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__)
210 /** A flag for opening a file and requesting synchronous data writes.
211 * This is only used when writing a meta page. It's not strictly needed;
212 * we could just do a normal write and then immediately perform a flush.
213 * But if this flag is available it saves us an extra system call.
215 * @note If O_DSYNC is undefined but exists in /usr/include,
216 * preferably set some compiler flag to get the definition.
217 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
220 # define MDB_DSYNC O_DSYNC
224 /** Function for flushing the data of a file. Define this to fsync
225 * if fdatasync() is not supported.
227 #ifndef MDB_FDATASYNC
228 # define MDB_FDATASYNC fdatasync
231 /** A page number in the database.
232 * Note that 64 bit page numbers are overkill, since pages themselves
233 * already represent 12-13 bits of addressable memory, and the OS will
234 * always limit applications to a maximum of 63 bits of address space.
236 * @note In the #MDB_node structure, we only store 48 bits of this value,
237 * which thus limits us to only 60 bits of addressable data.
241 /** A transaction ID.
242 * See struct MDB_txn.mt_txnid for details.
246 /** @defgroup debug Debug Macros
250 /** Enable debug output.
251 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
252 * read from and written to the database (used for free space management).
257 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
258 # define DPRINTF (void) /* Vararg macros may be unsupported */
260 /** Print a debug message with printf formatting. */
261 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
262 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)
264 # define DPRINTF(fmt, ...) ((void) 0)
266 /** Print a debug string.
267 * The string is printed literally, with no format processing.
269 #define DPUTS(arg) DPRINTF("%s", arg)
272 /** A default memory page size.
273 * The actual size is platform-dependent, but we use this for
274 * boot-strapping. We probably should not be using this any more.
275 * The #GET_PAGESIZE() macro is used to get the actual size.
277 * Note that we don't currently support Huge pages. On Linux,
278 * regular data files cannot use Huge pages, and in general
279 * Huge pages aren't actually pageable. We rely on the OS
280 * demand-pager to read our data and page it out when memory
281 * pressure from other processes is high. So until OSs have
282 * actual paging support for Huge pages, they're not viable.
284 #define MDB_PAGESIZE 4096
286 /** The minimum number of keys required in a database page.
287 * Setting this to a larger value will place a smaller bound on the
288 * maximum size of a data item. Data items larger than this size will
289 * be pushed into overflow pages instead of being stored directly in
290 * the B-tree node. This value used to default to 4. With a page size
291 * of 4096 bytes that meant that any item larger than 1024 bytes would
292 * go into an overflow page. That also meant that on average 2-3KB of
293 * each overflow page was wasted space. The value cannot be lower than
294 * 2 because then there would no longer be a tree structure. With this
295 * value, items larger than 2KB will go into overflow pages, and on
296 * average only 1KB will be wasted.
298 #define MDB_MINKEYS 2
300 /** A stamp that identifies a file as an MDB file.
301 * There's nothing special about this value other than that it is easily
302 * recognizable, and it will reflect any byte order mismatches.
304 #define MDB_MAGIC 0xBEEFC0DE
306 /** The version number for a database's file format. */
307 #define MDB_VERSION 1
309 /** The maximum size of a key in the database.
310 * While data items have essentially unbounded size, we require that
311 * keys all fit onto a regular page. This limit could be raised a bit
312 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
314 #define MAXKEYSIZE 511
319 * This is used for printing a hex dump of a key's contents.
321 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
322 /** Display a key in hex.
324 * Invoke a function to display a key in hex.
326 #define DKEY(x) mdb_dkey(x, kbuf)
328 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
332 /** @defgroup lazylock Lazy Locking
333 * Macros for locks that aren't actually needed.
334 * The DB view is always consistent because all writes are wrapped in
335 * the wmutex. Finer-grained locks aren't necessary.
339 /** Use lazy locking. I.e., don't lock these accesses at all. */
343 /** Grab the reader lock */
344 #define LAZY_MUTEX_LOCK(x)
345 /** Release the reader lock */
346 #define LAZY_MUTEX_UNLOCK(x)
347 /** Release the DB table reader/writer lock */
348 #define LAZY_RWLOCK_UNLOCK(x)
349 /** Grab the DB table write lock */
350 #define LAZY_RWLOCK_WRLOCK(x)
351 /** Grab the DB table read lock */
352 #define LAZY_RWLOCK_RDLOCK(x)
353 /** Declare the DB table rwlock. Should not be followed by ';'. */
354 #define LAZY_RWLOCK_DEF(x)
355 /** Initialize the DB table rwlock */
356 #define LAZY_RWLOCK_INIT(x,y)
357 /** Destroy the DB table rwlock */
358 #define LAZY_RWLOCK_DESTROY(x)
360 #define LAZY_MUTEX_LOCK(x) pthread_mutex_lock(x)
361 #define LAZY_MUTEX_UNLOCK(x) pthread_mutex_unlock(x)
362 #define LAZY_RWLOCK_UNLOCK(x) pthread_rwlock_unlock(x)
363 #define LAZY_RWLOCK_WRLOCK(x) pthread_rwlock_wrlock(x)
364 #define LAZY_RWLOCK_RDLOCK(x) pthread_rwlock_rdlock(x)
365 #define LAZY_RWLOCK_DEF(x) pthread_rwlock_t x;
366 #define LAZY_RWLOCK_INIT(x,y) pthread_rwlock_init(x,y)
367 #define LAZY_RWLOCK_DESTROY(x) pthread_rwlock_destroy(x)
371 /** An invalid page number.
372 * Mainly used to denote an empty tree.
374 #define P_INVALID (~0UL)
376 /** Test if a flag \b f is set in a flag word \b w. */
377 #define F_ISSET(w, f) (((w) & (f)) == (f))
379 /** Used for offsets within a single page.
380 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
383 typedef uint16_t indx_t;
385 /** Default size of memory map.
386 * This is certainly too small for any actual applications. Apps should always set
387 * the size explicitly using #mdb_env_set_mapsize().
389 #define DEFAULT_MAPSIZE 1048576
391 /** @defgroup readers Reader Lock Table
392 * Readers don't acquire any locks for their data access. Instead, they
393 * simply record their transaction ID in the reader table. The reader
394 * mutex is needed just to find an empty slot in the reader table. The
395 * slot's address is saved in thread-specific data so that subsequent read
396 * transactions started by the same thread need no further locking to proceed.
398 * Since the database uses multi-version concurrency control, readers don't
399 * actually need any locking. This table is used to keep track of which
400 * readers are using data from which old transactions, so that we'll know
401 * when a particular old transaction is no longer in use. Old transactions
402 * that have discarded any data pages can then have those pages reclaimed
403 * for use by a later write transaction.
405 * The lock table is constructed such that reader slots are aligned with the
406 * processor's cache line size. Any slot is only ever used by one thread.
407 * This alignment guarantees that there will be no contention or cache
408 * thrashing as threads update their own slot info, and also eliminates
409 * any need for locking when accessing a slot.
411 * A writer thread will scan every slot in the table to determine the oldest
412 * outstanding reader transaction. Any freed pages older than this will be
413 * reclaimed by the writer. The writer doesn't use any locks when scanning
414 * this table. This means that there's no guarantee that the writer will
415 * see the most up-to-date reader info, but that's not required for correct
416 * operation - all we need is to know the upper bound on the oldest reader,
417 * we don't care at all about the newest reader. So the only consequence of
418 * reading stale information here is that old pages might hang around a
419 * while longer before being reclaimed. That's actually good anyway, because
420 * the longer we delay reclaiming old pages, the more likely it is that a
421 * string of contiguous pages can be found after coalescing old pages from
422 * many old transactions together.
424 * @todo We don't actually do such coalescing yet, we grab pages from one
425 * old transaction at a time.
428 /** Number of slots in the reader table.
429 * This value was chosen somewhat arbitrarily. 126 readers plus a
430 * couple mutexes fit exactly into 8KB on my development machine.
431 * Applications should set the table size using #mdb_env_set_maxreaders().
433 #define DEFAULT_READERS 126
435 /** The size of a CPU cache line in bytes. We want our lock structures
436 * aligned to this size to avoid false cache line sharing in the
438 * This value works for most CPUs. For Itanium this should be 128.
