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);
1228 last = txn->mt_env->me_pglast + 1;
1230 key.mv_data = &last;
1231 key.mv_size = sizeof(last);
1232 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1235 last = *(txnid_t *)key.mv_data;
1240 oldest = txn->mt_txnid - 1;
1241 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1242 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1243 if (mr && mr < oldest)
1248 if (oldest > last) {
1249 /* It's usable, grab it.
1254 if (!txn->mt_env->me_pgfirst) {
1255 mdb_node_read(txn, leaf, &data);
1257 idl = (ID *) data.mv_data;
1258 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1259 mop->mo_next = txn->mt_env->me_pghead;
1260 mop->mo_txnid = last;
1261 txn->mt_env->me_pglast = last;
1262 if (!txn->mt_env->me_pgfirst)
1263 txn->mt_env->me_pgfirst = last;
1264 txn->mt_env->me_pghead = mop;
1265 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1270 DPRINTF("IDL read txn %zu root %zu num %zu",
1271 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1272 for (i=0; i<idl[0]; i++) {
1273 DPRINTF("IDL %zu", idl[i+1]);
1280 if (txn->mt_env->me_pghead) {
1281 MDB_oldpages *mop = txn->mt_env->me_pghead;
1283 /* FIXME: For now, always use fresh pages. We
1284 * really ought to search the free list for a
1289 /* peel pages off tail, so we only have to truncate the list */
1290 pgno = MDB_IDL_LAST(mop->mo_pages);
1291 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1293 if (mop->mo_pages[2] > mop->mo_pages[1])
1294 mop->mo_pages[0] = 0;
1298 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1299 txn->mt_env->me_pghead = mop->mo_next;
1306 if (pgno == P_INVALID) {
1307 /* DB size is maxed out */
1308 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1309 DPUTS("DB size maxed out");
1313 if (txn->mt_env->me_dpages && num == 1) {
1314 np = txn->mt_env->me_dpages;
1315 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1316 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1317 txn->mt_env->me_dpages = np->mp_next;
1319 size_t sz = txn->mt_env->me_psize * num;
1320 if ((np = malloc(sz)) == NULL)
1322 VGMEMP_ALLOC(txn->mt_env, np, sz);
1324 if (pgno == P_INVALID) {
1325 np->mp_pgno = txn->mt_next_pgno;
1326 txn->mt_next_pgno += num;
1330 mid.mid = np->mp_pgno;
1332 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1337 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1338 * @param[in] mc cursor pointing to the page to be touched
1339 * @return 0 on success, non-zero on failure.
1342 mdb_page_touch(MDB_cursor *mc)
1344 MDB_page *mp = mc->mc_pg[mc->mc_top];
1347 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1349 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1351 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1352 assert(mp->mp_pgno != np->mp_pgno);
1353 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1355 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1358 mp->mp_flags |= P_DIRTY;
1361 /* Adjust other cursors pointing to mp */
1362 if (mc->mc_flags & C_SUB) {
1363 MDB_cursor *m2, *m3;
1364 MDB_dbi dbi = mc->mc_dbi-1;
1366 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1367 if (m2 == mc) continue;
1368 m3 = &m2->mc_xcursor->mx_cursor;
1369 if (m3->mc_snum < mc->mc_snum) continue;
1370 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1371 m3->mc_pg[mc->mc_top] = mp;
1377 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1378 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1379 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1380 m2->mc_pg[mc->mc_top] = mp;
1384 mc->mc_pg[mc->mc_top] = mp;
1385 /** If this page has a parent, update the parent to point to
1389 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1391 mc->mc_db->md_root = mp->mp_pgno;
1392 } else if (mc->mc_txn->mt_parent) {
1395 /* If txn has a parent, make sure the page is in our
1398 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1399 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1400 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1401 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1402 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1403 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1404 mc->mc_pg[mc->mc_top] = mp;
1410 np = mdb_page_malloc(mc);
1411 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1412 mid.mid = np->mp_pgno;
1414 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1422 mdb_env_sync(MDB_env *env, int force)
1425 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1426 if (MDB_FDATASYNC(env->me_fd))
1432 /** Make shadow copies of all of parent txn's cursors */
1434 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1436 MDB_cursor *mc, *m2;
1437 unsigned int i, j, size;
1439 for (i=0;i<src->mt_numdbs; i++) {
1440 if (src->mt_cursors[i]) {
1441 size = sizeof(MDB_cursor);
1442 if (src->mt_cursors[i]->mc_xcursor)
1443 size += sizeof(MDB_xcursor);
1444 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1451 mc->mc_db = &dst->mt_dbs[i];
1452 mc->mc_dbx = m2->mc_dbx;
1453 mc->mc_dbflag = &dst->mt_dbflags[i];
1454 mc->mc_snum = m2->mc_snum;
1455 mc->mc_top = m2->mc_top;
1456 mc->mc_flags = m2->mc_flags | C_SHADOW;
1457 for (j=0; j<mc->mc_snum; j++) {
1458 mc->mc_pg[j] = m2->mc_pg[j];
1459 mc->mc_ki[j] = m2->mc_ki[j];
1461 if (m2->mc_xcursor) {
1462 MDB_xcursor *mx, *mx2;
1463 mx = (MDB_xcursor *)(mc+1);
1464 mc->mc_xcursor = mx;
1465 mx2 = m2->mc_xcursor;
1466 mx->mx_db = mx2->mx_db;
1467 mx->mx_dbx = mx2->mx_dbx;
1468 mx->mx_dbflag = mx2->mx_dbflag;
1469 mx->mx_cursor.mc_txn = dst;
1470 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1471 mx->mx_cursor.mc_db = &mx->mx_db;
1472 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1473 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1474 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1475 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1476 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1477 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1478 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1479 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1482 mc->mc_xcursor = NULL;
1484 mc->mc_next = dst->mt_cursors[i];
1485 dst->mt_cursors[i] = mc;
1492 /** Merge shadow cursors back into parent's */
1494 mdb_cursor_merge(MDB_txn *txn)
1497 for (i=0; i<txn->mt_numdbs; i++) {
1498 if (txn->mt_cursors[i]) {
1500 while ((mc = txn->mt_cursors[i])) {
1501 txn->mt_cursors[i] = mc->mc_next;
1502 if (mc->mc_flags & C_SHADOW) {
1503 MDB_cursor *m2 = mc->mc_orig;
1505 m2->mc_snum = mc->mc_snum;
1506 m2->mc_top = mc->mc_top;
1507 for (j=0; j<mc->mc_snum; j++) {
1508 m2->mc_pg[j] = mc->mc_pg[j];
1509 m2->mc_ki[j] = mc->mc_ki[j];
1512 if (mc->mc_flags & C_ALLOCD)
1520 mdb_txn_reset0(MDB_txn *txn);
1522 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1523 * @param[in] txn the transaction handle to initialize
1524 * @return 0 on success, non-zero on failure. This can only
1525 * fail for read-only transactions, and then only if the
1526 * reader table is full.
1529 mdb_txn_renew0(MDB_txn *txn)
1531 MDB_env *env = txn->mt_env;
1534 if (txn->mt_flags & MDB_TXN_RDONLY) {
1535 MDB_reader *r = pthread_getspecific(env->me_txkey);
1538 pid_t pid = getpid();
1539 pthread_t tid = pthread_self();
1542 for (i=0; i<env->me_txns->mti_numreaders; i++)
1543 if (env->me_txns->mti_readers[i].mr_pid == 0)
1545 if (i == env->me_maxreaders) {
1546 UNLOCK_MUTEX_R(env);
1549 env->me_txns->mti_readers[i].mr_pid = pid;
1550 env->me_txns->mti_readers[i].mr_tid = tid;
1551 if (i >= env->me_txns->mti_numreaders)
1552 env->me_txns->mti_numreaders = i+1;
1553 UNLOCK_MUTEX_R(env);
1554 r = &env->me_txns->mti_readers[i];
1555 pthread_setspecific(env->me_txkey, r);
1557 txn->mt_toggle = env->me_txns->mti_me_toggle;
1558 txn->mt_txnid = env->me_txns->mti_txnid;
1559 /* This happens if a different process was the
1560 * last writer to the DB.
1562 if (env->me_wtxnid < txn->mt_txnid)
1563 mt_dbflag = DB_STALE;
1564 r->mr_txnid = txn->mt_txnid;
1565 txn->mt_u.reader = r;
1569 txn->mt_txnid = env->me_txns->mti_txnid;
1570 if (env->me_wtxnid < txn->mt_txnid)
1571 mt_dbflag = DB_STALE;
1573 txn->mt_toggle = env->me_txns->mti_me_toggle;
1574 txn->mt_u.dirty_list = env->me_dirty_list;
1575 txn->mt_u.dirty_list[0].mid = 0;
1576 txn->mt_free_pgs = env->me_free_pgs;
1577 txn->mt_free_pgs[0] = 0;
1578 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1582 /* Copy the DB arrays */
1583 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1584 txn->mt_numdbs = env->me_numdbs;
1585 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1586 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1587 if (txn->mt_numdbs > 2)
1588 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1589 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1590 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1592 memset(txn->mt_dbflags, mt_dbflag, env->me_numdbs);
1598 mdb_txn_renew(MDB_txn *txn)
1605 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1606 DPUTS("environment had fatal error, must shutdown!");
1610 rc = mdb_txn_renew0(txn);
1611 if (rc == MDB_SUCCESS) {
1612 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1613 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1614 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1620 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1625 if (env->me_flags & MDB_FATAL_ERROR) {
1626 DPUTS("environment had fatal error, must shutdown!");
1630 /* parent already has an active child txn */
1631 if (parent->mt_child) {
1635 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1636 if (!(flags & MDB_RDONLY))
1637 size += env->me_maxdbs * sizeof(MDB_cursor *);
1639 if ((txn = calloc(1, size)) == NULL) {
1640 DPRINTF("calloc: %s", strerror(ErrCode()));
1643 txn->mt_dbs = (MDB_db *)(txn+1);
1644 if (flags & MDB_RDONLY) {
1645 txn->mt_flags |= MDB_TXN_RDONLY;
1646 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1648 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1649 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1654 txn->mt_free_pgs = mdb_midl_alloc();
1655 if (!txn->mt_free_pgs) {
1659 txn->mt_u.dirty_list = malloc(sizeof(ID2)*MDB_IDL_UM_SIZE);
1660 if (!txn->mt_u.dirty_list) {
1661 free(txn->mt_free_pgs);
1665 txn->mt_txnid = parent->mt_txnid;
1666 txn->mt_toggle = parent->mt_toggle;
1667 txn->mt_u.dirty_list[0].mid = 0;
1668 txn->mt_free_pgs[0] = 0;
1669 txn->mt_next_pgno = parent->mt_next_pgno;
1670 parent->mt_child = txn;
1671 txn->mt_parent = parent;
1672 txn->mt_numdbs = parent->mt_numdbs;
1673 txn->mt_dbxs = parent->mt_dbxs;
1674 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1675 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1676 mdb_cursor_shadow(parent, txn);
1679 rc = mdb_txn_renew0(txn);
1685 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1686 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1687 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1693 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1694 * @param[in] txn the transaction handle to reset
1697 mdb_txn_reset0(MDB_txn *txn)
1699 MDB_env *env = txn->mt_env;
1701 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1702 txn->mt_u.reader->mr_txnid = 0;
1708 /* close(free) all cursors */
1709 for (i=0; i<txn->mt_numdbs; i++) {
1710 if (txn->mt_cursors[i]) {
1712 while ((mc = txn->mt_cursors[i])) {
1713 txn->mt_cursors[i] = mc->mc_next;
1714 if (mc->mc_flags & C_ALLOCD)
1720 /* return all dirty pages to dpage list */
1721 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1722 dp = txn->mt_u.dirty_list[i].mptr;
1723 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1724 dp->mp_next = txn->mt_env->me_dpages;
1725 VGMEMP_FREE(txn->mt_env, dp);
1726 txn->mt_env->me_dpages = dp;
1728 /* large pages just get freed directly */
1729 VGMEMP_FREE(txn->mt_env, dp);
1734 if (txn->mt_parent) {
1735 txn->mt_parent->mt_child = NULL;
1736 free(txn->mt_free_pgs);
1737 free(txn->mt_u.dirty_list);
1740 if (mdb_midl_shrink(&txn->mt_free_pgs))
1741 env->me_free_pgs = txn->mt_free_pgs;
1744 while ((mop = txn->mt_env->me_pghead)) {
1745 txn->mt_env->me_pghead = mop->mo_next;
1748 txn->mt_env->me_pgfirst = 0;
1749 txn->mt_env->me_pglast = 0;
1752 /* The writer mutex was locked in mdb_txn_begin. */
1753 UNLOCK_MUTEX_W(env);
1758 mdb_txn_reset(MDB_txn *txn)
1763 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1764 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1765 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1767 mdb_txn_reset0(txn);
1771 mdb_txn_abort(MDB_txn *txn)
1776 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1777 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1778 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1781 mdb_txn_abort(txn->mt_child);
1783 mdb_txn_reset0(txn);
1788 mdb_txn_commit(MDB_txn *txn)
1796 pgno_t next, freecnt;
1799 assert(txn != NULL);
1800 assert(txn->mt_env != NULL);
1802 if (txn->mt_child) {
1803 mdb_txn_commit(txn->mt_child);
1804 txn->mt_child = NULL;
1809 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1810 if (txn->mt_numdbs > env->me_numdbs) {
1811 /* update the DB tables */
1812 int toggle = !env->me_db_toggle;
1816 ip = &env->me_dbs[toggle][env->me_numdbs];
1817 jp = &txn->mt_dbs[env->me_numdbs];
1818 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1819 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
1823 env->me_db_toggle = toggle;
1824 env->me_numdbs = txn->mt_numdbs;
1825 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1831 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1832 DPUTS("error flag is set, can't commit");
1834 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1839 /* Merge (and close) our cursors with parent's */
1840 mdb_cursor_merge(txn);
1842 if (txn->mt_parent) {
1848 /* Update parent's DB table */
1849 ip = &txn->mt_parent->mt_dbs[2];
1850 jp = &txn->mt_dbs[2];
1851 for (i = 2; i < txn->mt_numdbs; i++) {
1852 if (ip->md_root != jp->md_root)
1856 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1858 /* Append our free list to parent's */
1859 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1861 mdb_midl_free(txn->mt_free_pgs);
1863 /* Merge our dirty list with parent's */
1864 dst = txn->mt_parent->mt_u.dirty_list;
1865 src = txn->mt_u.dirty_list;
1866 x = mdb_mid2l_search(dst, src[1].mid);
1867 for (y=1; y<=src[0].mid; y++) {
1868 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1872 dst[x].mptr = src[y].mptr;
1875 for (; y<=src[0].mid; y++) {
1876 if (++x >= MDB_IDL_UM_MAX) {
1883 free(txn->mt_u.dirty_list);
1884 txn->mt_parent->mt_child = NULL;
1889 if (txn != env->me_txn) {
1890 DPUTS("attempt to commit unknown transaction");
1895 if (!txn->mt_u.dirty_list[0].mid)
1898 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1899 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1901 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1903 /* should only be one record now */
1904 if (env->me_pghead) {
1905 /* make sure first page of freeDB is touched and on freelist */
1906 mdb_page_search(&mc, NULL, 1);
1909 /* Delete IDLs we used from the free list */
1910 if (env->me_pgfirst) {
1915 key.mv_size = sizeof(cur);
1916 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
1919 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
1920 mdb_cursor_del(&mc, 0);
1922 env->me_pgfirst = 0;
1926 /* save to free list */
1928 freecnt = txn->mt_free_pgs[0];
1929 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1932 /* make sure last page of freeDB is touched and on freelist */
1933 key.mv_size = MAXKEYSIZE+1;
1935 mdb_page_search(&mc, &key, 1);
1937 mdb_midl_sort(txn->mt_free_pgs);
1941 ID *idl = txn->mt_free_pgs;
1942 DPRINTF("IDL write txn %zu root %zu num %zu",
1943 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1944 for (i=0; i<idl[0]; i++) {
1945 DPRINTF("IDL %zu", idl[i+1]);
1949 /* write to last page of freeDB */
1950 key.mv_size = sizeof(pgno_t);
1951 key.mv_data = &txn->mt_txnid;
1952 data.mv_data = txn->mt_free_pgs;
1953 /* The free list can still grow during this call,
1954 * despite the pre-emptive touches above. So check
1955 * and make sure the entire thing got written.
