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 * can be raised to page size / #MDB_MINKEYS - (64-bit ? 66 : 44).
390 * Note that data items in an #MDB_DUPSORT database are actually keys
391 * of a subDB, so they're also limited to this size.
393 #ifndef MDB_MAXKEYSIZE
394 #define MDB_MAXKEYSIZE 511
397 /** @brief The maximum size of a data item.
399 * We only store a 32 bit value for node sizes.
401 #define MAXDATASIZE 0xffffffffUL
406 * This is used for printing a hex dump of a key's contents.
408 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
409 /** Display a key in hex.
411 * Invoke a function to display a key in hex.
413 #define DKEY(x) mdb_dkey(x, kbuf)
419 /** An invalid page number.
420 * Mainly used to denote an empty tree.
422 #define P_INVALID (~(pgno_t)0)
424 /** Test if the flags \b f are set in a flag word \b w. */
425 #define F_ISSET(w, f) (((w) & (f)) == (f))
427 /** Round \b n up to an even number. */
428 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
430 /** Used for offsets within a single page.
431 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
434 typedef uint16_t indx_t;
436 /** Default size of memory map.
437 * This is certainly too small for any actual applications. Apps should always set
438 * the size explicitly using #mdb_env_set_mapsize().
440 #define DEFAULT_MAPSIZE 1048576
442 /** @defgroup readers Reader Lock Table
443 * Readers don't acquire any locks for their data access. Instead, they
444 * simply record their transaction ID in the reader table. The reader
445 * mutex is needed just to find an empty slot in the reader table. The
446 * slot's address is saved in thread-specific data so that subsequent read
447 * transactions started by the same thread need no further locking to proceed.
449 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
451 * No reader table is used if the database is on a read-only filesystem, or
452 * if #MDB_NOLOCK is set.
454 * Since the database uses multi-version concurrency control, readers don't
455 * actually need any locking. This table is used to keep track of which
456 * readers are using data from which old transactions, so that we'll know
457 * when a particular old transaction is no longer in use. Old transactions
458 * that have discarded any data pages can then have those pages reclaimed
459 * for use by a later write transaction.
461 * The lock table is constructed such that reader slots are aligned with the
462 * processor's cache line size. Any slot is only ever used by one thread.
463 * This alignment guarantees that there will be no contention or cache
464 * thrashing as threads update their own slot info, and also eliminates
465 * any need for locking when accessing a slot.
467 * A writer thread will scan every slot in the table to determine the oldest
468 * outstanding reader transaction. Any freed pages older than this will be
469 * reclaimed by the writer. The writer doesn't use any locks when scanning
470 * this table. This means that there's no guarantee that the writer will
471 * see the most up-to-date reader info, but that's not required for correct
472 * operation - all we need is to know the upper bound on the oldest reader,
473 * we don't care at all about the newest reader. So the only consequence of
474 * reading stale information here is that old pages might hang around a
475 * while longer before being reclaimed. That's actually good anyway, because
476 * the longer we delay reclaiming old pages, the more likely it is that a
477 * string of contiguous pages can be found after coalescing old pages from
478 * many old transactions together.
481 /** Number of slots in the reader table.
482 * This value was chosen somewhat arbitrarily. 126 readers plus a
483 * couple mutexes fit exactly into 8KB on my development machine.
484 * Applications should set the table size using #mdb_env_set_maxreaders().
486 #define DEFAULT_READERS 126
488 /** The size of a CPU cache line in bytes. We want our lock structures
489 * aligned to this size to avoid false cache line sharing in the
491 * This value works for most CPUs. For Itanium this should be 128.
497 /** The information we store in a single slot of the reader table.
498 * In addition to a transaction ID, we also record the process and
499 * thread ID that owns a slot, so that we can detect stale information,
500 * e.g. threads or processes that went away without cleaning up.
501 * @note We currently don't check for stale records. We simply re-init
502 * the table when we know that we're the only process opening the
505 typedef struct MDB_rxbody {
506 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
507 * Multiple readers that start at the same time will probably have the
508 * same ID here. Again, it's not important to exclude them from
509 * anything; all we need to know is which version of the DB they
510 * started from so we can avoid overwriting any data used in that
511 * particular version.
514 /** The process ID of the process owning this reader txn. */
516 /** The thread ID of the thread owning this txn. */
520 /** The actual reader record, with cacheline padding. */
521 typedef struct MDB_reader {
524 /** shorthand for mrb_txnid */
525 #define mr_txnid mru.mrx.mrb_txnid
526 #define mr_pid mru.mrx.mrb_pid
527 #define mr_tid mru.mrx.mrb_tid
528 /** cache line alignment */
529 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
533 /** The header for the reader table.
534 * The table resides in a memory-mapped file. (This is a different file
535 * than is used for the main database.)
537 * For POSIX the actual mutexes reside in the shared memory of this
538 * mapped file. On Windows, mutexes are named objects allocated by the
539 * kernel; we store the mutex names in this mapped file so that other
540 * processes can grab them. This same approach is also used on
541 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
542 * process-shared POSIX mutexes. For these cases where a named object
543 * is used, the object name is derived from a 64 bit FNV hash of the
544 * environment pathname. As such, naming collisions are extremely
545 * unlikely. If a collision occurs, the results are unpredictable.
547 typedef struct MDB_txbody {
548 /** Stamp identifying this as an MDB file. It must be set
551 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
553 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
554 char mtb_rmname[MNAME_LEN];
556 /** Mutex protecting access to this table.
557 * This is the reader lock that #LOCK_MUTEX_R acquires.
559 pthread_mutex_t mtb_mutex;
561 /** The ID of the last transaction committed to the database.
562 * This is recorded here only for convenience; the value can always
563 * be determined by reading the main database meta pages.
566 /** The number of slots that have been used in the reader table.
567 * This always records the maximum count, it is not decremented
568 * when readers release their slots.
570 unsigned mtb_numreaders;
573 /** The actual reader table definition. */
574 typedef struct MDB_txninfo {
577 #define mti_magic mt1.mtb.mtb_magic
578 #define mti_format mt1.mtb.mtb_format
579 #define mti_mutex mt1.mtb.mtb_mutex
580 #define mti_rmname mt1.mtb.mtb_rmname
581 #define mti_txnid mt1.mtb.mtb_txnid
582 #define mti_numreaders mt1.mtb.mtb_numreaders
583 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
586 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
587 char mt2_wmname[MNAME_LEN];
588 #define mti_wmname mt2.mt2_wmname
590 pthread_mutex_t mt2_wmutex;
591 #define mti_wmutex mt2.mt2_wmutex
593 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
595 MDB_reader mti_readers[1];
598 /** Lockfile format signature: version, features and field layout */
599 #define MDB_LOCK_FORMAT \
601 ((MDB_LOCK_VERSION) \
602 /* Flags which describe functionality */ \
603 + (((MDB_PIDLOCK) != 0) << 16)))
606 /** Common header for all page types.
607 * Overflow records occupy a number of contiguous pages with no
608 * headers on any page after the first.
610 typedef struct MDB_page {
611 #define mp_pgno mp_p.p_pgno
612 #define mp_next mp_p.p_next
614 pgno_t p_pgno; /**< page number */
615 void * p_next; /**< for in-memory list of freed structs */
618 /** @defgroup mdb_page Page Flags
620 * Flags for the page headers.
623 #define P_BRANCH 0x01 /**< branch page */
624 #define P_LEAF 0x02 /**< leaf page */
625 #define P_OVERFLOW 0x04 /**< overflow page */
626 #define P_META 0x08 /**< meta page */
627 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
628 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
629 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
630 #define P_KEEP 0x8000 /**< leave this page alone during spill */
632 uint16_t mp_flags; /**< @ref mdb_page */
633 #define mp_lower mp_pb.pb.pb_lower
634 #define mp_upper mp_pb.pb.pb_upper
635 #define mp_pages mp_pb.pb_pages
638 indx_t pb_lower; /**< lower bound of free space */
639 indx_t pb_upper; /**< upper bound of free space */
641 uint32_t pb_pages; /**< number of overflow pages */
643 indx_t mp_ptrs[1]; /**< dynamic size */
646 /** Size of the page header, excluding dynamic data at the end */
647 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
649 /** Address of first usable data byte in a page, after the header */
650 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
652 /** Number of nodes on a page */
653 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
655 /** The amount of space remaining in the page */
656 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
658 /** The percentage of space used in the page, in tenths of a percent. */
659 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
660 ((env)->me_psize - PAGEHDRSZ))
661 /** The minimum page fill factor, in tenths of a percent.
662 * Pages emptier than this are candidates for merging.
664 #define FILL_THRESHOLD 250
666 /** Test if a page is a leaf page */
667 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
668 /** Test if a page is a LEAF2 page */
669 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
670 /** Test if a page is a branch page */
671 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
672 /** Test if a page is an overflow page */
673 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
674 /** Test if a page is a sub page */
675 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
677 /** The number of overflow pages needed to store the given size. */
678 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
680 /** Header for a single key/data pair within a page.
681 * We guarantee 2-byte alignment for nodes.
683 typedef struct MDB_node {
684 /** lo and hi are used for data size on leaf nodes and for
685 * child pgno on branch nodes. On 64 bit platforms, flags
686 * is also used for pgno. (Branch nodes have no flags).
687 * They are in host byte order in case that lets some
688 * accesses be optimized into a 32-bit word access.
690 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
691 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
692 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
693 /** @defgroup mdb_node Node Flags
695 * Flags for node headers.
698 #define F_BIGDATA 0x01 /**< data put on overflow page */
699 #define F_SUBDATA 0x02 /**< data is a sub-database */
700 #define F_DUPDATA 0x04 /**< data has duplicates */
702 /** valid flags for #mdb_node_add() */
703 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
706 unsigned short mn_flags; /**< @ref mdb_node */
707 unsigned short mn_ksize; /**< key size */
708 char mn_data[1]; /**< key and data are appended here */
711 /** Size of the node header, excluding dynamic data at the end */
712 #define NODESIZE offsetof(MDB_node, mn_data)
714 /** Bit position of top word in page number, for shifting mn_flags */
715 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
717 /** Size of a node in a branch page with a given key.
718 * This is just the node header plus the key, there is no data.
720 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
722 /** Size of a node in a leaf page with a given key and data.
723 * This is node header plus key plus data size.
725 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
727 /** Address of node \b i in page \b p */
728 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
730 /** Address of the key for the node */
731 #define NODEKEY(node) (void *)((node)->mn_data)
733 /** Address of the data for a node */
734 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
736 /** Get the page number pointed to by a branch node */
737 #define NODEPGNO(node) \
738 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
739 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
740 /** Set the page number in a branch node */
741 #define SETPGNO(node,pgno) do { \
742 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
743 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
745 /** Get the size of the data in a leaf node */
746 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
747 /** Set the size of the data for a leaf node */
748 #define SETDSZ(node,size) do { \
749 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
750 /** The size of a key in a node */
751 #define NODEKSZ(node) ((node)->mn_ksize)
753 /** Copy a page number from src to dst */
755 #define COPY_PGNO(dst,src) dst = src
757 #if SIZE_MAX > 4294967295UL
758 #define COPY_PGNO(dst,src) do { \
759 unsigned short *s, *d; \
760 s = (unsigned short *)&(src); \
761 d = (unsigned short *)&(dst); \
768 #define COPY_PGNO(dst,src) do { \
769 unsigned short *s, *d; \
770 s = (unsigned short *)&(src); \
771 d = (unsigned short *)&(dst); \
777 /** The address of a key in a LEAF2 page.
778 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
779 * There are no node headers, keys are stored contiguously.
781 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
783 /** Set the \b node's key into \b keyptr, if requested. */
784 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
785 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
787 /** Set the \b node's key into \b key. */
788 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
790 /** Information about a single database in the environment. */
791 typedef struct MDB_db {
792 uint32_t md_pad; /**< also ksize for LEAF2 pages */
793 uint16_t md_flags; /**< @ref mdb_dbi_open */
794 uint16_t md_depth; /**< depth of this tree */
795 pgno_t md_branch_pages; /**< number of internal pages */
796 pgno_t md_leaf_pages; /**< number of leaf pages */
797 pgno_t md_overflow_pages; /**< number of overflow pages */
798 size_t md_entries; /**< number of data items */
799 pgno_t md_root; /**< the root page of this tree */
802 /** mdb_dbi_open flags */
803 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
804 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
805 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
806 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
808 /** Handle for the DB used to track free pages. */
810 /** Handle for the default DB. */
813 /** Meta page content.
814 * A meta page is the start point for accessing a database snapshot.
815 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
817 typedef struct MDB_meta {
818 /** Stamp identifying this as an MDB file. It must be set
821 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
823 void *mm_address; /**< address for fixed mapping */
824 size_t mm_mapsize; /**< size of mmap region */
825 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
826 /** The size of pages used in this DB */
827 #define mm_psize mm_dbs[0].md_pad
828 /** Any persistent environment flags. @ref mdb_env */
829 #define mm_flags mm_dbs[0].md_flags
830 pgno_t mm_last_pg; /**< last used page in file */
831 txnid_t mm_txnid; /**< txnid that committed this page */
834 /** Buffer for a stack-allocated meta page.
835 * The members define size and alignment, and silence type
836 * aliasing warnings. They are not used directly; that could
837 * mean incorrectly using several union members in parallel.
839 typedef union MDB_metabuf {
842 char mm_pad[PAGEHDRSZ];
847 /** Auxiliary DB info.
848 * The information here is mostly static/read-only. There is
849 * only a single copy of this record in the environment.
851 typedef struct MDB_dbx {
852 MDB_val md_name; /**< name of the database */
853 MDB_cmp_func *md_cmp; /**< function for comparing keys */
854 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
855 MDB_rel_func *md_rel; /**< user relocate function */
856 void *md_relctx; /**< user-provided context for md_rel */
859 /** A database transaction.
860 * Every operation requires a transaction handle.
863 MDB_txn *mt_parent; /**< parent of a nested txn */
864 MDB_txn *mt_child; /**< nested txn under this txn */
865 pgno_t mt_next_pgno; /**< next unallocated page */
866 /** The ID of this transaction. IDs are integers incrementing from 1.
867 * Only committed write transactions increment the ID. If a transaction
868 * aborts, the ID may be re-used by the next writer.
871 MDB_env *mt_env; /**< the DB environment */
872 /** The list of pages that became unused during this transaction.
875 /** The sorted list of dirty pages we temporarily wrote to disk
876 * because the dirty list was full. page numbers in here are
877 * shifted left by 1, deleted slots have the LSB set.
879 MDB_IDL mt_spill_pgs;
881 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
883 /** For read txns: This thread/txn's reader table slot, or NULL. */
886 /** Array of records for each DB known in the environment. */
888 /** Array of MDB_db records for each known DB */
890 /** @defgroup mt_dbflag Transaction DB Flags
894 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
895 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
896 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
897 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
899 /** In write txns, array of cursors for each DB */
900 MDB_cursor **mt_cursors;
901 /** Array of flags for each DB */
902 unsigned char *mt_dbflags;
903 /** Number of DB records in use. This number only ever increments;
904 * we don't decrement it when individual DB handles are closed.
908 /** @defgroup mdb_txn Transaction Flags
912 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
913 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
914 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
915 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
917 unsigned int mt_flags; /**< @ref mdb_txn */
918 /** dirty_list room: Array size - #dirty pages visible to this txn.
919 * Includes ancestor txns' dirty pages not hidden by other txns'
920 * dirty/spilled pages. Thus commit(nested txn) has room to merge
921 * dirty_list into mt_parent after freeing hidden mt_parent pages.
923 unsigned int mt_dirty_room;
926 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
927 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
928 * raise this on a 64 bit machine.
930 #define CURSOR_STACK 32
934 /** Cursors are used for all DB operations.
935 * A cursor holds a path of (page pointer, key index) from the DB
936 * root to a position in the DB, plus other state. #MDB_DUPSORT
937 * cursors include an xcursor to the current data item. Write txns
938 * track their cursors and keep them up to date when data moves.
939 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
940 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
943 /** Next cursor on this DB in this txn */
945 /** Backup of the original cursor if this cursor is a shadow */
946 MDB_cursor *mc_backup;
947 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
948 struct MDB_xcursor *mc_xcursor;
949 /** The transaction that owns this cursor */
951 /** The database handle this cursor operates on */
953 /** The database record for this cursor */
955 /** The database auxiliary record for this cursor */
957 /** The @ref mt_dbflag for this database */
958 unsigned char *mc_dbflag;
959 unsigned short mc_snum; /**< number of pushed pages */
960 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
961 /** @defgroup mdb_cursor Cursor Flags
963 * Cursor state flags.
966 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
967 #define C_EOF 0x02 /**< No more data */
968 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
969 #define C_DEL 0x08 /**< last op was a cursor_del */
970 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
971 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
973 unsigned int mc_flags; /**< @ref mdb_cursor */
974 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
975 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
978 /** Context for sorted-dup records.
979 * We could have gone to a fully recursive design, with arbitrarily
980 * deep nesting of sub-databases. But for now we only handle these
981 * levels - main DB, optional sub-DB, sorted-duplicate DB.
983 typedef struct MDB_xcursor {
984 /** A sub-cursor for traversing the Dup DB */
985 MDB_cursor mx_cursor;
986 /** The database record for this Dup DB */
988 /** The auxiliary DB record for this Dup DB */
990 /** The @ref mt_dbflag for this Dup DB */
991 unsigned char mx_dbflag;
994 /** State of FreeDB old pages, stored in the MDB_env */
995 typedef struct MDB_pgstate {
996 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
997 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1000 /** The database environment. */
1002 HANDLE me_fd; /**< The main data file */
1003 HANDLE me_lfd; /**< The lock file */
1004 HANDLE me_mfd; /**< just for writing the meta pages */
1005 /** Failed to update the meta page. Probably an I/O error. */
1006 #define MDB_FATAL_ERROR 0x80000000U
1007 /** Some fields are initialized. */
1008 #define MDB_ENV_ACTIVE 0x20000000U
1009 /** me_txkey is set */
1010 #define MDB_ENV_TXKEY 0x10000000U
1011 /** Have liveness lock in reader table */
1012 #define MDB_LIVE_READER 0x08000000U
1013 uint32_t me_flags; /**< @ref mdb_env */
1014 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1015 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1016 unsigned int me_maxreaders; /**< size of the reader table */
1017 unsigned int me_numreaders; /**< max numreaders set by this env */
1018 MDB_dbi me_numdbs; /**< number of DBs opened */
1019 MDB_dbi me_maxdbs; /**< size of the DB table */
1020 MDB_PID_T me_pid; /**< process ID of this env */
1021 char *me_path; /**< path to the DB files */
1022 char *me_map; /**< the memory map of the data file */
1023 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1024 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1025 void *me_pbuf; /**< scratch area for DUPSORT put() */
1026 MDB_txn *me_txn; /**< current write transaction */
1027 size_t me_mapsize; /**< size of the data memory map */
1028 off_t me_size; /**< current file size */
1029 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1030 MDB_dbx *me_dbxs; /**< array of static DB info */
1031 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1032 pthread_key_t me_txkey; /**< thread-key for readers */
1033 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1034 # define me_pglast me_pgstate.mf_pglast
1035 # define me_pghead me_pgstate.mf_pghead
1036 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1037 /** IDL of pages that became unused in a write txn */
1038 MDB_IDL me_free_pgs;
1039 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1040 MDB_ID2L me_dirty_list;
1041 /** Max number of freelist items that can fit in a single overflow page */
1043 /** Max size of a node on a page */
1044 unsigned int me_nodemax;
1046 int me_pidquery; /**< Used in OpenProcess */
1047 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1049 #elif defined(MDB_USE_POSIX_SEM)
1050 sem_t *me_rmutex; /* Shared mutexes are not supported */
1055 /** Nested transaction */
1056 typedef struct MDB_ntxn {
1057 MDB_txn mnt_txn; /**< the transaction */
1058 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1061 /** max number of pages to commit in one writev() call */
1062 #define MDB_COMMIT_PAGES 64
1063 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1064 #undef MDB_COMMIT_PAGES
1065 #define MDB_COMMIT_PAGES IOV_MAX
1068 /* max bytes to write in one call */
1069 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1071 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1072 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1073 static int mdb_page_touch(MDB_cursor *mc);
1075 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1076 static int mdb_page_search_root(MDB_cursor *mc,
1077 MDB_val *key, int modify);
1078 #define MDB_PS_MODIFY 1
1079 #define MDB_PS_ROOTONLY 2
1080 #define MDB_PS_FIRST 4
1081 #define MDB_PS_LAST 8
1082 static int mdb_page_search(MDB_cursor *mc,
1083 MDB_val *key, int flags);
1084 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1086 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1087 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1088 pgno_t newpgno, unsigned int nflags);
1090 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1091 static int mdb_env_pick_meta(const MDB_env *env);
1092 static int mdb_env_write_meta(MDB_txn *txn);
1093 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1094 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1096 static void mdb_env_close0(MDB_env *env, int excl);
1098 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1099 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1100 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1101 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1102 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1103 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1104 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1105 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1106 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1108 static int mdb_rebalance(MDB_cursor *mc);
1109 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1111 static void mdb_cursor_pop(MDB_cursor *mc);
1112 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1114 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1115 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1116 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1117 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1118 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1120 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1121 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1123 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1124 static void mdb_xcursor_init0(MDB_cursor *mc);
1125 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1127 static int mdb_drop0(MDB_cursor *mc, int subs);
1128 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1131 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1135 static SECURITY_DESCRIPTOR mdb_null_sd;
1136 static SECURITY_ATTRIBUTES mdb_all_sa;
1137 static int mdb_sec_inited;
1140 /** Return the library version info. */
1142 mdb_version(int *major, int *minor, int *patch)
1144 if (major) *major = MDB_VERSION_MAJOR;
1145 if (minor) *minor = MDB_VERSION_MINOR;
1146 if (patch) *patch = MDB_VERSION_PATCH;
1147 return MDB_VERSION_STRING;
1150 /** Table of descriptions for MDB @ref errors */
1151 static char *const mdb_errstr[] = {
1152 "MDB_KEYEXIST: Key/data pair already exists",
1153 "MDB_NOTFOUND: No matching key/data pair found",
1154 "MDB_PAGE_NOTFOUND: Requested page not found",
1155 "MDB_CORRUPTED: Located page was wrong type",
1156 "MDB_PANIC: Update of meta page failed",
1157 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1158 "MDB_INVALID: File is not an MDB file",
1159 "MDB_MAP_FULL: Environment mapsize limit reached",
1160 "MDB_DBS_FULL: Environment maxdbs limit reached",
1161 "MDB_READERS_FULL: Environment maxreaders limit reached",
1162 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1163 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1164 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1165 "MDB_PAGE_FULL: Internal error - page has no more space",
1166 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1167 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1168 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1169 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1170 "MDB_BAD_VALSIZE: Too big key/data, key is empty, or wrong DUPFIXED size",
1174 mdb_strerror(int err)
1178 return ("Successful return: 0");
1180 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1181 i = err - MDB_KEYEXIST;
1182 return mdb_errstr[i];
1185 return strerror(err);
1189 /** Display a key in hexadecimal and return the address of the result.
