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 maximum size of a key we can write to a database.
387 * We require that keys all fit onto a regular page. This limit
388 * could be raised a bit further if needed; to something just
389 * under (page size / #MDB_MINKEYS / 3).
391 * Note that data items in an #MDB_DUPSORT database are actually keys
392 * of a subDB, so they're also limited to this size.
394 #ifndef MDB_MAXKEYSIZE
395 #define MDB_MAXKEYSIZE 511
398 /** @brief The maximum size of a data item.
400 * We only store a 32 bit value for node sizes.
402 #define MAXDATASIZE 0xffffffffUL
407 * This is used for printing a hex dump of a key's contents.
409 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
410 /** Display a key in hex.
412 * Invoke a function to display a key in hex.
414 #define DKEY(x) mdb_dkey(x, kbuf)
420 /** An invalid page number.
421 * Mainly used to denote an empty tree.
423 #define P_INVALID (~(pgno_t)0)
425 /** Test if the flags \b f are set in a flag word \b w. */
426 #define F_ISSET(w, f) (((w) & (f)) == (f))
428 /** Round \b n up to an even number. */
429 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
431 /** Used for offsets within a single page.
432 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
435 typedef uint16_t indx_t;
437 /** Default size of memory map.
438 * This is certainly too small for any actual applications. Apps should always set
439 * the size explicitly using #mdb_env_set_mapsize().
441 #define DEFAULT_MAPSIZE 1048576
443 /** @defgroup readers Reader Lock Table
444 * Readers don't acquire any locks for their data access. Instead, they
445 * simply record their transaction ID in the reader table. The reader
446 * mutex is needed just to find an empty slot in the reader table. The
447 * slot's address is saved in thread-specific data so that subsequent read
448 * transactions started by the same thread need no further locking to proceed.
450 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
452 * No reader table is used if the database is on a read-only filesystem, or
453 * if #MDB_NOLOCK is set.
455 * Since the database uses multi-version concurrency control, readers don't
456 * actually need any locking. This table is used to keep track of which
457 * readers are using data from which old transactions, so that we'll know
458 * when a particular old transaction is no longer in use. Old transactions
459 * that have discarded any data pages can then have those pages reclaimed
460 * for use by a later write transaction.
462 * The lock table is constructed such that reader slots are aligned with the
463 * processor's cache line size. Any slot is only ever used by one thread.
464 * This alignment guarantees that there will be no contention or cache
465 * thrashing as threads update their own slot info, and also eliminates
466 * any need for locking when accessing a slot.
468 * A writer thread will scan every slot in the table to determine the oldest
469 * outstanding reader transaction. Any freed pages older than this will be
470 * reclaimed by the writer. The writer doesn't use any locks when scanning
471 * this table. This means that there's no guarantee that the writer will
472 * see the most up-to-date reader info, but that's not required for correct
473 * operation - all we need is to know the upper bound on the oldest reader,
474 * we don't care at all about the newest reader. So the only consequence of
475 * reading stale information here is that old pages might hang around a
476 * while longer before being reclaimed. That's actually good anyway, because
477 * the longer we delay reclaiming old pages, the more likely it is that a
478 * string of contiguous pages can be found after coalescing old pages from
479 * many old transactions together.
482 /** Number of slots in the reader table.
483 * This value was chosen somewhat arbitrarily. 126 readers plus a
484 * couple mutexes fit exactly into 8KB on my development machine.
485 * Applications should set the table size using #mdb_env_set_maxreaders().
487 #define DEFAULT_READERS 126
489 /** The size of a CPU cache line in bytes. We want our lock structures
490 * aligned to this size to avoid false cache line sharing in the
492 * This value works for most CPUs. For Itanium this should be 128.
498 /** The information we store in a single slot of the reader table.
499 * In addition to a transaction ID, we also record the process and
500 * thread ID that owns a slot, so that we can detect stale information,
501 * e.g. threads or processes that went away without cleaning up.
502 * @note We currently don't check for stale records. We simply re-init
503 * the table when we know that we're the only process opening the
506 typedef struct MDB_rxbody {
507 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
508 * Multiple readers that start at the same time will probably have the
509 * same ID here. Again, it's not important to exclude them from
510 * anything; all we need to know is which version of the DB they
511 * started from so we can avoid overwriting any data used in that
512 * particular version.
515 /** The process ID of the process owning this reader txn. */
517 /** The thread ID of the thread owning this txn. */
521 /** The actual reader record, with cacheline padding. */
522 typedef struct MDB_reader {
525 /** shorthand for mrb_txnid */
526 #define mr_txnid mru.mrx.mrb_txnid
527 #define mr_pid mru.mrx.mrb_pid
528 #define mr_tid mru.mrx.mrb_tid
529 /** cache line alignment */
530 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
534 /** The header for the reader table.
535 * The table resides in a memory-mapped file. (This is a different file
536 * than is used for the main database.)
538 * For POSIX the actual mutexes reside in the shared memory of this
539 * mapped file. On Windows, mutexes are named objects allocated by the
540 * kernel; we store the mutex names in this mapped file so that other
541 * processes can grab them. This same approach is also used on
542 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
543 * process-shared POSIX mutexes. For these cases where a named object
544 * is used, the object name is derived from a 64 bit FNV hash of the
545 * environment pathname. As such, naming collisions are extremely
546 * unlikely. If a collision occurs, the results are unpredictable.
548 typedef struct MDB_txbody {
549 /** Stamp identifying this as an MDB file. It must be set
552 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
554 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
555 char mtb_rmname[MNAME_LEN];
557 /** Mutex protecting access to this table.
558 * This is the reader lock that #LOCK_MUTEX_R acquires.
560 pthread_mutex_t mtb_mutex;
562 /** The ID of the last transaction committed to the database.
563 * This is recorded here only for convenience; the value can always
564 * be determined by reading the main database meta pages.
567 /** The number of slots that have been used in the reader table.
568 * This always records the maximum count, it is not decremented
569 * when readers release their slots.
571 unsigned mtb_numreaders;
574 /** The actual reader table definition. */
575 typedef struct MDB_txninfo {
578 #define mti_magic mt1.mtb.mtb_magic
579 #define mti_format mt1.mtb.mtb_format
580 #define mti_mutex mt1.mtb.mtb_mutex
581 #define mti_rmname mt1.mtb.mtb_rmname
582 #define mti_txnid mt1.mtb.mtb_txnid
583 #define mti_numreaders mt1.mtb.mtb_numreaders
584 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
587 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
588 char mt2_wmname[MNAME_LEN];
589 #define mti_wmname mt2.mt2_wmname
591 pthread_mutex_t mt2_wmutex;
592 #define mti_wmutex mt2.mt2_wmutex
594 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
596 MDB_reader mti_readers[1];
599 /** Lockfile format signature: version, features and field layout */
600 #define MDB_LOCK_FORMAT \
602 ((MDB_LOCK_VERSION) \
603 /* Flags which describe functionality */ \
604 + (((MDB_PIDLOCK) != 0) << 16)))
607 /** Common header for all page types.
608 * Overflow records occupy a number of contiguous pages with no
609 * headers on any page after the first.
611 typedef struct MDB_page {
612 #define mp_pgno mp_p.p_pgno
613 #define mp_next mp_p.p_next
615 pgno_t p_pgno; /**< page number */
616 void * p_next; /**< for in-memory list of freed structs */
619 /** @defgroup mdb_page Page Flags
621 * Flags for the page headers.
624 #define P_BRANCH 0x01 /**< branch page */
625 #define P_LEAF 0x02 /**< leaf page */
626 #define P_OVERFLOW 0x04 /**< overflow page */
627 #define P_META 0x08 /**< meta page */
628 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
629 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
630 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
631 #define P_KEEP 0x8000 /**< leave this page alone during spill */
633 uint16_t mp_flags; /**< @ref mdb_page */
634 #define mp_lower mp_pb.pb.pb_lower
635 #define mp_upper mp_pb.pb.pb_upper
636 #define mp_pages mp_pb.pb_pages
639 indx_t pb_lower; /**< lower bound of free space */
640 indx_t pb_upper; /**< upper bound of free space */
642 uint32_t pb_pages; /**< number of overflow pages */
644 indx_t mp_ptrs[1]; /**< dynamic size */
647 /** Size of the page header, excluding dynamic data at the end */
648 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
650 /** Address of first usable data byte in a page, after the header */
651 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
653 /** Number of nodes on a page */
654 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
656 /** The amount of space remaining in the page */
657 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
659 /** The percentage of space used in the page, in tenths of a percent. */
660 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
661 ((env)->me_psize - PAGEHDRSZ))
662 /** The minimum page fill factor, in tenths of a percent.
663 * Pages emptier than this are candidates for merging.
665 #define FILL_THRESHOLD 250
667 /** Test if a page is a leaf page */
668 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
669 /** Test if a page is a LEAF2 page */
670 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
671 /** Test if a page is a branch page */
672 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
673 /** Test if a page is an overflow page */
674 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
675 /** Test if a page is a sub page */
676 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
678 /** The number of overflow pages needed to store the given size. */
679 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
681 /** Header for a single key/data pair within a page.
682 * We guarantee 2-byte alignment for nodes.
684 typedef struct MDB_node {
685 /** lo and hi are used for data size on leaf nodes and for
686 * child pgno on branch nodes. On 64 bit platforms, flags
687 * is also used for pgno. (Branch nodes have no flags).
688 * They are in host byte order in case that lets some
689 * accesses be optimized into a 32-bit word access.
691 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
692 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
693 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
694 /** @defgroup mdb_node Node Flags
696 * Flags for node headers.
699 #define F_BIGDATA 0x01 /**< data put on overflow page */
700 #define F_SUBDATA 0x02 /**< data is a sub-database */
701 #define F_DUPDATA 0x04 /**< data has duplicates */
703 /** valid flags for #mdb_node_add() */
704 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
707 unsigned short mn_flags; /**< @ref mdb_node */
708 unsigned short mn_ksize; /**< key size */
709 char mn_data[1]; /**< key and data are appended here */
712 /** Size of the node header, excluding dynamic data at the end */
713 #define NODESIZE offsetof(MDB_node, mn_data)
715 /** Bit position of top word in page number, for shifting mn_flags */
716 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
718 /** Size of a node in a branch page with a given key.
719 * This is just the node header plus the key, there is no data.
721 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
723 /** Size of a node in a leaf page with a given key and data.
724 * This is node header plus key plus data size.
726 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
728 /** Address of node \b i in page \b p */
729 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
731 /** Address of the key for the node */
732 #define NODEKEY(node) (void *)((node)->mn_data)
734 /** Address of the data for a node */
735 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
737 /** Get the page number pointed to by a branch node */
738 #define NODEPGNO(node) \
739 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
740 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
741 /** Set the page number in a branch node */
742 #define SETPGNO(node,pgno) do { \
743 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
744 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
746 /** Get the size of the data in a leaf node */
747 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
748 /** Set the size of the data for a leaf node */
749 #define SETDSZ(node,size) do { \
750 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
751 /** The size of a key in a node */
752 #define NODEKSZ(node) ((node)->mn_ksize)
754 /** Copy a page number from src to dst */
756 #define COPY_PGNO(dst,src) dst = src
758 #if SIZE_MAX > 4294967295UL
759 #define COPY_PGNO(dst,src) do { \
760 unsigned short *s, *d; \
761 s = (unsigned short *)&(src); \
762 d = (unsigned short *)&(dst); \
769 #define COPY_PGNO(dst,src) do { \
770 unsigned short *s, *d; \
771 s = (unsigned short *)&(src); \
772 d = (unsigned short *)&(dst); \
778 /** The address of a key in a LEAF2 page.
779 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
780 * There are no node headers, keys are stored contiguously.
782 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
784 /** Set the \b node's key into \b keyptr, if requested. */
785 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
786 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
788 /** Set the \b node's key into \b key. */
789 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
791 /** Information about a single database in the environment. */
792 typedef struct MDB_db {
793 uint32_t md_pad; /**< also ksize for LEAF2 pages */
794 uint16_t md_flags; /**< @ref mdb_dbi_open */
795 uint16_t md_depth; /**< depth of this tree */
796 pgno_t md_branch_pages; /**< number of internal pages */
797 pgno_t md_leaf_pages; /**< number of leaf pages */
798 pgno_t md_overflow_pages; /**< number of overflow pages */
799 size_t md_entries; /**< number of data items */
800 pgno_t md_root; /**< the root page of this tree */
803 /** mdb_dbi_open flags */
804 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
805 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
806 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
807 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
809 /** Handle for the DB used to track free pages. */
811 /** Handle for the default DB. */
814 /** Meta page content.
815 * A meta page is the start point for accessing a database snapshot.
816 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
818 typedef struct MDB_meta {
819 /** Stamp identifying this as an MDB file. It must be set
822 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
824 void *mm_address; /**< address for fixed mapping */
825 size_t mm_mapsize; /**< size of mmap region */
826 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
827 /** The size of pages used in this DB */
828 #define mm_psize mm_dbs[0].md_pad
829 /** Any persistent environment flags. @ref mdb_env */
830 #define mm_flags mm_dbs[0].md_flags
831 pgno_t mm_last_pg; /**< last used page in file */
832 txnid_t mm_txnid; /**< txnid that committed this page */
835 /** Buffer for a stack-allocated meta page.
836 * The members define size and alignment, and silence type
837 * aliasing warnings. They are not used directly; that could
838 * mean incorrectly using several union members in parallel.
840 typedef union MDB_metabuf {
843 char mm_pad[PAGEHDRSZ];
848 /** Auxiliary DB info.
849 * The information here is mostly static/read-only. There is
850 * only a single copy of this record in the environment.
852 typedef struct MDB_dbx {
853 MDB_val md_name; /**< name of the database */
854 MDB_cmp_func *md_cmp; /**< function for comparing keys */
855 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
856 MDB_rel_func *md_rel; /**< user relocate function */
857 void *md_relctx; /**< user-provided context for md_rel */
860 /** A database transaction.
861 * Every operation requires a transaction handle.
864 MDB_txn *mt_parent; /**< parent of a nested txn */
865 MDB_txn *mt_child; /**< nested txn under this txn */
866 pgno_t mt_next_pgno; /**< next unallocated page */
867 /** The ID of this transaction. IDs are integers incrementing from 1.
868 * Only committed write transactions increment the ID. If a transaction
869 * aborts, the ID may be re-used by the next writer.
872 MDB_env *mt_env; /**< the DB environment */
873 /** The list of pages that became unused during this transaction.
876 /** The sorted list of dirty pages we temporarily wrote to disk
877 * because the dirty list was full. page numbers in here are
878 * shifted left by 1, deleted slots have the LSB set.
880 MDB_IDL mt_spill_pgs;
882 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
884 /** For read txns: This thread/txn's reader table slot, or NULL. */
887 /** Array of records for each DB known in the environment. */
889 /** Array of MDB_db records for each known DB */
891 /** @defgroup mt_dbflag Transaction DB Flags
895 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
896 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
897 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
898 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
900 /** In write txns, array of cursors for each DB */
901 MDB_cursor **mt_cursors;
902 /** Array of flags for each DB */
903 unsigned char *mt_dbflags;
904 /** Number of DB records in use. This number only ever increments;
905 * we don't decrement it when individual DB handles are closed.
909 /** @defgroup mdb_txn Transaction Flags
913 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
914 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
915 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
916 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
918 unsigned int mt_flags; /**< @ref mdb_txn */
919 /** dirty_list room: Array size - #dirty pages visible to this txn.
920 * Includes ancestor txns' dirty pages not hidden by other txns'
921 * dirty/spilled pages. Thus commit(nested txn) has room to merge
922 * dirty_list into mt_parent after freeing hidden mt_parent pages.
924 unsigned int mt_dirty_room;
927 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
928 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
929 * raise this on a 64 bit machine.
931 #define CURSOR_STACK 32
935 /** Cursors are used for all DB operations.
936 * A cursor holds a path of (page pointer, key index) from the DB
937 * root to a position in the DB, plus other state. #MDB_DUPSORT
938 * cursors include an xcursor to the current data item. Write txns
939 * track their cursors and keep them up to date when data moves.
940 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
941 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
944 /** Next cursor on this DB in this txn */
946 /** Backup of the original cursor if this cursor is a shadow */
947 MDB_cursor *mc_backup;
948 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
949 struct MDB_xcursor *mc_xcursor;
950 /** The transaction that owns this cursor */
952 /** The database handle this cursor operates on */
954 /** The database record for this cursor */
956 /** The database auxiliary record for this cursor */
958 /** The @ref mt_dbflag for this database */
959 unsigned char *mc_dbflag;
960 unsigned short mc_snum; /**< number of pushed pages */
961 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
962 /** @defgroup mdb_cursor Cursor Flags
964 * Cursor state flags.
967 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
968 #define C_EOF 0x02 /**< No more data */
969 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
970 #define C_DEL 0x08 /**< last op was a cursor_del */
971 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
972 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
974 unsigned int mc_flags; /**< @ref mdb_cursor */
975 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
976 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
979 /** Context for sorted-dup records.
980 * We could have gone to a fully recursive design, with arbitrarily
981 * deep nesting of sub-databases. But for now we only handle these
982 * levels - main DB, optional sub-DB, sorted-duplicate DB.
984 typedef struct MDB_xcursor {
985 /** A sub-cursor for traversing the Dup DB */
986 MDB_cursor mx_cursor;
987 /** The database record for this Dup DB */
989 /** The auxiliary DB record for this Dup DB */
991 /** The @ref mt_dbflag for this Dup DB */
992 unsigned char mx_dbflag;
995 /** State of FreeDB old pages, stored in the MDB_env */
996 typedef struct MDB_pgstate {
997 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
998 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1001 /** The database environment. */
1003 HANDLE me_fd; /**< The main data file */
1004 HANDLE me_lfd; /**< The lock file */
1005 HANDLE me_mfd; /**< just for writing the meta pages */
1006 /** Failed to update the meta page. Probably an I/O error. */
1007 #define MDB_FATAL_ERROR 0x80000000U
1008 /** Some fields are initialized. */
1009 #define MDB_ENV_ACTIVE 0x20000000U
1010 /** me_txkey is set */
1011 #define MDB_ENV_TXKEY 0x10000000U
1012 /** Have liveness lock in reader table */
1013 #define MDB_LIVE_READER 0x08000000U
1014 uint32_t me_flags; /**< @ref mdb_env */
1015 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1016 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1017 unsigned int me_maxreaders; /**< size of the reader table */
1018 unsigned int me_numreaders; /**< max numreaders set by this env */
1019 MDB_dbi me_numdbs; /**< number of DBs opened */
1020 MDB_dbi me_maxdbs; /**< size of the DB table */
1021 MDB_PID_T me_pid; /**< process ID of this env */
1022 char *me_path; /**< path to the DB files */
1023 char *me_map; /**< the memory map of the data file */
1024 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1025 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1026 void *me_pbuf; /**< scratch area for DUPSORT put() */
1027 MDB_txn *me_txn; /**< current write transaction */
1028 size_t me_mapsize; /**< size of the data memory map */
1029 off_t me_size; /**< current file size */
1030 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1031 MDB_dbx *me_dbxs; /**< array of static DB info */
1032 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1033 pthread_key_t me_txkey; /**< thread-key for readers */
1034 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1035 # define me_pglast me_pgstate.mf_pglast
1036 # define me_pghead me_pgstate.mf_pghead
1037 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1038 /** IDL of pages that became unused in a write txn */
1039 MDB_IDL me_free_pgs;
1040 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1041 MDB_ID2L me_dirty_list;
1042 /** Max number of freelist items that can fit in a single overflow page */
1044 /** Max size of a node on a page */
1045 unsigned int me_nodemax;
1047 int me_pidquery; /**< Used in OpenProcess */
1048 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1050 #elif defined(MDB_USE_POSIX_SEM)
1051 sem_t *me_rmutex; /* Shared mutexes are not supported */
1056 /** Nested transaction */
1057 typedef struct MDB_ntxn {
1058 MDB_txn mnt_txn; /**< the transaction */
1059 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1062 /** max number of pages to commit in one writev() call */
1063 #define MDB_COMMIT_PAGES 64
1064 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1065 #undef MDB_COMMIT_PAGES
1066 #define MDB_COMMIT_PAGES IOV_MAX
1069 /* max bytes to write in one call */
1070 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1072 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1073 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1074 static int mdb_page_touch(MDB_cursor *mc);
1076 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1077 static int mdb_page_search_root(MDB_cursor *mc,
1078 MDB_val *key, int modify);
1079 #define MDB_PS_MODIFY 1
1080 #define MDB_PS_ROOTONLY 2
1081 #define MDB_PS_FIRST 4
1082 #define MDB_PS_LAST 8
1083 static int mdb_page_search(MDB_cursor *mc,
1084 MDB_val *key, int flags);
1085 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1087 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1088 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1089 pgno_t newpgno, unsigned int nflags);
1091 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1092 static int mdb_env_pick_meta(const MDB_env *env);
1093 static int mdb_env_write_meta(MDB_txn *txn);
1094 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1095 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1097 static void mdb_env_close0(MDB_env *env, int excl);
1099 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1100 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1101 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1102 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1103 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1104 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1105 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1106 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1107 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1109 static int mdb_rebalance(MDB_cursor *mc);
1110 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1112 static void mdb_cursor_pop(MDB_cursor *mc);
1113 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1115 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1116 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1117 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1118 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1119 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1121 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1122 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1124 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1125 static void mdb_xcursor_init0(MDB_cursor *mc);
1126 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1128 static int mdb_drop0(MDB_cursor *mc, int subs);
1129 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1132 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1136 static SECURITY_DESCRIPTOR mdb_null_sd;
1137 static SECURITY_ATTRIBUTES mdb_all_sa;
1138 static int mdb_sec_inited;
1141 /** Return the library version info. */
1143 mdb_version(int *major, int *minor, int *patch)
1145 if (major) *major = MDB_VERSION_MAJOR;
1146 if (minor) *minor = MDB_VERSION_MINOR;
1147 if (patch) *patch = MDB_VERSION_PATCH;
1148 return MDB_VERSION_STRING;
1151 /** Table of descriptions for MDB @ref errors */
1152 static char *const mdb_errstr[] = {
1153 "MDB_KEYEXIST: Key/data pair already exists",
1154 "MDB_NOTFOUND: No matching key/data pair found",
1155 "MDB_PAGE_NOTFOUND: Requested page not found",
1156 "MDB_CORRUPTED: Located page was wrong type",
1157 "MDB_PANIC: Update of meta page failed",
1158 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1159 "MDB_INVALID: File is not an MDB file",
1160 "MDB_MAP_FULL: Environment mapsize limit reached",
1161 "MDB_DBS_FULL: Environment maxdbs limit reached",
1162 "MDB_READERS_FULL: Environment maxreaders limit reached",
1163 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1164 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1165 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1166 "MDB_PAGE_FULL: Internal error - page has no more space",
1167 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1168 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1169 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1170 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1171 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1175 mdb_strerror(int err)
1179 return ("Successful return: 0");
1181 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1182 i = err - MDB_KEYEXIST;
1183 return mdb_errstr[i];
1186 return strerror(err);
1190 /** Display a key in hexadecimal and return the address of the result.
