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 2011-2012 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.
35 #include <sys/types.h>
37 #include <sys/param.h>
43 #ifdef HAVE_SYS_FILE_H
60 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
61 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
67 #include <semaphore.h>
72 #include <valgrind/memcheck.h>
73 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
74 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
75 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
76 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
77 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
79 #define VGMEMP_CREATE(h,r,z)
80 #define VGMEMP_ALLOC(h,a,s)
81 #define VGMEMP_FREE(h,a)
82 #define VGMEMP_DESTROY(h)
83 #define VGMEMP_DEFINED(a,s)
87 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
88 /* Solaris just defines one or the other */
89 # define LITTLE_ENDIAN 1234
90 # define BIG_ENDIAN 4321
91 # ifdef _LITTLE_ENDIAN
92 # define BYTE_ORDER LITTLE_ENDIAN
94 # define BYTE_ORDER BIG_ENDIAN
97 # define BYTE_ORDER __BYTE_ORDER
101 #ifndef LITTLE_ENDIAN
102 #define LITTLE_ENDIAN __LITTLE_ENDIAN
105 #define BIG_ENDIAN __BIG_ENDIAN
108 #if defined(__i386) || defined(__x86_64)
109 #define MISALIGNED_OK 1
115 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
116 # error "Unknown or unsupported endianness (BYTE_ORDER)"
117 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
118 # error "Two's complement, reasonably sized integer types, please"
121 /** @defgroup internal MDB Internals
124 /** @defgroup compat Windows Compatibility Macros
125 * A bunch of macros to minimize the amount of platform-specific ifdefs
126 * needed throughout the rest of the code. When the features this library
127 * needs are similar enough to POSIX to be hidden in a one-or-two line
128 * replacement, this macro approach is used.
132 #define pthread_t DWORD
133 #define pthread_mutex_t HANDLE
134 #define pthread_key_t DWORD
135 #define pthread_self() GetCurrentThreadId()
136 #define pthread_key_create(x,y) (*(x) = TlsAlloc())
137 #define pthread_key_delete(x) TlsFree(x)
138 #define pthread_getspecific(x) TlsGetValue(x)
139 #define pthread_setspecific(x,y) TlsSetValue(x,y)
140 #define pthread_mutex_unlock(x) ReleaseMutex(x)
141 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
142 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
143 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
144 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
145 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
146 #define getpid() GetCurrentProcessId()
147 #define fdatasync(fd) (!FlushFileBuffers(fd))
148 #define ErrCode() GetLastError()
149 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
150 #define close(fd) CloseHandle(fd)
151 #define munmap(ptr,len) UnmapViewOfFile(ptr)
154 #define LOCK_MUTEX_R(env) sem_wait((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) sem_wait((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
158 #define fdatasync(fd) fsync(fd)
161 #define fdatasync(fd) fsync(fd)
163 /** Lock the reader mutex.
165 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
166 /** Unlock the reader mutex.
168 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
170 /** Lock the writer mutex.
171 * Only a single write transaction is allowed at a time. Other writers
172 * will block waiting for this mutex.
174 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
175 /** Unlock the writer mutex.
177 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
178 #endif /* __APPLE__ */
180 /** Get the error code for the last failed system function.
182 #define ErrCode() errno
184 /** An abstraction for a file handle.
185 * On POSIX systems file handles are small integers. On Windows
186 * they're opaque pointers.
190 /** A value for an invalid file handle.
191 * Mainly used to initialize file variables and signify that they are
194 #define INVALID_HANDLE_VALUE (-1)
196 /** Get the size of a memory page for the system.
197 * This is the basic size that the platform's memory manager uses, and is
198 * fundamental to the use of memory-mapped files.
200 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
203 #if defined(_WIN32) || defined(__APPLE__)
206 #define MNAME_LEN (sizeof(pthread_mutex_t))
212 /** A flag for opening a file and requesting synchronous data writes.
213 * This is only used when writing a meta page. It's not strictly needed;
214 * we could just do a normal write and then immediately perform a flush.
215 * But if this flag is available it saves us an extra system call.
217 * @note If O_DSYNC is undefined but exists in /usr/include,
218 * preferably set some compiler flag to get the definition.
219 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
222 # define MDB_DSYNC O_DSYNC
226 /** Function for flushing the data of a file. Define this to fsync
227 * if fdatasync() is not supported.
229 #ifndef MDB_FDATASYNC
230 # define MDB_FDATASYNC fdatasync
233 /** A page number in the database.
234 * Note that 64 bit page numbers are overkill, since pages themselves
235 * already represent 12-13 bits of addressable memory, and the OS will
236 * always limit applications to a maximum of 63 bits of address space.
238 * @note In the #MDB_node structure, we only store 48 bits of this value,
239 * which thus limits us to only 60 bits of addressable data.
243 /** A transaction ID.
244 * See struct MDB_txn.mt_txnid for details.
248 /** @defgroup debug Debug Macros
252 /** Enable debug output.
253 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
254 * read from and written to the database (used for free space management).
259 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
260 # define DPRINTF (void) /* Vararg macros may be unsupported */
262 /** Print a debug message with printf formatting. */
263 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
264 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
266 # define DPRINTF(fmt, ...) ((void) 0)
268 /** Print a debug string.
269 * The string is printed literally, with no format processing.
271 #define DPUTS(arg) DPRINTF("%s", arg)
274 /** A default memory page size.
275 * The actual size is platform-dependent, but we use this for
276 * boot-strapping. We probably should not be using this any more.
277 * The #GET_PAGESIZE() macro is used to get the actual size.
279 * Note that we don't currently support Huge pages. On Linux,
280 * regular data files cannot use Huge pages, and in general
281 * Huge pages aren't actually pageable. We rely on the OS
282 * demand-pager to read our data and page it out when memory
283 * pressure from other processes is high. So until OSs have
284 * actual paging support for Huge pages, they're not viable.
286 #define MDB_PAGESIZE 4096
288 /** The minimum number of keys required in a database page.
289 * Setting this to a larger value will place a smaller bound on the
290 * maximum size of a data item. Data items larger than this size will
291 * be pushed into overflow pages instead of being stored directly in
292 * the B-tree node. This value used to default to 4. With a page size
293 * of 4096 bytes that meant that any item larger than 1024 bytes would
294 * go into an overflow page. That also meant that on average 2-3KB of
295 * each overflow page was wasted space. The value cannot be lower than
296 * 2 because then there would no longer be a tree structure. With this
297 * value, items larger than 2KB will go into overflow pages, and on
298 * average only 1KB will be wasted.
300 #define MDB_MINKEYS 2
302 /** A stamp that identifies a file as an MDB file.
303 * There's nothing special about this value other than that it is easily
304 * recognizable, and it will reflect any byte order mismatches.
306 #define MDB_MAGIC 0xBEEFC0DE
308 /** The version number for a database's file format. */
309 #define MDB_VERSION 1
311 /** The maximum size of a key in the database.
312 * While data items have essentially unbounded size, we require that
313 * keys all fit onto a regular page. This limit could be raised a bit
314 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
316 #define MAXKEYSIZE 511
321 * This is used for printing a hex dump of a key's contents.
323 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
324 /** Display a key in hex.
326 * Invoke a function to display a key in hex.
328 #define DKEY(x) mdb_dkey(x, kbuf)
330 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
334 /** @defgroup lazylock Lazy Locking
335 * Macros for locks that aren't actually needed.
336 * The DB view is always consistent because all writes are wrapped in
337 * the wmutex. Finer-grained locks aren't necessary.
341 /** Use lazy locking. I.e., don't lock these accesses at all. */
345 /** Grab the reader lock */
346 #define LAZY_MUTEX_LOCK(x)
347 /** Release the reader lock */
348 #define LAZY_MUTEX_UNLOCK(x)
349 /** Release the DB table reader/writer lock */
350 #define LAZY_RWLOCK_UNLOCK(x)
351 /** Grab the DB table write lock */
352 #define LAZY_RWLOCK_WRLOCK(x)
353 /** Grab the DB table read lock */
354 #define LAZY_RWLOCK_RDLOCK(x)
355 /** Declare the DB table rwlock. Should not be followed by ';'. */
356 #define LAZY_RWLOCK_DEF(x)
357 /** Initialize the DB table rwlock */
358 #define LAZY_RWLOCK_INIT(x,y)
359 /** Destroy the DB table rwlock */
360 #define LAZY_RWLOCK_DESTROY(x)
362 #define LAZY_MUTEX_LOCK(x) pthread_mutex_lock(x)
363 #define LAZY_MUTEX_UNLOCK(x) pthread_mutex_unlock(x)
364 #define LAZY_RWLOCK_UNLOCK(x) pthread_rwlock_unlock(x)
365 #define LAZY_RWLOCK_WRLOCK(x) pthread_rwlock_wrlock(x)
366 #define LAZY_RWLOCK_RDLOCK(x) pthread_rwlock_rdlock(x)
367 #define LAZY_RWLOCK_DEF(x) pthread_rwlock_t x;
368 #define LAZY_RWLOCK_INIT(x,y) pthread_rwlock_init(x,y)
369 #define LAZY_RWLOCK_DESTROY(x) pthread_rwlock_destroy(x)
373 /** An invalid page number.
374 * Mainly used to denote an empty tree.
376 #define P_INVALID (~0UL)
378 /** Test if a flag \b f is set in a flag word \b w. */
379 #define F_ISSET(w, f) (((w) & (f)) == (f))
381 /** Used for offsets within a single page.
382 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
385 typedef uint16_t indx_t;
387 /** Default size of memory map.
388 * This is certainly too small for any actual applications. Apps should always set
389 * the size explicitly using #mdb_env_set_mapsize().
391 #define DEFAULT_MAPSIZE 1048576
393 /** @defgroup readers Reader Lock Table
394 * Readers don't acquire any locks for their data access. Instead, they
395 * simply record their transaction ID in the reader table. The reader
396 * mutex is needed just to find an empty slot in the reader table. The
397 * slot's address is saved in thread-specific data so that subsequent read
398 * transactions started by the same thread need no further locking to proceed.
400 * Since the database uses multi-version concurrency control, readers don't
401 * actually need any locking. This table is used to keep track of which
402 * readers are using data from which old transactions, so that we'll know
403 * when a particular old transaction is no longer in use. Old transactions
404 * that have discarded any data pages can then have those pages reclaimed
405 * for use by a later write transaction.
407 * The lock table is constructed such that reader slots are aligned with the
408 * processor's cache line size. Any slot is only ever used by one thread.
409 * This alignment guarantees that there will be no contention or cache
410 * thrashing as threads update their own slot info, and also eliminates
411 * any need for locking when accessing a slot.
413 * A writer thread will scan every slot in the table to determine the oldest
414 * outstanding reader transaction. Any freed pages older than this will be
415 * reclaimed by the writer. The writer doesn't use any locks when scanning
416 * this table. This means that there's no guarantee that the writer will
417 * see the most up-to-date reader info, but that's not required for correct
418 * operation - all we need is to know the upper bound on the oldest reader,
419 * we don't care at all about the newest reader. So the only consequence of
420 * reading stale information here is that old pages might hang around a
421 * while longer before being reclaimed. That's actually good anyway, because
422 * the longer we delay reclaiming old pages, the more likely it is that a
423 * string of contiguous pages can be found after coalescing old pages from
424 * many old transactions together.
426 * @todo We don't actually do such coalescing yet, we grab pages from one
427 * old transaction at a time.
430 /** Number of slots in the reader table.
431 * This value was chosen somewhat arbitrarily. 126 readers plus a
432 * couple mutexes fit exactly into 8KB on my development machine.
433 * Applications should set the table size using #mdb_env_set_maxreaders().
435 #define DEFAULT_READERS 126
437 /** The size of a CPU cache line in bytes. We want our lock structures
438 * aligned to this size to avoid false cache line sharing in the
440 * This value works for most CPUs. For Itanium this should be 128.
446 /** The information we store in a single slot of the reader table.
447 * In addition to a transaction ID, we also record the process and
448 * thread ID that owns a slot, so that we can detect stale information,
449 * e.g. threads or processes that went away without cleaning up.
450 * @note We currently don't check for stale records. We simply re-init
451 * the table when we know that we're the only process opening the
454 typedef struct MDB_rxbody {
455 /** The current Transaction ID when this transaction began.
456 * Multiple readers that start at the same time will probably have the
457 * same ID here. Again, it's not important to exclude them from
458 * anything; all we need to know is which version of the DB they
459 * started from so we can avoid overwriting any data used in that
460 * particular version.
463 /** The process ID of the process owning this reader txn. */
465 /** The thread ID of the thread owning this txn. */
469 /** The actual reader record, with cacheline padding. */
470 typedef struct MDB_reader {
473 /** shorthand for mrb_txnid */
474 #define mr_txnid mru.mrx.mrb_txnid
475 #define mr_pid mru.mrx.mrb_pid
476 #define mr_tid mru.mrx.mrb_tid
477 /** cache line alignment */
478 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
482 /** The header for the reader table.
483 * The table resides in a memory-mapped file. (This is a different file
484 * than is used for the main database.)
486 * For POSIX the actual mutexes reside in the shared memory of this
487 * mapped file. On Windows, mutexes are named objects allocated by the
488 * kernel; we store the mutex names in this mapped file so that other
489 * processes can grab them. This same approach is also used on
490 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
491 * process-shared POSIX mutexes. For these cases where a named object
492 * is used, the object name is derived from a 64 bit FNV hash of the
493 * environment pathname. As such, naming collisions are extremely
494 * unlikely. If a collision occurs, the results are unpredictable.
496 typedef struct MDB_txbody {
497 /** Stamp identifying this as an MDB file. It must be set
500 /** Version number of this lock file. Must be set to #MDB_VERSION. */
501 uint32_t mtb_version;
502 #if defined(_WIN32) || defined(__APPLE__)
503 char mtb_rmname[MNAME_LEN];
505 /** Mutex protecting access to this table.
506 * This is the reader lock that #LOCK_MUTEX_R acquires.
508 pthread_mutex_t mtb_mutex;
510 /** The ID of the last transaction committed to the database.
511 * This is recorded here only for convenience; the value can always
512 * be determined by reading the main database meta pages.
515 /** The number of slots that have been used in the reader table.
516 * This always records the maximum count, it is not decremented
517 * when readers release their slots.
519 unsigned mtb_numreaders;
520 /** The ID of the most recent meta page in the database.
521 * This is recorded here only for convenience; the value can always
522 * be determined by reading the main database meta pages.
524 uint32_t mtb_me_toggle;
527 /** The actual reader table definition. */
528 typedef struct MDB_txninfo {
531 #define mti_magic mt1.mtb.mtb_magic
532 #define mti_version mt1.mtb.mtb_version
533 #define mti_mutex mt1.mtb.mtb_mutex
534 #define mti_rmname mt1.mtb.mtb_rmname
535 #define mti_txnid mt1.mtb.mtb_txnid
536 #define mti_numreaders mt1.mtb.mtb_numreaders
537 #define mti_me_toggle mt1.mtb.mtb_me_toggle
538 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
541 #if defined(_WIN32) || defined(__APPLE__)
542 char mt2_wmname[MNAME_LEN];
543 #define mti_wmname mt2.mt2_wmname
545 pthread_mutex_t mt2_wmutex;
546 #define mti_wmutex mt2.mt2_wmutex
548 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
550 MDB_reader mti_readers[1];
554 /** Common header for all page types.
555 * Overflow records occupy a number of contiguous pages with no
556 * headers on any page after the first.
558 typedef struct MDB_page {
559 #define mp_pgno mp_p.p_pgno
560 #define mp_next mp_p.p_next
562 pgno_t p_pgno; /**< page number */
563 void * p_next; /**< for in-memory list of freed structs */
566 /** @defgroup mdb_page Page Flags
568 * Flags for the page headers.
571 #define P_BRANCH 0x01 /**< branch page */
572 #define P_LEAF 0x02 /**< leaf page */
573 #define P_OVERFLOW 0x04 /**< overflow page */
574 #define P_META 0x08 /**< meta page */
575 #define P_DIRTY 0x10 /**< dirty page */
576 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
577 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
579 uint16_t mp_flags; /**< @ref mdb_page */
580 #define mp_lower mp_pb.pb.pb_lower
581 #define mp_upper mp_pb.pb.pb_upper
582 #define mp_pages mp_pb.pb_pages
585 indx_t pb_lower; /**< lower bound of free space */
586 indx_t pb_upper; /**< upper bound of free space */
588 uint32_t pb_pages; /**< number of overflow pages */
590 indx_t mp_ptrs[1]; /**< dynamic size */
593 /** Size of the page header, excluding dynamic data at the end */
594 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
596 /** Address of first usable data byte in a page, after the header */
597 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
599 /** Number of nodes on a page */
600 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
602 /** The amount of space remaining in the page */
603 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
605 /** The percentage of space used in the page, in tenths of a percent. */
606 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
607 ((env)->me_psize - PAGEHDRSZ))
608 /** The minimum page fill factor, in tenths of a percent.
609 * Pages emptier than this are candidates for merging.
611 #define FILL_THRESHOLD 250
613 /** Test if a page is a leaf page */
614 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
615 /** Test if a page is a LEAF2 page */
616 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
617 /** Test if a page is a branch page */
618 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
619 /** Test if a page is an overflow page */
620 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
621 /** Test if a page is a sub page */
622 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
624 /** The number of overflow pages needed to store the given size. */
625 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
627 /** Header for a single key/data pair within a page.