1190 * @param[in] key the key to display
1191 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1192 * @return The key in hexadecimal form.
1195 mdb_dkey(MDB_val *key, char *buf)
1198 unsigned char *c = key->mv_data;
1204 if (key->mv_size > MDB_MAXKEYSIZE)
1205 return "MDB_MAXKEYSIZE";
1206 /* may want to make this a dynamic check: if the key is mostly
1207 * printable characters, print it as-is instead of converting to hex.
1211 for (i=0; i<key->mv_size; i++)
1212 ptr += sprintf(ptr, "%02x", *c++);
1214 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1219 /** Display all the keys in the page. */
1221 mdb_page_list(MDB_page *mp)
1224 unsigned int i, nkeys, nsize, total = 0;
1228 nkeys = NUMKEYS(mp);
1229 fprintf(stderr, "Page %"Z"u numkeys %d\n", mp->mp_pgno, nkeys);
1230 for (i=0; i<nkeys; i++) {
1231 node = NODEPTR(mp, i);
1232 key.mv_size = node->mn_ksize;
1233 key.mv_data = node->mn_data;
1234 nsize = NODESIZE + key.mv_size;
1235 if (IS_BRANCH(mp)) {
1236 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1240 if (F_ISSET(node->mn_flags, F_BIGDATA))
1241 nsize += sizeof(pgno_t);
1243 nsize += NODEDSZ(node);
1245 nsize += sizeof(indx_t);
1246 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1248 total = EVEN(total);
1250 fprintf(stderr, "Total: %d\n", total);
1254 mdb_cursor_chk(MDB_cursor *mc)
1260 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1261 for (i=0; i<mc->mc_top; i++) {
1263 node = NODEPTR(mp, mc->mc_ki[i]);
1264 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1267 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1273 /** Count all the pages in each DB and in the freelist
1274 * and make sure it matches the actual number of pages
1277 static void mdb_audit(MDB_txn *txn)
1281 MDB_ID freecount, count;
1286 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1287 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1288 freecount += *(MDB_ID *)data.mv_data;
1291 for (i = 0; i<txn->mt_numdbs; i++) {
1293 mdb_cursor_init(&mc, txn, i, &mx);
1294 if (txn->mt_dbs[i].md_root == P_INVALID)
1296 count += txn->mt_dbs[i].md_branch_pages +
1297 txn->mt_dbs[i].md_leaf_pages +
1298 txn->mt_dbs[i].md_overflow_pages;
1299 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1300 mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1304 mp = mc.mc_pg[mc.mc_top];
1305 for (j=0; j<NUMKEYS(mp); j++) {
1306 MDB_node *leaf = NODEPTR(mp, j);
1307 if (leaf->mn_flags & F_SUBDATA) {
1309 memcpy(&db, NODEDATA(leaf), sizeof(db));
1310 count += db.md_branch_pages + db.md_leaf_pages +
1311 db.md_overflow_pages;
1315 while (mdb_cursor_sibling(&mc, 1) == 0);
1318 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1319 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1320 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1326 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1328 return txn->mt_dbxs[dbi].md_cmp(a, b);
1332 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1334 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1337 /** Allocate memory for a page.
1338 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1341 mdb_page_malloc(MDB_txn *txn, unsigned num)
1343 MDB_env *env = txn->mt_env;
1344 MDB_page *ret = env->me_dpages;
1345 size_t psize = env->me_psize, sz = psize, off;
1346 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1347 * For a single page alloc, we init everything after the page header.
1348 * For multi-page, we init the final page; if the caller needed that
1349 * many pages they will be filling in at least up to the last page.
1353 VGMEMP_ALLOC(env, ret, sz);
1354 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1355 env->me_dpages = ret->mp_next;
1358 psize -= off = PAGEHDRSZ;
1363 if ((ret = malloc(sz)) != NULL) {
1364 if (!(env->me_flags & MDB_NOMEMINIT)) {
1365 memset((char *)ret + off, 0, psize);
1368 VGMEMP_ALLOC(env, ret, sz);
1373 /** Free a single page.
1374 * Saves single pages to a list, for future reuse.
1375 * (This is not used for multi-page overflow pages.)
1378 mdb_page_free(MDB_env *env, MDB_page *mp)
1380 mp->mp_next = env->me_dpages;
1381 VGMEMP_FREE(env, mp);
1382 env->me_dpages = mp;
1385 /** Free a dirty page */
1387 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1389 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1390 mdb_page_free(env, dp);
1392 /* large pages just get freed directly */
1393 VGMEMP_FREE(env, dp);
1398 /** Return all dirty pages to dpage list */
1400 mdb_dlist_free(MDB_txn *txn)
1402 MDB_env *env = txn->mt_env;
1403 MDB_ID2L dl = txn->mt_u.dirty_list;
1404 unsigned i, n = dl[0].mid;
1406 for (i = 1; i <= n; i++) {
1407 mdb_dpage_free(env, dl[i].mptr);
1412 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1413 * @param[in] mc A cursor handle for the current operation.
1414 * @param[in] pflags Flags of the pages to update:
1415 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1416 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1417 * @return 0 on success, non-zero on failure.
1420 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1422 enum { Mask = P_SUBP|P_DIRTY|P_KEEP };
1423 MDB_txn *txn = mc->mc_txn;
1429 int rc = MDB_SUCCESS, level;
1431 /* Mark pages seen by cursors */
1432 if (mc->mc_flags & C_UNTRACK)
1433 mc = NULL; /* will find mc in mt_cursors */
1434 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1435 for (; mc; mc=mc->mc_next) {
1436 if (!(mc->mc_flags & C_INITIALIZED))
1438 for (m3 = mc;; m3 = &mx->mx_cursor) {
1440 for (j=0; j<m3->mc_snum; j++) {
1442 if ((mp->mp_flags & Mask) == pflags)
1443 mp->mp_flags ^= P_KEEP;
1445 mx = m3->mc_xcursor;
1446 /* Proceed to mx if it is at a sub-database */
1447 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1449 if (! (mp && (mp->mp_flags & P_LEAF)))
1451 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1452 if (!(leaf->mn_flags & F_SUBDATA))
1461 /* Mark dirty root pages */
1462 for (i=0; i<txn->mt_numdbs; i++) {
1463 if (txn->mt_dbflags[i] & DB_DIRTY) {
1464 pgno_t pgno = txn->mt_dbs[i].md_root;
1465 if (pgno == P_INVALID)
1467 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1469 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1470 dp->mp_flags ^= P_KEEP;
1478 static int mdb_page_flush(MDB_txn *txn, int keep);
1480 /** Spill pages from the dirty list back to disk.
1481 * This is intended to prevent running into #MDB_TXN_FULL situations,
1482 * but note that they may still occur in a few cases:
1483 * 1) our estimate of the txn size could be too small. Currently this
1484 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1485 * 2) child txns may run out of space if their parents dirtied a
1486 * lot of pages and never spilled them. TODO: we probably should do
1487 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1488 * the parent's dirty_room is below a given threshold.
1490 * Otherwise, if not using nested txns, it is expected that apps will
1491 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1492 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1493 * If the txn never references them again, they can be left alone.
1494 * If the txn only reads them, they can be used without any fuss.
1495 * If the txn writes them again, they can be dirtied immediately without
1496 * going thru all of the work of #mdb_page_touch(). Such references are
1497 * handled by #mdb_page_unspill().
1499 * Also note, we never spill DB root pages, nor pages of active cursors,
1500 * because we'll need these back again soon anyway. And in nested txns,
1501 * we can't spill a page in a child txn if it was already spilled in a
1502 * parent txn. That would alter the parent txns' data even though
1503 * the child hasn't committed yet, and we'd have no way to undo it if
1504 * the child aborted.
1506 * @param[in] m0 cursor A cursor handle identifying the transaction and
1507 * database for which we are checking space.
1508 * @param[in] key For a put operation, the key being stored.
1509 * @param[in] data For a put operation, the data being stored.
1510 * @return 0 on success, non-zero on failure.
1513 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1515 MDB_txn *txn = m0->mc_txn;
1517 MDB_ID2L dl = txn->mt_u.dirty_list;
1518 unsigned int i, j, need;
1521 if (m0->mc_flags & C_SUB)
1524 /* Estimate how much space this op will take */
1525 i = m0->mc_db->md_depth;
1526 /* Named DBs also dirty the main DB */
1527 if (m0->mc_dbi > MAIN_DBI)
1528 i += txn->mt_dbs[MAIN_DBI].md_depth;
1529 /* For puts, roughly factor in the key+data size */
1531 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1532 i += i; /* double it for good measure */
1535 if (txn->mt_dirty_room > i)
1538 if (!txn->mt_spill_pgs) {
1539 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1540 if (!txn->mt_spill_pgs)
1543 /* purge deleted slots */
1544 MDB_IDL sl = txn->mt_spill_pgs;
1545 unsigned int num = sl[0];
1547 for (i=1; i<=num; i++) {
1554 /* Preserve pages which may soon be dirtied again */
1555 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1558 /* Less aggressive spill - we originally spilled the entire dirty list,
1559 * with a few exceptions for cursor pages and DB root pages. But this
1560 * turns out to be a lot of wasted effort because in a large txn many
1561 * of those pages will need to be used again. So now we spill only 1/8th
1562 * of the dirty pages. Testing revealed this to be a good tradeoff,
1563 * better than 1/2, 1/4, or 1/10.
1565 if (need < MDB_IDL_UM_MAX / 8)
1566 need = MDB_IDL_UM_MAX / 8;
1568 /* Save the page IDs of all the pages we're flushing */
1569 /* flush from the tail forward, this saves a lot of shifting later on. */
1570 for (i=dl[0].mid; i && need; i--) {
1571 MDB_ID pn = dl[i].mid << 1;
1573 if (dp->mp_flags & P_KEEP)
1575 /* Can't spill twice, make sure it's not already in a parent's
1578 if (txn->mt_parent) {
1580 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1581 if (tx2->mt_spill_pgs) {
1582 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1583 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1584 dp->mp_flags |= P_KEEP;
1592 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1596 mdb_midl_sort(txn->mt_spill_pgs);
1598 /* Flush the spilled part of dirty list */
1599 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1602 /* Reset any dirty pages we kept that page_flush didn't see */
1603 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1606 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1610 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1612 mdb_find_oldest(MDB_txn *txn)
1615 txnid_t mr, oldest = txn->mt_txnid - 1;
1616 if (txn->mt_env->me_txns) {
1617 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1618 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1629 /** Add a page to the txn's dirty list */
1631 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1634 int (*insert)(MDB_ID2L, MDB_ID2 *);
1636 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1637 insert = mdb_mid2l_append;
1639 insert = mdb_mid2l_insert;
1641 mid.mid = mp->mp_pgno;
1643 insert(txn->mt_u.dirty_list, &mid);
1644 txn->mt_dirty_room--;
1647 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1648 * me_pghead and mt_next_pgno.
1650 * If there are free pages available from older transactions, they
1651 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1652 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1653 * and move me_pglast to say which records were consumed. Only this
1654 * function can create me_pghead and move me_pglast/mt_next_pgno.
1655 * @param[in] mc cursor A cursor handle identifying the transaction and
1656 * database for which we are allocating.
1657 * @param[in] num the number of pages to allocate.
1658 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1659 * will always be satisfied by a single contiguous chunk of memory.
1660 * @return 0 on success, non-zero on failure.
1663 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1665 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1666 /* Get at most <Max_retries> more freeDB records once me_pghead
1667 * has enough pages. If not enough, use new pages from the map.
1668 * If <Paranoid> and mc is updating the freeDB, only get new
1669 * records if me_pghead is empty. Then the freelist cannot play
1670 * catch-up with itself by growing while trying to save it.
1672 enum { Paranoid = 1, Max_retries = 500 };
1674 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1676 int rc, retry = Max_retries;
1677 MDB_txn *txn = mc->mc_txn;
1678 MDB_env *env = txn->mt_env;
1679 pgno_t pgno, *mop = env->me_pghead;
1680 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1682 txnid_t oldest = 0, last;
1688 /* If our dirty list is already full, we can't do anything */
1689 if (txn->mt_dirty_room == 0)
1690 return MDB_TXN_FULL;
1692 for (op = MDB_FIRST;; op = MDB_NEXT) {
1695 pgno_t *idl, old_id, new_id;
1697 /* Seek a big enough contiguous page range. Prefer
1698 * pages at the tail, just truncating the list.
1704 if (mop[i-n2] == pgno+n2)
1707 if (Max_retries < INT_MAX && --retry < 0)
1711 if (op == MDB_FIRST) { /* 1st iteration */
1712 /* Prepare to fetch more and coalesce */
1713 oldest = mdb_find_oldest(txn);
1714 last = env->me_pglast;
1715 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1718 key.mv_data = &last; /* will look up last+1 */
1719 key.mv_size = sizeof(last);
1721 if (Paranoid && mc->mc_dbi == FREE_DBI)
1724 if (Paranoid && retry < 0 && mop_len)
1728 /* Do not fetch more if the record will be too recent */
1731 rc = mdb_cursor_get(&m2, &key, NULL, op);
1733 if (rc == MDB_NOTFOUND)
1737 last = *(txnid_t*)key.mv_data;
1740 np = m2.mc_pg[m2.mc_top];
1741 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1742 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1745 idl = (MDB_ID *) data.mv_data;
1748 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1751 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1753 mop = env->me_pghead;
1755 env->me_pglast = last;
1757 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1758 last, txn->mt_dbs[FREE_DBI].md_root, i));
1760 DPRINTF(("IDL %"Z"u", idl[k]));
1762 /* Merge in descending sorted order */
1765 mop[0] = (pgno_t)-1;
1769 for (; old_id < new_id; old_id = mop[--j])
1776 /* Use new pages from the map when nothing suitable in the freeDB */
1778 pgno = txn->mt_next_pgno;
1779 if (pgno + num >= env->me_maxpg) {
1780 DPUTS("DB size maxed out");
1781 return MDB_MAP_FULL;
1785 if (env->me_flags & MDB_WRITEMAP) {
1786 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1788 if (!(np = mdb_page_malloc(txn, num)))
1792 mop[0] = mop_len -= num;
1793 /* Move any stragglers down */
1794 for (j = i-num; j < mop_len; )
1795 mop[++j] = mop[++i];
1797 txn->mt_next_pgno = pgno + num;
1800 mdb_page_dirty(txn, np);
1806 /** Copy the used portions of a non-overflow page.
1807 * @param[in] dst page to copy into
1808 * @param[in] src page to copy from
1809 * @param[in] psize size of a page
1812 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1814 enum { Align = sizeof(pgno_t) };
1815 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1817 /* If page isn't full, just copy the used portion. Adjust
1818 * alignment so memcpy may copy words instead of bytes.
1820 if ((unused &= -Align) && !IS_LEAF2(src)) {
1822 memcpy(dst, src, (lower + (Align-1)) & -Align);
1823 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1826 memcpy(dst, src, psize - unused);
1830 /** Pull a page off the txn's spill list, if present.
1831 * If a page being referenced was spilled to disk in this txn, bring
1832 * it back and make it dirty/writable again.
1833 * @param[in] txn the transaction handle.
1834 * @param[in] mp the page being referenced.
1835 * @param[out] ret the writable page, if any. ret is unchanged if
1836 * mp wasn't spilled.
1839 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
1841 MDB_env *env = txn->mt_env;
1844 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
1846 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
1847 if (!tx2->mt_spill_pgs)
1849 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
1850 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
1853 if (txn->mt_dirty_room == 0)
1854 return MDB_TXN_FULL;
1855 if (IS_OVERFLOW(mp))
1859 if (env->me_flags & MDB_WRITEMAP) {
1862 np = mdb_page_malloc(txn, num);
1866 memcpy(np, mp, num * env->me_psize);
1868 mdb_page_copy(np, mp, env->me_psize);
1871 /* If in current txn, this page is no longer spilled.
1872 * If it happens to be the last page, truncate the spill list.
1873 * Otherwise mark it as deleted by setting the LSB.
1875 if (x == txn->mt_spill_pgs[0])
1876 txn->mt_spill_pgs[0]--;
1878 txn->mt_spill_pgs[x] |= 1;
1879 } /* otherwise, if belonging to a parent txn, the
1880 * page remains spilled until child commits
1883 mdb_page_dirty(txn, np);
1884 np->mp_flags |= P_DIRTY;
1892 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1893 * @param[in] mc cursor pointing to the page to be touched
1894 * @return 0 on success, non-zero on failure.
1897 mdb_page_touch(MDB_cursor *mc)
1899 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1900 MDB_txn *txn = mc->mc_txn;
1901 MDB_cursor *m2, *m3;
1905 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1906 if (txn->mt_flags & MDB_TXN_SPILLS) {
1908 rc = mdb_page_unspill(txn, mp, &np);
1914 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1915 (rc = mdb_page_alloc(mc, 1, &np)))
1918 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
1919 mp->mp_pgno, pgno));
1920 assert(mp->mp_pgno != pgno);
1921 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1922 /* Update the parent page, if any, to point to the new page */
1924 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1925 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1926 SETPGNO(node, pgno);
1928 mc->mc_db->md_root = pgno;
1930 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1931 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1933 /* If txn has a parent, make sure the page is in our
1937 unsigned x = mdb_mid2l_search(dl, pgno);
1938 if (x <= dl[0].mid && dl[x].mid == pgno) {
1939 if (mp != dl[x].mptr) { /* bad cursor? */
1940 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1941 return MDB_CORRUPTED;
1946 assert(dl[0].mid < MDB_IDL_UM_MAX);
1948 np = mdb_page_malloc(txn, 1);
1953 mdb_mid2l_insert(dl, &mid);
1958 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1960 np->mp_flags |= P_DIRTY;
1963 /* Adjust cursors pointing to mp */
1964 mc->mc_pg[mc->mc_top] = np;
1965 m2 = txn->mt_cursors[mc->mc_dbi];
1966 if (mc->mc_flags & C_SUB) {
1967 for (; m2; m2=m2->mc_next) {
1968 m3 = &m2->mc_xcursor->mx_cursor;
1969 if (m3->mc_snum < mc->mc_snum) continue;
1970 if (m3->mc_pg[mc->mc_top] == mp)
1971 m3->mc_pg[mc->mc_top] = np;
1974 for (; m2; m2=m2->mc_next) {
1975 if (m2->mc_snum < mc->mc_snum) continue;
1976 if (m2->mc_pg[mc->mc_top] == mp) {
1977 m2->mc_pg[mc->mc_top] = np;
1978 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1979 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1981 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1982 if (!(leaf->mn_flags & F_SUBDATA))
1983 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1992 mdb_env_sync(MDB_env *env, int force)
1995 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1996 if (env->me_flags & MDB_WRITEMAP) {
1997 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1998 ? MS_ASYNC : MS_SYNC;
1999 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2002 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2006 if (MDB_FDATASYNC(env->me_fd))
2013 /** Back up parent txn's cursors, then grab the originals for tracking */
2015 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2017 MDB_cursor *mc, *bk;
2022 for (i = src->mt_numdbs; --i >= 0; ) {
2023 if ((mc = src->mt_cursors[i]) != NULL) {
2024 size = sizeof(MDB_cursor);
2026 size += sizeof(MDB_xcursor);
2027 for (; mc; mc = bk->mc_next) {
2033 mc->mc_db = &dst->mt_dbs[i];
2034 /* Kill pointers into src - and dst to reduce abuse: The
2035 * user may not use mc until dst ends. Otherwise we'd...
2037 mc->mc_txn = NULL; /* ...set this to dst */
2038 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2039 if ((mx = mc->mc_xcursor) != NULL) {
2040 *(MDB_xcursor *)(bk+1) = *mx;
2041 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2043 mc->mc_next = dst->mt_cursors[i];
2044 dst->mt_cursors[i] = mc;
2051 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2052 * @param[in] txn the transaction handle.
2053 * @param[in] merge true to keep changes to parent cursors, false to revert.
2054 * @return 0 on success, non-zero on failure.
2057 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2059 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2063 for (i = txn->mt_numdbs; --i >= 0; ) {
2064 for (mc = cursors[i]; mc; mc = next) {
2066 if ((bk = mc->mc_backup) != NULL) {
2068 /* Commit changes to parent txn */
2069 mc->mc_next = bk->mc_next;
2070 mc->mc_backup = bk->mc_backup;
2071 mc->mc_txn = bk->mc_txn;
2072 mc->mc_db = bk->mc_db;
2073 mc->mc_dbflag = bk->mc_dbflag;
2074 if ((mx = mc->mc_xcursor) != NULL)
2075 mx->mx_cursor.mc_txn = bk->mc_txn;
2077 /* Abort nested txn */
2079 if ((mx = mc->mc_xcursor) != NULL)
2080 *mx = *(MDB_xcursor *)(bk+1);
2084 /* Only malloced cursors are permanently tracked. */
2092 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2095 mdb_txn_reset0(MDB_txn *txn, const char *act);
2097 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2103 Pidset = F_SETLK, Pidcheck = F_GETLK
2107 /** Set or check a pid lock. Set returns 0 on success.