1191 * @param[in] key the key to display
1192 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1193 * @return The key in hexadecimal form.
1196 mdb_dkey(MDB_val *key, char *buf)
1199 unsigned char *c = key->mv_data;
1205 if (key->mv_size > MDB_MAXKEYSIZE)
1206 return "MDB_MAXKEYSIZE";
1207 /* may want to make this a dynamic check: if the key is mostly
1208 * printable characters, print it as-is instead of converting to hex.
1212 for (i=0; i<key->mv_size; i++)
1213 ptr += sprintf(ptr, "%02x", *c++);
1215 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1220 /** Display all the keys in the page. */
1222 mdb_page_list(MDB_page *mp)
1225 unsigned int i, nkeys, nsize, total = 0;
1229 nkeys = NUMKEYS(mp);
1230 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1231 for (i=0; i<nkeys; i++) {
1232 node = NODEPTR(mp, i);
1233 key.mv_size = node->mn_ksize;
1234 key.mv_data = node->mn_data;
1235 nsize = NODESIZE + key.mv_size;
1236 if (IS_BRANCH(mp)) {
1237 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1241 if (F_ISSET(node->mn_flags, F_BIGDATA))
1242 nsize += sizeof(pgno_t);
1244 nsize += NODEDSZ(node);
1246 nsize += sizeof(indx_t);
1247 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1249 total = EVEN(total);
1251 fprintf(stderr, "Total: %d\n", total);
1255 mdb_cursor_chk(MDB_cursor *mc)
1261 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1262 for (i=0; i<mc->mc_top; i++) {
1264 node = NODEPTR(mp, mc->mc_ki[i]);
1265 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1268 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1274 /** Count all the pages in each DB and in the freelist
1275 * and make sure it matches the actual number of pages
1278 static void mdb_audit(MDB_txn *txn)
1282 MDB_ID freecount, count;
1287 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1288 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1289 freecount += *(MDB_ID *)data.mv_data;
1292 for (i = 0; i<txn->mt_numdbs; i++) {
1294 mdb_cursor_init(&mc, txn, i, &mx);
1295 if (txn->mt_dbs[i].md_root == P_INVALID)
1297 count += txn->mt_dbs[i].md_branch_pages +
1298 txn->mt_dbs[i].md_leaf_pages +
1299 txn->mt_dbs[i].md_overflow_pages;
1300 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1301 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1305 mp = mc.mc_pg[mc.mc_top];
1306 for (j=0; j<NUMKEYS(mp); j++) {
1307 MDB_node *leaf = NODEPTR(mp, j);
1308 if (leaf->mn_flags & F_SUBDATA) {
1310 memcpy(&db, NODEDATA(leaf), sizeof(db));
1311 count += db.md_branch_pages + db.md_leaf_pages +
1312 db.md_overflow_pages;
1316 while (mdb_cursor_sibling(&mc, 1) == 0);
1319 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1320 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1321 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1327 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1329 return txn->mt_dbxs[dbi].md_cmp(a, b);
1333 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1335 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1338 /** Allocate memory for a page.
1339 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1342 mdb_page_malloc(MDB_txn *txn, unsigned num)
1344 MDB_env *env = txn->mt_env;
1345 MDB_page *ret = env->me_dpages;
1346 size_t psize = env->me_psize, sz = psize, off;
1347 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1348 * For a single page alloc, we init everything after the page header.
1349 * For multi-page, we init the final page; if the caller needed that
1350 * many pages they will be filling in at least up to the last page.
1354 VGMEMP_ALLOC(env, ret, sz);
1355 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1356 env->me_dpages = ret->mp_next;
1359 psize -= off = PAGEHDRSZ;
1364 if ((ret = malloc(sz)) != NULL) {
1365 if (!(env->me_flags & MDB_NOMEMINIT)) {
1366 memset((char *)ret + off, 0, psize);
1369 VGMEMP_ALLOC(env, ret, sz);
1374 /** Free a single page.
1375 * Saves single pages to a list, for future reuse.
1376 * (This is not used for multi-page overflow pages.)
1379 mdb_page_free(MDB_env *env, MDB_page *mp)
1381 mp->mp_next = env->me_dpages;
1382 VGMEMP_FREE(env, mp);
1383 env->me_dpages = mp;
1386 /** Free a dirty page */
1388 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1390 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1391 mdb_page_free(env, dp);
1393 /* large pages just get freed directly */
1394 VGMEMP_FREE(env, dp);
1399 /** Return all dirty pages to dpage list */
1401 mdb_dlist_free(MDB_txn *txn)
1403 MDB_env *env = txn->mt_env;
1404 MDB_ID2L dl = txn->mt_u.dirty_list;
1405 unsigned i, n = dl[0].mid;
1407 for (i = 1; i <= n; i++) {
1408 mdb_dpage_free(env, dl[i].mptr);
1413 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1414 * @param[in] mc A cursor handle for the current operation.
1415 * @param[in] pflags Flags of the pages to update:
1416 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1417 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1418 * @return 0 on success, non-zero on failure.
1421 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1423 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1424 MDB_txn *txn = mc->mc_txn;
1430 int rc = MDB_SUCCESS, level;
1432 /* Mark pages seen by cursors */
1433 if (mc->mc_flags & C_UNTRACK)
1434 mc = NULL; /* will find mc in mt_cursors */
1435 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1436 for (; mc; mc=mc->mc_next) {
1437 if (!(mc->mc_flags & C_INITIALIZED))
1439 for (m3 = mc;; m3 = &mx->mx_cursor) {
1441 for (j=0; j<m3->mc_snum; j++) {
1443 if ((mp->mp_flags & Mask) == pflags)
1444 mp->mp_flags ^= P_KEEP;
1446 mx = m3->mc_xcursor;
1447 /* Proceed to mx if it is at a sub-database */
1448 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1450 if (! (mp && (mp->mp_flags & P_LEAF)))
1452 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1453 if (!(leaf->mn_flags & F_SUBDATA))
1462 /* Mark dirty root pages */
1463 for (i=0; i<txn->mt_numdbs; i++) {
1464 if (txn->mt_dbflags[i] & DB_DIRTY) {
1465 pgno_t pgno = txn->mt_dbs[i].md_root;
1466 if (pgno == P_INVALID)
1468 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1470 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1471 dp->mp_flags ^= P_KEEP;
1479 static int mdb_page_flush(MDB_txn *txn, int keep);
1481 /** Spill pages from the dirty list back to disk.
1482 * This is intended to prevent running into #MDB_TXN_FULL situations,
1483 * but note that they may still occur in a few cases:
1484 * 1) our estimate of the txn size could be too small. Currently this
1485 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1486 * 2) child txns may run out of space if their parents dirtied a
1487 * lot of pages and never spilled them. TODO: we probably should do
1488 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1489 * the parent's dirty_room is below a given threshold.
1491 * Otherwise, if not using nested txns, it is expected that apps will
1492 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1493 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1494 * If the txn never references them again, they can be left alone.
1495 * If the txn only reads them, they can be used without any fuss.
1496 * If the txn writes them again, they can be dirtied immediately without
1497 * going thru all of the work of #mdb_page_touch(). Such references are
1498 * handled by #mdb_page_unspill().
1500 * Also note, we never spill DB root pages, nor pages of active cursors,
1501 * because we'll need these back again soon anyway. And in nested txns,
1502 * we can't spill a page in a child txn if it was already spilled in a
1503 * parent txn. That would alter the parent txns' data even though
1504 * the child hasn't committed yet, and we'd have no way to undo it if
1505 * the child aborted.
1507 * @param[in] m0 cursor A cursor handle identifying the transaction and
1508 * database for which we are checking space.
1509 * @param[in] key For a put operation, the key being stored.
1510 * @param[in] data For a put operation, the data being stored.
1511 * @return 0 on success, non-zero on failure.
1514 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1516 MDB_txn *txn = m0->mc_txn;
1518 MDB_ID2L dl = txn->mt_u.dirty_list;
1519 unsigned int i, j, need;
1522 if (m0->mc_flags & C_SUB)
1525 /* Estimate how much space this op will take */
1526 i = m0->mc_db->md_depth;
1527 /* Named DBs also dirty the main DB */
1528 if (m0->mc_dbi > MAIN_DBI)
1529 i += txn->mt_dbs[MAIN_DBI].md_depth;
1530 /* For puts, roughly factor in the key+data size */
1532 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1533 i += i; /* double it for good measure */
1536 if (txn->mt_dirty_room > i)
1539 if (!txn->mt_spill_pgs) {
1540 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1541 if (!txn->mt_spill_pgs)
1544 /* purge deleted slots */
1545 MDB_IDL sl = txn->mt_spill_pgs;
1546 unsigned int num = sl[0];
1548 for (i=1; i<=num; i++) {
1555 /* Preserve pages which may soon be dirtied again */
1556 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1559 /* Less aggressive spill - we originally spilled the entire dirty list,
1560 * with a few exceptions for cursor pages and DB root pages. But this
1561 * turns out to be a lot of wasted effort because in a large txn many
1562 * of those pages will need to be used again. So now we spill only 1/8th
1563 * of the dirty pages. Testing revealed this to be a good tradeoff,
1564 * better than 1/2, 1/4, or 1/10.
1566 if (need < MDB_IDL_UM_MAX / 8)
1567 need = MDB_IDL_UM_MAX / 8;
1569 /* Save the page IDs of all the pages we're flushing */
1570 /* flush from the tail forward, this saves a lot of shifting later on. */
1571 for (i=dl[0].mid; i && need; i--) {
1572 MDB_ID pn = dl[i].mid << 1;
1574 if (dp->mp_flags & P_KEEP)
1576 /* Can't spill twice, make sure it's not already in a parent's
1579 if (txn->mt_parent) {
1581 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1582 if (tx2->mt_spill_pgs) {
1583 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1584 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1585 dp->mp_flags |= P_KEEP;
1593 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1597 mdb_midl_sort(txn->mt_spill_pgs);
1599 /* Flush the spilled part of dirty list */
1600 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1603 /* Reset any dirty pages we kept that page_flush didn't see */
1604 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1607 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1611 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1613 mdb_find_oldest(MDB_txn *txn)
1616 txnid_t mr, oldest = txn->mt_txnid - 1;
1617 if (txn->mt_env->me_txns) {
1618 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1619 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1630 /** Add a page to the txn's dirty list */
1632 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1635 int (*insert)(MDB_ID2L, MDB_ID2 *);
1637 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1638 insert = mdb_mid2l_append;
1640 insert = mdb_mid2l_insert;
1642 mid.mid = mp->mp_pgno;
1644 insert(txn->mt_u.dirty_list, &mid);
1645 txn->mt_dirty_room--;
1648 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1649 * me_pghead and mt_next_pgno.
1651 * If there are free pages available from older transactions, they
1652 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1653 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1654 * and move me_pglast to say which records were consumed. Only this
1655 * function can create me_pghead and move me_pglast/mt_next_pgno.
1656 * @param[in] mc cursor A cursor handle identifying the transaction and
1657 * database for which we are allocating.
1658 * @param[in] num the number of pages to allocate.
1659 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1660 * will always be satisfied by a single contiguous chunk of memory.
1661 * @return 0 on success, non-zero on failure.
1664 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1666 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1667 /* Get at most <Max_retries> more freeDB records once me_pghead
1668 * has enough pages. If not enough, use new pages from the map.
1669 * If <Paranoid> and mc is updating the freeDB, only get new
1670 * records if me_pghead is empty. Then the freelist cannot play
1671 * catch-up with itself by growing while trying to save it.
1673 enum { Paranoid = 1, Max_retries = 500 };
1675 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1677 int rc, retry = Max_retries;
1678 MDB_txn *txn = mc->mc_txn;
1679 MDB_env *env = txn->mt_env;
1680 pgno_t pgno, *mop = env->me_pghead;
1681 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1683 txnid_t oldest = 0, last;
1689 /* If our dirty list is already full, we can't do anything */
1690 if (txn->mt_dirty_room == 0)
1691 return MDB_TXN_FULL;
1693 for (op = MDB_FIRST;; op = MDB_NEXT) {
1696 pgno_t *idl, old_id, new_id;
1698 /* Seek a big enough contiguous page range. Prefer
1699 * pages at the tail, just truncating the list.
1705 if (mop[i-n2] == pgno+n2)
1708 if (Max_retries < INT_MAX && --retry < 0)
1712 if (op == MDB_FIRST) { /* 1st iteration */
1713 /* Prepare to fetch more and coalesce */
1714 oldest = mdb_find_oldest(txn);
1715 last = env->me_pglast;
1716 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1719 key.mv_data = &last; /* will look up last+1 */
1720 key.mv_size = sizeof(last);
1722 if (Paranoid && mc->mc_dbi == FREE_DBI)
1725 if (Paranoid && retry < 0 && mop_len)
1729 /* Do not fetch more if the record will be too recent */
1732 rc = mdb_cursor_get(&m2, &key, NULL, op);
1734 if (rc == MDB_NOTFOUND)
1738 last = *(txnid_t*)key.mv_data;
1741 np = m2.mc_pg[m2.mc_top];
1742 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1743 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1746 idl = (MDB_ID *) data.mv_data;
1749 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1752 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1754 mop = env->me_pghead;
1756 env->me_pglast = last;
1758 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1759 last, txn->mt_dbs[FREE_DBI].md_root, i));
1761 DPRINTF(("IDL %"Z"u", idl[k]));
1763 /* Merge in descending sorted order */
1766 mop[0] = (pgno_t)-1;
1770 for (; old_id < new_id; old_id = mop[--j])
1777 /* Use new pages from the map when nothing suitable in the freeDB */
1779 pgno = txn->mt_next_pgno;
1780 if (pgno + num >= env->me_maxpg) {
1781 DPUTS("DB size maxed out");
1782 return MDB_MAP_FULL;
1786 if (env->me_flags & MDB_WRITEMAP) {
1787 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1789 if (!(np = mdb_page_malloc(txn, num)))
1793 mop[0] = mop_len -= num;
1794 /* Move any stragglers down */
1795 for (j = i-num; j < mop_len; )
1796 mop[++j] = mop[++i];
1798 txn->mt_next_pgno = pgno + num;
1801 mdb_page_dirty(txn, np);
1807 /** Copy the used portions of a non-overflow page.
1808 * @param[in] dst page to copy into
1809 * @param[in] src page to copy from
1810 * @param[in] psize size of a page
1813 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1815 enum { Align = sizeof(pgno_t) };
1816 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1818 /* If page isn't full, just copy the used portion. Adjust
1819 * alignment so memcpy may copy words instead of bytes.
1821 if ((unused &= -Align) && !IS_LEAF2(src)) {
1823 memcpy(dst, src, (lower + (Align-1)) & -Align);
1824 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1827 memcpy(dst, src, psize - unused);
1831 /** Pull a page off the txn's spill list, if present.
1832 * If a page being referenced was spilled to disk in this txn, bring
1833 * it back and make it dirty/writable again.
1834 * @param[in] txn the transaction handle.
1835 * @param[in] mp the page being referenced.
1836 * @param[out] ret the writable page, if any. ret is unchanged if
1837 * mp wasn't spilled.
1840 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1842 MDB_env *env = txn->mt_env;
1845 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1847 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1848 if (!tx2->mt_spill_pgs)
1850 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1851 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1854 if (txn->mt_dirty_room == 0)
1855 return MDB_TXN_FULL;
1856 if (IS_OVERFLOW(mp))
1860 if (env->me_flags & MDB_WRITEMAP) {
1863 np = mdb_page_malloc(txn, num);
1867 memcpy(np, mp, num * env->me_psize);
1869 mdb_page_copy(np, mp, env->me_psize);
1872 /* If in current txn, this page is no longer spilled.
1873 * If it happens to be the last page, truncate the spill list.
1874 * Otherwise mark it as deleted by setting the LSB.
1876 if (x == txn->mt_spill_pgs[0])
1877 txn->mt_spill_pgs[0]--;
1879 txn->mt_spill_pgs[x] |= 1;
1880 } /* otherwise, if belonging to a parent txn, the
1881 * page remains spilled until child commits
1884 mdb_page_dirty(txn, np);
1885 np->mp_flags |= P_DIRTY;
1893 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1894 * @param[in] mc cursor pointing to the page to be touched
1895 * @return 0 on success, non-zero on failure.
1898 mdb_page_touch(MDB_cursor *mc)
1900 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1901 MDB_txn *txn = mc->mc_txn;
1902 MDB_cursor *m2, *m3;
1906 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1907 if (txn->mt_flags & MDB_TXN_SPILLS) {
1909 rc = mdb_page_unspill(txn, mp, &np);
1915 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1916 (rc = mdb_page_alloc(mc, 1, &np)))
1919 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
1920 mp->mp_pgno, pgno));
1921 assert(mp->mp_pgno != pgno);
1922 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1923 /* Update the parent page, if any, to point to the new page */
1925 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1926 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1927 SETPGNO(node, pgno);
1929 mc->mc_db->md_root = pgno;
1931 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1932 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1934 /* If txn has a parent, make sure the page is in our
1938 unsigned x = mdb_mid2l_search(dl, pgno);
1939 if (x <= dl[0].mid && dl[x].mid == pgno) {
1940 if (mp != dl[x].mptr) { /* bad cursor? */
1941 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1942 return MDB_CORRUPTED;
1947 assert(dl[0].mid < MDB_IDL_UM_MAX);
1949 np = mdb_page_malloc(txn, 1);
1954 mdb_mid2l_insert(dl, &mid);
1959 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1961 np->mp_flags |= P_DIRTY;
1964 /* Adjust cursors pointing to mp */
1965 mc->mc_pg[mc->mc_top] = np;
1966 m2 = txn->mt_cursors[mc->mc_dbi];
1967 if (mc->mc_flags & C_SUB) {
1968 for (; m2; m2=m2->mc_next) {
1969 m3 = &m2->mc_xcursor->mx_cursor;
1970 if (m3->mc_snum < mc->mc_snum) continue;
1971 if (m3->mc_pg[mc->mc_top] == mp)
1972 m3->mc_pg[mc->mc_top] = np;
1975 for (; m2; m2=m2->mc_next) {
1976 if (m2->mc_snum < mc->mc_snum) continue;
1977 if (m2->mc_pg[mc->mc_top] == mp) {
1978 m2->mc_pg[mc->mc_top] = np;
1979 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1980 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1982 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1983 if (!(leaf->mn_flags & F_SUBDATA))
1984 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1993 mdb_env_sync(MDB_env *env, int force)
1996 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1997 if (env->me_flags & MDB_WRITEMAP) {
1998 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1999 ? MS_ASYNC : MS_SYNC;
2000 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2003 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2007 if (MDB_FDATASYNC(env->me_fd))
2014 /** Back up parent txn's cursors, then grab the originals for tracking */
2016 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2018 MDB_cursor *mc, *bk;
2023 for (i = src->mt_numdbs; --i >= 0; ) {
2024 if ((mc = src->mt_cursors[i]) != NULL) {
2025 size = sizeof(MDB_cursor);
2027 size += sizeof(MDB_xcursor);
2028 for (; mc; mc = bk->mc_next) {
2034 mc->mc_db = &dst->mt_dbs[i];
2035 /* Kill pointers into src - and dst to reduce abuse: The
2036 * user may not use mc until dst ends. Otherwise we'd...
2038 mc->mc_txn = NULL; /* ...set this to dst */
2039 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2040 if ((mx = mc->mc_xcursor) != NULL) {
2041 *(MDB_xcursor *)(bk+1) = *mx;
2042 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2044 mc->mc_next = dst->mt_cursors[i];
2045 dst->mt_cursors[i] = mc;
2052 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2053 * @param[in] txn the transaction handle.
2054 * @param[in] merge true to keep changes to parent cursors, false to revert.
2055 * @return 0 on success, non-zero on failure.