1958 freecnt = txn->mt_free_pgs[0];
1959 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1960 rc = mdb_cursor_put(&mc, &key, &data, 0);
1965 } while (freecnt != txn->mt_free_pgs[0]);
1967 /* should only be one record now */
1969 if (env->me_pghead) {
1975 mop = env->me_pghead;
1977 key.mv_size = sizeof(id);
1979 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1980 data.mv_data = mop->mo_pages;
1981 orig = mop->mo_pages[0];
1982 /* These steps may grow the freelist again
1983 * due to freed overflow pages...
1985 mdb_cursor_put(&mc, &key, &data, 0);
1986 if (mop == env->me_pghead) {
1987 /* could have been used again here */
1988 if (mop->mo_pages[0] != orig) {
1989 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1990 data.mv_data = mop->mo_pages;
1992 mdb_cursor_put(&mc, &key, &data, 0);
1994 env->me_pghead = NULL;
1997 /* was completely used up */
1998 mdb_cursor_del(&mc, 0);
2002 env->me_pgfirst = 0;
2005 /* Check for growth of freelist again */
2006 if (freecnt != txn->mt_free_pgs[0])
2009 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2010 if (mdb_midl_shrink(&txn->mt_free_pgs))
2011 env->me_free_pgs = txn->mt_free_pgs;
2014 /* Update DB root pointers. Their pages have already been
2015 * touched so this is all in-place and cannot fail.
2020 data.mv_size = sizeof(MDB_db);
2022 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2023 for (i = 2; i < txn->mt_numdbs; i++) {
2024 if (txn->mt_dbflags[i] & DB_DIRTY) {
2025 data.mv_data = &txn->mt_dbs[i];
2026 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2034 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2040 /* Windows actually supports scatter/gather I/O, but only on
2041 * unbuffered file handles. Since we're relying on the OS page
2042 * cache for all our data, that's self-defeating. So we just
2043 * write pages one at a time. We use the ov structure to set
2044 * the write offset, to at least save the overhead of a Seek
2048 memset(&ov, 0, sizeof(ov));
2049 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2051 dp = txn->mt_u.dirty_list[i].mptr;
2052 DPRINTF("committing page %zu", dp->mp_pgno);
2053 size = dp->mp_pgno * env->me_psize;
2054 ov.Offset = size & 0xffffffff;
2055 ov.OffsetHigh = size >> 16;
2056 ov.OffsetHigh >>= 16;
2057 /* clear dirty flag */
2058 dp->mp_flags &= ~P_DIRTY;
2059 wsize = env->me_psize;
2060 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2061 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2064 DPRINTF("WriteFile: %d", n);
2071 struct iovec iov[MDB_COMMIT_PAGES];
2075 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2076 dp = txn->mt_u.dirty_list[i].mptr;
2077 if (dp->mp_pgno != next) {
2079 rc = writev(env->me_fd, iov, n);
2083 DPUTS("short write, filesystem full?");
2085 DPRINTF("writev: %s", strerror(n));
2092 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2095 DPRINTF("committing page %zu", dp->mp_pgno);
2096 iov[n].iov_len = env->me_psize;
2097 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2098 iov[n].iov_base = (char *)dp;
2099 size += iov[n].iov_len;
2100 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2101 /* clear dirty flag */
2102 dp->mp_flags &= ~P_DIRTY;
2103 if (++n >= MDB_COMMIT_PAGES) {
2113 rc = writev(env->me_fd, iov, n);
2117 DPUTS("short write, filesystem full?");
2119 DPRINTF("writev: %s", strerror(n));
2126 /* Drop the dirty pages.
2128 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2129 dp = txn->mt_u.dirty_list[i].mptr;
2130 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2131 dp->mp_next = txn->mt_env->me_dpages;
2132 VGMEMP_FREE(txn->mt_env, dp);
2133 txn->mt_env->me_dpages = dp;
2135 VGMEMP_FREE(txn->mt_env, dp);
2138 txn->mt_u.dirty_list[i].mid = 0;
2140 txn->mt_u.dirty_list[0].mid = 0;
2142 if ((n = mdb_env_sync(env, 0)) != 0 ||
2143 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2147 env->me_wtxnid = txn->mt_txnid;
2151 /* update the DB tables */
2153 int toggle = !env->me_db_toggle;
2157 ip = &env->me_dbs[toggle][2];
2158 jp = &txn->mt_dbs[2];
2159 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
2160 for (i = 2; i < txn->mt_numdbs; i++) {
2161 if (ip->md_root != jp->md_root)
2166 env->me_db_toggle = toggle;
2167 env->me_numdbs = txn->mt_numdbs;
2168 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
2171 UNLOCK_MUTEX_W(env);
2177 /** Read the environment parameters of a DB environment before
2178 * mapping it into memory.
2179 * @param[in] env the environment handle
2180 * @param[out] meta address of where to store the meta information
2181 * @return 0 on success, non-zero on failure.
2184 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2191 /* We don't know the page size yet, so use a minimum value.
2195 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2197 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2202 else if (rc != MDB_PAGESIZE) {
2206 DPRINTF("read: %s", strerror(err));
2210 p = (MDB_page *)&pbuf;
2212 if (!F_ISSET(p->mp_flags, P_META)) {
2213 DPRINTF("page %zu not a meta page", p->mp_pgno);
2218 if (m->mm_magic != MDB_MAGIC) {
2219 DPUTS("meta has invalid magic");
2223 if (m->mm_version != MDB_VERSION) {
2224 DPRINTF("database is version %u, expected version %u",
2225 m->mm_version, MDB_VERSION);
2226 return MDB_VERSION_MISMATCH;
2229 memcpy(meta, m, sizeof(*m));
2233 /** Write the environment parameters of a freshly created DB environment.
2234 * @param[in] env the environment handle
2235 * @param[out] meta address of where to store the meta information
2236 * @return 0 on success, non-zero on failure.
2239 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2246 DPUTS("writing new meta page");
2248 GET_PAGESIZE(psize);
2250 meta->mm_magic = MDB_MAGIC;
2251 meta->mm_version = MDB_VERSION;
2252 meta->mm_psize = psize;
2253 meta->mm_last_pg = 1;
2254 meta->mm_flags = env->me_flags & 0xffff;
2255 meta->mm_flags |= MDB_INTEGERKEY;
2256 meta->mm_dbs[0].md_root = P_INVALID;
2257 meta->mm_dbs[1].md_root = P_INVALID;
2259 p = calloc(2, psize);
2261 p->mp_flags = P_META;
2264 memcpy(m, meta, sizeof(*meta));
2266 q = (MDB_page *)((char *)p + psize);
2269 q->mp_flags = P_META;
2272 memcpy(m, meta, sizeof(*meta));
2277 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2278 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2281 rc = write(env->me_fd, p, psize * 2);
2282 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2288 /** Update the environment info to commit a transaction.
2289 * @param[in] txn the transaction that's being committed
2290 * @return 0 on success, non-zero on failure.
2293 mdb_env_write_meta(MDB_txn *txn)
2296 MDB_meta meta, metab;
2298 int rc, len, toggle;
2304 assert(txn != NULL);
2305 assert(txn->mt_env != NULL);
2307 toggle = !txn->mt_toggle;
2308 DPRINTF("writing meta page %d for root page %zu",
2309 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2313 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2314 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2316 ptr = (char *)&meta;
2317 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2318 len = sizeof(MDB_meta) - off;
2321 meta.mm_dbs[0] = txn->mt_dbs[0];
2322 meta.mm_dbs[1] = txn->mt_dbs[1];
2323 meta.mm_last_pg = txn->mt_next_pgno - 1;
2324 meta.mm_txnid = txn->mt_txnid;
2327 off += env->me_psize;
2330 /* Write to the SYNC fd */
2333 memset(&ov, 0, sizeof(ov));
2335 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2338 rc = pwrite(env->me_mfd, ptr, len, off);
2343 DPUTS("write failed, disk error?");
2344 /* On a failure, the pagecache still contains the new data.
2345 * Write some old data back, to prevent it from being used.
2346 * Use the non-SYNC fd; we know it will fail anyway.
2348 meta.mm_last_pg = metab.mm_last_pg;
2349 meta.mm_txnid = metab.mm_txnid;
2351 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2353 r2 = pwrite(env->me_fd, ptr, len, off);
2355 env->me_flags |= MDB_FATAL_ERROR;
2358 /* Memory ordering issues are irrelevant; since the entire writer
2359 * is wrapped by wmutex, all of these changes will become visible
2360 * after the wmutex is unlocked. Since the DB is multi-version,
2361 * readers will get consistent data regardless of how fresh or
2362 * how stale their view of these values is.
2364 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
2365 txn->mt_env->me_txns->mti_me_toggle = toggle;
2366 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2367 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
2372 /** Check both meta pages to see which one is newer.
2373 * @param[in] env the environment handle
2374 * @param[out] which address of where to store the meta toggle ID
2375 * @return 0 on success, non-zero on failure.
