt_numdbs; i++) {
1309 mdb_cursor_init(&mc, txn, i, &mx);
1310 if (txn->mt_dbs[i].md_root == P_INVALID)
1312 count += txn->mt_dbs[i].md_branch_pages +
1313 txn->mt_dbs[i].md_leaf_pages +
1314 txn->mt_dbs[i].md_overflow_pages;
1315 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1316 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1320 mp = mc.mc_pg[mc.mc_top];
1321 for (j=0; j<NUMKEYS(mp); j++) {
1322 MDB_node *leaf = NODEPTR(mp, j);
1323 if (leaf->mn_flags & F_SUBDATA) {
1325 memcpy(&db, NODEDATA(leaf), sizeof(db));
1326 count += db.md_branch_pages + db.md_leaf_pages +
1327 db.md_overflow_pages;
1331 while (mdb_cursor_sibling(&mc, 1) == 0);
1334 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1335 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1336 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1342 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1344 return txn->mt_dbxs[dbi].md_cmp(a, b);
1348 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1350 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1353 /** Allocate memory for a page.
1354 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1357 mdb_page_malloc(MDB_txn *txn, unsigned num)
1359 MDB_env *env = txn->mt_env;
1360 MDB_page *ret = env->me_dpages;
1361 size_t psize = env->me_psize, sz = psize, off;
1362 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1363 * For a single page alloc, we init everything after the page header.
1364 * For multi-page, we init the final page; if the caller needed that
1365 * many pages they will be filling in at least up to the last page.
1369 VGMEMP_ALLOC(env, ret, sz);
1370 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1371 env->me_dpages = ret->mp_next;
1374 psize -= off = PAGEHDRSZ;
1379 if ((ret = malloc(sz)) != NULL) {
1380 if (!(env->me_flags & MDB_NOMEMINIT)) {
1381 memset((char *)ret + off, 0, psize);
1384 VGMEMP_ALLOC(env, ret, sz);
1389 /** Free a single page.
1390 * Saves single pages to a list, for future reuse.
1391 * (This is not used for multi-page overflow pages.)
1394 mdb_page_free(MDB_env *env, MDB_page *mp)
1396 mp->mp_next = env->me_dpages;
1397 VGMEMP_FREE(env, mp);
1398 env->me_dpages = mp;
1401 /** Free a dirty page */
1403 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1405 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1406 mdb_page_free(env, dp);
1408 /* large pages just get freed directly */
1409 VGMEMP_FREE(env, dp);
1414 /** Return all dirty pages to dpage list */
1416 mdb_dlist_free(MDB_txn *txn)
1418 MDB_env *env = txn->mt_env;
1419 MDB_ID2L dl = txn->mt_u.dirty_list;
1420 unsigned i, n = dl[0].mid;
1422 for (i = 1; i <= n; i++) {
1423 mdb_dpage_free(env, dl[i].mptr);
1428 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1429 * @param[in] mc A cursor handle for the current operation.
1430 * @param[in] pflags Flags of the pages to update:
1431 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1432 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1433 * @return 0 on success, non-zero on failure.
1436 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1438 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1439 MDB_txn *txn = mc->mc_txn;
1445 int rc = MDB_SUCCESS, level;
1447 /* Mark pages seen by cursors */
1448 if (mc->mc_flags & C_UNTRACK)
1449 mc = NULL; /* will find mc in mt_cursors */
1450 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1451 for (; mc; mc=mc->mc_next) {
1452 if (!(mc->mc_flags & C_INITIALIZED))
1454 for (m3 = mc;; m3 = &mx->mx_cursor) {
1456 for (j=0; j<m3->mc_snum; j++) {
1458 if ((mp->mp_flags & Mask) == pflags)
1459 mp->mp_flags ^= P_KEEP;
1461 mx = m3->mc_xcursor;
1462 /* Proceed to mx if it is at a sub-database */
1463 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1465 if (! (mp && (mp->mp_flags & P_LEAF)))
1467 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1468 if (!(leaf->mn_flags & F_SUBDATA))
1477 /* Mark dirty root pages */
1478 for (i=0; i<txn->mt_numdbs; i++) {
1479 if (txn->mt_dbflags[i] & DB_DIRTY) {
1480 pgno_t pgno = txn->mt_dbs[i].md_root;
1481 if (pgno == P_INVALID)
1483 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1485 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1486 dp->mp_flags ^= P_KEEP;
1494 static int mdb_page_flush(MDB_txn *txn, int keep);
1496 /** Spill pages from the dirty list back to disk.
1497 * This is intended to prevent running into #MDB_TXN_FULL situations,
1498 * but note that they may still occur in a few cases:
1499 * 1) our estimate of the txn size could be too small. Currently this
1500 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1501 * 2) child txns may run out of space if their parents dirtied a
1502 * lot of pages and never spilled them. TODO: we probably should do
1503 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1504 * the parent's dirty_room is below a given threshold.
1506 * Otherwise, if not using nested txns, it is expected that apps will
1507 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1508 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1509 * If the txn never references them again, they can be left alone.
1510 * If the txn only reads them, they can be used without any fuss.
1511 * If the txn writes them again, they can be dirtied immediately without
1512 * going thru all of the work of #mdb_page_touch(). Such references are
1513 * handled by #mdb_page_unspill().
1515 * Also note, we never spill DB root pages, nor pages of active cursors,
1516 * because we'll need these back again soon anyway. And in nested txns,
1517 * we can't spill a page in a child txn if it was already spilled in a
1518 * parent txn. That would alter the parent txns' data even though
1519 * the child hasn't committed yet, and we'd have no way to undo it if
1520 * the child aborted.
1522 * @param[in] m0 cursor A cursor handle identifying the transaction and
1523 * database for which we are checking space.
1524 * @param[in] key For a put operation, the key being stored.
1525 * @param[in] data For a put operation, the data being stored.
1526 * @return 0 on success, non-zero on failure.
1529 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1531 MDB_txn *txn = m0->mc_txn;
1533 MDB_ID2L dl = txn->mt_u.dirty_list;
1534 unsigned int i, j, need;
1537 if (m0->mc_flags & C_SUB)
1540 /* Estimate how much space this op will take */
1541 i = m0->mc_db->md_depth;
1542 /* Named DBs also dirty the main DB */
1543 if (m0->mc_dbi > MAIN_DBI)
1544 i += txn->mt_dbs[MAIN_DBI].md_depth;
1545 /* For puts, roughly factor in the key+data size */
1547 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1548 i += i; /* double it for good measure */
1551 if (txn->mt_dirty_room > i)
1554 if (!txn->mt_spill_pgs) {
1555 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1556 if (!txn->mt_spill_pgs)
1559 /* purge deleted slots */
1560 MDB_IDL sl = txn->mt_spill_pgs;
1561 unsigned int num = sl[0];
1563 for (i=1; i<=num; i++) {
1570 /* Preserve pages which may soon be dirtied again */
1571 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1574 /* Less aggressive spill - we originally spilled the entire dirty list,
1575 * with a few exceptions for cursor pages and DB root pages. But this
1576 * turns out to be a lot of wasted effort because in a large txn many
1577 * of those pages will need to be used again. So now we spill only 1/8th
1578 * of the dirty pages. Testing revealed this to be a good tradeoff,
1579 * better than 1/2, 1/4, or 1/10.
1581 if (need < MDB_IDL_UM_MAX / 8)
1582 need = MDB_IDL_UM_MAX / 8;
1584 /* Save the page IDs of all the pages we're flushing */
1585 /* flush from the tail forward, this saves a lot of shifting later on. */
1586 for (i=dl[0].mid; i && need; i--) {
1587 MDB_ID pn = dl[i].mid << 1;
1589 if (dp->mp_flags & P_KEEP)
1591 /* Can't spill twice, make sure it's not already in a parent's
1594 if (txn->mt_parent) {
1596 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1597 if (tx2->mt_spill_pgs) {
1598 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1599 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1600 dp->mp_flags |= P_KEEP;
1608 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1612 mdb_midl_sort(txn->mt_spill_pgs);
1614 /* Flush the spilled part of dirty list */
1615 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1618 /* Reset any dirty pages we kept that page_flush didn't see */
1619 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1622 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1626 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1628 mdb_find_oldest(MDB_txn *txn)
1631 txnid_t mr, oldest = txn->mt_txnid - 1;
1632 if (txn->mt_env->me_txns) {
1633 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1634 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1645 /** Add a page to the txn's dirty list */
1647 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1650 int (*insert)(MDB_ID2L, MDB_ID2 *);
1652 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1653 insert = mdb_mid2l_append;
1655 insert = mdb_mid2l_insert;
1657 mid.mid = mp->mp_pgno;
1659 insert(txn->mt_u.dirty_list, &mid);
1660 txn->mt_dirty_room--;
1663 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1664 * me_pghead and mt_next_pgno.
1666 * If there are free pages available from older transactions, they
1667 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1668 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1669 * and move me_pglast to say which records were consumed. Only this
1670 * function can create me_pghead and move me_pglast/mt_next_pgno.
1671 * @param[in] mc cursor A cursor handle identifying the transaction and
1672 * database for which we are allocating.
1673 * @param[in] num the number of pages to allocate.
1674 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1675 * will always be satisfied by a single contiguous chunk of memory.
1676 * @return 0 on success, non-zero on failure.
1679 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1681 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1682 /* Get at most <Max_retries> more freeDB records once me_pghead
1683 * has enough pages. If not enough, use new pages from the map.
1684 * If <Paranoid> and mc is updating the freeDB, only get new
1685 * records if me_pghead is empty. Then the freelist cannot play
1686 * catch-up with itself by growing while trying to save it.
1688 enum { Paranoid = 1, Max_retries = 500 };
1690 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1692 int rc, retry = Max_retries;
1693 MDB_txn *txn = mc->mc_txn;
1694 MDB_env *env = txn->mt_env;
1695 pgno_t pgno, *mop = env->me_pghead;
1696 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1698 txnid_t oldest = 0, last;
1704 /* If our dirty list is already full, we can't do anything */
1705 if (txn->mt_dirty_room == 0)
1706 return MDB_TXN_FULL;
1708 for (op = MDB_FIRST;; op = MDB_NEXT) {
1711 pgno_t *idl, old_id, new_id;
1713 /* Seek a big enough contiguous page range. Prefer
1714 * pages at the tail, just truncating the list.
1720 if (mop[i-n2] == pgno+n2)
1723 if (Max_retries < INT_MAX && --retry < 0)
1727 if (op == MDB_FIRST) { /* 1st iteration */
1728 /* Prepare to fetch more and coalesce */
1729 oldest = mdb_find_oldest(txn);
1730 last = env->me_pglast;
1731 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1734 key.mv_data = &last; /* will look up last+1 */
1735 key.mv_size = sizeof(last);
1737 if (Paranoid && mc->mc_dbi == FREE_DBI)
1740 if (Paranoid && retry < 0 && mop_len)
1744 /* Do not fetch more if the record will be too recent */
1747 rc = mdb_cursor_get(&m2, &key, NULL, op);
1749 if (rc == MDB_NOTFOUND)
1753 last = *(txnid_t*)key.mv_data;
1756 np = m2.mc_pg[m2.mc_top];
1757 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1758 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1761 idl = (MDB_ID *) data.mv_data;
1764 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1767 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1769 mop = env->me_pghead;
1771 env->me_pglast = last;
1773 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1774 last, txn->mt_dbs[FREE_DBI].md_root, i));
1776 DPRINTF(("IDL %"Z"u", idl[k]));
1778 /* Merge in descending sorted order */
1781 mop[0] = (pgno_t)-1;
1785 for (; old_id < new_id; old_id = mop[--j])
1792 /* Use new pages from the map when nothing suitable in the freeDB */
1794 pgno = txn->mt_next_pgno;
1795 if (pgno + num >= env->me_maxpg) {
1796 DPUTS("DB size maxed out");
1797 return MDB_MAP_FULL;
1801 if (env->me_flags & MDB_WRITEMAP) {
1802 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1804 if (!(np = mdb_page_malloc(txn, num)))
1808 mop[0] = mop_len -= num;
1809 /* Move any stragglers down */
1810 for (j = i-num; j < mop_len; )
1811 mop[++j] = mop[++i];
1813 txn->mt_next_pgno = pgno + num;
1816 mdb_page_dirty(txn, np);
1822 /** Copy the used portions of a non-overflow page.
1823 * @param[in] dst page to copy into
1824 * @param[in] src page to copy from
1825 * @param[in] psize size of a page
1828 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1830 enum { Align = sizeof(pgno_t) };
1831 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1833 /* If page isn't full, just copy the used portion. Adjust
1834 * alignment so memcpy may copy words instead of bytes.
1836 if ((unused &= -Align) && !IS_LEAF2(src)) {
1838 memcpy(dst, src, (lower + (Align-1)) & -Align);
1839 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1842 memcpy(dst, src, psize - unused);
1846 /** Pull a page off the txn's spill list, if present.
1847 * If a page being referenced was spilled to disk in this txn, bring
1848 * it back and make it dirty/writable again.
1849 * @param[in] txn the transaction handle.
1850 * @param[in] mp the page being referenced.
1851 * @param[out] ret the writable page, if any. ret is unchanged if
1852 * mp wasn't spilled.
1855 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1857 MDB_env *env = txn->mt_env;
1860 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1862 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1863 if (!tx2->mt_spill_pgs)
1865 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1866 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1869 if (txn->mt_dirty_room == 0)
1870 return MDB_TXN_FULL;
1871 if (IS_OVERFLOW(mp))
1875 if (env->me_flags & MDB_WRITEMAP) {
1878 np = mdb_page_malloc(txn, num);
1882 memcpy(np, mp, num * env->me_psize);
1884 mdb_page_copy(np, mp, env->me_psize);
1887 /* If in current txn, this page is no longer spilled.
1888 * If it happens to be the last page, truncate the spill list.
1889 * Otherwise mark it as deleted by setting the LSB.
1891 if (x == txn->mt_spill_pgs[0])
1892 txn->mt_spill_pgs[0]--;
1894 txn->mt_spill_pgs[x] |= 1;
1895 } /* otherwise, if belonging to a parent txn, the
1896 * page remains spilled until child commits
1899 mdb_page_dirty(txn, np);
1900 np->mp_flags |= P_DIRTY;
1908 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1909 * @param[in] mc cursor pointing to the page to be touched
1910 * @return 0 on success, non-zero on failure.
1913 mdb_page_touch(MDB_cursor *mc)
1915 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1916 MDB_txn *txn = mc->mc_txn;
1917 MDB_cursor *m2, *m3;
1921 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1922 if (txn->mt_flags & MDB_TXN_SPILLS) {
1924 rc = mdb_page_unspill(txn, mp, &np);
1930 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1931 (rc = mdb_page_alloc(mc, 1, &np)))
1934 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
1935 mp->mp_pgno, pgno));
1936 assert(mp->mp_pgno != pgno);
1937 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1938 /* Update the parent page, if any, to point to the new page */
1940 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1941 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1942 SETPGNO(node, pgno);
1944 mc->mc_db->md_root = pgno;
1946 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1947 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1949 /* If txn has a parent, make sure the page is in our
1953 unsigned x = mdb_mid2l_search(dl, pgno);
1954 if (x <= dl[0].mid && dl[x].mid == pgno) {
1955 if (mp != dl[x].mptr) { /* bad cursor? */
1956 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1957 return MDB_CORRUPTED;
1962 assert(dl[0].mid < MDB_IDL_UM_MAX);
1964 np = mdb_page_malloc(txn, 1);
1969 mdb_mid2l_insert(dl, &mid);
1974 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1976 np->mp_flags |= P_DIRTY;
1979 /* Adjust cursors pointing to mp */
1980 mc->mc_pg[mc->mc_top] = np;
1981 m2 = txn->mt_cursors[mc->mc_dbi];
1982 if (mc->mc_flags & C_SUB) {
1983 for (; m2; m2=m2->mc_next) {
1984 m3 = &m2->mc_xcursor->mx_cursor;
1985 if (m3->mc_snum < mc->mc_snum) continue;
1986 if (m3->mc_pg[mc->mc_top] == mp)
1987 m3->mc_pg[mc->mc_top] = np;
1990 for (; m2; m2=m2->mc_next) {
1991 if (m2->mc_snum < mc->mc_snum) continue;
1992 if (m2->mc_pg[mc->mc_top] == mp) {
1993 m2->mc_pg[mc->mc_top] = np;
1994 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1995 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1997 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1998 if (!(leaf->mn_flags & F_SUBDATA))
1999 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2008 mdb_env_sync(MDB_env *env, int force)
2011 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2012 if (env->me_flags & MDB_WRITEMAP) {
2013 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2014 ? MS_ASYNC : MS_SYNC;
2015 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2018 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2022 if (MDB_FDATASYNC(env->me_fd))
2029 /** Back up parent txn's cursors, then grab the originals for tracking */
2031 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2033 MDB_cursor *mc, *bk;
2038 for (i = src->mt_numdbs; --i >= 0; ) {
2039 if ((mc = src->mt_cursors[i]) != NULL) {
2040 size = sizeof(MDB_cursor);
2042 size += sizeof(MDB_xcursor);
2043 for (; mc; mc = bk->mc_next) {
2049 mc->mc_db = &dst->mt_dbs[i];
2050 /* Kill pointers into src - and dst to reduce abuse: The
2051 * user may not use mc until dst ends. Otherwise we'd...
2053 mc->mc_txn = NULL; /* ...set this to dst */
2054 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2055 if ((mx = mc->mc_xcursor) != NULL) {
2056 *(MDB_xcursor *)(bk+1) = *mx;
2057 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2059 mc->mc_next = dst->mt_cursors[i];
2060 dst->mt_cursors[i] = mc;
2067 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2068 * @param[in] txn the transaction handle.
2069 * @param[in] merge true to keep changes to parent cursors, false to revert.
2070 * @return 0 on success, non-zero on failure.
2073 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2075 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2079 for (i = txn->mt_numdbs; --i >= 0; ) {
2080 for (mc = cursors[i]; mc; mc = next) {
2082 if ((bk = mc->mc_backup) != NULL) {
2084 /* Commit changes to parent txn */
2085 mc->mc_next = bk->mc_next;
2086 mc->mc_backup = bk->mc_backup;
2087 mc->mc_txn = bk->mc_txn;
2088 mc->mc_db = bk->mc_db;
2089 mc->mc_dbflag = bk->mc_dbflag;
2090 if ((mx = mc->mc_xcursor) != NULL)
2091 mx->mx_cursor.mc_txn = bk->mc_txn;
2093 /* Abort nested txn */
2095 if ((mx = mc->mc_xcursor) != NULL)
2096 *mx = *(MDB_xcursor *)(bk+1);
2100 /* Only malloced cursors are permanently tracked. */
2108 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2111 mdb_txn_reset0(MDB_txn *txn, const char *act);
2113 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2119 Pidset = F_SETLK, Pidcheck = F_GETLK
2123 /** Set or check a pid lock. Set returns 0 on success.