2058 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2060 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2064 for (i = txn->mt_numdbs; --i >= 0; ) {
2065 for (mc = cursors[i]; mc; mc = next) {
2067 if ((bk = mc->mc_backup) != NULL) {
2069 /* Commit changes to parent txn */
2070 mc->mc_next = bk->mc_next;
2071 mc->mc_backup = bk->mc_backup;
2072 mc->mc_txn = bk->mc_txn;
2073 mc->mc_db = bk->mc_db;
2074 mc->mc_dbflag = bk->mc_dbflag;
2075 if ((mx = mc->mc_xcursor) != NULL)
2076 mx->mx_cursor.mc_txn = bk->mc_txn;
2078 /* Abort nested txn */
2080 if ((mx = mc->mc_xcursor) != NULL)
2081 *mx = *(MDB_xcursor *)(bk+1);
2085 /* Only malloced cursors are permanently tracked. */
2093 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2096 mdb_txn_reset0(MDB_txn *txn, const char *act);
2098 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2104 Pidset = F_SETLK, Pidcheck = F_GETLK
2108 /** Set or check a pid lock. Set returns 0 on success.
2109 * Check returns 0 if the process is certainly dead, nonzero if it may
2110 * be alive (the lock exists or an error happened so we do not know).
2112 * On Windows Pidset is a no-op, we merely check for the existence
2113 * of the process with the given pid. On POSIX we use a single byte
2114 * lock on the lockfile, set at an offset equal to the pid.
2117 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2119 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2122 if (op == Pidcheck) {
2123 h = OpenProcess(env->me_pidquery, FALSE, pid);
2124 /* No documented "no such process" code, but other program use this: */
2126 return ErrCode() != ERROR_INVALID_PARAMETER;
2127 /* A process exists until all handles to it close. Has it exited? */
2128 ret = WaitForSingleObject(h, 0) != 0;
2135 struct flock lock_info;
2136 memset(&lock_info, 0, sizeof(lock_info));
2137 lock_info.l_type = F_WRLCK;
2138 lock_info.l_whence = SEEK_SET;
2139 lock_info.l_start = pid;
2140 lock_info.l_len = 1;
2141 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2142 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2144 } else if ((rc = ErrCode()) == EINTR) {
2152 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2153 * @param[in] txn the transaction handle to initialize
2154 * @return 0 on success, non-zero on failure.
2157 mdb_txn_renew0(MDB_txn *txn)
2159 MDB_env *env = txn->mt_env;
2160 MDB_txninfo *ti = env->me_txns;
2164 int rc, new_notls = 0;
2167 txn->mt_numdbs = env->me_numdbs;
2168 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2170 if (txn->mt_flags & MDB_TXN_RDONLY) {
2172 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2173 txn->mt_txnid = meta->mm_txnid;
2174 txn->mt_u.reader = NULL;
2176 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2177 pthread_getspecific(env->me_txkey);
2179 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2180 return MDB_BAD_RSLOT;
2182 MDB_PID_T pid = env->me_pid;
2183 pthread_t tid = pthread_self();
2185 if (!(env->me_flags & MDB_LIVE_READER)) {
2186 rc = mdb_reader_pid(env, Pidset, pid);
2188 UNLOCK_MUTEX_R(env);
2191 env->me_flags |= MDB_LIVE_READER;
2195 nr = ti->mti_numreaders;
2196 for (i=0; i<nr; i++)
2197 if (ti->mti_readers[i].mr_pid == 0)
2199 if (i == env->me_maxreaders) {
2200 UNLOCK_MUTEX_R(env);
2201 return MDB_READERS_FULL;
2203 ti->mti_readers[i].mr_pid = pid;
2204 ti->mti_readers[i].mr_tid = tid;
2206 ti->mti_numreaders = ++nr;
2207 /* Save numreaders for un-mutexed mdb_env_close() */
2208 env->me_numreaders = nr;
2209 UNLOCK_MUTEX_R(env);
2211 r = &ti->mti_readers[i];
2212 new_notls = (env->me_flags & MDB_NOTLS);
2213 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2218 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2219 txn->mt_u.reader = r;
2220 meta = env->me_metas[txn->mt_txnid & 1];
2226 txn->mt_txnid = ti->mti_txnid;
2227 meta = env->me_metas[txn->mt_txnid & 1];
2229 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2230 txn->mt_txnid = meta->mm_txnid;
2234 if (txn->mt_txnid == mdb_debug_start)
2237 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2238 txn->mt_u.dirty_list = env->me_dirty_list;
2239 txn->mt_u.dirty_list[0].mid = 0;
2240 txn->mt_free_pgs = env->me_free_pgs;
2241 txn->mt_free_pgs[0] = 0;
2242 txn->mt_spill_pgs = NULL;
2246 /* Copy the DB info and flags */
2247 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2249 /* Moved to here to avoid a data race in read TXNs */
2250 txn->mt_next_pgno = meta->mm_last_pg+1;
2252 for (i=2; i<txn->mt_numdbs; i++) {
2253 x = env->me_dbflags[i];
2254 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2255 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2257 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2259 if (env->me_maxpg < txn->mt_next_pgno) {
2260 mdb_txn_reset0(txn, "renew0-mapfail");
2262 txn->mt_u.reader->mr_pid = 0;
2263 txn->mt_u.reader = NULL;
2265 return MDB_MAP_RESIZED;
2272 mdb_txn_renew(MDB_txn *txn)
2276 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2279 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2280 DPUTS("environment had fatal error, must shutdown!");
2284 rc = mdb_txn_renew0(txn);
2285 if (rc == MDB_SUCCESS) {
2286 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2287 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2288 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2294 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2298 int rc, size, tsize = sizeof(MDB_txn);
2300 if (env->me_flags & MDB_FATAL_ERROR) {
2301 DPUTS("environment had fatal error, must shutdown!");
2304 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2307 /* Nested transactions: Max 1 child, write txns only, no writemap */
2308 if (parent->mt_child ||
2309 (flags & MDB_RDONLY) ||
2310 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2311 (env->me_flags & MDB_WRITEMAP))
2313 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2315 tsize = sizeof(MDB_ntxn);
2317 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2318 if (!(flags & MDB_RDONLY))
2319 size += env->me_maxdbs * sizeof(MDB_cursor *);
2321 if ((txn = calloc(1, size)) == NULL) {
2322 DPRINTF(("calloc: %s", strerror(ErrCode())));
2325 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2326 if (flags & MDB_RDONLY) {
2327 txn->mt_flags |= MDB_TXN_RDONLY;
2328 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2330 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2331 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2337 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2338 if (!txn->mt_u.dirty_list ||
2339 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2341 free(txn->mt_u.dirty_list);
2345 txn->mt_txnid = parent->mt_txnid;
2346 txn->mt_dirty_room = parent->mt_dirty_room;
2347 txn->mt_u.dirty_list[0].mid = 0;
2348 txn->mt_spill_pgs = NULL;
2349 txn->mt_next_pgno = parent->mt_next_pgno;
2350 parent->mt_child = txn;
2351 txn->mt_parent = parent;
2352 txn->mt_numdbs = parent->mt_numdbs;
2353 txn->mt_flags = parent->mt_flags;
2354 txn->mt_dbxs = parent->mt_dbxs;
2355 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2356 /* Copy parent's mt_dbflags, but clear DB_NEW */
2357 for (i=0; i<txn->mt_numdbs; i++)
2358 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2360 ntxn = (MDB_ntxn *)txn;
2361 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2362 if (env->me_pghead) {
2363 size = MDB_IDL_SIZEOF(env->me_pghead);
2364 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2366 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2371 rc = mdb_cursor_shadow(parent, txn);
2373 mdb_txn_reset0(txn, "beginchild-fail");
2375 rc = mdb_txn_renew0(txn);
2381 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2382 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2383 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2390 mdb_txn_env(MDB_txn *txn)
2392 if(!txn) return NULL;
2396 /** Export or close DBI handles opened in this txn. */
2398 mdb_dbis_update(MDB_txn *txn, int keep)
2401 MDB_dbi n = txn->mt_numdbs;
2402 MDB_env *env = txn->mt_env;
2403 unsigned char *tdbflags = txn->mt_dbflags;
2405 for (i = n; --i >= 2;) {
2406 if (tdbflags[i] & DB_NEW) {
2408 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2410 char *ptr = env->me_dbxs[i].md_name.mv_data;
2411 env->me_dbxs[i].md_name.mv_data = NULL;
2412 env->me_dbxs[i].md_name.mv_size = 0;
2413 env->me_dbflags[i] = 0;
2418 if (keep && env->me_numdbs < n)
2422 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2423 * May be called twice for readonly txns: First reset it, then abort.
2424 * @param[in] txn the transaction handle to reset
2425 * @param[in] act why the transaction is being reset
2428 mdb_txn_reset0(MDB_txn *txn, const char *act)
2430 MDB_env *env = txn->mt_env;
2432 /* Close any DBI handles opened in this txn */
2433 mdb_dbis_update(txn, 0);
2435 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2436 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2437 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2439 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2440 if (txn->mt_u.reader) {
2441 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2442 if (!(env->me_flags & MDB_NOTLS))
2443 txn->mt_u.reader = NULL; /* txn does not own reader */
2445 txn->mt_numdbs = 0; /* close nothing if called again */
2446 txn->mt_dbxs = NULL; /* mark txn as reset */
2448 mdb_cursors_close(txn, 0);
2450 if (!(env->me_flags & MDB_WRITEMAP)) {
2451 mdb_dlist_free(txn);
2453 mdb_midl_free(env->me_pghead);
2455 if (txn->mt_parent) {
2456 txn->mt_parent->mt_child = NULL;
2457 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2458 mdb_midl_free(txn->mt_free_pgs);
2459 mdb_midl_free(txn->mt_spill_pgs);
2460 free(txn->mt_u.dirty_list);
2464 if (mdb_midl_shrink(&txn->mt_free_pgs))
2465 env->me_free_pgs = txn->mt_free_pgs;
2466 env->me_pghead = NULL;
2470 /* The writer mutex was locked in mdb_txn_begin. */
2472 UNLOCK_MUTEX_W(env);
2477 mdb_txn_reset(MDB_txn *txn)
2482 /* This call is only valid for read-only txns */
2483 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2486 mdb_txn_reset0(txn, "reset");
2490 mdb_txn_abort(MDB_txn *txn)
2496 mdb_txn_abort(txn->mt_child);
2498 mdb_txn_reset0(txn, "abort");
2499 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2500 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2501 txn->mt_u.reader->mr_pid = 0;
2506 /** Save the freelist as of this transaction to the freeDB.
2507 * This changes the freelist. Keep trying until it stabilizes.
2510 mdb_freelist_save(MDB_txn *txn)
2512 /* env->me_pghead[] can grow and shrink during this call.
2513 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2514 * Page numbers cannot disappear from txn->mt_free_pgs[].
2517 MDB_env *env = txn->mt_env;
2518 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2519 txnid_t pglast = 0, head_id = 0;
2520 pgno_t freecnt = 0, *free_pgs, *mop;
2521 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2523 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2525 if (env->me_pghead) {
2526 /* Make sure first page of freeDB is touched and on freelist */
2527 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2528 if (rc && rc != MDB_NOTFOUND)
2532 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2533 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2534 ? SSIZE_MAX : maxfree_1pg;
2537 /* Come back here after each Put() in case freelist changed */
2542 /* If using records from freeDB which we have not yet
2543 * deleted, delete them and any we reserved for me_pghead.
2545 while (pglast < env->me_pglast) {
2546 rc = mdb_cursor_first(&mc, &key, NULL);
2549 pglast = head_id = *(txnid_t *)key.mv_data;
2550 total_room = head_room = 0;
2551 assert(pglast <= env->me_pglast);
2552 rc = mdb_cursor_del(&mc, 0);
2557 /* Save the IDL of pages freed by this txn, to a single record */
2558 if (freecnt < txn->mt_free_pgs[0]) {
2560 /* Make sure last page of freeDB is touched and on freelist */
2561 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2562 if (rc && rc != MDB_NOTFOUND)
2565 free_pgs = txn->mt_free_pgs;
2566 /* Write to last page of freeDB */
2567 key.mv_size = sizeof(txn->mt_txnid);
2568 key.mv_data = &txn->mt_txnid;
2570 freecnt = free_pgs[0];
2571 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2572 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2575 /* Retry if mt_free_pgs[] grew during the Put() */
2576 free_pgs = txn->mt_free_pgs;
2577 } while (freecnt < free_pgs[0]);
2578 mdb_midl_sort(free_pgs);
2579 memcpy(data.mv_data, free_pgs, data.mv_size);
2582 unsigned int i = free_pgs[0];
2583 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2584 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2586 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2592 mop = env->me_pghead;
2593 mop_len = mop ? mop[0] : 0;
2595 /* Reserve records for me_pghead[]. Split it if multi-page,
2596 * to avoid searching freeDB for a page range. Use keys in
2597 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2599 if (total_room >= mop_len) {
2600 if (total_room == mop_len || --more < 0)
2602 } else if (head_room >= maxfree_1pg && head_id > 1) {
2603 /* Keep current record (overflow page), add a new one */
2607 /* (Re)write {key = head_id, IDL length = head_room} */
2608 total_room -= head_room;
2609 head_room = mop_len - total_room;
2610 if (head_room > maxfree_1pg && head_id > 1) {
2611 /* Overflow multi-page for part of me_pghead */
2612 head_room /= head_id; /* amortize page sizes */
2613 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2614 } else if (head_room < 0) {
2615 /* Rare case, not bothering to delete this record */
2618 key.mv_size = sizeof(head_id);
2619 key.mv_data = &head_id;
2620 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2621 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2624 /* IDL is initially empty, zero out at least the length */
2625 pgs = (pgno_t *)data.mv_data;
2626 j = head_room > clean_limit ? head_room : 0;
2630 total_room += head_room;
2633 /* Fill in the reserved me_pghead records */
2639 rc = mdb_cursor_first(&mc, &key, &data);
2640 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2641 unsigned flags = MDB_CURRENT;
2642 txnid_t id = *(txnid_t *)key.mv_data;
2643 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2646 assert(len >= 0 && id <= env->me_pglast);
2648 if (len > mop_len) {
2650 data.mv_size = (len + 1) * sizeof(MDB_ID);
2653 data.mv_data = mop -= len;
2656 rc = mdb_cursor_put(&mc, &key, &data, flags);
2658 if (rc || !(mop_len -= len))
2665 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2666 * @param[in] txn the transaction that's being committed
2667 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2668 * @return 0 on success, non-zero on failure.
2671 mdb_page_flush(MDB_txn *txn, int keep)
2673 MDB_env *env = txn->mt_env;
2674 MDB_ID2L dl = txn->mt_u.dirty_list;
2675 unsigned psize = env->me_psize, j;
2676 int i, pagecount = dl[0].mid, rc;
2677 size_t size = 0, pos = 0;
2679 MDB_page *dp = NULL;
2683 struct iovec iov[MDB_COMMIT_PAGES];
2684 ssize_t wpos = 0, wsize = 0, wres;
2685 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2691 if (env->me_flags & MDB_WRITEMAP) {
2692 /* Clear dirty flags */
2693 while (++i <= pagecount) {
2695 /* Don't flush this page yet */
2696 if (dp->mp_flags & P_KEEP) {
2697 dp->mp_flags ^= P_KEEP;
2701 dp->mp_flags &= ~P_DIRTY;
2706 /* Write the pages */
2708 if (++i <= pagecount) {
2710 /* Don't flush this page yet */
2711 if (dp->mp_flags & P_KEEP) {
2712 dp->mp_flags ^= P_KEEP;
2717 /* clear dirty flag */
2718 dp->mp_flags &= ~P_DIRTY;
2721 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2726 /* Windows actually supports scatter/gather I/O, but only on
2727 * unbuffered file handles. Since we're relying on the OS page
2728 * cache for all our data, that's self-defeating. So we just
2729 * write pages one at a time. We use the ov structure to set
2730 * the write offset, to at least save the overhead of a Seek
2733 DPRINTF(("committing page %"Z"u", pgno));
2734 memset(&ov, 0, sizeof(ov));
2735 ov.Offset = pos & 0xffffffff;
2736 ov.OffsetHigh = pos >> 16 >> 16;
2737 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2739 DPRINTF(("WriteFile: %d", rc));
2743 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2744 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2746 /* Write previous page(s) */
2747 #ifdef MDB_USE_PWRITEV
2748 wres = pwritev(env->me_fd, iov, n, wpos);
2751 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2753 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2755 DPRINTF(("lseek: %s", strerror(rc)));
2758 wres = writev(env->me_fd, iov, n);
2761 if (wres != wsize) {
2764 DPRINTF(("Write error: %s", strerror(rc)));
2766 rc = EIO; /* TODO: Use which error code? */
2767 DPUTS("short write, filesystem full?");
2778 DPRINTF(("committing page %"Z"u", pgno));
2779 next_pos = pos + size;
2780 iov[n].iov_len = size;
2781 iov[n].iov_base = (char *)dp;
2787 for (i = keep; ++i <= pagecount; ) {
2789 /* This is a page we skipped above */
2792 dl[j].mid = dp->mp_pgno;
2795 mdb_dpage_free(env, dp);
2800 txn->mt_dirty_room += i - j;
2806 mdb_txn_commit(MDB_txn *txn)
2812 assert(txn != NULL);
2813 assert(txn->mt_env != NULL);
2815 if (txn->mt_child) {
2816 rc = mdb_txn_commit(txn->mt_child);
2817 txn->mt_child = NULL;
2824 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2825 mdb_dbis_update(txn, 1);
2826 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2831 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2832 DPUTS("error flag is set, can't commit");
2834 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2839 if (txn->mt_parent) {
2840 MDB_txn *parent = txn->mt_parent;
2843 unsigned x, y, len, ps_len;
2845 /* Append our free list to parent's */
2846 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2849 mdb_midl_free(txn->mt_free_pgs);
2850 /* Failures after this must either undo the changes
2851 * to the parent or set MDB_TXN_ERROR in the parent.
2854 parent->mt_next_pgno = txn->mt_next_pgno;
2855 parent->mt_flags = txn->mt_flags;
2857 /* Merge our cursors into parent's and close them */
2858 mdb_cursors_close(txn, 1);
2860 /* Update parent's DB table. */
2861 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2862 parent->mt_numdbs = txn->mt_numdbs;
2863 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2864 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2865 for (i=2; i<txn->mt_numdbs; i++) {
2866 /* preserve parent's DB_NEW status */
2867 x = parent->mt_dbflags[i] & DB_NEW;
2868 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2871 dst = parent->mt_u.dirty_list;
2872 src = txn->mt_u.dirty_list;
2873 /* Remove anything in our dirty list from parent's spill list */
2874 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2876 pspill[0] = (pgno_t)-1;
2877 /* Mark our dirty pages as deleted in parent spill list */
2878 for (i=0, len=src[0].mid; ++i <= len; ) {
2879 MDB_ID pn = src[i].mid << 1;
2880 while (pn > pspill[x])
2882 if (pn == pspill[x]) {
2887 /* Squash deleted pagenums if we deleted any */
2888 for (x=y; ++x <= ps_len; )
2889 if (!(pspill[x] & 1))
2890 pspill[++y] = pspill[x];
2894 /* Find len = length of merging our dirty list with parent's */
2896 dst[0].mid = 0; /* simplify loops */
2897 if (parent->mt_parent) {
2898 len = x + src[0].mid;
2899 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2900 for (i = x; y && i; y--) {
2901 pgno_t yp = src[y].mid;
2902 while (yp < dst[i].mid)
2904 if (yp == dst[i].mid) {
2909 } else { /* Simplify the above for single-ancestor case */
2910 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2912 /* Merge our dirty list with parent's */
2914 for (i = len; y; dst[i--] = src[y--]) {
2915 pgno_t yp = src[y].mid;
2916 while (yp < dst[x].mid)
2917 dst[i--] = dst[x--];
2918 if (yp == dst[x].mid)
2919 free(dst[x--].mptr);
2923 free(txn->mt_u.dirty_list);
2924 parent->mt_dirty_room = txn->mt_dirty_room;
2925 if (txn->mt_spill_pgs) {
2926 if (parent->mt_spill_pgs) {
2927 /* TODO: Prevent failure here, so parent does not fail */
2928 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2930 parent->mt_flags |= MDB_TXN_ERROR;
2931 mdb_midl_free(txn->mt_spill_pgs);
2932 mdb_midl_sort(parent->mt_spill_pgs);
2934 parent->mt_spill_pgs = txn->mt_spill_pgs;
2938 parent->mt_child = NULL;
2939 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2944 if (txn != env->me_txn) {
2945 DPUTS("attempt to commit unknown transaction");
2950 mdb_cursors_close(txn, 0);
2952 if (!txn->mt_u.dirty_list[0].mid &&
2953 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2956 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2957 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2959 /* Update DB root pointers */
2960 if (txn->mt_numdbs > 2) {
2964 data.mv_size = sizeof(MDB_db);
2966 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2967 for (i = 2; i < txn->mt_numdbs; i++) {
2968 if (txn->mt_dbflags[i] & DB_DIRTY) {
2969 data.mv_data = &txn->mt_dbs[i];
2970 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2977 rc = mdb_freelist_save(txn);
2981 mdb_midl_free(env->me_pghead);
2982 env->me_pghead = NULL;
2983 if (mdb_midl_shrink(&txn->mt_free_pgs))
2984 env->me_free_pgs = txn->mt_free_pgs;
2990 if ((rc = mdb_page_flush(txn, 0)) ||
2991 (rc = mdb_env_sync(env, 0)) ||
2992 (rc = mdb_env_write_meta(txn)))
2998 mdb_dbis_update(txn, 1);
3001 UNLOCK_MUTEX_W(env);
3011 /** Read the environment parameters of a DB environment before
3012 * mapping it into memory.
3013 * @param[in] env the environment handle
3014 * @param[out] meta address of where to store the meta information
3015 * @return 0 on success, non-zero on failure.
3018 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3024 enum { Size = sizeof(pbuf) };
3026 /* We don't know the page size yet, so use a minimum value.