628 * We guarantee 2-byte alignment for nodes.
630 typedef struct MDB_node {
631 /** lo and hi are used for data size on leaf nodes and for
632 * child pgno on branch nodes. On 64 bit platforms, flags
633 * is also used for pgno. (Branch nodes have no flags).
634 * They are in host byte order in case that lets some
635 * accesses be optimized into a 32-bit word access.
637 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
638 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
639 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
640 /** @defgroup mdb_node Node Flags
642 * Flags for node headers.
645 #define F_BIGDATA 0x01 /**< data put on overflow page */
646 #define F_SUBDATA 0x02 /**< data is a sub-database */
647 #define F_DUPDATA 0x04 /**< data has duplicates */
649 /** valid flags for #mdb_node_add() */
650 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
653 unsigned short mn_flags; /**< @ref mdb_node */
654 unsigned short mn_ksize; /**< key size */
655 char mn_data[1]; /**< key and data are appended here */
658 /** Size of the node header, excluding dynamic data at the end */
659 #define NODESIZE offsetof(MDB_node, mn_data)
661 /** Bit position of top word in page number, for shifting mn_flags */
662 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
664 /** Size of a node in a branch page with a given key.
665 * This is just the node header plus the key, there is no data.
667 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
669 /** Size of a node in a leaf page with a given key and data.
670 * This is node header plus key plus data size.
672 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
674 /** Address of node \b i in page \b p */
675 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
677 /** Address of the key for the node */
678 #define NODEKEY(node) (void *)((node)->mn_data)
680 /** Address of the data for a node */
681 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
683 /** Get the page number pointed to by a branch node */
684 #define NODEPGNO(node) \
685 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
686 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
687 /** Set the page number in a branch node */
688 #define SETPGNO(node,pgno) do { \
689 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
690 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
692 /** Get the size of the data in a leaf node */
693 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
694 /** Set the size of the data for a leaf node */
695 #define SETDSZ(node,size) do { \
696 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
697 /** The size of a key in a node */
698 #define NODEKSZ(node) ((node)->mn_ksize)
700 /** Copy a page number from src to dst */
702 #define COPY_PGNO(dst,src) dst = src
704 #if SIZE_MAX > 4294967295UL
705 #define COPY_PGNO(dst,src) do { \
706 unsigned short *s, *d; \
707 s = (unsigned short *)&(src); \
708 d = (unsigned short *)&(dst); \
715 #define COPY_PGNO(dst,src) do { \
716 unsigned short *s, *d; \
717 s = (unsigned short *)&(src); \
718 d = (unsigned short *)&(dst); \
724 /** The address of a key in a LEAF2 page.
725 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
726 * There are no node headers, keys are stored contiguously.
728 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
730 /** Set the \b node's key into \b key, if requested. */
731 #define MDB_SET_KEY(node, key) { if ((key) != NULL) { \
732 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
734 /** Information about a single database in the environment. */
735 typedef struct MDB_db {
736 uint32_t md_pad; /**< also ksize for LEAF2 pages */
737 uint16_t md_flags; /**< @ref mdb_open */
738 uint16_t md_depth; /**< depth of this tree */
739 pgno_t md_branch_pages; /**< number of internal pages */
740 pgno_t md_leaf_pages; /**< number of leaf pages */
741 pgno_t md_overflow_pages; /**< number of overflow pages */
742 size_t md_entries; /**< number of data items */
743 pgno_t md_root; /**< the root page of this tree */
746 /** Handle for the DB used to track free pages. */
748 /** Handle for the default DB. */
751 /** Meta page content. */
752 typedef struct MDB_meta {
753 /** Stamp identifying this as an MDB file. It must be set
756 /** Version number of this lock file. Must be set to #MDB_VERSION. */
758 void *mm_address; /**< address for fixed mapping */
759 size_t mm_mapsize; /**< size of mmap region */
760 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
761 /** The size of pages used in this DB */
762 #define mm_psize mm_dbs[0].md_pad
763 /** Any persistent environment flags. @ref mdb_env */
764 #define mm_flags mm_dbs[0].md_flags
765 pgno_t mm_last_pg; /**< last used page in file */
766 txnid_t mm_txnid; /**< txnid that committed this page */
769 /** Buffer for a stack-allocated dirty page.
770 * The members define size and alignment, and silence type
771 * aliasing warnings. They are not used directly; that could
772 * mean incorrectly using several union members in parallel.
774 typedef union MDB_pagebuf {
775 char mb_raw[MDB_PAGESIZE];
778 char mm_pad[PAGEHDRSZ];
783 /** Auxiliary DB info.
784 * The information here is mostly static/read-only. There is
785 * only a single copy of this record in the environment.
787 typedef struct MDB_dbx {
788 MDB_val md_name; /**< name of the database */
789 MDB_cmp_func *md_cmp; /**< function for comparing keys */
790 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
791 MDB_rel_func *md_rel; /**< user relocate function */
792 void *md_relctx; /**< user-provided context for md_rel */
795 /** A database transaction.
796 * Every operation requires a transaction handle.
799 MDB_txn *mt_parent; /**< parent of a nested txn */
800 MDB_txn *mt_child; /**< nested txn under this txn */
801 pgno_t mt_next_pgno; /**< next unallocated page */
802 /** The ID of this transaction. IDs are integers incrementing from 1.
803 * Only committed write transactions increment the ID. If a transaction
804 * aborts, the ID may be re-used by the next writer.
807 MDB_env *mt_env; /**< the DB environment */
808 /** The list of pages that became unused during this transaction.
812 ID2L dirty_list; /**< modified pages */
813 MDB_reader *reader; /**< this thread's slot in the reader table */
815 /** Array of records for each DB known in the environment. */
817 /** Array of MDB_db records for each known DB */
819 /** @defgroup mt_dbflag Transaction DB Flags
823 #define DB_DIRTY 0x01 /**< DB was written in this txn */
824 #define DB_STALE 0x02 /**< DB record is older than txnID */
826 /** Array of cursors for each DB */
827 MDB_cursor **mt_cursors;
828 /** Array of flags for each DB */
829 unsigned char *mt_dbflags;
830 /** Number of DB records in use. This number only ever increments;
831 * we don't decrement it when individual DB handles are closed.
835 /** @defgroup mdb_txn Transaction Flags
839 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
840 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
842 unsigned int mt_flags; /**< @ref mdb_txn */
843 /** Tracks which of the two meta pages was used at the start
844 * of this transaction.
846 unsigned int mt_toggle;
849 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
850 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
851 * raise this on a 64 bit machine.
853 #define CURSOR_STACK 32
857 /** Cursors are used for all DB operations */
859 /** Next cursor on this DB in this txn */
861 /** Original cursor if this is a shadow */
863 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
864 struct MDB_xcursor *mc_xcursor;
865 /** The transaction that owns this cursor */
867 /** The database handle this cursor operates on */
869 /** The database record for this cursor */
871 /** The database auxiliary record for this cursor */
873 /** The @ref mt_dbflag for this database */
874 unsigned char *mc_dbflag;
875 unsigned short mc_snum; /**< number of pushed pages */
876 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
877 /** @defgroup mdb_cursor Cursor Flags
879 * Cursor state flags.
882 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
883 #define C_EOF 0x02 /**< No more data */
884 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
885 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
886 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
888 unsigned int mc_flags; /**< @ref mdb_cursor */
889 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
890 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
893 /** Context for sorted-dup records.
894 * We could have gone to a fully recursive design, with arbitrarily
895 * deep nesting of sub-databases. But for now we only handle these
896 * levels - main DB, optional sub-DB, sorted-duplicate DB.
898 typedef struct MDB_xcursor {
899 /** A sub-cursor for traversing the Dup DB */
900 MDB_cursor mx_cursor;
901 /** The database record for this Dup DB */
903 /** The auxiliary DB record for this Dup DB */
905 /** The @ref mt_dbflag for this Dup DB */
906 unsigned char mx_dbflag;
909 /** A set of pages freed by an earlier transaction. */
910 typedef struct MDB_oldpages {
911 /** Usually we only read one record from the FREEDB at a time, but
912 * in case we read more, this will chain them together.
914 struct MDB_oldpages *mo_next;
915 /** The ID of the transaction in which these pages were freed. */
917 /** An #IDL of the pages */
918 pgno_t mo_pages[1]; /* dynamic */
921 /** The database environment. */
923 HANDLE me_fd; /**< The main data file */
924 HANDLE me_lfd; /**< The lock file */
925 HANDLE me_mfd; /**< just for writing the meta pages */
926 /** Failed to update the meta page. Probably an I/O error. */
927 #define MDB_FATAL_ERROR 0x80000000U
928 uint32_t me_flags; /**< @ref mdb_env */
929 uint32_t me_extrapad; /**< unused for now */
930 unsigned int me_maxreaders; /**< size of the reader table */
931 MDB_dbi me_numdbs; /**< number of DBs opened */
932 MDB_dbi me_maxdbs; /**< size of the DB table */
933 char *me_path; /**< path to the DB files */
934 char *me_map; /**< the memory map of the data file */
935 MDB_txninfo *me_txns; /**< the memory map of the lock file */
936 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
937 MDB_txn *me_txn; /**< current write transaction */
938 size_t me_mapsize; /**< size of the data memory map */
939 off_t me_size; /**< current file size */
940 pgno_t me_maxpg; /**< me_mapsize / me_psize */
941 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
942 unsigned int me_db_toggle; /**< which DB table is current */
943 txnid_t me_wtxnid; /**< ID of last txn we committed */
944 txnid_t me_pgfirst; /**< ID of first old page record we used */
945 txnid_t me_pglast; /**< ID of last old page record we used */
946 MDB_dbx *me_dbxs; /**< array of static DB info */
947 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
948 MDB_oldpages *me_pghead; /**< list of old page records */
949 pthread_key_t me_txkey; /**< thread-key for readers */
950 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
951 /** IDL of pages that became unused in a write txn */
953 /** ID2L of pages that were written during a write txn */
954 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
955 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
956 LAZY_RWLOCK_DEF(me_dblock)
958 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
962 sem_t *me_rmutex; /* Apple doesn't support shared mutexes */
966 /** max number of pages to commit in one writev() call */
967 #define MDB_COMMIT_PAGES 64
968 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
969 #undef MDB_COMMIT_PAGES
970 #define MDB_COMMIT_PAGES IOV_MAX
973 static MDB_page *mdb_page_alloc(MDB_cursor *mc, int num);
974 static MDB_page *mdb_page_new(MDB_cursor *mc, uint32_t flags, int num);
975 static int mdb_page_touch(MDB_cursor *mc);
977 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
978 static int mdb_page_search_root(MDB_cursor *mc,
979 MDB_val *key, int modify);
980 static int mdb_page_search(MDB_cursor *mc,
981 MDB_val *key, int modify);
982 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
983 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
984 pgno_t newpgno, unsigned int nflags);
986 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
987 static int mdb_env_read_meta(MDB_env *env, int *which);
988 static int mdb_env_write_meta(MDB_txn *txn);
990 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
991 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
992 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
993 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
994 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
995 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
996 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
997 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
998 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1000 static int mdb_rebalance(MDB_cursor *mc);
1001 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
1003 static void mdb_cursor_pop(MDB_cursor *mc);
1004 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1006 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1007 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1008 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1009 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1010 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1012 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1013 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1015 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1016 static void mdb_xcursor_init0(MDB_cursor *mc);
1017 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1019 static int mdb_drop0(MDB_cursor *mc, int subs);
1020 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1023 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1027 static SECURITY_DESCRIPTOR mdb_null_sd;
1028 static SECURITY_ATTRIBUTES mdb_all_sa;
1029 static int mdb_sec_inited;
1032 /** Return the library version info. */
1034 mdb_version(int *major, int *minor, int *patch)
1036 if (major) *major = MDB_VERSION_MAJOR;
1037 if (minor) *minor = MDB_VERSION_MINOR;
1038 if (patch) *patch = MDB_VERSION_PATCH;
1039 return MDB_VERSION_STRING;
1042 /** Table of descriptions for MDB @ref errors */
1043 static char *const mdb_errstr[] = {
1044 "MDB_KEYEXIST: Key/data pair already exists",
1045 "MDB_NOTFOUND: No matching key/data pair found",
1046 "MDB_PAGE_NOTFOUND: Requested page not found",
1047 "MDB_CORRUPTED: Located page was wrong type",
1048 "MDB_PANIC: Update of meta page failed",
1049 "MDB_VERSION_MISMATCH: Database environment version mismatch"
1053 mdb_strerror(int err)
1056 return ("Successful return: 0");
1058 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
1059 return mdb_errstr[err - MDB_KEYEXIST];
1061 return strerror(err);
1065 /** Display a key in hexadecimal and return the address of the result.
1066 * @param[in] key the key to display
1067 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1068 * @return The key in hexadecimal form.
1071 mdb_dkey(MDB_val *key, char *buf)
1074 unsigned char *c = key->mv_data;
1076 if (key->mv_size > MAXKEYSIZE)
1077 return "MAXKEYSIZE";
1078 /* may want to make this a dynamic check: if the key is mostly
1079 * printable characters, print it as-is instead of converting to hex.
1083 for (i=0; i<key->mv_size; i++)
1084 ptr += sprintf(ptr, "%02x", *c++);
1086 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1091 /** Display all the keys in the page. */
1093 mdb_page_keys(MDB_page *mp)
1096 unsigned int i, nkeys;
1100 nkeys = NUMKEYS(mp);
1101 DPRINTF("numkeys %d", nkeys);
1102 for (i=0; i<nkeys; i++) {
1103 node = NODEPTR(mp, i);
1104 key.mv_size = node->mn_ksize;
1105 key.mv_data = node->mn_data;
1106 DPRINTF("key %d: %s", i, DKEY(&key));
1112 /** Count all the pages in each DB and in the freelist
1113 * and make sure it matches the actual number of pages
1116 static void mdb_audit(MDB_txn *txn)
1121 ID freecount, count;
1124 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1125 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1126 freecount += *(ID *)data.mv_data;
1127 freecount += txn->mt_dbs[0].md_branch_pages + txn->mt_dbs[0].md_leaf_pages +
1128 txn->mt_dbs[0].md_overflow_pages;
1131 for (i = 0; i<txn->mt_numdbs; i++) {
1132 count += txn->mt_dbs[i].md_branch_pages +
1133 txn->mt_dbs[i].md_leaf_pages +
1134 txn->mt_dbs[i].md_overflow_pages;
1135 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1137 mdb_cursor_init(&mc, txn, i, &mx);
1138 mdb_page_search(&mc, NULL, 0);
1142 mp = mc.mc_pg[mc.mc_top];
1143 for (j=0; j<NUMKEYS(mp); j++) {
1144 MDB_node *leaf = NODEPTR(mp, j);
1145 if (leaf->mn_flags & F_SUBDATA) {
1147 memcpy(&db, NODEDATA(leaf), sizeof(db));
1148 count += db.md_branch_pages + db.md_leaf_pages +
1149 db.md_overflow_pages;
1153 while (mdb_cursor_sibling(&mc, 1) == 0);
1156 assert(freecount + count + 2 >= txn->mt_next_pgno - 1);
1161 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1163 return txn->mt_dbxs[dbi].md_cmp(a, b);
1167 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1169 if (txn->mt_dbxs[dbi].md_dcmp)
1170 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1172 return EINVAL; /* too bad you can't distinguish this from a valid result */
1175 /** Allocate a single page.
1176 * Re-use old malloc'd pages first, otherwise just malloc.
1179 mdb_page_malloc(MDB_cursor *mc) {
1181 size_t sz = mc->mc_txn->mt_env->me_psize;
1182 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1183 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1184 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1185 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1186 } else if ((ret = malloc(sz)) != NULL) {
1187 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1192 /** Allocate pages for writing.
1193 * If there are free pages available from older transactions, they
1194 * will be re-used first. Otherwise a new page will be allocated.
1195 * @param[in] mc cursor A cursor handle identifying the transaction and
1196 * database for which we are allocating.
1197 * @param[in] num the number of pages to allocate.
1198 * @return Address of the allocated page(s). Requests for multiple pages
1199 * will always be satisfied by a single contiguous chunk of memory.
1202 mdb_page_alloc(MDB_cursor *mc, int num)
1204 MDB_txn *txn = mc->mc_txn;
1206 pgno_t pgno = P_INVALID;
1209 if (txn->mt_txnid > 2) {
1211 if (!txn->mt_env->me_pghead &&
1212 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1213 /* See if there's anything in the free DB */
1217 txnid_t *kptr, oldest, last;
1219 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1220 if (!txn->mt_env->me_pgfirst) {
1221 mdb_page_search(&m2, NULL, 0);
1222 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1223 kptr = (txnid_t *)NODEKEY(leaf);
1230 last = txn->mt_env->me_pglast + 1;
1232 key.mv_data = &last;
1233 key.mv_size = sizeof(last);
1234 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1237 last = *(txnid_t *)key.mv_data;
1242 oldest = txn->mt_txnid - 1;
1243 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1244 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1245 if (mr && mr < oldest)
1250 if (oldest > last) {
1251 /* It's usable, grab it.