2124 * Check returns 0 if the process is certainly dead, nonzero if it may
2125 * be alive (the lock exists or an error happened so we do not know).
2127 * On Windows Pidset is a no-op, we merely check for the existence
2128 * of the process with the given pid. On POSIX we use a single byte
2129 * lock on the lockfile, set at an offset equal to the pid.
2132 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2134 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2137 if (op == Pidcheck) {
2138 h = OpenProcess(env->me_pidquery, FALSE, pid);
2139 /* No documented "no such process" code, but other program use this: */
2141 return ErrCode() != ERROR_INVALID_PARAMETER;
2142 /* A process exists until all handles to it close. Has it exited? */
2143 ret = WaitForSingleObject(h, 0) != 0;
2150 struct flock lock_info;
2151 memset(&lock_info, 0, sizeof(lock_info));
2152 lock_info.l_type = F_WRLCK;
2153 lock_info.l_whence = SEEK_SET;
2154 lock_info.l_start = pid;
2155 lock_info.l_len = 1;
2156 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2157 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2159 } else if ((rc = ErrCode()) == EINTR) {
2167 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2168 * @param[in] txn the transaction handle to initialize
2169 * @return 0 on success, non-zero on failure.
2172 mdb_txn_renew0(MDB_txn *txn)
2174 MDB_env *env = txn->mt_env;
2175 MDB_txninfo *ti = env->me_txns;
2179 int rc, new_notls = 0;
2182 txn->mt_numdbs = env->me_numdbs;
2183 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2185 if (txn->mt_flags & MDB_TXN_RDONLY) {
2187 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2188 txn->mt_txnid = meta->mm_txnid;
2189 txn->mt_u.reader = NULL;
2191 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2192 pthread_getspecific(env->me_txkey);
2194 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2195 return MDB_BAD_RSLOT;
2197 MDB_PID_T pid = env->me_pid;
2198 pthread_t tid = pthread_self();
2200 if (!(env->me_flags & MDB_LIVE_READER)) {
2201 rc = mdb_reader_pid(env, Pidset, pid);
2203 UNLOCK_MUTEX_R(env);
2206 env->me_flags |= MDB_LIVE_READER;
2210 nr = ti->mti_numreaders;
2211 for (i=0; i<nr; i++)
2212 if (ti->mti_readers[i].mr_pid == 0)
2214 if (i == env->me_maxreaders) {
2215 UNLOCK_MUTEX_R(env);
2216 return MDB_READERS_FULL;
2218 ti->mti_readers[i].mr_pid = pid;
2219 ti->mti_readers[i].mr_tid = tid;
2221 ti->mti_numreaders = ++nr;
2222 /* Save numreaders for un-mutexed mdb_env_close() */
2223 env->me_numreaders = nr;
2224 UNLOCK_MUTEX_R(env);
2226 r = &ti->mti_readers[i];
2227 new_notls = (env->me_flags & MDB_NOTLS);
2228 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2233 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2234 txn->mt_u.reader = r;
2235 meta = env->me_metas[txn->mt_txnid & 1];
2241 txn->mt_txnid = ti->mti_txnid;
2242 meta = env->me_metas[txn->mt_txnid & 1];
2244 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2245 txn->mt_txnid = meta->mm_txnid;
2249 if (txn->mt_txnid == mdb_debug_start)
2252 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2253 txn->mt_u.dirty_list = env->me_dirty_list;
2254 txn->mt_u.dirty_list[0].mid = 0;
2255 txn->mt_free_pgs = env->me_free_pgs;
2256 txn->mt_free_pgs[0] = 0;
2257 txn->mt_spill_pgs = NULL;
2261 /* Copy the DB info and flags */
2262 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2264 /* Moved to here to avoid a data race in read TXNs */
2265 txn->mt_next_pgno = meta->mm_last_pg+1;
2267 for (i=2; i<txn->mt_numdbs; i++) {
2268 x = env->me_dbflags[i];
2269 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2270 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2272 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2274 if (env->me_maxpg < txn->mt_next_pgno) {
2275 mdb_txn_reset0(txn, "renew0-mapfail");
2277 txn->mt_u.reader->mr_pid = 0;
2278 txn->mt_u.reader = NULL;
2280 return MDB_MAP_RESIZED;
2287 mdb_txn_renew(MDB_txn *txn)
2291 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2294 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2295 DPUTS("environment had fatal error, must shutdown!");
2299 rc = mdb_txn_renew0(txn);
2300 if (rc == MDB_SUCCESS) {
2301 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2302 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2303 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2309 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2313 int rc, size, tsize = sizeof(MDB_txn);
2315 if (env->me_flags & MDB_FATAL_ERROR) {
2316 DPUTS("environment had fatal error, must shutdown!");
2319 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2322 /* Nested transactions: Max 1 child, write txns only, no writemap */
2323 if (parent->mt_child ||
2324 (flags & MDB_RDONLY) ||
2325 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2326 (env->me_flags & MDB_WRITEMAP))
2328 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2330 tsize = sizeof(MDB_ntxn);
2332 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2333 if (!(flags & MDB_RDONLY))
2334 size += env->me_maxdbs * sizeof(MDB_cursor *);
2336 if ((txn = calloc(1, size)) == NULL) {
2337 DPRINTF(("calloc: %s", strerror(ErrCode())));
2340 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2341 if (flags & MDB_RDONLY) {
2342 txn->mt_flags |= MDB_TXN_RDONLY;
2343 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2345 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2346 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2352 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2353 if (!txn->mt_u.dirty_list ||
2354 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2356 free(txn->mt_u.dirty_list);
2360 txn->mt_txnid = parent->mt_txnid;
2361 txn->mt_dirty_room = parent->mt_dirty_room;
2362 txn->mt_u.dirty_list[0].mid = 0;
2363 txn->mt_spill_pgs = NULL;
2364 txn->mt_next_pgno = parent->mt_next_pgno;
2365 parent->mt_child = txn;
2366 txn->mt_parent = parent;
2367 txn->mt_numdbs = parent->mt_numdbs;
2368 txn->mt_flags = parent->mt_flags;
2369 txn->mt_dbxs = parent->mt_dbxs;
2370 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2371 /* Copy parent's mt_dbflags, but clear DB_NEW */
2372 for (i=0; i<txn->mt_numdbs; i++)
2373 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2375 ntxn = (MDB_ntxn *)txn;
2376 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2377 if (env->me_pghead) {
2378 size = MDB_IDL_SIZEOF(env->me_pghead);
2379 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2381 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2386 rc = mdb_cursor_shadow(parent, txn);
2388 mdb_txn_reset0(txn, "beginchild-fail");
2390 rc = mdb_txn_renew0(txn);
2396 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2397 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2398 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2405 mdb_txn_env(MDB_txn *txn)
2407 if(!txn) return NULL;
2411 /** Export or close DBI handles opened in this txn. */
2413 mdb_dbis_update(MDB_txn *txn, int keep)
2416 MDB_dbi n = txn->mt_numdbs;
2417 MDB_env *env = txn->mt_env;
2418 unsigned char *tdbflags = txn->mt_dbflags;
2420 for (i = n; --i >= 2;) {
2421 if (tdbflags[i] & DB_NEW) {
2423 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2425 char *ptr = env->me_dbxs[i].md_name.mv_data;
2426 env->me_dbxs[i].md_name.mv_data = NULL;
2427 env->me_dbxs[i].md_name.mv_size = 0;
2428 env->me_dbflags[i] = 0;
2433 if (keep && env->me_numdbs < n)
2437 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2438 * May be called twice for readonly txns: First reset it, then abort.
2439 * @param[in] txn the transaction handle to reset
2440 * @param[in] act why the transaction is being reset
2443 mdb_txn_reset0(MDB_txn *txn, const char *act)
2445 MDB_env *env = txn->mt_env;
2447 /* Close any DBI handles opened in this txn */
2448 mdb_dbis_update(txn, 0);
2450 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2451 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2452 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2454 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2455 if (txn->mt_u.reader) {
2456 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2457 if (!(env->me_flags & MDB_NOTLS))
2458 txn->mt_u.reader = NULL; /* txn does not own reader */
2460 txn->mt_numdbs = 0; /* close nothing if called again */
2461 txn->mt_dbxs = NULL; /* mark txn as reset */
2463 mdb_cursors_close(txn, 0);
2465 if (!(env->me_flags & MDB_WRITEMAP)) {
2466 mdb_dlist_free(txn);
2468 mdb_midl_free(env->me_pghead);
2470 if (txn->mt_parent) {
2471 txn->mt_parent->mt_child = NULL;
2472 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2473 mdb_midl_free(txn->mt_free_pgs);
2474 mdb_midl_free(txn->mt_spill_pgs);
2475 free(txn->mt_u.dirty_list);
2479 if (mdb_midl_shrink(&txn->mt_free_pgs))
2480 env->me_free_pgs = txn->mt_free_pgs;
2481 env->me_pghead = NULL;
2485 /* The writer mutex was locked in mdb_txn_begin. */
2487 UNLOCK_MUTEX_W(env);
2492 mdb_txn_reset(MDB_txn *txn)
2497 /* This call is only valid for read-only txns */
2498 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2501 mdb_txn_reset0(txn, "reset");
2505 mdb_txn_abort(MDB_txn *txn)
2511 mdb_txn_abort(txn->mt_child);
2513 mdb_txn_reset0(txn, "abort");
2514 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2515 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2516 txn->mt_u.reader->mr_pid = 0;
2521 /** Save the freelist as of this transaction to the freeDB.
2522 * This changes the freelist. Keep trying until it stabilizes.
2525 mdb_freelist_save(MDB_txn *txn)
2527 /* env->me_pghead[] can grow and shrink during this call.
2528 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2529 * Page numbers cannot disappear from txn->mt_free_pgs[].
2532 MDB_env *env = txn->mt_env;
2533 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2534 txnid_t pglast = 0, head_id = 0;
2535 pgno_t freecnt = 0, *free_pgs, *mop;
2536 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2538 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2540 if (env->me_pghead) {
2541 /* Make sure first page of freeDB is touched and on freelist */
2542 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2543 if (rc && rc != MDB_NOTFOUND)
2547 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2548 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2549 ? SSIZE_MAX : maxfree_1pg;
2552 /* Come back here after each Put() in case freelist changed */
2557 /* If using records from freeDB which we have not yet
2558 * deleted, delete them and any we reserved for me_pghead.
2560 while (pglast < env->me_pglast) {
2561 rc = mdb_cursor_first(&mc, &key, NULL);
2564 pglast = head_id = *(txnid_t *)key.mv_data;
2565 total_room = head_room = 0;
2566 assert(pglast <= env->me_pglast);
2567 rc = mdb_cursor_del(&mc, 0);
2572 /* Save the IDL of pages freed by this txn, to a single record */
2573 if (freecnt < txn->mt_free_pgs[0]) {
2575 /* Make sure last page of freeDB is touched and on freelist */
2576 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2577 if (rc && rc != MDB_NOTFOUND)
2580 free_pgs = txn->mt_free_pgs;
2581 /* Write to last page of freeDB */
2582 key.mv_size = sizeof(txn->mt_txnid);
2583 key.mv_data = &txn->mt_txnid;
2585 freecnt = free_pgs[0];
2586 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2587 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2590 /* Retry if mt_free_pgs[] grew during the Put() */
2591 free_pgs = txn->mt_free_pgs;
2592 } while (freecnt < free_pgs[0]);
2593 mdb_midl_sort(free_pgs);
2594 memcpy(data.mv_data, free_pgs, data.mv_size);
2597 unsigned int i = free_pgs[0];
2598 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2599 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2601 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2607 mop = env->me_pghead;
2608 mop_len = mop ? mop[0] : 0;
2610 /* Reserve records for me_pghead[]. Split it if multi-page,
2611 * to avoid searching freeDB for a page range. Use keys in
2612 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2614 if (total_room >= mop_len) {
2615 if (total_room == mop_len || --more < 0)
2617 } else if (head_room >= maxfree_1pg && head_id > 1) {
2618 /* Keep current record (overflow page), add a new one */
2622 /* (Re)write {key = head_id, IDL length = head_room} */
2623 total_room -= head_room;
2624 head_room = mop_len - total_room;
2625 if (head_room > maxfree_1pg && head_id > 1) {
2626 /* Overflow multi-page for part of me_pghead */
2627 head_room /= head_id; /* amortize page sizes */
2628 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2629 } else if (head_room < 0) {
2630 /* Rare case, not bothering to delete this record */
2633 key.mv_size = sizeof(head_id);
2634 key.mv_data = &head_id;
2635 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2636 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2639 /* IDL is initially empty, zero out at least the length */
2640 pgs = (pgno_t *)data.mv_data;
2641 j = head_room > clean_limit ? head_room : 0;
2645 total_room += head_room;
2648 /* Fill in the reserved me_pghead records */
2654 rc = mdb_cursor_first(&mc, &key, &data);
2655 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2656 unsigned flags = MDB_CURRENT;
2657 txnid_t id = *(txnid_t *)key.mv_data;
2658 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2661 assert(len >= 0 && id <= env->me_pglast);
2663 if (len > mop_len) {
2665 data.mv_size = (len + 1) * sizeof(MDB_ID);
2668 data.mv_data = mop -= len;
2671 rc = mdb_cursor_put(&mc, &key, &data, flags);
2673 if (rc || !(mop_len -= len))
2680 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2681 * @param[in] txn the transaction that's being committed
2682 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2683 * @return 0 on success, non-zero on failure.
2686 mdb_page_flush(MDB_txn *txn, int keep)
2688 MDB_env *env = txn->mt_env;
2689 MDB_ID2L dl = txn->mt_u.dirty_list;
2690 unsigned psize = env->me_psize, j;
2691 int i, pagecount = dl[0].mid, rc;
2692 size_t size = 0, pos = 0;
2694 MDB_page *dp = NULL;
2698 struct iovec iov[MDB_COMMIT_PAGES];
2699 ssize_t wpos = 0, wsize = 0, wres;
2700 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2706 if (env->me_flags & MDB_WRITEMAP) {
2707 /* Clear dirty flags */
2708 while (++i <= pagecount) {
2710 /* Don't flush this page yet */
2711 if (dp->mp_flags & P_KEEP) {
2712 dp->mp_flags ^= P_KEEP;
2716 dp->mp_flags &= ~P_DIRTY;
2721 /* Write the pages */
2723 if (++i <= pagecount) {
2725 /* Don't flush this page yet */
2726 if (dp->mp_flags & P_KEEP) {
2727 dp->mp_flags ^= P_KEEP;
2732 /* clear dirty flag */
2733 dp->mp_flags &= ~P_DIRTY;
2736 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2741 /* Windows actually supports scatter/gather I/O, but only on
2742 * unbuffered file handles. Since we're relying on the OS page
2743 * cache for all our data, that's self-defeating. So we just
2744 * write pages one at a time. We use the ov structure to set
2745 * the write offset, to at least save the overhead of a Seek
2748 DPRINTF(("committing page %"Z"u", pgno));
2749 memset(&ov, 0, sizeof(ov));
2750 ov.Offset = pos & 0xffffffff;
2751 ov.OffsetHigh = pos >> 16 >> 16;
2752 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2754 DPRINTF(("WriteFile: %d", rc));
2758 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2759 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2761 /* Write previous page(s) */
2762 #ifdef MDB_USE_PWRITEV
2763 wres = pwritev(env->me_fd, iov, n, wpos);
2766 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2768 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2770 DPRINTF(("lseek: %s", strerror(rc)));
2773 wres = writev(env->me_fd, iov, n);
2776 if (wres != wsize) {
2779 DPRINTF(("Write error: %s", strerror(rc)));
2781 rc = EIO; /* TODO: Use which error code? */
2782 DPUTS("short write, filesystem full?");
2793 DPRINTF(("committing page %"Z"u", pgno));
2794 next_pos = pos + size;
2795 iov[n].iov_len = size;
2796 iov[n].iov_base = (char *)dp;
2802 for (i = keep; ++i <= pagecount; ) {
2804 /* This is a page we skipped above */
2807 dl[j].mid = dp->mp_pgno;
2810 mdb_dpage_free(env, dp);
2815 txn->mt_dirty_room += i - j;
2821 mdb_txn_commit(MDB_txn *txn)
2827 assert(txn != NULL);
2828 assert(txn->mt_env != NULL);
2830 if (txn->mt_child) {
2831 rc = mdb_txn_commit(txn->mt_child);
2832 txn->mt_child = NULL;
2839 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2840 mdb_dbis_update(txn, 1);
2841 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2846 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2847 DPUTS("error flag is set, can't commit");
2849 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2854 if (txn->mt_parent) {
2855 MDB_txn *parent = txn->mt_parent;
2858 unsigned x, y, len, ps_len;
2860 /* Append our free list to parent's */
2861 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2864 mdb_midl_free(txn->mt_free_pgs);
2865 /* Failures after this must either undo the changes
2866 * to the parent or set MDB_TXN_ERROR in the parent.