2378 mdb_env_read_meta(MDB_env *env, int *which)
2382 assert(env != NULL);
2384 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2387 DPRINTF("Using meta page %d", toggle);
2394 mdb_env_create(MDB_env **env)
2398 e = calloc(1, sizeof(MDB_env));
2402 e->me_free_pgs = mdb_midl_alloc();
2403 if (!e->me_free_pgs) {
2407 e->me_maxreaders = DEFAULT_READERS;
2409 e->me_fd = INVALID_HANDLE_VALUE;
2410 e->me_lfd = INVALID_HANDLE_VALUE;
2411 e->me_mfd = INVALID_HANDLE_VALUE;
2412 VGMEMP_CREATE(e,0,0);
2418 mdb_env_set_mapsize(MDB_env *env, size_t size)
2422 env->me_mapsize = size;
2424 env->me_maxpg = env->me_mapsize / env->me_psize;
2429 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2433 env->me_maxdbs = dbs;
2438 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2440 if (env->me_map || readers < 1)
2442 env->me_maxreaders = readers;
2447 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2449 if (!env || !readers)
2451 *readers = env->me_maxreaders;
2455 /** Further setup required for opening an MDB environment
2458 mdb_env_open2(MDB_env *env, unsigned int flags)
2460 int i, newenv = 0, toggle;
2464 env->me_flags = flags;
2466 memset(&meta, 0, sizeof(meta));
2468 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2471 DPUTS("new mdbenv");
2475 if (!env->me_mapsize) {
2476 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2482 LONG sizelo, sizehi;
2483 sizelo = env->me_mapsize & 0xffffffff;
2484 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2486 /* Windows won't create mappings for zero length files.
2487 * Just allocate the maxsize right now.
2490 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2491 if (!SetEndOfFile(env->me_fd))
2493 SetFilePointer(env->me_fd, 0, NULL, 0);
2495 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2496 sizehi, sizelo, NULL);
2499 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2507 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2509 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2511 if (env->me_map == MAP_FAILED) {
2518 meta.mm_mapsize = env->me_mapsize;
2519 if (flags & MDB_FIXEDMAP)
2520 meta.mm_address = env->me_map;
2521 i = mdb_env_init_meta(env, &meta);
2522 if (i != MDB_SUCCESS) {
2523 munmap(env->me_map, env->me_mapsize);
2527 env->me_psize = meta.mm_psize;
2529 env->me_maxpg = env->me_mapsize / env->me_psize;
2531 p = (MDB_page *)env->me_map;
2532 env->me_metas[0] = METADATA(p);
2533 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2535 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
2538 DPRINTF("opened database version %u, pagesize %u",
2539 env->me_metas[toggle]->mm_version, env->me_psize);
2540 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
2541 DPRINTF("entries: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
2542 DPRINTF("branch pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
2543 DPRINTF("leaf pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
2544 DPRINTF("overflow pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
2545 DPRINTF("root: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
2551 /** Release a reader thread's slot in the reader lock table.
2552 * This function is called automatically when a thread exits.
2553 * Windows doesn't support destructor callbacks for thread-specific storage,
2554 * so this function is not compiled there.
2555 * @param[in] ptr This points to the slot in the reader lock table.
2558 mdb_env_reader_dest(void *ptr)
2560 MDB_reader *reader = ptr;
2562 reader->mr_txnid = 0;
2568 /** Downgrade the exclusive lock on the region back to shared */
2570 mdb_env_share_locks(MDB_env *env)
2574 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2576 env->me_txns->mti_me_toggle = toggle;
2577 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2582 /* First acquire a shared lock. The Unlock will
2583 * then release the existing exclusive lock.
2585 memset(&ov, 0, sizeof(ov));
2586 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2587 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2591 struct flock lock_info;
2592 /* The shared lock replaces the existing lock */
2593 memset((void *)&lock_info, 0, sizeof(lock_info));
2594 lock_info.l_type = F_RDLCK;
2595 lock_info.l_whence = SEEK_SET;
2596 lock_info.l_start = 0;
2597 lock_info.l_len = 1;
2598 fcntl(env->me_lfd, F_SETLK, &lock_info);
2602 #if defined(_WIN32) || defined(__APPLE__)
2604 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2606 * @(#) $Revision: 5.1 $
2607 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2608 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2610 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2614 * Please do not copyright this code. This code is in the public domain.
2616 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2617 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2618 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2619 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2620 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2621 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2622 * PERFORMANCE OF THIS SOFTWARE.
2625 * chongo <Landon Curt Noll> /\oo/\
2626 * http://www.isthe.com/chongo/
2628 * Share and Enjoy! :-)
2631 typedef unsigned long long mdb_hash_t;
2632 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2634 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2635 * @param[in] str string to hash
2636 * @param[in] hval initial value for hash
2637 * @return 64 bit hash
2639 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2640 * hval arg on the first call.
2643 mdb_hash_str(char *str, mdb_hash_t hval)
2645 unsigned char *s = (unsigned char *)str; /* unsigned string */
2647 * FNV-1a hash each octet of the string
2650 /* xor the bottom with the current octet */
2651 hval ^= (mdb_hash_t)*s++;
2653 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2654 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2655 (hval << 7) + (hval << 8) + (hval << 40);
2657 /* return our new hash value */
2661 /** Hash the string and output the hash in hex.
2662 * @param[in] str string to hash
2663 * @param[out] hexbuf an array of 17 chars to hold the hash
2666 mdb_hash_hex(char *str, char *hexbuf)
2669 mdb_hash_t h = mdb_hash_str(str, MDB_HASH_INIT);
2670 for (i=0; i<8; i++) {
2671 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2677 /** Open and/or initialize the lock region for the environment.
2678 * @param[in] env The MDB environment.
2679 * @param[in] lpath The pathname of the file used for the lock region.
2680 * @param[in] mode The Unix permissions for the file, if we create it.
2681 * @param[out] excl Set to true if we got an exclusive lock on the region.
2682 * @return 0 on success, non-zero on failure.
2685 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2693 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2694 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2695 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2699 /* Try to get exclusive lock. If we succeed, then
2700 * nobody is using the lock region and we should initialize it.
2703 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2707 memset(&ov, 0, sizeof(ov));
2708 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2714 size = GetFileSize(env->me_lfd, NULL);
2716 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2720 /* Try to get exclusive lock. If we succeed, then
2721 * nobody is using the lock region and we should initialize it.
2724 struct flock lock_info;
2725 memset((void *)&lock_info, 0, sizeof(lock_info));
2726 lock_info.l_type = F_WRLCK;
2727 lock_info.l_whence = SEEK_SET;
2728 lock_info.l_start = 0;
2729 lock_info.l_len = 1;
2730 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2734 lock_info.l_type = F_RDLCK;
2735 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2742 size = lseek(env->me_lfd, 0, SEEK_END);
2744 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2745 if (size < rsize && *excl) {
2747 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2748 if (!SetEndOfFile(env->me_lfd)) {
2753 if (ftruncate(env->me_lfd, rsize) != 0) {
2760 size = rsize - sizeof(MDB_txninfo);
2761 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2766 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2772 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2774 if (!env->me_txns) {
2779 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2781 if (m == MAP_FAILED) {
2782 env->me_txns = NULL;
2792 if (!mdb_sec_inited) {
2793 InitializeSecurityDescriptor(&mdb_null_sd,
2794 SECURITY_DESCRIPTOR_REVISION);
2795 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2796 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2797 mdb_all_sa.bInheritHandle = FALSE;
2798 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2801 mdb_hash_hex(lpath, hexbuf);
2802 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2803 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2804 if (!env->me_rmutex) {
2808 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2809 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2810 if (!env->me_wmutex) {
2817 mdb_hash_hex(lpath, hexbuf);
2818 sprintf(env->me_txns->mti_rmname, "MDBr%s", hexbuf);
2819 if (sem_unlink(env->me_txns->mti_rmname)) {
2821 if (rc != ENOENT && rc != EINVAL)
2824 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2825 if (!env->me_rmutex) {
2829 sprintf(env->me_txns->mti_wmname, "MDBw%s", hexbuf);
2830 if (sem_unlink(env->me_txns->mti_wmname)) {
2832 if (rc != ENOENT && rc != EINVAL)
2835 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2836 if (!env->me_wmutex) {
2840 #else /* __APPLE__ */
2841 pthread_mutexattr_t mattr;
2843 pthread_mutexattr_init(&mattr);
2844 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2848 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2849 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2850 #endif /* __APPLE__ */
2852 env->me_txns->mti_version = MDB_VERSION;
2853 env->me_txns->mti_magic = MDB_MAGIC;
2854 env->me_txns->mti_txnid = 0;
2855 env->me_txns->mti_numreaders = 0;
2856 env->me_txns->mti_me_toggle = 0;
2859 if (env->me_txns->mti_magic != MDB_MAGIC) {
2860 DPUTS("lock region has invalid magic");
2864 if (env->me_txns->mti_version != MDB_VERSION) {
2865 DPRINTF("lock region is version %u, expected version %u",
2866 env->me_txns->mti_version, MDB_VERSION);
2867 rc = MDB_VERSION_MISMATCH;
2871 if (rc != EACCES && rc != EAGAIN) {
2875 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2876 if (!env->me_rmutex) {
2880 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2881 if (!env->me_wmutex) {
2887 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2888 if (!env->me_rmutex) {
2892 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2893 if (!env->me_wmutex) {
2903 env->me_lfd = INVALID_HANDLE_VALUE;
2908 /** The name of the lock file in the DB environment */
2909 #define LOCKNAME "/lock.mdb"
2910 /** The name of the data file in the DB environment */
2911 #define DATANAME "/data.mdb"
2912 /** The suffix of the lock file when no subdir is used */
2913 #define LOCKSUFF "-lock"
2916 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2918 int oflags, rc, len, excl;
2919 char *lpath, *dpath;
2922 if (flags & MDB_NOSUBDIR) {
2923 rc = len + sizeof(LOCKSUFF) + len + 1;
2925 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
2930 if (flags & MDB_NOSUBDIR) {
2931 dpath = lpath + len + sizeof(LOCKSUFF);
2932 sprintf(lpath, "%s" LOCKSUFF, path);
2933 strcpy(dpath, path);
2935 dpath = lpath + len + sizeof(LOCKNAME);
2936 sprintf(lpath, "%s" LOCKNAME, path);
2937 sprintf(dpath, "%s" DATANAME, path);
2940 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2945 if (F_ISSET(flags, MDB_RDONLY)) {
2946 oflags = GENERIC_READ;
2947 len = OPEN_EXISTING;
2949 oflags = GENERIC_READ|GENERIC_WRITE;
2952 mode = FILE_ATTRIBUTE_NORMAL;
2953 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2954 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2959 if (F_ISSET(flags, MDB_RDONLY))
2962 oflags = O_RDWR | O_CREAT;
2964 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2970 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2971 /* synchronous fd for meta writes */
2973 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2974 mode |= FILE_FLAG_WRITE_THROUGH;
2975 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2976 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2981 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2982 oflags |= MDB_DSYNC;
2983 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
2988 env->me_path = strdup(path);
2989 DPRINTF("opened dbenv %p", (void *) env);
2990 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
2991 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
2993 mdb_env_share_locks(env);
2994 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
2995 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
2996 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
3002 if (env->me_fd != INVALID_HANDLE_VALUE) {
3004 env->me_fd = INVALID_HANDLE_VALUE;
3006 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3008 env->me_lfd = INVALID_HANDLE_VALUE;
3016 mdb_env_close(MDB_env *env)
3023 VGMEMP_DESTROY(env);
3024 while (env->me_dpages) {
3025 dp = env->me_dpages;
3026 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3027 env->me_dpages = dp->mp_next;
3031 free(env->me_dbs[1]);
3032 free(env->me_dbs[0]);
3036 LAZY_RWLOCK_DESTROY(&env->me_dblock);
3037 pthread_key_delete(env->me_txkey);
3040 munmap(env->me_map, env->me_mapsize);
3045 pid_t pid = getpid();
3047 for (i=0; i<env->me_txns->mti_numreaders; i++)
3048 if (env->me_txns->mti_readers[i].mr_pid == pid)
3049 env->me_txns->mti_readers[i].mr_pid = 0;
3050 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3053 mdb_midl_free(env->me_free_pgs);
3057 /** Compare two items pointing at aligned size_t's */
3059 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3061 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3062 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3065 /** Compare two items pointing at aligned int's */
3067 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3069 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3070 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3073 /** Compare two items pointing at ints of unknown alignment.
3074 * Nodes and keys are guaranteed to be 2-byte aligned.
3077 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3079 #if BYTE_ORDER == LITTLE_ENDIAN
3080 unsigned short *u, *c;
3083 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3084 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3087 } while(!x && u > (unsigned short *)a->mv_data);
3090 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3094 /** Compare two items lexically */
3096 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3103 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3109 diff = memcmp(a->mv_data, b->mv_data, len);
3110 return diff ? diff : len_diff<0 ? -1 : len_diff;
3113 /** Compare two items in reverse byte order */
3115 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3117 const unsigned char *p1, *p2, *p1_lim;
3121 p1_lim = (const unsigned char *)a->mv_data;
3122 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3123 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3125 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3131 while (p1 > p1_lim) {
3132 diff = *--p1 - *--p2;
3136 return len_diff<0 ? -1 : len_diff;
3139 /** Search for key within a page, using binary search.
3140 * Returns the smallest entry larger or equal to the key.
3141 * If exactp is non-null, stores whether the found entry was an exact match
3142 * in *exactp (1 or 0).
3143 * Updates the cursor index with the index of the found entry.
3144 * If no entry larger or equal to the key is found, returns NULL.