1256 if (!txn->mt_env->me_pgfirst) {
1257 mdb_node_read(txn, leaf, &data);
1259 txn->mt_env->me_pglast = last;
1260 if (!txn->mt_env->me_pgfirst)
1261 txn->mt_env->me_pgfirst = last;
1262 idl = (ID *) data.mv_data;
1263 /* We might have a zero-length IDL due to freelist growth
1264 * during a prior commit
1266 if (!idl[0]) goto again;
1267 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1268 mop->mo_next = txn->mt_env->me_pghead;
1269 mop->mo_txnid = last;
1270 txn->mt_env->me_pghead = mop;
1271 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1276 DPRINTF("IDL read txn %zu root %zu num %zu",
1277 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1278 for (i=0; i<idl[0]; i++) {
1279 DPRINTF("IDL %zu", idl[i+1]);
1286 if (txn->mt_env->me_pghead) {
1287 MDB_oldpages *mop = txn->mt_env->me_pghead;
1289 /* FIXME: For now, always use fresh pages. We
1290 * really ought to search the free list for a
1295 /* peel pages off tail, so we only have to truncate the list */
1296 pgno = MDB_IDL_LAST(mop->mo_pages);
1297 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1299 if (mop->mo_pages[2] > mop->mo_pages[1])
1300 mop->mo_pages[0] = 0;
1304 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1305 txn->mt_env->me_pghead = mop->mo_next;
1312 if (pgno == P_INVALID) {
1313 /* DB size is maxed out */
1314 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1315 DPUTS("DB size maxed out");
1319 if (txn->mt_env->me_dpages && num == 1) {
1320 np = txn->mt_env->me_dpages;
1321 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1322 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1323 txn->mt_env->me_dpages = np->mp_next;
1325 size_t sz = txn->mt_env->me_psize * num;
1326 if ((np = malloc(sz)) == NULL)
1328 VGMEMP_ALLOC(txn->mt_env, np, sz);
1330 if (pgno == P_INVALID) {
1331 np->mp_pgno = txn->mt_next_pgno;
1332 txn->mt_next_pgno += num;
1336 mid.mid = np->mp_pgno;
1338 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1343 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1344 * @param[in] mc cursor pointing to the page to be touched
1345 * @return 0 on success, non-zero on failure.
1348 mdb_page_touch(MDB_cursor *mc)
1350 MDB_page *mp = mc->mc_pg[mc->mc_top];
1353 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1355 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1357 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1358 assert(mp->mp_pgno != np->mp_pgno);
1359 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1361 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1364 mp->mp_flags |= P_DIRTY;
1367 /* Adjust other cursors pointing to mp */
1368 if (mc->mc_flags & C_SUB) {
1369 MDB_cursor *m2, *m3;
1370 MDB_dbi dbi = mc->mc_dbi-1;
1372 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1373 if (m2 == mc) continue;
1374 m3 = &m2->mc_xcursor->mx_cursor;
1375 if (m3->mc_snum < mc->mc_snum) continue;
1376 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1377 m3->mc_pg[mc->mc_top] = mp;
1383 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1384 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1385 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1386 m2->mc_pg[mc->mc_top] = mp;
1390 mc->mc_pg[mc->mc_top] = mp;
1391 /** If this page has a parent, update the parent to point to
1395 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1397 mc->mc_db->md_root = mp->mp_pgno;
1398 } else if (mc->mc_txn->mt_parent) {
1401 /* If txn has a parent, make sure the page is in our
1404 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1405 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1406 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1407 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1408 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1409 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1410 mc->mc_pg[mc->mc_top] = mp;
1416 np = mdb_page_malloc(mc);
1417 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1418 mid.mid = np->mp_pgno;
1420 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1428 mdb_env_sync(MDB_env *env, int force)
1431 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1432 if (MDB_FDATASYNC(env->me_fd))
1438 /** Make shadow copies of all of parent txn's cursors */
1440 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1442 MDB_cursor *mc, *m2;
1443 unsigned int i, j, size;
1445 for (i=0;i<src->mt_numdbs; i++) {
1446 if (src->mt_cursors[i]) {
1447 size = sizeof(MDB_cursor);
1448 if (src->mt_cursors[i]->mc_xcursor)
1449 size += sizeof(MDB_xcursor);
1450 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1457 mc->mc_db = &dst->mt_dbs[i];
1458 mc->mc_dbx = m2->mc_dbx;
1459 mc->mc_dbflag = &dst->mt_dbflags[i];
1460 mc->mc_snum = m2->mc_snum;
1461 mc->mc_top = m2->mc_top;
1462 mc->mc_flags = m2->mc_flags | C_SHADOW;
1463 for (j=0; j<mc->mc_snum; j++) {
1464 mc->mc_pg[j] = m2->mc_pg[j];
1465 mc->mc_ki[j] = m2->mc_ki[j];
1467 if (m2->mc_xcursor) {
1468 MDB_xcursor *mx, *mx2;
1469 mx = (MDB_xcursor *)(mc+1);
1470 mc->mc_xcursor = mx;
1471 mx2 = m2->mc_xcursor;
1472 mx->mx_db = mx2->mx_db;
1473 mx->mx_dbx = mx2->mx_dbx;
1474 mx->mx_dbflag = mx2->mx_dbflag;
1475 mx->mx_cursor.mc_txn = dst;
1476 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1477 mx->mx_cursor.mc_db = &mx->mx_db;
1478 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1479 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1480 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1481 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1482 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1483 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1484 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1485 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1488 mc->mc_xcursor = NULL;
1490 mc->mc_next = dst->mt_cursors[i];
1491 dst->mt_cursors[i] = mc;
1498 /** Merge shadow cursors back into parent's */
1500 mdb_cursor_merge(MDB_txn *txn)
1503 for (i=0; i<txn->mt_numdbs; i++) {
1504 if (txn->mt_cursors[i]) {
1506 while ((mc = txn->mt_cursors[i])) {
1507 txn->mt_cursors[i] = mc->mc_next;
1508 if (mc->mc_flags & C_SHADOW) {
1509 MDB_cursor *m2 = mc->mc_orig;
1511 m2->mc_snum = mc->mc_snum;
1512 m2->mc_top = mc->mc_top;
1513 for (j=0; j<mc->mc_snum; j++) {
1514 m2->mc_pg[j] = mc->mc_pg[j];
1515 m2->mc_ki[j] = mc->mc_ki[j];
1518 if (mc->mc_flags & C_ALLOCD)
1526 mdb_txn_reset0(MDB_txn *txn);
1528 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1529 * @param[in] txn the transaction handle to initialize
1530 * @return 0 on success, non-zero on failure. This can only
1531 * fail for read-only transactions, and then only if the
1532 * reader table is full.
1535 mdb_txn_renew0(MDB_txn *txn)
1537 MDB_env *env = txn->mt_env;
1540 if (txn->mt_flags & MDB_TXN_RDONLY) {
1541 MDB_reader *r = pthread_getspecific(env->me_txkey);
1544 pid_t pid = getpid();
1545 pthread_t tid = pthread_self();
1548 for (i=0; i<env->me_txns->mti_numreaders; i++)
1549 if (env->me_txns->mti_readers[i].mr_pid == 0)
1551 if (i == env->me_maxreaders) {
1552 UNLOCK_MUTEX_R(env);
1555 env->me_txns->mti_readers[i].mr_pid = pid;
1556 env->me_txns->mti_readers[i].mr_tid = tid;
1557 if (i >= env->me_txns->mti_numreaders)
1558 env->me_txns->mti_numreaders = i+1;
1559 UNLOCK_MUTEX_R(env);
1560 r = &env->me_txns->mti_readers[i];
1561 pthread_setspecific(env->me_txkey, r);
1563 txn->mt_toggle = env->me_txns->mti_me_toggle;
1564 txn->mt_txnid = env->me_txns->mti_txnid;
1565 /* This happens if a different process was the
1566 * last writer to the DB.
1568 if (env->me_wtxnid < txn->mt_txnid)
1569 mt_dbflag = DB_STALE;
1570 r->mr_txnid = txn->mt_txnid;
1571 txn->mt_u.reader = r;
1575 txn->mt_txnid = env->me_txns->mti_txnid;
1576 if (env->me_wtxnid < txn->mt_txnid)
1577 mt_dbflag = DB_STALE;
1579 txn->mt_toggle = env->me_txns->mti_me_toggle;
1580 txn->mt_u.dirty_list = env->me_dirty_list;
1581 txn->mt_u.dirty_list[0].mid = 0;
1582 txn->mt_free_pgs = env->me_free_pgs;
1583 txn->mt_free_pgs[0] = 0;
1584 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1588 /* Copy the DB arrays */
1589 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1590 txn->mt_numdbs = env->me_numdbs;
1591 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1592 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1593 if (txn->mt_numdbs > 2)
1594 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1595 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1596 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1598 memset(txn->mt_dbflags, mt_dbflag, env->me_numdbs);
1604 mdb_txn_renew(MDB_txn *txn)
1611 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1612 DPUTS("environment had fatal error, must shutdown!");
1616 rc = mdb_txn_renew0(txn);
1617 if (rc == MDB_SUCCESS) {
1618 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1619 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1620 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1626 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1631 if (env->me_flags & MDB_FATAL_ERROR) {
1632 DPUTS("environment had fatal error, must shutdown!");
1636 /* parent already has an active child txn */
1637 if (parent->mt_child) {
1641 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1642 if (!(flags & MDB_RDONLY))
1643 size += env->me_maxdbs * sizeof(MDB_cursor *);
1645 if ((txn = calloc(1, size)) == NULL) {
1646 DPRINTF("calloc: %s", strerror(ErrCode()));
1649 txn->mt_dbs = (MDB_db *)(txn+1);
1650 if (flags & MDB_RDONLY) {
1651 txn->mt_flags |= MDB_TXN_RDONLY;
1652 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1654 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1655 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1660 txn->mt_free_pgs = mdb_midl_alloc();
1661 if (!txn->mt_free_pgs) {
1665 txn->mt_u.dirty_list = malloc(sizeof(ID2)*MDB_IDL_UM_SIZE);
1666 if (!txn->mt_u.dirty_list) {
1667 free(txn->mt_free_pgs);
1671 txn->mt_txnid = parent->mt_txnid;
1672 txn->mt_toggle = parent->mt_toggle;
1673 txn->mt_u.dirty_list[0].mid = 0;
1674 txn->mt_free_pgs[0] = 0;
1675 txn->mt_next_pgno = parent->mt_next_pgno;
1676 parent->mt_child = txn;
1677 txn->mt_parent = parent;
1678 txn->mt_numdbs = parent->mt_numdbs;
1679 txn->mt_dbxs = parent->mt_dbxs;
1680 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1681 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1682 mdb_cursor_shadow(parent, txn);
1685 rc = mdb_txn_renew0(txn);
1691 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1692 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1693 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1699 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1700 * @param[in] txn the transaction handle to reset
1703 mdb_txn_reset0(MDB_txn *txn)
1705 MDB_env *env = txn->mt_env;
1707 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1708 txn->mt_u.reader->mr_txnid = 0;
1714 /* close(free) all cursors */
1715 for (i=0; i<txn->mt_numdbs; i++) {
1716 if (txn->mt_cursors[i]) {
1718 while ((mc = txn->mt_cursors[i])) {
1719 txn->mt_cursors[i] = mc->mc_next;
1720 if (mc->mc_flags & C_ALLOCD)
1726 /* return all dirty pages to dpage list */
1727 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1728 dp = txn->mt_u.dirty_list[i].mptr;
1729 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1730 dp->mp_next = txn->mt_env->me_dpages;
1731 VGMEMP_FREE(txn->mt_env, dp);
1732 txn->mt_env->me_dpages = dp;
1734 /* large pages just get freed directly */
1735 VGMEMP_FREE(txn->mt_env, dp);
1740 if (txn->mt_parent) {
1741 txn->mt_parent->mt_child = NULL;
1742 free(txn->mt_free_pgs);
1743 free(txn->mt_u.dirty_list);
1746 if (mdb_midl_shrink(&txn->mt_free_pgs))
1747 env->me_free_pgs = txn->mt_free_pgs;
1750 while ((mop = txn->mt_env->me_pghead)) {
1751 txn->mt_env->me_pghead = mop->mo_next;
1754 txn->mt_env->me_pgfirst = 0;
1755 txn->mt_env->me_pglast = 0;
1758 /* The writer mutex was locked in mdb_txn_begin. */
1759 UNLOCK_MUTEX_W(env);
1764 mdb_txn_reset(MDB_txn *txn)
1769 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1770 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1771 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1773 mdb_txn_reset0(txn);
1777 mdb_txn_abort(MDB_txn *txn)
1782 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1783 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1784 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1787 mdb_txn_abort(txn->mt_child);
1789 mdb_txn_reset0(txn);
1794 mdb_txn_commit(MDB_txn *txn)
1802 pgno_t next, freecnt;
1805 assert(txn != NULL);
1806 assert(txn->mt_env != NULL);
1808 if (txn->mt_child) {
1809 mdb_txn_commit(txn->mt_child);
1810 txn->mt_child = NULL;
1815 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1816 if (txn->mt_numdbs > env->me_numdbs) {
1817 /* update the DB tables */
1818 int toggle = !env->me_db_toggle;
1822 ip = &env->me_dbs[toggle][env->me_numdbs];
1823 jp = &txn->mt_dbs[env->me_numdbs];
1824 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1825 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
1829 env->me_db_toggle = toggle;
1830 env->me_numdbs = txn->mt_numdbs;
1831 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1837 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1838 DPUTS("error flag is set, can't commit");
1840 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1845 /* Merge (and close) our cursors with parent's */
1846 mdb_cursor_merge(txn);
1848 if (txn->mt_parent) {
1854 /* Update parent's DB table */
1855 ip = &txn->mt_parent->mt_dbs[2];
1856 jp = &txn->mt_dbs[2];
1857 for (i = 2; i < txn->mt_numdbs; i++) {
1858 if (ip->md_root != jp->md_root)
1862 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1864 /* Append our free list to parent's */
1865 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1867 mdb_midl_free(txn->mt_free_pgs);
1869 /* Merge our dirty list with parent's */
1870 dst = txn->mt_parent->mt_u.dirty_list;
1871 src = txn->mt_u.dirty_list;
1872 x = mdb_mid2l_search(dst, src[1].mid);
1873 for (y=1; y<=src[0].mid; y++) {
1874 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1878 dst[x].mptr = src[y].mptr;
1881 for (; y<=src[0].mid; y++) {
1882 if (++x >= MDB_IDL_UM_MAX) {
1889 free(txn->mt_u.dirty_list);
1890 txn->mt_parent->mt_child = NULL;
1895 if (txn != env->me_txn) {
1896 DPUTS("attempt to commit unknown transaction");
1901 if (!txn->mt_u.dirty_list[0].mid)
1904 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1905 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1907 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1909 /* should only be one record now */
1910 if (env->me_pghead) {
1911 /* make sure first page of freeDB is touched and on freelist */
1912 mdb_page_search(&mc, NULL, 1);
1915 /* Delete IDLs we used from the free list */
1916 if (env->me_pgfirst) {
1921 key.mv_size = sizeof(cur);
1922 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
1925 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
1926 mdb_cursor_del(&mc, 0);
1928 env->me_pgfirst = 0;
1932 /* save to free list */
1934 freecnt = txn->mt_free_pgs[0];
1935 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1938 /* make sure last page of freeDB is touched and on freelist */
1939 key.mv_size = MAXKEYSIZE+1;
1941 mdb_page_search(&mc, &key, 1);
1943 mdb_midl_sort(txn->mt_free_pgs);
1947 ID *idl = txn->mt_free_pgs;
1948 DPRINTF("IDL write txn %zu root %zu num %zu",
1949 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1950 for (i=0; i<idl[0]; i++) {
1951 DPRINTF("IDL %zu", idl[i+1]);
1955 /* write to last page of freeDB */
1956 key.mv_size = sizeof(pgno_t);
1957 key.mv_data = &txn->mt_txnid;
1958 data.mv_data = txn->mt_free_pgs;
1959 /* The free list can still grow during this call,
1960 * despite the pre-emptive touches above. So check
1961 * and make sure the entire thing got written.
1964 freecnt = txn->mt_free_pgs[0];
1965 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1966 rc = mdb_cursor_put(&mc, &key, &data, 0);
1971 } while (freecnt != txn->mt_free_pgs[0]);
1973 /* should only be one record now */
1975 if (env->me_pghead) {
1981 mop = env->me_pghead;
1983 key.mv_size = sizeof(id);
1985 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1986 data.mv_data = mop->mo_pages;
1987 orig = mop->mo_pages[0];
1988 /* These steps may grow the freelist again
1989 * due to freed overflow pages...
1991 mdb_cursor_put(&mc, &key, &data, 0);
1992 if (mop == env->me_pghead) {
1993 /* could have been used again here */
1994 if (mop->mo_pages[0] != orig) {
1995 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1996 data.mv_data = mop->mo_pages;
1998 mdb_cursor_put(&mc, &key, &data, 0);
2000 env->me_pghead = NULL;
2003 /* was completely used up */
2004 mdb_cursor_del(&mc, 0);
2008 env->me_pgfirst = 0;
2011 /* Check for growth of freelist again */
2012 if (freecnt != txn->mt_free_pgs[0])
2015 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2016 if (mdb_midl_shrink(&txn->mt_free_pgs))
2017 env->me_free_pgs = txn->mt_free_pgs;
2020 /* Update DB root pointers. Their pages have already been
2021 * touched so this is all in-place and cannot fail.