2869 parent->mt_next_pgno = txn->mt_next_pgno;
2870 parent->mt_flags = txn->mt_flags;
2872 /* Merge our cursors into parent's and close them */
2873 mdb_cursors_close(txn, 1);
2875 /* Update parent's DB table. */
2876 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2877 parent->mt_numdbs = txn->mt_numdbs;
2878 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2879 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2880 for (i=2; i<txn->mt_numdbs; i++) {
2881 /* preserve parent's DB_NEW status */
2882 x = parent->mt_dbflags[i] & DB_NEW;
2883 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2886 dst = parent->mt_u.dirty_list;
2887 src = txn->mt_u.dirty_list;
2888 /* Remove anything in our dirty list from parent's spill list */
2889 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2891 pspill[0] = (pgno_t)-1;
2892 /* Mark our dirty pages as deleted in parent spill list */
2893 for (i=0, len=src[0].mid; ++i <= len; ) {
2894 MDB_ID pn = src[i].mid << 1;
2895 while (pn > pspill[x])
2897 if (pn == pspill[x]) {
2902 /* Squash deleted pagenums if we deleted any */
2903 for (x=y; ++x <= ps_len; )
2904 if (!(pspill[x] & 1))
2905 pspill[++y] = pspill[x];
2909 /* Find len = length of merging our dirty list with parent's */
2911 dst[0].mid = 0; /* simplify loops */
2912 if (parent->mt_parent) {
2913 len = x + src[0].mid;
2914 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2915 for (i = x; y && i; y--) {
2916 pgno_t yp = src[y].mid;
2917 while (yp < dst[i].mid)
2919 if (yp == dst[i].mid) {
2924 } else { /* Simplify the above for single-ancestor case */
2925 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2927 /* Merge our dirty list with parent's */
2929 for (i = len; y; dst[i--] = src[y--]) {
2930 pgno_t yp = src[y].mid;
2931 while (yp < dst[x].mid)
2932 dst[i--] = dst[x--];
2933 if (yp == dst[x].mid)
2934 free(dst[x--].mptr);
2938 free(txn->mt_u.dirty_list);
2939 parent->mt_dirty_room = txn->mt_dirty_room;
2940 if (txn->mt_spill_pgs) {
2941 if (parent->mt_spill_pgs) {
2942 /* TODO: Prevent failure here, so parent does not fail */
2943 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2945 parent->mt_flags |= MDB_TXN_ERROR;
2946 mdb_midl_free(txn->mt_spill_pgs);
2947 mdb_midl_sort(parent->mt_spill_pgs);
2949 parent->mt_spill_pgs = txn->mt_spill_pgs;
2953 parent->mt_child = NULL;
2954 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2959 if (txn != env->me_txn) {
2960 DPUTS("attempt to commit unknown transaction");
2965 mdb_cursors_close(txn, 0);
2967 if (!txn->mt_u.dirty_list[0].mid &&
2968 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2971 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2972 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2974 /* Update DB root pointers */
2975 if (txn->mt_numdbs > 2) {
2979 data.mv_size = sizeof(MDB_db);
2981 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2982 for (i = 2; i < txn->mt_numdbs; i++) {
2983 if (txn->mt_dbflags[i] & DB_DIRTY) {
2984 data.mv_data = &txn->mt_dbs[i];
2985 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2992 rc = mdb_freelist_save(txn);
2996 mdb_midl_free(env->me_pghead);
2997 env->me_pghead = NULL;
2998 if (mdb_midl_shrink(&txn->mt_free_pgs))
2999 env->me_free_pgs = txn->mt_free_pgs;
3005 if ((rc = mdb_page_flush(txn, 0)) ||
3006 (rc = mdb_env_sync(env, 0)) ||
3007 (rc = mdb_env_write_meta(txn)))
3013 mdb_dbis_update(txn, 1);
3016 UNLOCK_MUTEX_W(env);
3026 /** Read the environment parameters of a DB environment before
3027 * mapping it into memory.
3028 * @param[in] env the environment handle
3029 * @param[out] meta address of where to store the meta information
3030 * @return 0 on success, non-zero on failure.
3033 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3039 enum { Size = sizeof(pbuf) };
3041 /* We don't know the page size yet, so use a minimum value.
3042 * Read both meta pages so we can use the latest one.
3045 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3049 memset(&ov, 0, sizeof(ov));
3051 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3052 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3055 rc = pread(env->me_fd, &pbuf, Size, off);
3058 if (rc == 0 && off == 0)
3060 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3061 DPRINTF(("read: %s", mdb_strerror(rc)));
3065 p = (MDB_page *)&pbuf;
3067 if (!F_ISSET(p->mp_flags, P_META)) {
3068 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3073 if (m->mm_magic != MDB_MAGIC) {
3074 DPUTS("meta has invalid magic");
3078 if (m->mm_version != MDB_DATA_VERSION) {
3079 DPRINTF(("database is version %u, expected version %u",
3080 m->mm_version, MDB_DATA_VERSION));
3081 return MDB_VERSION_MISMATCH;
3084 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3090 /** Write the environment parameters of a freshly created DB environment.
3091 * @param[in] env the environment handle
3092 * @param[out] meta address of where to store the meta information
3093 * @return 0 on success, non-zero on failure.
3096 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3104 memset(&ov, 0, sizeof(ov));
3105 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3107 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3110 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3111 len = pwrite(fd, ptr, size, pos); \
3112 rc = (len >= 0); } while(0)
3115 DPUTS("writing new meta page");
3117 psize = env->me_psize;
3119 meta->mm_magic = MDB_MAGIC;
3120 meta->mm_version = MDB_DATA_VERSION;
3121 meta->mm_mapsize = env->me_mapsize;
3122 meta->mm_psize = psize;
3123 meta->mm_last_pg = 1;
3124 meta->mm_flags = env->me_flags & 0xffff;
3125 meta->mm_flags |= MDB_INTEGERKEY;
3126 meta->mm_dbs[0].md_root = P_INVALID;
3127 meta->mm_dbs[1].md_root = P_INVALID;
3129 p = calloc(2, psize);
3131 p->mp_flags = P_META;
3132 *(MDB_meta *)METADATA(p) = *meta;
3134 q = (MDB_page *)((char *)p + psize);
3136 q->mp_flags = P_META;
3137 *(MDB_meta *)METADATA(q) = *meta;
3139 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3142 else if ((unsigned) len == psize * 2)
3150 /** Update the environment info to commit a transaction.
3151 * @param[in] txn the transaction that's being committed
3152 * @return 0 on success, non-zero on failure.
3155 mdb_env_write_meta(MDB_txn *txn)
3158 MDB_meta meta, metab, *mp;
3160 int rc, len, toggle;
3169 assert(txn != NULL);
3170 assert(txn->mt_env != NULL);
3172 toggle = txn->mt_txnid & 1;
3173 DPRINTF(("writing meta page %d for root page %"Z"u",
3174 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3177 mp = env->me_metas[toggle];
3179 if (env->me_flags & MDB_WRITEMAP) {
3180 /* Persist any increases of mapsize config */
3181 if (env->me_mapsize > mp->mm_mapsize)
3182 mp->mm_mapsize = env->me_mapsize;
3183 mp->mm_dbs[0] = txn->mt_dbs[0];
3184 mp->mm_dbs[1] = txn->mt_dbs[1];
3185 mp->mm_last_pg = txn->mt_next_pgno - 1;
3186 mp->mm_txnid = txn->mt_txnid;
3187 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3188 unsigned meta_size = env->me_psize;
3189 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3192 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3193 if (meta_size < env->me_os_psize)
3194 meta_size += meta_size;
3199 if (MDB_MSYNC(ptr, meta_size, rc)) {
3206 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3207 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3209 ptr = (char *)&meta;
3210 if (env->me_mapsize > mp->mm_mapsize) {
3211 /* Persist any increases of mapsize config */
3212 meta.mm_mapsize = env->me_mapsize;
3213 off = offsetof(MDB_meta, mm_mapsize);
3215 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3217 len = sizeof(MDB_meta) - off;
3220 meta.mm_dbs[0] = txn->mt_dbs[0];
3221 meta.mm_dbs[1] = txn->mt_dbs[1];
3222 meta.mm_last_pg = txn->mt_next_pgno - 1;
3223 meta.mm_txnid = txn->mt_txnid;
3226 off += env->me_psize;
3229 /* Write to the SYNC fd */
3230 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3231 env->me_fd : env->me_mfd;
3234 memset(&ov, 0, sizeof(ov));
3236 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3240 rc = pwrite(mfd, ptr, len, off);
3243 rc = rc < 0 ? ErrCode() : EIO;
3244 DPUTS("write failed, disk error?");
3245 /* On a failure, the pagecache still contains the new data.
3246 * Write some old data back, to prevent it from being used.
3247 * Use the non-SYNC fd; we know it will fail anyway.
3249 meta.mm_last_pg = metab.mm_last_pg;
3250 meta.mm_txnid = metab.mm_txnid;
3252 memset(&ov, 0, sizeof(ov));
3254 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3256 r2 = pwrite(env->me_fd, ptr, len, off);
3257 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3260 env->me_flags |= MDB_FATAL_ERROR;
3264 /* Memory ordering issues are irrelevant; since the entire writer
3265 * is wrapped by wmutex, all of these changes will become visible
3266 * after the wmutex is unlocked. Since the DB is multi-version,
3267 * readers will get consistent data regardless of how fresh or
3268 * how stale their view of these values is.
3271 env->me_txns->mti_txnid = txn->mt_txnid;
3276 /** Check both meta pages to see which one is newer.
3277 * @param[in] env the environment handle
3278 * @return meta toggle (0 or 1).
3281 mdb_env_pick_meta(const MDB_env *env)
3283 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3287 mdb_env_create(MDB_env **env)
3291 e = calloc(1, sizeof(MDB_env));
3295 e->me_maxreaders = DEFAULT_READERS;
3296 e->me_maxdbs = e->me_numdbs = 2;
3297 e->me_fd = INVALID_HANDLE_VALUE;
3298 e->me_lfd = INVALID_HANDLE_VALUE;
3299 e->me_mfd = INVALID_HANDLE_VALUE;
3300 #ifdef MDB_USE_POSIX_SEM
3301 e->me_rmutex = SEM_FAILED;
3302 e->me_wmutex = SEM_FAILED;
3304 e->me_pid = getpid();
3305 GET_PAGESIZE(e->me_os_psize);
3306 VGMEMP_CREATE(e,0,0);
3312 mdb_env_map(MDB_env *env, void *addr, int newsize)
3315 unsigned int flags = env->me_flags;
3319 LONG sizelo, sizehi;
3320 sizelo = env->me_mapsize & 0xffffffff;
3321 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3323 /* Windows won't create mappings for zero length files.
3324 * Just allocate the maxsize right now.
3327 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3328 || !SetEndOfFile(env->me_fd)
3329 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3332 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3333 PAGE_READWRITE : PAGE_READONLY,
3334 sizehi, sizelo, NULL);
3337 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3338 FILE_MAP_WRITE : FILE_MAP_READ,
3339 0, 0, env->me_mapsize, addr);
3340 rc = env->me_map ? 0 : ErrCode();
3345 int prot = PROT_READ;
3346 if (flags & MDB_WRITEMAP) {
3348 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3351 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3353 if (env->me_map == MAP_FAILED) {
3358 if (flags & MDB_NORDAHEAD) {
3359 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3361 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3363 #ifdef POSIX_MADV_RANDOM
3364 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3365 #endif /* POSIX_MADV_RANDOM */
3366 #endif /* MADV_RANDOM */
3370 /* Can happen because the address argument to mmap() is just a
3371 * hint. mmap() can pick another, e.g. if the range is in use.
3372 * The MAP_FIXED flag would prevent that, but then mmap could
3373 * instead unmap existing pages to make room for the new map.
3375 if (addr && env->me_map != addr)
3376 return EBUSY; /* TODO: Make a new MDB_* error code? */
3378 p = (MDB_page *)env->me_map;
3379 env->me_metas[0] = METADATA(p);
3380 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3386 mdb_env_set_mapsize(MDB_env *env, size_t size)
3388 /* If env is already open, caller is responsible for making
3389 * sure there are no active txns.
3397 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3398 else if (size < env->me_mapsize) {
3399 /* If the configured size is smaller, make sure it's
3400 * still big enough. Silently round up to minimum if not.
3402 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3406 munmap(env->me_map, env->me_mapsize);
3407 env->me_mapsize = size;
3408 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3409 rc = mdb_env_map(env, old, 1);
3413 env->me_mapsize = size;
3415 env->me_maxpg = env->me_mapsize / env->me_psize;
3420 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3424 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3429 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3431 if (env->me_map || readers < 1)
3433 env->me_maxreaders = readers;
3438 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3440 if (!env || !readers)
3442 *readers = env->me_maxreaders;
3446 /** Further setup required for opening an MDB environment
3449 mdb_env_open2(MDB_env *env)
3451 unsigned int flags = env->me_flags;
3452 int i, newenv = 0, rc;
3456 /* See if we should use QueryLimited */
3458 if ((rc & 0xff) > 5)
3459 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3461 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3464 memset(&meta, 0, sizeof(meta));
3466 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3469 DPUTS("new mdbenv");
3471 env->me_psize = env->me_os_psize;
3472 if (env->me_psize > MAX_PAGESIZE)
3473 env->me_psize = MAX_PAGESIZE;
3475 env->me_psize = meta.mm_psize;
3478 /* Was a mapsize configured? */
3479 if (!env->me_mapsize) {
3480 /* If this is a new environment, take the default,
3481 * else use the size recorded in the existing env.
3483 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3484 } else if (env->me_mapsize < meta.mm_mapsize) {
3485 /* If the configured size is smaller, make sure it's
3486 * still big enough. Silently round up to minimum if not.
3488 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3489 if (env->me_mapsize < minsize)
3490 env->me_mapsize = minsize;
3493 rc = mdb_env_map(env, meta.mm_address, newenv);
3498 if (flags & MDB_FIXEDMAP)
3499 meta.mm_address = env->me_map;
3500 i = mdb_env_init_meta(env, &meta);
3501 if (i != MDB_SUCCESS) {
3506 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3507 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3509 #if !(MDB_MAXKEYSIZE)
3510 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3512 env->me_maxpg = env->me_mapsize / env->me_psize;
3516 int toggle = mdb_env_pick_meta(env);
3517 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3519 DPRINTF(("opened database version %u, pagesize %u",
3520 env->me_metas[0]->mm_version, env->me_psize));
3521 DPRINTF(("using meta page %d", toggle));
3522 DPRINTF(("depth: %u", db->md_depth));
3523 DPRINTF(("entries: %"Z"u", db->md_entries));
3524 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3525 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3526 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3527 DPRINTF(("root: %"Z"u", db->md_root));
3535 /** Release a reader thread's slot in the reader lock table.
3536 * This function is called automatically when a thread exits.
3537 * @param[in] ptr This points to the slot in the reader lock table.
3540 mdb_env_reader_dest(void *ptr)
3542 MDB_reader *reader = ptr;
3548 /** Junk for arranging thread-specific callbacks on Windows. This is
3549 * necessarily platform and compiler-specific. Windows supports up
3550 * to 1088 keys. Let's assume nobody opens more than 64 environments
3551 * in a single process, for now. They can override this if needed.
3553 #ifndef MAX_TLS_KEYS
3554 #define MAX_TLS_KEYS 64
3556 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3557 static int mdb_tls_nkeys;
3559 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3563 case DLL_PROCESS_ATTACH: break;
3564 case DLL_THREAD_ATTACH: break;
3565 case DLL_THREAD_DETACH:
3566 for (i=0; i<mdb_tls_nkeys; i++) {
3567 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3568 mdb_env_reader_dest(r);
3571 case DLL_PROCESS_DETACH: break;
3576 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3578 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3582 /* Force some symbol references.
3583 * _tls_used forces the linker to create the TLS directory if not already done
3584 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3586 #pragma comment(linker, "/INCLUDE:_tls_used")
3587 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3588 #pragma const_seg(".CRT$XLB")
3589 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3590 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3593 #pragma comment(linker, "/INCLUDE:__tls_used")
3594 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3595 #pragma data_seg(".CRT$XLB")
3596 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3598 #endif /* WIN 32/64 */
3599 #endif /* !__GNUC__ */
3602 /** Downgrade the exclusive lock on the region back to shared */
3604 mdb_env_share_locks(MDB_env *env, int *excl)
3606 int rc = 0, toggle = mdb_env_pick_meta(env);
3608 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3613 /* First acquire a shared lock. The Unlock will
3614 * then release the existing exclusive lock.
3616 memset(&ov, 0, sizeof(ov));
3617 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3620 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3626 struct flock lock_info;
3627 /* The shared lock replaces the existing lock */
3628 memset((void *)&lock_info, 0, sizeof(lock_info));
3629 lock_info.l_type = F_RDLCK;
3630 lock_info.l_whence = SEEK_SET;
3631 lock_info.l_start = 0;
3632 lock_info.l_len = 1;
3633 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3634 (rc = ErrCode()) == EINTR) ;
3635 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3642 /** Try to get exlusive lock, otherwise shared.
3643 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3646 mdb_env_excl_lock(MDB_env *env, int *excl)
3650 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3654 memset(&ov, 0, sizeof(ov));
3655 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3662 struct flock lock_info;
3663 memset((void *)&lock_info, 0, sizeof(lock_info));
3664 lock_info.l_type = F_WRLCK;
3665 lock_info.l_whence = SEEK_SET;
3666 lock_info.l_start = 0;
3667 lock_info.l_len = 1;
3668 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3669 (rc = ErrCode()) == EINTR) ;
3673 # ifdef MDB_USE_POSIX_SEM
3674 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3677 lock_info.l_type = F_RDLCK;
3678 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3679 (rc = ErrCode()) == EINTR) ;
3689 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3691 * @(#) $Revision: 5.1 $
3692 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3693 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3695 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3699 * Please do not copyright this code. This code is in the public domain.
3701 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3702 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3703 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3704 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3705 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3706 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3707 * PERFORMANCE OF THIS SOFTWARE.
3710 * chongo <Landon Curt Noll> /\oo/\
3711 * http://www.isthe.com/chongo/
3713 * Share and Enjoy! :-)
3716 typedef unsigned long long mdb_hash_t;
3717 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3719 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3720 * @param[in] val value to hash
3721 * @param[in] hval initial value for hash
3722 * @return 64 bit hash
3724 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3725 * hval arg on the first call.
3728 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3730 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3731 unsigned char *end = s + val->mv_size;
3733 * FNV-1a hash each octet of the string
3736 /* xor the bottom with the current octet */
3737 hval ^= (mdb_hash_t)*s++;
3739 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3740 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3741 (hval << 7) + (hval << 8) + (hval << 40);
3743 /* return our new hash value */
3747 /** Hash the string and output the encoded hash.
3748 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3749 * very short name limits. We don't care about the encoding being reversible,
3750 * we just want to preserve as many bits of the input as possible in a
3751 * small printable string.
3752 * @param[in] str string to hash
3753 * @param[out] encbuf an array of 11 chars to hold the hash
3755 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3758 mdb_pack85(unsigned long l, char *out)
3762 for (i=0; i<5; i++) {
3763 *out++ = mdb_a85[l % 85];
3769 mdb_hash_enc(MDB_val *val, char *encbuf)
3771 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3773 mdb_pack85(h, encbuf);
3774 mdb_pack85(h>>32, encbuf+5);
3779 /** Open and/or initialize the lock region for the environment.
3780 * @param[in] env The MDB environment.
3781 * @param[in] lpath The pathname of the file used for the lock region.
3782 * @param[in] mode The Unix permissions for the file, if we create it.
3783 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3784 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3785 * @return 0 on success, non-zero on failure.