444 /** The information we store in a single slot of the reader table.
445 * In addition to a transaction ID, we also record the process and
446 * thread ID that owns a slot, so that we can detect stale information,
447 * e.g. threads or processes that went away without cleaning up.
448 * @note We currently don't check for stale records. We simply re-init
449 * the table when we know that we're the only process opening the
452 typedef struct MDB_rxbody {
453 /** The current Transaction ID when this transaction began.
454 * Multiple readers that start at the same time will probably have the
455 * same ID here. Again, it's not important to exclude them from
456 * anything; all we need to know is which version of the DB they
457 * started from so we can avoid overwriting any data used in that
458 * particular version.
461 /** The process ID of the process owning this reader txn. */
463 /** The thread ID of the thread owning this txn. */
467 /** The actual reader record, with cacheline padding. */
468 typedef struct MDB_reader {
471 /** shorthand for mrb_txnid */
472 #define mr_txnid mru.mrx.mrb_txnid
473 #define mr_pid mru.mrx.mrb_pid
474 #define mr_tid mru.mrx.mrb_tid
475 /** cache line alignment */
476 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
480 /** The header for the reader table.
481 * The table resides in a memory-mapped file. (This is a different file
482 * than is used for the main database.)
484 * For POSIX the actual mutexes reside in the shared memory of this
485 * mapped file. On Windows, mutexes are named objects allocated by the
486 * kernel; we store the mutex names in this mapped file so that other
487 * processes can grab them. This same approach is also used on
488 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
489 * process-shared POSIX mutexes. For these cases where a named object
490 * is used, the object name is derived from a 64 bit FNV hash of the
491 * environment pathname. As such, naming collisions are extremely
492 * unlikely. If a collision occurs, the results are unpredictable.
494 typedef struct MDB_txbody {
495 /** Stamp identifying this as an MDB file. It must be set
498 /** Version number of this lock file. Must be set to #MDB_VERSION. */
499 uint32_t mtb_version;
500 #if defined(_WIN32) || defined(__APPLE__)
501 char mtb_rmname[MNAME_LEN];
503 /** Mutex protecting access to this table.
504 * This is the reader lock that #LOCK_MUTEX_R acquires.
506 pthread_mutex_t mtb_mutex;
508 /** The ID of the last transaction committed to the database.
509 * This is recorded here only for convenience; the value can always
510 * be determined by reading the main database meta pages.
513 /** The number of slots that have been used in the reader table.
514 * This always records the maximum count, it is not decremented
515 * when readers release their slots.
517 unsigned mtb_numreaders;
518 /** The ID of the most recent meta page in the database.
519 * This is recorded here only for convenience; the value can always
520 * be determined by reading the main database meta pages.
522 uint32_t mtb_me_toggle;
525 /** The actual reader table definition. */
526 typedef struct MDB_txninfo {
529 #define mti_magic mt1.mtb.mtb_magic
530 #define mti_version mt1.mtb.mtb_version
531 #define mti_mutex mt1.mtb.mtb_mutex
532 #define mti_rmname mt1.mtb.mtb_rmname
533 #define mti_txnid mt1.mtb.mtb_txnid
534 #define mti_numreaders mt1.mtb.mtb_numreaders
535 #define mti_me_toggle mt1.mtb.mtb_me_toggle
536 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
539 #if defined(_WIN32) || defined(__APPLE__)
540 char mt2_wmname[MNAME_LEN];
541 #define mti_wmname mt2.mt2_wmname
543 pthread_mutex_t mt2_wmutex;
544 #define mti_wmutex mt2.mt2_wmutex
546 char pad[(sizeof(pthread_mutex_t)+CACHELINE-1) & ~(CACHELINE-1)];
548 MDB_reader mti_readers[1];
552 /** Common header for all page types.
553 * Overflow records occupy a number of contiguous pages with no
554 * headers on any page after the first.
556 typedef struct MDB_page {
557 #define mp_pgno mp_p.p_pgno
558 #define mp_next mp_p.p_next
560 pgno_t p_pgno; /**< page number */
561 void * p_next; /**< for in-memory list of freed structs */
564 /** @defgroup mdb_page Page Flags
566 * Flags for the page headers.
569 #define P_BRANCH 0x01 /**< branch page */
570 #define P_LEAF 0x02 /**< leaf page */
571 #define P_OVERFLOW 0x04 /**< overflow page */
572 #define P_META 0x08 /**< meta page */
573 #define P_DIRTY 0x10 /**< dirty page */
574 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
575 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
577 uint16_t mp_flags; /**< @ref mdb_page */
578 #define mp_lower mp_pb.pb.pb_lower
579 #define mp_upper mp_pb.pb.pb_upper
580 #define mp_pages mp_pb.pb_pages
583 indx_t pb_lower; /**< lower bound of free space */
584 indx_t pb_upper; /**< upper bound of free space */
586 uint32_t pb_pages; /**< number of overflow pages */
588 indx_t mp_ptrs[1]; /**< dynamic size */
591 /** Size of the page header, excluding dynamic data at the end */
592 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
594 /** Address of first usable data byte in a page, after the header */
595 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
597 /** Number of nodes on a page */
598 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
600 /** The amount of space remaining in the page */
601 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
603 /** The percentage of space used in the page, in tenths of a percent. */
604 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
605 ((env)->me_psize - PAGEHDRSZ))
606 /** The minimum page fill factor, in tenths of a percent.
607 * Pages emptier than this are candidates for merging.
609 #define FILL_THRESHOLD 250
611 /** Test if a page is a leaf page */
612 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
613 /** Test if a page is a LEAF2 page */
614 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
615 /** Test if a page is a branch page */
616 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
617 /** Test if a page is an overflow page */
618 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
619 /** Test if a page is a sub page */
620 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
622 /** The number of overflow pages needed to store the given size. */
623 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
625 /** Header for a single key/data pair within a page.
626 * We guarantee 2-byte alignment for nodes.
628 typedef struct MDB_node {
629 /** lo and hi are used for data size on leaf nodes and for
630 * child pgno on branch nodes. On 64 bit platforms, flags
631 * is also used for pgno. (Branch nodes have no flags).
632 * They are in host byte order in case that lets some
633 * accesses be optimized into a 32-bit word access.
635 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
636 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
637 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
638 /** @defgroup mdb_node Node Flags
640 * Flags for node headers.
643 #define F_BIGDATA 0x01 /**< data put on overflow page */
644 #define F_SUBDATA 0x02 /**< data is a sub-database */
645 #define F_DUPDATA 0x04 /**< data has duplicates */
647 /** valid flags for #mdb_node_add() */
648 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
651 unsigned short mn_flags; /**< @ref mdb_node */
652 unsigned short mn_ksize; /**< key size */
653 char mn_data[1]; /**< key and data are appended here */
656 /** Size of the node header, excluding dynamic data at the end */
657 #define NODESIZE offsetof(MDB_node, mn_data)
659 /** Bit position of top word in page number, for shifting mn_flags */
660 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
662 /** Size of a node in a branch page with a given key.
663 * This is just the node header plus the key, there is no data.
665 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
667 /** Size of a node in a leaf page with a given key and data.
668 * This is node header plus key plus data size.
670 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
672 /** Address of node \b i in page \b p */
673 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
675 /** Address of the key for the node */
676 #define NODEKEY(node) (void *)((node)->mn_data)
678 /** Address of the data for a node */
679 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
681 /** Get the page number pointed to by a branch node */
682 #define NODEPGNO(node) \
683 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
684 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
685 /** Set the page number in a branch node */
686 #define SETPGNO(node,pgno) do { \
687 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
688 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
690 /** Get the size of the data in a leaf node */
691 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
692 /** Set the size of the data for a leaf node */
693 #define SETDSZ(node,size) do { \
694 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
695 /** The size of a key in a node */
696 #define NODEKSZ(node) ((node)->mn_ksize)
698 /** Copy a page number from src to dst */
700 #define COPY_PGNO(dst,src) dst = src
702 #if SIZE_MAX > 4294967295UL
703 #define COPY_PGNO(dst,src) do { \
704 unsigned short *s, *d; \
705 s = (unsigned short *)&(src); \
706 d = (unsigned short *)&(dst); \
713 #define COPY_PGNO(dst,src) do { \
714 unsigned short *s, *d; \
715 s = (unsigned short *)&(src); \
716 d = (unsigned short *)&(dst); \
722 /** The address of a key in a LEAF2 page.