2108 * Check returns 0 if the process is certainly dead, nonzero if it may
2109 * be alive (the lock exists or an error happened so we do not know).
2111 * On Windows Pidset is a no-op, we merely check for the existence
2112 * of the process with the given pid. On POSIX we use a single byte
2113 * lock on the lockfile, set at an offset equal to the pid.
2116 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2118 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2121 if (op == Pidcheck) {
2122 h = OpenProcess(env->me_pidquery, FALSE, pid);
2123 /* No documented "no such process" code, but other program use this: */
2125 return ErrCode() != ERROR_INVALID_PARAMETER;
2126 /* A process exists until all handles to it close. Has it exited? */
2127 ret = WaitForSingleObject(h, 0) != 0;
2134 struct flock lock_info;
2135 memset(&lock_info, 0, sizeof(lock_info));
2136 lock_info.l_type = F_WRLCK;
2137 lock_info.l_whence = SEEK_SET;
2138 lock_info.l_start = pid;
2139 lock_info.l_len = 1;
2140 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2141 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2143 } else if ((rc = ErrCode()) == EINTR) {
2151 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2152 * @param[in] txn the transaction handle to initialize
2153 * @return 0 on success, non-zero on failure.
2156 mdb_txn_renew0(MDB_txn *txn)
2158 MDB_env *env = txn->mt_env;
2159 MDB_txninfo *ti = env->me_txns;
2163 int rc, new_notls = 0;
2166 txn->mt_numdbs = env->me_numdbs;
2167 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2169 if (txn->mt_flags & MDB_TXN_RDONLY) {
2171 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2172 txn->mt_txnid = meta->mm_txnid;
2173 txn->mt_u.reader = NULL;
2175 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2176 pthread_getspecific(env->me_txkey);
2178 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2179 return MDB_BAD_RSLOT;
2181 MDB_PID_T pid = env->me_pid;
2182 pthread_t tid = pthread_self();
2184 if (!(env->me_flags & MDB_LIVE_READER)) {
2185 rc = mdb_reader_pid(env, Pidset, pid);
2187 UNLOCK_MUTEX_R(env);
2190 env->me_flags |= MDB_LIVE_READER;
2194 nr = ti->mti_numreaders;
2195 for (i=0; i<nr; i++)
2196 if (ti->mti_readers[i].mr_pid == 0)
2198 if (i == env->me_maxreaders) {
2199 UNLOCK_MUTEX_R(env);
2200 return MDB_READERS_FULL;
2202 ti->mti_readers[i].mr_pid = pid;
2203 ti->mti_readers[i].mr_tid = tid;
2205 ti->mti_numreaders = ++nr;
2206 /* Save numreaders for un-mutexed mdb_env_close() */
2207 env->me_numreaders = nr;
2208 UNLOCK_MUTEX_R(env);
2210 r = &ti->mti_readers[i];
2211 new_notls = (env->me_flags & MDB_NOTLS);
2212 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2217 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2218 txn->mt_u.reader = r;
2219 meta = env->me_metas[txn->mt_txnid & 1];
2225 txn->mt_txnid = ti->mti_txnid;
2226 meta = env->me_metas[txn->mt_txnid & 1];
2228 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2229 txn->mt_txnid = meta->mm_txnid;
2233 if (txn->mt_txnid == mdb_debug_start)
2236 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2237 txn->mt_u.dirty_list = env->me_dirty_list;
2238 txn->mt_u.dirty_list[0].mid = 0;
2239 txn->mt_free_pgs = env->me_free_pgs;
2240 txn->mt_free_pgs[0] = 0;
2241 txn->mt_spill_pgs = NULL;
2245 /* Copy the DB info and flags */
2246 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2248 /* Moved to here to avoid a data race in read TXNs */
2249 txn->mt_next_pgno = meta->mm_last_pg+1;
2251 for (i=2; i<txn->mt_numdbs; i++) {
2252 x = env->me_dbflags[i];
2253 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2254 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2256 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2258 if (env->me_maxpg < txn->mt_next_pgno) {
2259 mdb_txn_reset0(txn, "renew0-mapfail");
2261 txn->mt_u.reader->mr_pid = 0;
2262 txn->mt_u.reader = NULL;
2264 return MDB_MAP_RESIZED;
2271 mdb_txn_renew(MDB_txn *txn)
2275 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2278 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2279 DPUTS("environment had fatal error, must shutdown!");
2283 rc = mdb_txn_renew0(txn);
2284 if (rc == MDB_SUCCESS) {
2285 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2286 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2287 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2293 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2297 int rc, size, tsize = sizeof(MDB_txn);
2299 if (env->me_flags & MDB_FATAL_ERROR) {
2300 DPUTS("environment had fatal error, must shutdown!");
2303 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2306 /* Nested transactions: Max 1 child, write txns only, no writemap */
2307 if (parent->mt_child ||
2308 (flags & MDB_RDONLY) ||
2309 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2310 (env->me_flags & MDB_WRITEMAP))
2312 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2314 tsize = sizeof(MDB_ntxn);
2316 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2317 if (!(flags & MDB_RDONLY))
2318 size += env->me_maxdbs * sizeof(MDB_cursor *);
2320 if ((txn = calloc(1, size)) == NULL) {
2321 DPRINTF(("calloc: %s", strerror(ErrCode())));
2324 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2325 if (flags & MDB_RDONLY) {
2326 txn->mt_flags |= MDB_TXN_RDONLY;
2327 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2329 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2330 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2336 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2337 if (!txn->mt_u.dirty_list ||
2338 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2340 free(txn->mt_u.dirty_list);
2344 txn->mt_txnid = parent->mt_txnid;
2345 txn->mt_dirty_room = parent->mt_dirty_room;
2346 txn->mt_u.dirty_list[0].mid = 0;
2347 txn->mt_spill_pgs = NULL;
2348 txn->mt_next_pgno = parent->mt_next_pgno;
2349 parent->mt_child = txn;
2350 txn->mt_parent = parent;
2351 txn->mt_numdbs = parent->mt_numdbs;
2352 txn->mt_flags = parent->mt_flags;
2353 txn->mt_dbxs = parent->mt_dbxs;
2354 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2355 /* Copy parent's mt_dbflags, but clear DB_NEW */
2356 for (i=0; i<txn->mt_numdbs; i++)
2357 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2359 ntxn = (MDB_ntxn *)txn;
2360 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2361 if (env->me_pghead) {
2362 size = MDB_IDL_SIZEOF(env->me_pghead);
2363 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2365 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2370 rc = mdb_cursor_shadow(parent, txn);
2372 mdb_txn_reset0(txn, "beginchild-fail");
2374 rc = mdb_txn_renew0(txn);
2380 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2381 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2382 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2389 mdb_txn_env(MDB_txn *txn)
2391 if(!txn) return NULL;
2395 /** Export or close DBI handles opened in this txn. */
2397 mdb_dbis_update(MDB_txn *txn, int keep)
2400 MDB_dbi n = txn->mt_numdbs;
2401 MDB_env *env = txn->mt_env;
2402 unsigned char *tdbflags = txn->mt_dbflags;
2404 for (i = n; --i >= 2;) {
2405 if (tdbflags[i] & DB_NEW) {
2407 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2409 char *ptr = env->me_dbxs[i].md_name.mv_data;
2410 env->me_dbxs[i].md_name.mv_data = NULL;
2411 env->me_dbxs[i].md_name.mv_size = 0;
2412 env->me_dbflags[i] = 0;
2417 if (keep && env->me_numdbs < n)
2421 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2422 * May be called twice for readonly txns: First reset it, then abort.
2423 * @param[in] txn the transaction handle to reset
2424 * @param[in] act why the transaction is being reset
2427 mdb_txn_reset0(MDB_txn *txn, const char *act)
2429 MDB_env *env = txn->mt_env;
2431 /* Close any DBI handles opened in this txn */
2432 mdb_dbis_update(txn, 0);
2434 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2435 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2436 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2438 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2439 if (txn->mt_u.reader) {
2440 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2441 if (!(env->me_flags & MDB_NOTLS))
2442 txn->mt_u.reader = NULL; /* txn does not own reader */
2444 txn->mt_numdbs = 0; /* close nothing if called again */
2445 txn->mt_dbxs = NULL; /* mark txn as reset */
2447 mdb_cursors_close(txn, 0);
2449 if (!(env->me_flags & MDB_WRITEMAP)) {
2450 mdb_dlist_free(txn);
2452 mdb_midl_free(env->me_pghead);
2454 if (txn->mt_parent) {
2455 txn->mt_parent->mt_child = NULL;
2456 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2457 mdb_midl_free(txn->mt_free_pgs);
2458 mdb_midl_free(txn->mt_spill_pgs);
2459 free(txn->mt_u.dirty_list);
2463 if (mdb_midl_shrink(&txn->mt_free_pgs))
2464 env->me_free_pgs = txn->mt_free_pgs;
2465 env->me_pghead = NULL;
2469 /* The writer mutex was locked in mdb_txn_begin. */
2471 UNLOCK_MUTEX_W(env);
2476 mdb_txn_reset(MDB_txn *txn)
2481 /* This call is only valid for read-only txns */
2482 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2485 mdb_txn_reset0(txn, "reset");
2489 mdb_txn_abort(MDB_txn *txn)
2495 mdb_txn_abort(txn->mt_child);
2497 mdb_txn_reset0(txn, "abort");
2498 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2499 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2500 txn->mt_u.reader->mr_pid = 0;
2505 /** Save the freelist as of this transaction to the freeDB.
2506 * This changes the freelist. Keep trying until it stabilizes.
2509 mdb_freelist_save(MDB_txn *txn)
2511 /* env->me_pghead[] can grow and shrink during this call.
2512 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2513 * Page numbers cannot disappear from txn->mt_free_pgs[].
2516 MDB_env *env = txn->mt_env;
2517 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2518 txnid_t pglast = 0, head_id = 0;
2519 pgno_t freecnt = 0, *free_pgs, *mop;
2520 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2522 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2524 if (env->me_pghead) {
2525 /* Make sure first page of freeDB is touched and on freelist */
2526 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2527 if (rc && rc != MDB_NOTFOUND)
2531 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2532 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2533 ? SSIZE_MAX : maxfree_1pg;
2536 /* Come back here after each Put() in case freelist changed */
2541 /* If using records from freeDB which we have not yet
2542 * deleted, delete them and any we reserved for me_pghead.
2544 while (pglast < env->me_pglast) {
2545 rc = mdb_cursor_first(&mc, &key, NULL);
2548 pglast = head_id = *(txnid_t *)key.mv_data;
2549 total_room = head_room = 0;
2550 assert(pglast <= env->me_pglast);
2551 rc = mdb_cursor_del(&mc, 0);
2556 /* Save the IDL of pages freed by this txn, to a single record */
2557 if (freecnt < txn->mt_free_pgs[0]) {
2559 /* Make sure last page of freeDB is touched and on freelist */
2560 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2561 if (rc && rc != MDB_NOTFOUND)
2564 free_pgs = txn->mt_free_pgs;
2565 /* Write to last page of freeDB */
2566 key.mv_size = sizeof(txn->mt_txnid);
2567 key.mv_data = &txn->mt_txnid;
2569 freecnt = free_pgs[0];
2570 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2571 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2574 /* Retry if mt_free_pgs[] grew during the Put() */
2575 free_pgs = txn->mt_free_pgs;
2576 } while (freecnt < free_pgs[0]);
2577 mdb_midl_sort(free_pgs);
2578 memcpy(data.mv_data, free_pgs, data.mv_size);
2581 unsigned int i = free_pgs[0];
2582 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2583 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2585 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2591 mop = env->me_pghead;
2592 mop_len = mop ? mop[0] : 0;
2594 /* Reserve records for me_pghead[]. Split it if multi-page,
2595 * to avoid searching freeDB for a page range. Use keys in
2596 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2598 if (total_room >= mop_len) {
2599 if (total_room == mop_len || --more < 0)
2601 } else if (head_room >= maxfree_1pg && head_id > 1) {
2602 /* Keep current record (overflow page), add a new one */
2606 /* (Re)write {key = head_id, IDL length = head_room} */
2607 total_room -= head_room;
2608 head_room = mop_len - total_room;
2609 if (head_room > maxfree_1pg && head_id > 1) {
2610 /* Overflow multi-page for part of me_pghead */
2611 head_room /= head_id; /* amortize page sizes */
2612 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2613 } else if (head_room < 0) {
2614 /* Rare case, not bothering to delete this record */
2617 key.mv_size = sizeof(head_id);
2618 key.mv_data = &head_id;
2619 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2620 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2623 /* IDL is initially empty, zero out at least the length */
2624 pgs = (pgno_t *)data.mv_data;
2625 j = head_room > clean_limit ? head_room : 0;
2629 total_room += head_room;
2632 /* Fill in the reserved me_pghead records */
2638 rc = mdb_cursor_first(&mc, &key, &data);
2639 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2640 unsigned flags = MDB_CURRENT;
2641 txnid_t id = *(txnid_t *)key.mv_data;
2642 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2645 assert(len >= 0 && id <= env->me_pglast);
2647 if (len > mop_len) {
2649 data.mv_size = (len + 1) * sizeof(MDB_ID);
2652 data.mv_data = mop -= len;
2655 rc = mdb_cursor_put(&mc, &key, &data, flags);
2657 if (rc || !(mop_len -= len))
2664 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2665 * @param[in] txn the transaction that's being committed
2666 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2667 * @return 0 on success, non-zero on failure.
2670 mdb_page_flush(MDB_txn *txn, int keep)
2672 MDB_env *env = txn->mt_env;
2673 MDB_ID2L dl = txn->mt_u.dirty_list;
2674 unsigned psize = env->me_psize, j;
2675 int i, pagecount = dl[0].mid, rc;
2676 size_t size = 0, pos = 0;
2678 MDB_page *dp = NULL;
2682 struct iovec iov[MDB_COMMIT_PAGES];
2683 ssize_t wpos = 0, wsize = 0, wres;
2684 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2690 if (env->me_flags & MDB_WRITEMAP) {
2691 /* Clear dirty flags */
2692 while (++i <= pagecount) {
2694 /* Don't flush this page yet */
2695 if (dp->mp_flags & P_KEEP) {
2696 dp->mp_flags ^= P_KEEP;
2700 dp->mp_flags &= ~P_DIRTY;
2705 /* Write the pages */
2707 if (++i <= pagecount) {
2709 /* Don't flush this page yet */
2710 if (dp->mp_flags & P_KEEP) {
2711 dp->mp_flags ^= P_KEEP;
2716 /* clear dirty flag */
2717 dp->mp_flags &= ~P_DIRTY;
2720 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2725 /* Windows actually supports scatter/gather I/O, but only on
2726 * unbuffered file handles. Since we're relying on the OS page
2727 * cache for all our data, that's self-defeating. So we just
2728 * write pages one at a time. We use the ov structure to set
2729 * the write offset, to at least save the overhead of a Seek
2732 DPRINTF(("committing page %"Z"u", pgno));
2733 memset(&ov, 0, sizeof(ov));
2734 ov.Offset = pos & 0xffffffff;
2735 ov.OffsetHigh = pos >> 16 >> 16;
2736 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2738 DPRINTF(("WriteFile: %d", rc));
2742 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2743 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
2745 /* Write previous page(s) */
2746 #ifdef MDB_USE_PWRITEV
2747 wres = pwritev(env->me_fd, iov, n, wpos);
2750 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
2752 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
2754 DPRINTF(("lseek: %s", strerror(rc)));
2757 wres = writev(env->me_fd, iov, n);
2760 if (wres != wsize) {
2763 DPRINTF(("Write error: %s", strerror(rc)));
2765 rc = EIO; /* TODO: Use which error code? */
2766 DPUTS("short write, filesystem full?");
2777 DPRINTF(("committing page %"Z"u", pgno));
2778 next_pos = pos + size;
2779 iov[n].iov_len = size;
2780 iov[n].iov_base = (char *)dp;
2786 for (i = keep; ++i <= pagecount; ) {
2788 /* This is a page we skipped above */
2791 dl[j].mid = dp->mp_pgno;
2794 mdb_dpage_free(env, dp);
2799 txn->mt_dirty_room += i - j;
2805 mdb_txn_commit(MDB_txn *txn)
2811 assert(txn != NULL);
2812 assert(txn->mt_env != NULL);
2814 if (txn->mt_child) {
2815 rc = mdb_txn_commit(txn->mt_child);
2816 txn->mt_child = NULL;
2823 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2824 mdb_dbis_update(txn, 1);
2825 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2830 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2831 DPUTS("error flag is set, can't commit");
2833 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2838 if (txn->mt_parent) {
2839 MDB_txn *parent = txn->mt_parent;
2842 unsigned x, y, len, ps_len;
2844 /* Append our free list to parent's */
2845 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2848 mdb_midl_free(txn->mt_free_pgs);
2849 /* Failures after this must either undo the changes
2850 * to the parent or set MDB_TXN_ERROR in the parent.
2853 parent->mt_next_pgno = txn->mt_next_pgno;
2854 parent->mt_flags = txn->mt_flags;
2856 /* Merge our cursors into parent's and close them */
2857 mdb_cursors_close(txn, 1);
2859 /* Update parent's DB table. */
2860 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2861 parent->mt_numdbs = txn->mt_numdbs;
2862 parent->mt_dbflags[0] = txn->mt_dbflags[0];
2863 parent->mt_dbflags[1] = txn->mt_dbflags[1];
2864 for (i=2; i<txn->mt_numdbs; i++) {
2865 /* preserve parent's DB_NEW status */
2866 x = parent->mt_dbflags[i] & DB_NEW;
2867 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2870 dst = parent->mt_u.dirty_list;
2871 src = txn->mt_u.dirty_list;
2872 /* Remove anything in our dirty list from parent's spill list */
2873 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
2875 pspill[0] = (pgno_t)-1;
2876 /* Mark our dirty pages as deleted in parent spill list */
2877 for (i=0, len=src[0].mid; ++i <= len; ) {
2878 MDB_ID pn = src[i].mid << 1;
2879 while (pn > pspill[x])
2881 if (pn == pspill[x]) {
2886 /* Squash deleted pagenums if we deleted any */
2887 for (x=y; ++x <= ps_len; )
2888 if (!(pspill[x] & 1))
2889 pspill[++y] = pspill[x];
2893 /* Find len = length of merging our dirty list with parent's */
2895 dst[0].mid = 0; /* simplify loops */
2896 if (parent->mt_parent) {
2897 len = x + src[0].mid;
2898 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2899 for (i = x; y && i; y--) {
2900 pgno_t yp = src[y].mid;
2901 while (yp < dst[i].mid)
2903 if (yp == dst[i].mid) {
2908 } else { /* Simplify the above for single-ancestor case */
2909 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2911 /* Merge our dirty list with parent's */
2913 for (i = len; y; dst[i--] = src[y--]) {
2914 pgno_t yp = src[y].mid;
2915 while (yp < dst[x].mid)
2916 dst[i--] = dst[x--];
2917 if (yp == dst[x].mid)
2918 free(dst[x--].mptr);
2922 free(txn->mt_u.dirty_list);
2923 parent->mt_dirty_room = txn->mt_dirty_room;
2924 if (txn->mt_spill_pgs) {
2925 if (parent->mt_spill_pgs) {
2926 /* TODO: Prevent failure here, so parent does not fail */
2927 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
2929 parent->mt_flags |= MDB_TXN_ERROR;
2930 mdb_midl_free(txn->mt_spill_pgs);
2931 mdb_midl_sort(parent->mt_spill_pgs);
2933 parent->mt_spill_pgs = txn->mt_spill_pgs;
2937 parent->mt_child = NULL;
2938 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2943 if (txn != env->me_txn) {
2944 DPUTS("attempt to commit unknown transaction");
2949 mdb_cursors_close(txn, 0);
2951 if (!txn->mt_u.dirty_list[0].mid &&
2952 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
2955 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
2956 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
2958 /* Update DB root pointers */
2959 if (txn->mt_numdbs > 2) {
2963 data.mv_size = sizeof(MDB_db);
2965 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2966 for (i = 2; i < txn->mt_numdbs; i++) {
2967 if (txn->mt_dbflags[i] & DB_DIRTY) {
2968 data.mv_data = &txn->mt_dbs[i];
2969 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2976 rc = mdb_freelist_save(txn);
2980 mdb_midl_free(env->me_pghead);
2981 env->me_pghead = NULL;
2982 if (mdb_midl_shrink(&txn->mt_free_pgs))
2983 env->me_free_pgs = txn->mt_free_pgs;
2989 if ((rc = mdb_page_flush(txn, 0)) ||
2990 (rc = mdb_env_sync(env, 0)) ||
2991 (rc = mdb_env_write_meta(txn)))
2997 mdb_dbis_update(txn, 1);
3000 UNLOCK_MUTEX_W(env);
3010 /** Read the environment parameters of a DB environment before
3011 * mapping it into memory.
3012 * @param[in] env the environment handle
3013 * @param[out] meta address of where to store the meta information
3014 * @return 0 on success, non-zero on failure.
3017 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3023 enum { Size = sizeof(pbuf) };
3025 /* We don't know the page size yet, so use a minimum value.
3026 * Read both meta pages so we can use the latest one.
3029 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3033 memset(&ov, 0, sizeof(ov));
3035 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3036 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3039 rc = pread(env->me_fd, &pbuf, Size, off);
3042 if (rc == 0 && off == 0)
3044 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3045 DPRINTF(("read: %s", mdb_strerror(rc)));
3049 p = (MDB_page *)&pbuf;
3051 if (!F_ISSET(p->mp_flags, P_META)) {
3052 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3057 if (m->mm_magic != MDB_MAGIC) {
3058 DPUTS("meta has invalid magic");
3062 if (m->mm_version != MDB_DATA_VERSION) {
3063 DPRINTF(("database is version %u, expected version %u",
3064 m->mm_version, MDB_DATA_VERSION));
3065 return MDB_VERSION_MISMATCH;
3068 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3074 /** Write the environment parameters of a freshly created DB environment.