3788 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3791 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3793 # define MDB_ERRCODE_ROFS EROFS
3794 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3795 # define MDB_CLOEXEC O_CLOEXEC
3798 # define MDB_CLOEXEC 0
3805 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3806 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3807 FILE_ATTRIBUTE_NORMAL, NULL);
3809 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3811 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3813 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3818 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3819 /* Lose record locks when exec*() */
3820 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3821 fcntl(env->me_lfd, F_SETFD, fdflags);
3824 if (!(env->me_flags & MDB_NOTLS)) {
3825 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3828 env->me_flags |= MDB_ENV_TXKEY;
3830 /* Windows TLS callbacks need help finding their TLS info. */
3831 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3835 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3839 /* Try to get exclusive lock. If we succeed, then
3840 * nobody is using the lock region and we should initialize it.
3842 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3845 size = GetFileSize(env->me_lfd, NULL);
3847 size = lseek(env->me_lfd, 0, SEEK_END);
3848 if (size == -1) goto fail_errno;
3850 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3851 if (size < rsize && *excl > 0) {
3853 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3854 || !SetEndOfFile(env->me_lfd))
3857 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3861 size = rsize - sizeof(MDB_txninfo);
3862 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3867 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3869 if (!mh) goto fail_errno;
3870 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3872 if (!env->me_txns) goto fail_errno;
3874 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3876 if (m == MAP_FAILED) goto fail_errno;
3882 BY_HANDLE_FILE_INFORMATION stbuf;
3891 if (!mdb_sec_inited) {
3892 InitializeSecurityDescriptor(&mdb_null_sd,
3893 SECURITY_DESCRIPTOR_REVISION);
3894 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3895 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3896 mdb_all_sa.bInheritHandle = FALSE;
3897 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3900 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3901 idbuf.volume = stbuf.dwVolumeSerialNumber;
3902 idbuf.nhigh = stbuf.nFileIndexHigh;
3903 idbuf.nlow = stbuf.nFileIndexLow;
3904 val.mv_data = &idbuf;
3905 val.mv_size = sizeof(idbuf);
3906 mdb_hash_enc(&val, encbuf);
3907 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3908 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3909 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3910 if (!env->me_rmutex) goto fail_errno;
3911 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3912 if (!env->me_wmutex) goto fail_errno;
3913 #elif defined(MDB_USE_POSIX_SEM)
3922 #if defined(__NetBSD__)
3923 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3925 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3926 idbuf.dev = stbuf.st_dev;
3927 idbuf.ino = stbuf.st_ino;
3928 val.mv_data = &idbuf;
3929 val.mv_size = sizeof(idbuf);
3930 mdb_hash_enc(&val, encbuf);
3931 #ifdef MDB_SHORT_SEMNAMES
3932 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3934 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3935 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3936 /* Clean up after a previous run, if needed: Try to
3937 * remove both semaphores before doing anything else.
3939 sem_unlink(env->me_txns->mti_rmname);
3940 sem_unlink(env->me_txns->mti_wmname);
3941 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3942 O_CREAT|O_EXCL, mode, 1);
3943 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3944 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3945 O_CREAT|O_EXCL, mode, 1);
3946 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3947 #else /* MDB_USE_POSIX_SEM */
3948 pthread_mutexattr_t mattr;
3950 if ((rc = pthread_mutexattr_init(&mattr))
3951 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3952 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3953 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3955 pthread_mutexattr_destroy(&mattr);
3956 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3958 env->me_txns->mti_magic = MDB_MAGIC;
3959 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3960 env->me_txns->mti_txnid = 0;
3961 env->me_txns->mti_numreaders = 0;
3964 if (env->me_txns->mti_magic != MDB_MAGIC) {
3965 DPUTS("lock region has invalid magic");
3969 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3970 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3971 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3972 rc = MDB_VERSION_MISMATCH;
3976 if (rc && rc != EACCES && rc != EAGAIN) {
3980 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3981 if (!env->me_rmutex) goto fail_errno;
3982 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3983 if (!env->me_wmutex) goto fail_errno;
3984 #elif defined(MDB_USE_POSIX_SEM)
3985 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3986 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3987 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3988 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3999 /** The name of the lock file in the DB environment */
4000 #define LOCKNAME "/lock.mdb"
4001 /** The name of the data file in the DB environment */
4002 #define DATANAME "/data.mdb"
4003 /** The suffix of the lock file when no subdir is used */
4004 #define LOCKSUFF "-lock"
4005 /** Only a subset of the @ref mdb_env flags can be changed
4006 * at runtime. Changing other flags requires closing the
4007 * environment and re-opening it with the new flags.
4009 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4010 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4011 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4014 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4016 int oflags, rc, len, excl = -1;
4017 char *lpath, *dpath;
4019 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4023 if (flags & MDB_NOSUBDIR) {
4024 rc = len + sizeof(LOCKSUFF) + len + 1;
4026 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4031 if (flags & MDB_NOSUBDIR) {
4032 dpath = lpath + len + sizeof(LOCKSUFF);
4033 sprintf(lpath, "%s" LOCKSUFF, path);
4034 strcpy(dpath, path);
4036 dpath = lpath + len + sizeof(LOCKNAME);
4037 sprintf(lpath, "%s" LOCKNAME, path);
4038 sprintf(dpath, "%s" DATANAME, path);
4042 flags |= env->me_flags;
4043 if (flags & MDB_RDONLY) {
4044 /* silently ignore WRITEMAP when we're only getting read access */
4045 flags &= ~MDB_WRITEMAP;
4047 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4048 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4051 env->me_flags = flags |= MDB_ENV_ACTIVE;
4055 env->me_path = strdup(path);
4056 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4057 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4058 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4063 /* For RDONLY, get lockfile after we know datafile exists */
4064 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4065 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4071 if (F_ISSET(flags, MDB_RDONLY)) {
4072 oflags = GENERIC_READ;
4073 len = OPEN_EXISTING;
4075 oflags = GENERIC_READ|GENERIC_WRITE;
4078 mode = FILE_ATTRIBUTE_NORMAL;
4079 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4080 NULL, len, mode, NULL);
4082 if (F_ISSET(flags, MDB_RDONLY))
4085 oflags = O_RDWR | O_CREAT;
4087 env->me_fd = open(dpath, oflags, mode);
4089 if (env->me_fd == INVALID_HANDLE_VALUE) {
4094 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4095 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4100 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4101 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4102 env->me_mfd = env->me_fd;
4104 /* Synchronous fd for meta writes. Needed even with
4105 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4108 len = OPEN_EXISTING;
4109 env->me_mfd = CreateFile(dpath, oflags,
4110 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4111 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4114 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4116 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4121 DPRINTF(("opened dbenv %p", (void *) env));
4123 rc = mdb_env_share_locks(env, &excl);
4127 if (!((flags & MDB_RDONLY) ||
4128 (env->me_pbuf = calloc(1, env->me_psize))))
4134 mdb_env_close0(env, excl);
4140 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4142 mdb_env_close0(MDB_env *env, int excl)
4146 if (!(env->me_flags & MDB_ENV_ACTIVE))
4149 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4150 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4151 free(env->me_dbxs[i].md_name.mv_data);
4154 free(env->me_dbflags);
4157 free(env->me_dirty_list);
4158 mdb_midl_free(env->me_free_pgs);
4160 if (env->me_flags & MDB_ENV_TXKEY) {
4161 pthread_key_delete(env->me_txkey);
4163 /* Delete our key from the global list */
4164 for (i=0; i<mdb_tls_nkeys; i++)
4165 if (mdb_tls_keys[i] == env->me_txkey) {
4166 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4174 munmap(env->me_map, env->me_mapsize);
4176 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4177 (void) close(env->me_mfd);
4178 if (env->me_fd != INVALID_HANDLE_VALUE)
4179 (void) close(env->me_fd);
4181 MDB_PID_T pid = env->me_pid;
4182 /* Clearing readers is done in this function because
4183 * me_txkey with its destructor must be disabled first.
4185 for (i = env->me_numreaders; --i >= 0; )
4186 if (env->me_txns->mti_readers[i].mr_pid == pid)
4187 env->me_txns->mti_readers[i].mr_pid = 0;
4189 if (env->me_rmutex) {
4190 CloseHandle(env->me_rmutex);
4191 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4193 /* Windows automatically destroys the mutexes when
4194 * the last handle closes.
4196 #elif defined(MDB_USE_POSIX_SEM)
4197 if (env->me_rmutex != SEM_FAILED) {
4198 sem_close(env->me_rmutex);
4199 if (env->me_wmutex != SEM_FAILED)
4200 sem_close(env->me_wmutex);
4201 /* If we have the filelock: If we are the
4202 * only remaining user, clean up semaphores.
4205 mdb_env_excl_lock(env, &excl);
4207 sem_unlink(env->me_txns->mti_rmname);
4208 sem_unlink(env->me_txns->mti_wmname);
4212 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4214 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4217 /* Unlock the lockfile. Windows would have unlocked it
4218 * after closing anyway, but not necessarily at once.
4220 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4223 (void) close(env->me_lfd);
4226 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4230 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4232 MDB_txn *txn = NULL;
4238 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4242 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4245 /* Do the lock/unlock of the reader mutex before starting the
4246 * write txn. Otherwise other read txns could block writers.
4248 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4253 /* We must start the actual read txn after blocking writers */
4254 mdb_txn_reset0(txn, "reset-stage1");
4256 /* Temporarily block writers until we snapshot the meta pages */
4259 rc = mdb_txn_renew0(txn);
4261 UNLOCK_MUTEX_W(env);
4266 wsize = env->me_psize * 2;
4270 DO_WRITE(rc, fd, ptr, w2, len);
4274 } else if (len > 0) {
4280 /* Non-blocking or async handles are not supported */
4286 UNLOCK_MUTEX_W(env);
4291 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4293 if (wsize > MAX_WRITE)
4297 DO_WRITE(rc, fd, ptr, w2, len);
4301 } else if (len > 0) {
4318 mdb_env_copy(MDB_env *env, const char *path)
4322 HANDLE newfd = INVALID_HANDLE_VALUE;
4324 if (env->me_flags & MDB_NOSUBDIR) {
4325 lpath = (char *)path;
4328 len += sizeof(DATANAME);
4329 lpath = malloc(len);
4332 sprintf(lpath, "%s" DATANAME, path);
4335 /* The destination path must exist, but the destination file must not.
4336 * We don't want the OS to cache the writes, since the source data is
4337 * already in the OS cache.
4340 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4341 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4343 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4345 if (newfd == INVALID_HANDLE_VALUE) {
4351 /* Set O_DIRECT if the file system supports it */
4352 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4353 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4355 #ifdef F_NOCACHE /* __APPLE__ */
4356 rc = fcntl(newfd, F_NOCACHE, 1);
4363 rc = mdb_env_copyfd(env, newfd);
4366 if (!(env->me_flags & MDB_NOSUBDIR))
4368 if (newfd != INVALID_HANDLE_VALUE)
4369 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4376 mdb_env_close(MDB_env *env)
4383 VGMEMP_DESTROY(env);
4384 while ((dp = env->me_dpages) != NULL) {
4385 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4386 env->me_dpages = dp->mp_next;
4390 mdb_env_close0(env, 0);
4394 /** Compare two items pointing at aligned size_t's */
4396 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4398 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4399 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4402 /** Compare two items pointing at aligned unsigned int's */
4404 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4406 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4407 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4410 /** Compare two items pointing at unsigned ints of unknown alignment.
4411 * Nodes and keys are guaranteed to be 2-byte aligned.
4414 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4416 #if BYTE_ORDER == LITTLE_ENDIAN
4417 unsigned short *u, *c;
4420 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4421 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4424 } while(!x && u > (unsigned short *)a->mv_data);
4427 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4431 /** Compare two items lexically */
4433 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4440 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4446 diff = memcmp(a->mv_data, b->mv_data, len);
4447 return diff ? diff : len_diff<0 ? -1 : len_diff;
4450 /** Compare two items in reverse byte order */
4452 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4454 const unsigned char *p1, *p2, *p1_lim;
4458 p1_lim = (const unsigned char *)a->mv_data;
4459 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4460 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4462 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4468 while (p1 > p1_lim) {
4469 diff = *--p1 - *--p2;
4473 return len_diff<0 ? -1 : len_diff;
4476 /** Search for key within a page, using binary search.
4477 * Returns the smallest entry larger or equal to the key.
4478 * If exactp is non-null, stores whether the found entry was an exact match
4479 * in *exactp (1 or 0).
4480 * Updates the cursor index with the index of the found entry.
4481 * If no entry larger or equal to the key is found, returns NULL.
4484 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4486 unsigned int i = 0, nkeys;
4489 MDB_page *mp = mc->mc_pg[mc->mc_top];
4490 MDB_node *node = NULL;
4495 nkeys = NUMKEYS(mp);
4500 COPY_PGNO(pgno, mp->mp_pgno);
4501 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4502 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4509 low = IS_LEAF(mp) ? 0 : 1;
4511 cmp = mc->mc_dbx->md_cmp;
4513 /* Branch pages have no data, so if using integer keys,
4514 * alignment is guaranteed. Use faster mdb_cmp_int.
4516 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4517 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4524 nodekey.mv_size = mc->mc_db->md_pad;
4525 node = NODEPTR(mp, 0); /* fake */
4526 while (low <= high) {
4527 i = (low + high) >> 1;
4528 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4529 rc = cmp(key, &nodekey);
4530 DPRINTF(("found leaf index %u [%s], rc = %i",
4531 i, DKEY(&nodekey), rc));
4540 while (low <= high) {
4541 i = (low + high) >> 1;
4543 node = NODEPTR(mp, i);
4544 nodekey.mv_size = NODEKSZ(node);
4545 nodekey.mv_data = NODEKEY(node);
4547 rc = cmp(key, &nodekey);
4550 DPRINTF(("found leaf index %u [%s], rc = %i",
4551 i, DKEY(&nodekey), rc));
4553 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4554 i, DKEY(&nodekey), NODEPGNO(node), rc));
4565 if (rc > 0) { /* Found entry is less than the key. */
4566 i++; /* Skip to get the smallest entry larger than key. */
4568 node = NODEPTR(mp, i);
4571 *exactp = (rc == 0);
4572 /* store the key index */
4573 mc->mc_ki[mc->mc_top] = i;
4575 /* There is no entry larger or equal to the key. */
4578 /* nodeptr is fake for LEAF2 */
4584 mdb_cursor_adjust(MDB_cursor *mc, func)
4588 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4589 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4596 /** Pop a page off the top of the cursor's stack. */
4598 mdb_cursor_pop(MDB_cursor *mc)
4602 MDB_page *top = mc->mc_pg[mc->mc_top];
4608 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4609 DDBI(mc), (void *) mc));
4613 /** Push a page onto the top of the cursor's stack. */
4615 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4617 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4618 DDBI(mc), (void *) mc));
4620 if (mc->mc_snum >= CURSOR_STACK) {
4621 assert(mc->mc_snum < CURSOR_STACK);
4622 return MDB_CURSOR_FULL;
4625 mc->mc_top = mc->mc_snum++;
4626 mc->mc_pg[mc->mc_top] = mp;
4627 mc->mc_ki[mc->mc_top] = 0;
4632 /** Find the address of the page corresponding to a given page number.
4633 * @param[in] txn the transaction for this access.
4634 * @param[in] pgno the page number for the page to retrieve.
4635 * @param[out] ret address of a pointer where the page's address will be stored.
4636 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4637 * @return 0 on success, non-zero on failure.
4640 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4642 MDB_env *env = txn->mt_env;
4646 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4650 MDB_ID2L dl = tx2->mt_u.dirty_list;
4652 /* Spilled pages were dirtied in this txn and flushed
4653 * because the dirty list got full. Bring this page
4654 * back in from the map (but don't unspill it here,
4655 * leave that unless page_touch happens again).
4657 if (tx2->mt_spill_pgs) {
4658 MDB_ID pn = pgno << 1;
4659 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4660 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4661 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4666 unsigned x = mdb_mid2l_search(dl, pgno);
4667 if (x <= dl[0].mid && dl[x].mid == pgno) {
4673 } while ((tx2 = tx2->mt_parent) != NULL);
4676 if (pgno < txn->mt_next_pgno) {
4678 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4680 DPRINTF(("page %"Z"u not found", pgno));
4682 return MDB_PAGE_NOTFOUND;
4692 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4693 * The cursor is at the root page, set up the rest of it.
4696 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4698 MDB_page *mp = mc->mc_pg[mc->mc_top];
4702 while (IS_BRANCH(mp)) {
4706 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4707 assert(NUMKEYS(mp) > 1);
4708 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4710 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4712 if (flags & MDB_PS_LAST)
4713 i = NUMKEYS(mp) - 1;
4716 node = mdb_node_search(mc, key, &exact);
4718 i = NUMKEYS(mp) - 1;
4720 i = mc->mc_ki[mc->mc_top];
4726 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4729 assert(i < NUMKEYS(mp));
4730 node = NODEPTR(mp, i);
4732 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4735 mc->mc_ki[mc->mc_top] = i;
4736 if ((rc = mdb_cursor_push(mc, mp)))
4739 if (flags & MDB_PS_MODIFY) {
4740 if ((rc = mdb_page_touch(mc)) != 0)
4742 mp = mc->mc_pg[mc->mc_top];
4747 DPRINTF(("internal error, index points to a %02X page!?",
4749 return MDB_CORRUPTED;
4752 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4753 key ? DKEY(key) : "null"));
4754 mc->mc_flags |= C_INITIALIZED;
4755 mc->mc_flags &= ~C_EOF;
4760 /** Search for the lowest key under the current branch page.
4761 * This just bypasses a NUMKEYS check in the current page
4762 * before calling mdb_page_search_root(), because the callers
4763 * are all in situations where the current page is known to
4767 mdb_page_search_lowest(MDB_cursor *mc)
4769 MDB_page *mp = mc->mc_pg[mc->mc_top];
4770 MDB_node *node = NODEPTR(mp, 0);
4773 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4776 mc->mc_ki[mc->mc_top] = 0;
4777 if ((rc = mdb_cursor_push(mc, mp)))
4779 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4782 /** Search for the page a given key should be in.
4783 * Push it and its parent pages on the cursor stack.
4784 * @param[in,out] mc the cursor for this operation.
4785 * @param[in] key the key to search for, or NULL for first/last page.
4786 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4787 * are touched (updated with new page numbers).
4788 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4789 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4790 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4791 * @return 0 on success, non-zero on failure.
4794 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4799 /* Make sure the txn is still viable, then find the root from
4800 * the txn's db table and set it as the root of the cursor's stack.
4802 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4803 DPUTS("transaction has failed, must abort");
4806 /* Make sure we're using an up-to-date root */
4807 if (*mc->mc_dbflag & DB_STALE) {
4809 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4810 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4817 MDB_node *leaf = mdb_node_search(&mc2,
4818 &mc->mc_dbx->md_name, &exact);
4820 return MDB_NOTFOUND;
4821 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4824 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4826 /* The txn may not know this DBI, or another process may
4827 * have dropped and recreated the DB with other flags.