723 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
724 * There are no node headers, keys are stored contiguously.
726 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
728 /** Set the \b node's key into \b key, if requested. */
729 #define MDB_SET_KEY(node, key) { if ((key) != NULL) { \
730 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
732 /** Information about a single database in the environment. */
733 typedef struct MDB_db {
734 uint32_t md_pad; /**< also ksize for LEAF2 pages */
735 uint16_t md_flags; /**< @ref mdb_open */
736 uint16_t md_depth; /**< depth of this tree */
737 pgno_t md_branch_pages; /**< number of internal pages */
738 pgno_t md_leaf_pages; /**< number of leaf pages */
739 pgno_t md_overflow_pages; /**< number of overflow pages */
740 size_t md_entries; /**< number of data items */
741 pgno_t md_root; /**< the root page of this tree */
744 /** Handle for the DB used to track free pages. */
746 /** Handle for the default DB. */
749 /** Meta page content. */
750 typedef struct MDB_meta {
751 /** Stamp identifying this as an MDB file. It must be set
754 /** Version number of this lock file. Must be set to #MDB_VERSION. */
756 void *mm_address; /**< address for fixed mapping */
757 size_t mm_mapsize; /**< size of mmap region */
758 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
759 /** The size of pages used in this DB */
760 #define mm_psize mm_dbs[0].md_pad
761 /** Any persistent environment flags. @ref mdb_env */
762 #define mm_flags mm_dbs[0].md_flags
763 pgno_t mm_last_pg; /**< last used page in file */
764 txnid_t mm_txnid; /**< txnid that committed this page */
767 /** Auxiliary DB info.
768 * The information here is mostly static/read-only. There is
769 * only a single copy of this record in the environment.
771 typedef struct MDB_dbx {
772 MDB_val md_name; /**< name of the database */
773 MDB_cmp_func *md_cmp; /**< function for comparing keys */
774 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
775 MDB_rel_func *md_rel; /**< user relocate function */
776 void *md_relctx; /**< user-provided context for md_rel */
779 /** A database transaction.
780 * Every operation requires a transaction handle.
783 MDB_txn *mt_parent; /**< parent of a nested txn */
784 MDB_txn *mt_child; /**< nested txn under this txn */
785 pgno_t mt_next_pgno; /**< next unallocated page */
786 /** The ID of this transaction. IDs are integers incrementing from 1.
787 * Only committed write transactions increment the ID. If a transaction
788 * aborts, the ID may be re-used by the next writer.
791 MDB_env *mt_env; /**< the DB environment */
792 /** The list of pages that became unused during this transaction.
796 ID2L dirty_list; /**< modified pages */
797 MDB_reader *reader; /**< this thread's slot in the reader table */
799 /** Array of records for each DB known in the environment. */
801 /** Array of MDB_db records for each known DB */
803 /** @defgroup mt_dbflag Transaction DB Flags
807 #define DB_DIRTY 0x01 /**< DB was written in this txn */
808 #define DB_STALE 0x02 /**< DB record is older than txnID */
810 /** Array of cursors for each DB */
811 MDB_cursor **mt_cursors;
812 /** Array of flags for each DB */
813 unsigned char *mt_dbflags;
814 /** Number of DB records in use. This number only ever increments;
815 * we don't decrement it when individual DB handles are closed.
819 /** @defgroup mdb_txn Transaction Flags
823 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
824 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
826 unsigned int mt_flags; /**< @ref mdb_txn */
827 /** Tracks which of the two meta pages was used at the start
828 * of this transaction.
830 unsigned int mt_toggle;
833 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
834 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
835 * raise this on a 64 bit machine.
837 #define CURSOR_STACK 32
841 /** Cursors are used for all DB operations */
843 /** Next cursor on this DB in this txn */
845 /** Original cursor if this is a shadow */
847 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
848 struct MDB_xcursor *mc_xcursor;
849 /** The transaction that owns this cursor */
851 /** The database handle this cursor operates on */
853 /** The database record for this cursor */
855 /** The database auxiliary record for this cursor */
857 /** The @ref mt_dbflag for this database */
858 unsigned char *mc_dbflag;
859 unsigned short mc_snum; /**< number of pushed pages */
860 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
861 /** @defgroup mdb_cursor Cursor Flags
863 * Cursor state flags.
866 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
867 #define C_EOF 0x02 /**< No more data */
868 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
869 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
870 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
872 unsigned int mc_flags; /**< @ref mdb_cursor */
873 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
874 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
877 /** Context for sorted-dup records.
878 * We could have gone to a fully recursive design, with arbitrarily
879 * deep nesting of sub-databases. But for now we only handle these
880 * levels - main DB, optional sub-DB, sorted-duplicate DB.
882 typedef struct MDB_xcursor {
883 /** A sub-cursor for traversing the Dup DB */
884 MDB_cursor mx_cursor;
885 /** The database record for this Dup DB */
887 /** The auxiliary DB record for this Dup DB */
889 /** The @ref mt_dbflag for this Dup DB */
890 unsigned char mx_dbflag;
893 /** A set of pages freed by an earlier transaction. */
894 typedef struct MDB_oldpages {
895 /** Usually we only read one record from the FREEDB at a time, but
896 * in case we read more, this will chain them together.
898 struct MDB_oldpages *mo_next;
899 /** The ID of the transaction in which these pages were freed. */
901 /** An #IDL of the pages */
902 pgno_t mo_pages[1]; /* dynamic */
905 /** The database environment. */
907 HANDLE me_fd; /**< The main data file */
908 HANDLE me_lfd; /**< The lock file */
909 HANDLE me_mfd; /**< just for writing the meta pages */
910 /** Failed to update the meta page. Probably an I/O error. */
911 #define MDB_FATAL_ERROR 0x80000000U
912 uint32_t me_flags; /**< @ref mdb_env */
913 uint32_t me_extrapad; /**< unused for now */
914 unsigned int me_maxreaders; /**< size of the reader table */
915 MDB_dbi me_numdbs; /**< number of DBs opened */
916 MDB_dbi me_maxdbs; /**< size of the DB table */
917 char *me_path; /**< path to the DB files */
918 char *me_map; /**< the memory map of the data file */
919 MDB_txninfo *me_txns; /**< the memory map of the lock file */
920 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
921 MDB_txn *me_txn; /**< current write transaction */
922 size_t me_mapsize; /**< size of the data memory map */
923 off_t me_size; /**< current file size */
924 pgno_t me_maxpg; /**< me_mapsize / me_psize */
925 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
926 unsigned int me_db_toggle; /**< which DB table is current */
927 txnid_t me_wtxnid; /**< ID of last txn we committed */
928 MDB_dbx *me_dbxs; /**< array of static DB info */
929 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
930 MDB_oldpages *me_pghead; /**< list of old page records */
931 pthread_key_t me_txkey; /**< thread-key for readers */
932 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
933 /** IDL of pages that became unused in a write txn */
935 /** ID2L of pages that were written during a write txn */
936 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
937 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
938 LAZY_RWLOCK_DEF(me_dblock)
940 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
944 sem_t *me_rmutex; /* Apple doesn't support shared mutexes */
948 /** max number of pages to commit in one writev() call */
949 #define MDB_COMMIT_PAGES 64
950 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
951 #undef MDB_COMMIT_PAGES
952 #define MDB_COMMIT_PAGES IOV_MAX
955 static MDB_page *mdb_page_alloc(MDB_cursor *mc, int num);
956 static MDB_page *mdb_page_new(MDB_cursor *mc, uint32_t flags, int num);
957 static int mdb_page_touch(MDB_cursor *mc);
959 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
960 static int mdb_page_search_root(MDB_cursor *mc,
961 MDB_val *key, int modify);
962 static int mdb_page_search(MDB_cursor *mc,
963 MDB_val *key, int modify);
964 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
965 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
966 pgno_t newpgno, unsigned int nflags);
968 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
969 static int mdb_env_read_meta(MDB_env *env, int *which);
970 static int mdb_env_write_meta(MDB_txn *txn);
972 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
973 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
974 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
975 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
976 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
977 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
978 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
979 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
980 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
982 static int mdb_rebalance(MDB_cursor *mc);
983 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
985 static void mdb_cursor_pop(MDB_cursor *mc);
986 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
988 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
989 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
990 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
991 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
992 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
994 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
995 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
997 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
998 static void mdb_xcursor_init0(MDB_cursor *mc);
999 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1001 static int mdb_drop0(MDB_cursor *mc, int subs);
1002 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1005 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1009 static SECURITY_DESCRIPTOR mdb_null_sd;
1010 static SECURITY_ATTRIBUTES mdb_all_sa;
1011 static int mdb_sec_inited;
1014 /** Return the library version info. */
1016 mdb_version(int *major, int *minor, int *patch)
1018 if (major) *major = MDB_VERSION_MAJOR;
1019 if (minor) *minor = MDB_VERSION_MINOR;
1020 if (patch) *patch = MDB_VERSION_PATCH;
1021 return MDB_VERSION_STRING;
1024 /** Table of descriptions for MDB @ref errors */
1025 static char *const mdb_errstr[] = {
1026 "MDB_KEYEXIST: Key/data pair already exists",
1027 "MDB_NOTFOUND: No matching key/data pair found",
1028 "MDB_PAGE_NOTFOUND: Requested page not found",
1029 "MDB_CORRUPTED: Located page was wrong type",
1030 "MDB_PANIC: Update of meta page failed",
1031 "MDB_VERSION_MISMATCH: Database environment version mismatch"
1035 mdb_strerror(int err)
1038 return ("Successful return: 0");
1040 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
1041 return mdb_errstr[err - MDB_KEYEXIST];
1043 return strerror(err);
1047 /** Display a key in hexadecimal and return the address of the result.