3147 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3149 unsigned int i = 0, nkeys;
3152 MDB_page *mp = mc->mc_pg[mc->mc_top];
3153 MDB_node *node = NULL;
3158 nkeys = NUMKEYS(mp);
3163 COPY_PGNO(pgno, mp->mp_pgno);
3164 DPRINTF("searching %u keys in %s %spage %zu",
3165 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3172 low = IS_LEAF(mp) ? 0 : 1;
3174 cmp = mc->mc_dbx->md_cmp;
3176 /* Branch pages have no data, so if using integer keys,
3177 * alignment is guaranteed. Use faster mdb_cmp_int.
3179 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3180 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3187 nodekey.mv_size = mc->mc_db->md_pad;
3188 node = NODEPTR(mp, 0); /* fake */
3189 while (low <= high) {
3190 i = (low + high) >> 1;
3191 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3192 rc = cmp(key, &nodekey);
3193 DPRINTF("found leaf index %u [%s], rc = %i",
3194 i, DKEY(&nodekey), rc);
3203 while (low <= high) {
3204 i = (low + high) >> 1;
3206 node = NODEPTR(mp, i);
3207 nodekey.mv_size = NODEKSZ(node);
3208 nodekey.mv_data = NODEKEY(node);
3210 rc = cmp(key, &nodekey);
3213 DPRINTF("found leaf index %u [%s], rc = %i",
3214 i, DKEY(&nodekey), rc);
3216 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3217 i, DKEY(&nodekey), NODEPGNO(node), rc);
3228 if (rc > 0) { /* Found entry is less than the key. */
3229 i++; /* Skip to get the smallest entry larger than key. */
3231 node = NODEPTR(mp, i);
3234 *exactp = (rc == 0);
3235 /* store the key index */
3236 mc->mc_ki[mc->mc_top] = i;
3238 /* There is no entry larger or equal to the key. */
3241 /* nodeptr is fake for LEAF2 */
3247 mdb_cursor_adjust(MDB_cursor *mc, func)
3251 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3252 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3259 /** Pop a page off the top of the cursor's stack. */
3261 mdb_cursor_pop(MDB_cursor *mc)
3266 top = mc->mc_pg[mc->mc_top];
3271 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3272 mc->mc_dbi, (void *) mc);
3276 /** Push a page onto the top of the cursor's stack. */
3278 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3280 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3281 mc->mc_dbi, (void *) mc);
3283 if (mc->mc_snum >= CURSOR_STACK) {
3284 assert(mc->mc_snum < CURSOR_STACK);
3288 mc->mc_top = mc->mc_snum++;
3289 mc->mc_pg[mc->mc_top] = mp;
3290 mc->mc_ki[mc->mc_top] = 0;
3295 /** Find the address of the page corresponding to a given page number.
3296 * @param[in] txn the transaction for this access.
3297 * @param[in] pgno the page number for the page to retrieve.
3298 * @param[out] ret address of a pointer where the page's address will be stored.
3299 * @return 0 on success, non-zero on failure.
3302 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3306 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3308 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3309 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3310 p = txn->mt_u.dirty_list[x].mptr;
3314 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
3315 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3319 DPRINTF("page %zu not found", pgno);
3322 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3325 /** Search for the page a given key should be in.
3326 * Pushes parent pages on the cursor stack. This function continues a
3327 * search on a cursor that has already been initialized. (Usually by
3328 * #mdb_page_search() but also by #mdb_node_move().)
3329 * @param[in,out] mc the cursor for this operation.
3330 * @param[in] key the key to search for. If NULL, search for the lowest
3331 * page. (This is used by #mdb_cursor_first().)
3332 * @param[in] modify If true, visited pages are updated with new page numbers.
3333 * @return 0 on success, non-zero on failure.
3336 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3338 MDB_page *mp = mc->mc_pg[mc->mc_top];
3343 while (IS_BRANCH(mp)) {
3347 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3348 assert(NUMKEYS(mp) > 1);
3349 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3351 if (key == NULL) /* Initialize cursor to first page. */
3353 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3354 /* cursor to last page */
3358 node = mdb_node_search(mc, key, &exact);
3360 i = NUMKEYS(mp) - 1;
3362 i = mc->mc_ki[mc->mc_top];
3371 DPRINTF("following index %u for key [%s]",
3373 assert(i < NUMKEYS(mp));
3374 node = NODEPTR(mp, i);
3376 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3379 mc->mc_ki[mc->mc_top] = i;
3380 if ((rc = mdb_cursor_push(mc, mp)))
3384 if ((rc = mdb_page_touch(mc)) != 0)
3386 mp = mc->mc_pg[mc->mc_top];
3391 DPRINTF("internal error, index points to a %02X page!?",
3393 return MDB_CORRUPTED;
3396 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3397 key ? DKEY(key) : NULL);
3402 /** Search for the page a given key should be in.
3403 * Pushes parent pages on the cursor stack. This function just sets up
3404 * the search; it finds the root page for \b mc's database and sets this
3405 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3406 * called to complete the search.
3407 * @param[in,out] mc the cursor for this operation.
3408 * @param[in] key the key to search for. If NULL, search for the lowest
3409 * page. (This is used by #mdb_cursor_first().)
3410 * @param[in] modify If true, visited pages are updated with new page numbers.
3411 * @return 0 on success, non-zero on failure.
3414 mdb_page_search(MDB_cursor *mc, MDB_val *key, int modify)
3419 /* Make sure the txn is still viable, then find the root from
3420 * the txn's db table.
3422 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3423 DPUTS("transaction has failed, must abort");
3426 /* Make sure we're using an up-to-date root */
3427 if (mc->mc_dbi > MAIN_DBI) {
3428 if ((*mc->mc_dbflag & DB_STALE) ||
3429 (modify && !(*mc->mc_dbflag & DB_DIRTY))) {
3431 unsigned char dbflag = 0;
3432 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3433 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, modify);
3436 if (*mc->mc_dbflag & DB_STALE) {
3439 MDB_node *leaf = mdb_node_search(&mc2,
3440 &mc->mc_dbx->md_name, &exact);
3442 return MDB_NOTFOUND;
3443 mdb_node_read(mc->mc_txn, leaf, &data);
3444 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3448 *mc->mc_dbflag = dbflag;
3451 root = mc->mc_db->md_root;
3453 if (root == P_INVALID) { /* Tree is empty. */
3454 DPUTS("tree is empty");
3455 return MDB_NOTFOUND;
3460 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3466 DPRINTF("db %u root page %zu has flags 0x%X",
3467 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3470 if ((rc = mdb_page_touch(mc)))
3474 return mdb_page_search_root(mc, key, modify);
3477 /** Return the data associated with a given node.
3478 * @param[in] txn The transaction for this operation.
3479 * @param[in] leaf The node being read.
3480 * @param[out] data Updated to point to the node's data.
3481 * @return 0 on success, non-zero on failure.
3484 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3486 MDB_page *omp; /* overflow page */
3490 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3491 data->mv_size = NODEDSZ(leaf);
3492 data->mv_data = NODEDATA(leaf);
3496 /* Read overflow data.
3498 data->mv_size = NODEDSZ(leaf);
3499 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3500 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3501 DPRINTF("read overflow page %zu failed", pgno);
3504 data->mv_data = METADATA(omp);
3510 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3511 MDB_val *key, MDB_val *data)
3520 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3522 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3525 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3529 mdb_cursor_init(&mc, txn, dbi, &mx);
3530 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3533 /** Find a sibling for a page.
3534 * Replaces the page at the top of the cursor's stack with the
3535 * specified sibling, if one exists.
3536 * @param[in] mc The cursor for this operation.
3537 * @param[in] move_right Non-zero if the right sibling is requested,
3538 * otherwise the left sibling.
3539 * @return 0 on success, non-zero on failure.
3542 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3548 if (mc->mc_snum < 2) {
3549 return MDB_NOTFOUND; /* root has no siblings */
3553 DPRINTF("parent page is page %zu, index %u",
3554 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3556 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3557 : (mc->mc_ki[mc->mc_top] == 0)) {
3558 DPRINTF("no more keys left, moving to %s sibling",
3559 move_right ? "right" : "left");
3560 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3564 mc->mc_ki[mc->mc_top]++;
3566 mc->mc_ki[mc->mc_top]--;
3567 DPRINTF("just moving to %s index key %u",
3568 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3570 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3572 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3573 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3576 mdb_cursor_push(mc, mp);
3581 /** Move the cursor to the next data item. */
3583 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3589 if (mc->mc_flags & C_EOF) {
3590 return MDB_NOTFOUND;
3593 assert(mc->mc_flags & C_INITIALIZED);
3595 mp = mc->mc_pg[mc->mc_top];
3597 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3598 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3599 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3600 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3601 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3602 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3606 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3607 if (op == MDB_NEXT_DUP)
3608 return MDB_NOTFOUND;
3612 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3614 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3615 DPUTS("=====> move to next sibling page");
3616 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3617 mc->mc_flags |= C_EOF;
3618 mc->mc_flags &= ~C_INITIALIZED;
3619 return MDB_NOTFOUND;
3621 mp = mc->mc_pg[mc->mc_top];
3622 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3624 mc->mc_ki[mc->mc_top]++;
3626 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3627 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3630 key->mv_size = mc->mc_db->md_pad;
3631 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3635 assert(IS_LEAF(mp));
3636 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3638 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3639 mdb_xcursor_init1(mc, leaf);
3642 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3645 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3646 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3647 if (rc != MDB_SUCCESS)
3652 MDB_SET_KEY(leaf, key);
3656 /** Move the cursor to the previous data item. */
3658 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3664 assert(mc->mc_flags & C_INITIALIZED);
3666 mp = mc->mc_pg[mc->mc_top];
3668 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3669 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3670 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3671 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3672 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3673 if (op != MDB_PREV || rc == MDB_SUCCESS)
3676 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3677 if (op == MDB_PREV_DUP)
3678 return MDB_NOTFOUND;
3683 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3685 if (mc->mc_ki[mc->mc_top] == 0) {
3686 DPUTS("=====> move to prev sibling page");
3687 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3688 mc->mc_flags &= ~C_INITIALIZED;
3689 return MDB_NOTFOUND;
3691 mp = mc->mc_pg[mc->mc_top];
3692 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3693 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3695 mc->mc_ki[mc->mc_top]--;
3697 mc->mc_flags &= ~C_EOF;
3699 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3700 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3703 key->mv_size = mc->mc_db->md_pad;
3704 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3708 assert(IS_LEAF(mp));
3709 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3711 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3712 mdb_xcursor_init1(mc, leaf);
3715 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3718 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3719 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3720 if (rc != MDB_SUCCESS)
3725 MDB_SET_KEY(leaf, key);
3729 /** Set the cursor on a specific data item. */
3731 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3732 MDB_cursor_op op, int *exactp)
3741 assert(key->mv_size > 0);
3743 /* See if we're already on the right page */
3744 if (mc->mc_flags & C_INITIALIZED) {
3747 mp = mc->mc_pg[mc->mc_top];
3749 mc->mc_ki[mc->mc_top] = 0;
3750 return MDB_NOTFOUND;
3752 if (mp->mp_flags & P_LEAF2) {
3753 nodekey.mv_size = mc->mc_db->md_pad;
3754 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3756 leaf = NODEPTR(mp, 0);
3757 MDB_SET_KEY(leaf, &nodekey);
3759 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3761 /* Probably happens rarely, but first node on the page
3762 * was the one we wanted.
3764 mc->mc_ki[mc->mc_top] = 0;
3765 leaf = NODEPTR(mp, 0);
3772 unsigned int nkeys = NUMKEYS(mp);
3774 if (mp->mp_flags & P_LEAF2) {
3775 nodekey.mv_data = LEAF2KEY(mp,
3776 nkeys-1, nodekey.mv_size);
3778 leaf = NODEPTR(mp, nkeys-1);
3779 MDB_SET_KEY(leaf, &nodekey);
3781 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3783 /* last node was the one we wanted */
3784 mc->mc_ki[mc->mc_top] = nkeys-1;
3785 leaf = NODEPTR(mp, nkeys-1);
3791 /* This is definitely the right page, skip search_page */
3796 /* If any parents have right-sibs, search.
3797 * Otherwise, there's nothing further.