3075 * @param[in] env the environment handle
3076 * @param[out] meta address of where to store the meta information
3077 * @return 0 on success, non-zero on failure.
3080 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3088 memset(&ov, 0, sizeof(ov));
3089 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3091 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3094 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3095 len = pwrite(fd, ptr, size, pos); \
3096 rc = (len >= 0); } while(0)
3099 DPUTS("writing new meta page");
3101 psize = env->me_psize;
3103 meta->mm_magic = MDB_MAGIC;
3104 meta->mm_version = MDB_DATA_VERSION;
3105 meta->mm_mapsize = env->me_mapsize;
3106 meta->mm_psize = psize;
3107 meta->mm_last_pg = 1;
3108 meta->mm_flags = env->me_flags & 0xffff;
3109 meta->mm_flags |= MDB_INTEGERKEY;
3110 meta->mm_dbs[0].md_root = P_INVALID;
3111 meta->mm_dbs[1].md_root = P_INVALID;
3113 p = calloc(2, psize);
3115 p->mp_flags = P_META;
3116 *(MDB_meta *)METADATA(p) = *meta;
3118 q = (MDB_page *)((char *)p + psize);
3120 q->mp_flags = P_META;
3121 *(MDB_meta *)METADATA(q) = *meta;
3123 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3126 else if ((unsigned) len == psize * 2)
3134 /** Update the environment info to commit a transaction.
3135 * @param[in] txn the transaction that's being committed
3136 * @return 0 on success, non-zero on failure.
3139 mdb_env_write_meta(MDB_txn *txn)
3142 MDB_meta meta, metab, *mp;
3144 int rc, len, toggle;
3153 assert(txn != NULL);
3154 assert(txn->mt_env != NULL);
3156 toggle = txn->mt_txnid & 1;
3157 DPRINTF(("writing meta page %d for root page %"Z"u",
3158 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3161 mp = env->me_metas[toggle];
3163 if (env->me_flags & MDB_WRITEMAP) {
3164 /* Persist any increases of mapsize config */
3165 if (env->me_mapsize > mp->mm_mapsize)
3166 mp->mm_mapsize = env->me_mapsize;
3167 mp->mm_dbs[0] = txn->mt_dbs[0];
3168 mp->mm_dbs[1] = txn->mt_dbs[1];
3169 mp->mm_last_pg = txn->mt_next_pgno - 1;
3170 mp->mm_txnid = txn->mt_txnid;
3171 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3172 unsigned meta_size = env->me_psize;
3173 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3176 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3177 if (meta_size < env->me_os_psize)
3178 meta_size += meta_size;
3183 if (MDB_MSYNC(ptr, meta_size, rc)) {
3190 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3191 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3193 ptr = (char *)&meta;
3194 if (env->me_mapsize > mp->mm_mapsize) {
3195 /* Persist any increases of mapsize config */
3196 meta.mm_mapsize = env->me_mapsize;
3197 off = offsetof(MDB_meta, mm_mapsize);
3199 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3201 len = sizeof(MDB_meta) - off;
3204 meta.mm_dbs[0] = txn->mt_dbs[0];
3205 meta.mm_dbs[1] = txn->mt_dbs[1];
3206 meta.mm_last_pg = txn->mt_next_pgno - 1;
3207 meta.mm_txnid = txn->mt_txnid;
3210 off += env->me_psize;
3213 /* Write to the SYNC fd */
3214 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3215 env->me_fd : env->me_mfd;
3218 memset(&ov, 0, sizeof(ov));
3220 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3224 rc = pwrite(mfd, ptr, len, off);
3227 rc = rc < 0 ? ErrCode() : EIO;
3228 DPUTS("write failed, disk error?");
3229 /* On a failure, the pagecache still contains the new data.
3230 * Write some old data back, to prevent it from being used.
3231 * Use the non-SYNC fd; we know it will fail anyway.
3233 meta.mm_last_pg = metab.mm_last_pg;
3234 meta.mm_txnid = metab.mm_txnid;
3236 memset(&ov, 0, sizeof(ov));
3238 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3240 r2 = pwrite(env->me_fd, ptr, len, off);
3241 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3244 env->me_flags |= MDB_FATAL_ERROR;
3248 /* Memory ordering issues are irrelevant; since the entire writer
3249 * is wrapped by wmutex, all of these changes will become visible
3250 * after the wmutex is unlocked. Since the DB is multi-version,
3251 * readers will get consistent data regardless of how fresh or
3252 * how stale their view of these values is.
3255 env->me_txns->mti_txnid = txn->mt_txnid;
3260 /** Check both meta pages to see which one is newer.
3261 * @param[in] env the environment handle
3262 * @return meta toggle (0 or 1).
3265 mdb_env_pick_meta(const MDB_env *env)
3267 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3271 mdb_env_create(MDB_env **env)
3275 e = calloc(1, sizeof(MDB_env));
3279 e->me_maxreaders = DEFAULT_READERS;
3280 e->me_maxdbs = e->me_numdbs = 2;
3281 e->me_fd = INVALID_HANDLE_VALUE;
3282 e->me_lfd = INVALID_HANDLE_VALUE;
3283 e->me_mfd = INVALID_HANDLE_VALUE;
3284 #ifdef MDB_USE_POSIX_SEM
3285 e->me_rmutex = SEM_FAILED;
3286 e->me_wmutex = SEM_FAILED;
3288 e->me_pid = getpid();
3289 GET_PAGESIZE(e->me_os_psize);
3290 VGMEMP_CREATE(e,0,0);
3296 mdb_env_map(MDB_env *env, void *addr, int newsize)
3299 unsigned int flags = env->me_flags;
3303 LONG sizelo, sizehi;
3304 sizelo = env->me_mapsize & 0xffffffff;
3305 sizehi = env->me_mapsize >> 16 >> 16; /* only needed on Win64 */
3307 /* Windows won't create mappings for zero length files.
3308 * Just allocate the maxsize right now.
3311 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3312 || !SetEndOfFile(env->me_fd)
3313 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3316 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3317 PAGE_READWRITE : PAGE_READONLY,
3318 sizehi, sizelo, NULL);
3321 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3322 FILE_MAP_WRITE : FILE_MAP_READ,
3323 0, 0, env->me_mapsize, addr);
3324 rc = env->me_map ? 0 : ErrCode();
3329 int prot = PROT_READ;
3330 if (flags & MDB_WRITEMAP) {
3332 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3335 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3337 if (env->me_map == MAP_FAILED) {
3342 if (flags & MDB_NORDAHEAD) {
3343 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3345 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3347 #ifdef POSIX_MADV_RANDOM
3348 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3349 #endif /* POSIX_MADV_RANDOM */
3350 #endif /* MADV_RANDOM */
3354 /* Can happen because the address argument to mmap() is just a
3355 * hint. mmap() can pick another, e.g. if the range is in use.
3356 * The MAP_FIXED flag would prevent that, but then mmap could
3357 * instead unmap existing pages to make room for the new map.
3359 if (addr && env->me_map != addr)
3360 return EBUSY; /* TODO: Make a new MDB_* error code? */
3362 p = (MDB_page *)env->me_map;
3363 env->me_metas[0] = METADATA(p);
3364 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3370 mdb_env_set_mapsize(MDB_env *env, size_t size)
3372 /* If env is already open, caller is responsible for making
3373 * sure there are no active txns.
3381 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3382 else if (size < env->me_mapsize) {
3383 /* If the configured size is smaller, make sure it's
3384 * still big enough. Silently round up to minimum if not.
3386 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3390 munmap(env->me_map, env->me_mapsize);
3391 env->me_mapsize = size;
3392 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3393 rc = mdb_env_map(env, old, 1);
3397 env->me_mapsize = size;
3399 env->me_maxpg = env->me_mapsize / env->me_psize;
3404 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3408 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3413 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3415 if (env->me_map || readers < 1)
3417 env->me_maxreaders = readers;
3422 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3424 if (!env || !readers)
3426 *readers = env->me_maxreaders;
3430 /** Further setup required for opening an MDB environment
3433 mdb_env_open2(MDB_env *env)
3435 unsigned int flags = env->me_flags;
3436 int i, newenv = 0, rc;
3440 /* See if we should use QueryLimited */
3442 if ((rc & 0xff) > 5)
3443 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3445 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3448 memset(&meta, 0, sizeof(meta));
3450 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3453 DPUTS("new mdbenv");
3455 env->me_psize = env->me_os_psize;
3456 if (env->me_psize > MAX_PAGESIZE)
3457 env->me_psize = MAX_PAGESIZE;
3459 env->me_psize = meta.mm_psize;
3462 /* Was a mapsize configured? */
3463 if (!env->me_mapsize) {
3464 /* If this is a new environment, take the default,
3465 * else use the size recorded in the existing env.
3467 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3468 } else if (env->me_mapsize < meta.mm_mapsize) {
3469 /* If the configured size is smaller, make sure it's
3470 * still big enough. Silently round up to minimum if not.
3472 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3473 if (env->me_mapsize < minsize)
3474 env->me_mapsize = minsize;
3477 rc = mdb_env_map(env, meta.mm_address, newenv);
3482 if (flags & MDB_FIXEDMAP)
3483 meta.mm_address = env->me_map;
3484 i = mdb_env_init_meta(env, &meta);
3485 if (i != MDB_SUCCESS) {
3489 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3490 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3493 env->me_maxpg = env->me_mapsize / env->me_psize;
3496 int toggle = mdb_env_pick_meta(env);
3497 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3499 DPRINTF(("opened database version %u, pagesize %u",
3500 env->me_metas[0]->mm_version, env->me_psize));
3501 DPRINTF(("using meta page %d", toggle));
3502 DPRINTF(("depth: %u", db->md_depth));
3503 DPRINTF(("entries: %"Z"u", db->md_entries));
3504 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3505 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3506 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3507 DPRINTF(("root: %"Z"u", db->md_root));
3515 /** Release a reader thread's slot in the reader lock table.
3516 * This function is called automatically when a thread exits.
3517 * @param[in] ptr This points to the slot in the reader lock table.
3520 mdb_env_reader_dest(void *ptr)
3522 MDB_reader *reader = ptr;
3528 /** Junk for arranging thread-specific callbacks on Windows. This is
3529 * necessarily platform and compiler-specific. Windows supports up
3530 * to 1088 keys. Let's assume nobody opens more than 64 environments
3531 * in a single process, for now. They can override this if needed.
3533 #ifndef MAX_TLS_KEYS
3534 #define MAX_TLS_KEYS 64
3536 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3537 static int mdb_tls_nkeys;
3539 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3543 case DLL_PROCESS_ATTACH: break;
3544 case DLL_THREAD_ATTACH: break;
3545 case DLL_THREAD_DETACH:
3546 for (i=0; i<mdb_tls_nkeys; i++) {
3547 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3548 mdb_env_reader_dest(r);
3551 case DLL_PROCESS_DETACH: break;
3556 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3558 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3562 /* Force some symbol references.
3563 * _tls_used forces the linker to create the TLS directory if not already done
3564 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3566 #pragma comment(linker, "/INCLUDE:_tls_used")
3567 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3568 #pragma const_seg(".CRT$XLB")
3569 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3570 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3573 #pragma comment(linker, "/INCLUDE:__tls_used")
3574 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3575 #pragma data_seg(".CRT$XLB")
3576 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3578 #endif /* WIN 32/64 */
3579 #endif /* !__GNUC__ */
3582 /** Downgrade the exclusive lock on the region back to shared */
3584 mdb_env_share_locks(MDB_env *env, int *excl)
3586 int rc = 0, toggle = mdb_env_pick_meta(env);
3588 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3593 /* First acquire a shared lock. The Unlock will
3594 * then release the existing exclusive lock.
3596 memset(&ov, 0, sizeof(ov));
3597 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3600 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3606 struct flock lock_info;
3607 /* The shared lock replaces the existing lock */
3608 memset((void *)&lock_info, 0, sizeof(lock_info));
3609 lock_info.l_type = F_RDLCK;
3610 lock_info.l_whence = SEEK_SET;
3611 lock_info.l_start = 0;
3612 lock_info.l_len = 1;
3613 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3614 (rc = ErrCode()) == EINTR) ;
3615 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3622 /** Try to get exlusive lock, otherwise shared.
3623 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3626 mdb_env_excl_lock(MDB_env *env, int *excl)
3630 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3634 memset(&ov, 0, sizeof(ov));
3635 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3642 struct flock lock_info;
3643 memset((void *)&lock_info, 0, sizeof(lock_info));
3644 lock_info.l_type = F_WRLCK;
3645 lock_info.l_whence = SEEK_SET;
3646 lock_info.l_start = 0;
3647 lock_info.l_len = 1;
3648 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3649 (rc = ErrCode()) == EINTR) ;
3653 # ifdef MDB_USE_POSIX_SEM
3654 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3657 lock_info.l_type = F_RDLCK;
3658 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3659 (rc = ErrCode()) == EINTR) ;
3669 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3671 * @(#) $Revision: 5.1 $
3672 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3673 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3675 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3679 * Please do not copyright this code. This code is in the public domain.
3681 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3682 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3683 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3684 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3685 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3686 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3687 * PERFORMANCE OF THIS SOFTWARE.
3690 * chongo <Landon Curt Noll> /\oo/\
3691 * http://www.isthe.com/chongo/
3693 * Share and Enjoy! :-)
3696 typedef unsigned long long mdb_hash_t;
3697 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3699 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3700 * @param[in] val value to hash
3701 * @param[in] hval initial value for hash
3702 * @return 64 bit hash
3704 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3705 * hval arg on the first call.
3708 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3710 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3711 unsigned char *end = s + val->mv_size;
3713 * FNV-1a hash each octet of the string
3716 /* xor the bottom with the current octet */
3717 hval ^= (mdb_hash_t)*s++;
3719 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3720 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3721 (hval << 7) + (hval << 8) + (hval << 40);
3723 /* return our new hash value */
3727 /** Hash the string and output the encoded hash.
3728 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
3729 * very short name limits. We don't care about the encoding being reversible,
3730 * we just want to preserve as many bits of the input as possible in a
3731 * small printable string.
3732 * @param[in] str string to hash
3733 * @param[out] encbuf an array of 11 chars to hold the hash
3735 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
3738 mdb_pack85(unsigned long l, char *out)
3742 for (i=0; i<5; i++) {
3743 *out++ = mdb_a85[l % 85];
3749 mdb_hash_enc(MDB_val *val, char *encbuf)
3751 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3753 mdb_pack85(h, encbuf);
3754 mdb_pack85(h>>32, encbuf+5);
3759 /** Open and/or initialize the lock region for the environment.
3760 * @param[in] env The MDB environment.
3761 * @param[in] lpath The pathname of the file used for the lock region.
3762 * @param[in] mode The Unix permissions for the file, if we create it.
3763 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3764 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3765 * @return 0 on success, non-zero on failure.
3768 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3771 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3773 # define MDB_ERRCODE_ROFS EROFS
3774 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3775 # define MDB_CLOEXEC O_CLOEXEC
3778 # define MDB_CLOEXEC 0
3785 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3786 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3787 FILE_ATTRIBUTE_NORMAL, NULL);
3789 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3791 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3793 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3798 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3799 /* Lose record locks when exec*() */
3800 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3801 fcntl(env->me_lfd, F_SETFD, fdflags);
3804 if (!(env->me_flags & MDB_NOTLS)) {
3805 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3808 env->me_flags |= MDB_ENV_TXKEY;
3810 /* Windows TLS callbacks need help finding their TLS info. */
3811 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3815 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3819 /* Try to get exclusive lock. If we succeed, then
3820 * nobody is using the lock region and we should initialize it.
3822 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3825 size = GetFileSize(env->me_lfd, NULL);
3827 size = lseek(env->me_lfd, 0, SEEK_END);
3828 if (size == -1) goto fail_errno;
3830 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3831 if (size < rsize && *excl > 0) {
3833 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
3834 || !SetEndOfFile(env->me_lfd))
3837 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3841 size = rsize - sizeof(MDB_txninfo);
3842 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3847 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3849 if (!mh) goto fail_errno;
3850 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3852 if (!env->me_txns) goto fail_errno;
3854 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3856 if (m == MAP_FAILED) goto fail_errno;
3862 BY_HANDLE_FILE_INFORMATION stbuf;
3871 if (!mdb_sec_inited) {
3872 InitializeSecurityDescriptor(&mdb_null_sd,
3873 SECURITY_DESCRIPTOR_REVISION);
3874 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3875 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3876 mdb_all_sa.bInheritHandle = FALSE;
3877 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3880 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3881 idbuf.volume = stbuf.dwVolumeSerialNumber;
3882 idbuf.nhigh = stbuf.nFileIndexHigh;
3883 idbuf.nlow = stbuf.nFileIndexLow;
3884 val.mv_data = &idbuf;
3885 val.mv_size = sizeof(idbuf);
3886 mdb_hash_enc(&val, encbuf);
3887 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
3888 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
3889 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3890 if (!env->me_rmutex) goto fail_errno;
3891 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3892 if (!env->me_wmutex) goto fail_errno;
3893 #elif defined(MDB_USE_POSIX_SEM)
3902 #if defined(__NetBSD__)
3903 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
3905 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3906 idbuf.dev = stbuf.st_dev;
3907 idbuf.ino = stbuf.st_ino;
3908 val.mv_data = &idbuf;
3909 val.mv_size = sizeof(idbuf);
3910 mdb_hash_enc(&val, encbuf);
3911 #ifdef MDB_SHORT_SEMNAMES
3912 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
3914 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
3915 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
3916 /* Clean up after a previous run, if needed: Try to
3917 * remove both semaphores before doing anything else.
3919 sem_unlink(env->me_txns->mti_rmname);
3920 sem_unlink(env->me_txns->mti_wmname);
3921 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3922 O_CREAT|O_EXCL, mode, 1);
3923 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3924 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3925 O_CREAT|O_EXCL, mode, 1);
3926 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3927 #else /* MDB_USE_POSIX_SEM */
3928 pthread_mutexattr_t mattr;
3930 if ((rc = pthread_mutexattr_init(&mattr))
3931 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3932 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3933 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3935 pthread_mutexattr_destroy(&mattr);
3936 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3938 env->me_txns->mti_magic = MDB_MAGIC;
3939 env->me_txns->mti_format = MDB_LOCK_FORMAT;
3940 env->me_txns->mti_txnid = 0;
3941 env->me_txns->mti_numreaders = 0;
3944 if (env->me_txns->mti_magic != MDB_MAGIC) {
3945 DPUTS("lock region has invalid magic");
3949 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
3950 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
3951 env->me_txns->mti_format, MDB_LOCK_FORMAT));
3952 rc = MDB_VERSION_MISMATCH;
3956 if (rc && rc != EACCES && rc != EAGAIN) {
3960 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3961 if (!env->me_rmutex) goto fail_errno;
3962 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3963 if (!env->me_wmutex) goto fail_errno;
3964 #elif defined(MDB_USE_POSIX_SEM)
3965 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3966 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3967 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3968 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3979 /** The name of the lock file in the DB environment */
3980 #define LOCKNAME "/lock.mdb"
3981 /** The name of the data file in the DB environment */
3982 #define DATANAME "/data.mdb"
3983 /** The suffix of the lock file when no subdir is used */
3984 #define LOCKSUFF "-lock"
3985 /** Only a subset of the @ref mdb_env flags can be changed
3986 * at runtime. Changing other flags requires closing the
3987 * environment and re-opening it with the new flags.
3989 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
3990 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
3991 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
3994 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3996 int oflags, rc, len, excl = -1;
3997 char *lpath, *dpath;
3999 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4003 if (flags & MDB_NOSUBDIR) {
4004 rc = len + sizeof(LOCKSUFF) + len + 1;
4006 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4011 if (flags & MDB_NOSUBDIR) {
4012 dpath = lpath + len + sizeof(LOCKSUFF);
4013 sprintf(lpath, "%s" LOCKSUFF, path);
4014 strcpy(dpath, path);
4016 dpath = lpath + len + sizeof(LOCKNAME);
4017 sprintf(lpath, "%s" LOCKNAME, path);
4018 sprintf(dpath, "%s" DATANAME, path);
4022 flags |= env->me_flags;
4023 if (flags & MDB_RDONLY) {
4024 /* silently ignore WRITEMAP when we're only getting read access */
4025 flags &= ~MDB_WRITEMAP;
4027 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4028 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4031 env->me_flags = flags |= MDB_ENV_ACTIVE;
4035 env->me_path = strdup(path);
4036 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4037 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4038 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4043 /* For RDONLY, get lockfile after we know datafile exists */
4044 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4045 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4051 if (F_ISSET(flags, MDB_RDONLY)) {
4052 oflags = GENERIC_READ;
4053 len = OPEN_EXISTING;
4055 oflags = GENERIC_READ|GENERIC_WRITE;
4058 mode = FILE_ATTRIBUTE_NORMAL;
4059 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4060 NULL, len, mode, NULL);
4062 if (F_ISSET(flags, MDB_RDONLY))
4065 oflags = O_RDWR | O_CREAT;
4067 env->me_fd = open(dpath, oflags, mode);
4069 if (env->me_fd == INVALID_HANDLE_VALUE) {
4074 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4075 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4080 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4081 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4082 env->me_mfd = env->me_fd;
4084 /* Synchronous fd for meta writes. Needed even with
4085 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4088 len = OPEN_EXISTING;
4089 env->me_mfd = CreateFile(dpath, oflags,
4090 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4091 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4094 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4096 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4101 DPRINTF(("opened dbenv %p", (void *) env));
4103 rc = mdb_env_share_locks(env, &excl);
4107 if (!((flags & MDB_RDONLY) ||
4108 (env->me_pbuf = calloc(1, env->me_psize))))
4114 mdb_env_close0(env, excl);
4120 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4122 mdb_env_close0(MDB_env *env, int excl)
4126 if (!(env->me_flags & MDB_ENV_ACTIVE))
4129 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4130 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4131 free(env->me_dbxs[i].md_name.mv_data);
4134 free(env->me_dbflags);
4137 free(env->me_dirty_list);
4138 mdb_midl_free(env->me_free_pgs);
4140 if (env->me_flags & MDB_ENV_TXKEY) {
4141 pthread_key_delete(env->me_txkey);
4143 /* Delete our key from the global list */
4144 for (i=0; i<mdb_tls_nkeys; i++)
4145 if (mdb_tls_keys[i] == env->me_txkey) {
4146 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4154 munmap(env->me_map, env->me_mapsize);
4156 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4157 (void) close(env->me_mfd);
4158 if (env->me_fd != INVALID_HANDLE_VALUE)
4159 (void) close(env->me_fd);
4161 MDB_PID_T pid = env->me_pid;
4162 /* Clearing readers is done in this function because
4163 * me_txkey with its destructor must be disabled first.