3027 * Read both meta pages so we can use the latest one.
3030 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3034 memset(&ov, 0, sizeof(ov));
3036 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3037 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3040 rc = pread(env->me_fd, &pbuf, Size, off);
3043 if (rc == 0 && off == 0)
3045 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3046 DPRINTF(("read: %s", mdb_strerror(rc)));
3050 p = (MDB_page *)&pbuf;
3052 if (!F_ISSET(p->mp_flags, P_META)) {
3053 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3058 if (m->mm_magic != MDB_MAGIC) {
3059 DPUTS("meta has invalid magic");
3063 if (m->mm_version != MDB_DATA_VERSION) {
3064 DPRINTF(("database is version %u, expected version %u",
3065 m->mm_version, MDB_DATA_VERSION));
3066 return MDB_VERSION_MISMATCH;
3069 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3075 /** Write the environment parameters of a freshly created DB environment.
3076 * @param[in] env the environment handle
3077 * @param[out] meta address of where to store the meta information
3078 * @return 0 on success, non-zero on failure.
3081 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3089 memset(&ov, 0, sizeof(ov));
3090 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3092 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3095 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3096 len = pwrite(fd, ptr, size, pos); \
3097 rc = (len >= 0); } while(0)
3100 DPUTS("writing new meta page");
3102 psize = env->me_psize;
3104 meta->mm_magic = MDB_MAGIC;
3105 meta->mm_version = MDB_DATA_VERSION;
3106 meta->mm_mapsize = env->me_mapsize;
3107 meta->mm_psize = psize;
3108 meta->mm_last_pg = 1;
3109 meta->mm_flags = env->me_flags & 0xffff;
3110 meta->mm_flags |= MDB_INTEGERKEY;
3111 meta->mm_dbs[0].md_root = P_INVALID;
3112 meta->mm_dbs[1].md_root = P_INVALID;
3114 p = calloc(2, psize);
3116 p->mp_flags = P_META;
3117 *(MDB_meta *)METADATA(p) = *meta;
3119 q = (MDB_page *)((char *)p + psize);
3121 q->mp_flags = P_META;
3122 *(MDB_meta *)METADATA(q) = *meta;
3124 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3127 else if ((unsigned) len == psize * 2)
3135 /** Update the environment info to commit a transaction.
3136 * @param[in] txn the transaction that's being committed
3137 * @return 0 on success, non-zero on failure.
3140 mdb_env_write_meta(MDB_txn *txn)
3143 MDB_meta meta, metab, *mp;
3145 int rc, len, toggle;
3154 assert(txn != NULL);
3155 assert(txn->mt_env != NULL);
3157 toggle = txn->mt_txnid & 1;
3158 DPRINTF(("writing meta page %d for root page %"Z"u",
3159 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3162 mp = env->me_metas[toggle];
3164 if (env->me_flags & MDB_WRITEMAP) {
3165 /* Persist any increases of mapsize config */
3166 if (env->me_mapsize > mp->mm_mapsize)
3167 mp->mm_mapsize = env->me_mapsize;
3168 mp->mm_dbs[0] = txn->mt_dbs[0];
3169 mp->mm_dbs[1] = txn->mt_dbs[1];
3170 mp->mm_last_pg = txn->mt_next_pgno - 1;
3171 mp->mm_txnid = txn->mt_txnid;
3172 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3173 unsigned meta_size = env->me_psize;
3174 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3177 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3178 if (meta_size < env->me_os_psize)
3179 meta_size += meta_size;
3184 if (MDB_MSYNC(ptr, meta_size, rc)) {
3191 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3192 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3194 ptr = (char *)&meta;
3195 if (env->me_mapsize > mp->mm_mapsize) {
3196 /* Persist any increases of mapsize config */
3197 meta.mm_mapsize = env->me_mapsize;
3198 off = offsetof(MDB_meta, mm_mapsize);
3200 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3202 len = sizeof(MDB_meta) - off;
3205 meta.mm_dbs[0] = txn->mt_dbs[0];
3206 meta.mm_dbs[1] = txn->mt_dbs[1];
3207 meta.mm_last_pg = txn->mt_next_pgno - 1;
3208 meta.mm_txnid = txn->mt_txnid;
3211 off += env->me_psize;
3214 /* Write to the SYNC fd */
3215 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3216 env->me_fd : env->me_mfd;
3219 memset(&ov, 0, sizeof(ov));
3221 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3225 rc = pwrite(mfd, ptr, len, off);
3228 rc = rc < 0 ? ErrCode() : EIO;
3229 DPUTS("write failed, disk error?");
3230 /* On a failure, the pagecache still contains the new data.
3231 * Write some old data back, to prevent it from being used.
3232 * Use the non-SYNC fd; we know it will fail anyway.
3234 meta.mm_last_pg = metab.mm_last_pg;
3235 meta.mm_txnid = metab.mm_txnid;
3237 memset(&ov, 0, sizeof(ov));
3239 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3241 r2 = pwrite(env->me_fd, ptr, len, off);
3242 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3245 env->me_flags |= MDB_FATAL_ERROR;
3249 /* Memory ordering issues are irrelevant; since the entire writer
3250 * is wrapped by wmutex, all of these changes will become visible
3251 * after the wmutex is unlocked. Since the DB is multi-version,
3252 * readers will get consistent data regardless of how fresh or
3253 * how stale their view of these values is.
3256 env->me_txns->mti_txnid = txn->mt_txnid;
3261 /** Check both meta pages to see which one is newer.
3262 * @param[in] env the environment handle
3263 * @return meta toggle (0 or 1).
3266 mdb_env_pick_meta(const MDB_env *env)
3268 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3272 mdb_env_create(MDB_env **env)
3276 e = calloc(1, sizeof(MDB_env));
3280 e->me_maxreaders = DEFAULT_READERS;
3281 e->me_maxdbs = e->me_numdbs = 2;
3282 e->me_fd = INVALID_HANDLE_VALUE;
3283 e->me_lfd = INVALID_HANDLE_VALUE;
3284 e->me_mfd = INVALID_HANDLE_VALUE;
3285 #ifdef MDB_USE_POSIX_SEM
3286 e->me_rmutex = SEM_FAILED;
3287 e->me_wmutex = SEM_FAILED;
3289 e->me_pid = getpid();
3290 GET_PAGESIZE(e->me_os_psize);
3291 VGMEMP_CREATE(e,0,0);
3297 mdb_env_map(MDB_env *env, void *addr, int newsize)
3300 unsigned int flags = env->me_flags;
3304 LONG sizelo, sizehi;
3305 sizelo = env->me_mapsize & 0xffffffff;
3306 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3308 /* Windows won't create mappings for zero length files.
3309 * Just allocate the maxsize right now.
3312 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3313 || !SetEndOfFile(env->me_fd)
3314 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3317 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3318 PAGE_READWRITE : PAGE_READONLY,
3319 sizehi, sizelo, NULL);
3322 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3323 FILE_MAP_WRITE : FILE_MAP_READ,
3324 0, 0, env->me_mapsize, addr);
3325 rc = env->me_map ? 0 : ErrCode();
3330 int prot = PROT_READ;
3331 if (flags & MDB_WRITEMAP) {
3333 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3336 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3338 if (env->me_map == MAP_FAILED) {
3343 if (flags & MDB_NORDAHEAD) {
3344 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3346 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3348 #ifdef POSIX_MADV_RANDOM
3349 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3350 #endif /* POSIX_MADV_RANDOM */
3351 #endif /* MADV_RANDOM */
3355 /* Can happen because the address argument to mmap() is just a
3356 * hint. mmap() can pick another, e.g. if the range is in use.
3357 * The MAP_FIXED flag would prevent that, but then mmap could
3358 * instead unmap existing pages to make room for the new map.
3360 if (addr && env->me_map != addr)
3361 return EBUSY; /* TODO: Make a new MDB_* error code? */
3363 p = (MDB_page *)env->me_map;
3364 env->me_metas[0] = METADATA(p);
3365 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3371 mdb_env_set_mapsize(MDB_env *env, size_t size)
3373 /* If env is already open, caller is responsible for making
3374 * sure there are no active txns.
3382 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3383 else if (size < env->me_mapsize) {
3384 /* If the configured size is smaller, make sure it's
3385 * still big enough. Silently round up to minimum if not.
3387 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3391 munmap(env->me_map, env->me_mapsize);
3392 env->me_mapsize = size;
3393 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3394 rc = mdb_env_map(env, old, 1);
3398 env->me_mapsize = size;
3400 env->me_maxpg = env->me_mapsize / env->me_psize;
3405 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3409 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3414 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3416 if (env->me_map || readers < 1)
3418 env->me_maxreaders = readers;
3423 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3425 if (!env || !readers)
3427 *readers = env->me_maxreaders;
3431 /** Further setup required for opening an MDB environment
3434 mdb_env_open2(MDB_env *env)
3436 unsigned int flags = env->me_flags;
3437 int i, newenv = 0, rc;
3441 /* See if we should use QueryLimited */
3443 if ((rc & 0xff) > 5)
3444 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3446 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3449 memset(&meta, 0, sizeof(meta));
3451 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3454 DPUTS("new mdbenv");
3456 env->me_psize = env->me_os_psize;
3457 if (env->me_psize > MAX_PAGESIZE)
3458 env->me_psize = MAX_PAGESIZE;
3460 env->me_psize = meta.mm_psize;
3463 /* Was a mapsize configured? */
3464 if (!env->me_mapsize) {
3465 /* If this is a new environment, take the default,
3466 * else use the size recorded in the existing env.
3468 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3469 } else if (env->me_mapsize < meta.mm_mapsize) {
3470 /* If the configured size is smaller, make sure it's
3471 * still big enough. Silently round up to minimum if not.
3473 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3474 if (env->me_mapsize < minsize)
3475 env->me_mapsize = minsize;
3478 rc = mdb_env_map(env, meta.mm_address, newenv);
3483 if (flags & MDB_FIXEDMAP)
3484 meta.mm_address = env->me_map;
3485 i = mdb_env_init_meta(env, &meta);
3486 if (i != MDB_SUCCESS) {
3490 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3491 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3494 env->me_maxpg = env->me_mapsize / env->me_psize;
3497 int toggle = mdb_env_pick_meta(env);
3498 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3500 DPRINTF(("opened database version %u, pagesize %u",
3501 env->me_metas[0]->mm_version, env->me_psize));
3502 DPRINTF(("using meta page %d", toggle));
3503 DPRINTF(("depth: %u", db->md_depth));
3504 DPRINTF(("entries: %"Z"u", db->md_entries));
3505 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3506 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3507 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3508 DPRINTF(("root: %"Z"u", db->md_root));
3516 /** Release a reader thread's slot in the reader lock table.
3517 * This function is called automatically when a thread exits.
3518 * @param[in] ptr This points to the slot in the reader lock table.
3521 mdb_env_reader_dest(void *ptr)
3523 MDB_reader *reader = ptr;
3529 /** Junk for arranging thread-specific callbacks on Windows. This is
3530 * necessarily platform and compiler-specific. Windows supports up
3531 * to 1088 keys. Let's assume nobody opens more than 64 environments
3532 * in a single process, for now. They can override this if needed.
3534 #ifndef MAX_TLS_KEYS
3535 #define MAX_TLS_KEYS 64
3537 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3538 static int mdb_tls_nkeys;
3540 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3544 case DLL_PROCESS_ATTACH: break;
3545 case DLL_THREAD_ATTACH: break;
3546 case DLL_THREAD_DETACH:
3547 for (i=0; i<mdb_tls_nkeys; i++) {
3548 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3549 mdb_env_reader_dest(r);
3552 case DLL_PROCESS_DETACH: break;
3557 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3559 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3563 /* Force some symbol references.
3564 * _tls_used forces the linker to create the TLS directory if not already done
3565 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3567 #pragma comment(linker, "/INCLUDE:_tls_used")
3568 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3569 #pragma const_seg(".CRT$XLB")
3570 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3571 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3574 #pragma comment(linker, "/INCLUDE:__tls_used")
3575 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3576 #pragma data_seg(".CRT$XLB")
3577 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3579 #endif /* WIN 32/64 */
3580 #endif /* !__GNUC__ */
3583 /** Downgrade the exclusive lock on the region back to shared */
3585 mdb_env_share_locks(MDB_env *env, int *excl)
3587 int rc = 0, toggle = mdb_env_pick_meta(env);
3589 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3594 /* First acquire a shared lock. The Unlock will
3595 * then release the existing exclusive lock.
3597 memset(&ov, 0, sizeof(ov));
3598 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3601 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3607 struct flock lock_info;
3608 /* The shared lock replaces the existing lock */
3609 memset((void *)&lock_info, 0, sizeof(lock_info));
3610 lock_info.l_type = F_RDLCK;
3611 lock_info.l_whence = SEEK_SET;
3612 lock_info.l_start = 0;
3613 lock_info.l_len = 1;
3614 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3615 (rc = ErrCode()) == EINTR) ;
3616 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3623 /** Try to get exlusive lock, otherwise shared.
3624 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3627 mdb_env_excl_lock(MDB_env *env, int *excl)
3631 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3635 memset(&ov, 0, sizeof(ov));
3636 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3643 struct flock lock_info;
3644 memset((void *)&lock_info, 0, sizeof(lock_info));
3645 lock_info.l_type = F_WRLCK;
3646 lock_info.l_whence = SEEK_SET;
3647 lock_info.l_start = 0;
3648 lock_info.l_len = 1;
3649 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3650 (rc = ErrCode()) == EINTR) ;
3654 # ifdef MDB_USE_POSIX_SEM
3655 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3658 lock_info.l_type = F_RDLCK;
3659 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3660 (rc = ErrCode()) == EINTR) ;
3670 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3672 * @(#) $Revision: 5.1 $
3673 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3674 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3676 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3680 * Please do not copyright this code. This code is in the public domain.
3682 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3683 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3684 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3685 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3686 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3687 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3688 * PERFORMANCE OF THIS SOFTWARE.
3691 * chongo <Landon Curt Noll> /\oo/\
3692 * http://www.isthe.com/chongo/
3694 * Share and Enjoy! :-)
3697 typedef unsigned long long mdb_hash_t;
3698 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3700 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3701 * @param[in] val value to hash
3702 * @param[in] hval initial value for hash
3703 * @return 64 bit hash
3705 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3706 * hval arg on the first call.
3709 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3711 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3712 unsigned char *end = s + val->mv_size;
3714 * FNV-1a hash each octet of the string
3717 /* xor the bottom with the current octet */
3718 hval ^= (mdb_hash_t)*s++;
3720 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3721 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3722 (hval << 7) + (hval << 8) + (hval << 40);
3724 /* return our new hash value */
3728 /** Hash the string and output the encoded hash.
3729 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3730 * very short name limits. We don't care about the encoding being reversible,
3731 * we just want to preserve as many bits of the input as possible in a
3732 * small printable string.
3733 * @param[in] str string to hash
3734 * @param[out] encbuf an array of 11 chars to hold the hash
3736 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3739 mdb_pack85(unsigned long l, char *out)
3743 for (i=0; i<5; i++) {
3744 *out++ = mdb_a85[l % 85];
3750 mdb_hash_enc(MDB_val *val, char *encbuf)
3752 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3754 mdb_pack85(h, encbuf);
3755 mdb_pack85(h>>32, encbuf+5);
3760 /** Open and/or initialize the lock region for the environment.
3761 * @param[in] env The MDB environment.
3762 * @param[in] lpath The pathname of the file used for the lock region.
3763 * @param[in] mode The Unix permissions for the file, if we create it.
3764 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3765 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3766 * @return 0 on success, non-zero on failure.
3769 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3772 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3774 # define MDB_ERRCODE_ROFS EROFS
3775 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3776 # define MDB_CLOEXEC O_CLOEXEC
3779 # define MDB_CLOEXEC 0
3786 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3787 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3788 FILE_ATTRIBUTE_NORMAL, NULL);
3790 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3792 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3794 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3799 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3800 /* Lose record locks when exec*() */
3801 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3802 fcntl(env->me_lfd, F_SETFD, fdflags);
3805 if (!(env->me_flags & MDB_NOTLS)) {
3806 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3809 env->me_flags |= MDB_ENV_TXKEY;
3811 /* Windows TLS callbacks need help finding their TLS info. */
3812 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3816 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3820 /* Try to get exclusive lock. If we succeed, then
3821 * nobody is using the lock region and we should initialize it.
3823 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3826 size = GetFileSize(env->me_lfd, NULL);
3828 size = lseek(env->me_lfd, 0, SEEK_END);
3829 if (size == -1) goto fail_errno;
3831 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3832 if (size < rsize && *excl > 0) {
3834 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3835 || !SetEndOfFile(env->me_lfd))
3838 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3842 size = rsize - sizeof(MDB_txninfo);
3843 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3848 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3850 if (!mh) goto fail_errno;
3851 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3853 if (!env->me_txns) goto fail_errno;
3855 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3857 if (m == MAP_FAILED) goto fail_errno;
3863 BY_HANDLE_FILE_INFORMATION stbuf;
3872 if (!mdb_sec_inited) {
3873 InitializeSecurityDescriptor(&mdb_null_sd,
3874 SECURITY_DESCRIPTOR_REVISION);
3875 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3876 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3877 mdb_all_sa.bInheritHandle = FALSE;
3878 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3881 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3882 idbuf.volume = stbuf.dwVolumeSerialNumber;
3883 idbuf.nhigh = stbuf.nFileIndexHigh;
3884 idbuf.nlow = stbuf.nFileIndexLow;
3885 val.mv_data = &idbuf;
3886 val.mv_size = sizeof(idbuf);
3887 mdb_hash_enc(&val, encbuf);
3888 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3889 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3890 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3891 if (!env->me_rmutex) goto fail_errno;
3892 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3893 if (!env->me_wmutex) goto fail_errno;
3894 #elif defined(MDB_USE_POSIX_SEM)
3903 #if defined(__NetBSD__)
3904 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3906 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3907 idbuf.dev = stbuf.st_dev;
3908 idbuf.ino = stbuf.st_ino;
3909 val.mv_data = &idbuf;
3910 val.mv_size = sizeof(idbuf);
3911 mdb_hash_enc(&val, encbuf);
3912 #ifdef MDB_SHORT_SEMNAMES
3913 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3915 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3916 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3917 /* Clean up after a previous run, if needed: Try to
3918 * remove both semaphores before doing anything else.
3920 sem_unlink(env->me_txns->mti_rmname);
3921 sem_unlink(env->me_txns->mti_wmname);
3922 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3923 O_CREAT|O_EXCL, mode, 1);
3924 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3925 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3926 O_CREAT|O_EXCL, mode, 1);
3927 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3928 #else /* MDB_USE_POSIX_SEM */
3929 pthread_mutexattr_t mattr;
3931 if ((rc = pthread_mutexattr_init(&mattr))
3932 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3933 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3934 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3936 pthread_mutexattr_destroy(&mattr);
3937 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3939 env->me_txns->mti_magic = MDB_MAGIC;
3940 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3941 env->me_txns->mti_txnid = 0;
3942 env->me_txns->mti_numreaders = 0;
3945 if (env->me_txns->mti_magic != MDB_MAGIC) {
3946 DPUTS("lock region has invalid magic");
3950 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3951 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3952 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3953 rc = MDB_VERSION_MISMATCH;
3957 if (rc && rc != EACCES && rc != EAGAIN) {
3961 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3962 if (!env->me_rmutex) goto fail_errno;
3963 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3964 if (!env->me_wmutex) goto fail_errno;
3965 #elif defined(MDB_USE_POSIX_SEM)
3966 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3967 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3968 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3969 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3980 /** The name of the lock file in the DB environment */
3981 #define LOCKNAME "/lock.mdb"
3982 /** The name of the data file in the DB environment */
3983 #define DATANAME "/data.mdb"
3984 /** The suffix of the lock file when no subdir is used */
3985 #define LOCKSUFF "-lock"
3986 /** Only a subset of the @ref mdb_env flags can be changed
3987 * at runtime. Changing other flags requires closing the
3988 * environment and re-opening it with the new flags.