4829 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4830 return MDB_INCOMPATIBLE;
4831 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4833 *mc->mc_dbflag &= ~DB_STALE;
4835 root = mc->mc_db->md_root;
4837 if (root == P_INVALID) { /* Tree is empty. */
4838 DPUTS("tree is empty");
4839 return MDB_NOTFOUND;
4844 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4845 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4851 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4852 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4854 if (flags & MDB_PS_MODIFY) {
4855 if ((rc = mdb_page_touch(mc)))
4859 if (flags & MDB_PS_ROOTONLY)
4862 return mdb_page_search_root(mc, key, flags);
4866 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4868 MDB_txn *txn = mc->mc_txn;
4869 pgno_t pg = mp->mp_pgno;
4870 unsigned x = 0, ovpages = mp->mp_pages;
4871 MDB_env *env = txn->mt_env;
4872 MDB_IDL sl = txn->mt_spill_pgs;
4873 MDB_ID pn = pg << 1;
4876 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4877 /* If the page is dirty or on the spill list we just acquired it,
4878 * so we should give it back to our current free list, if any.
4879 * Otherwise put it onto the list of pages we freed in this txn.
4881 * Won't create me_pghead: me_pglast must be inited along with it.
4882 * Unsupported in nested txns: They would need to hide the page
4883 * range in ancestor txns' dirty and spilled lists.
4885 if (env->me_pghead &&
4887 ((mp->mp_flags & P_DIRTY) ||
4888 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4892 MDB_ID2 *dl, ix, iy;
4893 rc = mdb_midl_need(&env->me_pghead, ovpages);
4896 if (!(mp->mp_flags & P_DIRTY)) {
4897 /* This page is no longer spilled */
4904 /* Remove from dirty list */
4905 dl = txn->mt_u.dirty_list;
4907 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4915 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4916 txn->mt_flags |= MDB_TXN_ERROR;
4917 return MDB_CORRUPTED;
4920 if (!(env->me_flags & MDB_WRITEMAP))
4921 mdb_dpage_free(env, mp);
4923 /* Insert in me_pghead */
4924 mop = env->me_pghead;
4925 j = mop[0] + ovpages;
4926 for (i = mop[0]; i && mop[i] < pg; i--)
4932 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4936 mc->mc_db->md_overflow_pages -= ovpages;
4940 /** Return the data associated with a given node.
4941 * @param[in] txn The transaction for this operation.
4942 * @param[in] leaf The node being read.
4943 * @param[out] data Updated to point to the node's data.
4944 * @return 0 on success, non-zero on failure.
4947 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4949 MDB_page *omp; /* overflow page */
4953 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4954 data->mv_size = NODEDSZ(leaf);
4955 data->mv_data = NODEDATA(leaf);
4959 /* Read overflow data.
4961 data->mv_size = NODEDSZ(leaf);
4962 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4963 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4964 DPRINTF(("read overflow page %"Z"u failed", pgno));
4967 data->mv_data = METADATA(omp);
4973 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4974 MDB_val *key, MDB_val *data)
4983 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4985 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4988 if (txn->mt_flags & MDB_TXN_ERROR)
4991 mdb_cursor_init(&mc, txn, dbi, &mx);
4992 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4995 /** Find a sibling for a page.
4996 * Replaces the page at the top of the cursor's stack with the
4997 * specified sibling, if one exists.
4998 * @param[in] mc The cursor for this operation.
4999 * @param[in] move_right Non-zero if the right sibling is requested,
5000 * otherwise the left sibling.
5001 * @return 0 on success, non-zero on failure.
5004 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5010 if (mc->mc_snum < 2) {
5011 return MDB_NOTFOUND; /* root has no siblings */
5015 DPRINTF(("parent page is page %"Z"u, index %u",
5016 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5018 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5019 : (mc->mc_ki[mc->mc_top] == 0)) {
5020 DPRINTF(("no more keys left, moving to %s sibling",
5021 move_right ? "right" : "left"));
5022 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5023 /* undo cursor_pop before returning */
5030 mc->mc_ki[mc->mc_top]++;
5032 mc->mc_ki[mc->mc_top]--;
5033 DPRINTF(("just moving to %s index key %u",
5034 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5036 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
5038 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5039 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5040 /* mc will be inconsistent if caller does mc_snum++ as above */
5041 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5045 mdb_cursor_push(mc, mp);
5047 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5052 /** Move the cursor to the next data item. */
5054 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5060 if (mc->mc_flags & C_EOF) {
5061 return MDB_NOTFOUND;
5064 assert(mc->mc_flags & C_INITIALIZED);
5066 mp = mc->mc_pg[mc->mc_top];
5068 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5069 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5070 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5071 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5072 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5073 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5074 if (rc == MDB_SUCCESS)
5075 MDB_GET_KEY(leaf, key);
5080 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5081 if (op == MDB_NEXT_DUP)
5082 return MDB_NOTFOUND;
5086 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5087 if (mc->mc_flags & C_DEL)
5090 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5091 DPUTS("=====> move to next sibling page");
5092 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5093 mc->mc_flags |= C_EOF;
5096 mp = mc->mc_pg[mc->mc_top];
5097 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5099 mc->mc_ki[mc->mc_top]++;
5102 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5103 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5106 key->mv_size = mc->mc_db->md_pad;
5107 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5111 assert(IS_LEAF(mp));
5112 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5114 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5115 mdb_xcursor_init1(mc, leaf);
5118 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5121 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5122 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5123 if (rc != MDB_SUCCESS)
5128 MDB_GET_KEY(leaf, key);
5132 /** Move the cursor to the previous data item. */
5134 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5140 assert(mc->mc_flags & C_INITIALIZED);
5142 mp = mc->mc_pg[mc->mc_top];
5144 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5145 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5146 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5147 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5148 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5149 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5150 if (rc == MDB_SUCCESS)
5151 MDB_GET_KEY(leaf, key);
5155 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5156 if (op == MDB_PREV_DUP)
5157 return MDB_NOTFOUND;
5162 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5164 if (mc->mc_ki[mc->mc_top] == 0) {
5165 DPUTS("=====> move to prev sibling page");
5166 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5169 mp = mc->mc_pg[mc->mc_top];
5170 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5171 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5173 mc->mc_ki[mc->mc_top]--;
5175 mc->mc_flags &= ~C_EOF;
5177 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5178 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5181 key->mv_size = mc->mc_db->md_pad;
5182 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5186 assert(IS_LEAF(mp));
5187 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5189 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5190 mdb_xcursor_init1(mc, leaf);
5193 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5196 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5197 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5198 if (rc != MDB_SUCCESS)
5203 MDB_GET_KEY(leaf, key);
5207 /** Set the cursor on a specific data item. */
5209 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5210 MDB_cursor_op op, int *exactp)
5214 MDB_node *leaf = NULL;
5219 if (key->mv_size == 0)
5220 return MDB_BAD_VALSIZE;
5223 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5225 /* See if we're already on the right page */
5226 if (mc->mc_flags & C_INITIALIZED) {
5229 mp = mc->mc_pg[mc->mc_top];
5231 mc->mc_ki[mc->mc_top] = 0;
5232 return MDB_NOTFOUND;
5234 if (mp->mp_flags & P_LEAF2) {
5235 nodekey.mv_size = mc->mc_db->md_pad;
5236 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5238 leaf = NODEPTR(mp, 0);
5239 MDB_GET_KEY2(leaf, nodekey);
5241 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5243 /* Probably happens rarely, but first node on the page
5244 * was the one we wanted.
5246 mc->mc_ki[mc->mc_top] = 0;
5253 unsigned int nkeys = NUMKEYS(mp);
5255 if (mp->mp_flags & P_LEAF2) {
5256 nodekey.mv_data = LEAF2KEY(mp,
5257 nkeys-1, nodekey.mv_size);
5259 leaf = NODEPTR(mp, nkeys-1);
5260 MDB_GET_KEY2(leaf, nodekey);
5262 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5264 /* last node was the one we wanted */
5265 mc->mc_ki[mc->mc_top] = nkeys-1;
5271 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5272 /* This is definitely the right page, skip search_page */
5273 if (mp->mp_flags & P_LEAF2) {
5274 nodekey.mv_data = LEAF2KEY(mp,
5275 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5277 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5278 MDB_GET_KEY2(leaf, nodekey);
5280 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5282 /* current node was the one we wanted */
5292 /* If any parents have right-sibs, search.
5293 * Otherwise, there's nothing further.
5295 for (i=0; i<mc->mc_top; i++)
5297 NUMKEYS(mc->mc_pg[i])-1)
5299 if (i == mc->mc_top) {
5300 /* There are no other pages */
5301 mc->mc_ki[mc->mc_top] = nkeys;
5302 return MDB_NOTFOUND;
5306 /* There are no other pages */
5307 mc->mc_ki[mc->mc_top] = 0;
5308 if (op == MDB_SET_RANGE) {
5312 return MDB_NOTFOUND;
5316 rc = mdb_page_search(mc, key, 0);
5317 if (rc != MDB_SUCCESS)
5320 mp = mc->mc_pg[mc->mc_top];
5321 assert(IS_LEAF(mp));
5324 leaf = mdb_node_search(mc, key, exactp);
5325 if (exactp != NULL && !*exactp) {
5326 /* MDB_SET specified and not an exact match. */
5327 return MDB_NOTFOUND;
5331 DPUTS("===> inexact leaf not found, goto sibling");
5332 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5333 return rc; /* no entries matched */
5334 mp = mc->mc_pg[mc->mc_top];
5335 assert(IS_LEAF(mp));
5336 leaf = NODEPTR(mp, 0);
5340 mc->mc_flags |= C_INITIALIZED;
5341 mc->mc_flags &= ~C_EOF;
5344 key->mv_size = mc->mc_db->md_pad;
5345 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5349 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5350 mdb_xcursor_init1(mc, leaf);
5353 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5354 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5355 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5358 if (op == MDB_GET_BOTH) {
5364 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5365 if (rc != MDB_SUCCESS)
5368 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5370 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5372 rc = mc->mc_dbx->md_dcmp(data, &d2);
5374 if (op == MDB_GET_BOTH || rc > 0)
5375 return MDB_NOTFOUND;
5382 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5383 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5388 /* The key already matches in all other cases */
5389 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5390 MDB_GET_KEY(leaf, key);
5391 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5396 /** Move the cursor to the first item in the database. */
5398 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5404 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5406 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5407 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5408 if (rc != MDB_SUCCESS)
5411 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5413 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5414 mc->mc_flags |= C_INITIALIZED;
5415 mc->mc_flags &= ~C_EOF;
5417 mc->mc_ki[mc->mc_top] = 0;
5419 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5420 key->mv_size = mc->mc_db->md_pad;
5421 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5426 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5427 mdb_xcursor_init1(mc, leaf);
5428 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5432 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5436 MDB_GET_KEY(leaf, key);
5440 /** Move the cursor to the last item in the database. */
5442 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5448 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5450 if (!(mc->mc_flags & C_EOF)) {
5452 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5453 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5454 if (rc != MDB_SUCCESS)
5457 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5460 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5461 mc->mc_flags |= C_INITIALIZED|C_EOF;
5462 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5464 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5465 key->mv_size = mc->mc_db->md_pad;
5466 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5471 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5472 mdb_xcursor_init1(mc, leaf);
5473 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5477 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5482 MDB_GET_KEY(leaf, key);
5487 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5492 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5496 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5500 case MDB_GET_CURRENT:
5501 if (!(mc->mc_flags & C_INITIALIZED)) {
5504 MDB_page *mp = mc->mc_pg[mc->mc_top];
5505 int nkeys = NUMKEYS(mp);
5506 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5507 mc->mc_ki[mc->mc_top] = nkeys;
5513 key->mv_size = mc->mc_db->md_pad;
5514 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5516 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5517 MDB_GET_KEY(leaf, key);
5519 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5520 if (mc->mc_flags & C_DEL)
5521 mdb_xcursor_init1(mc, leaf);
5522 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5524 rc = mdb_node_read(mc->mc_txn, leaf, data);
5531 case MDB_GET_BOTH_RANGE:
5536 if (mc->mc_xcursor == NULL) {
5537 rc = MDB_INCOMPATIBLE;
5547 rc = mdb_cursor_set(mc, key, data, op,
5548 op == MDB_SET_RANGE ? NULL : &exact);
5551 case MDB_GET_MULTIPLE:
5552 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5556 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5557 rc = MDB_INCOMPATIBLE;
5561 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5562 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5565 case MDB_NEXT_MULTIPLE:
5570 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5571 rc = MDB_INCOMPATIBLE;
5574 if (!(mc->mc_flags & C_INITIALIZED))
5575 rc = mdb_cursor_first(mc, key, data);
5577 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5578 if (rc == MDB_SUCCESS) {
5579 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5582 mx = &mc->mc_xcursor->mx_cursor;
5583 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5585 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5586 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5594 case MDB_NEXT_NODUP:
5595 if (!(mc->mc_flags & C_INITIALIZED))
5596 rc = mdb_cursor_first(mc, key, data);
5598 rc = mdb_cursor_next(mc, key, data, op);
5602 case MDB_PREV_NODUP:
5603 if (!(mc->mc_flags & C_INITIALIZED)) {
5604 rc = mdb_cursor_last(mc, key, data);
5607 mc->mc_flags |= C_INITIALIZED;
5608 mc->mc_ki[mc->mc_top]++;
5610 rc = mdb_cursor_prev(mc, key, data, op);
5613 rc = mdb_cursor_first(mc, key, data);
5616 mfunc = mdb_cursor_first;
5618 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5622 if (mc->mc_xcursor == NULL) {
5623 rc = MDB_INCOMPATIBLE;
5626 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5630 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5633 rc = mdb_cursor_last(mc, key, data);
5636 mfunc = mdb_cursor_last;
5639 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5644 if (mc->mc_flags & C_DEL)
5645 mc->mc_flags ^= C_DEL;
5650 /** Touch all the pages in the cursor stack. Set mc_top.
5651 * Makes sure all the pages are writable, before attempting a write operation.
5652 * @param[in] mc The cursor to operate on.
5655 mdb_cursor_touch(MDB_cursor *mc)
5657 int rc = MDB_SUCCESS;
5659 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5662 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5663 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5666 *mc->mc_dbflag |= DB_DIRTY;
5671 rc = mdb_page_touch(mc);
5672 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5673 mc->mc_top = mc->mc_snum-1;
5678 /** Do not spill pages to disk if txn is getting full, may fail instead */
5679 #define MDB_NOSPILL 0x8000
5682 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5685 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5686 MDB_env *env = mc->mc_txn->mt_env;
5687 MDB_node *leaf = NULL;
5690 MDB_val xdata, *rdata, dkey, olddata;
5692 int do_sub = 0, insert;
5693 unsigned int mcount = 0, dcount = 0, nospill;
5696 unsigned int nflags;
5699 /* Check this first so counter will always be zero on any
5702 if (flags & MDB_MULTIPLE) {
5703 dcount = data[1].mv_size;
5704 data[1].mv_size = 0;
5705 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5706 return MDB_INCOMPATIBLE;
5709 nospill = flags & MDB_NOSPILL;
5710 flags &= ~MDB_NOSPILL;
5712 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5713 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5715 if (flags != MDB_CURRENT && key->mv_size-1 >= ENV_MAXKEY(env))
5716 return MDB_BAD_VALSIZE;
5718 #if SIZE_MAX > MAXDATASIZE
5719 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5720 return MDB_BAD_VALSIZE;
5722 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5723 return MDB_BAD_VALSIZE;
5726 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5727 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5731 if (flags == MDB_CURRENT) {
5732 if (!(mc->mc_flags & C_INITIALIZED))
5735 } else if (mc->mc_db->md_root == P_INVALID) {
5736 /* new database, cursor has nothing to point to */
5739 mc->mc_flags &= ~C_INITIALIZED;
5744 if (flags & MDB_APPEND) {
5746 rc = mdb_cursor_last(mc, &k2, &d2);
5748 rc = mc->mc_dbx->md_cmp(key, &k2);
5751 mc->mc_ki[mc->mc_top]++;
5753 /* new key is <= last key */
5758 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5760 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5761 DPRINTF(("duplicate key [%s]", DKEY(key)));
5763 return MDB_KEYEXIST;
5765 if (rc && rc != MDB_NOTFOUND)
5769 if (mc->mc_flags & C_DEL)
5770 mc->mc_flags ^= C_DEL;
5772 /* Cursor is positioned, check for room in the dirty list */
5774 if (flags & MDB_MULTIPLE) {
5776 xdata.mv_size = data->mv_size * dcount;
5780 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5784 if (rc == MDB_NO_ROOT) {
5786 /* new database, write a root leaf page */
5787 DPUTS("allocating new root leaf page");
5788 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5791 mdb_cursor_push(mc, np);
5792 mc->mc_db->md_root = np->mp_pgno;
5793 mc->mc_db->md_depth++;
5794 *mc->mc_dbflag |= DB_DIRTY;
5795 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5797 np->mp_flags |= P_LEAF2;
5798 mc->mc_flags |= C_INITIALIZED;
5800 /* make sure all cursor pages are writable */
5801 rc2 = mdb_cursor_touch(mc);
5808 /* The key does not exist */
5809 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5810 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5811 LEAFSIZE(key, data) > env->me_nodemax)
5813 /* Too big for a node, insert in sub-DB */
5814 fp_flags = P_LEAF|P_DIRTY;
5816 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5817 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5821 /* there's only a key anyway, so this is a no-op */
5822 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5823 unsigned int ksize = mc->mc_db->md_pad;
5824 if (key->mv_size != ksize)
5825 return MDB_BAD_VALSIZE;
5826 if (flags == MDB_CURRENT) {
5827 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5828 memcpy(ptr, key->mv_data, ksize);
5834 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5835 olddata.mv_size = NODEDSZ(leaf);
5836 olddata.mv_data = NODEDATA(leaf);
5839 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5840 /* Prepare (sub-)page/sub-DB to accept the new item,
5841 * if needed. fp: old sub-page or a header faking
5842 * it. mp: new (sub-)page. offset: growth in page
5843 * size. xdata: node data with new page or DB.