4165 for (i = env->me_numreaders; --i >= 0; )
4166 if (env->me_txns->mti_readers[i].mr_pid == pid)
4167 env->me_txns->mti_readers[i].mr_pid = 0;
4169 if (env->me_rmutex) {
4170 CloseHandle(env->me_rmutex);
4171 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4173 /* Windows automatically destroys the mutexes when
4174 * the last handle closes.
4176 #elif defined(MDB_USE_POSIX_SEM)
4177 if (env->me_rmutex != SEM_FAILED) {
4178 sem_close(env->me_rmutex);
4179 if (env->me_wmutex != SEM_FAILED)
4180 sem_close(env->me_wmutex);
4181 /* If we have the filelock: If we are the
4182 * only remaining user, clean up semaphores.
4185 mdb_env_excl_lock(env, &excl);
4187 sem_unlink(env->me_txns->mti_rmname);
4188 sem_unlink(env->me_txns->mti_wmname);
4192 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4194 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4197 /* Unlock the lockfile. Windows would have unlocked it
4198 * after closing anyway, but not necessarily at once.
4200 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4203 (void) close(env->me_lfd);
4206 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4210 mdb_env_copyfd(MDB_env *env, HANDLE fd)
4212 MDB_txn *txn = NULL;
4218 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
4222 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
4225 /* Do the lock/unlock of the reader mutex before starting the
4226 * write txn. Otherwise other read txns could block writers.
4228 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
4233 /* We must start the actual read txn after blocking writers */
4234 mdb_txn_reset0(txn, "reset-stage1");
4236 /* Temporarily block writers until we snapshot the meta pages */
4239 rc = mdb_txn_renew0(txn);
4241 UNLOCK_MUTEX_W(env);
4246 wsize = env->me_psize * 2;
4250 DO_WRITE(rc, fd, ptr, w2, len);
4254 } else if (len > 0) {
4260 /* Non-blocking or async handles are not supported */
4266 UNLOCK_MUTEX_W(env);
4271 wsize = txn->mt_next_pgno * env->me_psize - wsize;
4273 if (wsize > MAX_WRITE)
4277 DO_WRITE(rc, fd, ptr, w2, len);
4281 } else if (len > 0) {
4298 mdb_env_copy(MDB_env *env, const char *path)
4302 HANDLE newfd = INVALID_HANDLE_VALUE;
4304 if (env->me_flags & MDB_NOSUBDIR) {
4305 lpath = (char *)path;
4308 len += sizeof(DATANAME);
4309 lpath = malloc(len);
4312 sprintf(lpath, "%s" DATANAME, path);
4315 /* The destination path must exist, but the destination file must not.
4316 * We don't want the OS to cache the writes, since the source data is
4317 * already in the OS cache.
4320 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
4321 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
4323 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
4325 if (newfd == INVALID_HANDLE_VALUE) {
4331 /* Set O_DIRECT if the file system supports it */
4332 if ((rc = fcntl(newfd, F_GETFL)) != -1)
4333 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
4335 #ifdef F_NOCACHE /* __APPLE__ */
4336 rc = fcntl(newfd, F_NOCACHE, 1);
4343 rc = mdb_env_copyfd(env, newfd);
4346 if (!(env->me_flags & MDB_NOSUBDIR))
4348 if (newfd != INVALID_HANDLE_VALUE)
4349 if (close(newfd) < 0 && rc == MDB_SUCCESS)
4356 mdb_env_close(MDB_env *env)
4363 VGMEMP_DESTROY(env);
4364 while ((dp = env->me_dpages) != NULL) {
4365 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4366 env->me_dpages = dp->mp_next;
4370 mdb_env_close0(env, 0);
4374 /** Compare two items pointing at aligned size_t's */
4376 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4378 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4379 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4382 /** Compare two items pointing at aligned unsigned int's */
4384 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4386 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4387 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4390 /** Compare two items pointing at unsigned ints of unknown alignment.
4391 * Nodes and keys are guaranteed to be 2-byte aligned.
4394 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4396 #if BYTE_ORDER == LITTLE_ENDIAN
4397 unsigned short *u, *c;
4400 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4401 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4404 } while(!x && u > (unsigned short *)a->mv_data);
4407 return memcmp(a->mv_data, b->mv_data, a->mv_size);
4411 /** Compare two items lexically */
4413 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4420 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4426 diff = memcmp(a->mv_data, b->mv_data, len);
4427 return diff ? diff : len_diff<0 ? -1 : len_diff;
4430 /** Compare two items in reverse byte order */
4432 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4434 const unsigned char *p1, *p2, *p1_lim;
4438 p1_lim = (const unsigned char *)a->mv_data;
4439 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4440 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4442 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4448 while (p1 > p1_lim) {
4449 diff = *--p1 - *--p2;
4453 return len_diff<0 ? -1 : len_diff;
4456 /** Search for key within a page, using binary search.
4457 * Returns the smallest entry larger or equal to the key.
4458 * If exactp is non-null, stores whether the found entry was an exact match
4459 * in *exactp (1 or 0).
4460 * Updates the cursor index with the index of the found entry.
4461 * If no entry larger or equal to the key is found, returns NULL.
4464 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4466 unsigned int i = 0, nkeys;
4469 MDB_page *mp = mc->mc_pg[mc->mc_top];
4470 MDB_node *node = NULL;
4475 nkeys = NUMKEYS(mp);
4480 COPY_PGNO(pgno, mp->mp_pgno);
4481 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4482 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4489 low = IS_LEAF(mp) ? 0 : 1;
4491 cmp = mc->mc_dbx->md_cmp;
4493 /* Branch pages have no data, so if using integer keys,
4494 * alignment is guaranteed. Use faster mdb_cmp_int.
4496 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4497 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4504 nodekey.mv_size = mc->mc_db->md_pad;
4505 node = NODEPTR(mp, 0); /* fake */
4506 while (low <= high) {
4507 i = (low + high) >> 1;
4508 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4509 rc = cmp(key, &nodekey);
4510 DPRINTF(("found leaf index %u [%s], rc = %i",
4511 i, DKEY(&nodekey), rc));
4520 while (low <= high) {
4521 i = (low + high) >> 1;
4523 node = NODEPTR(mp, i);
4524 nodekey.mv_size = NODEKSZ(node);
4525 nodekey.mv_data = NODEKEY(node);
4527 rc = cmp(key, &nodekey);
4530 DPRINTF(("found leaf index %u [%s], rc = %i",
4531 i, DKEY(&nodekey), rc));
4533 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4534 i, DKEY(&nodekey), NODEPGNO(node), rc));
4545 if (rc > 0) { /* Found entry is less than the key. */
4546 i++; /* Skip to get the smallest entry larger than key. */
4548 node = NODEPTR(mp, i);
4551 *exactp = (rc == 0);
4552 /* store the key index */
4553 mc->mc_ki[mc->mc_top] = i;
4555 /* There is no entry larger or equal to the key. */
4558 /* nodeptr is fake for LEAF2 */
4564 mdb_cursor_adjust(MDB_cursor *mc, func)
4568 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4569 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4576 /** Pop a page off the top of the cursor's stack. */
4578 mdb_cursor_pop(MDB_cursor *mc)
4582 MDB_page *top = mc->mc_pg[mc->mc_top];
4588 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4589 DDBI(mc), (void *) mc));
4593 /** Push a page onto the top of the cursor's stack. */
4595 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4597 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4598 DDBI(mc), (void *) mc));
4600 if (mc->mc_snum >= CURSOR_STACK) {
4601 assert(mc->mc_snum < CURSOR_STACK);
4602 return MDB_CURSOR_FULL;
4605 mc->mc_top = mc->mc_snum++;
4606 mc->mc_pg[mc->mc_top] = mp;
4607 mc->mc_ki[mc->mc_top] = 0;
4612 /** Find the address of the page corresponding to a given page number.
4613 * @param[in] txn the transaction for this access.
4614 * @param[in] pgno the page number for the page to retrieve.
4615 * @param[out] ret address of a pointer where the page's address will be stored.
4616 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4617 * @return 0 on success, non-zero on failure.
4620 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4622 MDB_env *env = txn->mt_env;
4626 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4630 MDB_ID2L dl = tx2->mt_u.dirty_list;
4632 /* Spilled pages were dirtied in this txn and flushed
4633 * because the dirty list got full. Bring this page
4634 * back in from the map (but don't unspill it here,
4635 * leave that unless page_touch happens again).
4637 if (tx2->mt_spill_pgs) {
4638 MDB_ID pn = pgno << 1;
4639 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4640 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4641 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4646 unsigned x = mdb_mid2l_search(dl, pgno);
4647 if (x <= dl[0].mid && dl[x].mid == pgno) {
4653 } while ((tx2 = tx2->mt_parent) != NULL);
4656 if (pgno < txn->mt_next_pgno) {
4658 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4660 DPRINTF(("page %"Z"u not found", pgno));
4662 return MDB_PAGE_NOTFOUND;
4672 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4673 * The cursor is at the root page, set up the rest of it.
4676 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4678 MDB_page *mp = mc->mc_pg[mc->mc_top];
4682 while (IS_BRANCH(mp)) {
4686 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4687 assert(NUMKEYS(mp) > 1);
4688 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4690 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4692 if (flags & MDB_PS_LAST)
4693 i = NUMKEYS(mp) - 1;
4696 node = mdb_node_search(mc, key, &exact);
4698 i = NUMKEYS(mp) - 1;
4700 i = mc->mc_ki[mc->mc_top];
4706 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4709 assert(i < NUMKEYS(mp));
4710 node = NODEPTR(mp, i);
4712 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4715 mc->mc_ki[mc->mc_top] = i;
4716 if ((rc = mdb_cursor_push(mc, mp)))
4719 if (flags & MDB_PS_MODIFY) {
4720 if ((rc = mdb_page_touch(mc)) != 0)
4722 mp = mc->mc_pg[mc->mc_top];
4727 DPRINTF(("internal error, index points to a %02X page!?",
4729 return MDB_CORRUPTED;
4732 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4733 key ? DKEY(key) : "null"));
4734 mc->mc_flags |= C_INITIALIZED;
4735 mc->mc_flags &= ~C_EOF;
4740 /** Search for the lowest key under the current branch page.
4741 * This just bypasses a NUMKEYS check in the current page
4742 * before calling mdb_page_search_root(), because the callers
4743 * are all in situations where the current page is known to
4747 mdb_page_search_lowest(MDB_cursor *mc)
4749 MDB_page *mp = mc->mc_pg[mc->mc_top];
4750 MDB_node *node = NODEPTR(mp, 0);
4753 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4756 mc->mc_ki[mc->mc_top] = 0;
4757 if ((rc = mdb_cursor_push(mc, mp)))
4759 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4762 /** Search for the page a given key should be in.
4763 * Push it and its parent pages on the cursor stack.
4764 * @param[in,out] mc the cursor for this operation.
4765 * @param[in] key the key to search for, or NULL for first/last page.
4766 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4767 * are touched (updated with new page numbers).
4768 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4769 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4770 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4771 * @return 0 on success, non-zero on failure.
4774 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4779 /* Make sure the txn is still viable, then find the root from
4780 * the txn's db table and set it as the root of the cursor's stack.
4782 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4783 DPUTS("transaction has failed, must abort");
4786 /* Make sure we're using an up-to-date root */
4787 if (*mc->mc_dbflag & DB_STALE) {
4789 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4790 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4797 MDB_node *leaf = mdb_node_search(&mc2,
4798 &mc->mc_dbx->md_name, &exact);
4800 return MDB_NOTFOUND;
4801 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4804 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4806 /* The txn may not know this DBI, or another process may
4807 * have dropped and recreated the DB with other flags.
4809 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4810 return MDB_INCOMPATIBLE;
4811 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4813 *mc->mc_dbflag &= ~DB_STALE;
4815 root = mc->mc_db->md_root;
4817 if (root == P_INVALID) { /* Tree is empty. */
4818 DPUTS("tree is empty");
4819 return MDB_NOTFOUND;
4824 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4825 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4831 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4832 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4834 if (flags & MDB_PS_MODIFY) {
4835 if ((rc = mdb_page_touch(mc)))
4839 if (flags & MDB_PS_ROOTONLY)
4842 return mdb_page_search_root(mc, key, flags);
4846 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4848 MDB_txn *txn = mc->mc_txn;
4849 pgno_t pg = mp->mp_pgno;
4850 unsigned x = 0, ovpages = mp->mp_pages;
4851 MDB_env *env = txn->mt_env;
4852 MDB_IDL sl = txn->mt_spill_pgs;
4853 MDB_ID pn = pg << 1;
4856 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
4857 /* If the page is dirty or on the spill list we just acquired it,
4858 * so we should give it back to our current free list, if any.
4859 * Otherwise put it onto the list of pages we freed in this txn.
4861 * Won't create me_pghead: me_pglast must be inited along with it.
4862 * Unsupported in nested txns: They would need to hide the page
4863 * range in ancestor txns' dirty and spilled lists.
4865 if (env->me_pghead &&
4867 ((mp->mp_flags & P_DIRTY) ||
4868 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
4872 MDB_ID2 *dl, ix, iy;
4873 rc = mdb_midl_need(&env->me_pghead, ovpages);
4876 if (!(mp->mp_flags & P_DIRTY)) {
4877 /* This page is no longer spilled */
4884 /* Remove from dirty list */
4885 dl = txn->mt_u.dirty_list;
4887 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
4895 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
4896 txn->mt_flags |= MDB_TXN_ERROR;
4897 return MDB_CORRUPTED;
4900 if (!(env->me_flags & MDB_WRITEMAP))
4901 mdb_dpage_free(env, mp);
4903 /* Insert in me_pghead */
4904 mop = env->me_pghead;
4905 j = mop[0] + ovpages;
4906 for (i = mop[0]; i && mop[i] < pg; i--)
4912 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4916 mc->mc_db->md_overflow_pages -= ovpages;
4920 /** Return the data associated with a given node.
4921 * @param[in] txn The transaction for this operation.
4922 * @param[in] leaf The node being read.
4923 * @param[out] data Updated to point to the node's data.
4924 * @return 0 on success, non-zero on failure.
4927 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4929 MDB_page *omp; /* overflow page */
4933 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4934 data->mv_size = NODEDSZ(leaf);
4935 data->mv_data = NODEDATA(leaf);
4939 /* Read overflow data.
4941 data->mv_size = NODEDSZ(leaf);
4942 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4943 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4944 DPRINTF(("read overflow page %"Z"u failed", pgno));
4947 data->mv_data = METADATA(omp);
4953 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4954 MDB_val *key, MDB_val *data)
4963 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
4965 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4968 if (txn->mt_flags & MDB_TXN_ERROR)
4971 mdb_cursor_init(&mc, txn, dbi, &mx);
4972 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4975 /** Find a sibling for a page.
4976 * Replaces the page at the top of the cursor's stack with the
4977 * specified sibling, if one exists.
4978 * @param[in] mc The cursor for this operation.
4979 * @param[in] move_right Non-zero if the right sibling is requested,
4980 * otherwise the left sibling.
4981 * @return 0 on success, non-zero on failure.
4984 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4990 if (mc->mc_snum < 2) {
4991 return MDB_NOTFOUND; /* root has no siblings */
4995 DPRINTF(("parent page is page %"Z"u, index %u",
4996 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
4998 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4999 : (mc->mc_ki[mc->mc_top] == 0)) {
5000 DPRINTF(("no more keys left, moving to %s sibling",
5001 move_right ? "right" : "left"));
5002 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5003 /* undo cursor_pop before returning */
5010 mc->mc_ki[mc->mc_top]++;
5012 mc->mc_ki[mc->mc_top]--;
5013 DPRINTF(("just moving to %s index key %u",
5014 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5016 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
5018 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5019 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5020 /* mc will be inconsistent if caller does mc_snum++ as above */
5021 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5025 mdb_cursor_push(mc, mp);
5027 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5032 /** Move the cursor to the next data item. */
5034 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5040 if (mc->mc_flags & C_EOF) {
5041 return MDB_NOTFOUND;
5044 assert(mc->mc_flags & C_INITIALIZED);
5046 mp = mc->mc_pg[mc->mc_top];
5048 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5049 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5050 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5051 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5052 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5053 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5054 if (rc == MDB_SUCCESS)
5055 MDB_GET_KEY(leaf, key);
5060 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5061 if (op == MDB_NEXT_DUP)
5062 return MDB_NOTFOUND;
5066 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5067 if (mc->mc_flags & C_DEL)
5070 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5071 DPUTS("=====> move to next sibling page");
5072 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5073 mc->mc_flags |= C_EOF;
5076 mp = mc->mc_pg[mc->mc_top];
5077 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5079 mc->mc_ki[mc->mc_top]++;
5082 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5083 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5086 key->mv_size = mc->mc_db->md_pad;
5087 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5091 assert(IS_LEAF(mp));
5092 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5094 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5095 mdb_xcursor_init1(mc, leaf);
5098 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5101 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5102 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5103 if (rc != MDB_SUCCESS)
5108 MDB_GET_KEY(leaf, key);
5112 /** Move the cursor to the previous data item. */
5114 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5120 assert(mc->mc_flags & C_INITIALIZED);
5122 mp = mc->mc_pg[mc->mc_top];
5124 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5125 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5126 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5127 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5128 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5129 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5130 if (rc == MDB_SUCCESS)
5131 MDB_GET_KEY(leaf, key);
5135 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5136 if (op == MDB_PREV_DUP)
5137 return MDB_NOTFOUND;
5142 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p", mp->mp_pgno, (void *) mc));
5144 if (mc->mc_ki[mc->mc_top] == 0) {
5145 DPUTS("=====> move to prev sibling page");
5146 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5149 mp = mc->mc_pg[mc->mc_top];
5150 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5151 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5153 mc->mc_ki[mc->mc_top]--;
5155 mc->mc_flags &= ~C_EOF;
5157 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5158 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5161 key->mv_size = mc->mc_db->md_pad;
5162 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5166 assert(IS_LEAF(mp));
5167 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5169 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5170 mdb_xcursor_init1(mc, leaf);
5173 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5176 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5177 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5178 if (rc != MDB_SUCCESS)
5183 MDB_GET_KEY(leaf, key);
5187 /** Set the cursor on a specific data item. */
5189 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5190 MDB_cursor_op op, int *exactp)
5194 MDB_node *leaf = NULL;
5199 if (key->mv_size == 0)
5200 return MDB_BAD_VALSIZE;
5203 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5205 /* See if we're already on the right page */
5206 if (mc->mc_flags & C_INITIALIZED) {
5209 mp = mc->mc_pg[mc->mc_top];
5211 mc->mc_ki[mc->mc_top] = 0;
5212 return MDB_NOTFOUND;
5214 if (mp->mp_flags & P_LEAF2) {
5215 nodekey.mv_size = mc->mc_db->md_pad;
5216 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5218 leaf = NODEPTR(mp, 0);
5219 MDB_GET_KEY2(leaf, nodekey);
5221 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5223 /* Probably happens rarely, but first node on the page
5224 * was the one we wanted.
5226 mc->mc_ki[mc->mc_top] = 0;
5233 unsigned int nkeys = NUMKEYS(mp);
5235 if (mp->mp_flags & P_LEAF2) {
5236 nodekey.mv_data = LEAF2KEY(mp,
5237 nkeys-1, nodekey.mv_size);
5239 leaf = NODEPTR(mp, nkeys-1);
5240 MDB_GET_KEY2(leaf, nodekey);
5242 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5244 /* last node was the one we wanted */
5245 mc->mc_ki[mc->mc_top] = nkeys-1;
5251 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5252 /* This is definitely the right page, skip search_page */
5253 if (mp->mp_flags & P_LEAF2) {
5254 nodekey.mv_data = LEAF2KEY(mp,
5255 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5257 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5258 MDB_GET_KEY2(leaf, nodekey);
5260 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5262 /* current node was the one we wanted */
5272 /* If any parents have right-sibs, search.
5273 * Otherwise, there's nothing further.