3990 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
3991 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
3992 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
3995 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3997 int oflags, rc, len, excl = -1;
3998 char *lpath, *dpath;
4000 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4004 if (flags & MDB_NOSUBDIR) {
4005 rc = len + sizeof(LOCKSUFF) + len + 1;
4007 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4012 if (flags & MDB_NOSUBDIR) {
4013 dpath = lpath + len + sizeof(LOCKSUFF);
4014 sprintf(lpath, "%s" LOCKSUFF, path);
4015 strcpy(dpath, path);
4017 dpath = lpath + len + sizeof(LOCKNAME);
4018 sprintf(lpath, "%s" LOCKNAME, path);
4019 sprintf(dpath, "%s" DATANAME, path);
4023 flags |= env->me_flags;
4024 if (flags & MDB_RDONLY) {
4025 /* silently ignore WRITEMAP when we're only getting read access */
4026 flags &= ~MDB_WRITEMAP;
4028 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4029 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4032 env->me_flags = flags |= MDB_ENV_ACTIVE;
4036 env->me_path = strdup(path);
4037 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4038 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4039 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4044 /* For RDONLY, get lockfile after we know datafile exists */
4045 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4046 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4052 if (F_ISSET(flags, MDB_RDONLY)) {
4053 oflags = GENERIC_READ;
4054 len = OPEN_EXISTING;
4056 oflags = GENERIC_READ|GENERIC_WRITE;
4059 mode = FILE_ATTRIBUTE_NORMAL;
4060 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4061 NULL, len, mode, NULL);
4063 if (F_ISSET(flags, MDB_RDONLY))
4066 oflags = O_RDWR | O_CREAT;
4068 env->me_fd = open(dpath, oflags, mode);
4070 if (env->me_fd == INVALID_HANDLE_VALUE) {
4075 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4076 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4081 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4082 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4083 env->me_mfd = env->me_fd;
4085 /* Synchronous fd for meta writes. Needed even with
4086 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4089 len = OPEN_EXISTING;
4090 env->me_mfd = CreateFile(dpath, oflags,
4091 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4092 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4095 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4097 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4102 DPRINTF(("opened dbenv %p", (void *) env));
4104 rc = mdb_env_share_locks(env, &excl);
4108 if (!((flags & MDB_RDONLY) ||
4109 (env->me_pbuf = calloc(1, env->me_psize))))
4115 mdb_env_close0(env, excl);
4121 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4123 mdb_env_close0(MDB_env *env, int excl)
4127 if (!(env->me_flags & MDB_ENV_ACTIVE))
4130 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4131 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4132 free(env->me_dbxs[i].md_name.mv_data);
4135 free(env->me_dbflags);
4138 free(env->me_dirty_list);
4139 mdb_midl_free(env->me_free_pgs);
4141 if (env->me_flags & MDB_ENV_TXKEY) {
4142 pthread_key_delete(env->me_txkey);
4144 /* Delete our key from the global list */
4145 for (i=0; i<mdb_tls_nkeys; i++)
4146 if (mdb_tls_keys[i] == env->me_txkey) {
4147 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4155 munmap(env->me_map, env->me_mapsize);
4157 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4158 (void) close(env->me_mfd);
4159 if (env->me_fd != INVALID_HANDLE_VALUE)
4160 (void) close(env->me_fd);
4162 MDB_PID_T pid = env->me_pid;
4163 /* Clearing readers is done in this function because
4164 * me_txkey with its destructor must be disabled first.
4166 for (i = env->me_numreaders; --i >= 0; )
4167 if (env->me_txns->mti_readers[i].mr_pid == pid)
4168 env->me_txns->mti_readers[i].mr_pid = 0;
4170 if (env->me_rmutex) {
4171 CloseHandle(env->me_rmutex);
4172 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4174 /* Windows automatically destroys the mutexes when
4175 * the last handle closes.
4177 #elif defined(MDB_USE_POSIX_SEM)
4178 if (env->me_rmutex != SEM_FAILED) {
4179 sem_close(env->me_rmutex);
4180 if (env->me_wmutex != SEM_FAILED)
4181 sem_close(env->me_wmutex);
4182 /* If we have the filelock: If we are the
4183 * only remaining user, clean up semaphores.
4186 mdb_env_excl_lock(env, &excl);
4188 sem_unlink(env->me_txns->mti_rmname);
4189 sem_unlink(env->me_txns->mti_wmname);
4193 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4195 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4198 /* Unlock the lockfile. Windows would have unlocked it
4199 * after closing anyway, but not necessarily at once.
4201 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4204 (void) close(env->me_lfd);
4207 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4211 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4213 MDB_txn *txn = NULL;
4219 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4223 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4226 /* Do the lock/unlock of the reader mutex before starting the
4227 * write txn. Otherwise other read txns could block writers.
4229 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4234 /* We must start the actual read txn after blocking writers */
4235 mdb_txn_reset0(txn, "reset-stage1");
4237 /* Temporarily block writers until we snapshot the meta pages */
4240 rc = mdb_txn_renew0(txn);
4242 UNLOCK_MUTEX_W(env);
4247 wsize = env->me_psize * 2;
4251 DO_WRITE(rc, fd, ptr, w2, len);
4255 } else if (len > 0) {
4261 /* Non-blocking or async handles are not supported */
4267 UNLOCK_MUTEX_W(env);
4272 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4274 if (wsize > MAX_WRITE)
4278 DO_WRITE(rc, fd, ptr, w2, len);
4282 } else if (len > 0) {
4299 mdb_env_copy(MDB_env *env, const char *path)
4303 HANDLE newfd = INVALID_HANDLE_VALUE;
4305 if (env->me_flags & MDB_NOSUBDIR) {
4306 lpath = (char *)path;
4309 len += sizeof(DATANAME);
4310 lpath = malloc(len);
4313 sprintf(lpath, "%s" DATANAME, path);
4316 /* The destination path must exist, but the destination file must not.
4317 * We don't want the OS to cache the writes, since the source data is
4318 * already in the OS cache.
4321 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4322 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4324 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4326 if (newfd == INVALID_HANDLE_VALUE) {
4332 /* Set O_DIRECT if the file system supports it */
4333 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4334 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4336 #ifdef F_NOCACHE /* __APPLE__ */
4337 rc = fcntl(newfd, F_NOCACHE, 1);
4344 rc = mdb_env_copyfd(env, newfd);
4347 if (!(env->me_flags & MDB_NOSUBDIR))
4349 if (newfd != INVALID_HANDLE_VALUE)
4350 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4357 mdb_env_close(MDB_env *env)
4364 VGMEMP_DESTROY(env);
4365 while ((dp = env->me_dpages) != NULL) {
4366 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4367 env->me_dpages = dp->mp_next;
4371 mdb_env_close0(env, 0);
4375 /** Compare two items pointing at aligned size_t's */
4377 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4379 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4380 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4383 /** Compare two items pointing at aligned unsigned int's */
4385 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4387 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4388 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4391 /** Compare two items pointing at unsigned ints of unknown alignment.
4392 * Nodes and keys are guaranteed to be 2-byte aligned.
4395 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4397 #if BYTE_ORDER == LITTLE_ENDIAN
4398 unsigned short *u, *c;
4401 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4402 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4405 } while(!x && u > (unsigned short *)a->mv_data);
4408 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4412 /** Compare two items lexically */
4414 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4421 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4427 diff = memcmp(a->mv_data, b->mv_data, len);
4428 return diff ? diff : len_diff<0 ? -1 : len_diff;
4431 /** Compare two items in reverse byte order */
4433 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4435 const unsigned char *p1, *p2, *p1_lim;
4439 p1_lim = (const unsigned char *)a->mv_data;
4440 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4441 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4443 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4449 while (p1 > p1_lim) {
4450 diff = *--p1 - *--p2;
4454 return len_diff<0 ? -1 : len_diff;
4457 /** Search for key within a page, using binary search.
4458 * Returns the smallest entry larger or equal to the key.
4459 * If exactp is non-null, stores whether the found entry was an exact match
4460 * in *exactp (1 or 0).
4461 * Updates the cursor index with the index of the found entry.
4462 * If no entry larger or equal to the key is found, returns NULL.
4465 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4467 unsigned int i = 0, nkeys;
4470 MDB_page *mp = mc->mc_pg[mc->mc_top];
4471 MDB_node *node = NULL;
4476 nkeys = NUMKEYS(mp);
4481 COPY_PGNO(pgno, mp->mp_pgno);
4482 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4483 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4490 low = IS_LEAF(mp) ? 0 : 1;
4492 cmp = mc->mc_dbx->md_cmp;
4494 /* Branch pages have no data, so if using integer keys,
4495 * alignment is guaranteed. Use faster mdb_cmp_int.
4497 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4498 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4505 nodekey.mv_size = mc->mc_db->md_pad;
4506 node = NODEPTR(mp, 0); /* fake */
4507 while (low <= high) {
4508 i = (low + high) >> 1;
4509 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4510 rc = cmp(key, &nodekey);
4511 DPRINTF(("found leaf index %u [%s], rc = %i",
4512 i, DKEY(&nodekey), rc));
4521 while (low <= high) {
4522 i = (low + high) >> 1;
4524 node = NODEPTR(mp, i);
4525 nodekey.mv_size = NODEKSZ(node);
4526 nodekey.mv_data = NODEKEY(node);
4528 rc = cmp(key, &nodekey);
4531 DPRINTF(("found leaf index %u [%s], rc = %i",
4532 i, DKEY(&nodekey), rc));
4534 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4535 i, DKEY(&nodekey), NODEPGNO(node), rc));
4546 if (rc > 0) { /* Found entry is less than the key. */
4547 i++; /* Skip to get the smallest entry larger than key. */
4549 node = NODEPTR(mp, i);
4552 *exactp = (rc == 0);
4553 /* store the key index */
4554 mc->mc_ki[mc->mc_top] = i;
4556 /* There is no entry larger or equal to the key. */
4559 /* nodeptr is fake for LEAF2 */
4565 mdb_cursor_adjust(MDB_cursor *mc, func)
4569 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4570 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4577 /** Pop a page off the top of the cursor's stack. */
4579 mdb_cursor_pop(MDB_cursor *mc)
4583 MDB_page *top = mc->mc_pg[mc->mc_top];
4589 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4590 DDBI(mc), (void *) mc));
4594 /** Push a page onto the top of the cursor's stack. */
4596 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4598 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4599 DDBI(mc), (void *) mc));
4601 if (mc->mc_snum >= CURSOR_STACK) {
4602 assert(mc->mc_snum < CURSOR_STACK);
4603 return MDB_CURSOR_FULL;
4606 mc->mc_top = mc->mc_snum++;
4607 mc->mc_pg[mc->mc_top] = mp;
4608 mc->mc_ki[mc->mc_top] = 0;
4613 /** Find the address of the page corresponding to a given page number.
4614 * @param[in] txn the transaction for this access.
4615 * @param[in] pgno the page number for the page to retrieve.
4616 * @param[out] ret address of a pointer where the page's address will be stored.
4617 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4618 * @return 0 on success, non-zero on failure.
4621 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4623 MDB_env *env = txn->mt_env;
4627 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4631 MDB_ID2L dl = tx2->mt_u.dirty_list;
4633 /* Spilled pages were dirtied in this txn and flushed
4634 * because the dirty list got full. Bring this page
4635 * back in from the map (but don't unspill it here,
4636 * leave that unless page_touch happens again).
4638 if (tx2->mt_spill_pgs) {
4639 MDB_ID pn = pgno << 1;
4640 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4641 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4642 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4647 unsigned x = mdb_mid2l_search(dl, pgno);
4648 if (x <= dl[0].mid && dl[x].mid == pgno) {
4654 } while ((tx2 = tx2->mt_parent) != NULL);
4657 if (pgno < txn->mt_next_pgno) {
4659 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4661 DPRINTF(("page %"Z"u not found", pgno));
4663 return MDB_PAGE_NOTFOUND;
4673 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4674 * The cursor is at the root page, set up the rest of it.
4677 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4679 MDB_page *mp = mc->mc_pg[mc->mc_top];
4683 while (IS_BRANCH(mp)) {
4687 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4688 assert(NUMKEYS(mp) > 1);
4689 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4691 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4693 if (flags & MDB_PS_LAST)
4694 i = NUMKEYS(mp) - 1;
4697 node = mdb_node_search(mc, key, &exact);
4699 i = NUMKEYS(mp) - 1;
4701 i = mc->mc_ki[mc->mc_top];
4707 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4710 assert(i < NUMKEYS(mp));
4711 node = NODEPTR(mp, i);
4713 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4716 mc->mc_ki[mc->mc_top] = i;
4717 if ((rc = mdb_cursor_push(mc, mp)))
4720 if (flags & MDB_PS_MODIFY) {
4721 if ((rc = mdb_page_touch(mc)) != 0)
4723 mp = mc->mc_pg[mc->mc_top];
4728 DPRINTF(("internal error, index points to a %02X page!?",
4730 return MDB_CORRUPTED;
4733 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4734 key ? DKEY(key) : "null"));
4735 mc->mc_flags |= C_INITIALIZED;
4736 mc->mc_flags &= ~C_EOF;
4741 /** Search for the lowest key under the current branch page.
4742 * This just bypasses a NUMKEYS check in the current page
4743 * before calling mdb_page_search_root(), because the callers
4744 * are all in situations where the current page is known to
4748 mdb_page_search_lowest(MDB_cursor *mc)
4750 MDB_page *mp = mc->mc_pg[mc->mc_top];
4751 MDB_node *node = NODEPTR(mp, 0);
4754 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4757 mc->mc_ki[mc->mc_top] = 0;
4758 if ((rc = mdb_cursor_push(mc, mp)))
4760 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4763 /** Search for the page a given key should be in.
4764 * Push it and its parent pages on the cursor stack.
4765 * @param[in,out] mc the cursor for this operation.
4766 * @param[in] key the key to search for, or NULL for first/last page.
4767 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4768 * are touched (updated with new page numbers).
4769 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4770 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4771 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4772 * @return 0 on success, non-zero on failure.
4775 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4780 /* Make sure the txn is still viable, then find the root from
4781 * the txn's db table and set it as the root of the cursor's stack.
4783 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4784 DPUTS("transaction has failed, must abort");
4787 /* Make sure we're using an up-to-date root */
4788 if (*mc->mc_dbflag & DB_STALE) {
4790 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4791 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4798 MDB_node *leaf = mdb_node_search(&mc2,
4799 &mc->mc_dbx->md_name, &exact);
4801 return MDB_NOTFOUND;
4802 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4805 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4807 /* The txn may not know this DBI, or another process may
4808 * have dropped and recreated the DB with other flags.
4810 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4811 return MDB_INCOMPATIBLE;
4812 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4814 *mc->mc_dbflag &= ~DB_STALE;
4816 root = mc->mc_db->md_root;
4818 if (root == P_INVALID) { /* Tree is empty. */
4819 DPUTS("tree is empty");
4820 return MDB_NOTFOUND;
4825 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4826 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4832 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4833 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4835 if (flags & MDB_PS_MODIFY) {
4836 if ((rc = mdb_page_touch(mc)))
4840 if (flags & MDB_PS_ROOTONLY)
4843 return mdb_page_search_root(mc, key, flags);
4847 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4849 MDB_txn *txn = mc->mc_txn;
4850 pgno_t pg = mp->mp_pgno;
4851 unsigned x = 0, ovpages = mp->mp_pages;
4852 MDB_env *env = txn->mt_env;
4853 MDB_IDL sl = txn->mt_spill_pgs;
4854 MDB_ID pn = pg << 1;
4857 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4858 /* If the page is dirty or on the spill list we just acquired it,
4859 * so we should give it back to our current free list, if any.
4860 * Otherwise put it onto the list of pages we freed in this txn.
4862 * Won't create me_pghead: me_pglast must be inited along with it.
4863 * Unsupported in nested txns: They would need to hide the page
4864 * range in ancestor txns' dirty and spilled lists.
4866 if (env->me_pghead &&
4868 ((mp->mp_flags & P_DIRTY) ||
4869 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4873 MDB_ID2 *dl, ix, iy;
4874 rc = mdb_midl_need(&env->me_pghead, ovpages);
4877 if (!(mp->mp_flags & P_DIRTY)) {
4878 /* This page is no longer spilled */
4885 /* Remove from dirty list */
4886 dl = txn->mt_u.dirty_list;
4888 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4896 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4897 txn->mt_flags |= MDB_TXN_ERROR;
4898 return MDB_CORRUPTED;
4901 if (!(env->me_flags & MDB_WRITEMAP))
4902 mdb_dpage_free(env, mp);
4904 /* Insert in me_pghead */
4905 mop = env->me_pghead;
4906 j = mop[0] + ovpages;
4907 for (i = mop[0]; i && mop[i] < pg; i--)
4913 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4917 mc->mc_db->md_overflow_pages -= ovpages;
4921 /** Return the data associated with a given node.
4922 * @param[in] txn The transaction for this operation.
4923 * @param[in] leaf The node being read.
4924 * @param[out] data Updated to point to the node's data.
4925 * @return 0 on success, non-zero on failure.
4928 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4930 MDB_page *omp; /* overflow page */
4934 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4935 data->mv_size = NODEDSZ(leaf);
4936 data->mv_data = NODEDATA(leaf);
4940 /* Read overflow data.
4942 data->mv_size = NODEDSZ(leaf);
4943 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4944 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4945 DPRINTF(("read overflow page %"Z"u failed", pgno));
4948 data->mv_data = METADATA(omp);
4954 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4955 MDB_val *key, MDB_val *data)
4964 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4966 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4969 if (txn->mt_flags & MDB_TXN_ERROR)
4972 mdb_cursor_init(&mc, txn, dbi, &mx);
4973 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4976 /** Find a sibling for a page.
4977 * Replaces the page at the top of the cursor's stack with the
4978 * specified sibling, if one exists.
4979 * @param[in] mc The cursor for this operation.
4980 * @param[in] move_right Non-zero if the right sibling is requested,
4981 * otherwise the left sibling.
4982 * @return 0 on success, non-zero on failure.
4985 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4991 if (mc->mc_snum < 2) {
4992 return MDB_NOTFOUND; /* root has no siblings */
4996 DPRINTF(("parent page is page %"Z"u, index %u",
4997 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4999 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5000 : (mc->mc_ki[mc->mc_top] == 0)) {
5001 DPRINTF(("no more keys left, moving to %s sibling",
5002 move_right ? "right" : "left"));
5003 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5004 /* undo cursor_pop before returning */
5011 mc->mc_ki[mc->mc_top]++;
5013 mc->mc_ki[mc->mc_top]--;
5014 DPRINTF(("just moving to %s index key %u",
5015 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5017 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
5019 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5020 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5021 /* mc will be inconsistent if caller does mc_snum++ as above */
5022 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5026 mdb_cursor_push(mc, mp);
5028 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5033 /** Move the cursor to the next data item. */
5035 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5041 if (mc->mc_flags & C_EOF) {
5042 return MDB_NOTFOUND;
5045 assert(mc->mc_flags & C_INITIALIZED);
5047 mp = mc->mc_pg[mc->mc_top];
5049 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5050 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5051 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5052 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5053 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5054 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5055 if (rc == MDB_SUCCESS)
5056 MDB_GET_KEY(leaf, key);
5061 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5062 if (op == MDB_NEXT_DUP)
5063 return MDB_NOTFOUND;
5067 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5068 if (mc->mc_flags & C_DEL)
5071 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5072 DPUTS("=====> move to next sibling page");
5073 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5074 mc->mc_flags |= C_EOF;
5077 mp = mc->mc_pg[mc->mc_top];
5078 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5080 mc->mc_ki[mc->mc_top]++;
5083 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5084 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5087 key->mv_size = mc->mc_db->md_pad;
5088 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5092 assert(IS_LEAF(mp));
5093 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5095 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5096 mdb_xcursor_init1(mc, leaf);
5099 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5102 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5103 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5104 if (rc != MDB_SUCCESS)
5109 MDB_GET_KEY(leaf, key);
5113 /** Move the cursor to the previous data item. */
5115 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5121 assert(mc->mc_flags & C_INITIALIZED);
5123 mp = mc->mc_pg[mc->mc_top];
5125 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5126 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5127 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5128 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5129 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5130 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5131 if (rc == MDB_SUCCESS)
5132 MDB_GET_KEY(leaf, key);
5136 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5137 if (op == MDB_PREV_DUP)
5138 return MDB_NOTFOUND;
5143 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5145 if (mc->mc_ki[mc->mc_top] == 0) {
5146 DPUTS("=====> move to prev sibling page");
5147 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5150 mp = mc->mc_pg[mc->mc_top];
5151 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5152 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5154 mc->mc_ki[mc->mc_top]--;
5156 mc->mc_flags &= ~C_EOF;
5158 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5159 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5162 key->mv_size = mc->mc_db->md_pad;
5163 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5167 assert(IS_LEAF(mp));
5168 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5170 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5171 mdb_xcursor_init1(mc, leaf);
5174 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5177 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5178 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5179 if (rc != MDB_SUCCESS)
5184 MDB_GET_KEY(leaf, key);
5188 /** Set the cursor on a specific data item. */
5190 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5191 MDB_cursor_op op, int *exactp)
5195 MDB_node *leaf = NULL;
5200 if (key->mv_size == 0)
5201 return MDB_BAD_VALSIZE;
5204 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5206 /* See if we're already on the right page */
5207 if (mc->mc_flags & C_INITIALIZED) {
5210 mp = mc->mc_pg[mc->mc_top];
5212 mc->mc_ki[mc->mc_top] = 0;
5213 return MDB_NOTFOUND;
5215 if (mp->mp_flags & P_LEAF2) {
5216 nodekey.mv_size = mc->mc_db->md_pad;
5217 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5219 leaf = NODEPTR(mp, 0);
5220 MDB_GET_KEY2(leaf, nodekey);
5222 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5224 /* Probably happens rarely, but first node on the page
5225 * was the one we wanted.
5227 mc->mc_ki[mc->mc_top] = 0;
5234 unsigned int nkeys = NUMKEYS(mp);
5236 if (mp->mp_flags & P_LEAF2) {
5237 nodekey.mv_data = LEAF2KEY(mp,
5238 nkeys-1, nodekey.mv_size);
5240 leaf = NODEPTR(mp, nkeys-1);
5241 MDB_GET_KEY2(leaf, nodekey);
5243 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5245 /* last node was the one we wanted */
5246 mc->mc_ki[mc->mc_top] = nkeys-1;
5252 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5253 /* This is definitely the right page, skip search_page */
5254 if (mp->mp_flags & P_LEAF2) {
5255 nodekey.mv_data = LEAF2KEY(mp,
5256 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5258 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5259 MDB_GET_KEY2(leaf, nodekey);
5261 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5263 /* current node was the one we wanted */
5273 /* If any parents have right-sibs, search.
5274 * Otherwise, there's nothing further.