5845 unsigned i, offset = 0;
5846 mp = fp = xdata.mv_data = env->me_pbuf;
5847 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5849 /* Was a single item before, must convert now */
5850 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5851 /* Just overwrite the current item */
5852 if (flags == MDB_CURRENT)
5855 #if UINT_MAX < SIZE_MAX
5856 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5857 #ifdef MISALIGNED_OK
5858 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5860 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5863 /* if data matches, skip it */
5864 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5865 if (flags & MDB_NODUPDATA)
5867 else if (flags & MDB_MULTIPLE)
5874 /* Back up original data item */
5875 dkey.mv_size = olddata.mv_size;
5876 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5878 /* Make sub-page header for the dup items, with dummy body */
5879 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5880 fp->mp_lower = PAGEHDRSZ;
5881 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5882 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5883 fp->mp_flags |= P_LEAF2;
5884 fp->mp_pad = data->mv_size;
5885 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
5887 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
5888 (dkey.mv_size & 1) + (data->mv_size & 1);
5890 fp->mp_upper = xdata.mv_size;
5891 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
5892 } else if (leaf->mn_flags & F_SUBDATA) {
5893 /* Data is on sub-DB, just store it */
5894 flags |= F_DUPDATA|F_SUBDATA;
5897 /* Data is on sub-page */
5898 fp = olddata.mv_data;
5901 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5902 offset = EVEN(NODESIZE + sizeof(indx_t) +
5906 offset = fp->mp_pad;
5907 if (SIZELEFT(fp) < offset) {
5908 offset *= 4; /* space for 4 more */
5911 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
5913 fp->mp_flags |= P_DIRTY;
5914 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
5915 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5919 xdata.mv_size = olddata.mv_size + offset;
5922 fp_flags = fp->mp_flags;
5923 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
5924 /* Too big for a sub-page, convert to sub-DB */
5925 fp_flags &= ~P_SUBP;
5927 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5928 fp_flags |= P_LEAF2;
5929 dummy.md_pad = fp->mp_pad;
5930 dummy.md_flags = MDB_DUPFIXED;
5931 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5932 dummy.md_flags |= MDB_INTEGERKEY;
5938 dummy.md_branch_pages = 0;
5939 dummy.md_leaf_pages = 1;
5940 dummy.md_overflow_pages = 0;
5941 dummy.md_entries = NUMKEYS(fp);
5942 xdata.mv_size = sizeof(MDB_db);
5943 xdata.mv_data = &dummy;
5944 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5946 offset = env->me_psize - olddata.mv_size;
5947 flags |= F_DUPDATA|F_SUBDATA;
5948 dummy.md_root = mp->mp_pgno;
5951 mp->mp_flags = fp_flags | P_DIRTY;
5952 mp->mp_pad = fp->mp_pad;
5953 mp->mp_lower = fp->mp_lower;
5954 mp->mp_upper = fp->mp_upper + offset;
5955 if (fp_flags & P_LEAF2) {
5956 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5958 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
5959 olddata.mv_size - fp->mp_upper);
5960 for (i=0; i<NUMKEYS(fp); i++)
5961 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5968 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5972 /* overflow page overwrites need special handling */
5973 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5976 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
5978 memcpy(&pg, olddata.mv_data, sizeof(pg));
5979 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5981 ovpages = omp->mp_pages;
5983 /* Is the ov page large enough? */
5984 if (ovpages >= dpages) {
5985 if (!(omp->mp_flags & P_DIRTY) &&
5986 (level || (env->me_flags & MDB_WRITEMAP)))
5988 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5991 level = 0; /* dirty in this txn or clean */
5994 if (omp->mp_flags & P_DIRTY) {
5995 /* yes, overwrite it. Note in this case we don't
5996 * bother to try shrinking the page if the new data
5997 * is smaller than the overflow threshold.
6000 /* It is writable only in a parent txn */
6001 size_t sz = (size_t) env->me_psize * ovpages, off;
6002 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6008 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6009 if (!(flags & MDB_RESERVE)) {
6010 /* Copy end of page, adjusting alignment so
6011 * compiler may copy words instead of bytes.
6013 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6014 memcpy((size_t *)((char *)np + off),
6015 (size_t *)((char *)omp + off), sz - off);
6018 memcpy(np, omp, sz); /* Copy beginning of page */
6021 SETDSZ(leaf, data->mv_size);
6022 if (F_ISSET(flags, MDB_RESERVE))
6023 data->mv_data = METADATA(omp);
6025 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6029 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6031 } else if (data->mv_size == olddata.mv_size) {
6032 /* same size, just replace it. Note that we could
6033 * also reuse this node if the new data is smaller,
6034 * but instead we opt to shrink the node in that case.
6036 if (F_ISSET(flags, MDB_RESERVE))
6037 data->mv_data = olddata.mv_data;
6038 else if (data->mv_size)
6039 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6041 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6044 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6045 mc->mc_db->md_entries--;
6051 nflags = flags & NODE_ADD_FLAGS;
6052 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6053 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6054 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6055 nflags &= ~MDB_APPEND;
6057 nflags |= MDB_SPLIT_REPLACE;
6058 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6060 /* There is room already in this leaf page. */
6061 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6062 if (rc == 0 && !do_sub && insert) {
6063 /* Adjust other cursors pointing to mp */
6064 MDB_cursor *m2, *m3;
6065 MDB_dbi dbi = mc->mc_dbi;
6066 unsigned i = mc->mc_top;
6067 MDB_page *mp = mc->mc_pg[i];
6069 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6070 if (mc->mc_flags & C_SUB)
6071 m3 = &m2->mc_xcursor->mx_cursor;
6074 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6075 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6082 if (rc != MDB_SUCCESS)
6083 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6085 /* Now store the actual data in the child DB. Note that we're
6086 * storing the user data in the keys field, so there are strict
6087 * size limits on dupdata. The actual data fields of the child
6088 * DB are all zero size.
6095 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6096 if (flags & MDB_CURRENT) {
6097 xflags = MDB_CURRENT|MDB_NOSPILL;
6099 mdb_xcursor_init1(mc, leaf);
6100 xflags = (flags & MDB_NODUPDATA) ?
6101 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6103 /* converted, write the original data first */
6105 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6109 /* Adjust other cursors pointing to mp */
6111 unsigned i = mc->mc_top;
6112 MDB_page *mp = mc->mc_pg[i];
6114 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6115 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6116 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6117 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6118 mdb_xcursor_init1(m2, leaf);
6122 /* we've done our job */
6125 if (flags & MDB_APPENDDUP)
6126 xflags |= MDB_APPEND;
6127 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6128 if (flags & F_SUBDATA) {
6129 void *db = NODEDATA(leaf);
6130 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6133 /* sub-writes might have failed so check rc again.
6134 * Don't increment count if we just replaced an existing item.
6136 if (!rc && !(flags & MDB_CURRENT))
6137 mc->mc_db->md_entries++;
6138 if (flags & MDB_MULTIPLE) {
6142 /* let caller know how many succeeded, if any */
6143 data[1].mv_size = mcount;
6144 if (mcount < dcount) {
6145 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6152 /* If we succeeded and the key didn't exist before, make sure
6153 * the cursor is marked valid.
6156 mc->mc_flags |= C_INITIALIZED;
6161 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6167 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6168 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6170 if (!(mc->mc_flags & C_INITIALIZED))
6173 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6174 return MDB_NOTFOUND;
6176 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6179 rc = mdb_cursor_touch(mc);
6183 mp = mc->mc_pg[mc->mc_top];
6184 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6186 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6187 if (!(flags & MDB_NODUPDATA)) {
6188 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6189 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6191 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6192 /* If sub-DB still has entries, we're done */
6193 if (mc->mc_xcursor->mx_db.md_entries) {
6194 if (leaf->mn_flags & F_SUBDATA) {
6195 /* update subDB info */
6196 void *db = NODEDATA(leaf);
6197 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6200 /* shrink fake page */
6201 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6202 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6203 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6204 /* fix other sub-DB cursors pointed at this fake page */
6205 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6206 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6207 if (m2->mc_pg[mc->mc_top] == mp &&
6208 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6209 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6212 mc->mc_db->md_entries--;
6213 mc->mc_flags |= C_DEL;
6216 /* otherwise fall thru and delete the sub-DB */
6219 if (leaf->mn_flags & F_SUBDATA) {
6220 /* add all the child DB's pages to the free list */
6221 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6222 if (rc == MDB_SUCCESS) {
6223 mc->mc_db->md_entries -=
6224 mc->mc_xcursor->mx_db.md_entries;
6229 return mdb_cursor_del0(mc, leaf);
6232 /** Allocate and initialize new pages for a database.
6233 * @param[in] mc a cursor on the database being added to.
6234 * @param[in] flags flags defining what type of page is being allocated.
6235 * @param[in] num the number of pages to allocate. This is usually 1,
6236 * unless allocating overflow pages for a large record.
6237 * @param[out] mp Address of a page, or NULL on failure.
6238 * @return 0 on success, non-zero on failure.
6241 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6246 if ((rc = mdb_page_alloc(mc, num, &np)))
6248 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6249 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6250 np->mp_flags = flags | P_DIRTY;
6251 np->mp_lower = PAGEHDRSZ;
6252 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6255 mc->mc_db->md_branch_pages++;
6256 else if (IS_LEAF(np))
6257 mc->mc_db->md_leaf_pages++;
6258 else if (IS_OVERFLOW(np)) {
6259 mc->mc_db->md_overflow_pages += num;
6267 /** Calculate the size of a leaf node.
6268 * The size depends on the environment's page size; if a data item
6269 * is too large it will be put onto an overflow page and the node
6270 * size will only include the key and not the data. Sizes are always
6271 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6272 * of the #MDB_node headers.
6273 * @param[in] env The environment handle.
6274 * @param[in] key The key for the node.
6275 * @param[in] data The data for the node.
6276 * @return The number of bytes needed to store the node.
6279 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6283 sz = LEAFSIZE(key, data);
6284 if (sz > env->me_nodemax) {
6285 /* put on overflow page */
6286 sz -= data->mv_size - sizeof(pgno_t);
6289 return EVEN(sz + sizeof(indx_t));
6292 /** Calculate the size of a branch node.
6293 * The size should depend on the environment's page size but since
6294 * we currently don't support spilling large keys onto overflow
6295 * pages, it's simply the size of the #MDB_node header plus the
6296 * size of the key. Sizes are always rounded up to an even number
6297 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6298 * @param[in] env The environment handle.
6299 * @param[in] key The key for the node.
6300 * @return The number of bytes needed to store the node.
6303 mdb_branch_size(MDB_env *env, MDB_val *key)
6308 if (sz > env->me_nodemax) {
6309 /* put on overflow page */
6310 /* not implemented */
6311 /* sz -= key->size - sizeof(pgno_t); */
6314 return sz + sizeof(indx_t);
6317 /** Add a node to the page pointed to by the cursor.
6318 * @param[in] mc The cursor for this operation.
6319 * @param[in] indx The index on the page where the new node should be added.
6320 * @param[in] key The key for the new node.
6321 * @param[in] data The data for the new node, if any.
6322 * @param[in] pgno The page number, if adding a branch node.
6323 * @param[in] flags Flags for the node.
6324 * @return 0 on success, non-zero on failure. Possible errors are:
6326 * <li>ENOMEM - failed to allocate overflow pages for the node.
6327 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6328 * should never happen since all callers already calculate the
6329 * page's free space before calling this function.
6333 mdb_node_add(MDB_cursor *mc, indx_t indx,
6334 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6337 size_t node_size = NODESIZE;
6341 MDB_page *mp = mc->mc_pg[mc->mc_top];
6342 MDB_page *ofp = NULL; /* overflow page */
6345 assert(mp->mp_upper >= mp->mp_lower);
6347 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6348 IS_LEAF(mp) ? "leaf" : "branch",
6349 IS_SUBP(mp) ? "sub-" : "",
6350 mp->mp_pgno, indx, data ? data->mv_size : 0,
6351 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6354 /* Move higher keys up one slot. */
6355 int ksize = mc->mc_db->md_pad, dif;
6356 char *ptr = LEAF2KEY(mp, indx, ksize);
6357 dif = NUMKEYS(mp) - indx;
6359 memmove(ptr+ksize, ptr, dif*ksize);
6360 /* insert new key */
6361 memcpy(ptr, key->mv_data, ksize);
6363 /* Just using these for counting */
6364 mp->mp_lower += sizeof(indx_t);
6365 mp->mp_upper -= ksize - sizeof(indx_t);
6369 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6371 node_size += key->mv_size;
6374 if (F_ISSET(flags, F_BIGDATA)) {
6375 /* Data already on overflow page. */
6376 node_size += sizeof(pgno_t);
6377 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6378 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6380 /* Put data on overflow page. */
6381 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6382 data->mv_size, node_size+data->mv_size));
6383 node_size = EVEN(node_size + sizeof(pgno_t));
6384 if ((ssize_t)node_size > room)
6386 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6388 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6392 node_size += data->mv_size;
6395 node_size = EVEN(node_size);
6396 if ((ssize_t)node_size > room)
6400 /* Move higher pointers up one slot. */
6401 for (i = NUMKEYS(mp); i > indx; i--)
6402 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6404 /* Adjust free space offsets. */
6405 ofs = mp->mp_upper - node_size;
6406 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6407 mp->mp_ptrs[indx] = ofs;
6409 mp->mp_lower += sizeof(indx_t);
6411 /* Write the node data. */
6412 node = NODEPTR(mp, indx);
6413 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6414 node->mn_flags = flags;
6416 SETDSZ(node,data->mv_size);
6421 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6426 if (F_ISSET(flags, F_BIGDATA))
6427 memcpy(node->mn_data + key->mv_size, data->mv_data,
6429 else if (F_ISSET(flags, MDB_RESERVE))
6430 data->mv_data = node->mn_data + key->mv_size;
6432 memcpy(node->mn_data + key->mv_size, data->mv_data,
6435 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6437 if (F_ISSET(flags, MDB_RESERVE))
6438 data->mv_data = METADATA(ofp);
6440 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6447 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6448 mp->mp_pgno, NUMKEYS(mp)));
6449 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6450 DPRINTF(("node size = %"Z"u", node_size));
6451 return MDB_PAGE_FULL;
6454 /** Delete the specified node from a page.
6455 * @param[in] mp The page to operate on.
6456 * @param[in] indx The index of the node to delete.
6457 * @param[in] ksize The size of a node. Only used if the page is
6458 * part of a #MDB_DUPFIXED database.
6461 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6464 indx_t i, j, numkeys, ptr;
6471 COPY_PGNO(pgno, mp->mp_pgno);
6472 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6473 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6476 assert(indx < NUMKEYS(mp));
6479 int x = NUMKEYS(mp) - 1 - indx;
6480 base = LEAF2KEY(mp, indx, ksize);
6482 memmove(base, base + ksize, x * ksize);
6483 mp->mp_lower -= sizeof(indx_t);
6484 mp->mp_upper += ksize - sizeof(indx_t);
6488 node = NODEPTR(mp, indx);
6489 sz = NODESIZE + node->mn_ksize;
6491 if (F_ISSET(node->mn_flags, F_BIGDATA))
6492 sz += sizeof(pgno_t);
6494 sz += NODEDSZ(node);
6498 ptr = mp->mp_ptrs[indx];
6499 numkeys = NUMKEYS(mp);
6500 for (i = j = 0; i < numkeys; i++) {
6502 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6503 if (mp->mp_ptrs[i] < ptr)
6504 mp->mp_ptrs[j] += sz;
6509 base = (char *)mp + mp->mp_upper;
6510 memmove(base + sz, base, ptr - mp->mp_upper);
6512 mp->mp_lower -= sizeof(indx_t);
6516 /** Compact the main page after deleting a node on a subpage.
6517 * @param[in] mp The main page to operate on.
6518 * @param[in] indx The index of the subpage on the main page.
6521 mdb_node_shrink(MDB_page *mp, indx_t indx)
6527 indx_t i, numkeys, ptr;
6529 node = NODEPTR(mp, indx);
6530 sp = (MDB_page *)NODEDATA(node);
6531 delta = SIZELEFT(sp);
6532 xp = (MDB_page *)((char *)sp + delta);
6534 /* shift subpage upward */
6536 nsize = NUMKEYS(sp) * sp->mp_pad;
6538 return; /* do not make the node uneven-sized */
6539 memmove(METADATA(xp), METADATA(sp), nsize);
6542 numkeys = NUMKEYS(sp);
6543 for (i=numkeys-1; i>=0; i--)
6544 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6546 xp->mp_upper = sp->mp_lower;
6547 xp->mp_lower = sp->mp_lower;
6548 xp->mp_flags = sp->mp_flags;
6549 xp->mp_pad = sp->mp_pad;
6550 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6552 nsize = NODEDSZ(node) - delta;
6553 SETDSZ(node, nsize);
6555 /* shift lower nodes upward */
6556 ptr = mp->mp_ptrs[indx];
6557 numkeys = NUMKEYS(mp);
6558 for (i = 0; i < numkeys; i++) {
6559 if (mp->mp_ptrs[i] <= ptr)
6560 mp->mp_ptrs[i] += delta;
6563 base = (char *)mp + mp->mp_upper;
6564 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6565 mp->mp_upper += delta;
6568 /** Initial setup of a sorted-dups cursor.
6569 * Sorted duplicates are implemented as a sub-database for the given key.
6570 * The duplicate data items are actually keys of the sub-database.
6571 * Operations on the duplicate data items are performed using a sub-cursor
6572 * initialized when the sub-database is first accessed. This function does
6573 * the preliminary setup of the sub-cursor, filling in the fields that
6574 * depend only on the parent DB.
6575 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6578 mdb_xcursor_init0(MDB_cursor *mc)
6580 MDB_xcursor *mx = mc->mc_xcursor;
6582 mx->mx_cursor.mc_xcursor = NULL;
6583 mx->mx_cursor.mc_txn = mc->mc_txn;
6584 mx->mx_cursor.mc_db = &mx->mx_db;
6585 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6586 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6587 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6588 mx->mx_cursor.mc_snum = 0;
6589 mx->mx_cursor.mc_top = 0;
6590 mx->mx_cursor.mc_flags = C_SUB;
6591 mx->mx_dbx.md_name.mv_size = 0;
6592 mx->mx_dbx.md_name.mv_data = NULL;
6593 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6594 mx->mx_dbx.md_dcmp = NULL;
6595 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6598 /** Final setup of a sorted-dups cursor.
6599 * Sets up the fields that depend on the data from the main cursor.
6600 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6601 * @param[in] node The data containing the #MDB_db record for the
6602 * sorted-dup database.