2026 data.mv_size = sizeof(MDB_db);
2028 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2029 for (i = 2; i < txn->mt_numdbs; i++) {
2030 if (txn->mt_dbflags[i] & DB_DIRTY) {
2031 data.mv_data = &txn->mt_dbs[i];
2032 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2040 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2046 /* Windows actually supports scatter/gather I/O, but only on
2047 * unbuffered file handles. Since we're relying on the OS page
2048 * cache for all our data, that's self-defeating. So we just
2049 * write pages one at a time. We use the ov structure to set
2050 * the write offset, to at least save the overhead of a Seek
2054 memset(&ov, 0, sizeof(ov));
2055 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2057 dp = txn->mt_u.dirty_list[i].mptr;
2058 DPRINTF("committing page %zu", dp->mp_pgno);
2059 size = dp->mp_pgno * env->me_psize;
2060 ov.Offset = size & 0xffffffff;
2061 ov.OffsetHigh = size >> 16;
2062 ov.OffsetHigh >>= 16;
2063 /* clear dirty flag */
2064 dp->mp_flags &= ~P_DIRTY;
2065 wsize = env->me_psize;
2066 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2067 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2070 DPRINTF("WriteFile: %d", n);
2077 struct iovec iov[MDB_COMMIT_PAGES];
2081 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2082 dp = txn->mt_u.dirty_list[i].mptr;
2083 if (dp->mp_pgno != next) {
2085 rc = writev(env->me_fd, iov, n);
2089 DPUTS("short write, filesystem full?");
2091 DPRINTF("writev: %s", strerror(n));
2098 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2101 DPRINTF("committing page %zu", dp->mp_pgno);
2102 iov[n].iov_len = env->me_psize;
2103 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2104 iov[n].iov_base = (char *)dp;
2105 size += iov[n].iov_len;
2106 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2107 /* clear dirty flag */
2108 dp->mp_flags &= ~P_DIRTY;
2109 if (++n >= MDB_COMMIT_PAGES) {
2119 rc = writev(env->me_fd, iov, n);
2123 DPUTS("short write, filesystem full?");
2125 DPRINTF("writev: %s", strerror(n));
2132 /* Drop the dirty pages.
2134 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2135 dp = txn->mt_u.dirty_list[i].mptr;
2136 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2137 dp->mp_next = txn->mt_env->me_dpages;
2138 VGMEMP_FREE(txn->mt_env, dp);
2139 txn->mt_env->me_dpages = dp;
2141 VGMEMP_FREE(txn->mt_env, dp);
2144 txn->mt_u.dirty_list[i].mid = 0;
2146 txn->mt_u.dirty_list[0].mid = 0;
2148 if ((n = mdb_env_sync(env, 0)) != 0 ||
2149 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2153 env->me_wtxnid = txn->mt_txnid;
2157 /* update the DB tables */
2159 int toggle = !env->me_db_toggle;
2163 ip = &env->me_dbs[toggle][2];
2164 jp = &txn->mt_dbs[2];
2165 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
2166 for (i = 2; i < txn->mt_numdbs; i++) {
2167 if (ip->md_root != jp->md_root)
2172 env->me_db_toggle = toggle;
2173 env->me_numdbs = txn->mt_numdbs;
2174 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
2177 UNLOCK_MUTEX_W(env);
2183 /** Read the environment parameters of a DB environment before
2184 * mapping it into memory.
2185 * @param[in] env the environment handle
2186 * @param[out] meta address of where to store the meta information
2187 * @return 0 on success, non-zero on failure.
2190 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2197 /* We don't know the page size yet, so use a minimum value.
2201 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2203 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2208 else if (rc != MDB_PAGESIZE) {
2212 DPRINTF("read: %s", strerror(err));
2216 p = (MDB_page *)&pbuf;
2218 if (!F_ISSET(p->mp_flags, P_META)) {
2219 DPRINTF("page %zu not a meta page", p->mp_pgno);
2224 if (m->mm_magic != MDB_MAGIC) {
2225 DPUTS("meta has invalid magic");
2229 if (m->mm_version != MDB_VERSION) {
2230 DPRINTF("database is version %u, expected version %u",
2231 m->mm_version, MDB_VERSION);
2232 return MDB_VERSION_MISMATCH;
2235 memcpy(meta, m, sizeof(*m));
2239 /** Write the environment parameters of a freshly created DB environment.
2240 * @param[in] env the environment handle
2241 * @param[out] meta address of where to store the meta information
2242 * @return 0 on success, non-zero on failure.
2245 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2252 DPUTS("writing new meta page");
2254 GET_PAGESIZE(psize);
2256 meta->mm_magic = MDB_MAGIC;
2257 meta->mm_version = MDB_VERSION;
2258 meta->mm_psize = psize;
2259 meta->mm_last_pg = 1;
2260 meta->mm_flags = env->me_flags & 0xffff;
2261 meta->mm_flags |= MDB_INTEGERKEY;
2262 meta->mm_dbs[0].md_root = P_INVALID;
2263 meta->mm_dbs[1].md_root = P_INVALID;
2265 p = calloc(2, psize);
2267 p->mp_flags = P_META;
2270 memcpy(m, meta, sizeof(*meta));
2272 q = (MDB_page *)((char *)p + psize);
2275 q->mp_flags = P_META;
2278 memcpy(m, meta, sizeof(*meta));
2283 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2284 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2287 rc = write(env->me_fd, p, psize * 2);
2288 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2294 /** Update the environment info to commit a transaction.
2295 * @param[in] txn the transaction that's being committed
2296 * @return 0 on success, non-zero on failure.
2299 mdb_env_write_meta(MDB_txn *txn)
2302 MDB_meta meta, metab;
2304 int rc, len, toggle;
2310 assert(txn != NULL);
2311 assert(txn->mt_env != NULL);
2313 toggle = !txn->mt_toggle;
2314 DPRINTF("writing meta page %d for root page %zu",
2315 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2319 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2320 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2322 ptr = (char *)&meta;
2323 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2324 len = sizeof(MDB_meta) - off;
2327 meta.mm_dbs[0] = txn->mt_dbs[0];
2328 meta.mm_dbs[1] = txn->mt_dbs[1];
2329 meta.mm_last_pg = txn->mt_next_pgno - 1;
2330 meta.mm_txnid = txn->mt_txnid;
2333 off += env->me_psize;
2336 /* Write to the SYNC fd */
2339 memset(&ov, 0, sizeof(ov));
2341 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2344 rc = pwrite(env->me_mfd, ptr, len, off);
2349 DPUTS("write failed, disk error?");
2350 /* On a failure, the pagecache still contains the new data.
2351 * Write some old data back, to prevent it from being used.
2352 * Use the non-SYNC fd; we know it will fail anyway.
2354 meta.mm_last_pg = metab.mm_last_pg;
2355 meta.mm_txnid = metab.mm_txnid;
2357 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2359 r2 = pwrite(env->me_fd, ptr, len, off);
2361 env->me_flags |= MDB_FATAL_ERROR;
2364 /* Memory ordering issues are irrelevant; since the entire writer
2365 * is wrapped by wmutex, all of these changes will become visible
2366 * after the wmutex is unlocked. Since the DB is multi-version,
2367 * readers will get consistent data regardless of how fresh or
2368 * how stale their view of these values is.
2370 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
2371 txn->mt_env->me_txns->mti_me_toggle = toggle;
2372 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2373 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
2378 /** Check both meta pages to see which one is newer.
2379 * @param[in] env the environment handle
2380 * @param[out] which address of where to store the meta toggle ID
2381 * @return 0 on success, non-zero on failure.
2384 mdb_env_read_meta(MDB_env *env, int *which)
2388 assert(env != NULL);
2390 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2393 DPRINTF("Using meta page %d", toggle);
2400 mdb_env_create(MDB_env **env)
2404 e = calloc(1, sizeof(MDB_env));
2408 e->me_free_pgs = mdb_midl_alloc();
2409 if (!e->me_free_pgs) {
2413 e->me_maxreaders = DEFAULT_READERS;
2415 e->me_fd = INVALID_HANDLE_VALUE;
2416 e->me_lfd = INVALID_HANDLE_VALUE;
2417 e->me_mfd = INVALID_HANDLE_VALUE;
2418 VGMEMP_CREATE(e,0,0);
2424 mdb_env_set_mapsize(MDB_env *env, size_t size)
2428 env->me_mapsize = size;
2430 env->me_maxpg = env->me_mapsize / env->me_psize;
2435 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2439 env->me_maxdbs = dbs;
2444 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2446 if (env->me_map || readers < 1)
2448 env->me_maxreaders = readers;
2453 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2455 if (!env || !readers)
2457 *readers = env->me_maxreaders;
2461 /** Further setup required for opening an MDB environment
2464 mdb_env_open2(MDB_env *env, unsigned int flags)
2466 int i, newenv = 0, toggle;
2470 env->me_flags = flags;
2472 memset(&meta, 0, sizeof(meta));
2474 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2477 DPUTS("new mdbenv");
2481 if (!env->me_mapsize) {
2482 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2488 LONG sizelo, sizehi;
2489 sizelo = env->me_mapsize & 0xffffffff;
2490 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2492 /* Windows won't create mappings for zero length files.
2493 * Just allocate the maxsize right now.
2496 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2497 if (!SetEndOfFile(env->me_fd))
2499 SetFilePointer(env->me_fd, 0, NULL, 0);
2501 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2502 sizehi, sizelo, NULL);
2505 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2513 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2515 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2517 if (env->me_map == MAP_FAILED) {
2524 meta.mm_mapsize = env->me_mapsize;
2525 if (flags & MDB_FIXEDMAP)
2526 meta.mm_address = env->me_map;
2527 i = mdb_env_init_meta(env, &meta);
2528 if (i != MDB_SUCCESS) {
2529 munmap(env->me_map, env->me_mapsize);
2533 env->me_psize = meta.mm_psize;
2535 env->me_maxpg = env->me_mapsize / env->me_psize;
2537 p = (MDB_page *)env->me_map;
2538 env->me_metas[0] = METADATA(p);
2539 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2541 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
2544 DPRINTF("opened database version %u, pagesize %u",
2545 env->me_metas[toggle]->mm_version, env->me_psize);
2546 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
2547 DPRINTF("entries: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
2548 DPRINTF("branch pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
2549 DPRINTF("leaf pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
2550 DPRINTF("overflow pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
2551 DPRINTF("root: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
2557 /** Release a reader thread's slot in the reader lock table.
2558 * This function is called automatically when a thread exits.
2559 * Windows doesn't support destructor callbacks for thread-specific storage,
2560 * so this function is not compiled there.
2561 * @param[in] ptr This points to the slot in the reader lock table.
2564 mdb_env_reader_dest(void *ptr)
2566 MDB_reader *reader = ptr;
2568 reader->mr_txnid = 0;
2574 /** Downgrade the exclusive lock on the region back to shared */
2576 mdb_env_share_locks(MDB_env *env)
2580 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2582 env->me_txns->mti_me_toggle = toggle;
2583 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2588 /* First acquire a shared lock. The Unlock will
2589 * then release the existing exclusive lock.
2591 memset(&ov, 0, sizeof(ov));
2592 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2593 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2597 struct flock lock_info;
2598 /* The shared lock replaces the existing lock */
2599 memset((void *)&lock_info, 0, sizeof(lock_info));
2600 lock_info.l_type = F_RDLCK;
2601 lock_info.l_whence = SEEK_SET;
2602 lock_info.l_start = 0;
2603 lock_info.l_len = 1;
2604 fcntl(env->me_lfd, F_SETLK, &lock_info);
2608 #if defined(_WIN32) || defined(__APPLE__)
2610 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2612 * @(#) $Revision: 5.1 $
2613 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2614 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2616 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2620 * Please do not copyright this code. This code is in the public domain.
2622 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2623 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2624 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2625 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2626 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2627 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2628 * PERFORMANCE OF THIS SOFTWARE.
2631 * chongo <Landon Curt Noll> /\oo/\
2632 * http://www.isthe.com/chongo/
2634 * Share and Enjoy! :-)
2637 typedef unsigned long long mdb_hash_t;
2638 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2640 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2641 * @param[in] str string to hash
2642 * @param[in] hval initial value for hash
2643 * @return 64 bit hash
2645 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2646 * hval arg on the first call.
2649 mdb_hash_str(char *str, mdb_hash_t hval)
2651 unsigned char *s = (unsigned char *)str; /* unsigned string */
2653 * FNV-1a hash each octet of the string
2656 /* xor the bottom with the current octet */
2657 hval ^= (mdb_hash_t)*s++;
2659 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2660 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2661 (hval << 7) + (hval << 8) + (hval << 40);
2663 /* return our new hash value */
2667 /** Hash the string and output the hash in hex.
2668 * @param[in] str string to hash
2669 * @param[out] hexbuf an array of 17 chars to hold the hash
2672 mdb_hash_hex(char *str, char *hexbuf)
2675 mdb_hash_t h = mdb_hash_str(str, MDB_HASH_INIT);
2676 for (i=0; i<8; i++) {
2677 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2683 /** Open and/or initialize the lock region for the environment.
2684 * @param[in] env The MDB environment.
2685 * @param[in] lpath The pathname of the file used for the lock region.
2686 * @param[in] mode The Unix permissions for the file, if we create it.
2687 * @param[out] excl Set to true if we got an exclusive lock on the region.
2688 * @return 0 on success, non-zero on failure.
2691 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2699 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2700 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2701 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2705 /* Try to get exclusive lock. If we succeed, then
2706 * nobody is using the lock region and we should initialize it.
2709 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2713 memset(&ov, 0, sizeof(ov));
2714 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2720 size = GetFileSize(env->me_lfd, NULL);
2722 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2726 /* Try to get exclusive lock. If we succeed, then
2727 * nobody is using the lock region and we should initialize it.
2730 struct flock lock_info;
2731 memset((void *)&lock_info, 0, sizeof(lock_info));
2732 lock_info.l_type = F_WRLCK;
2733 lock_info.l_whence = SEEK_SET;
2734 lock_info.l_start = 0;
2735 lock_info.l_len = 1;
2736 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2740 lock_info.l_type = F_RDLCK;
2741 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2748 size = lseek(env->me_lfd, 0, SEEK_END);
2750 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2751 if (size < rsize && *excl) {
2753 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2754 if (!SetEndOfFile(env->me_lfd)) {
2759 if (ftruncate(env->me_lfd, rsize) != 0) {
2766 size = rsize - sizeof(MDB_txninfo);
2767 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2772 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2778 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2780 if (!env->me_txns) {
2785 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2787 if (m == MAP_FAILED) {
2788 env->me_txns = NULL;
2798 if (!mdb_sec_inited) {
2799 InitializeSecurityDescriptor(&mdb_null_sd,
2800 SECURITY_DESCRIPTOR_REVISION);
2801 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2802 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2803 mdb_all_sa.bInheritHandle = FALSE;
2804 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2807 mdb_hash_hex(lpath, hexbuf);
2808 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2809 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2810 if (!env->me_rmutex) {
2814 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2815 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2816 if (!env->me_wmutex) {
2823 mdb_hash_hex(lpath, hexbuf);
2824 sprintf(env->me_txns->mti_rmname, "MDBr%s", hexbuf);
2825 if (sem_unlink(env->me_txns->mti_rmname)) {
2827 if (rc != ENOENT && rc != EINVAL)
2830 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2831 if (!env->me_rmutex) {
2835 sprintf(env->me_txns->mti_wmname, "MDBw%s", hexbuf);
2836 if (sem_unlink(env->me_txns->mti_wmname)) {
2838 if (rc != ENOENT && rc != EINVAL)
2841 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2842 if (!env->me_wmutex) {
2846 #else /* __APPLE__ */
2847 pthread_mutexattr_t mattr;
2849 pthread_mutexattr_init(&mattr);
2850 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2854 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2855 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2856 #endif /* __APPLE__ */
2858 env->me_txns->mti_version = MDB_VERSION;
2859 env->me_txns->mti_magic = MDB_MAGIC;
2860 env->me_txns->mti_txnid = 0;
2861 env->me_txns->mti_numreaders = 0;
2862 env->me_txns->mti_me_toggle = 0;
2865 if (env->me_txns->mti_magic != MDB_MAGIC) {
2866 DPUTS("lock region has invalid magic");
2870 if (env->me_txns->mti_version != MDB_VERSION) {
2871 DPRINTF("lock region is version %u, expected version %u",
2872 env->me_txns->mti_version, MDB_VERSION);
2873 rc = MDB_VERSION_MISMATCH;
2877 if (rc != EACCES && rc != EAGAIN) {
2881 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2882 if (!env->me_rmutex) {
2886 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2887 if (!env->me_wmutex) {
2893 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2894 if (!env->me_rmutex) {
2898 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2899 if (!env->me_wmutex) {
2909 env->me_lfd = INVALID_HANDLE_VALUE;
2914 /** The name of the lock file in the DB environment */
2915 #define LOCKNAME "/lock.mdb"
2916 /** The name of the data file in the DB environment */
2917 #define DATANAME "/data.mdb"
2918 /** The suffix of the lock file when no subdir is used */
2919 #define LOCKSUFF "-lock"
2922 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2924 int oflags, rc, len, excl;
2925 char *lpath, *dpath;
2928 if (flags & MDB_NOSUBDIR) {
2929 rc = len + sizeof(LOCKSUFF) + len + 1;
2931 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
2936 if (flags & MDB_NOSUBDIR) {
2937 dpath = lpath + len + sizeof(LOCKSUFF);
2938 sprintf(lpath, "%s" LOCKSUFF, path);
2939 strcpy(dpath, path);
2941 dpath = lpath + len + sizeof(LOCKNAME);
2942 sprintf(lpath, "%s" LOCKNAME, path);
2943 sprintf(dpath, "%s" DATANAME, path);
2946 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2951 if (F_ISSET(flags, MDB_RDONLY)) {
2952 oflags = GENERIC_READ;
2953 len = OPEN_EXISTING;
2955 oflags = GENERIC_READ|GENERIC_WRITE;
2958 mode = FILE_ATTRIBUTE_NORMAL;
2959 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2960 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2965 if (F_ISSET(flags, MDB_RDONLY))
2968 oflags = O_RDWR | O_CREAT;
2970 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2976 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2977 /* synchronous fd for meta writes */
2979 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2980 mode |= FILE_FLAG_WRITE_THROUGH;
2981 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2982 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2987 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2988 oflags |= MDB_DSYNC;
2989 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
2994 env->me_path = strdup(path);
2995 DPRINTF("opened dbenv %p", (void *) env);
2996 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
2997 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
2999 mdb_env_share_locks(env);
3000 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3001 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
3002 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
3008 if (env->me_fd != INVALID_HANDLE_VALUE) {
3010 env->me_fd = INVALID_HANDLE_VALUE;
3012 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3014 env->me_lfd = INVALID_HANDLE_VALUE;
3022 mdb_env_close(MDB_env *env)
3029 VGMEMP_DESTROY(env);
3030 while (env->me_dpages) {
3031 dp = env->me_dpages;
3032 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3033 env->me_dpages = dp->mp_next;
3037 free(env->me_dbs[1]);
3038 free(env->me_dbs[0]);
3042 LAZY_RWLOCK_DESTROY(&env->me_dblock);
3043 pthread_key_delete(env->me_txkey);
3046 munmap(env->me_map, env->me_mapsize);
3051 pid_t pid = getpid();
3053 for (i=0; i<env->me_txns->mti_numreaders; i++)
3054 if (env->me_txns->mti_readers[i].mr_pid == pid)
3055 env->me_txns->mti_readers[i].mr_pid = 0;
3056 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3059 mdb_midl_free(env->me_free_pgs);
3063 /** Compare two items pointing at aligned size_t's */
3065 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3067 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3068 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3071 /** Compare two items pointing at aligned int's */
3073 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3075 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3076 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3079 /** Compare two items pointing at ints of unknown alignment.