5276 for (i=0; i<mc->mc_top; i++)
5278 NUMKEYS(mc->mc_pg[i])-1)
5280 if (i == mc->mc_top) {
5281 /* There are no other pages */
5282 mc->mc_ki[mc->mc_top] = nkeys;
5283 return MDB_NOTFOUND;
5287 /* There are no other pages */
5288 mc->mc_ki[mc->mc_top] = 0;
5289 if (op == MDB_SET_RANGE) {
5293 return MDB_NOTFOUND;
5297 rc = mdb_page_search(mc, key, 0);
5298 if (rc != MDB_SUCCESS)
5301 mp = mc->mc_pg[mc->mc_top];
5302 assert(IS_LEAF(mp));
5305 leaf = mdb_node_search(mc, key, exactp);
5306 if (exactp != NULL && !*exactp) {
5307 /* MDB_SET specified and not an exact match. */
5308 return MDB_NOTFOUND;
5312 DPUTS("===> inexact leaf not found, goto sibling");
5313 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5314 return rc; /* no entries matched */
5315 mp = mc->mc_pg[mc->mc_top];
5316 assert(IS_LEAF(mp));
5317 leaf = NODEPTR(mp, 0);
5321 mc->mc_flags |= C_INITIALIZED;
5322 mc->mc_flags &= ~C_EOF;
5325 key->mv_size = mc->mc_db->md_pad;
5326 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5330 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5331 mdb_xcursor_init1(mc, leaf);
5334 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5335 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5336 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5339 if (op == MDB_GET_BOTH) {
5345 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5346 if (rc != MDB_SUCCESS)
5349 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5351 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5353 rc = mc->mc_dbx->md_dcmp(data, &d2);
5355 if (op == MDB_GET_BOTH || rc > 0)
5356 return MDB_NOTFOUND;
5363 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5364 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5369 /* The key already matches in all other cases */
5370 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5371 MDB_GET_KEY(leaf, key);
5372 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5377 /** Move the cursor to the first item in the database. */
5379 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5385 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5387 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5388 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5389 if (rc != MDB_SUCCESS)
5392 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5394 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5395 mc->mc_flags |= C_INITIALIZED;
5396 mc->mc_flags &= ~C_EOF;
5398 mc->mc_ki[mc->mc_top] = 0;
5400 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5401 key->mv_size = mc->mc_db->md_pad;
5402 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5407 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5408 mdb_xcursor_init1(mc, leaf);
5409 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5413 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5417 MDB_GET_KEY(leaf, key);
5421 /** Move the cursor to the last item in the database. */
5423 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5429 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5431 if (!(mc->mc_flags & C_EOF)) {
5433 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5434 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5435 if (rc != MDB_SUCCESS)
5438 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5441 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5442 mc->mc_flags |= C_INITIALIZED|C_EOF;
5443 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5445 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5446 key->mv_size = mc->mc_db->md_pad;
5447 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5452 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5453 mdb_xcursor_init1(mc, leaf);
5454 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5458 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5463 MDB_GET_KEY(leaf, key);
5468 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5473 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5477 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5481 case MDB_GET_CURRENT:
5482 if (!(mc->mc_flags & C_INITIALIZED)) {
5485 MDB_page *mp = mc->mc_pg[mc->mc_top];
5486 int nkeys = NUMKEYS(mp);
5487 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5488 mc->mc_ki[mc->mc_top] = nkeys;
5494 key->mv_size = mc->mc_db->md_pad;
5495 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5497 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5498 MDB_GET_KEY(leaf, key);
5500 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5501 if (mc->mc_flags & C_DEL)
5502 mdb_xcursor_init1(mc, leaf);
5503 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5505 rc = mdb_node_read(mc->mc_txn, leaf, data);
5512 case MDB_GET_BOTH_RANGE:
5517 if (mc->mc_xcursor == NULL) {
5518 rc = MDB_INCOMPATIBLE;
5528 rc = mdb_cursor_set(mc, key, data, op,
5529 op == MDB_SET_RANGE ? NULL : &exact);
5532 case MDB_GET_MULTIPLE:
5533 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5537 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5538 rc = MDB_INCOMPATIBLE;
5542 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5543 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5546 case MDB_NEXT_MULTIPLE:
5551 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5552 rc = MDB_INCOMPATIBLE;
5555 if (!(mc->mc_flags & C_INITIALIZED))
5556 rc = mdb_cursor_first(mc, key, data);
5558 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5559 if (rc == MDB_SUCCESS) {
5560 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5563 mx = &mc->mc_xcursor->mx_cursor;
5564 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5566 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5567 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5575 case MDB_NEXT_NODUP:
5576 if (!(mc->mc_flags & C_INITIALIZED))
5577 rc = mdb_cursor_first(mc, key, data);
5579 rc = mdb_cursor_next(mc, key, data, op);
5583 case MDB_PREV_NODUP:
5584 if (!(mc->mc_flags & C_INITIALIZED)) {
5585 rc = mdb_cursor_last(mc, key, data);
5588 mc->mc_flags |= C_INITIALIZED;
5589 mc->mc_ki[mc->mc_top]++;
5591 rc = mdb_cursor_prev(mc, key, data, op);
5594 rc = mdb_cursor_first(mc, key, data);
5597 mfunc = mdb_cursor_first;
5599 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5603 if (mc->mc_xcursor == NULL) {
5604 rc = MDB_INCOMPATIBLE;
5607 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5611 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5614 rc = mdb_cursor_last(mc, key, data);
5617 mfunc = mdb_cursor_last;
5620 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5625 if (mc->mc_flags & C_DEL)
5626 mc->mc_flags ^= C_DEL;
5631 /** Touch all the pages in the cursor stack. Set mc_top.
5632 * Makes sure all the pages are writable, before attempting a write operation.
5633 * @param[in] mc The cursor to operate on.
5636 mdb_cursor_touch(MDB_cursor *mc)
5638 int rc = MDB_SUCCESS;
5640 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5643 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5644 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5647 *mc->mc_dbflag |= DB_DIRTY;
5652 rc = mdb_page_touch(mc);
5653 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5654 mc->mc_top = mc->mc_snum-1;
5659 /** Do not spill pages to disk if txn is getting full, may fail instead */
5660 #define MDB_NOSPILL 0x8000
5663 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5666 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5667 MDB_env *env = mc->mc_txn->mt_env;
5668 MDB_node *leaf = NULL;
5669 MDB_val xdata, *rdata, dkey;
5671 int do_sub = 0, insert = 0;
5672 unsigned int mcount = 0, dcount = 0, nospill;
5675 unsigned int nflags;
5678 /* Check this first so counter will always be zero on any
5681 if (flags & MDB_MULTIPLE) {
5682 dcount = data[1].mv_size;
5683 data[1].mv_size = 0;
5684 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5685 return MDB_INCOMPATIBLE;
5688 nospill = flags & MDB_NOSPILL;
5689 flags &= ~MDB_NOSPILL;
5691 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5692 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5694 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5695 return MDB_BAD_VALSIZE;
5697 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5698 return MDB_BAD_VALSIZE;
5700 #if SIZE_MAX > MAXDATASIZE
5701 if (data->mv_size > MAXDATASIZE)
5702 return MDB_BAD_VALSIZE;
5705 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5706 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5710 if (flags == MDB_CURRENT) {
5711 if (!(mc->mc_flags & C_INITIALIZED))
5714 } else if (mc->mc_db->md_root == P_INVALID) {
5715 /* new database, cursor has nothing to point to */
5718 mc->mc_flags &= ~C_INITIALIZED;
5723 if (flags & MDB_APPEND) {
5725 rc = mdb_cursor_last(mc, &k2, &d2);
5727 rc = mc->mc_dbx->md_cmp(key, &k2);
5730 mc->mc_ki[mc->mc_top]++;
5732 /* new key is <= last key */
5737 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5739 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5740 DPRINTF(("duplicate key [%s]", DKEY(key)));
5742 return MDB_KEYEXIST;
5744 if (rc && rc != MDB_NOTFOUND)
5748 if (mc->mc_flags & C_DEL)
5749 mc->mc_flags ^= C_DEL;
5751 /* Cursor is positioned, check for room in the dirty list */
5753 if (flags & MDB_MULTIPLE) {
5755 xdata.mv_size = data->mv_size * dcount;
5759 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5763 if (rc == MDB_NO_ROOT) {
5765 /* new database, write a root leaf page */
5766 DPUTS("allocating new root leaf page");
5767 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5770 mdb_cursor_push(mc, np);
5771 mc->mc_db->md_root = np->mp_pgno;
5772 mc->mc_db->md_depth++;
5773 *mc->mc_dbflag |= DB_DIRTY;
5774 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5776 np->mp_flags |= P_LEAF2;
5777 mc->mc_flags |= C_INITIALIZED;
5779 /* make sure all cursor pages are writable */
5780 rc2 = mdb_cursor_touch(mc);
5785 /* The key already exists */
5786 if (rc == MDB_SUCCESS) {
5790 /* there's only a key anyway, so this is a no-op */
5791 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5792 unsigned int ksize = mc->mc_db->md_pad;
5793 if (key->mv_size != ksize)
5794 return MDB_BAD_VALSIZE;
5795 if (flags == MDB_CURRENT) {
5796 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5797 memcpy(ptr, key->mv_data, ksize);
5803 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5804 olddata.mv_size = NODEDSZ(leaf);
5805 olddata.mv_data = NODEDATA(leaf);
5808 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5809 mp = fp = xdata.mv_data = env->me_pbuf;
5810 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5812 /* Was a single item before, must convert now */
5813 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5814 /* Just overwrite the current item */
5815 if (flags == MDB_CURRENT)
5818 #if UINT_MAX < SIZE_MAX
5819 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5820 #ifdef MISALIGNED_OK
5821 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5823 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5826 /* if data matches, skip it */
5827 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5828 if (flags & MDB_NODUPDATA)
5830 else if (flags & MDB_MULTIPLE)
5837 /* Back up original data item */
5838 dkey.mv_size = olddata.mv_size;
5839 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5841 /* Make sub-page header for the dup items, with dummy body */
5842 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5843 fp->mp_lower = PAGEHDRSZ;
5844 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5845 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5846 fp->mp_flags |= P_LEAF2;
5847 fp->mp_pad = data->mv_size;
5848 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
5850 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
5851 (dkey.mv_size & 1) + (data->mv_size & 1);
5853 fp->mp_upper = xdata.mv_size;
5854 } else if (leaf->mn_flags & F_SUBDATA) {
5855 /* Data is on sub-DB, just store it */
5856 flags |= F_DUPDATA|F_SUBDATA;
5859 /* See if we need to convert from fake page to subDB */
5860 unsigned int offset;
5864 fp = olddata.mv_data;
5867 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5868 offset = EVEN(NODESIZE + sizeof(indx_t) +
5872 offset = fp->mp_pad;
5873 if (SIZELEFT(fp) < offset) {
5874 offset *= 4; /* space for 4 more */
5877 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
5879 fp->mp_flags |= P_DIRTY;
5880 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
5881 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5885 fp_flags = fp->mp_flags;
5886 xdata.mv_size = olddata.mv_size + offset;
5887 if (NODESIZE+NODEKSZ(leaf)+xdata.mv_size > env->me_nodemax) {
5888 /* yes, convert it */
5889 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5890 dummy.md_pad = fp->mp_pad;
5891 dummy.md_flags = MDB_DUPFIXED;
5892 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5893 dummy.md_flags |= MDB_INTEGERKEY;
5899 dummy.md_branch_pages = 0;
5900 dummy.md_leaf_pages = 1;
5901 dummy.md_overflow_pages = 0;
5902 dummy.md_entries = NUMKEYS(fp);
5903 xdata.mv_size = sizeof(MDB_db);
5904 xdata.mv_data = &dummy;
5905 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5907 offset = env->me_psize - olddata.mv_size;
5908 flags |= F_DUPDATA|F_SUBDATA;
5909 dummy.md_root = mp->mp_pgno;
5910 fp_flags &= ~P_SUBP;
5912 mp->mp_flags = fp_flags | P_DIRTY;
5913 mp->mp_pad = fp->mp_pad;
5914 mp->mp_lower = fp->mp_lower;
5915 mp->mp_upper = fp->mp_upper + offset;
5917 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5919 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
5920 olddata.mv_size - fp->mp_upper);
5921 for (i=0; i<NUMKEYS(fp); i++)
5922 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5929 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5933 /* overflow page overwrites need special handling */
5934 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5937 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
5939 memcpy(&pg, olddata.mv_data, sizeof(pg));
5940 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5942 ovpages = omp->mp_pages;
5944 /* Is the ov page large enough? */
5945 if (ovpages >= dpages) {
5946 if (!(omp->mp_flags & P_DIRTY) &&
5947 (level || (env->me_flags & MDB_WRITEMAP)))
5949 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5952 level = 0; /* dirty in this txn or clean */
5955 if (omp->mp_flags & P_DIRTY) {
5956 /* yes, overwrite it. Note in this case we don't
5957 * bother to try shrinking the page if the new data
5958 * is smaller than the overflow threshold.
5961 /* It is writable only in a parent txn */
5962 size_t sz = (size_t) env->me_psize * ovpages, off;
5963 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5969 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5970 if (!(flags & MDB_RESERVE)) {
5971 /* Copy end of page, adjusting alignment so
5972 * compiler may copy words instead of bytes.
5974 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5975 memcpy((size_t *)((char *)np + off),
5976 (size_t *)((char *)omp + off), sz - off);
5979 memcpy(np, omp, sz); /* Copy beginning of page */
5982 SETDSZ(leaf, data->mv_size);
5983 if (F_ISSET(flags, MDB_RESERVE))
5984 data->mv_data = METADATA(omp);
5986 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5990 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5992 } else if (data->mv_size == olddata.mv_size) {
5993 /* same size, just replace it. Note that we could
5994 * also reuse this node if the new data is smaller,
5995 * but instead we opt to shrink the node in that case.
5997 if (F_ISSET(flags, MDB_RESERVE))
5998 data->mv_data = olddata.mv_data;
5999 else if (data->mv_size)
6000 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6002 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6005 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6006 mc->mc_db->md_entries--;
6008 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6015 nflags = flags & NODE_ADD_FLAGS;
6016 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6017 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6018 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6019 nflags &= ~MDB_APPEND;
6021 nflags |= MDB_SPLIT_REPLACE;
6022 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6024 /* There is room already in this leaf page. */
6025 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6026 if (rc == 0 && !do_sub && insert) {
6027 /* Adjust other cursors pointing to mp */
6028 MDB_cursor *m2, *m3;
6029 MDB_dbi dbi = mc->mc_dbi;
6030 unsigned i = mc->mc_top;
6031 MDB_page *mp = mc->mc_pg[i];
6033 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6034 if (mc->mc_flags & C_SUB)
6035 m3 = &m2->mc_xcursor->mx_cursor;
6038 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6039 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6046 if (rc != MDB_SUCCESS)
6047 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6049 /* Now store the actual data in the child DB. Note that we're
6050 * storing the user data in the keys field, so there are strict
6051 * size limits on dupdata. The actual data fields of the child
6052 * DB are all zero size.
6059 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6060 if (flags & MDB_CURRENT) {
6061 xflags = MDB_CURRENT|MDB_NOSPILL;
6063 mdb_xcursor_init1(mc, leaf);
6064 xflags = (flags & MDB_NODUPDATA) ?
6065 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6067 /* converted, write the original data first */
6069 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6073 /* Adjust other cursors pointing to mp */
6075 unsigned i = mc->mc_top;
6076 MDB_page *mp = mc->mc_pg[i];
6078 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6079 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6080 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6081 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6082 mdb_xcursor_init1(m2, leaf);
6086 /* we've done our job */
6089 if (flags & MDB_APPENDDUP)
6090 xflags |= MDB_APPEND;
6091 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6092 if (flags & F_SUBDATA) {
6093 void *db = NODEDATA(leaf);
6094 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6097 /* sub-writes might have failed so check rc again.
6098 * Don't increment count if we just replaced an existing item.
6100 if (!rc && !(flags & MDB_CURRENT))
6101 mc->mc_db->md_entries++;
6102 if (flags & MDB_MULTIPLE) {
6106 /* let caller know how many succeeded, if any */
6107 data[1].mv_size = mcount;
6108 if (mcount < dcount) {
6109 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6116 /* If we succeeded and the key didn't exist before, make sure
6117 * the cursor is marked valid.
6120 mc->mc_flags |= C_INITIALIZED;
6125 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6131 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6132 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6134 if (!(mc->mc_flags & C_INITIALIZED))
6137 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6138 return MDB_NOTFOUND;
6140 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6143 rc = mdb_cursor_touch(mc);
6147 mp = mc->mc_pg[mc->mc_top];
6148 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6150 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6151 if (!(flags & MDB_NODUPDATA)) {
6152 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6153 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6155 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6156 /* If sub-DB still has entries, we're done */
6157 if (mc->mc_xcursor->mx_db.md_entries) {
6158 if (leaf->mn_flags & F_SUBDATA) {
6159 /* update subDB info */
6160 void *db = NODEDATA(leaf);
6161 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6164 /* shrink fake page */
6165 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6166 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6167 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6168 /* fix other sub-DB cursors pointed at this fake page */
6169 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6170 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6171 if (m2->mc_pg[mc->mc_top] == mp &&
6172 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6173 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6176 mc->mc_db->md_entries--;
6177 mc->mc_flags |= C_DEL;
6180 /* otherwise fall thru and delete the sub-DB */
6183 if (leaf->mn_flags & F_SUBDATA) {
6184 /* add all the child DB's pages to the free list */
6185 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6186 if (rc == MDB_SUCCESS) {
6187 mc->mc_db->md_entries -=
6188 mc->mc_xcursor->mx_db.md_entries;
6193 return mdb_cursor_del0(mc, leaf);
6196 /** Allocate and initialize new pages for a database.
6197 * @param[in] mc a cursor on the database being added to.
6198 * @param[in] flags flags defining what type of page is being allocated.
6199 * @param[in] num the number of pages to allocate. This is usually 1,
6200 * unless allocating overflow pages for a large record.
6201 * @param[out] mp Address of a page, or NULL on failure.
6202 * @return 0 on success, non-zero on failure.
6205 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6210 if ((rc = mdb_page_alloc(mc, num, &np)))
6212 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6213 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6214 np->mp_flags = flags | P_DIRTY;
6215 np->mp_lower = PAGEHDRSZ;
6216 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6219 mc->mc_db->md_branch_pages++;
6220 else if (IS_LEAF(np))
6221 mc->mc_db->md_leaf_pages++;
6222 else if (IS_OVERFLOW(np)) {
6223 mc->mc_db->md_overflow_pages += num;
6231 /** Calculate the size of a leaf node.
6232 * The size depends on the environment's page size; if a data item
6233 * is too large it will be put onto an overflow page and the node
6234 * size will only include the key and not the data. Sizes are always
6235 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6236 * of the #MDB_node headers.
6237 * @param[in] env The environment handle.
6238 * @param[in] key The key for the node.
6239 * @param[in] data The data for the node.
6240 * @return The number of bytes needed to store the node.
6243 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6247 sz = LEAFSIZE(key, data);
6248 if (sz > env->me_nodemax) {
6249 /* put on overflow page */
6250 sz -= data->mv_size - sizeof(pgno_t);
6253 return EVEN(sz + sizeof(indx_t));
6256 /** Calculate the size of a branch node.
6257 * The size should depend on the environment's page size but since
6258 * we currently don't support spilling large keys onto overflow
6259 * pages, it's simply the size of the #MDB_node header plus the
6260 * size of the key. Sizes are always rounded up to an even number
6261 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6262 * @param[in] env The environment handle.
6263 * @param[in] key The key for the node.
6264 * @return The number of bytes needed to store the node.
6267 mdb_branch_size(MDB_env *env, MDB_val *key)
6272 if (sz > env->me_nodemax) {
6273 /* put on overflow page */
6274 /* not implemented */
6275 /* sz -= key->size - sizeof(pgno_t); */
6278 return sz + sizeof(indx_t);
6281 /** Add a node to the page pointed to by the cursor.
6282 * @param[in] mc The cursor for this operation.
6283 * @param[in] indx The index on the page where the new node should be added.
6284 * @param[in] key The key for the new node.
6285 * @param[in] data The data for the new node, if any.
6286 * @param[in] pgno The page number, if adding a branch node.
6287 * @param[in] flags Flags for the node.
6288 * @return 0 on success, non-zero on failure. Possible errors are:
6290 * <li>ENOMEM - failed to allocate overflow pages for the node.
6291 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6292 * should never happen since all callers already calculate the
6293 * page's free space before calling this function.
6297 mdb_node_add(MDB_cursor *mc, indx_t indx,
6298 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6301 size_t node_size = NODESIZE;
6305 MDB_page *mp = mc->mc_pg[mc->mc_top];
6306 MDB_page *ofp = NULL; /* overflow page */
6309 assert(mp->mp_upper >= mp->mp_lower);
6311 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6312 IS_LEAF(mp) ? "leaf" : "branch",
6313 IS_SUBP(mp) ? "sub-" : "",
6314 mp->mp_pgno, indx, data ? data->mv_size : 0,
6315 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6318 /* Move higher keys up one slot. */
6319 int ksize = mc->mc_db->md_pad, dif;
6320 char *ptr = LEAF2KEY(mp, indx, ksize);
6321 dif = NUMKEYS(mp) - indx;
6323 memmove(ptr+ksize, ptr, dif*ksize);
6324 /* insert new key */
6325 memcpy(ptr, key->mv_data, ksize);
6327 /* Just using these for counting */
6328 mp->mp_lower += sizeof(indx_t);
6329 mp->mp_upper -= ksize - sizeof(indx_t);
6333 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6335 node_size += key->mv_size;
6338 if (F_ISSET(flags, F_BIGDATA)) {
6339 /* Data already on overflow page. */
6340 node_size += sizeof(pgno_t);
6341 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6342 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6344 /* Put data on overflow page. */
6345 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6346 data->mv_size, node_size+data->mv_size));
6347 node_size = EVEN(node_size + sizeof(pgno_t));
6348 if ((ssize_t)node_size > room)
6350 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6352 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6356 node_size += data->mv_size;
6359 node_size = EVEN(node_size);
6360 if ((ssize_t)node_size > room)
6364 /* Move higher pointers up one slot. */
6365 for (i = NUMKEYS(mp); i > indx; i--)
6366 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6368 /* Adjust free space offsets. */
6369 ofs = mp->mp_upper - node_size;
6370 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6371 mp->mp_ptrs[indx] = ofs;
6373 mp->mp_lower += sizeof(indx_t);
6375 /* Write the node data. */
6376 node = NODEPTR(mp, indx);
6377 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6378 node->mn_flags = flags;
6380 SETDSZ(node,data->mv_size);
6385 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6390 if (F_ISSET(flags, F_BIGDATA))
6391 memcpy(node->mn_data + key->mv_size, data->mv_data,
6393 else if (F_ISSET(flags, MDB_RESERVE))
6394 data->mv_data = node->mn_data + key->mv_size;
6396 memcpy(node->mn_data + key->mv_size, data->mv_data,
6399 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6401 if (F_ISSET(flags, MDB_RESERVE))
6402 data->mv_data = METADATA(ofp);
6404 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6411 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6412 mp->mp_pgno, NUMKEYS(mp)));
6413 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6414 DPRINTF(("node size = %"Z"u", node_size));
6415 return MDB_PAGE_FULL;
6418 /** Delete the specified node from a page.