3799 for (i=0; i<mc->mc_top; i++)
3801 NUMKEYS(mc->mc_pg[i])-1)
3803 if (i == mc->mc_top) {
3804 /* There are no other pages */
3805 mc->mc_ki[mc->mc_top] = nkeys;
3806 return MDB_NOTFOUND;
3810 /* There are no other pages */
3811 mc->mc_ki[mc->mc_top] = 0;
3812 return MDB_NOTFOUND;
3816 rc = mdb_page_search(mc, key, 0);
3817 if (rc != MDB_SUCCESS)
3820 mp = mc->mc_pg[mc->mc_top];
3821 assert(IS_LEAF(mp));
3824 leaf = mdb_node_search(mc, key, exactp);
3825 if (exactp != NULL && !*exactp) {
3826 /* MDB_SET specified and not an exact match. */
3827 return MDB_NOTFOUND;
3831 DPUTS("===> inexact leaf not found, goto sibling");
3832 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3833 return rc; /* no entries matched */
3834 mp = mc->mc_pg[mc->mc_top];
3835 assert(IS_LEAF(mp));
3836 leaf = NODEPTR(mp, 0);
3840 mc->mc_flags |= C_INITIALIZED;
3841 mc->mc_flags &= ~C_EOF;
3844 key->mv_size = mc->mc_db->md_pad;
3845 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3849 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3850 mdb_xcursor_init1(mc, leaf);
3853 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3854 if (op == MDB_SET || op == MDB_SET_RANGE) {
3855 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3858 if (op == MDB_GET_BOTH) {
3864 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3865 if (rc != MDB_SUCCESS)
3868 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3870 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3872 rc = mc->mc_dbx->md_dcmp(data, &d2);
3874 if (op == MDB_GET_BOTH || rc > 0)
3875 return MDB_NOTFOUND;
3880 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3881 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3886 /* The key already matches in all other cases */
3887 if (op == MDB_SET_RANGE)
3888 MDB_SET_KEY(leaf, key);
3889 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3894 /** Move the cursor to the first item in the database. */
3896 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3901 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3902 rc = mdb_page_search(mc, NULL, 0);
3903 if (rc != MDB_SUCCESS)
3906 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3908 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3909 mc->mc_flags |= C_INITIALIZED;
3910 mc->mc_flags &= ~C_EOF;
3912 mc->mc_ki[mc->mc_top] = 0;
3914 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3915 key->mv_size = mc->mc_db->md_pad;
3916 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3921 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3922 mdb_xcursor_init1(mc, leaf);
3923 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3928 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3929 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3933 MDB_SET_KEY(leaf, key);
3937 /** Move the cursor to the last item in the database. */
3939 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3945 lkey.mv_size = MAXKEYSIZE+1;
3946 lkey.mv_data = NULL;
3948 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3949 rc = mdb_page_search(mc, &lkey, 0);
3950 if (rc != MDB_SUCCESS)
3953 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3955 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3956 mc->mc_flags |= C_INITIALIZED;
3957 mc->mc_flags &= ~C_EOF;
3959 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3961 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3962 key->mv_size = mc->mc_db->md_pad;
3963 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3968 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3969 mdb_xcursor_init1(mc, leaf);
3970 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3975 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3976 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3981 MDB_SET_KEY(leaf, key);
3986 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3996 case MDB_GET_BOTH_RANGE:
3997 if (data == NULL || mc->mc_xcursor == NULL) {
4004 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4006 } else if (op == MDB_SET_RANGE)
4007 rc = mdb_cursor_set(mc, key, data, op, NULL);
4009 rc = mdb_cursor_set(mc, key, data, op, &exact);
4011 case MDB_GET_MULTIPLE:
4013 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4014 !(mc->mc_flags & C_INITIALIZED)) {
4019 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4020 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4023 case MDB_NEXT_MULTIPLE:
4025 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4029 if (!(mc->mc_flags & C_INITIALIZED))
4030 rc = mdb_cursor_first(mc, key, data);
4032 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4033 if (rc == MDB_SUCCESS) {
4034 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4037 mx = &mc->mc_xcursor->mx_cursor;
4038 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4040 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4041 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4049 case MDB_NEXT_NODUP:
4050 if (!(mc->mc_flags & C_INITIALIZED))
4051 rc = mdb_cursor_first(mc, key, data);
4053 rc = mdb_cursor_next(mc, key, data, op);
4057 case MDB_PREV_NODUP:
4058 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
4059 rc = mdb_cursor_last(mc, key, data);
4061 rc = mdb_cursor_prev(mc, key, data, op);
4064 rc = mdb_cursor_first(mc, key, data);
4068 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4069 !(mc->mc_flags & C_INITIALIZED) ||
4070 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4074 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4077 rc = mdb_cursor_last(mc, key, data);
4081 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4082 !(mc->mc_flags & C_INITIALIZED) ||
4083 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4087 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4090 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4098 /** Touch all the pages in the cursor stack.
4099 * Makes sure all the pages are writable, before attempting a write operation.
4100 * @param[in] mc The cursor to operate on.
4103 mdb_cursor_touch(MDB_cursor *mc)
4107 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4109 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4110 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 1);
4113 *mc->mc_dbflag = DB_DIRTY;
4115 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4116 rc = mdb_page_touch(mc);
4120 mc->mc_top = mc->mc_snum-1;
4125 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4128 MDB_node *leaf = NULL;
4129 MDB_val xdata, *rdata, dkey;
4133 unsigned int mcount = 0;
4137 char dbuf[MAXKEYSIZE+1];
4138 unsigned int nflags;
4141 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4144 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4145 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4149 if (flags == MDB_CURRENT) {
4150 if (!(mc->mc_flags & C_INITIALIZED))
4153 } else if (mc->mc_db->md_root == P_INVALID) {
4155 /* new database, write a root leaf page */
4156 DPUTS("allocating new root leaf page");
4157 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4161 mdb_cursor_push(mc, np);
4162 mc->mc_db->md_root = np->mp_pgno;
4163 mc->mc_db->md_depth++;
4164 *mc->mc_dbflag = DB_DIRTY;
4165 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4167 np->mp_flags |= P_LEAF2;
4168 mc->mc_flags |= C_INITIALIZED;
4174 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4175 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4176 DPRINTF("duplicate key [%s]", DKEY(key));
4178 return MDB_KEYEXIST;
4180 if (rc && rc != MDB_NOTFOUND)
4184 /* Cursor is positioned, now make sure all pages are writable */
4185 rc2 = mdb_cursor_touch(mc);
4190 /* The key already exists */
4191 if (rc == MDB_SUCCESS) {
4192 /* there's only a key anyway, so this is a no-op */
4193 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4194 unsigned int ksize = mc->mc_db->md_pad;
4195 if (key->mv_size != ksize)
4197 if (flags == MDB_CURRENT) {
4198 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4199 memcpy(ptr, key->mv_data, ksize);
4204 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4207 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4208 /* Was a single item before, must convert now */
4210 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4211 /* Just overwrite the current item */
4212 if (flags == MDB_CURRENT)
4215 dkey.mv_size = NODEDSZ(leaf);
4216 dkey.mv_data = NODEDATA(leaf);
4217 #if UINT_MAX < SIZE_MAX
4218 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4219 #ifdef MISALIGNED_OK
4220 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4222 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4225 /* if data matches, ignore it */
4226 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4227 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4229 /* create a fake page for the dup items */
4230 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4231 dkey.mv_data = dbuf;
4232 fp = (MDB_page *)&pbuf;
4233 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4234 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4235 fp->mp_lower = PAGEHDRSZ;
4236 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4237 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4238 fp->mp_flags |= P_LEAF2;
4239 fp->mp_pad = data->mv_size;
4241 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4242 (dkey.mv_size & 1) + (data->mv_size & 1);
4244 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4247 xdata.mv_size = fp->mp_upper;
4252 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4253 /* See if we need to convert from fake page to subDB */
4255 unsigned int offset;
4258 fp = NODEDATA(leaf);
4259 if (flags == MDB_CURRENT) {
4260 fp->mp_flags |= P_DIRTY;
4261 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4262 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4266 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4267 offset = fp->mp_pad;
4269 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4271 offset += offset & 1;
4272 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4273 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4275 /* yes, convert it */
4277 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4278 dummy.md_pad = fp->mp_pad;
4279 dummy.md_flags = MDB_DUPFIXED;
4280 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4281 dummy.md_flags |= MDB_INTEGERKEY;
4284 dummy.md_branch_pages = 0;
4285 dummy.md_leaf_pages = 1;
4286 dummy.md_overflow_pages = 0;
4287 dummy.md_entries = NUMKEYS(fp);
4289 xdata.mv_size = sizeof(MDB_db);
4290 xdata.mv_data = &dummy;
4291 mp = mdb_page_alloc(mc, 1);
4294 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4295 flags |= F_DUPDATA|F_SUBDATA;
4296 dummy.md_root = mp->mp_pgno;
4298 /* no, just grow it */
4300 xdata.mv_size = NODEDSZ(leaf) + offset;
4301 xdata.mv_data = &pbuf;
4302 mp = (MDB_page *)&pbuf;
4303 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4306 mp->mp_flags = fp->mp_flags | P_DIRTY;
4307 mp->mp_pad = fp->mp_pad;
4308 mp->mp_lower = fp->mp_lower;
4309 mp->mp_upper = fp->mp_upper + offset;
4311 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4313 nsize = NODEDSZ(leaf) - fp->mp_upper;
4314 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4315 for (i=0; i<NUMKEYS(fp); i++)
4316 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4318 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4322 /* data is on sub-DB, just store it */
4323 flags |= F_DUPDATA|F_SUBDATA;
4327 /* overflow page overwrites need special handling */
4328 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4331 int ovpages, dpages;
4333 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4334 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4335 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4336 mdb_page_get(mc->mc_txn, pg, &omp);
4337 /* Is the ov page writable and large enough? */
4338 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4339 /* yes, overwrite it. Note in this case we don't
4340 * bother to try shrinking the node if the new data
4341 * is smaller than the overflow threshold.
4343 if (F_ISSET(flags, MDB_RESERVE))
4344 data->mv_data = METADATA(omp);
4346 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4349 /* no, free ovpages */
4351 mc->mc_db->md_overflow_pages -= ovpages;
4352 for (i=0; i<ovpages; i++) {
4353 DPRINTF("freed ov page %zu", pg);
4354 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4358 } else if (NODEDSZ(leaf) == data->mv_size) {
4359 /* same size, just replace it. Note that we could
4360 * also reuse this node if the new data is smaller,
4361 * but instead we opt to shrink the node in that case.
4363 if (F_ISSET(flags, MDB_RESERVE))
4364 data->mv_data = NODEDATA(leaf);
4366 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4369 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4370 mc->mc_db->md_entries--;
4372 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4378 nflags = flags & NODE_ADD_FLAGS;
4379 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4380 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4381 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4382 nflags &= ~MDB_APPEND;
4383 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4385 /* There is room already in this leaf page. */
4386 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4387 if (rc == 0 && !do_sub) {
4388 /* Adjust other cursors pointing to mp */
4389 MDB_cursor *m2, *m3;
4390 MDB_dbi dbi = mc->mc_dbi;
4391 unsigned i = mc->mc_top;
4392 MDB_page *mp = mc->mc_pg[i];
4394 if (mc->mc_flags & C_SUB)
4397 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4398 if (mc->mc_flags & C_SUB)
4399 m3 = &m2->mc_xcursor->mx_cursor;
4402 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4403 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4410 if (rc != MDB_SUCCESS)
4411 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4413 /* Now store the actual data in the child DB. Note that we're
4414 * storing the user data in the keys field, so there are strict
4415 * size limits on dupdata. The actual data fields of the child
4416 * DB are all zero size.
4423 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4424 if (flags & MDB_CURRENT) {
4425 xflags = MDB_CURRENT;
4427 mdb_xcursor_init1(mc, leaf);
4428 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4430 /* converted, write the original data first */
4432 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4436 /* Adjust other cursors pointing to mp */
4438 unsigned i = mc->mc_top;
4439 MDB_page *mp = mc->mc_pg[i];
4441 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4442 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4443 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4444 mdb_xcursor_init1(m2, leaf);
4449 xflags |= (flags & MDB_APPEND);
4450 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4451 if (flags & F_SUBDATA) {
4452 void *db = NODEDATA(leaf);
4453 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4456 /* sub-writes might have failed so check rc again.
4457 * Don't increment count if we just replaced an existing item.
4459 if (!rc && !(flags & MDB_CURRENT))
4460 mc->mc_db->md_entries++;
4461 if (flags & MDB_MULTIPLE) {
4463 if (mcount < data[1].mv_size) {
4464 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4465 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4475 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4480 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4483 if (!mc->mc_flags & C_INITIALIZED)
4486 rc = mdb_cursor_touch(mc);
4490 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4492 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4493 if (flags != MDB_NODUPDATA) {
4494 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4495 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4497 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4498 /* If sub-DB still has entries, we're done */
4499 if (mc->mc_xcursor->mx_db.md_entries) {
4500 if (leaf->mn_flags & F_SUBDATA) {
4501 /* update subDB info */
4502 void *db = NODEDATA(leaf);
4503 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4505 /* shrink fake page */
4506 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4508 mc->mc_db->md_entries--;
4511 /* otherwise fall thru and delete the sub-DB */
4514 if (leaf->mn_flags & F_SUBDATA) {
4515 /* add all the child DB's pages to the free list */
4516 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4517 if (rc == MDB_SUCCESS) {
4518 mc->mc_db->md_entries -=
4519 mc->mc_xcursor->mx_db.md_entries;
4524 return mdb_cursor_del0(mc, leaf);
4527 /** Allocate and initialize new pages for a database.
4528 * @param[in] mc a cursor on the database being added to.
4529 * @param[in] flags flags defining what type of page is being allocated.
4530 * @param[in] num the number of pages to allocate. This is usually 1,
4531 * unless allocating overflow pages for a large record.
4532 * @return Address of a page, or NULL on failure.