6605 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6607 MDB_xcursor *mx = mc->mc_xcursor;
6609 if (node->mn_flags & F_SUBDATA) {
6610 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6611 mx->mx_cursor.mc_pg[0] = 0;
6612 mx->mx_cursor.mc_snum = 0;
6613 mx->mx_cursor.mc_top = 0;
6614 mx->mx_cursor.mc_flags = C_SUB;
6616 MDB_page *fp = NODEDATA(node);
6617 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6618 mx->mx_db.md_flags = 0;
6619 mx->mx_db.md_depth = 1;
6620 mx->mx_db.md_branch_pages = 0;
6621 mx->mx_db.md_leaf_pages = 1;
6622 mx->mx_db.md_overflow_pages = 0;
6623 mx->mx_db.md_entries = NUMKEYS(fp);
6624 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6625 mx->mx_cursor.mc_snum = 1;
6626 mx->mx_cursor.mc_top = 0;
6627 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6628 mx->mx_cursor.mc_pg[0] = fp;
6629 mx->mx_cursor.mc_ki[0] = 0;
6630 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6631 mx->mx_db.md_flags = MDB_DUPFIXED;
6632 mx->mx_db.md_pad = fp->mp_pad;
6633 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6634 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6637 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6638 mx->mx_db.md_root));
6639 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6640 #if UINT_MAX < SIZE_MAX
6641 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6642 #ifdef MISALIGNED_OK
6643 mx->mx_dbx.md_cmp = mdb_cmp_long;
6645 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6650 /** Initialize a cursor for a given transaction and database. */
6652 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6655 mc->mc_backup = NULL;
6658 mc->mc_db = &txn->mt_dbs[dbi];
6659 mc->mc_dbx = &txn->mt_dbxs[dbi];
6660 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6665 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6667 mc->mc_xcursor = mx;
6668 mdb_xcursor_init0(mc);
6670 mc->mc_xcursor = NULL;
6672 if (*mc->mc_dbflag & DB_STALE) {
6673 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6678 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6681 size_t size = sizeof(MDB_cursor);
6683 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6686 if (txn->mt_flags & MDB_TXN_ERROR)
6689 /* Allow read access to the freelist */
6690 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6693 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6694 size += sizeof(MDB_xcursor);
6696 if ((mc = malloc(size)) != NULL) {
6697 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6698 if (txn->mt_cursors) {
6699 mc->mc_next = txn->mt_cursors[dbi];
6700 txn->mt_cursors[dbi] = mc;
6701 mc->mc_flags |= C_UNTRACK;
6713 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6715 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6718 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6721 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6725 /* Return the count of duplicate data items for the current key */
6727 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6731 if (mc == NULL || countp == NULL)
6734 if (mc->mc_xcursor == NULL)
6735 return MDB_INCOMPATIBLE;
6737 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6738 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6741 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6744 *countp = mc->mc_xcursor->mx_db.md_entries;
6750 mdb_cursor_close(MDB_cursor *mc)
6752 if (mc && !mc->mc_backup) {
6753 /* remove from txn, if tracked */
6754 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6755 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6756 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6758 *prev = mc->mc_next;
6765 mdb_cursor_txn(MDB_cursor *mc)
6767 if (!mc) return NULL;
6772 mdb_cursor_dbi(MDB_cursor *mc)
6778 /** Replace the key for a node with a new key.
6779 * @param[in] mc Cursor pointing to the node to operate on.
6780 * @param[in] key The new key to use.
6781 * @return 0 on success, non-zero on failure.
6784 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6790 int delta, ksize, oksize;
6791 indx_t ptr, i, numkeys, indx;
6794 indx = mc->mc_ki[mc->mc_top];
6795 mp = mc->mc_pg[mc->mc_top];
6796 node = NODEPTR(mp, indx);
6797 ptr = mp->mp_ptrs[indx];
6801 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6802 k2.mv_data = NODEKEY(node);
6803 k2.mv_size = node->mn_ksize;
6804 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6806 mdb_dkey(&k2, kbuf2),
6812 /* Sizes must be 2-byte aligned. */
6813 ksize = EVEN(key->mv_size);
6814 oksize = EVEN(node->mn_ksize);
6815 delta = ksize - oksize;
6817 /* Shift node contents if EVEN(key length) changed. */
6819 if (delta > 0 && SIZELEFT(mp) < delta) {
6821 /* not enough space left, do a delete and split */
6822 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6823 pgno = NODEPGNO(node);
6824 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6825 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6828 numkeys = NUMKEYS(mp);
6829 for (i = 0; i < numkeys; i++) {
6830 if (mp->mp_ptrs[i] <= ptr)
6831 mp->mp_ptrs[i] -= delta;
6834 base = (char *)mp + mp->mp_upper;
6835 len = ptr - mp->mp_upper + NODESIZE;
6836 memmove(base - delta, base, len);
6837 mp->mp_upper -= delta;
6839 node = NODEPTR(mp, indx);
6842 /* But even if no shift was needed, update ksize */
6843 if (node->mn_ksize != key->mv_size)
6844 node->mn_ksize = key->mv_size;
6847 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6853 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6855 /** Move a node from csrc to cdst.
6858 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6865 unsigned short flags;
6869 /* Mark src and dst as dirty. */
6870 if ((rc = mdb_page_touch(csrc)) ||
6871 (rc = mdb_page_touch(cdst)))
6874 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6875 key.mv_size = csrc->mc_db->md_pad;
6876 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6878 data.mv_data = NULL;
6882 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6883 assert(!((size_t)srcnode&1));
6884 srcpg = NODEPGNO(srcnode);
6885 flags = srcnode->mn_flags;
6886 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6887 unsigned int snum = csrc->mc_snum;
6889 /* must find the lowest key below src */
6890 mdb_page_search_lowest(csrc);
6891 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6892 key.mv_size = csrc->mc_db->md_pad;
6893 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6895 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6896 key.mv_size = NODEKSZ(s2);
6897 key.mv_data = NODEKEY(s2);
6899 csrc->mc_snum = snum--;
6900 csrc->mc_top = snum;
6902 key.mv_size = NODEKSZ(srcnode);
6903 key.mv_data = NODEKEY(srcnode);
6905 data.mv_size = NODEDSZ(srcnode);
6906 data.mv_data = NODEDATA(srcnode);
6908 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6909 unsigned int snum = cdst->mc_snum;
6912 /* must find the lowest key below dst */
6913 mdb_page_search_lowest(cdst);
6914 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6915 bkey.mv_size = cdst->mc_db->md_pad;
6916 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6918 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6919 bkey.mv_size = NODEKSZ(s2);
6920 bkey.mv_data = NODEKEY(s2);
6922 cdst->mc_snum = snum--;
6923 cdst->mc_top = snum;
6924 mdb_cursor_copy(cdst, &mn);
6926 rc = mdb_update_key(&mn, &bkey);
6931 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6932 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6933 csrc->mc_ki[csrc->mc_top],
6935 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6936 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6938 /* Add the node to the destination page.
6940 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6941 if (rc != MDB_SUCCESS)
6944 /* Delete the node from the source page.
6946 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6949 /* Adjust other cursors pointing to mp */
6950 MDB_cursor *m2, *m3;
6951 MDB_dbi dbi = csrc->mc_dbi;
6952 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6954 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6955 if (csrc->mc_flags & C_SUB)
6956 m3 = &m2->mc_xcursor->mx_cursor;
6959 if (m3 == csrc) continue;
6960 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6961 csrc->mc_ki[csrc->mc_top]) {
6962 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6963 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6968 /* Update the parent separators.
6970 if (csrc->mc_ki[csrc->mc_top] == 0) {
6971 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6972 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6973 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6975 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6976 key.mv_size = NODEKSZ(srcnode);
6977 key.mv_data = NODEKEY(srcnode);
6979 DPRINTF(("update separator for source page %"Z"u to [%s]",
6980 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6981 mdb_cursor_copy(csrc, &mn);
6984 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6987 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6989 indx_t ix = csrc->mc_ki[csrc->mc_top];
6990 nullkey.mv_size = 0;
6991 csrc->mc_ki[csrc->mc_top] = 0;
6992 rc = mdb_update_key(csrc, &nullkey);
6993 csrc->mc_ki[csrc->mc_top] = ix;
6994 assert(rc == MDB_SUCCESS);
6998 if (cdst->mc_ki[cdst->mc_top] == 0) {
6999 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7000 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7001 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7003 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7004 key.mv_size = NODEKSZ(srcnode);
7005 key.mv_data = NODEKEY(srcnode);
7007 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7008 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7009 mdb_cursor_copy(cdst, &mn);
7012 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7015 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7017 indx_t ix = cdst->mc_ki[cdst->mc_top];
7018 nullkey.mv_size = 0;
7019 cdst->mc_ki[cdst->mc_top] = 0;
7020 rc = mdb_update_key(cdst, &nullkey);
7021 cdst->mc_ki[cdst->mc_top] = ix;
7022 assert(rc == MDB_SUCCESS);
7029 /** Merge one page into another.
7030 * The nodes from the page pointed to by \b csrc will
7031 * be copied to the page pointed to by \b cdst and then
7032 * the \b csrc page will be freed.
7033 * @param[in] csrc Cursor pointing to the source page.
7034 * @param[in] cdst Cursor pointing to the destination page.
7037 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7045 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7046 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7048 assert(csrc->mc_snum > 1); /* can't merge root page */
7049 assert(cdst->mc_snum > 1);
7051 /* Mark dst as dirty. */
7052 if ((rc = mdb_page_touch(cdst)))
7055 /* Move all nodes from src to dst.
7057 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7058 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7059 key.mv_size = csrc->mc_db->md_pad;
7060 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7061 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7062 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7063 if (rc != MDB_SUCCESS)
7065 key.mv_data = (char *)key.mv_data + key.mv_size;
7068 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7069 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7070 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7071 unsigned int snum = csrc->mc_snum;
7073 /* must find the lowest key below src */
7074 mdb_page_search_lowest(csrc);
7075 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7076 key.mv_size = csrc->mc_db->md_pad;
7077 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7079 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7080 key.mv_size = NODEKSZ(s2);
7081 key.mv_data = NODEKEY(s2);
7083 csrc->mc_snum = snum--;
7084 csrc->mc_top = snum;
7086 key.mv_size = srcnode->mn_ksize;
7087 key.mv_data = NODEKEY(srcnode);
7090 data.mv_size = NODEDSZ(srcnode);
7091 data.mv_data = NODEDATA(srcnode);
7092 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7093 if (rc != MDB_SUCCESS)
7098 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7099 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7100 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7102 /* Unlink the src page from parent and add to free list.
7104 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
7105 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
7108 rc = mdb_update_key(csrc, &key);
7114 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7115 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7118 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7119 csrc->mc_db->md_leaf_pages--;
7121 csrc->mc_db->md_branch_pages--;
7123 /* Adjust other cursors pointing to mp */
7124 MDB_cursor *m2, *m3;
7125 MDB_dbi dbi = csrc->mc_dbi;
7126 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7128 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7129 if (csrc->mc_flags & C_SUB)
7130 m3 = &m2->mc_xcursor->mx_cursor;
7133 if (m3 == csrc) continue;
7134 if (m3->mc_snum < csrc->mc_snum) continue;
7135 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7136 m3->mc_pg[csrc->mc_top] = mp;
7137 m3->mc_ki[csrc->mc_top] += nkeys;
7141 mdb_cursor_pop(csrc);
7143 return mdb_rebalance(csrc);
7146 /** Copy the contents of a cursor.
7147 * @param[in] csrc The cursor to copy from.
7148 * @param[out] cdst The cursor to copy to.
7151 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7155 cdst->mc_txn = csrc->mc_txn;
7156 cdst->mc_dbi = csrc->mc_dbi;
7157 cdst->mc_db = csrc->mc_db;
7158 cdst->mc_dbx = csrc->mc_dbx;
7159 cdst->mc_snum = csrc->mc_snum;
7160 cdst->mc_top = csrc->mc_top;
7161 cdst->mc_flags = csrc->mc_flags;
7163 for (i=0; i<csrc->mc_snum; i++) {
7164 cdst->mc_pg[i] = csrc->mc_pg[i];
7165 cdst->mc_ki[i] = csrc->mc_ki[i];
7169 /** Rebalance the tree after a delete operation.
7170 * @param[in] mc Cursor pointing to the page where rebalancing
7172 * @return 0 on success, non-zero on failure.
7175 mdb_rebalance(MDB_cursor *mc)
7179 unsigned int ptop, minkeys;
7182 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7186 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7187 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7188 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7189 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7190 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7194 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7195 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7198 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7199 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7205 if (mc->mc_snum < 2) {
7206 MDB_page *mp = mc->mc_pg[0];
7208 DPUTS("Can't rebalance a subpage, ignoring");
7211 if (NUMKEYS(mp) == 0) {
7212 DPUTS("tree is completely empty");
7213 mc->mc_db->md_root = P_INVALID;
7214 mc->mc_db->md_depth = 0;
7215 mc->mc_db->md_leaf_pages = 0;
7216 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7219 /* Adjust cursors pointing to mp */
7222 mc->mc_flags &= ~C_INITIALIZED;
7224 MDB_cursor *m2, *m3;
7225 MDB_dbi dbi = mc->mc_dbi;
7227 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7228 if (mc->mc_flags & C_SUB)
7229 m3 = &m2->mc_xcursor->mx_cursor;
7232 if (m3->mc_snum < mc->mc_snum) continue;
7233 if (m3->mc_pg[0] == mp) {
7236 m3->mc_flags &= ~C_INITIALIZED;
7240 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7241 DPUTS("collapsing root page!");
7242 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7245 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7246 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7249 mc->mc_db->md_depth--;
7250 mc->mc_db->md_branch_pages--;
7251 mc->mc_ki[0] = mc->mc_ki[1];
7253 /* Adjust other cursors pointing to mp */
7254 MDB_cursor *m2, *m3;
7255 MDB_dbi dbi = mc->mc_dbi;
7257 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7258 if (mc->mc_flags & C_SUB)
7259 m3 = &m2->mc_xcursor->mx_cursor;
7262 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7263 if (m3->mc_pg[0] == mp) {
7267 for (i=0; i<m3->mc_snum; i++) {
7268 m3->mc_pg[i] = m3->mc_pg[i+1];
7269 m3->mc_ki[i] = m3->mc_ki[i+1];
7275 DPUTS("root page doesn't need rebalancing");
7279 /* The parent (branch page) must have at least 2 pointers,
7280 * otherwise the tree is invalid.
7282 ptop = mc->mc_top-1;
7283 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7285 /* Leaf page fill factor is below the threshold.
7286 * Try to move keys from left or right neighbor, or
7287 * merge with a neighbor page.
7292 mdb_cursor_copy(mc, &mn);
7293 mn.mc_xcursor = NULL;
7295 if (mc->mc_ki[ptop] == 0) {
7296 /* We're the leftmost leaf in our parent.
7298 DPUTS("reading right neighbor");
7300 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7301 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7304 mn.mc_ki[mn.mc_top] = 0;
7305 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7307 /* There is at least one neighbor to the left.
7309 DPUTS("reading left neighbor");
7311 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7312 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7315 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7316 mc->mc_ki[mc->mc_top] = 0;
7319 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7320 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7321 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7323 /* If the neighbor page is above threshold and has enough keys,
7324 * move one key from it. Otherwise we should try to merge them.
7325 * (A branch page must never have less than 2 keys.)
7327 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7328 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7329 return mdb_node_move(&mn, mc);
7331 if (mc->mc_ki[ptop] == 0)
7332 rc = mdb_page_merge(&mn, mc);
7334 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7335 rc = mdb_page_merge(mc, &mn);
7336 mdb_cursor_copy(&mn, mc);
7338 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7343 /** Complete a delete operation started by #mdb_cursor_del(). */
7345 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7352 mp = mc->mc_pg[mc->mc_top];
7353 ki = mc->mc_ki[mc->mc_top];
7355 /* add overflow pages to free list */
7356 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7360 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7361 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7362 (rc = mdb_ovpage_free(mc, omp)))
7365 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7366 mc->mc_db->md_entries--;
7367 rc = mdb_rebalance(mc);
7368 if (rc != MDB_SUCCESS)
7369 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7372 MDB_dbi dbi = mc->mc_dbi;
7374 mp = mc->mc_pg[mc->mc_top];
7375 nkeys = NUMKEYS(mp);
7377 /* if mc points past last node in page, find next sibling */
7378 if (mc->mc_ki[mc->mc_top] >= nkeys)
7379 mdb_cursor_sibling(mc, 1);
7381 /* Adjust other cursors pointing to mp */
7382 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7383 if (m2 == mc || m2->mc_snum < mc->mc_snum)
7385 if (!(m2->mc_flags & C_INITIALIZED))
7387 if (m2->mc_pg[mc->mc_top] == mp) {
7388 if (m2->mc_ki[mc->mc_top] >= ki) {
7389 m2->mc_flags |= C_DEL;
7390 if (m2->mc_ki[mc->mc_top] > ki)
7391 m2->mc_ki[mc->mc_top]--;
7393 if (m2->mc_ki[mc->mc_top] >= nkeys)
7394 mdb_cursor_sibling(m2, 1);
7397 mc->mc_flags |= C_DEL;
7404 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7405 MDB_val *key, MDB_val *data)
7410 MDB_val rdata, *xdata;
7414 assert(key != NULL);
7416 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7418 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7421 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7422 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7424 mdb_cursor_init(&mc, txn, dbi, &mx);
7427 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7428 /* must ignore any data */
7439 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7441 /* let mdb_page_split know about this cursor if needed:
7442 * delete will trigger a rebalance; if it needs to move
7443 * a node from one page to another, it will have to
7444 * update the parent's separator key(s). If the new sepkey
7445 * is larger than the current one, the parent page may
7446 * run out of space, triggering a split. We need this
7447 * cursor to be consistent until the end of the rebalance.
7449 mc.mc_flags |= C_UNTRACK;
7450 mc.mc_next = txn->mt_cursors[dbi];
7451 txn->mt_cursors[dbi] = &mc;
7452 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7453 txn->mt_cursors[dbi] = mc.mc_next;
7458 /** Split a page and insert a new node.
7459 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7460 * The cursor will be updated to point to the actual page and index where
7461 * the node got inserted after the split.
7462 * @param[in] newkey The key for the newly inserted node.
7463 * @param[in] newdata The data for the newly inserted node.
7464 * @param[in] newpgno The page number, if the new node is a branch node.
7465 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7466 * @return 0 on success, non-zero on failure.