3080 * Nodes and keys are guaranteed to be 2-byte aligned.
3083 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3085 #if BYTE_ORDER == LITTLE_ENDIAN
3086 unsigned short *u, *c;
3089 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3090 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3093 } while(!x && u > (unsigned short *)a->mv_data);
3096 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3100 /** Compare two items lexically */
3102 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3109 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3115 diff = memcmp(a->mv_data, b->mv_data, len);
3116 return diff ? diff : len_diff<0 ? -1 : len_diff;
3119 /** Compare two items in reverse byte order */
3121 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3123 const unsigned char *p1, *p2, *p1_lim;
3127 p1_lim = (const unsigned char *)a->mv_data;
3128 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3129 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3131 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3137 while (p1 > p1_lim) {
3138 diff = *--p1 - *--p2;
3142 return len_diff<0 ? -1 : len_diff;
3145 /** Search for key within a page, using binary search.
3146 * Returns the smallest entry larger or equal to the key.
3147 * If exactp is non-null, stores whether the found entry was an exact match
3148 * in *exactp (1 or 0).
3149 * Updates the cursor index with the index of the found entry.
3150 * If no entry larger or equal to the key is found, returns NULL.
3153 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3155 unsigned int i = 0, nkeys;
3158 MDB_page *mp = mc->mc_pg[mc->mc_top];
3159 MDB_node *node = NULL;
3164 nkeys = NUMKEYS(mp);
3169 COPY_PGNO(pgno, mp->mp_pgno);
3170 DPRINTF("searching %u keys in %s %spage %zu",
3171 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3178 low = IS_LEAF(mp) ? 0 : 1;
3180 cmp = mc->mc_dbx->md_cmp;
3182 /* Branch pages have no data, so if using integer keys,
3183 * alignment is guaranteed. Use faster mdb_cmp_int.
3185 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3186 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3193 nodekey.mv_size = mc->mc_db->md_pad;
3194 node = NODEPTR(mp, 0); /* fake */
3195 while (low <= high) {
3196 i = (low + high) >> 1;
3197 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3198 rc = cmp(key, &nodekey);
3199 DPRINTF("found leaf index %u [%s], rc = %i",
3200 i, DKEY(&nodekey), rc);
3209 while (low <= high) {
3210 i = (low + high) >> 1;
3212 node = NODEPTR(mp, i);
3213 nodekey.mv_size = NODEKSZ(node);
3214 nodekey.mv_data = NODEKEY(node);
3216 rc = cmp(key, &nodekey);
3219 DPRINTF("found leaf index %u [%s], rc = %i",
3220 i, DKEY(&nodekey), rc);
3222 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3223 i, DKEY(&nodekey), NODEPGNO(node), rc);
3234 if (rc > 0) { /* Found entry is less than the key. */
3235 i++; /* Skip to get the smallest entry larger than key. */
3237 node = NODEPTR(mp, i);
3240 *exactp = (rc == 0);
3241 /* store the key index */
3242 mc->mc_ki[mc->mc_top] = i;
3244 /* There is no entry larger or equal to the key. */
3247 /* nodeptr is fake for LEAF2 */
3253 mdb_cursor_adjust(MDB_cursor *mc, func)
3257 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3258 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3265 /** Pop a page off the top of the cursor's stack. */
3267 mdb_cursor_pop(MDB_cursor *mc)
3272 top = mc->mc_pg[mc->mc_top];
3277 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3278 mc->mc_dbi, (void *) mc);
3282 /** Push a page onto the top of the cursor's stack. */
3284 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3286 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3287 mc->mc_dbi, (void *) mc);
3289 if (mc->mc_snum >= CURSOR_STACK) {
3290 assert(mc->mc_snum < CURSOR_STACK);
3294 mc->mc_top = mc->mc_snum++;
3295 mc->mc_pg[mc->mc_top] = mp;
3296 mc->mc_ki[mc->mc_top] = 0;
3301 /** Find the address of the page corresponding to a given page number.
3302 * @param[in] txn the transaction for this access.
3303 * @param[in] pgno the page number for the page to retrieve.
3304 * @param[out] ret address of a pointer where the page's address will be stored.
3305 * @return 0 on success, non-zero on failure.
3308 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3312 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3314 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3315 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3316 p = txn->mt_u.dirty_list[x].mptr;
3320 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
3321 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3325 DPRINTF("page %zu not found", pgno);
3328 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3331 /** Search for the page a given key should be in.
3332 * Pushes parent pages on the cursor stack. This function continues a
3333 * search on a cursor that has already been initialized. (Usually by
3334 * #mdb_page_search() but also by #mdb_node_move().)
3335 * @param[in,out] mc the cursor for this operation.
3336 * @param[in] key the key to search for. If NULL, search for the lowest
3337 * page. (This is used by #mdb_cursor_first().)
3338 * @param[in] modify If true, visited pages are updated with new page numbers.
3339 * @return 0 on success, non-zero on failure.
3342 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3344 MDB_page *mp = mc->mc_pg[mc->mc_top];
3349 while (IS_BRANCH(mp)) {
3353 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3354 assert(NUMKEYS(mp) > 1);
3355 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3357 if (key == NULL) /* Initialize cursor to first page. */
3359 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3360 /* cursor to last page */
3364 node = mdb_node_search(mc, key, &exact);
3366 i = NUMKEYS(mp) - 1;
3368 i = mc->mc_ki[mc->mc_top];
3377 DPRINTF("following index %u for key [%s]",
3379 assert(i < NUMKEYS(mp));
3380 node = NODEPTR(mp, i);
3382 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3385 mc->mc_ki[mc->mc_top] = i;
3386 if ((rc = mdb_cursor_push(mc, mp)))
3390 if ((rc = mdb_page_touch(mc)) != 0)
3392 mp = mc->mc_pg[mc->mc_top];
3397 DPRINTF("internal error, index points to a %02X page!?",
3399 return MDB_CORRUPTED;
3402 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3403 key ? DKEY(key) : NULL);
3408 /** Search for the page a given key should be in.
3409 * Pushes parent pages on the cursor stack. This function just sets up
3410 * the search; it finds the root page for \b mc's database and sets this
3411 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3412 * called to complete the search.
3413 * @param[in,out] mc the cursor for this operation.
3414 * @param[in] key the key to search for. If NULL, search for the lowest
3415 * page. (This is used by #mdb_cursor_first().)
3416 * @param[in] modify If true, visited pages are updated with new page numbers.
3417 * @return 0 on success, non-zero on failure.
3420 mdb_page_search(MDB_cursor *mc, MDB_val *key, int modify)
3425 /* Make sure the txn is still viable, then find the root from
3426 * the txn's db table.
3428 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3429 DPUTS("transaction has failed, must abort");
3432 /* Make sure we're using an up-to-date root */
3433 if (mc->mc_dbi > MAIN_DBI) {
3434 if ((*mc->mc_dbflag & DB_STALE) ||
3435 (modify && !(*mc->mc_dbflag & DB_DIRTY))) {
3437 unsigned char dbflag = 0;
3438 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3439 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, modify);
3442 if (*mc->mc_dbflag & DB_STALE) {
3445 MDB_node *leaf = mdb_node_search(&mc2,
3446 &mc->mc_dbx->md_name, &exact);
3448 return MDB_NOTFOUND;
3449 mdb_node_read(mc->mc_txn, leaf, &data);
3450 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3454 *mc->mc_dbflag = dbflag;
3457 root = mc->mc_db->md_root;
3459 if (root == P_INVALID) { /* Tree is empty. */
3460 DPUTS("tree is empty");
3461 return MDB_NOTFOUND;
3466 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3472 DPRINTF("db %u root page %zu has flags 0x%X",
3473 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3476 if ((rc = mdb_page_touch(mc)))
3480 return mdb_page_search_root(mc, key, modify);
3483 /** Return the data associated with a given node.
3484 * @param[in] txn The transaction for this operation.
3485 * @param[in] leaf The node being read.
3486 * @param[out] data Updated to point to the node's data.
3487 * @return 0 on success, non-zero on failure.
3490 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3492 MDB_page *omp; /* overflow page */
3496 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3497 data->mv_size = NODEDSZ(leaf);
3498 data->mv_data = NODEDATA(leaf);
3502 /* Read overflow data.
3504 data->mv_size = NODEDSZ(leaf);
3505 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3506 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3507 DPRINTF("read overflow page %zu failed", pgno);
3510 data->mv_data = METADATA(omp);
3516 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3517 MDB_val *key, MDB_val *data)
3526 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3528 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3531 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3535 mdb_cursor_init(&mc, txn, dbi, &mx);
3536 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3539 /** Find a sibling for a page.
3540 * Replaces the page at the top of the cursor's stack with the
3541 * specified sibling, if one exists.
3542 * @param[in] mc The cursor for this operation.
3543 * @param[in] move_right Non-zero if the right sibling is requested,
3544 * otherwise the left sibling.
3545 * @return 0 on success, non-zero on failure.
3548 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3554 if (mc->mc_snum < 2) {
3555 return MDB_NOTFOUND; /* root has no siblings */
3559 DPRINTF("parent page is page %zu, index %u",
3560 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3562 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3563 : (mc->mc_ki[mc->mc_top] == 0)) {
3564 DPRINTF("no more keys left, moving to %s sibling",
3565 move_right ? "right" : "left");
3566 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3570 mc->mc_ki[mc->mc_top]++;
3572 mc->mc_ki[mc->mc_top]--;
3573 DPRINTF("just moving to %s index key %u",
3574 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3576 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3578 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3579 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3582 mdb_cursor_push(mc, mp);
3587 /** Move the cursor to the next data item. */
3589 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3595 if (mc->mc_flags & C_EOF) {
3596 return MDB_NOTFOUND;
3599 assert(mc->mc_flags & C_INITIALIZED);
3601 mp = mc->mc_pg[mc->mc_top];
3603 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3604 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3605 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3606 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3607 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3608 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3612 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3613 if (op == MDB_NEXT_DUP)
3614 return MDB_NOTFOUND;
3618 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3620 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3621 DPUTS("=====> move to next sibling page");
3622 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3623 mc->mc_flags |= C_EOF;
3624 mc->mc_flags &= ~C_INITIALIZED;
3625 return MDB_NOTFOUND;
3627 mp = mc->mc_pg[mc->mc_top];
3628 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3630 mc->mc_ki[mc->mc_top]++;
3632 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3633 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3636 key->mv_size = mc->mc_db->md_pad;
3637 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3641 assert(IS_LEAF(mp));
3642 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3644 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3645 mdb_xcursor_init1(mc, leaf);
3648 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3651 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3652 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3653 if (rc != MDB_SUCCESS)
3658 MDB_SET_KEY(leaf, key);
3662 /** Move the cursor to the previous data item. */
3664 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3670 assert(mc->mc_flags & C_INITIALIZED);
3672 mp = mc->mc_pg[mc->mc_top];
3674 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3675 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3676 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3677 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3678 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3679 if (op != MDB_PREV || rc == MDB_SUCCESS)
3682 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3683 if (op == MDB_PREV_DUP)
3684 return MDB_NOTFOUND;
3689 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3691 if (mc->mc_ki[mc->mc_top] == 0) {
3692 DPUTS("=====> move to prev sibling page");
3693 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3694 mc->mc_flags &= ~C_INITIALIZED;
3695 return MDB_NOTFOUND;
3697 mp = mc->mc_pg[mc->mc_top];
3698 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3699 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3701 mc->mc_ki[mc->mc_top]--;
3703 mc->mc_flags &= ~C_EOF;
3705 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3706 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3709 key->mv_size = mc->mc_db->md_pad;
3710 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3714 assert(IS_LEAF(mp));
3715 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3717 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3718 mdb_xcursor_init1(mc, leaf);
3721 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3724 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3725 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3726 if (rc != MDB_SUCCESS)
3731 MDB_SET_KEY(leaf, key);
3735 /** Set the cursor on a specific data item. */
3737 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3738 MDB_cursor_op op, int *exactp)
3747 assert(key->mv_size > 0);
3749 /* See if we're already on the right page */
3750 if (mc->mc_flags & C_INITIALIZED) {
3753 mp = mc->mc_pg[mc->mc_top];
3755 mc->mc_ki[mc->mc_top] = 0;
3756 return MDB_NOTFOUND;
3758 if (mp->mp_flags & P_LEAF2) {
3759 nodekey.mv_size = mc->mc_db->md_pad;
3760 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3762 leaf = NODEPTR(mp, 0);
3763 MDB_SET_KEY(leaf, &nodekey);
3765 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3767 /* Probably happens rarely, but first node on the page
3768 * was the one we wanted.
3770 mc->mc_ki[mc->mc_top] = 0;
3771 leaf = NODEPTR(mp, 0);
3778 unsigned int nkeys = NUMKEYS(mp);
3780 if (mp->mp_flags & P_LEAF2) {
3781 nodekey.mv_data = LEAF2KEY(mp,
3782 nkeys-1, nodekey.mv_size);
3784 leaf = NODEPTR(mp, nkeys-1);
3785 MDB_SET_KEY(leaf, &nodekey);
3787 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3789 /* last node was the one we wanted */
3790 mc->mc_ki[mc->mc_top] = nkeys-1;
3791 leaf = NODEPTR(mp, nkeys-1);
3797 /* This is definitely the right page, skip search_page */
3802 /* If any parents have right-sibs, search.
3803 * Otherwise, there's nothing further.