1048 * @param[in] key the key to display
1049 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1050 * @return The key in hexadecimal form.
1053 mdb_dkey(MDB_val *key, char *buf)
1056 unsigned char *c = key->mv_data;
1058 if (key->mv_size > MAXKEYSIZE)
1059 return "MAXKEYSIZE";
1060 /* may want to make this a dynamic check: if the key is mostly
1061 * printable characters, print it as-is instead of converting to hex.
1065 for (i=0; i<key->mv_size; i++)
1066 ptr += sprintf(ptr, "%02x", *c++);
1068 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1073 /** Display all the keys in the page. */
1075 mdb_page_keys(MDB_page *mp)
1078 unsigned int i, nkeys;
1082 nkeys = NUMKEYS(mp);
1083 DPRINTF("numkeys %d", nkeys);
1084 for (i=0; i<nkeys; i++) {
1085 node = NODEPTR(mp, i);
1086 key.mv_size = node->mn_ksize;
1087 key.mv_data = node->mn_data;
1088 DPRINTF("key %d: %s", i, DKEY(&key));
1094 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1096 return txn->mt_dbxs[dbi].md_cmp(a, b);
1100 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1102 if (txn->mt_dbxs[dbi].md_dcmp)
1103 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1105 return EINVAL; /* too bad you can't distinguish this from a valid result */
1108 /** Allocate a single page.
1109 * Re-use old malloc'd pages first, otherwise just malloc.
1112 mdb_page_malloc(MDB_cursor *mc) {
1114 size_t sz = mc->mc_txn->mt_env->me_psize;
1115 if (mc->mc_txn->mt_env->me_dpages) {
1116 ret = mc->mc_txn->mt_env->me_dpages;
1117 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1118 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1119 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1122 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1127 /** Allocate pages for writing.
1128 * If there are free pages available from older transactions, they
1129 * will be re-used first. Otherwise a new page will be allocated.
1130 * @param[in] mc cursor A cursor handle identifying the transaction and
1131 * database for which we are allocating.
1132 * @param[in] num the number of pages to allocate.
1133 * @return Address of the allocated page(s). Requests for multiple pages
1134 * will always be satisfied by a single contiguous chunk of memory.
1137 mdb_page_alloc(MDB_cursor *mc, int num)
1139 MDB_txn *txn = mc->mc_txn;
1141 pgno_t pgno = P_INVALID;
1144 if (txn->mt_txnid > 2) {
1146 if (!txn->mt_env->me_pghead && mc->mc_dbi != FREE_DBI &&
1147 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1148 /* See if there's anything in the free DB */
1151 txnid_t *kptr, oldest;
1153 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1154 mdb_page_search(&m2, NULL, 0);
1155 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1156 kptr = (txnid_t *)NODEKEY(leaf);
1160 oldest = txn->mt_txnid - 1;
1161 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1162 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1163 if (mr && mr < oldest)
1168 if (oldest > *kptr) {
1169 /* It's usable, grab it.
1175 mdb_node_read(txn, leaf, &data);
1176 idl = (ID *) data.mv_data;
1177 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1178 mop->mo_next = txn->mt_env->me_pghead;
1179 mop->mo_txnid = *kptr;
1180 txn->mt_env->me_pghead = mop;
1181 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1186 DPRINTF("IDL read txn %zu root %zu num %zu",
1187 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1188 for (i=0; i<idl[0]; i++) {
1189 DPRINTF("IDL %zu", idl[i+1]);
1193 /* drop this IDL from the DB */
1194 m2.mc_ki[m2.mc_top] = 0;
1195 m2.mc_flags = C_INITIALIZED;
1196 mdb_cursor_del(&m2, 0);
1199 if (txn->mt_env->me_pghead) {
1200 MDB_oldpages *mop = txn->mt_env->me_pghead;
1202 /* FIXME: For now, always use fresh pages. We
1203 * really ought to search the free list for a
1208 /* peel pages off tail, so we only have to truncate the list */
1209 pgno = MDB_IDL_LAST(mop->mo_pages);
1210 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1212 if (mop->mo_pages[2] > mop->mo_pages[1])
1213 mop->mo_pages[0] = 0;
1217 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1218 txn->mt_env->me_pghead = mop->mo_next;
1225 if (pgno == P_INVALID) {
1226 /* DB size is maxed out */
1227 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1228 DPUTS("DB size maxed out");
1232 if (txn->mt_env->me_dpages && num == 1) {
1233 np = txn->mt_env->me_dpages;
1234 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1235 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1236 txn->mt_env->me_dpages = np->mp_next;
1238 size_t sz = txn->mt_env->me_psize * num;
1239 if ((np = malloc(sz)) == NULL)
1241 VGMEMP_ALLOC(txn->mt_env, np, sz);
1243 if (pgno == P_INVALID) {
1244 np->mp_pgno = txn->mt_next_pgno;
1245 txn->mt_next_pgno += num;
1249 mid.mid = np->mp_pgno;
1251 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1256 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1257 * @param[in] mc cursor pointing to the page to be touched
1258 * @return 0 on success, non-zero on failure.