6419 * @param[in] mp The page to operate on.
6420 * @param[in] indx The index of the node to delete.
6421 * @param[in] ksize The size of a node. Only used if the page is
6422 * part of a #MDB_DUPFIXED database.
6425 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6428 indx_t i, j, numkeys, ptr;
6435 COPY_PGNO(pgno, mp->mp_pgno);
6436 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6437 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6440 assert(indx < NUMKEYS(mp));
6443 int x = NUMKEYS(mp) - 1 - indx;
6444 base = LEAF2KEY(mp, indx, ksize);
6446 memmove(base, base + ksize, x * ksize);
6447 mp->mp_lower -= sizeof(indx_t);
6448 mp->mp_upper += ksize - sizeof(indx_t);
6452 node = NODEPTR(mp, indx);
6453 sz = NODESIZE + node->mn_ksize;
6455 if (F_ISSET(node->mn_flags, F_BIGDATA))
6456 sz += sizeof(pgno_t);
6458 sz += NODEDSZ(node);
6462 ptr = mp->mp_ptrs[indx];
6463 numkeys = NUMKEYS(mp);
6464 for (i = j = 0; i < numkeys; i++) {
6466 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6467 if (mp->mp_ptrs[i] < ptr)
6468 mp->mp_ptrs[j] += sz;
6473 base = (char *)mp + mp->mp_upper;
6474 memmove(base + sz, base, ptr - mp->mp_upper);
6476 mp->mp_lower -= sizeof(indx_t);
6480 /** Compact the main page after deleting a node on a subpage.
6481 * @param[in] mp The main page to operate on.
6482 * @param[in] indx The index of the subpage on the main page.
6485 mdb_node_shrink(MDB_page *mp, indx_t indx)
6491 indx_t i, numkeys, ptr;
6493 node = NODEPTR(mp, indx);
6494 sp = (MDB_page *)NODEDATA(node);
6495 delta = SIZELEFT(sp);
6496 xp = (MDB_page *)((char *)sp + delta);
6498 /* shift subpage upward */
6500 nsize = NUMKEYS(sp) * sp->mp_pad;
6502 return; /* do not make the node uneven-sized */
6503 memmove(METADATA(xp), METADATA(sp), nsize);
6506 numkeys = NUMKEYS(sp);
6507 for (i=numkeys-1; i>=0; i--)
6508 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6510 xp->mp_upper = sp->mp_lower;
6511 xp->mp_lower = sp->mp_lower;
6512 xp->mp_flags = sp->mp_flags;
6513 xp->mp_pad = sp->mp_pad;
6514 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6516 nsize = NODEDSZ(node) - delta;
6517 SETDSZ(node, nsize);
6519 /* shift lower nodes upward */
6520 ptr = mp->mp_ptrs[indx];
6521 numkeys = NUMKEYS(mp);
6522 for (i = 0; i < numkeys; i++) {
6523 if (mp->mp_ptrs[i] <= ptr)
6524 mp->mp_ptrs[i] += delta;
6527 base = (char *)mp + mp->mp_upper;
6528 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6529 mp->mp_upper += delta;
6532 /** Initial setup of a sorted-dups cursor.
6533 * Sorted duplicates are implemented as a sub-database for the given key.
6534 * The duplicate data items are actually keys of the sub-database.
6535 * Operations on the duplicate data items are performed using a sub-cursor
6536 * initialized when the sub-database is first accessed. This function does
6537 * the preliminary setup of the sub-cursor, filling in the fields that
6538 * depend only on the parent DB.
6539 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6542 mdb_xcursor_init0(MDB_cursor *mc)
6544 MDB_xcursor *mx = mc->mc_xcursor;
6546 mx->mx_cursor.mc_xcursor = NULL;
6547 mx->mx_cursor.mc_txn = mc->mc_txn;
6548 mx->mx_cursor.mc_db = &mx->mx_db;
6549 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6550 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6551 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6552 mx->mx_cursor.mc_snum = 0;
6553 mx->mx_cursor.mc_top = 0;
6554 mx->mx_cursor.mc_flags = C_SUB;
6555 mx->mx_dbx.md_name.mv_size = 0;
6556 mx->mx_dbx.md_name.mv_data = NULL;
6557 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6558 mx->mx_dbx.md_dcmp = NULL;
6559 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6562 /** Final setup of a sorted-dups cursor.
6563 * Sets up the fields that depend on the data from the main cursor.
6564 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6565 * @param[in] node The data containing the #MDB_db record for the
6566 * sorted-dup database.
6569 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6571 MDB_xcursor *mx = mc->mc_xcursor;
6573 if (node->mn_flags & F_SUBDATA) {
6574 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6575 mx->mx_cursor.mc_pg[0] = 0;
6576 mx->mx_cursor.mc_snum = 0;
6577 mx->mx_cursor.mc_top = 0;
6578 mx->mx_cursor.mc_flags = C_SUB;
6580 MDB_page *fp = NODEDATA(node);
6581 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6582 mx->mx_db.md_flags = 0;
6583 mx->mx_db.md_depth = 1;
6584 mx->mx_db.md_branch_pages = 0;
6585 mx->mx_db.md_leaf_pages = 1;
6586 mx->mx_db.md_overflow_pages = 0;
6587 mx->mx_db.md_entries = NUMKEYS(fp);
6588 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6589 mx->mx_cursor.mc_snum = 1;
6590 mx->mx_cursor.mc_top = 0;
6591 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6592 mx->mx_cursor.mc_pg[0] = fp;
6593 mx->mx_cursor.mc_ki[0] = 0;
6594 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6595 mx->mx_db.md_flags = MDB_DUPFIXED;
6596 mx->mx_db.md_pad = fp->mp_pad;
6597 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6598 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6601 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6602 mx->mx_db.md_root));
6603 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6604 #if UINT_MAX < SIZE_MAX
6605 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6606 #ifdef MISALIGNED_OK
6607 mx->mx_dbx.md_cmp = mdb_cmp_long;
6609 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6614 /** Initialize a cursor for a given transaction and database. */
6616 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6619 mc->mc_backup = NULL;
6622 mc->mc_db = &txn->mt_dbs[dbi];
6623 mc->mc_dbx = &txn->mt_dbxs[dbi];
6624 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6629 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6631 mc->mc_xcursor = mx;
6632 mdb_xcursor_init0(mc);
6634 mc->mc_xcursor = NULL;
6636 if (*mc->mc_dbflag & DB_STALE) {
6637 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6642 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6645 size_t size = sizeof(MDB_cursor);
6647 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6650 if (txn->mt_flags & MDB_TXN_ERROR)
6653 /* Allow read access to the freelist */
6654 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6657 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6658 size += sizeof(MDB_xcursor);
6660 if ((mc = malloc(size)) != NULL) {
6661 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6662 if (txn->mt_cursors) {
6663 mc->mc_next = txn->mt_cursors[dbi];
6664 txn->mt_cursors[dbi] = mc;
6665 mc->mc_flags |= C_UNTRACK;
6677 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6679 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6682 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6685 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6689 /* Return the count of duplicate data items for the current key */
6691 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6695 if (mc == NULL || countp == NULL)
6698 if (mc->mc_xcursor == NULL)
6699 return MDB_INCOMPATIBLE;
6701 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6702 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6705 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6708 *countp = mc->mc_xcursor->mx_db.md_entries;
6714 mdb_cursor_close(MDB_cursor *mc)
6716 if (mc && !mc->mc_backup) {
6717 /* remove from txn, if tracked */
6718 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6719 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6720 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6722 *prev = mc->mc_next;
6729 mdb_cursor_txn(MDB_cursor *mc)
6731 if (!mc) return NULL;
6736 mdb_cursor_dbi(MDB_cursor *mc)
6742 /** Replace the key for a node with a new key.
6743 * @param[in] mc Cursor pointing to the node to operate on.
6744 * @param[in] key The new key to use.
6745 * @return 0 on success, non-zero on failure.
6748 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6754 int delta, ksize, oksize;
6755 indx_t ptr, i, numkeys, indx;
6758 indx = mc->mc_ki[mc->mc_top];
6759 mp = mc->mc_pg[mc->mc_top];
6760 node = NODEPTR(mp, indx);
6761 ptr = mp->mp_ptrs[indx];
6765 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6766 k2.mv_data = NODEKEY(node);
6767 k2.mv_size = node->mn_ksize;
6768 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6770 mdb_dkey(&k2, kbuf2),
6776 /* Sizes must be 2-byte aligned. */
6777 ksize = EVEN(key->mv_size);
6778 oksize = EVEN(node->mn_ksize);
6779 delta = ksize - oksize;
6781 /* Shift node contents if EVEN(key length) changed. */
6783 if (delta > 0 && SIZELEFT(mp) < delta) {
6785 /* not enough space left, do a delete and split */
6786 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6787 pgno = NODEPGNO(node);
6788 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6789 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6792 numkeys = NUMKEYS(mp);
6793 for (i = 0; i < numkeys; i++) {
6794 if (mp->mp_ptrs[i] <= ptr)
6795 mp->mp_ptrs[i] -= delta;
6798 base = (char *)mp + mp->mp_upper;
6799 len = ptr - mp->mp_upper + NODESIZE;
6800 memmove(base - delta, base, len);
6801 mp->mp_upper -= delta;
6803 node = NODEPTR(mp, indx);
6806 /* But even if no shift was needed, update ksize */
6807 if (node->mn_ksize != key->mv_size)
6808 node->mn_ksize = key->mv_size;
6811 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6817 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6819 /** Move a node from csrc to cdst.
6822 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6829 unsigned short flags;
6833 /* Mark src and dst as dirty. */
6834 if ((rc = mdb_page_touch(csrc)) ||
6835 (rc = mdb_page_touch(cdst)))
6838 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6839 key.mv_size = csrc->mc_db->md_pad;
6840 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6842 data.mv_data = NULL;
6846 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6847 assert(!((size_t)srcnode&1));
6848 srcpg = NODEPGNO(srcnode);
6849 flags = srcnode->mn_flags;
6850 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6851 unsigned int snum = csrc->mc_snum;
6853 /* must find the lowest key below src */
6854 mdb_page_search_lowest(csrc);
6855 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6856 key.mv_size = csrc->mc_db->md_pad;
6857 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6859 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6860 key.mv_size = NODEKSZ(s2);
6861 key.mv_data = NODEKEY(s2);
6863 csrc->mc_snum = snum--;
6864 csrc->mc_top = snum;
6866 key.mv_size = NODEKSZ(srcnode);
6867 key.mv_data = NODEKEY(srcnode);
6869 data.mv_size = NODEDSZ(srcnode);
6870 data.mv_data = NODEDATA(srcnode);
6872 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6873 unsigned int snum = cdst->mc_snum;
6876 /* must find the lowest key below dst */
6877 mdb_page_search_lowest(cdst);
6878 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6879 bkey.mv_size = cdst->mc_db->md_pad;
6880 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6882 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6883 bkey.mv_size = NODEKSZ(s2);
6884 bkey.mv_data = NODEKEY(s2);
6886 cdst->mc_snum = snum--;
6887 cdst->mc_top = snum;
6888 mdb_cursor_copy(cdst, &mn);
6890 rc = mdb_update_key(&mn, &bkey);
6895 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6896 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6897 csrc->mc_ki[csrc->mc_top],
6899 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6900 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6902 /* Add the node to the destination page.
6904 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6905 if (rc != MDB_SUCCESS)
6908 /* Delete the node from the source page.
6910 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6913 /* Adjust other cursors pointing to mp */
6914 MDB_cursor *m2, *m3;
6915 MDB_dbi dbi = csrc->mc_dbi;
6916 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6918 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6919 if (csrc->mc_flags & C_SUB)
6920 m3 = &m2->mc_xcursor->mx_cursor;
6923 if (m3 == csrc) continue;
6924 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6925 csrc->mc_ki[csrc->mc_top]) {
6926 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6927 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6932 /* Update the parent separators.
6934 if (csrc->mc_ki[csrc->mc_top] == 0) {
6935 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6936 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6937 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6939 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6940 key.mv_size = NODEKSZ(srcnode);
6941 key.mv_data = NODEKEY(srcnode);
6943 DPRINTF(("update separator for source page %"Z"u to [%s]",
6944 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6945 mdb_cursor_copy(csrc, &mn);
6948 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6951 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6953 indx_t ix = csrc->mc_ki[csrc->mc_top];
6954 nullkey.mv_size = 0;
6955 csrc->mc_ki[csrc->mc_top] = 0;
6956 rc = mdb_update_key(csrc, &nullkey);
6957 csrc->mc_ki[csrc->mc_top] = ix;
6958 assert(rc == MDB_SUCCESS);
6962 if (cdst->mc_ki[cdst->mc_top] == 0) {
6963 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6964 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6965 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6967 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6968 key.mv_size = NODEKSZ(srcnode);
6969 key.mv_data = NODEKEY(srcnode);
6971 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6972 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6973 mdb_cursor_copy(cdst, &mn);
6976 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6979 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6981 indx_t ix = cdst->mc_ki[cdst->mc_top];
6982 nullkey.mv_size = 0;
6983 cdst->mc_ki[cdst->mc_top] = 0;
6984 rc = mdb_update_key(cdst, &nullkey);
6985 cdst->mc_ki[cdst->mc_top] = ix;
6986 assert(rc == MDB_SUCCESS);
6993 /** Merge one page into another.
6994 * The nodes from the page pointed to by \b csrc will
6995 * be copied to the page pointed to by \b cdst and then
6996 * the \b csrc page will be freed.
6997 * @param[in] csrc Cursor pointing to the source page.
6998 * @param[in] cdst Cursor pointing to the destination page.
7001 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7009 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7010 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7012 assert(csrc->mc_snum > 1); /* can't merge root page */
7013 assert(cdst->mc_snum > 1);
7015 /* Mark dst as dirty. */
7016 if ((rc = mdb_page_touch(cdst)))
7019 /* Move all nodes from src to dst.
7021 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7022 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7023 key.mv_size = csrc->mc_db->md_pad;
7024 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7025 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7026 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7027 if (rc != MDB_SUCCESS)
7029 key.mv_data = (char *)key.mv_data + key.mv_size;
7032 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7033 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7034 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7035 unsigned int snum = csrc->mc_snum;
7037 /* must find the lowest key below src */
7038 mdb_page_search_lowest(csrc);
7039 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7040 key.mv_size = csrc->mc_db->md_pad;
7041 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7043 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7044 key.mv_size = NODEKSZ(s2);
7045 key.mv_data = NODEKEY(s2);
7047 csrc->mc_snum = snum--;
7048 csrc->mc_top = snum;
7050 key.mv_size = srcnode->mn_ksize;
7051 key.mv_data = NODEKEY(srcnode);
7054 data.mv_size = NODEDSZ(srcnode);
7055 data.mv_data = NODEDATA(srcnode);
7056 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7057 if (rc != MDB_SUCCESS)
7062 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7063 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7064 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7066 /* Unlink the src page from parent and add to free list.
7068 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
7069 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
7072 rc = mdb_update_key(csrc, &key);
7078 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7079 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7082 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7083 csrc->mc_db->md_leaf_pages--;
7085 csrc->mc_db->md_branch_pages--;
7087 /* Adjust other cursors pointing to mp */
7088 MDB_cursor *m2, *m3;
7089 MDB_dbi dbi = csrc->mc_dbi;
7090 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7092 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7093 if (csrc->mc_flags & C_SUB)
7094 m3 = &m2->mc_xcursor->mx_cursor;
7097 if (m3 == csrc) continue;
7098 if (m3->mc_snum < csrc->mc_snum) continue;
7099 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7100 m3->mc_pg[csrc->mc_top] = mp;
7101 m3->mc_ki[csrc->mc_top] += nkeys;
7105 mdb_cursor_pop(csrc);
7107 return mdb_rebalance(csrc);
7110 /** Copy the contents of a cursor.
7111 * @param[in] csrc The cursor to copy from.
7112 * @param[out] cdst The cursor to copy to.
7115 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7119 cdst->mc_txn = csrc->mc_txn;
7120 cdst->mc_dbi = csrc->mc_dbi;
7121 cdst->mc_db = csrc->mc_db;
7122 cdst->mc_dbx = csrc->mc_dbx;
7123 cdst->mc_snum = csrc->mc_snum;
7124 cdst->mc_top = csrc->mc_top;
7125 cdst->mc_flags = csrc->mc_flags;
7127 for (i=0; i<csrc->mc_snum; i++) {
7128 cdst->mc_pg[i] = csrc->mc_pg[i];
7129 cdst->mc_ki[i] = csrc->mc_ki[i];
7133 /** Rebalance the tree after a delete operation.
7134 * @param[in] mc Cursor pointing to the page where rebalancing
7136 * @return 0 on success, non-zero on failure.
7139 mdb_rebalance(MDB_cursor *mc)
7143 unsigned int ptop, minkeys;
7146 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7150 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7151 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7152 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7153 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7154 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7158 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7159 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7162 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7163 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7169 if (mc->mc_snum < 2) {
7170 MDB_page *mp = mc->mc_pg[0];
7172 DPUTS("Can't rebalance a subpage, ignoring");
7175 if (NUMKEYS(mp) == 0) {
7176 DPUTS("tree is completely empty");
7177 mc->mc_db->md_root = P_INVALID;
7178 mc->mc_db->md_depth = 0;
7179 mc->mc_db->md_leaf_pages = 0;
7180 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7183 /* Adjust cursors pointing to mp */
7186 mc->mc_flags &= ~C_INITIALIZED;
7188 MDB_cursor *m2, *m3;
7189 MDB_dbi dbi = mc->mc_dbi;
7191 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7192 if (mc->mc_flags & C_SUB)
7193 m3 = &m2->mc_xcursor->mx_cursor;
7196 if (m3->mc_snum < mc->mc_snum) continue;
7197 if (m3->mc_pg[0] == mp) {
7200 m3->mc_flags &= ~C_INITIALIZED;
7204 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7205 DPUTS("collapsing root page!");
7206 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7209 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7210 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7213 mc->mc_db->md_depth--;
7214 mc->mc_db->md_branch_pages--;
7215 mc->mc_ki[0] = mc->mc_ki[1];
7217 /* Adjust other cursors pointing to mp */
7218 MDB_cursor *m2, *m3;
7219 MDB_dbi dbi = mc->mc_dbi;
7221 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7222 if (mc->mc_flags & C_SUB)
7223 m3 = &m2->mc_xcursor->mx_cursor;
7226 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7227 if (m3->mc_pg[0] == mp) {
7231 for (i=0; i<m3->mc_snum; i++) {
7232 m3->mc_pg[i] = m3->mc_pg[i+1];
7233 m3->mc_ki[i] = m3->mc_ki[i+1];
7239 DPUTS("root page doesn't need rebalancing");
7243 /* The parent (branch page) must have at least 2 pointers,
7244 * otherwise the tree is invalid.
7246 ptop = mc->mc_top-1;
7247 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7249 /* Leaf page fill factor is below the threshold.
7250 * Try to move keys from left or right neighbor, or
7251 * merge with a neighbor page.
7256 mdb_cursor_copy(mc, &mn);
7257 mn.mc_xcursor = NULL;
7259 if (mc->mc_ki[ptop] == 0) {
7260 /* We're the leftmost leaf in our parent.
7262 DPUTS("reading right neighbor");
7264 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7265 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7268 mn.mc_ki[mn.mc_top] = 0;
7269 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7271 /* There is at least one neighbor to the left.
7273 DPUTS("reading left neighbor");
7275 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7276 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7279 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7280 mc->mc_ki[mc->mc_top] = 0;
7283 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7284 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7285 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7287 /* If the neighbor page is above threshold and has enough keys,
7288 * move one key from it. Otherwise we should try to merge them.
7289 * (A branch page must never have less than 2 keys.)