3805 for (i=0; i<mc->mc_top; i++)
3807 NUMKEYS(mc->mc_pg[i])-1)
3809 if (i == mc->mc_top) {
3810 /* There are no other pages */
3811 mc->mc_ki[mc->mc_top] = nkeys;
3812 return MDB_NOTFOUND;
3816 /* There are no other pages */
3817 mc->mc_ki[mc->mc_top] = 0;
3818 return MDB_NOTFOUND;
3822 rc = mdb_page_search(mc, key, 0);
3823 if (rc != MDB_SUCCESS)
3826 mp = mc->mc_pg[mc->mc_top];
3827 assert(IS_LEAF(mp));
3830 leaf = mdb_node_search(mc, key, exactp);
3831 if (exactp != NULL && !*exactp) {
3832 /* MDB_SET specified and not an exact match. */
3833 return MDB_NOTFOUND;
3837 DPUTS("===> inexact leaf not found, goto sibling");
3838 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3839 return rc; /* no entries matched */
3840 mp = mc->mc_pg[mc->mc_top];
3841 assert(IS_LEAF(mp));
3842 leaf = NODEPTR(mp, 0);
3846 mc->mc_flags |= C_INITIALIZED;
3847 mc->mc_flags &= ~C_EOF;
3850 key->mv_size = mc->mc_db->md_pad;
3851 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3855 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3856 mdb_xcursor_init1(mc, leaf);
3859 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3860 if (op == MDB_SET || op == MDB_SET_RANGE) {
3861 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3864 if (op == MDB_GET_BOTH) {
3870 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3871 if (rc != MDB_SUCCESS)
3874 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3876 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3878 rc = mc->mc_dbx->md_dcmp(data, &d2);
3880 if (op == MDB_GET_BOTH || rc > 0)
3881 return MDB_NOTFOUND;
3886 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3887 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3892 /* The key already matches in all other cases */
3893 if (op == MDB_SET_RANGE)
3894 MDB_SET_KEY(leaf, key);
3895 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3900 /** Move the cursor to the first item in the database. */
3902 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3907 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3908 rc = mdb_page_search(mc, NULL, 0);
3909 if (rc != MDB_SUCCESS)
3912 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3914 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3915 mc->mc_flags |= C_INITIALIZED;
3916 mc->mc_flags &= ~C_EOF;
3918 mc->mc_ki[mc->mc_top] = 0;
3920 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3921 key->mv_size = mc->mc_db->md_pad;
3922 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3927 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3928 mdb_xcursor_init1(mc, leaf);
3929 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3934 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3935 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3939 MDB_SET_KEY(leaf, key);
3943 /** Move the cursor to the last item in the database. */
3945 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3951 lkey.mv_size = MAXKEYSIZE+1;
3952 lkey.mv_data = NULL;
3954 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3955 rc = mdb_page_search(mc, &lkey, 0);
3956 if (rc != MDB_SUCCESS)
3959 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3961 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3962 mc->mc_flags |= C_INITIALIZED;
3963 mc->mc_flags &= ~C_EOF;
3965 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3967 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3968 key->mv_size = mc->mc_db->md_pad;
3969 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3974 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3975 mdb_xcursor_init1(mc, leaf);
3976 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3981 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3982 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3987 MDB_SET_KEY(leaf, key);
3992 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4002 case MDB_GET_BOTH_RANGE:
4003 if (data == NULL || mc->mc_xcursor == NULL) {
4010 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4012 } else if (op == MDB_SET_RANGE)
4013 rc = mdb_cursor_set(mc, key, data, op, NULL);
4015 rc = mdb_cursor_set(mc, key, data, op, &exact);
4017 case MDB_GET_MULTIPLE:
4019 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4020 !(mc->mc_flags & C_INITIALIZED)) {
4025 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4026 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4029 case MDB_NEXT_MULTIPLE:
4031 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4035 if (!(mc->mc_flags & C_INITIALIZED))
4036 rc = mdb_cursor_first(mc, key, data);
4038 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4039 if (rc == MDB_SUCCESS) {
4040 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4043 mx = &mc->mc_xcursor->mx_cursor;
4044 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4046 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4047 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4055 case MDB_NEXT_NODUP:
4056 if (!(mc->mc_flags & C_INITIALIZED))
4057 rc = mdb_cursor_first(mc, key, data);
4059 rc = mdb_cursor_next(mc, key, data, op);
4063 case MDB_PREV_NODUP:
4064 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
4065 rc = mdb_cursor_last(mc, key, data);
4067 rc = mdb_cursor_prev(mc, key, data, op);
4070 rc = mdb_cursor_first(mc, key, data);
4074 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4075 !(mc->mc_flags & C_INITIALIZED) ||
4076 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4080 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4083 rc = mdb_cursor_last(mc, key, data);
4087 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4088 !(mc->mc_flags & C_INITIALIZED) ||
4089 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4093 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4096 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4104 /** Touch all the pages in the cursor stack.
4105 * Makes sure all the pages are writable, before attempting a write operation.
4106 * @param[in] mc The cursor to operate on.
4109 mdb_cursor_touch(MDB_cursor *mc)
4113 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4115 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4116 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 1);
4119 *mc->mc_dbflag = DB_DIRTY;
4121 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4122 rc = mdb_page_touch(mc);
4126 mc->mc_top = mc->mc_snum-1;
4131 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4134 MDB_node *leaf = NULL;
4135 MDB_val xdata, *rdata, dkey;
4139 unsigned int mcount = 0;
4143 char dbuf[MAXKEYSIZE+1];
4144 unsigned int nflags;
4147 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4150 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4151 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4155 if (flags == MDB_CURRENT) {
4156 if (!(mc->mc_flags & C_INITIALIZED))
4159 } else if (mc->mc_db->md_root == P_INVALID) {
4161 /* new database, write a root leaf page */
4162 DPUTS("allocating new root leaf page");
4163 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4167 mdb_cursor_push(mc, np);
4168 mc->mc_db->md_root = np->mp_pgno;
4169 mc->mc_db->md_depth++;
4170 *mc->mc_dbflag = DB_DIRTY;
4171 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4173 np->mp_flags |= P_LEAF2;
4174 mc->mc_flags |= C_INITIALIZED;
4180 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4181 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4182 DPRINTF("duplicate key [%s]", DKEY(key));
4184 return MDB_KEYEXIST;
4186 if (rc && rc != MDB_NOTFOUND)
4190 /* Cursor is positioned, now make sure all pages are writable */
4191 rc2 = mdb_cursor_touch(mc);
4196 /* The key already exists */
4197 if (rc == MDB_SUCCESS) {
4198 /* there's only a key anyway, so this is a no-op */
4199 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4200 unsigned int ksize = mc->mc_db->md_pad;
4201 if (key->mv_size != ksize)
4203 if (flags == MDB_CURRENT) {
4204 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4205 memcpy(ptr, key->mv_data, ksize);
4210 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4213 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4214 /* Was a single item before, must convert now */
4216 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4217 /* Just overwrite the current item */
4218 if (flags == MDB_CURRENT)
4221 dkey.mv_size = NODEDSZ(leaf);
4222 dkey.mv_data = NODEDATA(leaf);
4223 #if UINT_MAX < SIZE_MAX
4224 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4225 #ifdef MISALIGNED_OK
4226 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4228 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4231 /* if data matches, ignore it */
4232 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4233 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4235 /* create a fake page for the dup items */
4236 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4237 dkey.mv_data = dbuf;
4238 fp = (MDB_page *)&pbuf;
4239 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4240 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4241 fp->mp_lower = PAGEHDRSZ;
4242 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4243 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4244 fp->mp_flags |= P_LEAF2;
4245 fp->mp_pad = data->mv_size;
4247 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4248 (dkey.mv_size & 1) + (data->mv_size & 1);
4250 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4253 xdata.mv_size = fp->mp_upper;
4258 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4259 /* See if we need to convert from fake page to subDB */
4261 unsigned int offset;
4264 fp = NODEDATA(leaf);
4265 if (flags == MDB_CURRENT) {
4266 fp->mp_flags |= P_DIRTY;
4267 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4268 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4272 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4273 offset = fp->mp_pad;
4275 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4277 offset += offset & 1;
4278 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4279 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4281 /* yes, convert it */
4283 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4284 dummy.md_pad = fp->mp_pad;
4285 dummy.md_flags = MDB_DUPFIXED;
4286 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4287 dummy.md_flags |= MDB_INTEGERKEY;
4290 dummy.md_branch_pages = 0;
4291 dummy.md_leaf_pages = 1;
4292 dummy.md_overflow_pages = 0;
4293 dummy.md_entries = NUMKEYS(fp);
4295 xdata.mv_size = sizeof(MDB_db);
4296 xdata.mv_data = &dummy;
4297 mp = mdb_page_alloc(mc, 1);
4300 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4301 flags |= F_DUPDATA|F_SUBDATA;
4302 dummy.md_root = mp->mp_pgno;
4304 /* no, just grow it */
4306 xdata.mv_size = NODEDSZ(leaf) + offset;
4307 xdata.mv_data = &pbuf;
4308 mp = (MDB_page *)&pbuf;
4309 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4312 mp->mp_flags = fp->mp_flags | P_DIRTY;
4313 mp->mp_pad = fp->mp_pad;
4314 mp->mp_lower = fp->mp_lower;
4315 mp->mp_upper = fp->mp_upper + offset;
4317 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4319 nsize = NODEDSZ(leaf) - fp->mp_upper;
4320 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4321 for (i=0; i<NUMKEYS(fp); i++)
4322 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4324 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4328 /* data is on sub-DB, just store it */
4329 flags |= F_DUPDATA|F_SUBDATA;
4333 /* overflow page overwrites need special handling */
4334 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4337 int ovpages, dpages;
4339 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4340 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4341 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4342 mdb_page_get(mc->mc_txn, pg, &omp);
4343 /* Is the ov page writable and large enough? */
4344 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4345 /* yes, overwrite it. Note in this case we don't
4346 * bother to try shrinking the node if the new data
4347 * is smaller than the overflow threshold.
4349 if (F_ISSET(flags, MDB_RESERVE))
4350 data->mv_data = METADATA(omp);
4352 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4355 /* no, free ovpages */
4357 mc->mc_db->md_overflow_pages -= ovpages;
4358 for (i=0; i<ovpages; i++) {
4359 DPRINTF("freed ov page %zu", pg);
4360 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4364 } else if (NODEDSZ(leaf) == data->mv_size) {
4365 /* same size, just replace it. Note that we could
4366 * also reuse this node if the new data is smaller,
4367 * but instead we opt to shrink the node in that case.
4369 if (F_ISSET(flags, MDB_RESERVE))
4370 data->mv_data = NODEDATA(leaf);
4372 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4375 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4376 mc->mc_db->md_entries--;
4378 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4384 nflags = flags & NODE_ADD_FLAGS;
4385 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4386 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4387 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4388 nflags &= ~MDB_APPEND;
4389 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4391 /* There is room already in this leaf page. */
4392 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4393 if (rc == 0 && !do_sub) {
4394 /* Adjust other cursors pointing to mp */
4395 MDB_cursor *m2, *m3;
4396 MDB_dbi dbi = mc->mc_dbi;
4397 unsigned i = mc->mc_top;
4398 MDB_page *mp = mc->mc_pg[i];
4400 if (mc->mc_flags & C_SUB)
4403 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4404 if (mc->mc_flags & C_SUB)
4405 m3 = &m2->mc_xcursor->mx_cursor;
4408 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4409 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4416 if (rc != MDB_SUCCESS)
4417 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4419 /* Now store the actual data in the child DB. Note that we're
4420 * storing the user data in the keys field, so there are strict
4421 * size limits on dupdata. The actual data fields of the child
4422 * DB are all zero size.
4429 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4430 if (flags & MDB_CURRENT) {
4431 xflags = MDB_CURRENT;
4433 mdb_xcursor_init1(mc, leaf);
4434 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4436 /* converted, write the original data first */
4438 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4442 /* Adjust other cursors pointing to mp */
4444 unsigned i = mc->mc_top;
4445 MDB_page *mp = mc->mc_pg[i];
4447 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4448 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4449 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4450 mdb_xcursor_init1(m2, leaf);
4455 xflags |= (flags & MDB_APPEND);
4456 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4457 if (flags & F_SUBDATA) {
4458 void *db = NODEDATA(leaf);
4459 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4462 /* sub-writes might have failed so check rc again.
4463 * Don't increment count if we just replaced an existing item.
4465 if (!rc && !(flags & MDB_CURRENT))
4466 mc->mc_db->md_entries++;
4467 if (flags & MDB_MULTIPLE) {
4469 if (mcount < data[1].mv_size) {
4470 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4471 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4481 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4486 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4489 if (!mc->mc_flags & C_INITIALIZED)
4492 rc = mdb_cursor_touch(mc);
4496 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4498 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4499 if (flags != MDB_NODUPDATA) {
4500 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4501 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4503 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4504 /* If sub-DB still has entries, we're done */
4505 if (mc->mc_xcursor->mx_db.md_entries) {
4506 if (leaf->mn_flags & F_SUBDATA) {
4507 /* update subDB info */
4508 void *db = NODEDATA(leaf);
4509 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4511 /* shrink fake page */
4512 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4514 mc->mc_db->md_entries--;
4517 /* otherwise fall thru and delete the sub-DB */
4520 if (leaf->mn_flags & F_SUBDATA) {
4521 /* add all the child DB's pages to the free list */
4522 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4523 if (rc == MDB_SUCCESS) {
4524 mc->mc_db->md_entries -=
4525 mc->mc_xcursor->mx_db.md_entries;
4530 return mdb_cursor_del0(mc, leaf);
4533 /** Allocate and initialize new pages for a database.
4534 * @param[in] mc a cursor on the database being added to.
4535 * @param[in] flags flags defining what type of page is being allocated.
4536 * @param[in] num the number of pages to allocate. This is usually 1,
4537 * unless allocating overflow pages for a large record.
4538 * @return Address of a page, or NULL on failure.
4541 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4545 if ((np = mdb_page_alloc(mc, num)) == NULL)
4547 DPRINTF("allocated new mpage %zu, page size %u",
4548 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4549 np->mp_flags = flags | P_DIRTY;
4550 np->mp_lower = PAGEHDRSZ;
4551 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4554 mc->mc_db->md_branch_pages++;
4555 else if (IS_LEAF(np))
4556 mc->mc_db->md_leaf_pages++;
4557 else if (IS_OVERFLOW(np)) {
4558 mc->mc_db->md_overflow_pages += num;
4565 /** Calculate the size of a leaf node.
4566 * The size depends on the environment's page size; if a data item
4567 * is too large it will be put onto an overflow page and the node
4568 * size will only include the key and not the data. Sizes are always
4569 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4570 * of the #MDB_node headers.
4571 * @param[in] env The environment handle.
4572 * @param[in] key The key for the node.
4573 * @param[in] data The data for the node.
4574 * @return The number of bytes needed to store the node.
4577 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4581 sz = LEAFSIZE(key, data);
4582 if (sz >= env->me_psize / MDB_MINKEYS) {
4583 /* put on overflow page */
4584 sz -= data->mv_size - sizeof(pgno_t);
4588 return sz + sizeof(indx_t);
4591 /** Calculate the size of a branch node.
4592 * The size should depend on the environment's page size but since
4593 * we currently don't support spilling large keys onto overflow
4594 * pages, it's simply the size of the #MDB_node header plus the
4595 * size of the key. Sizes are always rounded up to an even number
4596 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4597 * @param[in] env The environment handle.
4598 * @param[in] key The key for the node.
4599 * @return The number of bytes needed to store the node.
4602 mdb_branch_size(MDB_env *env, MDB_val *key)
4607 if (sz >= env->me_psize / MDB_MINKEYS) {
4608 /* put on overflow page */
4609 /* not implemented */
4610 /* sz -= key->size - sizeof(pgno_t); */
4613 return sz + sizeof(indx_t);
4616 /** Add a node to the page pointed to by the cursor.
4617 * @param[in] mc The cursor for this operation.
4618 * @param[in] indx The index on the page where the new node should be added.
4619 * @param[in] key The key for the new node.
4620 * @param[in] data The data for the new node, if any.
4621 * @param[in] pgno The page number, if adding a branch node.
4622 * @param[in] flags Flags for the node.
4623 * @return 0 on success, non-zero on failure. Possible errors are:
4625 * <li>ENOMEM - failed to allocate overflow pages for the node.
4626 * <li>ENOSPC - there is insufficient room in the page. This error
4627 * should never happen since all callers already calculate the
4628 * page's free space before calling this function.
4632 mdb_node_add(MDB_cursor *mc, indx_t indx,
4633 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4636 size_t node_size = NODESIZE;
4639 MDB_page *mp = mc->mc_pg[mc->mc_top];
4640 MDB_page *ofp = NULL; /* overflow page */
4643 assert(mp->mp_upper >= mp->mp_lower);
4645 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4646 IS_LEAF(mp) ? "leaf" : "branch",
4647 IS_SUBP(mp) ? "sub-" : "",
4648 mp->mp_pgno, indx, data ? data->mv_size : 0,
4649 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4652 /* Move higher keys up one slot. */
4653 int ksize = mc->mc_db->md_pad, dif;
4654 char *ptr = LEAF2KEY(mp, indx, ksize);
4655 dif = NUMKEYS(mp) - indx;
4657 memmove(ptr+ksize, ptr, dif*ksize);
4658 /* insert new key */
4659 memcpy(ptr, key->mv_data, ksize);
4661 /* Just using these for counting */
4662 mp->mp_lower += sizeof(indx_t);
4663 mp->mp_upper -= ksize - sizeof(indx_t);
4668 node_size += key->mv_size;
4672 if (F_ISSET(flags, F_BIGDATA)) {
4673 /* Data already on overflow page. */
4674 node_size += sizeof(pgno_t);
4675 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4676 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4677 /* Put data on overflow page. */
4678 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4679 data->mv_size, node_size+data->mv_size);
4680 node_size += sizeof(pgno_t);
4681 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4683 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4686 node_size += data->mv_size;
4689 node_size += node_size & 1;
4691 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4692 DPRINTF("not enough room in page %zu, got %u ptrs",
4693 mp->mp_pgno, NUMKEYS(mp));
4694 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4695 mp->mp_upper - mp->mp_lower);
4696 DPRINTF("node size = %zu", node_size);
4700 /* Move higher pointers up one slot. */
4701 for (i = NUMKEYS(mp); i > indx; i--)
4702 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4704 /* Adjust free space offsets. */
4705 ofs = mp->mp_upper - node_size;
4706 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4707 mp->mp_ptrs[indx] = ofs;
4709 mp->mp_lower += sizeof(indx_t);
4711 /* Write the node data. */
4712 node = NODEPTR(mp, indx);
4713 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4714 node->mn_flags = flags;
4716 SETDSZ(node,data->mv_size);
4721 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4726 if (F_ISSET(flags, F_BIGDATA))
4727 memcpy(node->mn_data + key->mv_size, data->mv_data,
4729 else if (F_ISSET(flags, MDB_RESERVE))
4730 data->mv_data = node->mn_data + key->mv_size;
4732 memcpy(node->mn_data + key->mv_size, data->mv_data,
4735 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4737 if (F_ISSET(flags, MDB_RESERVE))
4738 data->mv_data = METADATA(ofp);
4740 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4747 /** Delete the specified node from a page.