1261 mdb_page_touch(MDB_cursor *mc)
1263 MDB_page *mp = mc->mc_pg[mc->mc_top];
1266 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1268 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1270 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1271 assert(mp->mp_pgno != np->mp_pgno);
1272 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1274 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1277 mp->mp_flags |= P_DIRTY;
1280 /* Adjust other cursors pointing to mp */
1281 if (mc->mc_flags & C_SUB) {
1282 MDB_cursor *m2, *m3;
1283 MDB_dbi dbi = mc->mc_dbi-1;
1285 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1286 if (m2 == mc) continue;
1287 m3 = &m2->mc_xcursor->mx_cursor;
1288 if (m3->mc_snum < mc->mc_snum) continue;
1289 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1290 m3->mc_pg[mc->mc_top] = mp;
1296 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1297 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1298 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1299 m2->mc_pg[mc->mc_top] = mp;
1303 mc->mc_pg[mc->mc_top] = mp;
1304 /** If this page has a parent, update the parent to point to
1308 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1310 mc->mc_db->md_root = mp->mp_pgno;
1311 } else if (mc->mc_txn->mt_parent) {
1314 /* If txn has a parent, make sure the page is in our
1317 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1318 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1319 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1320 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1321 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1322 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1323 mc->mc_pg[mc->mc_top] = mp;
1329 np = mdb_page_malloc(mc);
1330 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1331 mid.mid = np->mp_pgno;
1333 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1341 mdb_env_sync(MDB_env *env, int force)
1344 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1345 if (MDB_FDATASYNC(env->me_fd))
1351 /** Make shadow copies of all of parent txn's cursors */
1353 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1355 MDB_cursor *mc, *m2;
1356 unsigned int i, j, size;
1358 for (i=0;i<src->mt_numdbs; i++) {
1359 if (src->mt_cursors[i]) {
1360 size = sizeof(MDB_cursor);
1361 if (src->mt_cursors[i]->mc_xcursor)
1362 size += sizeof(MDB_xcursor);
1363 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1370 mc->mc_db = &dst->mt_dbs[i];
1371 mc->mc_dbx = m2->mc_dbx;
1372 mc->mc_dbflag = &dst->mt_dbflags[i];
1373 mc->mc_snum = m2->mc_snum;
1374 mc->mc_top = m2->mc_top;
1375 mc->mc_flags = m2->mc_flags | C_SHADOW;
1376 for (j=0; j<mc->mc_snum; j++) {
1377 mc->mc_pg[j] = m2->mc_pg[j];
1378 mc->mc_ki[j] = m2->mc_ki[j];
1380 if (m2->mc_xcursor) {
1381 MDB_xcursor *mx, *mx2;
1382 mx = (MDB_xcursor *)(mc+1);
1383 mc->mc_xcursor = mx;
1384 mx2 = m2->mc_xcursor;
1385 mx->mx_db = mx2->mx_db;
1386 mx->mx_dbx = mx2->mx_dbx;
1387 mx->mx_dbflag = mx2->mx_dbflag;
1388 mx->mx_cursor.mc_txn = dst;
1389 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1390 mx->mx_cursor.mc_db = &mx->mx_db;
1391 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1392 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1393 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1394 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1395 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1396 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1397 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1398 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1401 mc->mc_xcursor = NULL;
1403 mc->mc_next = dst->mt_cursors[i];
1404 dst->mt_cursors[i] = mc;
1411 /** Merge shadow cursors back into parent's */
1413 mdb_cursor_merge(MDB_txn *txn)
1416 for (i=0; i<txn->mt_numdbs; i++) {
1417 if (txn->mt_cursors[i]) {
1419 while ((mc = txn->mt_cursors[i])) {
1420 txn->mt_cursors[i] = mc->mc_next;
1421 if (mc->mc_flags & C_SHADOW) {
1422 MDB_cursor *m2 = mc->mc_orig;
1424 m2->mc_snum = mc->mc_snum;
1425 m2->mc_top = mc->mc_top;
1426 for (j=0; j<mc->mc_snum; j++) {
1427 m2->mc_pg[j] = mc->mc_pg[j];
1428 m2->mc_ki[j] = mc->mc_ki[j];
1431 if (mc->mc_flags & C_ALLOCD)
1439 mdb_txn_reset0(MDB_txn *txn);
1441 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1442 * @param[in] txn the transaction handle to initialize
1443 * @return 0 on success, non-zero on failure. This can only
1444 * fail for read-only transactions, and then only if the
1445 * reader table is full.
1448 mdb_txn_renew0(MDB_txn *txn)
1450 MDB_env *env = txn->mt_env;
1453 if (txn->mt_flags & MDB_TXN_RDONLY) {
1454 MDB_reader *r = pthread_getspecific(env->me_txkey);
1457 pid_t pid = getpid();
1458 pthread_t tid = pthread_self();
1461 for (i=0; i<env->me_txns->mti_numreaders; i++)
1462 if (env->me_txns->mti_readers[i].mr_pid == 0)
1464 if (i == env->me_maxreaders) {
1465 UNLOCK_MUTEX_R(env);
1468 env->me_txns->mti_readers[i].mr_pid = pid;
1469 env->me_txns->mti_readers[i].mr_tid = tid;
1470 if (i >= env->me_txns->mti_numreaders)
1471 env->me_txns->mti_numreaders = i+1;
1472 UNLOCK_MUTEX_R(env);
1473 r = &env->me_txns->mti_readers[i];
1474 pthread_setspecific(env->me_txkey, r);
1476 txn->mt_toggle = env->me_txns->mti_me_toggle;
1477 txn->mt_txnid = env->me_txns->mti_txnid;
1478 /* This happens if a different process was the
1479 * last writer to the DB.
1481 if (env->me_wtxnid < txn->mt_txnid)
1482 mt_dbflag = DB_STALE;
1483 r->mr_txnid = txn->mt_txnid;
1484 txn->mt_u.reader = r;
1488 txn->mt_txnid = env->me_txns->mti_txnid;
1489 if (env->me_wtxnid < txn->mt_txnid)
1490 mt_dbflag = DB_STALE;
1492 txn->mt_toggle = env->me_txns->mti_me_toggle;
1493 txn->mt_u.dirty_list = env->me_dirty_list;
1494 txn->mt_u.dirty_list[0].mid = 0;
1495 txn->mt_free_pgs = env->me_free_pgs;
1496 txn->mt_free_pgs[0] = 0;
1497 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1501 /* Copy the DB arrays */
1502 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1503 txn->mt_numdbs = env->me_numdbs;
1504 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1505 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1506 if (txn->mt_numdbs > 2)
1507 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1508 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1509 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1511 memset(txn->mt_dbflags, mt_dbflag, env->me_numdbs);
1517 mdb_txn_renew(MDB_txn *txn)
1524 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1525 DPUTS("environment had fatal error, must shutdown!");
1529 rc = mdb_txn_renew0(txn);
1530 if (rc == MDB_SUCCESS) {
1531 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1532 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1533 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1539 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1544 if (env->me_flags & MDB_FATAL_ERROR) {
1545 DPUTS("environment had fatal error, must shutdown!");
1549 /* parent already has an active child txn */
1550 if (parent->mt_child) {
1554 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1555 if (!(flags & MDB_RDONLY))
1556 size += env->me_maxdbs * sizeof(MDB_cursor *);
1558 if ((txn = calloc(1, size)) == NULL) {
1559 DPRINTF("calloc: %s", strerror(ErrCode()));
1562 txn->mt_dbs = (MDB_db *)(txn+1);
1563 if (flags & MDB_RDONLY) {
1564 txn->mt_flags |= MDB_TXN_RDONLY;
1565 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1567 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1568 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1573 txn->mt_free_pgs = mdb_midl_alloc();
1574 if (!txn->mt_free_pgs) {
1578 txn->mt_u.dirty_list = malloc(sizeof(ID2)*MDB_IDL_UM_SIZE);
1579 if (!txn->mt_u.dirty_list) {
1580 free(txn->mt_free_pgs);
1584 txn->mt_txnid = parent->mt_txnid;
1585 txn->mt_toggle = parent->mt_toggle;
1586 txn->mt_u.dirty_list[0].mid = 0;
1587 txn->mt_free_pgs[0] = 0;
1588 txn->mt_next_pgno = parent->mt_next_pgno;
1589 parent->mt_child = txn;
1590 txn->mt_parent = parent;
1591 txn->mt_numdbs = parent->mt_numdbs;
1592 txn->mt_dbxs = parent->mt_dbxs;
1593 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1594 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1595 mdb_cursor_shadow(parent, txn);
1598 rc = mdb_txn_renew0(txn);
1604 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1605 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1606 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1612 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1613 * @param[in] txn the transaction handle to reset
1616 mdb_txn_reset0(MDB_txn *txn)