5275 for (i=0; i<mc->mc_top; i++)
5277 NUMKEYS(mc->mc_pg[i])-1)
5279 if (i == mc->mc_top) {
5280 /* There are no other pages */
5281 mc->mc_ki[mc->mc_top] = nkeys;
5282 return MDB_NOTFOUND;
5286 /* There are no other pages */
5287 mc->mc_ki[mc->mc_top] = 0;
5288 if (op == MDB_SET_RANGE) {
5292 return MDB_NOTFOUND;
5296 rc = mdb_page_search(mc, key, 0);
5297 if (rc != MDB_SUCCESS)
5300 mp = mc->mc_pg[mc->mc_top];
5301 assert(IS_LEAF(mp));
5304 leaf = mdb_node_search(mc, key, exactp);
5305 if (exactp != NULL && !*exactp) {
5306 /* MDB_SET specified and not an exact match. */
5307 return MDB_NOTFOUND;
5311 DPUTS("===> inexact leaf not found, goto sibling");
5312 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5313 return rc; /* no entries matched */
5314 mp = mc->mc_pg[mc->mc_top];
5315 assert(IS_LEAF(mp));
5316 leaf = NODEPTR(mp, 0);
5320 mc->mc_flags |= C_INITIALIZED;
5321 mc->mc_flags &= ~C_EOF;
5324 key->mv_size = mc->mc_db->md_pad;
5325 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5329 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5330 mdb_xcursor_init1(mc, leaf);
5333 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5334 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5335 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5338 if (op == MDB_GET_BOTH) {
5344 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5345 if (rc != MDB_SUCCESS)
5348 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5350 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5352 rc = mc->mc_dbx->md_dcmp(data, &d2);
5354 if (op == MDB_GET_BOTH || rc > 0)
5355 return MDB_NOTFOUND;
5362 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5363 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5368 /* The key already matches in all other cases */
5369 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5370 MDB_GET_KEY(leaf, key);
5371 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5376 /** Move the cursor to the first item in the database. */
5378 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5384 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5386 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5387 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5388 if (rc != MDB_SUCCESS)
5391 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5393 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5394 mc->mc_flags |= C_INITIALIZED;
5395 mc->mc_flags &= ~C_EOF;
5397 mc->mc_ki[mc->mc_top] = 0;
5399 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5400 key->mv_size = mc->mc_db->md_pad;
5401 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5406 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5407 mdb_xcursor_init1(mc, leaf);
5408 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5412 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5416 MDB_GET_KEY(leaf, key);
5420 /** Move the cursor to the last item in the database. */
5422 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5428 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5430 if (!(mc->mc_flags & C_EOF)) {
5432 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5433 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5434 if (rc != MDB_SUCCESS)
5437 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
5440 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5441 mc->mc_flags |= C_INITIALIZED|C_EOF;
5442 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5444 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5445 key->mv_size = mc->mc_db->md_pad;
5446 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5451 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5452 mdb_xcursor_init1(mc, leaf);
5453 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5457 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5462 MDB_GET_KEY(leaf, key);
5467 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5472 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5476 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5480 case MDB_GET_CURRENT:
5481 if (!(mc->mc_flags & C_INITIALIZED)) {
5484 MDB_page *mp = mc->mc_pg[mc->mc_top];
5485 int nkeys = NUMKEYS(mp);
5486 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5487 mc->mc_ki[mc->mc_top] = nkeys;
5493 key->mv_size = mc->mc_db->md_pad;
5494 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5496 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5497 MDB_GET_KEY(leaf, key);
5499 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5500 if (mc->mc_flags & C_DEL)
5501 mdb_xcursor_init1(mc, leaf);
5502 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5504 rc = mdb_node_read(mc->mc_txn, leaf, data);
5511 case MDB_GET_BOTH_RANGE:
5516 if (mc->mc_xcursor == NULL) {
5517 rc = MDB_INCOMPATIBLE;
5527 rc = mdb_cursor_set(mc, key, data, op,
5528 op == MDB_SET_RANGE ? NULL : &exact);
5531 case MDB_GET_MULTIPLE:
5532 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5536 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5537 rc = MDB_INCOMPATIBLE;
5541 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5542 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5545 case MDB_NEXT_MULTIPLE:
5550 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5551 rc = MDB_INCOMPATIBLE;
5554 if (!(mc->mc_flags & C_INITIALIZED))
5555 rc = mdb_cursor_first(mc, key, data);
5557 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5558 if (rc == MDB_SUCCESS) {
5559 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5562 mx = &mc->mc_xcursor->mx_cursor;
5563 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5565 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5566 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5574 case MDB_NEXT_NODUP:
5575 if (!(mc->mc_flags & C_INITIALIZED))
5576 rc = mdb_cursor_first(mc, key, data);
5578 rc = mdb_cursor_next(mc, key, data, op);
5582 case MDB_PREV_NODUP:
5583 if (!(mc->mc_flags & C_INITIALIZED)) {
5584 rc = mdb_cursor_last(mc, key, data);
5587 mc->mc_flags |= C_INITIALIZED;
5588 mc->mc_ki[mc->mc_top]++;
5590 rc = mdb_cursor_prev(mc, key, data, op);
5593 rc = mdb_cursor_first(mc, key, data);
5596 mfunc = mdb_cursor_first;
5598 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5602 if (mc->mc_xcursor == NULL) {
5603 rc = MDB_INCOMPATIBLE;
5606 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5610 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5613 rc = mdb_cursor_last(mc, key, data);
5616 mfunc = mdb_cursor_last;
5619 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5624 if (mc->mc_flags & C_DEL)
5625 mc->mc_flags ^= C_DEL;
5630 /** Touch all the pages in the cursor stack. Set mc_top.
5631 * Makes sure all the pages are writable, before attempting a write operation.
5632 * @param[in] mc The cursor to operate on.
5635 mdb_cursor_touch(MDB_cursor *mc)
5637 int rc = MDB_SUCCESS;
5639 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5642 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5643 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5646 *mc->mc_dbflag |= DB_DIRTY;
5651 rc = mdb_page_touch(mc);
5652 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5653 mc->mc_top = mc->mc_snum-1;
5658 /** Do not spill pages to disk if txn is getting full, may fail instead */
5659 #define MDB_NOSPILL 0x8000
5662 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5665 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5666 MDB_env *env = mc->mc_txn->mt_env;
5667 MDB_node *leaf = NULL;
5670 MDB_val xdata, *rdata, dkey, olddata;
5672 int do_sub = 0, insert;
5673 unsigned int mcount = 0, dcount = 0, nospill;
5676 unsigned int nflags;
5679 /* Check this first so counter will always be zero on any
5682 if (flags & MDB_MULTIPLE) {
5683 dcount = data[1].mv_size;
5684 data[1].mv_size = 0;
5685 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5686 return MDB_INCOMPATIBLE;
5689 nospill = flags & MDB_NOSPILL;
5690 flags &= ~MDB_NOSPILL;
5692 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5693 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5695 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5696 return MDB_BAD_VALSIZE;
5698 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5699 return MDB_BAD_VALSIZE;
5701 #if SIZE_MAX > MAXDATASIZE
5702 if (data->mv_size > MAXDATASIZE)
5703 return MDB_BAD_VALSIZE;
5706 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5707 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5711 if (flags == MDB_CURRENT) {
5712 if (!(mc->mc_flags & C_INITIALIZED))
5715 } else if (mc->mc_db->md_root == P_INVALID) {
5716 /* new database, cursor has nothing to point to */
5719 mc->mc_flags &= ~C_INITIALIZED;
5724 if (flags & MDB_APPEND) {
5726 rc = mdb_cursor_last(mc, &k2, &d2);
5728 rc = mc->mc_dbx->md_cmp(key, &k2);
5731 mc->mc_ki[mc->mc_top]++;
5733 /* new key is <= last key */
5738 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5740 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5741 DPRINTF(("duplicate key [%s]", DKEY(key)));
5743 return MDB_KEYEXIST;
5745 if (rc && rc != MDB_NOTFOUND)
5749 if (mc->mc_flags & C_DEL)
5750 mc->mc_flags ^= C_DEL;
5752 /* Cursor is positioned, check for room in the dirty list */
5754 if (flags & MDB_MULTIPLE) {
5756 xdata.mv_size = data->mv_size * dcount;
5760 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5764 if (rc == MDB_NO_ROOT) {
5766 /* new database, write a root leaf page */
5767 DPUTS("allocating new root leaf page");
5768 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5771 mdb_cursor_push(mc, np);
5772 mc->mc_db->md_root = np->mp_pgno;
5773 mc->mc_db->md_depth++;
5774 *mc->mc_dbflag |= DB_DIRTY;
5775 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5777 np->mp_flags |= P_LEAF2;
5778 mc->mc_flags |= C_INITIALIZED;
5780 /* make sure all cursor pages are writable */
5781 rc2 = mdb_cursor_touch(mc);
5788 /* The key does not exist */
5789 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5790 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5791 LEAFSIZE(key, data) > env->me_nodemax)
5793 /* Too big for a node, insert in sub-DB */
5794 fp_flags = P_LEAF|P_DIRTY;
5796 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5797 fp->mp_lower = fp->mp_upper = olddata.mv_size = PAGEHDRSZ;
5801 /* there's only a key anyway, so this is a no-op */
5802 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5803 unsigned int ksize = mc->mc_db->md_pad;
5804 if (key->mv_size != ksize)
5805 return MDB_BAD_VALSIZE;
5806 if (flags == MDB_CURRENT) {
5807 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5808 memcpy(ptr, key->mv_data, ksize);
5814 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5815 olddata.mv_size = NODEDSZ(leaf);
5816 olddata.mv_data = NODEDATA(leaf);
5819 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5820 /* Prepare (sub-)page/sub-DB to accept the new item,
5821 * if needed. fp: old sub-page or a header faking
5822 * it. mp: new (sub-)page. offset: growth in page
5823 * size. xdata: node data with new page or DB.
5825 unsigned i, offset = 0;
5826 mp = fp = xdata.mv_data = env->me_pbuf;
5827 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5829 /* Was a single item before, must convert now */
5830 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5831 /* Just overwrite the current item */
5832 if (flags == MDB_CURRENT)
5835 #if UINT_MAX < SIZE_MAX
5836 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5837 #ifdef MISALIGNED_OK
5838 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5840 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5843 /* if data matches, skip it */
5844 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5845 if (flags & MDB_NODUPDATA)
5847 else if (flags & MDB_MULTIPLE)
5854 /* Back up original data item */
5855 dkey.mv_size = olddata.mv_size;
5856 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
5858 /* Make sub-page header for the dup items, with dummy body */
5859 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5860 fp->mp_lower = PAGEHDRSZ;
5861 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5862 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5863 fp->mp_flags |= P_LEAF2;
5864 fp->mp_pad = data->mv_size;
5865 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
5867 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
5868 (dkey.mv_size & 1) + (data->mv_size & 1);
5870 fp->mp_upper = xdata.mv_size;
5871 olddata.mv_size = fp->mp_upper; /* pretend olddata is fp */
5872 } else if (leaf->mn_flags & F_SUBDATA) {
5873 /* Data is on sub-DB, just store it */
5874 flags |= F_DUPDATA|F_SUBDATA;
5877 /* Data is on sub-page */
5878 fp = olddata.mv_data;
5881 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5882 offset = EVEN(NODESIZE + sizeof(indx_t) +
5886 offset = fp->mp_pad;
5887 if (SIZELEFT(fp) < offset) {
5888 offset *= 4; /* space for 4 more */
5891 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
5893 fp->mp_flags |= P_DIRTY;
5894 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
5895 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5899 xdata.mv_size = olddata.mv_size + offset;
5902 fp_flags = fp->mp_flags;
5903 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
5904 /* Too big for a sub-page, convert to sub-DB */
5905 fp_flags &= ~P_SUBP;
5907 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5908 fp_flags |= P_LEAF2;
5909 dummy.md_pad = fp->mp_pad;
5910 dummy.md_flags = MDB_DUPFIXED;
5911 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5912 dummy.md_flags |= MDB_INTEGERKEY;
5918 dummy.md_branch_pages = 0;
5919 dummy.md_leaf_pages = 1;
5920 dummy.md_overflow_pages = 0;
5921 dummy.md_entries = NUMKEYS(fp);
5922 xdata.mv_size = sizeof(MDB_db);
5923 xdata.mv_data = &dummy;
5924 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5926 offset = env->me_psize - olddata.mv_size;
5927 flags |= F_DUPDATA|F_SUBDATA;
5928 dummy.md_root = mp->mp_pgno;
5931 mp->mp_flags = fp_flags | P_DIRTY;
5932 mp->mp_pad = fp->mp_pad;
5933 mp->mp_lower = fp->mp_lower;
5934 mp->mp_upper = fp->mp_upper + offset;
5935 if (fp_flags & P_LEAF2) {
5936 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5938 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper,
5939 olddata.mv_size - fp->mp_upper);
5940 for (i=0; i<NUMKEYS(fp); i++)
5941 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5948 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5952 /* overflow page overwrites need special handling */
5953 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5956 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
5958 memcpy(&pg, olddata.mv_data, sizeof(pg));
5959 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5961 ovpages = omp->mp_pages;
5963 /* Is the ov page large enough? */
5964 if (ovpages >= dpages) {
5965 if (!(omp->mp_flags & P_DIRTY) &&
5966 (level || (env->me_flags & MDB_WRITEMAP)))
5968 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
5971 level = 0; /* dirty in this txn or clean */
5974 if (omp->mp_flags & P_DIRTY) {
5975 /* yes, overwrite it. Note in this case we don't
5976 * bother to try shrinking the page if the new data
5977 * is smaller than the overflow threshold.
5980 /* It is writable only in a parent txn */
5981 size_t sz = (size_t) env->me_psize * ovpages, off;
5982 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5988 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5989 if (!(flags & MDB_RESERVE)) {
5990 /* Copy end of page, adjusting alignment so
5991 * compiler may copy words instead of bytes.
5993 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5994 memcpy((size_t *)((char *)np + off),
5995 (size_t *)((char *)omp + off), sz - off);
5998 memcpy(np, omp, sz); /* Copy beginning of page */
6001 SETDSZ(leaf, data->mv_size);
6002 if (F_ISSET(flags, MDB_RESERVE))
6003 data->mv_data = METADATA(omp);
6005 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6009 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6011 } else if (data->mv_size == olddata.mv_size) {
6012 /* same size, just replace it. Note that we could
6013 * also reuse this node if the new data is smaller,
6014 * but instead we opt to shrink the node in that case.
6016 if (F_ISSET(flags, MDB_RESERVE))
6017 data->mv_data = olddata.mv_data;
6018 else if (data->mv_size)
6019 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6021 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6024 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6025 mc->mc_db->md_entries--;
6031 nflags = flags & NODE_ADD_FLAGS;
6032 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6033 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6034 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6035 nflags &= ~MDB_APPEND;
6037 nflags |= MDB_SPLIT_REPLACE;
6038 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6040 /* There is room already in this leaf page. */
6041 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6042 if (rc == 0 && !do_sub && insert) {
6043 /* Adjust other cursors pointing to mp */
6044 MDB_cursor *m2, *m3;
6045 MDB_dbi dbi = mc->mc_dbi;
6046 unsigned i = mc->mc_top;
6047 MDB_page *mp = mc->mc_pg[i];
6049 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6050 if (mc->mc_flags & C_SUB)
6051 m3 = &m2->mc_xcursor->mx_cursor;
6054 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6055 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6062 if (rc != MDB_SUCCESS)
6063 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6065 /* Now store the actual data in the child DB. Note that we're
6066 * storing the user data in the keys field, so there are strict
6067 * size limits on dupdata. The actual data fields of the child
6068 * DB are all zero size.
6075 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6076 if (flags & MDB_CURRENT) {
6077 xflags = MDB_CURRENT|MDB_NOSPILL;
6079 mdb_xcursor_init1(mc, leaf);
6080 xflags = (flags & MDB_NODUPDATA) ?
6081 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6083 /* converted, write the original data first */
6085 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6089 /* Adjust other cursors pointing to mp */
6091 unsigned i = mc->mc_top;
6092 MDB_page *mp = mc->mc_pg[i];
6094 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6095 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6096 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6097 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6098 mdb_xcursor_init1(m2, leaf);
6102 /* we've done our job */
6105 if (flags & MDB_APPENDDUP)
6106 xflags |= MDB_APPEND;
6107 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6108 if (flags & F_SUBDATA) {
6109 void *db = NODEDATA(leaf);
6110 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6113 /* sub-writes might have failed so check rc again.
6114 * Don't increment count if we just replaced an existing item.
6116 if (!rc && !(flags & MDB_CURRENT))
6117 mc->mc_db->md_entries++;
6118 if (flags & MDB_MULTIPLE) {
6122 /* let caller know how many succeeded, if any */
6123 data[1].mv_size = mcount;
6124 if (mcount < dcount) {
6125 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6132 /* If we succeeded and the key didn't exist before, make sure
6133 * the cursor is marked valid.
6136 mc->mc_flags |= C_INITIALIZED;
6141 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6147 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6148 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6150 if (!(mc->mc_flags & C_INITIALIZED))
6153 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6154 return MDB_NOTFOUND;
6156 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6159 rc = mdb_cursor_touch(mc);
6163 mp = mc->mc_pg[mc->mc_top];
6164 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6166 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6167 if (!(flags & MDB_NODUPDATA)) {
6168 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6169 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6171 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6172 /* If sub-DB still has entries, we're done */
6173 if (mc->mc_xcursor->mx_db.md_entries) {
6174 if (leaf->mn_flags & F_SUBDATA) {
6175 /* update subDB info */
6176 void *db = NODEDATA(leaf);
6177 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6180 /* shrink fake page */
6181 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6182 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6183 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6184 /* fix other sub-DB cursors pointed at this fake page */
6185 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6186 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6187 if (m2->mc_pg[mc->mc_top] == mp &&
6188 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6189 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6192 mc->mc_db->md_entries--;
6193 mc->mc_flags |= C_DEL;
6196 /* otherwise fall thru and delete the sub-DB */
6199 if (leaf->mn_flags & F_SUBDATA) {
6200 /* add all the child DB's pages to the free list */
6201 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6202 if (rc == MDB_SUCCESS) {
6203 mc->mc_db->md_entries -=
6204 mc->mc_xcursor->mx_db.md_entries;
6209 return mdb_cursor_del0(mc, leaf);
6212 /** Allocate and initialize new pages for a database.
6213 * @param[in] mc a cursor on the database being added to.
6214 * @param[in] flags flags defining what type of page is being allocated.
6215 * @param[in] num the number of pages to allocate. This is usually 1,
6216 * unless allocating overflow pages for a large record.
6217 * @param[out] mp Address of a page, or NULL on failure.
6218 * @return 0 on success, non-zero on failure.
6221 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6226 if ((rc = mdb_page_alloc(mc, num, &np)))
6228 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6229 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6230 np->mp_flags = flags | P_DIRTY;
6231 np->mp_lower = PAGEHDRSZ;
6232 np->mp_upper = mc->mc_txn->mt_env->me_psize;
6235 mc->mc_db->md_branch_pages++;
6236 else if (IS_LEAF(np))
6237 mc->mc_db->md_leaf_pages++;
6238 else if (IS_OVERFLOW(np)) {
6239 mc->mc_db->md_overflow_pages += num;
6247 /** Calculate the size of a leaf node.
6248 * The size depends on the environment's page size; if a data item
6249 * is too large it will be put onto an overflow page and the node
6250 * size will only include the key and not the data. Sizes are always
6251 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6252 * of the #MDB_node headers.
6253 * @param[in] env The environment handle.
6254 * @param[in] key The key for the node.
6255 * @param[in] data The data for the node.
6256 * @return The number of bytes needed to store the node.
6259 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6263 sz = LEAFSIZE(key, data);
6264 if (sz > env->me_nodemax) {
6265 /* put on overflow page */
6266 sz -= data->mv_size - sizeof(pgno_t);
6269 return EVEN(sz + sizeof(indx_t));
6272 /** Calculate the size of a branch node.
6273 * The size should depend on the environment's page size but since
6274 * we currently don't support spilling large keys onto overflow
6275 * pages, it's simply the size of the #MDB_node header plus the
6276 * size of the key. Sizes are always rounded up to an even number
6277 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6278 * @param[in] env The environment handle.
6279 * @param[in] key The key for the node.
6280 * @return The number of bytes needed to store the node.
6283 mdb_branch_size(MDB_env *env, MDB_val *key)
6288 if (sz > env->me_nodemax) {
6289 /* put on overflow page */
6290 /* not implemented */
6291 /* sz -= key->size - sizeof(pgno_t); */
6294 return sz + sizeof(indx_t);
6297 /** Add a node to the page pointed to by the cursor.
6298 * @param[in] mc The cursor for this operation.
6299 * @param[in] indx The index on the page where the new node should be added.
6300 * @param[in] key The key for the new node.
6301 * @param[in] data The data for the new node, if any.
6302 * @param[in] pgno The page number, if adding a branch node.
6303 * @param[in] flags Flags for the node.
6304 * @return 0 on success, non-zero on failure. Possible errors are:
6306 * <li>ENOMEM - failed to allocate overflow pages for the node.
6307 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6308 * should never happen since all callers already calculate the
6309 * page's free space before calling this function.
6313 mdb_node_add(MDB_cursor *mc, indx_t indx,
6314 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6317 size_t node_size = NODESIZE;
6321 MDB_page *mp = mc->mc_pg[mc->mc_top];
6322 MDB_page *ofp = NULL; /* overflow page */
6325 assert(mp->mp_upper >= mp->mp_lower);
6327 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6328 IS_LEAF(mp) ? "leaf" : "branch",
6329 IS_SUBP(mp) ? "sub-" : "",
6330 mp->mp_pgno, indx, data ? data->mv_size : 0,
6331 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6334 /* Move higher keys up one slot. */
6335 int ksize = mc->mc_db->md_pad, dif;
6336 char *ptr = LEAF2KEY(mp, indx, ksize);
6337 dif = NUMKEYS(mp) - indx;
6339 memmove(ptr+ksize, ptr, dif*ksize);
6340 /* insert new key */
6341 memcpy(ptr, key->mv_data, ksize);
6343 /* Just using these for counting */
6344 mp->mp_lower += sizeof(indx_t);
6345 mp->mp_upper -= ksize - sizeof(indx_t);
6349 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6351 node_size += key->mv_size;
6354 if (F_ISSET(flags, F_BIGDATA)) {
6355 /* Data already on overflow page. */
6356 node_size += sizeof(pgno_t);
6357 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6358 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6360 /* Put data on overflow page. */
6361 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6362 data->mv_size, node_size+data->mv_size));
6363 node_size = EVEN(node_size + sizeof(pgno_t));
6364 if ((ssize_t)node_size > room)
6366 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6368 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6372 node_size += data->mv_size;
6375 node_size = EVEN(node_size);
6376 if ((ssize_t)node_size > room)
6380 /* Move higher pointers up one slot. */
6381 for (i = NUMKEYS(mp); i > indx; i--)
6382 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6384 /* Adjust free space offsets. */
6385 ofs = mp->mp_upper - node_size;
6386 assert(ofs >= mp->mp_lower + sizeof(indx_t));
6387 mp->mp_ptrs[indx] = ofs;
6389 mp->mp_lower += sizeof(indx_t);
6391 /* Write the node data. */
6392 node = NODEPTR(mp, indx);
6393 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6394 node->mn_flags = flags;
6396 SETDSZ(node,data->mv_size);
6401 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6406 if (F_ISSET(flags, F_BIGDATA))
6407 memcpy(node->mn_data + key->mv_size, data->mv_data,
6409 else if (F_ISSET(flags, MDB_RESERVE))
6410 data->mv_data = node->mn_data + key->mv_size;
6412 memcpy(node->mn_data + key->mv_size, data->mv_data,
6415 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6417 if (F_ISSET(flags, MDB_RESERVE))
6418 data->mv_data = METADATA(ofp);
6420 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6427 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6428 mp->mp_pgno, NUMKEYS(mp)));
6429 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6430 DPRINTF(("node size = %"Z"u", node_size));
6431 return MDB_PAGE_FULL;
6434 /** Delete the specified node from a page.