7469 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7470 unsigned int nflags)
7473 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7476 int i, j, split_indx, nkeys, pmax;
7477 MDB_env *env = mc->mc_txn->mt_env;
7479 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7480 MDB_page *copy = NULL;
7481 MDB_page *mp, *rp, *pp;
7486 mp = mc->mc_pg[mc->mc_top];
7487 newindx = mc->mc_ki[mc->mc_top];
7488 nkeys = NUMKEYS(mp);
7490 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7491 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7492 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7494 /* Create a right sibling. */
7495 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7497 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7499 if (mc->mc_snum < 2) {
7500 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7502 /* shift current top to make room for new parent */
7503 mc->mc_pg[1] = mc->mc_pg[0];
7504 mc->mc_ki[1] = mc->mc_ki[0];
7507 mc->mc_db->md_root = pp->mp_pgno;
7508 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7509 mc->mc_db->md_depth++;
7512 /* Add left (implicit) pointer. */
7513 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7514 /* undo the pre-push */
7515 mc->mc_pg[0] = mc->mc_pg[1];
7516 mc->mc_ki[0] = mc->mc_ki[1];
7517 mc->mc_db->md_root = mp->mp_pgno;
7518 mc->mc_db->md_depth--;
7525 ptop = mc->mc_top-1;
7526 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7529 mc->mc_flags |= C_SPLITTING;
7530 mdb_cursor_copy(mc, &mn);
7531 mn.mc_pg[mn.mc_top] = rp;
7532 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7534 if (nflags & MDB_APPEND) {
7535 mn.mc_ki[mn.mc_top] = 0;
7537 split_indx = newindx;
7541 split_indx = (nkeys+1) / 2;
7546 unsigned int lsize, rsize, ksize;
7547 /* Move half of the keys to the right sibling */
7549 x = mc->mc_ki[mc->mc_top] - split_indx;
7550 ksize = mc->mc_db->md_pad;
7551 split = LEAF2KEY(mp, split_indx, ksize);
7552 rsize = (nkeys - split_indx) * ksize;
7553 lsize = (nkeys - split_indx) * sizeof(indx_t);
7554 mp->mp_lower -= lsize;
7555 rp->mp_lower += lsize;
7556 mp->mp_upper += rsize - lsize;
7557 rp->mp_upper -= rsize - lsize;
7558 sepkey.mv_size = ksize;
7559 if (newindx == split_indx) {
7560 sepkey.mv_data = newkey->mv_data;
7562 sepkey.mv_data = split;
7565 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7566 memcpy(rp->mp_ptrs, split, rsize);
7567 sepkey.mv_data = rp->mp_ptrs;
7568 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7569 memcpy(ins, newkey->mv_data, ksize);
7570 mp->mp_lower += sizeof(indx_t);
7571 mp->mp_upper -= ksize - sizeof(indx_t);
7574 memcpy(rp->mp_ptrs, split, x * ksize);
7575 ins = LEAF2KEY(rp, x, ksize);
7576 memcpy(ins, newkey->mv_data, ksize);
7577 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7578 rp->mp_lower += sizeof(indx_t);
7579 rp->mp_upper -= ksize - sizeof(indx_t);
7580 mc->mc_ki[mc->mc_top] = x;
7581 mc->mc_pg[mc->mc_top] = rp;
7584 int psize, nsize, k;
7585 /* Maximum free space in an empty page */
7586 pmax = env->me_psize - PAGEHDRSZ;
7588 nsize = mdb_leaf_size(env, newkey, newdata);
7590 nsize = mdb_branch_size(env, newkey);
7591 nsize = EVEN(nsize);
7593 /* grab a page to hold a temporary copy */
7594 copy = mdb_page_malloc(mc->mc_txn, 1);
7597 copy->mp_pgno = mp->mp_pgno;
7598 copy->mp_flags = mp->mp_flags;
7599 copy->mp_lower = PAGEHDRSZ;
7600 copy->mp_upper = env->me_psize;
7602 /* prepare to insert */
7603 for (i=0, j=0; i<nkeys; i++) {
7605 copy->mp_ptrs[j++] = 0;
7607 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7610 /* When items are relatively large the split point needs
7611 * to be checked, because being off-by-one will make the
7612 * difference between success or failure in mdb_node_add.
7614 * It's also relevant if a page happens to be laid out
7615 * such that one half of its nodes are all "small" and
7616 * the other half of its nodes are "large." If the new
7617 * item is also "large" and falls on the half with
7618 * "large" nodes, it also may not fit.
7620 * As a final tweak, if the new item goes on the last
7621 * spot on the page (and thus, onto the new page), bias
7622 * the split so the new page is emptier than the old page.
7623 * This yields better packing during sequential inserts.
7625 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7626 /* Find split point */
7628 if (newindx <= split_indx || newindx >= nkeys) {
7630 k = newindx >= nkeys ? nkeys : split_indx+2;
7635 for (; i!=k; i+=j) {
7640 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7641 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7643 if (F_ISSET(node->mn_flags, F_BIGDATA))
7644 psize += sizeof(pgno_t);
7646 psize += NODEDSZ(node);
7648 psize = EVEN(psize);
7650 if (psize > pmax || i == k-j) {
7651 split_indx = i + (j<0);
7656 if (split_indx == newindx) {
7657 sepkey.mv_size = newkey->mv_size;
7658 sepkey.mv_data = newkey->mv_data;
7660 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7661 sepkey.mv_size = node->mn_ksize;
7662 sepkey.mv_data = NODEKEY(node);
7667 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7669 /* Copy separator key to the parent.
7671 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7675 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7678 if (mn.mc_snum == mc->mc_snum) {
7679 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7680 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7681 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7682 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7687 /* Right page might now have changed parent.
7688 * Check if left page also changed parent.
7690 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7691 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7692 for (i=0; i<ptop; i++) {
7693 mc->mc_pg[i] = mn.mc_pg[i];
7694 mc->mc_ki[i] = mn.mc_ki[i];
7696 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7697 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7701 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7704 mc->mc_flags ^= C_SPLITTING;
7705 if (rc != MDB_SUCCESS) {
7708 if (nflags & MDB_APPEND) {
7709 mc->mc_pg[mc->mc_top] = rp;
7710 mc->mc_ki[mc->mc_top] = 0;
7711 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7714 for (i=0; i<mc->mc_top; i++)
7715 mc->mc_ki[i] = mn.mc_ki[i];
7716 } else if (!IS_LEAF2(mp)) {
7718 mc->mc_pg[mc->mc_top] = rp;
7723 rkey.mv_data = newkey->mv_data;
7724 rkey.mv_size = newkey->mv_size;
7730 /* Update index for the new key. */
7731 mc->mc_ki[mc->mc_top] = j;
7733 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7734 rkey.mv_data = NODEKEY(node);
7735 rkey.mv_size = node->mn_ksize;
7737 xdata.mv_data = NODEDATA(node);
7738 xdata.mv_size = NODEDSZ(node);
7741 pgno = NODEPGNO(node);
7742 flags = node->mn_flags;
7745 if (!IS_LEAF(mp) && j == 0) {
7746 /* First branch index doesn't need key data. */
7750 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7752 /* return tmp page to freelist */
7753 mdb_page_free(env, copy);
7759 mc->mc_pg[mc->mc_top] = copy;
7764 } while (i != split_indx);
7766 nkeys = NUMKEYS(copy);
7767 for (i=0; i<nkeys; i++)
7768 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7769 mp->mp_lower = copy->mp_lower;
7770 mp->mp_upper = copy->mp_upper;
7771 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7772 env->me_psize - copy->mp_upper);
7774 /* reset back to original page */
7775 if (newindx < split_indx) {
7776 mc->mc_pg[mc->mc_top] = mp;
7777 if (nflags & MDB_RESERVE) {
7778 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7779 if (!(node->mn_flags & F_BIGDATA))
7780 newdata->mv_data = NODEDATA(node);
7783 mc->mc_pg[mc->mc_top] = rp;
7785 /* Make sure mc_ki is still valid.
7787 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7788 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7789 for (i=0; i<ptop; i++) {
7790 mc->mc_pg[i] = mn.mc_pg[i];
7791 mc->mc_ki[i] = mn.mc_ki[i];
7793 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7794 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7797 /* return tmp page to freelist */
7798 mdb_page_free(env, copy);
7802 /* Adjust other cursors pointing to mp */
7803 MDB_cursor *m2, *m3;
7804 MDB_dbi dbi = mc->mc_dbi;
7805 int fixup = NUMKEYS(mp);
7807 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7808 if (mc->mc_flags & C_SUB)
7809 m3 = &m2->mc_xcursor->mx_cursor;
7814 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7816 if (m3->mc_flags & C_SPLITTING)
7821 for (k=m3->mc_top; k>=0; k--) {
7822 m3->mc_ki[k+1] = m3->mc_ki[k];
7823 m3->mc_pg[k+1] = m3->mc_pg[k];
7825 if (m3->mc_ki[0] >= split_indx) {
7830 m3->mc_pg[0] = mc->mc_pg[0];
7834 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7835 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7836 m3->mc_ki[mc->mc_top]++;
7837 if (m3->mc_ki[mc->mc_top] >= fixup) {
7838 m3->mc_pg[mc->mc_top] = rp;
7839 m3->mc_ki[mc->mc_top] -= fixup;
7840 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7842 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7843 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7848 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
7853 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7854 MDB_val *key, MDB_val *data, unsigned int flags)
7859 assert(key != NULL);
7860 assert(data != NULL);
7862 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7865 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7868 mdb_cursor_init(&mc, txn, dbi, &mx);
7869 return mdb_cursor_put(&mc, key, data, flags);
7873 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7875 if ((flag & CHANGEABLE) != flag)
7878 env->me_flags |= flag;
7880 env->me_flags &= ~flag;
7885 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7890 *arg = env->me_flags;
7895 mdb_env_get_path(MDB_env *env, const char **arg)
7900 *arg = env->me_path;
7905 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
7914 /** Common code for #mdb_stat() and #mdb_env_stat().
7915 * @param[in] env the environment to operate in.
7916 * @param[in] db the #MDB_db record containing the stats to return.
7917 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7918 * @return 0, this function always succeeds.
7921 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7923 arg->ms_psize = env->me_psize;
7924 arg->ms_depth = db->md_depth;
7925 arg->ms_branch_pages = db->md_branch_pages;
7926 arg->ms_leaf_pages = db->md_leaf_pages;
7927 arg->ms_overflow_pages = db->md_overflow_pages;
7928 arg->ms_entries = db->md_entries;
7933 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7937 if (env == NULL || arg == NULL)
7940 toggle = mdb_env_pick_meta(env);
7942 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7946 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7950 if (env == NULL || arg == NULL)
7953 toggle = mdb_env_pick_meta(env);
7954 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7955 arg->me_mapsize = env->me_mapsize;
7956 arg->me_maxreaders = env->me_maxreaders;
7958 /* me_numreaders may be zero if this process never used any readers. Use
7959 * the shared numreader count if it exists.
7961 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7963 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7964 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7968 /** Set the default comparison functions for a database.
7969 * Called immediately after a database is opened to set the defaults.
7970 * The user can then override them with #mdb_set_compare() or
7971 * #mdb_set_dupsort().
7972 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7973 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7976 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7978 uint16_t f = txn->mt_dbs[dbi].md_flags;
7980 txn->mt_dbxs[dbi].md_cmp =
7981 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7982 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7984 txn->mt_dbxs[dbi].md_dcmp =
7985 !(f & MDB_DUPSORT) ? 0 :
7986 ((f & MDB_INTEGERDUP)
7987 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7988 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7991 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7996 int rc, dbflag, exact;
7997 unsigned int unused = 0;
8000 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8001 mdb_default_cmp(txn, FREE_DBI);
8004 if ((flags & VALID_FLAGS) != flags)
8006 if (txn->mt_flags & MDB_TXN_ERROR)
8012 if (flags & PERSISTENT_FLAGS) {
8013 uint16_t f2 = flags & PERSISTENT_FLAGS;
8014 /* make sure flag changes get committed */
8015 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8016 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8017 txn->mt_flags |= MDB_TXN_DIRTY;
8020 mdb_default_cmp(txn, MAIN_DBI);
8024 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8025 mdb_default_cmp(txn, MAIN_DBI);
8028 /* Is the DB already open? */
8030 for (i=2; i<txn->mt_numdbs; i++) {
8031 if (!txn->mt_dbxs[i].md_name.mv_size) {
8032 /* Remember this free slot */
8033 if (!unused) unused = i;
8036 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8037 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8043 /* If no free slot and max hit, fail */
8044 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8045 return MDB_DBS_FULL;
8047 /* Cannot mix named databases with some mainDB flags */
8048 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8049 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8051 /* Find the DB info */
8052 dbflag = DB_NEW|DB_VALID;
8055 key.mv_data = (void *)name;
8056 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8057 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8058 if (rc == MDB_SUCCESS) {
8059 /* make sure this is actually a DB */
8060 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8061 if (!(node->mn_flags & F_SUBDATA))
8062 return MDB_INCOMPATIBLE;
8063 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8064 /* Create if requested */
8066 data.mv_size = sizeof(MDB_db);
8067 data.mv_data = &dummy;
8068 memset(&dummy, 0, sizeof(dummy));
8069 dummy.md_root = P_INVALID;
8070 dummy.md_flags = flags & PERSISTENT_FLAGS;
8071 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8075 /* OK, got info, add to table */
8076 if (rc == MDB_SUCCESS) {
8077 unsigned int slot = unused ? unused : txn->mt_numdbs;
8078 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8079 txn->mt_dbxs[slot].md_name.mv_size = len;
8080 txn->mt_dbxs[slot].md_rel = NULL;
8081 txn->mt_dbflags[slot] = dbflag;
8082 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8084 mdb_default_cmp(txn, slot);
8093 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8095 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8098 if (txn->mt_dbflags[dbi] & DB_STALE) {
8101 /* Stale, must read the DB's root. cursor_init does it for us. */
8102 mdb_cursor_init(&mc, txn, dbi, &mx);
8104 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8107 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8110 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8112 ptr = env->me_dbxs[dbi].md_name.mv_data;
8113 env->me_dbxs[dbi].md_name.mv_data = NULL;
8114 env->me_dbxs[dbi].md_name.mv_size = 0;
8115 env->me_dbflags[dbi] = 0;
8119 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8121 /* We could return the flags for the FREE_DBI too but what's the point? */
8122 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8124 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8128 /** Add all the DB's pages to the free list.
8129 * @param[in] mc Cursor on the DB to free.
8130 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8131 * @return 0 on success, non-zero on failure.
8134 mdb_drop0(MDB_cursor *mc, int subs)
8138 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8139 if (rc == MDB_SUCCESS) {
8140 MDB_txn *txn = mc->mc_txn;
8145 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8146 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8149 mdb_cursor_copy(mc, &mx);
8150 while (mc->mc_snum > 0) {
8151 MDB_page *mp = mc->mc_pg[mc->mc_top];
8152 unsigned n = NUMKEYS(mp);
8154 for (i=0; i<n; i++) {
8155 ni = NODEPTR(mp, i);
8156 if (ni->mn_flags & F_BIGDATA) {
8159 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8160 rc = mdb_page_get(txn, pg, &omp, NULL);
8163 assert(IS_OVERFLOW(omp));
8164 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8168 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8169 mdb_xcursor_init1(mc, ni);
8170 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8176 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8178 for (i=0; i<n; i++) {
8180 ni = NODEPTR(mp, i);
8183 mdb_midl_xappend(txn->mt_free_pgs, pg);
8188 mc->mc_ki[mc->mc_top] = i;
8189 rc = mdb_cursor_sibling(mc, 1);
8191 /* no more siblings, go back to beginning
8192 * of previous level.
8196 for (i=1; i<mc->mc_snum; i++) {
8198 mc->mc_pg[i] = mx.mc_pg[i];
8203 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8204 } else if (rc == MDB_NOTFOUND) {
8210 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8212 MDB_cursor *mc, *m2;
8215 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8218 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8221 rc = mdb_cursor_open(txn, dbi, &mc);
8225 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8226 /* Invalidate the dropped DB's cursors */
8227 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8228 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8232 /* Can't delete the main DB */
8233 if (del && dbi > MAIN_DBI) {
8234 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8236 txn->mt_dbflags[dbi] = DB_STALE;
8237 mdb_dbi_close(txn->mt_env, dbi);
8240 /* reset the DB record, mark it dirty */
8241 txn->mt_dbflags[dbi] |= DB_DIRTY;
8242 txn->mt_dbs[dbi].md_depth = 0;
8243 txn->mt_dbs[dbi].md_branch_pages = 0;
8244 txn->mt_dbs[dbi].md_leaf_pages = 0;
8245 txn->mt_dbs[dbi].md_overflow_pages = 0;
8246 txn->mt_dbs[dbi].md_entries = 0;
8247 txn->mt_dbs[dbi].md_root = P_INVALID;
8249 txn->mt_flags |= MDB_TXN_DIRTY;
8252 mdb_cursor_close(mc);
8256 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8258 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8261 txn->mt_dbxs[dbi].md_cmp = cmp;
8265 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8267 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8270 txn->mt_dbxs[dbi].md_dcmp = cmp;
8274 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8276 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8279 txn->mt_dbxs[dbi].md_rel = rel;
8283 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8285 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8288 txn->mt_dbxs[dbi].md_relctx = ctx;
8292 int mdb_env_get_maxkeysize(MDB_env *env)
8294 return ENV_MAXKEY(env);
8297 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8299 unsigned int i, rdrs;
8302 int rc = 0, first = 1;
8306 if (!env->me_txns) {
8307 return func("(no reader locks)\n", ctx);
8309 rdrs = env->me_txns->mti_numreaders;
8310 mr = env->me_txns->mti_readers;
8311 for (i=0; i<rdrs; i++) {
8315 if (mr[i].mr_txnid == (txnid_t)-1) {
8316 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8318 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8322 rc = func(" pid thread txnid\n", ctx);
8326 rc = func(buf, ctx);
8332 rc = func("(no active readers)\n", ctx);
8337 /** Insert pid into list if not already present.
8338 * return -1 if already present.
8340 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8342 /* binary search of pid in list */
8344 unsigned cursor = 1;
8346 unsigned n = ids[0];
8349 unsigned pivot = n >> 1;
8350 cursor = base + pivot + 1;
8351 val = pid - ids[cursor];
8356 } else if ( val > 0 ) {
8361 /* found, so it's a duplicate */
8370 for (n = ids[0]; n > cursor; n--)
8376 int mdb_reader_check(MDB_env *env, int *dead)
8378 unsigned int i, j, rdrs;
8380 MDB_PID_T *pids, pid;
8389 rdrs = env->me_txns->mti_numreaders;
8390 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8394 mr = env->me_txns->mti_readers;
8395 for (i=0; i<rdrs; i++) {
8396 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8398 if (mdb_pid_insert(pids, pid) == 0) {
8399 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8401 /* Recheck, a new process may have reused pid */
8402 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8403 for (j=i; j<rdrs; j++)
8404 if (mr[j].mr_pid == pid) {
8405 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8406 (unsigned) pid, mr[j].mr_txnid));
8411 UNLOCK_MUTEX_R(env);