4748 * @param[in] mp The page to operate on.
4749 * @param[in] indx The index of the node to delete.
4750 * @param[in] ksize The size of a node. Only used if the page is
4751 * part of a #MDB_DUPFIXED database.
4754 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4757 indx_t i, j, numkeys, ptr;
4764 COPY_PGNO(pgno, mp->mp_pgno);
4765 DPRINTF("delete node %u on %s page %zu", indx,
4766 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4769 assert(indx < NUMKEYS(mp));
4772 int x = NUMKEYS(mp) - 1 - indx;
4773 base = LEAF2KEY(mp, indx, ksize);
4775 memmove(base, base + ksize, x * ksize);
4776 mp->mp_lower -= sizeof(indx_t);
4777 mp->mp_upper += ksize - sizeof(indx_t);
4781 node = NODEPTR(mp, indx);
4782 sz = NODESIZE + node->mn_ksize;
4784 if (F_ISSET(node->mn_flags, F_BIGDATA))
4785 sz += sizeof(pgno_t);
4787 sz += NODEDSZ(node);
4791 ptr = mp->mp_ptrs[indx];
4792 numkeys = NUMKEYS(mp);
4793 for (i = j = 0; i < numkeys; i++) {
4795 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4796 if (mp->mp_ptrs[i] < ptr)
4797 mp->mp_ptrs[j] += sz;
4802 base = (char *)mp + mp->mp_upper;
4803 memmove(base + sz, base, ptr - mp->mp_upper);
4805 mp->mp_lower -= sizeof(indx_t);
4809 /** Compact the main page after deleting a node on a subpage.
4810 * @param[in] mp The main page to operate on.
4811 * @param[in] indx The index of the subpage on the main page.
4814 mdb_node_shrink(MDB_page *mp, indx_t indx)
4821 indx_t i, numkeys, ptr;
4823 node = NODEPTR(mp, indx);
4824 sp = (MDB_page *)NODEDATA(node);
4825 osize = NODEDSZ(node);
4827 delta = sp->mp_upper - sp->mp_lower;
4828 SETDSZ(node, osize - delta);
4829 xp = (MDB_page *)((char *)sp + delta);
4831 /* shift subpage upward */
4833 nsize = NUMKEYS(sp) * sp->mp_pad;
4834 memmove(METADATA(xp), METADATA(sp), nsize);
4837 nsize = osize - sp->mp_upper;
4838 numkeys = NUMKEYS(sp);
4839 for (i=numkeys-1; i>=0; i--)
4840 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4842 xp->mp_upper = sp->mp_lower;
4843 xp->mp_lower = sp->mp_lower;
4844 xp->mp_flags = sp->mp_flags;
4845 xp->mp_pad = sp->mp_pad;
4846 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4848 /* shift lower nodes upward */
4849 ptr = mp->mp_ptrs[indx];
4850 numkeys = NUMKEYS(mp);
4851 for (i = 0; i < numkeys; i++) {
4852 if (mp->mp_ptrs[i] <= ptr)
4853 mp->mp_ptrs[i] += delta;
4856 base = (char *)mp + mp->mp_upper;
4857 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4858 mp->mp_upper += delta;
4861 /** Initial setup of a sorted-dups cursor.
4862 * Sorted duplicates are implemented as a sub-database for the given key.
4863 * The duplicate data items are actually keys of the sub-database.
4864 * Operations on the duplicate data items are performed using a sub-cursor
4865 * initialized when the sub-database is first accessed. This function does
4866 * the preliminary setup of the sub-cursor, filling in the fields that
4867 * depend only on the parent DB.
4868 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4871 mdb_xcursor_init0(MDB_cursor *mc)
4873 MDB_xcursor *mx = mc->mc_xcursor;
4875 mx->mx_cursor.mc_xcursor = NULL;
4876 mx->mx_cursor.mc_txn = mc->mc_txn;
4877 mx->mx_cursor.mc_db = &mx->mx_db;
4878 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
4879 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
4880 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
4881 mx->mx_cursor.mc_snum = 0;
4882 mx->mx_cursor.mc_top = 0;
4883 mx->mx_cursor.mc_flags = C_SUB;
4884 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
4885 mx->mx_dbx.md_dcmp = NULL;
4886 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
4889 /** Final setup of a sorted-dups cursor.
4890 * Sets up the fields that depend on the data from the main cursor.
4891 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4892 * @param[in] node The data containing the #MDB_db record for the
4893 * sorted-dup database.
4896 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
4898 MDB_xcursor *mx = mc->mc_xcursor;
4900 if (node->mn_flags & F_SUBDATA) {
4901 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
4902 mx->mx_cursor.mc_snum = 0;
4903 mx->mx_cursor.mc_flags = C_SUB;
4905 MDB_page *fp = NODEDATA(node);
4906 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
4907 mx->mx_db.md_flags = 0;
4908 mx->mx_db.md_depth = 1;
4909 mx->mx_db.md_branch_pages = 0;
4910 mx->mx_db.md_leaf_pages = 1;
4911 mx->mx_db.md_overflow_pages = 0;
4912 mx->mx_db.md_entries = NUMKEYS(fp);
4913 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
4914 mx->mx_cursor.mc_snum = 1;
4915 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
4916 mx->mx_cursor.mc_top = 0;
4917 mx->mx_cursor.mc_pg[0] = fp;
4918 mx->mx_cursor.mc_ki[0] = 0;
4919 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4920 mx->mx_db.md_flags = MDB_DUPFIXED;
4921 mx->mx_db.md_pad = fp->mp_pad;
4922 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4923 mx->mx_db.md_flags |= MDB_INTEGERKEY;
4926 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
4928 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
4930 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
4931 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
4932 #if UINT_MAX < SIZE_MAX
4933 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
4934 #ifdef MISALIGNED_OK
4935 mx->mx_dbx.md_cmp = mdb_cmp_long;
4937 mx->mx_dbx.md_cmp = mdb_cmp_cint;
4942 /** Initialize a cursor for a given transaction and database. */
4944 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
4949 mc->mc_db = &txn->mt_dbs[dbi];
4950 mc->mc_dbx = &txn->mt_dbxs[dbi];
4951 mc->mc_dbflag = &txn->mt_dbflags[dbi];
4955 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4957 mc->mc_xcursor = mx;
4958 mdb_xcursor_init0(mc);
4960 mc->mc_xcursor = NULL;
4965 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
4968 MDB_xcursor *mx = NULL;
4969 size_t size = sizeof(MDB_cursor);
4971 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
4974 /* Allow read access to the freelist */
4975 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
4978 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
4979 size += sizeof(MDB_xcursor);
4981 if ((mc = malloc(size)) != NULL) {
4982 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4983 mx = (MDB_xcursor *)(mc + 1);
4985 mdb_cursor_init(mc, txn, dbi, mx);
4986 if (txn->mt_cursors) {
4987 mc->mc_next = txn->mt_cursors[dbi];
4988 txn->mt_cursors[dbi] = mc;
4990 mc->mc_flags |= C_ALLOCD;
5000 /* Return the count of duplicate data items for the current key */
5002 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5006 if (mc == NULL || countp == NULL)
5009 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5012 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5013 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5016 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5019 *countp = mc->mc_xcursor->mx_db.md_entries;
5025 mdb_cursor_close(MDB_cursor *mc)
5028 /* remove from txn, if tracked */
5029 if (mc->mc_txn->mt_cursors) {
5030 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5031 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5033 *prev = mc->mc_next;
5035 if (mc->mc_flags & C_ALLOCD)
5041 mdb_cursor_txn(MDB_cursor *mc)
5043 if (!mc) return NULL;
5048 mdb_cursor_dbi(MDB_cursor *mc)
5054 /** Replace the key for a node with a new key.
5055 * @param[in] mp The page containing the node to operate on.
5056 * @param[in] indx The index of the node to operate on.
5057 * @param[in] key The new key to use.
5058 * @return 0 on success, non-zero on failure.
5061 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5067 indx_t ptr, i, numkeys;
5070 node = NODEPTR(mp, indx);
5071 ptr = mp->mp_ptrs[indx];
5075 char kbuf2[(MAXKEYSIZE*2+1)];
5076 k2.mv_data = NODEKEY(node);
5077 k2.mv_size = node->mn_ksize;
5078 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5080 mdb_dkey(&k2, kbuf2),
5086 delta0 = delta = key->mv_size - node->mn_ksize;
5088 /* Must be 2-byte aligned. If new key is
5089 * shorter by 1, the shift will be skipped.
5091 delta += (delta & 1);
5093 if (delta > 0 && SIZELEFT(mp) < delta) {
5094 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5098 numkeys = NUMKEYS(mp);
5099 for (i = 0; i < numkeys; i++) {
5100 if (mp->mp_ptrs[i] <= ptr)
5101 mp->mp_ptrs[i] -= delta;
5104 base = (char *)mp + mp->mp_upper;
5105 len = ptr - mp->mp_upper + NODESIZE;
5106 memmove(base - delta, base, len);
5107 mp->mp_upper -= delta;
5109 node = NODEPTR(mp, indx);
5112 /* But even if no shift was needed, update ksize */
5114 node->mn_ksize = key->mv_size;
5117 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5122 /** Move a node from csrc to cdst.
5125 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5131 unsigned short flags;
5135 /* Mark src and dst as dirty. */
5136 if ((rc = mdb_page_touch(csrc)) ||
5137 (rc = mdb_page_touch(cdst)))
5140 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5141 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5142 key.mv_size = csrc->mc_db->md_pad;
5143 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5145 data.mv_data = NULL;
5149 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5150 assert(!((long)srcnode&1));
5151 srcpg = NODEPGNO(srcnode);
5152 flags = srcnode->mn_flags;
5153 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5154 unsigned int snum = csrc->mc_snum;
5156 /* must find the lowest key below src */
5157 mdb_page_search_root(csrc, NULL, 0);
5158 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5159 key.mv_size = NODEKSZ(s2);
5160 key.mv_data = NODEKEY(s2);
5161 csrc->mc_snum = snum--;
5162 csrc->mc_top = snum;
5164 key.mv_size = NODEKSZ(srcnode);
5165 key.mv_data = NODEKEY(srcnode);
5167 data.mv_size = NODEDSZ(srcnode);
5168 data.mv_data = NODEDATA(srcnode);
5170 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5171 unsigned int snum = cdst->mc_snum;
5174 /* must find the lowest key below dst */
5175 mdb_page_search_root(cdst, NULL, 0);
5176 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5177 bkey.mv_size = NODEKSZ(s2);
5178 bkey.mv_data = NODEKEY(s2);
5179 cdst->mc_snum = snum--;
5180 cdst->mc_top = snum;
5181 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5184 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5185 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5186 csrc->mc_ki[csrc->mc_top],
5188 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5189 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5191 /* Add the node to the destination page.
5193 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5194 if (rc != MDB_SUCCESS)
5197 /* Delete the node from the source page.
5199 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5202 /* Adjust other cursors pointing to mp */
5203 MDB_cursor *m2, *m3;
5204 MDB_dbi dbi = csrc->mc_dbi;
5205 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5207 if (csrc->mc_flags & C_SUB)
5210 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5211 if (m2 == csrc) continue;
5212 if (csrc->mc_flags & C_SUB)
5213 m3 = &m2->mc_xcursor->mx_cursor;
5216 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5217 csrc->mc_ki[csrc->mc_top]) {
5218 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5219 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5224 /* Update the parent separators.
5226 if (csrc->mc_ki[csrc->mc_top] == 0) {
5227 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5228 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5229 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5231 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5232 key.mv_size = NODEKSZ(srcnode);
5233 key.mv_data = NODEKEY(srcnode);
5235 DPRINTF("update separator for source page %zu to [%s]",
5236 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5237 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5238 &key)) != MDB_SUCCESS)
5241 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5243 nullkey.mv_size = 0;
5244 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5245 assert(rc == MDB_SUCCESS);
5249 if (cdst->mc_ki[cdst->mc_top] == 0) {
5250 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5251 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5252 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5254 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5255 key.mv_size = NODEKSZ(srcnode);
5256 key.mv_data = NODEKEY(srcnode);
5258 DPRINTF("update separator for destination page %zu to [%s]",
5259 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5260 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5261 &key)) != MDB_SUCCESS)
5264 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5266 nullkey.mv_size = 0;
5267 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5268 assert(rc == MDB_SUCCESS);
5275 /** Merge one page into another.
5276 * The nodes from the page pointed to by \b csrc will
5277 * be copied to the page pointed to by \b cdst and then
5278 * the \b csrc page will be freed.
5279 * @param[in] csrc Cursor pointing to the source page.
5280 * @param[in] cdst Cursor pointing to the destination page.
5283 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5291 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5292 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5294 assert(csrc->mc_snum > 1); /* can't merge root page */
5295 assert(cdst->mc_snum > 1);
5297 /* Mark dst as dirty. */
5298 if ((rc = mdb_page_touch(cdst)))
5301 /* Move all nodes from src to dst.
5303 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5304 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5305 key.mv_size = csrc->mc_db->md_pad;
5306 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5307 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5308 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5309 if (rc != MDB_SUCCESS)
5311 key.mv_data = (char *)key.mv_data + key.mv_size;
5314 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5315 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5316 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5317 unsigned int snum = csrc->mc_snum;
5319 /* must find the lowest key below src */
5320 mdb_page_search_root(csrc, NULL, 0);
5321 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5322 key.mv_size = NODEKSZ(s2);
5323 key.mv_data = NODEKEY(s2);
5324 csrc->mc_snum = snum--;
5325 csrc->mc_top = snum;
5327 key.mv_size = srcnode->mn_ksize;
5328 key.mv_data = NODEKEY(srcnode);
5331 data.mv_size = NODEDSZ(srcnode);
5332 data.mv_data = NODEDATA(srcnode);
5333 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5334 if (rc != MDB_SUCCESS)
5339 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5340 cdst->mc_pg[cdst->mc_top]->mp_pgno, NUMKEYS(cdst->mc_pg[cdst->mc_top]), (float)PAGEFILL(cdst->mc_txn->mt_env, cdst->mc_pg[cdst->mc_top]) / 10);
5342 /* Unlink the src page from parent and add to free list.
5344 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5345 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5347 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5351 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5352 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5353 csrc->mc_db->md_leaf_pages--;
5355 csrc->mc_db->md_branch_pages--;
5357 /* Adjust other cursors pointing to mp */
5358 MDB_cursor *m2, *m3;
5359 MDB_dbi dbi = csrc->mc_dbi;
5360 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5362 if (csrc->mc_flags & C_SUB)
5365 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5366 if (csrc->mc_flags & C_SUB)
5367 m3 = &m2->mc_xcursor->mx_cursor;
5370 if (m3 == csrc) continue;
5371 if (m3->mc_snum < csrc->mc_snum) continue;
5372 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5373 m3->mc_pg[csrc->mc_top] = mp;
5374 m3->mc_ki[csrc->mc_top] += nkeys;
5378 mdb_cursor_pop(csrc);
5380 return mdb_rebalance(csrc);
5383 /** Copy the contents of a cursor.
5384 * @param[in] csrc The cursor to copy from.
5385 * @param[out] cdst The cursor to copy to.
5388 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5392 cdst->mc_txn = csrc->mc_txn;
5393 cdst->mc_dbi = csrc->mc_dbi;
5394 cdst->mc_db = csrc->mc_db;
5395 cdst->mc_dbx = csrc->mc_dbx;
5396 cdst->mc_snum = csrc->mc_snum;
5397 cdst->mc_top = csrc->mc_top;
5398 cdst->mc_flags = csrc->mc_flags;
5400 for (i=0; i<csrc->mc_snum; i++) {
5401 cdst->mc_pg[i] = csrc->mc_pg[i];
5402 cdst->mc_ki[i] = csrc->mc_ki[i];
5406 /** Rebalance the tree after a delete operation.
5407 * @param[in] mc Cursor pointing to the page where rebalancing
5409 * @return 0 on success, non-zero on failure.
5412 mdb_rebalance(MDB_cursor *mc)
5422 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5423 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5424 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5425 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5429 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5432 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5433 DPRINTF("no need to rebalance page %zu, above fill threshold",
5439 if (mc->mc_snum < 2) {
5440 MDB_page *mp = mc->mc_pg[0];
5441 if (NUMKEYS(mp) == 0) {
5442 DPUTS("tree is completely empty");
5443 mc->mc_db->md_root = P_INVALID;
5444 mc->mc_db->md_depth = 0;
5445 mc->mc_db->md_leaf_pages = 0;
5446 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5450 /* Adjust other cursors pointing to mp */
5451 MDB_cursor *m2, *m3;
5452 MDB_dbi dbi = mc->mc_dbi;
5454 if (mc->mc_flags & C_SUB)
5457 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5458 if (m2 == mc) continue;
5459 if (mc->mc_flags & C_SUB)
5460 m3 = &m2->mc_xcursor->mx_cursor;
5463 if (m3->mc_snum < mc->mc_snum) continue;
5464 if (m3->mc_pg[0] == mp) {
5470 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5471 DPUTS("collapsing root page!");
5472 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5473 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5474 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5477 mc->mc_db->md_depth--;
5478 mc->mc_db->md_branch_pages--;
5480 /* Adjust other cursors pointing to mp */
5481 MDB_cursor *m2, *m3;
5482 MDB_dbi dbi = mc->mc_dbi;
5484 if (mc->mc_flags & C_SUB)
5487 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5488 if (m2 == mc) continue;
5489 if (mc->mc_flags & C_SUB)
5490 m3 = &m2->mc_xcursor->mx_cursor;
5493 if (m3->mc_snum < mc->mc_snum) continue;
5494 if (m3->mc_pg[0] == mp) {
5495 m3->mc_pg[0] = mc->mc_pg[0];
5500 DPUTS("root page doesn't need rebalancing");
5504 /* The parent (branch page) must have at least 2 pointers,
5505 * otherwise the tree is invalid.