1618 MDB_env *env = txn->mt_env;
1620 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1621 txn->mt_u.reader->mr_txnid = 0;
1627 /* close(free) all cursors */
1628 for (i=0; i<txn->mt_numdbs; i++) {
1629 if (txn->mt_cursors[i]) {
1631 while ((mc = txn->mt_cursors[i])) {
1632 txn->mt_cursors[i] = mc->mc_next;
1633 if (mc->mc_flags & C_ALLOCD)
1639 /* return all dirty pages to dpage list */
1640 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1641 dp = txn->mt_u.dirty_list[i].mptr;
1642 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1643 dp->mp_next = txn->mt_env->me_dpages;
1644 VGMEMP_FREE(txn->mt_env, dp);
1645 txn->mt_env->me_dpages = dp;
1647 /* large pages just get freed directly */
1648 VGMEMP_FREE(txn->mt_env, dp);
1653 if (txn->mt_parent) {
1654 txn->mt_parent->mt_child = NULL;
1655 free(txn->mt_free_pgs);
1656 free(txn->mt_u.dirty_list);
1659 if (mdb_midl_shrink(&txn->mt_free_pgs))
1660 env->me_free_pgs = txn->mt_free_pgs;
1663 while ((mop = txn->mt_env->me_pghead)) {
1664 txn->mt_env->me_pghead = mop->mo_next;
1669 /* The writer mutex was locked in mdb_txn_begin. */
1670 UNLOCK_MUTEX_W(env);
1675 mdb_txn_reset(MDB_txn *txn)
1680 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1681 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1682 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1684 mdb_txn_reset0(txn);
1688 mdb_txn_abort(MDB_txn *txn)
1693 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1694 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1695 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1698 mdb_txn_abort(txn->mt_child);
1700 mdb_txn_reset0(txn);
1705 mdb_txn_commit(MDB_txn *txn)
1716 assert(txn != NULL);
1717 assert(txn->mt_env != NULL);
1719 if (txn->mt_child) {
1720 mdb_txn_commit(txn->mt_child);
1721 txn->mt_child = NULL;
1726 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1727 if (txn->mt_numdbs > env->me_numdbs) {
1728 /* update the DB tables */
1729 int toggle = !env->me_db_toggle;
1733 ip = &env->me_dbs[toggle][env->me_numdbs];
1734 jp = &txn->mt_dbs[env->me_numdbs];
1735 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1736 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
1740 env->me_db_toggle = toggle;
1741 env->me_numdbs = txn->mt_numdbs;
1742 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1748 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1749 DPUTS("error flag is set, can't commit");
1751 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1756 /* Merge (and close) our cursors with parent's */
1757 mdb_cursor_merge(txn);
1759 if (txn->mt_parent) {
1765 /* Update parent's DB table */
1766 ip = &txn->mt_parent->mt_dbs[2];
1767 jp = &txn->mt_dbs[2];
1768 for (i = 2; i < txn->mt_numdbs; i++) {
1769 if (ip->md_root != jp->md_root)
1773 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1775 /* Append our free list to parent's */
1776 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1778 mdb_midl_free(txn->mt_free_pgs);
1780 /* Merge our dirty list with parent's */
1781 dst = txn->mt_parent->mt_u.dirty_list;
1782 src = txn->mt_u.dirty_list;
1783 x = mdb_mid2l_search(dst, src[1].mid);
1784 for (y=1; y<=src[0].mid; y++) {
1785 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1789 dst[x].mptr = src[y].mptr;
1792 for (; y<=src[0].mid; y++) {
1793 if (++x >= MDB_IDL_UM_MAX) {
1800 free(txn->mt_u.dirty_list);
1801 txn->mt_parent->mt_child = NULL;
1806 if (txn != env->me_txn) {
1807 DPUTS("attempt to commit unknown transaction");
1812 if (!txn->mt_u.dirty_list[0].mid)
1815 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1816 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1818 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1820 /* should only be one record now */
1821 if (env->me_pghead) {
1822 /* make sure first page of freeDB is touched and on freelist */
1823 mdb_page_search(&mc, NULL, 1);
1825 /* save to free list */
1826 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1830 /* make sure last page of freeDB is touched and on freelist */
1831 key.mv_size = MAXKEYSIZE+1;
1833 mdb_page_search(&mc, &key, 1);
1835 mdb_midl_sort(txn->mt_free_pgs);
1839 ID *idl = txn->mt_free_pgs;
1840 DPRINTF("IDL write txn %zu root %zu num %zu",
1841 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1842 for (i=0; i<idl[0]; i++) {
1843 DPRINTF("IDL %zu", idl[i+1]);
1847 /* write to last page of freeDB */
1848 key.mv_size = sizeof(pgno_t);
1849 key.mv_data = &txn->mt_txnid;
1850 data.mv_data = txn->mt_free_pgs;
1851 /* The free list can still grow during this call,
1852 * despite the pre-emptive touches above. So check
1853 * and make sure the entire thing got written.
1856 i = txn->mt_free_pgs[0];
1857 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1858 rc = mdb_cursor_put(&mc, &key, &data, 0);
1863 } while (i != txn->mt_free_pgs[0]);
1864 if (mdb_midl_shrink(&txn->mt_free_pgs))
1865 env->me_free_pgs = txn->mt_free_pgs;
1867 /* should only be one record now */
1868 if (env->me_pghead) {
1872 mop = env->me_pghead;
1873 env->me_pghead = NULL;
1874 key.mv_size = sizeof(pgno_t);
1875 key.mv_data = &mop->mo_txnid;
1876 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1877 data.mv_data = mop->mo_pages;
1878 mdb_cursor_put(&mc, &key, &data, 0);
1882 /* Update DB root pointers. Their pages have already been
1883 * touched so this is all in-place and cannot fail.
1888 data.mv_size = sizeof(MDB_db);
1890 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
1891 for (i = 2; i < txn->mt_numdbs; i++) {
1892 if (txn->mt_dbflags[i] & DB_DIRTY) {
1893 data.mv_data = &txn->mt_dbs[i];
1894 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1899 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
1905 /* Windows actually supports scatter/gather I/O, but only on
1906 * unbuffered file handles. Since we're relying on the OS page
1907 * cache for all our data, that's self-defeating. So we just
1908 * write pages one at a time. We use the ov structure to set
1909 * the write offset, to at least save the overhead of a Seek
1913 memset(&ov, 0, sizeof(ov));
1914 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1916 dp = txn->mt_u.dirty_list[i].mptr;
1917 DPRINTF("committing page %zu", dp->mp_pgno);
1918 size = dp->mp_pgno * env->me_psize;
1919 ov.Offset = size & 0xffffffff;
1920 ov.OffsetHigh = size >> 16;
1921 ov.OffsetHigh >>= 16;
1922 /* clear dirty flag */
1923 dp->mp_flags &= ~P_DIRTY;
1924 wsize = env->me_psize;
1925 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
1926 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
1929 DPRINTF("WriteFile: %d", n);
1936 struct iovec iov[MDB_COMMIT_PAGES];
1940 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1941 dp = txn->mt_u.dirty_list[i].mptr;
1942 if (dp->mp_pgno != next) {
1944 DPRINTF("committing %u dirty pages", n);
1945 rc = writev(env->me_fd, iov, n);
1949 DPUTS("short write, filesystem full?");
1951 DPRINTF("writev: %s", strerror(n));
1958 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
1961 DPRINTF("committing page %zu", dp->mp_pgno);
1962 iov[n].iov_len = env->me_psize;
1963 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
1964 iov[n].iov_base = (char *)dp;
1965 size += iov[n].iov_len;
1966 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
1967 /* clear dirty flag */
1968 dp->mp_flags &= ~P_DIRTY;
1969 if (++n >= MDB_COMMIT_PAGES) {
1979 DPRINTF("committing %u dirty pages", n);
1980 rc = writev(env->me_fd, iov, n);
1984 DPUTS("short write, filesystem full?");
1986 DPRINTF("writev: %s", strerror(n));
1993 /* Drop the dirty pages.
1995 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1996 dp = txn->mt_u.dirty_list[i].mptr;
1997 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1998 dp->mp_next = txn->mt_env->me_dpages;
1999 VGMEMP_FREE(txn->mt_env, dp);
2000 txn->mt_env->me_dpages = dp;
2002 VGMEMP_FREE(txn->mt_env, dp);
2005 txn->mt_u.dirty_list[i].mid = 0;
2007 txn->mt_u.dirty_list[0].mid = 0;
2009 if ((n = mdb_env_sync(env, 0)) != 0 ||
2010 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2014 env->me_wtxnid = txn->mt_txnid;
2018 /* update the DB tables */
2020 int toggle = !env->me_db_toggle;
2024 ip = &env->me_dbs[toggle][2];
2025 jp = &txn->mt_dbs[2];
2026 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
2027 for (i = 2; i < txn->mt_numdbs; i++) {
2028 if (ip->md_root != jp->md_root)
2033 env->me_db_toggle = toggle;
2034 env->me_numdbs = txn->mt_numdbs;
2035 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
2038 UNLOCK_MUTEX_W(env);
2044 /** Read the environment parameters of a DB environment before
2045 * mapping it into memory.