7291 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7292 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7293 return mdb_node_move(&mn, mc);
7295 if (mc->mc_ki[ptop] == 0)
7296 rc = mdb_page_merge(&mn, mc);
7298 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7299 rc = mdb_page_merge(mc, &mn);
7300 mdb_cursor_copy(&mn, mc);
7302 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7307 /** Complete a delete operation started by #mdb_cursor_del(). */
7309 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7316 mp = mc->mc_pg[mc->mc_top];
7317 ki = mc->mc_ki[mc->mc_top];
7319 /* add overflow pages to free list */
7320 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7324 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7325 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7326 (rc = mdb_ovpage_free(mc, omp)))
7329 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7330 mc->mc_db->md_entries--;
7331 rc = mdb_rebalance(mc);
7332 if (rc != MDB_SUCCESS)
7333 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7336 MDB_dbi dbi = mc->mc_dbi;
7338 mp = mc->mc_pg[mc->mc_top];
7339 nkeys = NUMKEYS(mp);
7341 /* if mc points past last node in page, find next sibling */
7342 if (mc->mc_ki[mc->mc_top] >= nkeys)
7343 mdb_cursor_sibling(mc, 1);
7345 /* Adjust other cursors pointing to mp */
7346 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7347 if (m2 == mc || m2->mc_snum < mc->mc_snum)
7349 if (!(m2->mc_flags & C_INITIALIZED))
7351 if (m2->mc_pg[mc->mc_top] == mp) {
7352 if (m2->mc_ki[mc->mc_top] >= ki) {
7353 m2->mc_flags |= C_DEL;
7354 if (m2->mc_ki[mc->mc_top] > ki)
7355 m2->mc_ki[mc->mc_top]--;
7357 if (m2->mc_ki[mc->mc_top] >= nkeys)
7358 mdb_cursor_sibling(m2, 1);
7361 mc->mc_flags |= C_DEL;
7368 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7369 MDB_val *key, MDB_val *data)
7374 MDB_val rdata, *xdata;
7378 assert(key != NULL);
7380 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7382 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7385 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7386 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7388 mdb_cursor_init(&mc, txn, dbi, &mx);
7391 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7392 /* must ignore any data */
7403 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7405 /* let mdb_page_split know about this cursor if needed:
7406 * delete will trigger a rebalance; if it needs to move
7407 * a node from one page to another, it will have to
7408 * update the parent's separator key(s). If the new sepkey
7409 * is larger than the current one, the parent page may
7410 * run out of space, triggering a split. We need this
7411 * cursor to be consistent until the end of the rebalance.
7413 mc.mc_flags |= C_UNTRACK;
7414 mc.mc_next = txn->mt_cursors[dbi];
7415 txn->mt_cursors[dbi] = &mc;
7416 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7417 txn->mt_cursors[dbi] = mc.mc_next;
7422 /** Split a page and insert a new node.
7423 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7424 * The cursor will be updated to point to the actual page and index where
7425 * the node got inserted after the split.
7426 * @param[in] newkey The key for the newly inserted node.
7427 * @param[in] newdata The data for the newly inserted node.
7428 * @param[in] newpgno The page number, if the new node is a branch node.
7429 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7430 * @return 0 on success, non-zero on failure.
7433 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7434 unsigned int nflags)
7437 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7440 int i, j, split_indx, nkeys, pmax;
7441 MDB_env *env = mc->mc_txn->mt_env;
7443 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7444 MDB_page *copy = NULL;
7445 MDB_page *mp, *rp, *pp;
7450 mp = mc->mc_pg[mc->mc_top];
7451 newindx = mc->mc_ki[mc->mc_top];
7452 nkeys = NUMKEYS(mp);
7454 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7455 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7456 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7458 /* Create a right sibling. */
7459 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7461 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7463 if (mc->mc_snum < 2) {
7464 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7466 /* shift current top to make room for new parent */
7467 mc->mc_pg[1] = mc->mc_pg[0];
7468 mc->mc_ki[1] = mc->mc_ki[0];
7471 mc->mc_db->md_root = pp->mp_pgno;
7472 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7473 mc->mc_db->md_depth++;
7476 /* Add left (implicit) pointer. */
7477 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7478 /* undo the pre-push */
7479 mc->mc_pg[0] = mc->mc_pg[1];
7480 mc->mc_ki[0] = mc->mc_ki[1];
7481 mc->mc_db->md_root = mp->mp_pgno;
7482 mc->mc_db->md_depth--;
7489 ptop = mc->mc_top-1;
7490 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7493 mc->mc_flags |= C_SPLITTING;
7494 mdb_cursor_copy(mc, &mn);
7495 mn.mc_pg[mn.mc_top] = rp;
7496 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7498 if (nflags & MDB_APPEND) {
7499 mn.mc_ki[mn.mc_top] = 0;
7501 split_indx = newindx;
7505 split_indx = (nkeys+1) / 2;
7510 unsigned int lsize, rsize, ksize;
7511 /* Move half of the keys to the right sibling */
7513 x = mc->mc_ki[mc->mc_top] - split_indx;
7514 ksize = mc->mc_db->md_pad;
7515 split = LEAF2KEY(mp, split_indx, ksize);
7516 rsize = (nkeys - split_indx) * ksize;
7517 lsize = (nkeys - split_indx) * sizeof(indx_t);
7518 mp->mp_lower -= lsize;
7519 rp->mp_lower += lsize;
7520 mp->mp_upper += rsize - lsize;
7521 rp->mp_upper -= rsize - lsize;
7522 sepkey.mv_size = ksize;
7523 if (newindx == split_indx) {
7524 sepkey.mv_data = newkey->mv_data;
7526 sepkey.mv_data = split;
7529 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7530 memcpy(rp->mp_ptrs, split, rsize);
7531 sepkey.mv_data = rp->mp_ptrs;
7532 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7533 memcpy(ins, newkey->mv_data, ksize);
7534 mp->mp_lower += sizeof(indx_t);
7535 mp->mp_upper -= ksize - sizeof(indx_t);
7538 memcpy(rp->mp_ptrs, split, x * ksize);
7539 ins = LEAF2KEY(rp, x, ksize);
7540 memcpy(ins, newkey->mv_data, ksize);
7541 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7542 rp->mp_lower += sizeof(indx_t);
7543 rp->mp_upper -= ksize - sizeof(indx_t);
7544 mc->mc_ki[mc->mc_top] = x;
7545 mc->mc_pg[mc->mc_top] = rp;
7548 int psize, nsize, k;
7549 /* Maximum free space in an empty page */
7550 pmax = env->me_psize - PAGEHDRSZ;
7552 nsize = mdb_leaf_size(env, newkey, newdata);
7554 nsize = mdb_branch_size(env, newkey);
7555 nsize = EVEN(nsize);
7557 /* grab a page to hold a temporary copy */
7558 copy = mdb_page_malloc(mc->mc_txn, 1);
7561 copy->mp_pgno = mp->mp_pgno;
7562 copy->mp_flags = mp->mp_flags;
7563 copy->mp_lower = PAGEHDRSZ;
7564 copy->mp_upper = env->me_psize;
7566 /* prepare to insert */
7567 for (i=0, j=0; i<nkeys; i++) {
7569 copy->mp_ptrs[j++] = 0;
7571 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7574 /* When items are relatively large the split point needs
7575 * to be checked, because being off-by-one will make the
7576 * difference between success or failure in mdb_node_add.
7578 * It's also relevant if a page happens to be laid out
7579 * such that one half of its nodes are all "small" and
7580 * the other half of its nodes are "large." If the new
7581 * item is also "large" and falls on the half with
7582 * "large" nodes, it also may not fit.
7584 * As a final tweak, if the new item goes on the last
7585 * spot on the page (and thus, onto the new page), bias
7586 * the split so the new page is emptier than the old page.
7587 * This yields better packing during sequential inserts.
7589 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7590 /* Find split point */
7592 if (newindx <= split_indx || newindx >= nkeys) {
7594 k = newindx >= nkeys ? nkeys : split_indx+2;
7599 for (; i!=k; i+=j) {
7604 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7605 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7607 if (F_ISSET(node->mn_flags, F_BIGDATA))
7608 psize += sizeof(pgno_t);
7610 psize += NODEDSZ(node);
7612 psize = EVEN(psize);
7614 if (psize > pmax || i == k-j) {
7615 split_indx = i + (j<0);
7620 if (split_indx == newindx) {
7621 sepkey.mv_size = newkey->mv_size;
7622 sepkey.mv_data = newkey->mv_data;
7624 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7625 sepkey.mv_size = node->mn_ksize;
7626 sepkey.mv_data = NODEKEY(node);
7631 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7633 /* Copy separator key to the parent.
7635 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7639 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7642 if (mn.mc_snum == mc->mc_snum) {
7643 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7644 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7645 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7646 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7651 /* Right page might now have changed parent.
7652 * Check if left page also changed parent.
7654 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7655 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7656 for (i=0; i<ptop; i++) {
7657 mc->mc_pg[i] = mn.mc_pg[i];
7658 mc->mc_ki[i] = mn.mc_ki[i];
7660 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7661 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7665 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7668 mc->mc_flags ^= C_SPLITTING;
7669 if (rc != MDB_SUCCESS) {
7672 if (nflags & MDB_APPEND) {
7673 mc->mc_pg[mc->mc_top] = rp;
7674 mc->mc_ki[mc->mc_top] = 0;
7675 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7678 for (i=0; i<mc->mc_top; i++)
7679 mc->mc_ki[i] = mn.mc_ki[i];
7680 } else if (!IS_LEAF2(mp)) {
7682 mc->mc_pg[mc->mc_top] = rp;
7687 rkey.mv_data = newkey->mv_data;
7688 rkey.mv_size = newkey->mv_size;
7694 /* Update index for the new key. */
7695 mc->mc_ki[mc->mc_top] = j;
7697 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7698 rkey.mv_data = NODEKEY(node);
7699 rkey.mv_size = node->mn_ksize;
7701 xdata.mv_data = NODEDATA(node);
7702 xdata.mv_size = NODEDSZ(node);
7705 pgno = NODEPGNO(node);
7706 flags = node->mn_flags;
7709 if (!IS_LEAF(mp) && j == 0) {
7710 /* First branch index doesn't need key data. */
7714 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7716 /* return tmp page to freelist */
7717 mdb_page_free(env, copy);
7723 mc->mc_pg[mc->mc_top] = copy;
7728 } while (i != split_indx);
7730 nkeys = NUMKEYS(copy);
7731 for (i=0; i<nkeys; i++)
7732 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7733 mp->mp_lower = copy->mp_lower;
7734 mp->mp_upper = copy->mp_upper;
7735 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7736 env->me_psize - copy->mp_upper);
7738 /* reset back to original page */
7739 if (newindx < split_indx) {
7740 mc->mc_pg[mc->mc_top] = mp;
7741 if (nflags & MDB_RESERVE) {
7742 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7743 if (!(node->mn_flags & F_BIGDATA))
7744 newdata->mv_data = NODEDATA(node);
7747 mc->mc_pg[mc->mc_top] = rp;
7749 /* Make sure mc_ki is still valid.
7751 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7752 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7753 for (i=0; i<ptop; i++) {
7754 mc->mc_pg[i] = mn.mc_pg[i];
7755 mc->mc_ki[i] = mn.mc_ki[i];
7757 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7758 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7761 /* return tmp page to freelist */
7762 mdb_page_free(env, copy);
7766 /* Adjust other cursors pointing to mp */
7767 MDB_cursor *m2, *m3;
7768 MDB_dbi dbi = mc->mc_dbi;
7769 int fixup = NUMKEYS(mp);
7771 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7772 if (mc->mc_flags & C_SUB)
7773 m3 = &m2->mc_xcursor->mx_cursor;
7778 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7780 if (m3->mc_flags & C_SPLITTING)
7785 for (k=m3->mc_top; k>=0; k--) {
7786 m3->mc_ki[k+1] = m3->mc_ki[k];
7787 m3->mc_pg[k+1] = m3->mc_pg[k];
7789 if (m3->mc_ki[0] >= split_indx) {
7794 m3->mc_pg[0] = mc->mc_pg[0];
7798 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7799 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7800 m3->mc_ki[mc->mc_top]++;
7801 if (m3->mc_ki[mc->mc_top] >= fixup) {
7802 m3->mc_pg[mc->mc_top] = rp;
7803 m3->mc_ki[mc->mc_top] -= fixup;
7804 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7806 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7807 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7812 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
7817 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7818 MDB_val *key, MDB_val *data, unsigned int flags)
7823 assert(key != NULL);
7824 assert(data != NULL);
7826 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7829 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7832 mdb_cursor_init(&mc, txn, dbi, &mx);
7833 return mdb_cursor_put(&mc, key, data, flags);
7837 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7839 if ((flag & CHANGEABLE) != flag)
7842 env->me_flags |= flag;
7844 env->me_flags &= ~flag;
7849 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7854 *arg = env->me_flags;
7859 mdb_env_get_path(MDB_env *env, const char **arg)
7864 *arg = env->me_path;
7869 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
7878 /** Common code for #mdb_stat() and #mdb_env_stat().
7879 * @param[in] env the environment to operate in.
7880 * @param[in] db the #MDB_db record containing the stats to return.
7881 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7882 * @return 0, this function always succeeds.
7885 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7887 arg->ms_psize = env->me_psize;
7888 arg->ms_depth = db->md_depth;
7889 arg->ms_branch_pages = db->md_branch_pages;
7890 arg->ms_leaf_pages = db->md_leaf_pages;
7891 arg->ms_overflow_pages = db->md_overflow_pages;
7892 arg->ms_entries = db->md_entries;
7897 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7901 if (env == NULL || arg == NULL)
7904 toggle = mdb_env_pick_meta(env);
7906 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7910 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7914 if (env == NULL || arg == NULL)
7917 toggle = mdb_env_pick_meta(env);
7918 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7919 arg->me_mapsize = env->me_mapsize;
7920 arg->me_maxreaders = env->me_maxreaders;
7922 /* me_numreaders may be zero if this process never used any readers. Use
7923 * the shared numreader count if it exists.
7925 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7927 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7928 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7932 /** Set the default comparison functions for a database.
7933 * Called immediately after a database is opened to set the defaults.
7934 * The user can then override them with #mdb_set_compare() or
7935 * #mdb_set_dupsort().
7936 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7937 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7940 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7942 uint16_t f = txn->mt_dbs[dbi].md_flags;
7944 txn->mt_dbxs[dbi].md_cmp =
7945 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7946 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7948 txn->mt_dbxs[dbi].md_dcmp =
7949 !(f & MDB_DUPSORT) ? 0 :
7950 ((f & MDB_INTEGERDUP)
7951 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7952 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7955 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7960 int rc, dbflag, exact;
7961 unsigned int unused = 0;
7964 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7965 mdb_default_cmp(txn, FREE_DBI);
7968 if ((flags & VALID_FLAGS) != flags)
7970 if (txn->mt_flags & MDB_TXN_ERROR)
7976 if (flags & PERSISTENT_FLAGS) {
7977 uint16_t f2 = flags & PERSISTENT_FLAGS;
7978 /* make sure flag changes get committed */
7979 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7980 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7981 txn->mt_flags |= MDB_TXN_DIRTY;
7984 mdb_default_cmp(txn, MAIN_DBI);
7988 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7989 mdb_default_cmp(txn, MAIN_DBI);
7992 /* Is the DB already open? */
7994 for (i=2; i<txn->mt_numdbs; i++) {
7995 if (!txn->mt_dbxs[i].md_name.mv_size) {
7996 /* Remember this free slot */
7997 if (!unused) unused = i;
8000 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8001 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8007 /* If no free slot and max hit, fail */
8008 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8009 return MDB_DBS_FULL;
8011 /* Cannot mix named databases with some mainDB flags */
8012 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8013 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8015 /* Find the DB info */
8016 dbflag = DB_NEW|DB_VALID;
8019 key.mv_data = (void *)name;
8020 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8021 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8022 if (rc == MDB_SUCCESS) {
8023 /* make sure this is actually a DB */
8024 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8025 if (!(node->mn_flags & F_SUBDATA))
8026 return MDB_INCOMPATIBLE;
8027 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8028 /* Create if requested */
8030 data.mv_size = sizeof(MDB_db);
8031 data.mv_data = &dummy;
8032 memset(&dummy, 0, sizeof(dummy));
8033 dummy.md_root = P_INVALID;
8034 dummy.md_flags = flags & PERSISTENT_FLAGS;
8035 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8039 /* OK, got info, add to table */
8040 if (rc == MDB_SUCCESS) {
8041 unsigned int slot = unused ? unused : txn->mt_numdbs;
8042 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8043 txn->mt_dbxs[slot].md_name.mv_size = len;
8044 txn->mt_dbxs[slot].md_rel = NULL;
8045 txn->mt_dbflags[slot] = dbflag;
8046 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8048 mdb_default_cmp(txn, slot);
8057 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8059 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8062 if (txn->mt_dbflags[dbi] & DB_STALE) {
8065 /* Stale, must read the DB's root. cursor_init does it for us. */
8066 mdb_cursor_init(&mc, txn, dbi, &mx);
8068 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8071 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8074 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8076 ptr = env->me_dbxs[dbi].md_name.mv_data;
8077 env->me_dbxs[dbi].md_name.mv_data = NULL;
8078 env->me_dbxs[dbi].md_name.mv_size = 0;
8079 env->me_dbflags[dbi] = 0;
8083 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8085 /* We could return the flags for the FREE_DBI too but what's the point? */
8086 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8088 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8092 /** Add all the DB's pages to the free list.
8093 * @param[in] mc Cursor on the DB to free.
8094 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8095 * @return 0 on success, non-zero on failure.
8098 mdb_drop0(MDB_cursor *mc, int subs)
8102 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8103 if (rc == MDB_SUCCESS) {
8104 MDB_txn *txn = mc->mc_txn;
8109 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8110 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8113 mdb_cursor_copy(mc, &mx);
8114 while (mc->mc_snum > 0) {
8115 MDB_page *mp = mc->mc_pg[mc->mc_top];
8116 unsigned n = NUMKEYS(mp);
8118 for (i=0; i<n; i++) {
8119 ni = NODEPTR(mp, i);
8120 if (ni->mn_flags & F_BIGDATA) {
8123 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8124 rc = mdb_page_get(txn, pg, &omp, NULL);
8127 assert(IS_OVERFLOW(omp));
8128 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8132 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8133 mdb_xcursor_init1(mc, ni);
8134 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8140 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8142 for (i=0; i<n; i++) {
8144 ni = NODEPTR(mp, i);
8147 mdb_midl_xappend(txn->mt_free_pgs, pg);
8152 mc->mc_ki[mc->mc_top] = i;
8153 rc = mdb_cursor_sibling(mc, 1);
8155 /* no more siblings, go back to beginning
8156 * of previous level.
8160 for (i=1; i<mc->mc_snum; i++) {
8162 mc->mc_pg[i] = mx.mc_pg[i];
8167 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8168 } else if (rc == MDB_NOTFOUND) {
8174 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8176 MDB_cursor *mc, *m2;
8179 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8182 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8185 rc = mdb_cursor_open(txn, dbi, &mc);
8189 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8190 /* Invalidate the dropped DB's cursors */
8191 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8192 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8196 /* Can't delete the main DB */
8197 if (del && dbi > MAIN_DBI) {
8198 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8200 txn->mt_dbflags[dbi] = DB_STALE;
8201 mdb_dbi_close(txn->mt_env, dbi);
8204 /* reset the DB record, mark it dirty */
8205 txn->mt_dbflags[dbi] |= DB_DIRTY;
8206 txn->mt_dbs[dbi].md_depth = 0;
8207 txn->mt_dbs[dbi].md_branch_pages = 0;
8208 txn->mt_dbs[dbi].md_leaf_pages = 0;
8209 txn->mt_dbs[dbi].md_overflow_pages = 0;
8210 txn->mt_dbs[dbi].md_entries = 0;
8211 txn->mt_dbs[dbi].md_root = P_INVALID;
8213 txn->mt_flags |= MDB_TXN_DIRTY;
8216 mdb_cursor_close(mc);
8220 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8222 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8225 txn->mt_dbxs[dbi].md_cmp = cmp;
8229 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8231 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8234 txn->mt_dbxs[dbi].md_dcmp = cmp;
8238 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8240 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8243 txn->mt_dbxs[dbi].md_rel = rel;
8247 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8249 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8252 txn->mt_dbxs[dbi].md_relctx = ctx;
8256 int mdb_env_get_maxkeysize(MDB_env *env)
8258 return MDB_MAXKEYSIZE;
8261 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8263 unsigned int i, rdrs;
8270 if (!env->me_txns) {
8271 return func("(no reader locks)\n", ctx);
8273 rdrs = env->me_txns->mti_numreaders;
8274 mr = env->me_txns->mti_readers;
8275 for (i=0; i<rdrs; i++) {
8280 if (mr[i].mr_txnid == (txnid_t)-1) {
8281 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8283 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8287 func(" pid thread txnid\n", ctx);
8289 rc = func(buf, ctx);
8295 func("(no active readers)\n", ctx);
8300 /** Insert pid into list if not already present.
8301 * return -1 if already present.
8303 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8305 /* binary search of pid in list */
8307 unsigned cursor = 1;
8309 unsigned n = ids[0];
8312 unsigned pivot = n >> 1;
8313 cursor = base + pivot + 1;
8314 val = pid - ids[cursor];
8319 } else if ( val > 0 ) {
8324 /* found, so it's a duplicate */
8333 for (n = ids[0]; n > cursor; n--)
8339 int mdb_reader_check(MDB_env *env, int *dead)
8341 unsigned int i, j, rdrs;
8343 MDB_PID_T *pids, pid;
8352 rdrs = env->me_txns->mti_numreaders;
8353 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8357 mr = env->me_txns->mti_readers;
8358 for (i=0; i<rdrs; i++) {
8359 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8361 if (mdb_pid_insert(pids, pid) == 0) {
8362 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8364 /* Recheck, a new process may have reused pid */
8365 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8366 for (j=i; j<rdrs; j++)
8367 if (mr[j].mr_pid == pid) {
8368 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8369 (unsigned) pid, mr[j].mr_txnid));
8374 UNLOCK_MUTEX_R(env);
projects / openldap / blob