4535 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4539 if ((np = mdb_page_alloc(mc, num)) == NULL)
4541 DPRINTF("allocated new mpage %zu, page size %u",
4542 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4543 np->mp_flags = flags | P_DIRTY;
4544 np->mp_lower = PAGEHDRSZ;
4545 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4548 mc->mc_db->md_branch_pages++;
4549 else if (IS_LEAF(np))
4550 mc->mc_db->md_leaf_pages++;
4551 else if (IS_OVERFLOW(np)) {
4552 mc->mc_db->md_overflow_pages += num;
4559 /** Calculate the size of a leaf node.
4560 * The size depends on the environment's page size; if a data item
4561 * is too large it will be put onto an overflow page and the node
4562 * size will only include the key and not the data. Sizes are always
4563 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4564 * of the #MDB_node headers.
4565 * @param[in] env The environment handle.
4566 * @param[in] key The key for the node.
4567 * @param[in] data The data for the node.
4568 * @return The number of bytes needed to store the node.
4571 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4575 sz = LEAFSIZE(key, data);
4576 if (sz >= env->me_psize / MDB_MINKEYS) {
4577 /* put on overflow page */
4578 sz -= data->mv_size - sizeof(pgno_t);
4582 return sz + sizeof(indx_t);
4585 /** Calculate the size of a branch node.
4586 * The size should depend on the environment's page size but since
4587 * we currently don't support spilling large keys onto overflow
4588 * pages, it's simply the size of the #MDB_node header plus the
4589 * size of the key. Sizes are always rounded up to an even number
4590 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4591 * @param[in] env The environment handle.
4592 * @param[in] key The key for the node.
4593 * @return The number of bytes needed to store the node.
4596 mdb_branch_size(MDB_env *env, MDB_val *key)
4601 if (sz >= env->me_psize / MDB_MINKEYS) {
4602 /* put on overflow page */
4603 /* not implemented */
4604 /* sz -= key->size - sizeof(pgno_t); */
4607 return sz + sizeof(indx_t);
4610 /** Add a node to the page pointed to by the cursor.
4611 * @param[in] mc The cursor for this operation.
4612 * @param[in] indx The index on the page where the new node should be added.
4613 * @param[in] key The key for the new node.
4614 * @param[in] data The data for the new node, if any.
4615 * @param[in] pgno The page number, if adding a branch node.
4616 * @param[in] flags Flags for the node.
4617 * @return 0 on success, non-zero on failure. Possible errors are:
4619 * <li>ENOMEM - failed to allocate overflow pages for the node.
4620 * <li>ENOSPC - there is insufficient room in the page. This error
4621 * should never happen since all callers already calculate the
4622 * page's free space before calling this function.
4626 mdb_node_add(MDB_cursor *mc, indx_t indx,
4627 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4630 size_t node_size = NODESIZE;
4633 MDB_page *mp = mc->mc_pg[mc->mc_top];
4634 MDB_page *ofp = NULL; /* overflow page */
4637 assert(mp->mp_upper >= mp->mp_lower);
4639 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4640 IS_LEAF(mp) ? "leaf" : "branch",
4641 IS_SUBP(mp) ? "sub-" : "",
4642 mp->mp_pgno, indx, data ? data->mv_size : 0,
4643 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4646 /* Move higher keys up one slot. */
4647 int ksize = mc->mc_db->md_pad, dif;
4648 char *ptr = LEAF2KEY(mp, indx, ksize);
4649 dif = NUMKEYS(mp) - indx;
4651 memmove(ptr+ksize, ptr, dif*ksize);
4652 /* insert new key */
4653 memcpy(ptr, key->mv_data, ksize);
4655 /* Just using these for counting */
4656 mp->mp_lower += sizeof(indx_t);
4657 mp->mp_upper -= ksize - sizeof(indx_t);
4662 node_size += key->mv_size;
4666 if (F_ISSET(flags, F_BIGDATA)) {
4667 /* Data already on overflow page. */
4668 node_size += sizeof(pgno_t);
4669 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4670 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4671 /* Put data on overflow page. */
4672 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4673 data->mv_size, node_size+data->mv_size);
4674 node_size += sizeof(pgno_t);
4675 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4677 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4680 node_size += data->mv_size;
4683 node_size += node_size & 1;
4685 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4686 DPRINTF("not enough room in page %zu, got %u ptrs",
4687 mp->mp_pgno, NUMKEYS(mp));
4688 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4689 mp->mp_upper - mp->mp_lower);
4690 DPRINTF("node size = %zu", node_size);
4694 /* Move higher pointers up one slot. */
4695 for (i = NUMKEYS(mp); i > indx; i--)
4696 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4698 /* Adjust free space offsets. */
4699 ofs = mp->mp_upper - node_size;
4700 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4701 mp->mp_ptrs[indx] = ofs;
4703 mp->mp_lower += sizeof(indx_t);
4705 /* Write the node data. */
4706 node = NODEPTR(mp, indx);
4707 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4708 node->mn_flags = flags;
4710 SETDSZ(node,data->mv_size);
4715 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4720 if (F_ISSET(flags, F_BIGDATA))
4721 memcpy(node->mn_data + key->mv_size, data->mv_data,
4723 else if (F_ISSET(flags, MDB_RESERVE))
4724 data->mv_data = node->mn_data + key->mv_size;
4726 memcpy(node->mn_data + key->mv_size, data->mv_data,
4729 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4731 if (F_ISSET(flags, MDB_RESERVE))
4732 data->mv_data = METADATA(ofp);
4734 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4741 /** Delete the specified node from a page.
4742 * @param[in] mp The page to operate on.
4743 * @param[in] indx The index of the node to delete.
4744 * @param[in] ksize The size of a node. Only used if the page is
4745 * part of a #MDB_DUPFIXED database.
4748 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4751 indx_t i, j, numkeys, ptr;
4758 COPY_PGNO(pgno, mp->mp_pgno);
4759 DPRINTF("delete node %u on %s page %zu", indx,
4760 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4763 assert(indx < NUMKEYS(mp));
4766 int x = NUMKEYS(mp) - 1 - indx;
4767 base = LEAF2KEY(mp, indx, ksize);
4769 memmove(base, base + ksize, x * ksize);
4770 mp->mp_lower -= sizeof(indx_t);
4771 mp->mp_upper += ksize - sizeof(indx_t);
4775 node = NODEPTR(mp, indx);
4776 sz = NODESIZE + node->mn_ksize;
4778 if (F_ISSET(node->mn_flags, F_BIGDATA))
4779 sz += sizeof(pgno_t);
4781 sz += NODEDSZ(node);
4785 ptr = mp->mp_ptrs[indx];
4786 numkeys = NUMKEYS(mp);
4787 for (i = j = 0; i < numkeys; i++) {
4789 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4790 if (mp->mp_ptrs[i] < ptr)
4791 mp->mp_ptrs[j] += sz;
4796 base = (char *)mp + mp->mp_upper;
4797 memmove(base + sz, base, ptr - mp->mp_upper);
4799 mp->mp_lower -= sizeof(indx_t);
4803 /** Compact the main page after deleting a node on a subpage.
4804 * @param[in] mp The main page to operate on.
4805 * @param[in] indx The index of the subpage on the main page.
4808 mdb_node_shrink(MDB_page *mp, indx_t indx)
4815 indx_t i, numkeys, ptr;
4817 node = NODEPTR(mp, indx);
4818 sp = (MDB_page *)NODEDATA(node);
4819 osize = NODEDSZ(node);
4821 delta = sp->mp_upper - sp->mp_lower;
4822 SETDSZ(node, osize - delta);
4823 xp = (MDB_page *)((char *)sp + delta);
4825 /* shift subpage upward */
4827 nsize = NUMKEYS(sp) * sp->mp_pad;
4828 memmove(METADATA(xp), METADATA(sp), nsize);
4831 nsize = osize - sp->mp_upper;
4832 numkeys = NUMKEYS(sp);
4833 for (i=numkeys-1; i>=0; i--)
4834 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4836 xp->mp_upper = sp->mp_lower;
4837 xp->mp_lower = sp->mp_lower;
4838 xp->mp_flags = sp->mp_flags;
4839 xp->mp_pad = sp->mp_pad;
4840 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4842 /* shift lower nodes upward */
4843 ptr = mp->mp_ptrs[indx];
4844 numkeys = NUMKEYS(mp);
4845 for (i = 0; i < numkeys; i++) {
4846 if (mp->mp_ptrs[i] <= ptr)
4847 mp->mp_ptrs[i] += delta;
4850 base = (char *)mp + mp->mp_upper;
4851 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4852 mp->mp_upper += delta;
4855 /** Initial setup of a sorted-dups cursor.
4856 * Sorted duplicates are implemented as a sub-database for the given key.
4857 * The duplicate data items are actually keys of the sub-database.
4858 * Operations on the duplicate data items are performed using a sub-cursor
4859 * initialized when the sub-database is first accessed. This function does
4860 * the preliminary setup of the sub-cursor, filling in the fields that
4861 * depend only on the parent DB.
4862 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4865 mdb_xcursor_init0(MDB_cursor *mc)
4867 MDB_xcursor *mx = mc->mc_xcursor;
4869 mx->mx_cursor.mc_xcursor = NULL;
4870 mx->mx_cursor.mc_txn = mc->mc_txn;
4871 mx->mx_cursor.mc_db = &mx->mx_db;
4872 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
4873 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
4874 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
4875 mx->mx_cursor.mc_snum = 0;
4876 mx->mx_cursor.mc_top = 0;
4877 mx->mx_cursor.mc_flags = C_SUB;
4878 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
4879 mx->mx_dbx.md_dcmp = NULL;
4880 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
4883 /** Final setup of a sorted-dups cursor.
4884 * Sets up the fields that depend on the data from the main cursor.
4885 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4886 * @param[in] node The data containing the #MDB_db record for the
4887 * sorted-dup database.
4890 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
4892 MDB_xcursor *mx = mc->mc_xcursor;
4894 if (node->mn_flags & F_SUBDATA) {
4895 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
4896 mx->mx_cursor.mc_snum = 0;
4897 mx->mx_cursor.mc_flags = C_SUB;
4899 MDB_page *fp = NODEDATA(node);
4900 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
4901 mx->mx_db.md_flags = 0;
4902 mx->mx_db.md_depth = 1;
4903 mx->mx_db.md_branch_pages = 0;
4904 mx->mx_db.md_leaf_pages = 1;
4905 mx->mx_db.md_overflow_pages = 0;
4906 mx->mx_db.md_entries = NUMKEYS(fp);
4907 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
4908 mx->mx_cursor.mc_snum = 1;
4909 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
4910 mx->mx_cursor.mc_top = 0;
4911 mx->mx_cursor.mc_pg[0] = fp;
4912 mx->mx_cursor.mc_ki[0] = 0;
4913 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4914 mx->mx_db.md_flags = MDB_DUPFIXED;
4915 mx->mx_db.md_pad = fp->mp_pad;
4916 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4917 mx->mx_db.md_flags |= MDB_INTEGERKEY;
4920 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
4922 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
4924 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
4925 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
4926 #if UINT_MAX < SIZE_MAX
4927 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
4928 #ifdef MISALIGNED_OK
4929 mx->mx_dbx.md_cmp = mdb_cmp_long;
4931 mx->mx_dbx.md_cmp = mdb_cmp_cint;
4936 /** Initialize a cursor for a given transaction and database. */
4938 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
4943 mc->mc_db = &txn->mt_dbs[dbi];
4944 mc->mc_dbx = &txn->mt_dbxs[dbi];
4945 mc->mc_dbflag = &txn->mt_dbflags[dbi];
4949 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4951 mc->mc_xcursor = mx;
4952 mdb_xcursor_init0(mc);
4954 mc->mc_xcursor = NULL;
4959 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
4962 MDB_xcursor *mx = NULL;
4963 size_t size = sizeof(MDB_cursor);
4965 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
4968 /* Allow read access to the freelist */
4969 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
4972 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
4973 size += sizeof(MDB_xcursor);
4975 if ((mc = malloc(size)) != NULL) {
4976 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4977 mx = (MDB_xcursor *)(mc + 1);
4979 mdb_cursor_init(mc, txn, dbi, mx);
4980 if (txn->mt_cursors) {
4981 mc->mc_next = txn->mt_cursors[dbi];
4982 txn->mt_cursors[dbi] = mc;
4984 mc->mc_flags |= C_ALLOCD;
4994 /* Return the count of duplicate data items for the current key */
4996 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5000 if (mc == NULL || countp == NULL)
5003 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5006 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5007 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5010 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5013 *countp = mc->mc_xcursor->mx_db.md_entries;
5019 mdb_cursor_close(MDB_cursor *mc)
5022 /* remove from txn, if tracked */
5023 if (mc->mc_txn->mt_cursors) {
5024 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5025 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5027 *prev = mc->mc_next;
5029 if (mc->mc_flags & C_ALLOCD)
5035 mdb_cursor_txn(MDB_cursor *mc)
5037 if (!mc) return NULL;
5042 mdb_cursor_dbi(MDB_cursor *mc)
5048 /** Replace the key for a node with a new key.
5049 * @param[in] mp The page containing the node to operate on.
5050 * @param[in] indx The index of the node to operate on.
5051 * @param[in] key The new key to use.
5052 * @return 0 on success, non-zero on failure.