2046 * @param[in] env the environment handle
2047 * @param[out] meta address of where to store the meta information
2048 * @return 0 on success, non-zero on failure.
2051 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2053 char page[MDB_PAGESIZE];
2058 /* We don't know the page size yet, so use a minimum value.
2062 if (!ReadFile(env->me_fd, page, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2064 if ((rc = read(env->me_fd, page, MDB_PAGESIZE)) == 0)
2069 else if (rc != MDB_PAGESIZE) {
2073 DPRINTF("read: %s", strerror(err));
2077 p = (MDB_page *)page;
2079 if (!F_ISSET(p->mp_flags, P_META)) {
2080 DPRINTF("page %zu not a meta page", p->mp_pgno);
2085 if (m->mm_magic != MDB_MAGIC) {
2086 DPUTS("meta has invalid magic");
2090 if (m->mm_version != MDB_VERSION) {
2091 DPRINTF("database is version %u, expected version %u",
2092 m->mm_version, MDB_VERSION);
2093 return MDB_VERSION_MISMATCH;
2096 memcpy(meta, m, sizeof(*m));
2100 /** Write the environment parameters of a freshly created DB environment.
2101 * @param[in] env the environment handle
2102 * @param[out] meta address of where to store the meta information
2103 * @return 0 on success, non-zero on failure.
2106 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2113 DPUTS("writing new meta page");
2115 GET_PAGESIZE(psize);
2117 meta->mm_magic = MDB_MAGIC;
2118 meta->mm_version = MDB_VERSION;
2119 meta->mm_psize = psize;
2120 meta->mm_last_pg = 1;
2121 meta->mm_flags = env->me_flags & 0xffff;
2122 meta->mm_flags |= MDB_INTEGERKEY;
2123 meta->mm_dbs[0].md_root = P_INVALID;
2124 meta->mm_dbs[1].md_root = P_INVALID;
2126 p = calloc(2, psize);
2128 p->mp_flags = P_META;
2131 memcpy(m, meta, sizeof(*meta));
2133 q = (MDB_page *)((char *)p + psize);
2136 q->mp_flags = P_META;
2139 memcpy(m, meta, sizeof(*meta));
2144 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2145 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2148 rc = write(env->me_fd, p, psize * 2);
2149 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2155 /** Update the environment info to commit a transaction.
2156 * @param[in] txn the transaction that's being committed
2157 * @return 0 on success, non-zero on failure.
2160 mdb_env_write_meta(MDB_txn *txn)
2163 MDB_meta meta, metab;
2165 int rc, len, toggle;
2171 assert(txn != NULL);
2172 assert(txn->mt_env != NULL);
2174 toggle = !txn->mt_toggle;
2175 DPRINTF("writing meta page %d for root page %zu",
2176 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2180 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2181 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2183 ptr = (char *)&meta;
2184 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2185 len = sizeof(MDB_meta) - off;
2188 meta.mm_dbs[0] = txn->mt_dbs[0];
2189 meta.mm_dbs[1] = txn->mt_dbs[1];
2190 meta.mm_last_pg = txn->mt_next_pgno - 1;
2191 meta.mm_txnid = txn->mt_txnid;
2194 off += env->me_psize;
2197 /* Write to the SYNC fd */
2200 memset(&ov, 0, sizeof(ov));
2202 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2205 rc = pwrite(env->me_mfd, ptr, len, off);
2210 DPUTS("write failed, disk error?");
2211 /* On a failure, the pagecache still contains the new data.
2212 * Write some old data back, to prevent it from being used.
2213 * Use the non-SYNC fd; we know it will fail anyway.
2215 meta.mm_last_pg = metab.mm_last_pg;
2216 meta.mm_txnid = metab.mm_txnid;
2218 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2220 r2 = pwrite(env->me_fd, ptr, len, off);
2222 env->me_flags |= MDB_FATAL_ERROR;
2225 /* Memory ordering issues are irrelevant; since the entire writer
2226 * is wrapped by wmutex, all of these changes will become visible
2227 * after the wmutex is unlocked. Since the DB is multi-version,
2228 * readers will get consistent data regardless of how fresh or
2229 * how stale their view of these values is.
2231 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
2232 txn->mt_env->me_txns->mti_me_toggle = toggle;
2233 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2234 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
2239 /** Check both meta pages to see which one is newer.
2240 * @param[in] env the environment handle
2241 * @param[out] which address of where to store the meta toggle ID
2242 * @return 0 on success, non-zero on failure.
2245 mdb_env_read_meta(MDB_env *env, int *which)
2249 assert(env != NULL);
2251 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2254 DPRINTF("Using meta page %d", toggle);
2261 mdb_env_create(MDB_env **env)
2265 e = calloc(1, sizeof(MDB_env));
2269 e->me_free_pgs = mdb_midl_alloc();
2270 if (!e->me_free_pgs) {
2274 e->me_maxreaders = DEFAULT_READERS;
2276 e->me_fd = INVALID_HANDLE_VALUE;
2277 e->me_lfd = INVALID_HANDLE_VALUE;
2278 e->me_mfd = INVALID_HANDLE_VALUE;
2279 VGMEMP_CREATE(e,0,0);
2285 mdb_env_set_mapsize(MDB_env *env, size_t size)
2289 env->me_mapsize = size;
2291 env->me_maxpg = env->me_mapsize / env->me_psize;
2296 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2300 env->me_maxdbs = dbs;
2305 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2307 if (env->me_map || readers < 1)
2309 env->me_maxreaders = readers;
2314 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2316 if (!env || !readers)
2318 *readers = env->me_maxreaders;
2322 /** Further setup required for opening an MDB environment
2325 mdb_env_open2(MDB_env *env, unsigned int flags)
2327 int i, newenv = 0, toggle;
2331 env->me_flags = flags;
2333 memset(&meta, 0, sizeof(meta));
2335 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2338 DPUTS("new mdbenv");
2342 if (!env->me_mapsize) {
2343 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2349 LONG sizelo, sizehi;
2350 sizelo = env->me_mapsize & 0xffffffff;
2351 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2353 /* Windows won't create mappings for zero length files.
2354 * Just allocate the maxsize right now.
2357 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2358 if (!SetEndOfFile(env->me_fd))
2360 SetFilePointer(env->me_fd, 0, NULL, 0);
2362 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2363 sizehi, sizelo, NULL);
2366 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2374 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2376 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2378 if (env->me_map == MAP_FAILED)
2383 meta.mm_mapsize = env->me_mapsize;
2384 if (flags & MDB_FIXEDMAP)
2385 meta.mm_address = env->me_map;
2386 i = mdb_env_init_meta(env, &meta);
2387 if (i != MDB_SUCCESS) {
2388 munmap(env->me_map, env->me_mapsize);
2392 env->me_psize = meta.mm_psize;
2394 env->me_maxpg = env->me_mapsize / env->me_psize;
2396 p = (MDB_page *)env->me_map;
2397 env->me_metas[0] = METADATA(p);
2398 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2400 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
2403 DPRINTF("opened database version %u, pagesize %u",
2404 env->me_metas[toggle]->mm_version, env->me_psize);
2405 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
2406 DPRINTF("entries: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
2407 DPRINTF("branch pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
2408 DPRINTF("leaf pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
2409 DPRINTF("overflow pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
2410 DPRINTF("root: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
2416 /** Release a reader thread's slot in the reader lock table.
2417 * This function is called automatically when a thread exits.
2418 * Windows doesn't support destructor callbacks for thread-specific storage,
2419 * so this function is not compiled there.
2420 * @param[in] ptr This points to the slot in the reader lock table.
2423 mdb_env_reader_dest(void *ptr)
2425 MDB_reader *reader = ptr;
2427 reader->mr_txnid = 0;
2433 /** Downgrade the exclusive lock on the region back to shared */
2435 mdb_env_share_locks(MDB_env *env)
2439 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2441 env->me_txns->mti_me_toggle = toggle;
2442 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2447 /* First acquire a shared lock. The Unlock will
2448 * then release the existing exclusive lock.