6435 * @param[in] mp The page to operate on.
6436 * @param[in] indx The index of the node to delete.
6437 * @param[in] ksize The size of a node. Only used if the page is
6438 * part of a #MDB_DUPFIXED database.
6441 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
6444 indx_t i, j, numkeys, ptr;
6451 COPY_PGNO(pgno, mp->mp_pgno);
6452 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6453 IS_LEAF(mp) ? "leaf" : "branch", pgno));
6456 assert(indx < NUMKEYS(mp));
6459 int x = NUMKEYS(mp) - 1 - indx;
6460 base = LEAF2KEY(mp, indx, ksize);
6462 memmove(base, base + ksize, x * ksize);
6463 mp->mp_lower -= sizeof(indx_t);
6464 mp->mp_upper += ksize - sizeof(indx_t);
6468 node = NODEPTR(mp, indx);
6469 sz = NODESIZE + node->mn_ksize;
6471 if (F_ISSET(node->mn_flags, F_BIGDATA))
6472 sz += sizeof(pgno_t);
6474 sz += NODEDSZ(node);
6478 ptr = mp->mp_ptrs[indx];
6479 numkeys = NUMKEYS(mp);
6480 for (i = j = 0; i < numkeys; i++) {
6482 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6483 if (mp->mp_ptrs[i] < ptr)
6484 mp->mp_ptrs[j] += sz;
6489 base = (char *)mp + mp->mp_upper;
6490 memmove(base + sz, base, ptr - mp->mp_upper);
6492 mp->mp_lower -= sizeof(indx_t);
6496 /** Compact the main page after deleting a node on a subpage.
6497 * @param[in] mp The main page to operate on.
6498 * @param[in] indx The index of the subpage on the main page.
6501 mdb_node_shrink(MDB_page *mp, indx_t indx)
6507 indx_t i, numkeys, ptr;
6509 node = NODEPTR(mp, indx);
6510 sp = (MDB_page *)NODEDATA(node);
6511 delta = SIZELEFT(sp);
6512 xp = (MDB_page *)((char *)sp + delta);
6514 /* shift subpage upward */
6516 nsize = NUMKEYS(sp) * sp->mp_pad;
6518 return; /* do not make the node uneven-sized */
6519 memmove(METADATA(xp), METADATA(sp), nsize);
6522 numkeys = NUMKEYS(sp);
6523 for (i=numkeys-1; i>=0; i--)
6524 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6526 xp->mp_upper = sp->mp_lower;
6527 xp->mp_lower = sp->mp_lower;
6528 xp->mp_flags = sp->mp_flags;
6529 xp->mp_pad = sp->mp_pad;
6530 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6532 nsize = NODEDSZ(node) - delta;
6533 SETDSZ(node, nsize);
6535 /* shift lower nodes upward */
6536 ptr = mp->mp_ptrs[indx];
6537 numkeys = NUMKEYS(mp);
6538 for (i = 0; i < numkeys; i++) {
6539 if (mp->mp_ptrs[i] <= ptr)
6540 mp->mp_ptrs[i] += delta;
6543 base = (char *)mp + mp->mp_upper;
6544 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6545 mp->mp_upper += delta;
6548 /** Initial setup of a sorted-dups cursor.
6549 * Sorted duplicates are implemented as a sub-database for the given key.
6550 * The duplicate data items are actually keys of the sub-database.
6551 * Operations on the duplicate data items are performed using a sub-cursor
6552 * initialized when the sub-database is first accessed. This function does
6553 * the preliminary setup of the sub-cursor, filling in the fields that
6554 * depend only on the parent DB.
6555 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6558 mdb_xcursor_init0(MDB_cursor *mc)
6560 MDB_xcursor *mx = mc->mc_xcursor;
6562 mx->mx_cursor.mc_xcursor = NULL;
6563 mx->mx_cursor.mc_txn = mc->mc_txn;
6564 mx->mx_cursor.mc_db = &mx->mx_db;
6565 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6566 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6567 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6568 mx->mx_cursor.mc_snum = 0;
6569 mx->mx_cursor.mc_top = 0;
6570 mx->mx_cursor.mc_flags = C_SUB;
6571 mx->mx_dbx.md_name.mv_size = 0;
6572 mx->mx_dbx.md_name.mv_data = NULL;
6573 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6574 mx->mx_dbx.md_dcmp = NULL;
6575 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6578 /** Final setup of a sorted-dups cursor.
6579 * Sets up the fields that depend on the data from the main cursor.
6580 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6581 * @param[in] node The data containing the #MDB_db record for the
6582 * sorted-dup database.
6585 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6587 MDB_xcursor *mx = mc->mc_xcursor;
6589 if (node->mn_flags & F_SUBDATA) {
6590 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6591 mx->mx_cursor.mc_pg[0] = 0;
6592 mx->mx_cursor.mc_snum = 0;
6593 mx->mx_cursor.mc_top = 0;
6594 mx->mx_cursor.mc_flags = C_SUB;
6596 MDB_page *fp = NODEDATA(node);
6597 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6598 mx->mx_db.md_flags = 0;
6599 mx->mx_db.md_depth = 1;
6600 mx->mx_db.md_branch_pages = 0;
6601 mx->mx_db.md_leaf_pages = 1;
6602 mx->mx_db.md_overflow_pages = 0;
6603 mx->mx_db.md_entries = NUMKEYS(fp);
6604 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6605 mx->mx_cursor.mc_snum = 1;
6606 mx->mx_cursor.mc_top = 0;
6607 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6608 mx->mx_cursor.mc_pg[0] = fp;
6609 mx->mx_cursor.mc_ki[0] = 0;
6610 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6611 mx->mx_db.md_flags = MDB_DUPFIXED;
6612 mx->mx_db.md_pad = fp->mp_pad;
6613 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6614 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6617 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6618 mx->mx_db.md_root));
6619 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6620 #if UINT_MAX < SIZE_MAX
6621 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6622 #ifdef MISALIGNED_OK
6623 mx->mx_dbx.md_cmp = mdb_cmp_long;
6625 mx->mx_dbx.md_cmp = mdb_cmp_cint;
6630 /** Initialize a cursor for a given transaction and database. */
6632 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6635 mc->mc_backup = NULL;
6638 mc->mc_db = &txn->mt_dbs[dbi];
6639 mc->mc_dbx = &txn->mt_dbxs[dbi];
6640 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6645 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6647 mc->mc_xcursor = mx;
6648 mdb_xcursor_init0(mc);
6650 mc->mc_xcursor = NULL;
6652 if (*mc->mc_dbflag & DB_STALE) {
6653 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6658 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6661 size_t size = sizeof(MDB_cursor);
6663 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6666 if (txn->mt_flags & MDB_TXN_ERROR)
6669 /* Allow read access to the freelist */
6670 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6673 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6674 size += sizeof(MDB_xcursor);
6676 if ((mc = malloc(size)) != NULL) {
6677 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6678 if (txn->mt_cursors) {
6679 mc->mc_next = txn->mt_cursors[dbi];
6680 txn->mt_cursors[dbi] = mc;
6681 mc->mc_flags |= C_UNTRACK;
6693 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6695 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
6698 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6701 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6705 /* Return the count of duplicate data items for the current key */
6707 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6711 if (mc == NULL || countp == NULL)
6714 if (mc->mc_xcursor == NULL)
6715 return MDB_INCOMPATIBLE;
6717 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6718 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6721 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6724 *countp = mc->mc_xcursor->mx_db.md_entries;
6730 mdb_cursor_close(MDB_cursor *mc)
6732 if (mc && !mc->mc_backup) {
6733 /* remove from txn, if tracked */
6734 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6735 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6736 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6738 *prev = mc->mc_next;
6745 mdb_cursor_txn(MDB_cursor *mc)
6747 if (!mc) return NULL;
6752 mdb_cursor_dbi(MDB_cursor *mc)
6758 /** Replace the key for a node with a new key.
6759 * @param[in] mc Cursor pointing to the node to operate on.
6760 * @param[in] key The new key to use.
6761 * @return 0 on success, non-zero on failure.
6764 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6770 int delta, ksize, oksize;
6771 indx_t ptr, i, numkeys, indx;
6774 indx = mc->mc_ki[mc->mc_top];
6775 mp = mc->mc_pg[mc->mc_top];
6776 node = NODEPTR(mp, indx);
6777 ptr = mp->mp_ptrs[indx];
6781 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6782 k2.mv_data = NODEKEY(node);
6783 k2.mv_size = node->mn_ksize;
6784 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6786 mdb_dkey(&k2, kbuf2),
6792 /* Sizes must be 2-byte aligned. */
6793 ksize = EVEN(key->mv_size);
6794 oksize = EVEN(node->mn_ksize);
6795 delta = ksize - oksize;
6797 /* Shift node contents if EVEN(key length) changed. */
6799 if (delta > 0 && SIZELEFT(mp) < delta) {
6801 /* not enough space left, do a delete and split */
6802 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6803 pgno = NODEPGNO(node);
6804 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6805 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6808 numkeys = NUMKEYS(mp);
6809 for (i = 0; i < numkeys; i++) {
6810 if (mp->mp_ptrs[i] <= ptr)
6811 mp->mp_ptrs[i] -= delta;
6814 base = (char *)mp + mp->mp_upper;
6815 len = ptr - mp->mp_upper + NODESIZE;
6816 memmove(base - delta, base, len);
6817 mp->mp_upper -= delta;
6819 node = NODEPTR(mp, indx);
6822 /* But even if no shift was needed, update ksize */
6823 if (node->mn_ksize != key->mv_size)
6824 node->mn_ksize = key->mv_size;
6827 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6833 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6835 /** Move a node from csrc to cdst.
6838 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6845 unsigned short flags;
6849 /* Mark src and dst as dirty. */
6850 if ((rc = mdb_page_touch(csrc)) ||
6851 (rc = mdb_page_touch(cdst)))
6854 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6855 key.mv_size = csrc->mc_db->md_pad;
6856 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6858 data.mv_data = NULL;
6862 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6863 assert(!((size_t)srcnode&1));
6864 srcpg = NODEPGNO(srcnode);
6865 flags = srcnode->mn_flags;
6866 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6867 unsigned int snum = csrc->mc_snum;
6869 /* must find the lowest key below src */
6870 mdb_page_search_lowest(csrc);
6871 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6872 key.mv_size = csrc->mc_db->md_pad;
6873 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6875 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6876 key.mv_size = NODEKSZ(s2);
6877 key.mv_data = NODEKEY(s2);
6879 csrc->mc_snum = snum--;
6880 csrc->mc_top = snum;
6882 key.mv_size = NODEKSZ(srcnode);
6883 key.mv_data = NODEKEY(srcnode);
6885 data.mv_size = NODEDSZ(srcnode);
6886 data.mv_data = NODEDATA(srcnode);
6888 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6889 unsigned int snum = cdst->mc_snum;
6892 /* must find the lowest key below dst */
6893 mdb_page_search_lowest(cdst);
6894 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6895 bkey.mv_size = cdst->mc_db->md_pad;
6896 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6898 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6899 bkey.mv_size = NODEKSZ(s2);
6900 bkey.mv_data = NODEKEY(s2);
6902 cdst->mc_snum = snum--;
6903 cdst->mc_top = snum;
6904 mdb_cursor_copy(cdst, &mn);
6906 rc = mdb_update_key(&mn, &bkey);
6911 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
6912 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6913 csrc->mc_ki[csrc->mc_top],
6915 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6916 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
6918 /* Add the node to the destination page.
6920 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6921 if (rc != MDB_SUCCESS)
6924 /* Delete the node from the source page.
6926 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6929 /* Adjust other cursors pointing to mp */
6930 MDB_cursor *m2, *m3;
6931 MDB_dbi dbi = csrc->mc_dbi;
6932 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6934 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6935 if (csrc->mc_flags & C_SUB)
6936 m3 = &m2->mc_xcursor->mx_cursor;
6939 if (m3 == csrc) continue;
6940 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6941 csrc->mc_ki[csrc->mc_top]) {
6942 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6943 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6948 /* Update the parent separators.
6950 if (csrc->mc_ki[csrc->mc_top] == 0) {
6951 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6952 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6953 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6955 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6956 key.mv_size = NODEKSZ(srcnode);
6957 key.mv_data = NODEKEY(srcnode);
6959 DPRINTF(("update separator for source page %"Z"u to [%s]",
6960 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
6961 mdb_cursor_copy(csrc, &mn);
6964 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6967 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6969 indx_t ix = csrc->mc_ki[csrc->mc_top];
6970 nullkey.mv_size = 0;
6971 csrc->mc_ki[csrc->mc_top] = 0;
6972 rc = mdb_update_key(csrc, &nullkey);
6973 csrc->mc_ki[csrc->mc_top] = ix;
6974 assert(rc == MDB_SUCCESS);
6978 if (cdst->mc_ki[cdst->mc_top] == 0) {
6979 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6980 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6981 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6983 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6984 key.mv_size = NODEKSZ(srcnode);
6985 key.mv_data = NODEKEY(srcnode);
6987 DPRINTF(("update separator for destination page %"Z"u to [%s]",
6988 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
6989 mdb_cursor_copy(cdst, &mn);
6992 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6995 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6997 indx_t ix = cdst->mc_ki[cdst->mc_top];
6998 nullkey.mv_size = 0;
6999 cdst->mc_ki[cdst->mc_top] = 0;
7000 rc = mdb_update_key(cdst, &nullkey);
7001 cdst->mc_ki[cdst->mc_top] = ix;
7002 assert(rc == MDB_SUCCESS);
7009 /** Merge one page into another.
7010 * The nodes from the page pointed to by \b csrc will
7011 * be copied to the page pointed to by \b cdst and then
7012 * the \b csrc page will be freed.
7013 * @param[in] csrc Cursor pointing to the source page.
7014 * @param[in] cdst Cursor pointing to the destination page.
7017 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7025 DPRINTF(("merging page %"Z"u into %"Z"u", csrc->mc_pg[csrc->mc_top]->mp_pgno,
7026 cdst->mc_pg[cdst->mc_top]->mp_pgno));
7028 assert(csrc->mc_snum > 1); /* can't merge root page */
7029 assert(cdst->mc_snum > 1);
7031 /* Mark dst as dirty. */
7032 if ((rc = mdb_page_touch(cdst)))
7035 /* Move all nodes from src to dst.
7037 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
7038 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7039 key.mv_size = csrc->mc_db->md_pad;
7040 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
7041 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7042 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7043 if (rc != MDB_SUCCESS)
7045 key.mv_data = (char *)key.mv_data + key.mv_size;
7048 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
7049 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
7050 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7051 unsigned int snum = csrc->mc_snum;
7053 /* must find the lowest key below src */
7054 mdb_page_search_lowest(csrc);
7055 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7056 key.mv_size = csrc->mc_db->md_pad;
7057 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7059 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7060 key.mv_size = NODEKSZ(s2);
7061 key.mv_data = NODEKEY(s2);
7063 csrc->mc_snum = snum--;
7064 csrc->mc_top = snum;
7066 key.mv_size = srcnode->mn_ksize;
7067 key.mv_data = NODEKEY(srcnode);
7070 data.mv_size = NODEDSZ(srcnode);
7071 data.mv_data = NODEDATA(srcnode);
7072 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7073 if (rc != MDB_SUCCESS)
7078 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7079 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]),
7080 (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10));
7082 /* Unlink the src page from parent and add to free list.
7084 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
7085 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
7088 rc = mdb_update_key(csrc, &key);
7094 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
7095 csrc->mc_pg[csrc->mc_top]->mp_pgno);
7098 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
7099 csrc->mc_db->md_leaf_pages--;
7101 csrc->mc_db->md_branch_pages--;
7103 /* Adjust other cursors pointing to mp */
7104 MDB_cursor *m2, *m3;
7105 MDB_dbi dbi = csrc->mc_dbi;
7106 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
7108 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7109 if (csrc->mc_flags & C_SUB)
7110 m3 = &m2->mc_xcursor->mx_cursor;
7113 if (m3 == csrc) continue;
7114 if (m3->mc_snum < csrc->mc_snum) continue;
7115 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
7116 m3->mc_pg[csrc->mc_top] = mp;
7117 m3->mc_ki[csrc->mc_top] += nkeys;
7121 mdb_cursor_pop(csrc);
7123 return mdb_rebalance(csrc);
7126 /** Copy the contents of a cursor.
7127 * @param[in] csrc The cursor to copy from.
7128 * @param[out] cdst The cursor to copy to.
7131 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7135 cdst->mc_txn = csrc->mc_txn;
7136 cdst->mc_dbi = csrc->mc_dbi;
7137 cdst->mc_db = csrc->mc_db;
7138 cdst->mc_dbx = csrc->mc_dbx;
7139 cdst->mc_snum = csrc->mc_snum;
7140 cdst->mc_top = csrc->mc_top;
7141 cdst->mc_flags = csrc->mc_flags;
7143 for (i=0; i<csrc->mc_snum; i++) {
7144 cdst->mc_pg[i] = csrc->mc_pg[i];
7145 cdst->mc_ki[i] = csrc->mc_ki[i];
7149 /** Rebalance the tree after a delete operation.
7150 * @param[in] mc Cursor pointing to the page where rebalancing
7152 * @return 0 on success, non-zero on failure.
7155 mdb_rebalance(MDB_cursor *mc)
7159 unsigned int ptop, minkeys;
7162 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7166 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7167 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7168 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7169 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]),
7170 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7174 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7175 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7178 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
7179 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7185 if (mc->mc_snum < 2) {
7186 MDB_page *mp = mc->mc_pg[0];
7188 DPUTS("Can't rebalance a subpage, ignoring");
7191 if (NUMKEYS(mp) == 0) {
7192 DPUTS("tree is completely empty");
7193 mc->mc_db->md_root = P_INVALID;
7194 mc->mc_db->md_depth = 0;
7195 mc->mc_db->md_leaf_pages = 0;
7196 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7199 /* Adjust cursors pointing to mp */
7202 mc->mc_flags &= ~C_INITIALIZED;
7204 MDB_cursor *m2, *m3;
7205 MDB_dbi dbi = mc->mc_dbi;
7207 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7208 if (mc->mc_flags & C_SUB)
7209 m3 = &m2->mc_xcursor->mx_cursor;
7212 if (m3->mc_snum < mc->mc_snum) continue;
7213 if (m3->mc_pg[0] == mp) {
7216 m3->mc_flags &= ~C_INITIALIZED;
7220 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7221 DPUTS("collapsing root page!");
7222 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7225 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7226 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7229 mc->mc_db->md_depth--;
7230 mc->mc_db->md_branch_pages--;
7231 mc->mc_ki[0] = mc->mc_ki[1];
7233 /* Adjust other cursors pointing to mp */
7234 MDB_cursor *m2, *m3;
7235 MDB_dbi dbi = mc->mc_dbi;
7237 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7238 if (mc->mc_flags & C_SUB)
7239 m3 = &m2->mc_xcursor->mx_cursor;
7242 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7243 if (m3->mc_pg[0] == mp) {
7247 for (i=0; i<m3->mc_snum; i++) {
7248 m3->mc_pg[i] = m3->mc_pg[i+1];
7249 m3->mc_ki[i] = m3->mc_ki[i+1];
7255 DPUTS("root page doesn't need rebalancing");
7259 /* The parent (branch page) must have at least 2 pointers,
7260 * otherwise the tree is invalid.
7262 ptop = mc->mc_top-1;
7263 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
7265 /* Leaf page fill factor is below the threshold.
7266 * Try to move keys from left or right neighbor, or
7267 * merge with a neighbor page.
7272 mdb_cursor_copy(mc, &mn);
7273 mn.mc_xcursor = NULL;
7275 if (mc->mc_ki[ptop] == 0) {
7276 /* We're the leftmost leaf in our parent.
7278 DPUTS("reading right neighbor");
7280 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7281 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7284 mn.mc_ki[mn.mc_top] = 0;
7285 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7287 /* There is at least one neighbor to the left.
7289 DPUTS("reading left neighbor");
7291 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7292 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7295 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7296 mc->mc_ki[mc->mc_top] = 0;
7299 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7300 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7301 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7303 /* If the neighbor page is above threshold and has enough keys,
7304 * move one key from it. Otherwise we should try to merge them.
7305 * (A branch page must never have less than 2 keys.)