5055 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5057 indx_t ptr, i, numkeys;
5064 node = NODEPTR(mp, indx);
5065 ptr = mp->mp_ptrs[indx];
5069 char kbuf2[(MAXKEYSIZE*2+1)];
5070 k2.mv_data = NODEKEY(node);
5071 k2.mv_size = node->mn_ksize;
5072 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5074 mdb_dkey(&k2, kbuf2),
5080 delta = key->mv_size - node->mn_ksize;
5082 if (delta > 0 && SIZELEFT(mp) < delta) {
5083 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5087 numkeys = NUMKEYS(mp);
5088 for (i = 0; i < numkeys; i++) {
5089 if (mp->mp_ptrs[i] <= ptr)
5090 mp->mp_ptrs[i] -= delta;
5093 base = (char *)mp + mp->mp_upper;
5094 len = ptr - mp->mp_upper + NODESIZE;
5095 memmove(base - delta, base, len);
5096 mp->mp_upper -= delta;
5098 node = NODEPTR(mp, indx);
5099 node->mn_ksize = key->mv_size;
5103 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5108 /** Move a node from csrc to cdst.
5111 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5117 unsigned short flags;
5121 /* Mark src and dst as dirty. */
5122 if ((rc = mdb_page_touch(csrc)) ||
5123 (rc = mdb_page_touch(cdst)))
5126 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5127 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5128 key.mv_size = csrc->mc_db->md_pad;
5129 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5131 data.mv_data = NULL;
5135 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5136 assert(!((long)srcnode&1));
5137 srcpg = NODEPGNO(srcnode);
5138 flags = srcnode->mn_flags;
5139 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5140 unsigned int snum = csrc->mc_snum;
5142 /* must find the lowest key below src */
5143 mdb_page_search_root(csrc, NULL, 0);
5144 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5145 key.mv_size = NODEKSZ(s2);
5146 key.mv_data = NODEKEY(s2);
5147 csrc->mc_snum = snum--;
5148 csrc->mc_top = snum;
5150 key.mv_size = NODEKSZ(srcnode);
5151 key.mv_data = NODEKEY(srcnode);
5153 data.mv_size = NODEDSZ(srcnode);
5154 data.mv_data = NODEDATA(srcnode);
5156 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5157 unsigned int snum = cdst->mc_snum;
5160 /* must find the lowest key below dst */
5161 mdb_page_search_root(cdst, NULL, 0);
5162 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5163 bkey.mv_size = NODEKSZ(s2);
5164 bkey.mv_data = NODEKEY(s2);
5165 cdst->mc_snum = snum--;
5166 cdst->mc_top = snum;
5167 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5170 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5171 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5172 csrc->mc_ki[csrc->mc_top],
5174 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5175 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5177 /* Add the node to the destination page.
5179 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5180 if (rc != MDB_SUCCESS)
5183 /* Delete the node from the source page.
5185 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5188 /* Adjust other cursors pointing to mp */
5189 MDB_cursor *m2, *m3;
5190 MDB_dbi dbi = csrc->mc_dbi;
5191 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5193 if (csrc->mc_flags & C_SUB)
5196 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5197 if (m2 == csrc) continue;
5198 if (csrc->mc_flags & C_SUB)
5199 m3 = &m2->mc_xcursor->mx_cursor;
5202 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5203 csrc->mc_ki[csrc->mc_top]) {
5204 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5205 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5210 /* Update the parent separators.
5212 if (csrc->mc_ki[csrc->mc_top] == 0) {
5213 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5214 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5215 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5217 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5218 key.mv_size = NODEKSZ(srcnode);
5219 key.mv_data = NODEKEY(srcnode);
5221 DPRINTF("update separator for source page %zu to [%s]",
5222 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5223 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5224 &key)) != MDB_SUCCESS)
5227 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5229 nullkey.mv_size = 0;
5230 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5231 assert(rc == MDB_SUCCESS);
5235 if (cdst->mc_ki[cdst->mc_top] == 0) {
5236 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5237 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5238 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5240 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5241 key.mv_size = NODEKSZ(srcnode);
5242 key.mv_data = NODEKEY(srcnode);
5244 DPRINTF("update separator for destination page %zu to [%s]",
5245 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5246 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5247 &key)) != MDB_SUCCESS)
5250 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5252 nullkey.mv_size = 0;
5253 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5254 assert(rc == MDB_SUCCESS);
5261 /** Merge one page into another.
5262 * The nodes from the page pointed to by \b csrc will
5263 * be copied to the page pointed to by \b cdst and then
5264 * the \b csrc page will be freed.
5265 * @param[in] csrc Cursor pointing to the source page.
5266 * @param[in] cdst Cursor pointing to the destination page.
5269 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5277 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5278 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5280 assert(csrc->mc_snum > 1); /* can't merge root page */
5281 assert(cdst->mc_snum > 1);
5283 /* Mark dst as dirty. */
5284 if ((rc = mdb_page_touch(cdst)))
5287 /* Move all nodes from src to dst.
5289 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5290 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5291 key.mv_size = csrc->mc_db->md_pad;
5292 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5293 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5294 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5295 if (rc != MDB_SUCCESS)
5297 key.mv_data = (char *)key.mv_data + key.mv_size;
5300 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5301 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5302 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5303 unsigned int snum = csrc->mc_snum;
5305 /* must find the lowest key below src */
5306 mdb_page_search_root(csrc, NULL, 0);
5307 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5308 key.mv_size = NODEKSZ(s2);
5309 key.mv_data = NODEKEY(s2);
5310 csrc->mc_snum = snum--;
5311 csrc->mc_top = snum;
5313 key.mv_size = srcnode->mn_ksize;
5314 key.mv_data = NODEKEY(srcnode);
5317 data.mv_size = NODEDSZ(srcnode);
5318 data.mv_data = NODEDATA(srcnode);
5319 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5320 if (rc != MDB_SUCCESS)
5325 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5326 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);
5328 /* Unlink the src page from parent and add to free list.
5330 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5331 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5333 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5337 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5338 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5339 csrc->mc_db->md_leaf_pages--;
5341 csrc->mc_db->md_branch_pages--;
5343 /* Adjust other cursors pointing to mp */
5344 MDB_cursor *m2, *m3;
5345 MDB_dbi dbi = csrc->mc_dbi;
5346 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5348 if (csrc->mc_flags & C_SUB)
5351 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5352 if (csrc->mc_flags & C_SUB)
5353 m3 = &m2->mc_xcursor->mx_cursor;
5356 if (m3 == csrc) continue;
5357 if (m3->mc_snum < csrc->mc_snum) continue;
5358 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5359 m3->mc_pg[csrc->mc_top] = mp;
5360 m3->mc_ki[csrc->mc_top] += nkeys;
5364 mdb_cursor_pop(csrc);
5366 return mdb_rebalance(csrc);
5369 /** Copy the contents of a cursor.
5370 * @param[in] csrc The cursor to copy from.
5371 * @param[out] cdst The cursor to copy to.
5374 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5378 cdst->mc_txn = csrc->mc_txn;
5379 cdst->mc_dbi = csrc->mc_dbi;
5380 cdst->mc_db = csrc->mc_db;
5381 cdst->mc_dbx = csrc->mc_dbx;
5382 cdst->mc_snum = csrc->mc_snum;
5383 cdst->mc_top = csrc->mc_top;
5384 cdst->mc_flags = csrc->mc_flags;
5386 for (i=0; i<csrc->mc_snum; i++) {
5387 cdst->mc_pg[i] = csrc->mc_pg[i];
5388 cdst->mc_ki[i] = csrc->mc_ki[i];
5392 /** Rebalance the tree after a delete operation.
5393 * @param[in] mc Cursor pointing to the page where rebalancing
5395 * @return 0 on success, non-zero on failure.
5398 mdb_rebalance(MDB_cursor *mc)
5408 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5409 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5410 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5411 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5415 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5418 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5419 DPRINTF("no need to rebalance page %zu, above fill threshold",
5425 if (mc->mc_snum < 2) {
5426 MDB_page *mp = mc->mc_pg[0];
5427 if (NUMKEYS(mp) == 0) {
5428 DPUTS("tree is completely empty");
5429 mc->mc_db->md_root = P_INVALID;
5430 mc->mc_db->md_depth = 0;
5431 mc->mc_db->md_leaf_pages = 0;
5432 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5436 /* Adjust other cursors pointing to mp */
5437 MDB_cursor *m2, *m3;
5438 MDB_dbi dbi = mc->mc_dbi;
5440 if (mc->mc_flags & C_SUB)
5443 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5444 if (m2 == mc) continue;
5445 if (mc->mc_flags & C_SUB)
5446 m3 = &m2->mc_xcursor->mx_cursor;
5449 if (m3->mc_snum < mc->mc_snum) continue;
5450 if (m3->mc_pg[0] == mp) {
5456 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5457 DPUTS("collapsing root page!");
5458 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5459 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5460 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5463 mc->mc_db->md_depth--;
5464 mc->mc_db->md_branch_pages--;
5466 /* Adjust other cursors pointing to mp */
5467 MDB_cursor *m2, *m3;
5468 MDB_dbi dbi = mc->mc_dbi;
5470 if (mc->mc_flags & C_SUB)
5473 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5474 if (m2 == mc) continue;
5475 if (mc->mc_flags & C_SUB)
5476 m3 = &m2->mc_xcursor->mx_cursor;
5479 if (m3->mc_snum < mc->mc_snum) continue;
5480 if (m3->mc_pg[0] == mp) {
5481 m3->mc_pg[0] = mc->mc_pg[0];
5486 DPUTS("root page doesn't need rebalancing");
5490 /* The parent (branch page) must have at least 2 pointers,
5491 * otherwise the tree is invalid.
5493 ptop = mc->mc_top-1;
5494 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5496 /* Leaf page fill factor is below the threshold.
5497 * Try to move keys from left or right neighbor, or
5498 * merge with a neighbor page.
5503 mdb_cursor_copy(mc, &mn);
5504 mn.mc_xcursor = NULL;
5506 if (mc->mc_ki[ptop] == 0) {
5507 /* We're the leftmost leaf in our parent.
5509 DPUTS("reading right neighbor");
5511 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5512 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5514 mn.mc_ki[mn.mc_top] = 0;
5515 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5517 /* There is at least one neighbor to the left.
5519 DPUTS("reading left neighbor");
5521 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5522 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5524 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5525 mc->mc_ki[mc->mc_top] = 0;
5528 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5529 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);
5531 /* If the neighbor page is above threshold and has at least two
5532 * keys, move one key from it.
5534 * Otherwise we should try to merge them.
5536 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5537 return mdb_node_move(&mn, mc);
5538 else { /* FIXME: if (has_enough_room()) */
5539 mc->mc_flags &= ~C_INITIALIZED;
5540 if (mc->mc_ki[ptop] == 0)
5541 return mdb_page_merge(&mn, mc);
5543 return mdb_page_merge(mc, &mn);
5547 /** Complete a delete operation started by #mdb_cursor_del(). */
5549 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5553 /* add overflow pages to free list */
5554 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5558 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5559 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5560 mc->mc_db->md_overflow_pages -= ovpages;
5561 for (i=0; i<ovpages; i++) {
5562 DPRINTF("freed ov page %zu", pg);
5563 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5567 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5568 mc->mc_db->md_entries--;
5569 rc = mdb_rebalance(mc);
5570 if (rc != MDB_SUCCESS)
5571 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5577 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5578 MDB_val *key, MDB_val *data)
5583 MDB_val rdata, *xdata;
5587 assert(key != NULL);
5589 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5591 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5594 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5598 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5602 mdb_cursor_init(&mc, txn, dbi, &mx);
5613 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5615 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5619 /** Split a page and insert a new node.
5620 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5621 * The cursor will be updated to point to the actual page and index where
5622 * the node got inserted after the split.
5623 * @param[in] newkey The key for the newly inserted node.
5624 * @param[in] newdata The data for the newly inserted node.
5625 * @param[in] newpgno The page number, if the new node is a branch node.
5626 * @return 0 on success, non-zero on failure.