5507 ptop = mc->mc_top-1;
5508 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5510 /* Leaf page fill factor is below the threshold.
5511 * Try to move keys from left or right neighbor, or
5512 * merge with a neighbor page.
5517 mdb_cursor_copy(mc, &mn);
5518 mn.mc_xcursor = NULL;
5520 if (mc->mc_ki[ptop] == 0) {
5521 /* We're the leftmost leaf in our parent.
5523 DPUTS("reading right neighbor");
5525 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5526 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5528 mn.mc_ki[mn.mc_top] = 0;
5529 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5531 /* There is at least one neighbor to the left.
5533 DPUTS("reading left neighbor");
5535 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5536 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5538 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5539 mc->mc_ki[mc->mc_top] = 0;
5542 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5543 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10);
5545 /* If the neighbor page is above threshold and has at least two
5546 * keys, move one key from it.
5548 * Otherwise we should try to merge them.
5550 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5551 return mdb_node_move(&mn, mc);
5552 else { /* FIXME: if (has_enough_room()) */
5553 mc->mc_flags &= ~C_INITIALIZED;
5554 if (mc->mc_ki[ptop] == 0)
5555 return mdb_page_merge(&mn, mc);
5557 return mdb_page_merge(mc, &mn);
5561 /** Complete a delete operation started by #mdb_cursor_del(). */
5563 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5567 /* add overflow pages to free list */
5568 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5572 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5573 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5574 mc->mc_db->md_overflow_pages -= ovpages;
5575 for (i=0; i<ovpages; i++) {
5576 DPRINTF("freed ov page %zu", pg);
5577 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5581 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5582 mc->mc_db->md_entries--;
5583 rc = mdb_rebalance(mc);
5584 if (rc != MDB_SUCCESS)
5585 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5591 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5592 MDB_val *key, MDB_val *data)
5597 MDB_val rdata, *xdata;
5601 assert(key != NULL);
5603 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5605 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5608 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5612 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5616 mdb_cursor_init(&mc, txn, dbi, &mx);
5627 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5629 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5633 /** Split a page and insert a new node.
5634 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5635 * The cursor will be updated to point to the actual page and index where
5636 * the node got inserted after the split.
5637 * @param[in] newkey The key for the newly inserted node.
5638 * @param[in] newdata The data for the newly inserted node.
5639 * @param[in] newpgno The page number, if the new node is a branch node.
5640 * @return 0 on success, non-zero on failure.
5643 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5644 unsigned int nflags)
5647 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1;
5650 unsigned int i, j, split_indx, nkeys, pmax;
5652 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5654 MDB_page *mp, *rp, *pp;
5659 mp = mc->mc_pg[mc->mc_top];
5660 newindx = mc->mc_ki[mc->mc_top];
5662 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5663 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5664 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5666 /* Create a right sibling. */
5667 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5669 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5671 if (mc->mc_snum < 2) {
5672 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5674 /* shift current top to make room for new parent */
5675 mc->mc_pg[1] = mc->mc_pg[0];
5676 mc->mc_ki[1] = mc->mc_ki[0];
5679 mc->mc_db->md_root = pp->mp_pgno;
5680 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5681 mc->mc_db->md_depth++;
5684 /* Add left (implicit) pointer. */
5685 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5686 /* undo the pre-push */
5687 mc->mc_pg[0] = mc->mc_pg[1];
5688 mc->mc_ki[0] = mc->mc_ki[1];
5689 mc->mc_db->md_root = mp->mp_pgno;
5690 mc->mc_db->md_depth--;
5697 ptop = mc->mc_top-1;
5698 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5701 mdb_cursor_copy(mc, &mn);
5702 mn.mc_pg[mn.mc_top] = rp;
5703 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5705 if (nflags & MDB_APPEND) {
5706 mn.mc_ki[mn.mc_top] = 0;
5713 nkeys = NUMKEYS(mp);
5714 split_indx = nkeys / 2 + 1;
5719 unsigned int lsize, rsize, ksize;
5720 /* Move half of the keys to the right sibling */
5722 x = mc->mc_ki[mc->mc_top] - split_indx;
5723 ksize = mc->mc_db->md_pad;
5724 split = LEAF2KEY(mp, split_indx, ksize);
5725 rsize = (nkeys - split_indx) * ksize;
5726 lsize = (nkeys - split_indx) * sizeof(indx_t);
5727 mp->mp_lower -= lsize;
5728 rp->mp_lower += lsize;
5729 mp->mp_upper += rsize - lsize;
5730 rp->mp_upper -= rsize - lsize;
5731 sepkey.mv_size = ksize;
5732 if (newindx == split_indx) {
5733 sepkey.mv_data = newkey->mv_data;
5735 sepkey.mv_data = split;
5738 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5739 memcpy(rp->mp_ptrs, split, rsize);
5740 sepkey.mv_data = rp->mp_ptrs;
5741 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5742 memcpy(ins, newkey->mv_data, ksize);
5743 mp->mp_lower += sizeof(indx_t);
5744 mp->mp_upper -= ksize - sizeof(indx_t);
5747 memcpy(rp->mp_ptrs, split, x * ksize);
5748 ins = LEAF2KEY(rp, x, ksize);
5749 memcpy(ins, newkey->mv_data, ksize);
5750 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5751 rp->mp_lower += sizeof(indx_t);
5752 rp->mp_upper -= ksize - sizeof(indx_t);
5753 mc->mc_ki[mc->mc_top] = x;
5754 mc->mc_pg[mc->mc_top] = rp;
5759 /* For leaf pages, check the split point based on what
5760 * fits where, since otherwise mdb_node_add can fail.
5762 * This check is only needed when the data items are
5763 * relatively large, such that being off by one will
5764 * make the difference between success or failure.
5765 * When the size of the data items is much smaller than
5766 * one-half of a page, this check is irrelevant.
5768 if (IS_LEAF(mp) && nkeys < 4) {
5769 unsigned int psize, nsize;
5770 /* Maximum free space in an empty page */
5771 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5772 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5773 if (newindx <= split_indx) {
5776 for (i=0; i<split_indx; i++) {
5777 node = NODEPTR(mp, i);
5778 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5779 if (F_ISSET(node->mn_flags, F_BIGDATA))
5780 psize += sizeof(pgno_t);
5782 psize += NODEDSZ(node);
5785 if (i == split_indx - 1 && newindx == split_indx)
5794 for (i=nkeys-1; i>=split_indx; i--) {
5795 node = NODEPTR(mp, i);
5796 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5797 if (F_ISSET(node->mn_flags, F_BIGDATA))
5798 psize += sizeof(pgno_t);
5800 psize += NODEDSZ(node);
5810 /* First find the separating key between the split pages.
5811 * The case where newindx == split_indx is ambiguous; the
5812 * new item could go to the new page or stay on the original
5813 * page. If newpos == 1 it goes to the new page.
5815 if (newindx == split_indx && newpos) {
5816 sepkey.mv_size = newkey->mv_size;
5817 sepkey.mv_data = newkey->mv_data;
5819 node = NODEPTR(mp, split_indx);
5820 sepkey.mv_size = node->mn_ksize;
5821 sepkey.mv_data = NODEKEY(node);
5825 DPRINTF("separator is [%s]", DKEY(&sepkey));
5827 /* Copy separator key to the parent.
5829 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5832 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5834 /* Right page might now have changed parent.
5835 * Check if left page also changed parent.
5837 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
5838 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
5839 mc->mc_pg[ptop] = mn.mc_pg[ptop];
5840 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
5844 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
5847 if (rc != MDB_SUCCESS) {
5850 if (nflags & MDB_APPEND) {
5851 mc->mc_pg[mc->mc_top] = rp;
5852 mc->mc_ki[mc->mc_top] = 0;
5853 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
5862 /* Move half of the keys to the right sibling. */
5864 /* grab a page to hold a temporary copy */
5865 copy = mdb_page_malloc(mc);
5869 copy->mp_pgno = mp->mp_pgno;
5870 copy->mp_flags = mp->mp_flags;
5871 copy->mp_lower = PAGEHDRSZ;
5872 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
5873 mc->mc_pg[mc->mc_top] = copy;
5874 for (i = j = 0; i <= nkeys; j++) {
5875 if (i == split_indx) {
5876 /* Insert in right sibling. */
5877 /* Reset insert index for right sibling. */
5878 if (i != newindx || (newpos ^ ins_new)) {
5880 mc->mc_pg[mc->mc_top] = rp;
5884 if (i == newindx && !ins_new) {
5885 /* Insert the original entry that caused the split. */
5886 rkey.mv_data = newkey->mv_data;
5887 rkey.mv_size = newkey->mv_size;
5896 /* Update index for the new key. */
5897 mc->mc_ki[mc->mc_top] = j;
5898 } else if (i == nkeys) {
5901 node = NODEPTR(mp, i);
5902 rkey.mv_data = NODEKEY(node);
5903 rkey.mv_size = node->mn_ksize;
5905 xdata.mv_data = NODEDATA(node);
5906 xdata.mv_size = NODEDSZ(node);
5909 pgno = NODEPGNO(node);
5910 flags = node->mn_flags;
5915 if (!IS_LEAF(mp) && j == 0) {
5916 /* First branch index doesn't need key data. */
5920 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
5923 nkeys = NUMKEYS(copy);
5924 for (i=0; i<nkeys; i++)
5925 mp->mp_ptrs[i] = copy->mp_ptrs[i];
5926 mp->mp_lower = copy->mp_lower;
5927 mp->mp_upper = copy->mp_upper;
5928 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
5929 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
5931 /* reset back to original page */
5932 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
5933 mc->mc_pg[mc->mc_top] = mp;
5934 if (nflags & MDB_RESERVE) {
5935 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5936 if (!(node->mn_flags & F_BIGDATA))
5937 newdata->mv_data = NODEDATA(node);
5941 /* return tmp page to freelist */
5942 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
5943 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
5944 mc->mc_txn->mt_env->me_dpages = copy;
5947 /* Adjust other cursors pointing to mp */
5948 MDB_cursor *m2, *m3;
5949 MDB_dbi dbi = mc->mc_dbi;
5951 if (mc->mc_flags & C_SUB)
5954 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5955 if (m2 == mc) continue;
5956 if (mc->mc_flags & C_SUB)
5957 m3 = &m2->mc_xcursor->mx_cursor;
5960 if (!(m3->mc_flags & C_INITIALIZED))
5965 for (k=m3->mc_top; k>=0; k--) {
5966 m3->mc_ki[k+1] = m3->mc_ki[k];
5967 m3->mc_pg[k+1] = m3->mc_pg[k];
5969 m3->mc_ki[0] = mc->mc_ki[0];
5970 m3->mc_pg[0] = mc->mc_pg[0];
5974 if (m3->mc_pg[mc->mc_top] == mp) {
5975 if (m3->mc_ki[m3->mc_top] >= split_indx) {
5976 m3->mc_pg[m3->mc_top] = rp;
5977 m3->mc_ki[m3->mc_top] -= split_indx;
5986 mdb_put(MDB_txn *txn, MDB_dbi dbi,
5987 MDB_val *key, MDB_val *data, unsigned int flags)
5992 assert(key != NULL);
5993 assert(data != NULL);
5995 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5998 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6002 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6006 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6009 mdb_cursor_init(&mc, txn, dbi, &mx);
6010 return mdb_cursor_put(&mc, key, data, flags);
6013 /** Only a subset of the @ref mdb_env flags can be changed
6014 * at runtime. Changing other flags requires closing the environment
6015 * and re-opening it with the new flags.
6017 #define CHANGEABLE (MDB_NOSYNC)
6019 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6021 if ((flag & CHANGEABLE) != flag)
6024 env->me_flags |= flag;
6026 env->me_flags &= ~flag;
6031 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6036 *arg = env->me_flags;
6041 mdb_env_get_path(MDB_env *env, const char **arg)
6046 *arg = env->me_path;
6050 /** Common code for #mdb_stat() and #mdb_env_stat().
6051 * @param[in] env the environment to operate in.
6052 * @param[in] db the #MDB_db record containing the stats to return.
6053 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6054 * @return 0, this function always succeeds.
6057 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6059 arg->ms_psize = env->me_psize;
6060 arg->ms_depth = db->md_depth;
6061 arg->ms_branch_pages = db->md_branch_pages;
6062 arg->ms_leaf_pages = db->md_leaf_pages;
6063 arg->ms_overflow_pages = db->md_overflow_pages;
6064 arg->ms_entries = db->md_entries;
6069 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6073 if (env == NULL || arg == NULL)
6076 mdb_env_read_meta(env, &toggle);
6078 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6081 /** Set the default comparison functions for a database.
6082 * Called immediately after a database is opened to set the defaults.
6083 * The user can then override them with #mdb_set_compare() or
6084 * #mdb_set_dupsort().
6085 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6086 * @param[in] dbi A database handle returned by #mdb_open()
6089 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6091 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
6092 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memnr;
6093 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
6094 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_cint;
6096 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memn;
6098 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6099 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
6100 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
6101 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_int;
6103 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_cint;
6104 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
6105 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memnr;
6107 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memn;
6110 txn->mt_dbxs[dbi].md_dcmp = NULL;
6114 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6119 int rc, dbflag, exact;
6122 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6123 mdb_default_cmp(txn, FREE_DBI);
6129 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6130 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6131 mdb_default_cmp(txn, MAIN_DBI);
6135 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6136 mdb_default_cmp(txn, MAIN_DBI);
6139 /* Is the DB already open? */
6141 for (i=2; i<txn->mt_numdbs; i++) {
6142 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6143 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6149 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
6152 /* Find the DB info */
6156 key.mv_data = (void *)name;
6157 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6158 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6159 if (rc == MDB_SUCCESS) {
6160 /* make sure this is actually a DB */
6161 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6162 if (!(node->mn_flags & F_SUBDATA))
6164 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6165 /* Create if requested */
6167 data.mv_size = sizeof(MDB_db);
6168 data.mv_data = &dummy;
6169 memset(&dummy, 0, sizeof(dummy));
6170 dummy.md_root = P_INVALID;
6171 dummy.md_flags = flags & 0xffff;
6172 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6176 /* OK, got info, add to table */
6177 if (rc == MDB_SUCCESS) {
6178 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
6179 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
6180 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
6181 txn->mt_dbflags[txn->mt_numdbs] = dbflag;
6182 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
6183 *dbi = txn->mt_numdbs;
6184 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
6185 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
6186 mdb_default_cmp(txn, txn->mt_numdbs);
6193 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6195 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6198 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6201 void mdb_close(MDB_env *env, MDB_dbi dbi)
6204 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6206 ptr = env->me_dbxs[dbi].md_name.mv_data;
6207 env->me_dbxs[dbi].md_name.mv_data = NULL;
6208 env->me_dbxs[dbi].md_name.mv_size = 0;
6212 /** Add all the DB's pages to the free list.
6213 * @param[in] mc Cursor on the DB to free.
6214 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6215 * @return 0 on success, non-zero on failure.
6218 mdb_drop0(MDB_cursor *mc, int subs)
6222 rc = mdb_page_search(mc, NULL, 0);
6223 if (rc == MDB_SUCCESS) {
6228 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6229 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6232 mdb_cursor_copy(mc, &mx);
6233 while (mc->mc_snum > 0) {
6234 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6235 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6236 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6237 if (ni->mn_flags & F_SUBDATA) {
6238 mdb_xcursor_init1(mc, ni);
6239 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6245 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6247 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6250 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6255 rc = mdb_cursor_sibling(mc, 1);
6257 /* no more siblings, go back to beginning
6258 * of previous level. (stack was already popped
6259 * by mdb_cursor_sibling)
6261 for (i=1; i<mc->mc_top; i++)
6262 mc->mc_pg[i] = mx.mc_pg[i];
6266 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6267 mc->mc_db->md_root);
6272 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6277 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6280 rc = mdb_cursor_open(txn, dbi, &mc);
6284 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6288 /* Can't delete the main DB */
6289 if (del && dbi > MAIN_DBI) {
6290 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6292 mdb_close(txn->mt_env, dbi);
6294 txn->mt_dbflags[dbi] |= DB_DIRTY;
6295 txn->mt_dbs[dbi].md_depth = 0;
6296 txn->mt_dbs[dbi].md_branch_pages = 0;
6297 txn->mt_dbs[dbi].md_leaf_pages = 0;
6298 txn->mt_dbs[dbi].md_overflow_pages = 0;
6299 txn->mt_dbs[dbi].md_entries = 0;
6300 txn->mt_dbs[dbi].md_root = P_INVALID;
6303 mdb_cursor_close(mc);
6307 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6309 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6312 txn->mt_dbxs[dbi].md_cmp = cmp;
6316 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6318 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6321 txn->mt_dbxs[dbi].md_dcmp = cmp;
6325 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6327 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6330 txn->mt_dbxs[dbi].md_rel = rel;
6334 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6336 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6339 txn->mt_dbxs[dbi].md_relctx = ctx;