7307 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7308 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
7309 return mdb_node_move(&mn, mc);
7311 if (mc->mc_ki[ptop] == 0)
7312 rc = mdb_page_merge(&mn, mc);
7314 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7315 rc = mdb_page_merge(mc, &mn);
7316 mdb_cursor_copy(&mn, mc);
7318 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
7323 /** Complete a delete operation started by #mdb_cursor_del(). */
7325 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
7332 mp = mc->mc_pg[mc->mc_top];
7333 ki = mc->mc_ki[mc->mc_top];
7335 /* add overflow pages to free list */
7336 if (!IS_LEAF2(mp) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
7340 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
7341 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
7342 (rc = mdb_ovpage_free(mc, omp)))
7345 mdb_node_del(mp, ki, mc->mc_db->md_pad);
7346 mc->mc_db->md_entries--;
7347 rc = mdb_rebalance(mc);
7348 if (rc != MDB_SUCCESS)
7349 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7352 MDB_dbi dbi = mc->mc_dbi;
7354 mp = mc->mc_pg[mc->mc_top];
7355 nkeys = NUMKEYS(mp);
7357 /* if mc points past last node in page, find next sibling */
7358 if (mc->mc_ki[mc->mc_top] >= nkeys)
7359 mdb_cursor_sibling(mc, 1);
7361 /* Adjust other cursors pointing to mp */
7362 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7363 if (m2 == mc || m2->mc_snum < mc->mc_snum)
7365 if (!(m2->mc_flags & C_INITIALIZED))
7367 if (m2->mc_pg[mc->mc_top] == mp) {
7368 if (m2->mc_ki[mc->mc_top] >= ki) {
7369 m2->mc_flags |= C_DEL;
7370 if (m2->mc_ki[mc->mc_top] > ki)
7371 m2->mc_ki[mc->mc_top]--;
7373 if (m2->mc_ki[mc->mc_top] >= nkeys)
7374 mdb_cursor_sibling(m2, 1);
7377 mc->mc_flags |= C_DEL;
7384 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7385 MDB_val *key, MDB_val *data)
7390 MDB_val rdata, *xdata;
7394 assert(key != NULL);
7396 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7398 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7401 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7402 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7404 mdb_cursor_init(&mc, txn, dbi, &mx);
7407 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7408 /* must ignore any data */
7419 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7421 /* let mdb_page_split know about this cursor if needed:
7422 * delete will trigger a rebalance; if it needs to move
7423 * a node from one page to another, it will have to
7424 * update the parent's separator key(s). If the new sepkey
7425 * is larger than the current one, the parent page may
7426 * run out of space, triggering a split. We need this
7427 * cursor to be consistent until the end of the rebalance.
7429 mc.mc_flags |= C_UNTRACK;
7430 mc.mc_next = txn->mt_cursors[dbi];
7431 txn->mt_cursors[dbi] = &mc;
7432 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
7433 txn->mt_cursors[dbi] = mc.mc_next;
7438 /** Split a page and insert a new node.
7439 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7440 * The cursor will be updated to point to the actual page and index where
7441 * the node got inserted after the split.
7442 * @param[in] newkey The key for the newly inserted node.
7443 * @param[in] newdata The data for the newly inserted node.
7444 * @param[in] newpgno The page number, if the new node is a branch node.
7445 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7446 * @return 0 on success, non-zero on failure.
7449 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7450 unsigned int nflags)
7453 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7456 int i, j, split_indx, nkeys, pmax;
7457 MDB_env *env = mc->mc_txn->mt_env;
7459 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7460 MDB_page *copy = NULL;
7461 MDB_page *mp, *rp, *pp;
7466 mp = mc->mc_pg[mc->mc_top];
7467 newindx = mc->mc_ki[mc->mc_top];
7468 nkeys = NUMKEYS(mp);
7470 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7471 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7472 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7474 /* Create a right sibling. */
7475 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7477 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7479 if (mc->mc_snum < 2) {
7480 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7482 /* shift current top to make room for new parent */
7483 mc->mc_pg[1] = mc->mc_pg[0];
7484 mc->mc_ki[1] = mc->mc_ki[0];
7487 mc->mc_db->md_root = pp->mp_pgno;
7488 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7489 mc->mc_db->md_depth++;
7492 /* Add left (implicit) pointer. */
7493 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7494 /* undo the pre-push */
7495 mc->mc_pg[0] = mc->mc_pg[1];
7496 mc->mc_ki[0] = mc->mc_ki[1];
7497 mc->mc_db->md_root = mp->mp_pgno;
7498 mc->mc_db->md_depth--;
7505 ptop = mc->mc_top-1;
7506 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7509 mc->mc_flags |= C_SPLITTING;
7510 mdb_cursor_copy(mc, &mn);
7511 mn.mc_pg[mn.mc_top] = rp;
7512 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7514 if (nflags & MDB_APPEND) {
7515 mn.mc_ki[mn.mc_top] = 0;
7517 split_indx = newindx;
7521 split_indx = (nkeys+1) / 2;
7526 unsigned int lsize, rsize, ksize;
7527 /* Move half of the keys to the right sibling */
7529 x = mc->mc_ki[mc->mc_top] - split_indx;
7530 ksize = mc->mc_db->md_pad;
7531 split = LEAF2KEY(mp, split_indx, ksize);
7532 rsize = (nkeys - split_indx) * ksize;
7533 lsize = (nkeys - split_indx) * sizeof(indx_t);
7534 mp->mp_lower -= lsize;
7535 rp->mp_lower += lsize;
7536 mp->mp_upper += rsize - lsize;
7537 rp->mp_upper -= rsize - lsize;
7538 sepkey.mv_size = ksize;
7539 if (newindx == split_indx) {
7540 sepkey.mv_data = newkey->mv_data;
7542 sepkey.mv_data = split;
7545 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7546 memcpy(rp->mp_ptrs, split, rsize);
7547 sepkey.mv_data = rp->mp_ptrs;
7548 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7549 memcpy(ins, newkey->mv_data, ksize);
7550 mp->mp_lower += sizeof(indx_t);
7551 mp->mp_upper -= ksize - sizeof(indx_t);
7554 memcpy(rp->mp_ptrs, split, x * ksize);
7555 ins = LEAF2KEY(rp, x, ksize);
7556 memcpy(ins, newkey->mv_data, ksize);
7557 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7558 rp->mp_lower += sizeof(indx_t);
7559 rp->mp_upper -= ksize - sizeof(indx_t);
7560 mc->mc_ki[mc->mc_top] = x;
7561 mc->mc_pg[mc->mc_top] = rp;
7564 int psize, nsize, k;
7565 /* Maximum free space in an empty page */
7566 pmax = env->me_psize - PAGEHDRSZ;
7568 nsize = mdb_leaf_size(env, newkey, newdata);
7570 nsize = mdb_branch_size(env, newkey);
7571 nsize = EVEN(nsize);
7573 /* grab a page to hold a temporary copy */
7574 copy = mdb_page_malloc(mc->mc_txn, 1);
7577 copy->mp_pgno = mp->mp_pgno;
7578 copy->mp_flags = mp->mp_flags;
7579 copy->mp_lower = PAGEHDRSZ;
7580 copy->mp_upper = env->me_psize;
7582 /* prepare to insert */
7583 for (i=0, j=0; i<nkeys; i++) {
7585 copy->mp_ptrs[j++] = 0;
7587 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7590 /* When items are relatively large the split point needs
7591 * to be checked, because being off-by-one will make the
7592 * difference between success or failure in mdb_node_add.
7594 * It's also relevant if a page happens to be laid out
7595 * such that one half of its nodes are all "small" and
7596 * the other half of its nodes are "large." If the new
7597 * item is also "large" and falls on the half with
7598 * "large" nodes, it also may not fit.
7600 * As a final tweak, if the new item goes on the last
7601 * spot on the page (and thus, onto the new page), bias
7602 * the split so the new page is emptier than the old page.
7603 * This yields better packing during sequential inserts.
7605 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7606 /* Find split point */
7608 if (newindx <= split_indx || newindx >= nkeys) {
7610 k = newindx >= nkeys ? nkeys : split_indx+2;
7615 for (; i!=k; i+=j) {
7620 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7621 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7623 if (F_ISSET(node->mn_flags, F_BIGDATA))
7624 psize += sizeof(pgno_t);
7626 psize += NODEDSZ(node);
7628 psize = EVEN(psize);
7630 if (psize > pmax || i == k-j) {
7631 split_indx = i + (j<0);
7636 if (split_indx == newindx) {
7637 sepkey.mv_size = newkey->mv_size;
7638 sepkey.mv_data = newkey->mv_data;
7640 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx]);
7641 sepkey.mv_size = node->mn_ksize;
7642 sepkey.mv_data = NODEKEY(node);
7647 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7649 /* Copy separator key to the parent.
7651 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7655 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7658 if (mn.mc_snum == mc->mc_snum) {
7659 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7660 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7661 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7662 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7667 /* Right page might now have changed parent.
7668 * Check if left page also changed parent.
7670 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7671 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7672 for (i=0; i<ptop; i++) {
7673 mc->mc_pg[i] = mn.mc_pg[i];
7674 mc->mc_ki[i] = mn.mc_ki[i];
7676 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7677 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7681 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7684 mc->mc_flags ^= C_SPLITTING;
7685 if (rc != MDB_SUCCESS) {
7688 if (nflags & MDB_APPEND) {
7689 mc->mc_pg[mc->mc_top] = rp;
7690 mc->mc_ki[mc->mc_top] = 0;
7691 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7694 for (i=0; i<mc->mc_top; i++)
7695 mc->mc_ki[i] = mn.mc_ki[i];
7696 } else if (!IS_LEAF2(mp)) {
7698 mc->mc_pg[mc->mc_top] = rp;
7703 rkey.mv_data = newkey->mv_data;
7704 rkey.mv_size = newkey->mv_size;
7710 /* Update index for the new key. */
7711 mc->mc_ki[mc->mc_top] = j;
7713 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i]);
7714 rkey.mv_data = NODEKEY(node);
7715 rkey.mv_size = node->mn_ksize;
7717 xdata.mv_data = NODEDATA(node);
7718 xdata.mv_size = NODEDSZ(node);
7721 pgno = NODEPGNO(node);
7722 flags = node->mn_flags;
7725 if (!IS_LEAF(mp) && j == 0) {
7726 /* First branch index doesn't need key data. */
7730 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7732 /* return tmp page to freelist */
7733 mdb_page_free(env, copy);
7739 mc->mc_pg[mc->mc_top] = copy;
7744 } while (i != split_indx);
7746 nkeys = NUMKEYS(copy);
7747 for (i=0; i<nkeys; i++)
7748 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7749 mp->mp_lower = copy->mp_lower;
7750 mp->mp_upper = copy->mp_upper;
7751 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7752 env->me_psize - copy->mp_upper);
7754 /* reset back to original page */
7755 if (newindx < split_indx) {
7756 mc->mc_pg[mc->mc_top] = mp;
7757 if (nflags & MDB_RESERVE) {
7758 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7759 if (!(node->mn_flags & F_BIGDATA))
7760 newdata->mv_data = NODEDATA(node);
7763 mc->mc_pg[mc->mc_top] = rp;
7765 /* Make sure mc_ki is still valid.
7767 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7768 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7769 for (i=0; i<ptop; i++) {
7770 mc->mc_pg[i] = mn.mc_pg[i];
7771 mc->mc_ki[i] = mn.mc_ki[i];
7773 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7774 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7777 /* return tmp page to freelist */
7778 mdb_page_free(env, copy);
7782 /* Adjust other cursors pointing to mp */
7783 MDB_cursor *m2, *m3;
7784 MDB_dbi dbi = mc->mc_dbi;
7785 int fixup = NUMKEYS(mp);
7787 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7788 if (mc->mc_flags & C_SUB)
7789 m3 = &m2->mc_xcursor->mx_cursor;
7794 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7796 if (m3->mc_flags & C_SPLITTING)
7801 for (k=m3->mc_top; k>=0; k--) {
7802 m3->mc_ki[k+1] = m3->mc_ki[k];
7803 m3->mc_pg[k+1] = m3->mc_pg[k];
7805 if (m3->mc_ki[0] >= split_indx) {
7810 m3->mc_pg[0] = mc->mc_pg[0];
7814 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
7815 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7816 m3->mc_ki[mc->mc_top]++;
7817 if (m3->mc_ki[mc->mc_top] >= fixup) {
7818 m3->mc_pg[mc->mc_top] = rp;
7819 m3->mc_ki[mc->mc_top] -= fixup;
7820 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7822 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7823 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7828 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
7833 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7834 MDB_val *key, MDB_val *data, unsigned int flags)
7839 assert(key != NULL);
7840 assert(data != NULL);
7842 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7845 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7848 mdb_cursor_init(&mc, txn, dbi, &mx);
7849 return mdb_cursor_put(&mc, key, data, flags);
7853 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7855 if ((flag & CHANGEABLE) != flag)
7858 env->me_flags |= flag;
7860 env->me_flags &= ~flag;
7865 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7870 *arg = env->me_flags;
7875 mdb_env_get_path(MDB_env *env, const char **arg)
7880 *arg = env->me_path;
7885 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
7894 /** Common code for #mdb_stat() and #mdb_env_stat().
7895 * @param[in] env the environment to operate in.
7896 * @param[in] db the #MDB_db record containing the stats to return.
7897 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7898 * @return 0, this function always succeeds.
7901 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7903 arg->ms_psize = env->me_psize;
7904 arg->ms_depth = db->md_depth;
7905 arg->ms_branch_pages = db->md_branch_pages;
7906 arg->ms_leaf_pages = db->md_leaf_pages;
7907 arg->ms_overflow_pages = db->md_overflow_pages;
7908 arg->ms_entries = db->md_entries;
7913 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7917 if (env == NULL || arg == NULL)
7920 toggle = mdb_env_pick_meta(env);
7922 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7926 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7930 if (env == NULL || arg == NULL)
7933 toggle = mdb_env_pick_meta(env);
7934 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7935 arg->me_mapsize = env->me_mapsize;
7936 arg->me_maxreaders = env->me_maxreaders;
7938 /* me_numreaders may be zero if this process never used any readers. Use
7939 * the shared numreader count if it exists.
7941 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
7943 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7944 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7948 /** Set the default comparison functions for a database.
7949 * Called immediately after a database is opened to set the defaults.
7950 * The user can then override them with #mdb_set_compare() or
7951 * #mdb_set_dupsort().
7952 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7953 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7956 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7958 uint16_t f = txn->mt_dbs[dbi].md_flags;
7960 txn->mt_dbxs[dbi].md_cmp =
7961 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7962 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7964 txn->mt_dbxs[dbi].md_dcmp =
7965 !(f & MDB_DUPSORT) ? 0 :
7966 ((f & MDB_INTEGERDUP)
7967 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7968 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7971 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7976 int rc, dbflag, exact;
7977 unsigned int unused = 0;
7980 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7981 mdb_default_cmp(txn, FREE_DBI);
7984 if ((flags & VALID_FLAGS) != flags)
7986 if (txn->mt_flags & MDB_TXN_ERROR)
7992 if (flags & PERSISTENT_FLAGS) {
7993 uint16_t f2 = flags & PERSISTENT_FLAGS;
7994 /* make sure flag changes get committed */
7995 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7996 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7997 txn->mt_flags |= MDB_TXN_DIRTY;
8000 mdb_default_cmp(txn, MAIN_DBI);
8004 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8005 mdb_default_cmp(txn, MAIN_DBI);
8008 /* Is the DB already open? */
8010 for (i=2; i<txn->mt_numdbs; i++) {
8011 if (!txn->mt_dbxs[i].md_name.mv_size) {
8012 /* Remember this free slot */
8013 if (!unused) unused = i;
8016 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8017 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8023 /* If no free slot and max hit, fail */
8024 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8025 return MDB_DBS_FULL;
8027 /* Cannot mix named databases with some mainDB flags */
8028 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8029 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8031 /* Find the DB info */
8032 dbflag = DB_NEW|DB_VALID;
8035 key.mv_data = (void *)name;
8036 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8037 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8038 if (rc == MDB_SUCCESS) {
8039 /* make sure this is actually a DB */
8040 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8041 if (!(node->mn_flags & F_SUBDATA))
8042 return MDB_INCOMPATIBLE;
8043 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8044 /* Create if requested */
8046 data.mv_size = sizeof(MDB_db);
8047 data.mv_data = &dummy;
8048 memset(&dummy, 0, sizeof(dummy));
8049 dummy.md_root = P_INVALID;
8050 dummy.md_flags = flags & PERSISTENT_FLAGS;
8051 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8055 /* OK, got info, add to table */
8056 if (rc == MDB_SUCCESS) {
8057 unsigned int slot = unused ? unused : txn->mt_numdbs;
8058 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8059 txn->mt_dbxs[slot].md_name.mv_size = len;
8060 txn->mt_dbxs[slot].md_rel = NULL;
8061 txn->mt_dbflags[slot] = dbflag;
8062 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8064 mdb_default_cmp(txn, slot);
8073 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8075 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
8078 if (txn->mt_dbflags[dbi] & DB_STALE) {
8081 /* Stale, must read the DB's root. cursor_init does it for us. */
8082 mdb_cursor_init(&mc, txn, dbi, &mx);
8084 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8087 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8090 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8092 ptr = env->me_dbxs[dbi].md_name.mv_data;
8093 env->me_dbxs[dbi].md_name.mv_data = NULL;
8094 env->me_dbxs[dbi].md_name.mv_size = 0;
8095 env->me_dbflags[dbi] = 0;
8099 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8101 /* We could return the flags for the FREE_DBI too but what's the point? */
8102 if (txn == NULL || dbi < MAIN_DBI || dbi >= txn->mt_numdbs)
8104 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8108 /** Add all the DB's pages to the free list.
8109 * @param[in] mc Cursor on the DB to free.
8110 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8111 * @return 0 on success, non-zero on failure.
8114 mdb_drop0(MDB_cursor *mc, int subs)
8118 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8119 if (rc == MDB_SUCCESS) {
8120 MDB_txn *txn = mc->mc_txn;
8125 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8126 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8129 mdb_cursor_copy(mc, &mx);
8130 while (mc->mc_snum > 0) {
8131 MDB_page *mp = mc->mc_pg[mc->mc_top];
8132 unsigned n = NUMKEYS(mp);
8134 for (i=0; i<n; i++) {
8135 ni = NODEPTR(mp, i);
8136 if (ni->mn_flags & F_BIGDATA) {
8139 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8140 rc = mdb_page_get(txn, pg, &omp, NULL);
8143 assert(IS_OVERFLOW(omp));
8144 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8148 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8149 mdb_xcursor_init1(mc, ni);
8150 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8156 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8158 for (i=0; i<n; i++) {
8160 ni = NODEPTR(mp, i);
8163 mdb_midl_xappend(txn->mt_free_pgs, pg);
8168 mc->mc_ki[mc->mc_top] = i;
8169 rc = mdb_cursor_sibling(mc, 1);
8171 /* no more siblings, go back to beginning
8172 * of previous level.
8176 for (i=1; i<mc->mc_snum; i++) {
8178 mc->mc_pg[i] = mx.mc_pg[i];
8183 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8184 } else if (rc == MDB_NOTFOUND) {
8190 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8192 MDB_cursor *mc, *m2;
8195 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
8198 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8201 rc = mdb_cursor_open(txn, dbi, &mc);
8205 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
8206 /* Invalidate the dropped DB's cursors */
8207 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
8208 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
8212 /* Can't delete the main DB */
8213 if (del && dbi > MAIN_DBI) {
8214 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
8216 txn->mt_dbflags[dbi] = DB_STALE;
8217 mdb_dbi_close(txn->mt_env, dbi);
8220 /* reset the DB record, mark it dirty */
8221 txn->mt_dbflags[dbi] |= DB_DIRTY;
8222 txn->mt_dbs[dbi].md_depth = 0;
8223 txn->mt_dbs[dbi].md_branch_pages = 0;
8224 txn->mt_dbs[dbi].md_leaf_pages = 0;
8225 txn->mt_dbs[dbi].md_overflow_pages = 0;
8226 txn->mt_dbs[dbi].md_entries = 0;
8227 txn->mt_dbs[dbi].md_root = P_INVALID;
8229 txn->mt_flags |= MDB_TXN_DIRTY;
8232 mdb_cursor_close(mc);
8236 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8238 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8241 txn->mt_dbxs[dbi].md_cmp = cmp;
8245 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
8247 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8250 txn->mt_dbxs[dbi].md_dcmp = cmp;
8254 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
8256 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8259 txn->mt_dbxs[dbi].md_rel = rel;
8263 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
8265 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
8268 txn->mt_dbxs[dbi].md_relctx = ctx;
8272 int mdb_env_get_maxkeysize(MDB_env *env)
8274 return MDB_MAXKEYSIZE;
8277 int mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
8279 unsigned int i, rdrs;
8286 if (!env->me_txns) {
8287 return func("(no reader locks)\n", ctx);
8289 rdrs = env->me_txns->mti_numreaders;
8290 mr = env->me_txns->mti_readers;
8291 for (i=0; i<rdrs; i++) {
8296 if (mr[i].mr_txnid == (txnid_t)-1) {
8297 sprintf(buf, "%10d %"Z"x -\n", mr[i].mr_pid, tid);
8299 sprintf(buf, "%10d %"Z"x %"Z"u\n", mr[i].mr_pid, tid, mr[i].mr_txnid);
8303 func(" pid thread txnid\n", ctx);
8305 rc = func(buf, ctx);
8311 func("(no active readers)\n", ctx);
8316 /** Insert pid into list if not already present.
8317 * return -1 if already present.
8319 static int mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
8321 /* binary search of pid in list */
8323 unsigned cursor = 1;
8325 unsigned n = ids[0];
8328 unsigned pivot = n >> 1;
8329 cursor = base + pivot + 1;
8330 val = pid - ids[cursor];
8335 } else if ( val > 0 ) {
8340 /* found, so it's a duplicate */
8349 for (n = ids[0]; n > cursor; n--)
8355 int mdb_reader_check(MDB_env *env, int *dead)
8357 unsigned int i, j, rdrs;
8359 MDB_PID_T *pids, pid;
8368 rdrs = env->me_txns->mti_numreaders;
8369 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
8373 mr = env->me_txns->mti_readers;
8374 for (i=0; i<rdrs; i++) {
8375 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
8377 if (mdb_pid_insert(pids, pid) == 0) {
8378 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8380 /* Recheck, a new process may have reused pid */
8381 if (!mdb_reader_pid(env, Pidcheck, pid)) {
8382 for (j=i; j<rdrs; j++)
8383 if (mr[j].mr_pid == pid) {
8384 DPRINTF(("clear stale reader pid %u txn %"Z"d",
8385 (unsigned) pid, mr[j].mr_txnid));
8390 UNLOCK_MUTEX_R(env);