5629 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5630 unsigned int nflags)
5633 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0;
5636 unsigned int i, j, split_indx, nkeys, pmax;
5638 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5640 MDB_page *mp, *rp, *pp;
5645 mp = mc->mc_pg[mc->mc_top];
5646 newindx = mc->mc_ki[mc->mc_top];
5648 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5649 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5650 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5652 /* Create a right sibling. */
5653 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5655 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5657 if (mc->mc_snum < 2) {
5658 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5660 /* shift current top to make room for new parent */
5661 mc->mc_pg[1] = mc->mc_pg[0];
5662 mc->mc_ki[1] = mc->mc_ki[0];
5665 mc->mc_db->md_root = pp->mp_pgno;
5666 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5667 mc->mc_db->md_depth++;
5670 /* Add left (implicit) pointer. */
5671 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5672 /* undo the pre-push */
5673 mc->mc_pg[0] = mc->mc_pg[1];
5674 mc->mc_ki[0] = mc->mc_ki[1];
5675 mc->mc_db->md_root = mp->mp_pgno;
5676 mc->mc_db->md_depth--;
5683 ptop = mc->mc_top-1;
5684 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5687 mdb_cursor_copy(mc, &mn);
5688 mn.mc_pg[mn.mc_top] = rp;
5689 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5691 if (nflags & MDB_APPEND) {
5692 mn.mc_ki[mn.mc_top] = 0;
5699 nkeys = NUMKEYS(mp);
5700 split_indx = nkeys / 2 + 1;
5705 unsigned int lsize, rsize, ksize;
5706 /* Move half of the keys to the right sibling */
5708 x = mc->mc_ki[mc->mc_top] - split_indx;
5709 ksize = mc->mc_db->md_pad;
5710 split = LEAF2KEY(mp, split_indx, ksize);
5711 rsize = (nkeys - split_indx) * ksize;
5712 lsize = (nkeys - split_indx) * sizeof(indx_t);
5713 mp->mp_lower -= lsize;
5714 rp->mp_lower += lsize;
5715 mp->mp_upper += rsize - lsize;
5716 rp->mp_upper -= rsize - lsize;
5717 sepkey.mv_size = ksize;
5718 if (newindx == split_indx) {
5719 sepkey.mv_data = newkey->mv_data;
5721 sepkey.mv_data = split;
5724 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5725 memcpy(rp->mp_ptrs, split, rsize);
5726 sepkey.mv_data = rp->mp_ptrs;
5727 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5728 memcpy(ins, newkey->mv_data, ksize);
5729 mp->mp_lower += sizeof(indx_t);
5730 mp->mp_upper -= ksize - sizeof(indx_t);
5733 memcpy(rp->mp_ptrs, split, x * ksize);
5734 ins = LEAF2KEY(rp, x, ksize);
5735 memcpy(ins, newkey->mv_data, ksize);
5736 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5737 rp->mp_lower += sizeof(indx_t);
5738 rp->mp_upper -= ksize - sizeof(indx_t);
5739 mc->mc_ki[mc->mc_top] = x;
5740 mc->mc_pg[mc->mc_top] = rp;
5745 /* For leaf pages, check the split point based on what
5746 * fits where, since otherwise add_node can fail.
5749 unsigned int psize, nsize;
5750 /* Maximum free space in an empty page */
5751 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5752 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5753 if (newindx < split_indx) {
5755 for (i=0; i<split_indx; i++) {
5756 node = NODEPTR(mp, i);
5757 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5758 if (F_ISSET(node->mn_flags, F_BIGDATA))
5759 psize += sizeof(pgno_t);
5761 psize += NODEDSZ(node);
5770 for (i=nkeys-1; i>=split_indx; i--) {
5771 node = NODEPTR(mp, i);
5772 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5773 if (F_ISSET(node->mn_flags, F_BIGDATA))
5774 psize += sizeof(pgno_t);
5776 psize += NODEDSZ(node);
5786 /* First find the separating key between the split pages.
5788 if (newindx == split_indx) {
5789 sepkey.mv_size = newkey->mv_size;
5790 sepkey.mv_data = newkey->mv_data;
5792 node = NODEPTR(mp, split_indx);
5793 sepkey.mv_size = node->mn_ksize;
5794 sepkey.mv_data = NODEKEY(node);
5798 DPRINTF("separator is [%s]", DKEY(&sepkey));
5800 /* Copy separator key to the parent.
5802 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5805 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5807 /* Right page might now have changed parent.
5808 * Check if left page also changed parent.
5810 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
5811 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
5812 mc->mc_pg[ptop] = mn.mc_pg[ptop];
5813 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
5817 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
5820 if (rc != MDB_SUCCESS) {
5823 if (nflags & MDB_APPEND) {
5824 mc->mc_pg[mc->mc_top] = rp;
5825 mc->mc_ki[mc->mc_top] = 0;
5826 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
5835 /* Move half of the keys to the right sibling. */
5837 /* grab a page to hold a temporary copy */
5838 copy = mdb_page_malloc(mc);
5842 copy->mp_pgno = mp->mp_pgno;
5843 copy->mp_flags = mp->mp_flags;
5844 copy->mp_lower = PAGEHDRSZ;
5845 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
5846 mc->mc_pg[mc->mc_top] = copy;
5847 for (i = j = 0; i <= nkeys; j++) {
5848 if (i == split_indx) {
5849 /* Insert in right sibling. */
5850 /* Reset insert index for right sibling. */
5851 j = (i == newindx && ins_new);
5852 mc->mc_pg[mc->mc_top] = rp;
5855 if (i == newindx && !ins_new) {
5856 /* Insert the original entry that caused the split. */
5857 rkey.mv_data = newkey->mv_data;
5858 rkey.mv_size = newkey->mv_size;
5867 /* Update page and index for the new key. */
5869 mc->mc_pg[mc->mc_top] = copy;
5870 mc->mc_ki[mc->mc_top] = j;
5871 } else if (i == nkeys) {
5874 node = NODEPTR(mp, i);
5875 rkey.mv_data = NODEKEY(node);
5876 rkey.mv_size = node->mn_ksize;
5878 xdata.mv_data = NODEDATA(node);
5879 xdata.mv_size = NODEDSZ(node);
5882 pgno = NODEPGNO(node);
5883 flags = node->mn_flags;
5888 if (!IS_LEAF(mp) && j == 0) {
5889 /* First branch index doesn't need key data. */
5893 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
5896 nkeys = NUMKEYS(copy);
5897 for (i=0; i<nkeys; i++)
5898 mp->mp_ptrs[i] = copy->mp_ptrs[i];
5899 mp->mp_lower = copy->mp_lower;
5900 mp->mp_upper = copy->mp_upper;
5901 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
5902 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
5904 /* reset back to original page */
5905 if (!newindx || (newindx < split_indx)) {
5906 mc->mc_pg[mc->mc_top] = mp;
5907 if (nflags & MDB_RESERVE) {
5908 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5909 if (!(node->mn_flags & F_BIGDATA))
5910 newdata->mv_data = NODEDATA(node);
5914 /* return tmp page to freelist */
5915 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
5916 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
5917 mc->mc_txn->mt_env->me_dpages = copy;
5920 /* Adjust other cursors pointing to mp */
5921 MDB_cursor *m2, *m3;
5922 MDB_dbi dbi = mc->mc_dbi;
5924 if (mc->mc_flags & C_SUB)
5927 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5928 if (m2 == mc) continue;
5929 if (mc->mc_flags & C_SUB)
5930 m3 = &m2->mc_xcursor->mx_cursor;
5933 if (!(m3->mc_flags & C_INITIALIZED))
5938 for (k=m3->mc_top; k>=0; k--) {
5939 m3->mc_ki[k+1] = m3->mc_ki[k];
5940 m3->mc_pg[k+1] = m3->mc_pg[k];
5942 m3->mc_ki[0] = mc->mc_ki[0];
5943 m3->mc_pg[0] = mc->mc_pg[0];
5947 if (m3->mc_pg[mc->mc_top] == mp) {
5948 if (m3->mc_ki[m3->mc_top] >= split_indx) {
5949 m3->mc_pg[m3->mc_top] = rp;
5950 m3->mc_ki[m3->mc_top] -= split_indx;
5959 mdb_put(MDB_txn *txn, MDB_dbi dbi,
5960 MDB_val *key, MDB_val *data, unsigned int flags)
5965 assert(key != NULL);
5966 assert(data != NULL);
5968 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5971 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5975 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5979 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
5982 mdb_cursor_init(&mc, txn, dbi, &mx);
5983 return mdb_cursor_put(&mc, key, data, flags);
5986 /** Only a subset of the @ref mdb_env flags can be changed
5987 * at runtime. Changing other flags requires closing the environment
5988 * and re-opening it with the new flags.
5990 #define CHANGEABLE (MDB_NOSYNC)
5992 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
5994 if ((flag & CHANGEABLE) != flag)
5997 env->me_flags |= flag;
5999 env->me_flags &= ~flag;
6004 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6009 *arg = env->me_flags;
6014 mdb_env_get_path(MDB_env *env, const char **arg)
6019 *arg = env->me_path;
6023 /** Common code for #mdb_stat() and #mdb_env_stat().
6024 * @param[in] env the environment to operate in.
6025 * @param[in] db the #MDB_db record containing the stats to return.
6026 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6027 * @return 0, this function always succeeds.
6030 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6032 arg->ms_psize = env->me_psize;
6033 arg->ms_depth = db->md_depth;
6034 arg->ms_branch_pages = db->md_branch_pages;
6035 arg->ms_leaf_pages = db->md_leaf_pages;
6036 arg->ms_overflow_pages = db->md_overflow_pages;
6037 arg->ms_entries = db->md_entries;
6042 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6046 if (env == NULL || arg == NULL)
6049 mdb_env_read_meta(env, &toggle);
6051 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6054 /** Set the default comparison functions for a database.
6055 * Called immediately after a database is opened to set the defaults.
6056 * The user can then override them with #mdb_set_compare() or
6057 * #mdb_set_dupsort().
6058 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6059 * @param[in] dbi A database handle returned by #mdb_open()
6062 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6064 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
6065 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memnr;
6066 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
6067 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_cint;
6069 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memn;
6071 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6072 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
6073 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
6074 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_int;
6076 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_cint;
6077 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
6078 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memnr;
6080 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memn;
6083 txn->mt_dbxs[dbi].md_dcmp = NULL;
6087 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6092 int rc, dbflag, exact;
6095 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6096 mdb_default_cmp(txn, FREE_DBI);
6102 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6103 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6104 mdb_default_cmp(txn, MAIN_DBI);
6108 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6109 mdb_default_cmp(txn, MAIN_DBI);
6112 /* Is the DB already open? */
6114 for (i=2; i<txn->mt_numdbs; i++) {
6115 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6116 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6122 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
6125 /* Find the DB info */
6129 key.mv_data = (void *)name;
6130 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6131 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6132 if (rc == MDB_SUCCESS) {
6133 /* make sure this is actually a DB */
6134 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6135 if (!(node->mn_flags & F_SUBDATA))
6137 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6138 /* Create if requested */
6140 data.mv_size = sizeof(MDB_db);
6141 data.mv_data = &dummy;
6142 memset(&dummy, 0, sizeof(dummy));
6143 dummy.md_root = P_INVALID;
6144 dummy.md_flags = flags & 0xffff;
6145 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6149 /* OK, got info, add to table */
6150 if (rc == MDB_SUCCESS) {
6151 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
6152 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
6153 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
6154 txn->mt_dbflags[txn->mt_numdbs] = dbflag;
6155 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
6156 *dbi = txn->mt_numdbs;
6157 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
6158 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
6159 mdb_default_cmp(txn, txn->mt_numdbs);
6166 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6168 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6171 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6174 void mdb_close(MDB_env *env, MDB_dbi dbi)
6177 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6179 ptr = env->me_dbxs[dbi].md_name.mv_data;
6180 env->me_dbxs[dbi].md_name.mv_data = NULL;
6181 env->me_dbxs[dbi].md_name.mv_size = 0;
6185 /** Add all the DB's pages to the free list.
6186 * @param[in] mc Cursor on the DB to free.
6187 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6188 * @return 0 on success, non-zero on failure.
6191 mdb_drop0(MDB_cursor *mc, int subs)
6195 rc = mdb_page_search(mc, NULL, 0);
6196 if (rc == MDB_SUCCESS) {
6201 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6202 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6205 mdb_cursor_copy(mc, &mx);
6206 while (mc->mc_snum > 0) {
6207 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6208 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6209 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6210 if (ni->mn_flags & F_SUBDATA) {
6211 mdb_xcursor_init1(mc, ni);
6212 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6218 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6220 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6223 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6228 rc = mdb_cursor_sibling(mc, 1);
6230 /* no more siblings, go back to beginning
6231 * of previous level. (stack was already popped
6232 * by mdb_cursor_sibling)
6234 for (i=1; i<mc->mc_top; i++)
6235 mc->mc_pg[i] = mx.mc_pg[i];
6239 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6240 mc->mc_db->md_root);
6245 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6250 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6253 rc = mdb_cursor_open(txn, dbi, &mc);
6257 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6261 /* Can't delete the main DB */
6262 if (del && dbi > MAIN_DBI) {
6263 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6265 mdb_close(txn->mt_env, dbi);
6267 txn->mt_dbflags[dbi] |= DB_DIRTY;
6268 txn->mt_dbs[dbi].md_depth = 0;
6269 txn->mt_dbs[dbi].md_branch_pages = 0;
6270 txn->mt_dbs[dbi].md_leaf_pages = 0;
6271 txn->mt_dbs[dbi].md_overflow_pages = 0;
6272 txn->mt_dbs[dbi].md_entries = 0;
6273 txn->mt_dbs[dbi].md_root = P_INVALID;
6276 mdb_cursor_close(mc);
6280 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6282 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6285 txn->mt_dbxs[dbi].md_cmp = cmp;
6289 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6291 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6294 txn->mt_dbxs[dbi].md_dcmp = cmp;
6298 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6300 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6303 txn->mt_dbxs[dbi].md_rel = rel;
6307 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6309 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6312 txn->mt_dbxs[dbi].md_relctx = ctx;