2450 memset(&ov, 0, sizeof(ov));
2451 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2452 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2456 struct flock lock_info;
2457 /* The shared lock replaces the existing lock */
2458 memset((void *)&lock_info, 0, sizeof(lock_info));
2459 lock_info.l_type = F_RDLCK;
2460 lock_info.l_whence = SEEK_SET;
2461 lock_info.l_start = 0;
2462 lock_info.l_len = 1;
2463 fcntl(env->me_lfd, F_SETLK, &lock_info);
2467 #if defined(_WIN32) || defined(__APPLE__)
2469 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2471 * @(#) $Revision: 5.1 $
2472 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2473 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2475 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2479 * Please do not copyright this code. This code is in the public domain.
2481 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2482 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2483 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2484 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2485 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2486 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2487 * PERFORMANCE OF THIS SOFTWARE.
2490 * chongo <Landon Curt Noll> /\oo/\
2491 * http://www.isthe.com/chongo/
2493 * Share and Enjoy! :-)
2496 typedef unsigned long long mdb_hash_t;
2497 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2499 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2500 * @param[in] str string to hash
2501 * @param[in] hval initial value for hash
2502 * @return 64 bit hash
2504 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2505 * hval arg on the first call.
2508 mdb_hash_str(char *str, mdb_hash_t hval)
2510 unsigned char *s = (unsigned char *)str; /* unsigned string */
2512 * FNV-1a hash each octet of the string
2515 /* xor the bottom with the current octet */
2516 hval ^= (mdb_hash_t)*s++;
2518 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2519 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2520 (hval << 7) + (hval << 8) + (hval << 40);
2522 /* return our new hash value */
2526 /** Hash the string and output the hash in hex.
2527 * @param[in] str string to hash
2528 * @param[out] hexbuf an array of 17 chars to hold the hash
2531 mdb_hash_hex(char *str, char *hexbuf)
2534 mdb_hash_t h = mdb_hash_str(str, MDB_HASH_INIT);
2535 for (i=0; i<8; i++) {
2536 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2542 /** Open and/or initialize the lock region for the environment.
2543 * @param[in] env The MDB environment.
2544 * @param[in] lpath The pathname of the file used for the lock region.
2545 * @param[in] mode The Unix permissions for the file, if we create it.
2546 * @param[out] excl Set to true if we got an exclusive lock on the region.
2547 * @return 0 on success, non-zero on failure.
2550 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2558 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2559 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2560 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2564 /* Try to get exclusive lock. If we succeed, then
2565 * nobody is using the lock region and we should initialize it.
2568 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2572 memset(&ov, 0, sizeof(ov));
2573 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2579 size = GetFileSize(env->me_lfd, NULL);
2581 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2585 /* Try to get exclusive lock. If we succeed, then
2586 * nobody is using the lock region and we should initialize it.
2589 struct flock lock_info;
2590 memset((void *)&lock_info, 0, sizeof(lock_info));
2591 lock_info.l_type = F_WRLCK;
2592 lock_info.l_whence = SEEK_SET;
2593 lock_info.l_start = 0;
2594 lock_info.l_len = 1;
2595 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2599 lock_info.l_type = F_RDLCK;
2600 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2607 size = lseek(env->me_lfd, 0, SEEK_END);
2609 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2610 if (size < rsize && *excl) {
2612 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2613 if (!SetEndOfFile(env->me_lfd)) {
2618 if (ftruncate(env->me_lfd, rsize) != 0) {
2625 size = rsize - sizeof(MDB_txninfo);
2626 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2631 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2637 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2639 if (!env->me_txns) {
2645 env->me_txns = (MDB_txninfo *)mmap(0, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2647 if (env->me_txns == MAP_FAILED) {
2655 if (!mdb_sec_inited) {
2656 InitializeSecurityDescriptor(&mdb_null_sd,
2657 SECURITY_DESCRIPTOR_REVISION);
2658 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2659 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2660 mdb_all_sa.bInheritHandle = FALSE;
2661 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2664 mdb_hash_hex(lpath, hexbuf);
2665 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2666 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2667 if (!env->me_rmutex) {
2671 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2672 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2673 if (!env->me_wmutex) {
2680 mdb_hash_hex(lpath, hexbuf);
2681 sprintf(env->me_txns->mti_rmname, "MDBr%s", hexbuf);
2682 if (sem_unlink(env->me_txns->mti_rmname)) {
2684 if (rc != ENOENT && rc != EINVAL)
2687 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2688 if (!env->me_rmutex) {
2692 sprintf(env->me_txns->mti_wmname, "MDBw%s", hexbuf);
2693 if (sem_unlink(env->me_txns->mti_wmname)) {
2695 if (rc != ENOENT && rc != EINVAL)
2698 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2699 if (!env->me_wmutex) {
2703 #else /* __APPLE__ */
2704 pthread_mutexattr_t mattr;
2706 pthread_mutexattr_init(&mattr);
2707 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2711 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2712 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2713 #endif /* __APPLE__ */
2715 env->me_txns->mti_version = MDB_VERSION;
2716 env->me_txns->mti_magic = MDB_MAGIC;
2717 env->me_txns->mti_txnid = 0;
2718 env->me_txns->mti_numreaders = 0;
2719 env->me_txns->mti_me_toggle = 0;
2722 if (env->me_txns->mti_magic != MDB_MAGIC) {
2723 DPUTS("lock region has invalid magic");
2727 if (env->me_txns->mti_version != MDB_VERSION) {
2728 DPRINTF("lock region is version %u, expected version %u",
2729 env->me_txns->mti_version, MDB_VERSION);
2730 rc = MDB_VERSION_MISMATCH;
2734 if (rc != EACCES && rc != EAGAIN) {
2738 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2739 if (!env->me_rmutex) {
2743 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2744 if (!env->me_wmutex) {
2750 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2751 if (!env->me_rmutex) {
2755 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2756 if (!env->me_wmutex) {
2766 env->me_lfd = INVALID_HANDLE_VALUE;
2771 /** The name of the lock file in the DB environment */
2772 #define LOCKNAME "/lock.mdb"
2773 /** The name of the data file in the DB environment */
2774 #define DATANAME "/data.mdb"
2775 /** The suffix of the lock file when no subdir is used */
2776 #define LOCKSUFF "-lock"
2779 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2781 int oflags, rc, len, excl;
2782 char *lpath, *dpath;
2785 if (flags & MDB_NOSUBDIR) {
2786 rc = len + sizeof(LOCKSUFF) + len + 1;
2788 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
2793 if (flags & MDB_NOSUBDIR) {
2794 dpath = lpath + len + sizeof(LOCKSUFF);
2795 sprintf(lpath, "%s" LOCKSUFF, path);
2796 strcpy(dpath, path);
2798 dpath = lpath + len + sizeof(LOCKNAME);
2799 sprintf(lpath, "%s" LOCKNAME, path);
2800 sprintf(dpath, "%s" DATANAME, path);
2803 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2808 if (F_ISSET(flags, MDB_RDONLY)) {
2809 oflags = GENERIC_READ;
2810 len = OPEN_EXISTING;
2812 oflags = GENERIC_READ|GENERIC_WRITE;
2815 mode = FILE_ATTRIBUTE_NORMAL;
2816 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2817 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2822 if (F_ISSET(flags, MDB_RDONLY))
2825 oflags = O_RDWR | O_CREAT;
2827 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2833 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2834 /* synchronous fd for meta writes */
2836 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2837 mode |= FILE_FLAG_WRITE_THROUGH;
2838 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2839 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2844 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2845 oflags |= MDB_DSYNC;
2846 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
2851 env->me_path = strdup(path);
2852 DPRINTF("opened dbenv %p", (void *) env);
2853 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
2854 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
2856 mdb_env_share_locks(env);
2857 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
2858 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
2859 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
2865 if (env->me_fd != INVALID_HANDLE_VALUE) {
2867 env->me_fd = INVALID_HANDLE_VALUE;
2869 if (env->me_lfd != INVALID_HANDLE_VALUE) {
2871 env->me_lfd = INVALID_HANDLE_VALUE;
2879 mdb_env_close(MDB_env *env)
2886 VGMEMP_DESTROY(env);
2887 while (env->me_dpages) {
2888 dp = env->me_dpages;
2889 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
2890 env->me_dpages = dp->mp_next;
2894 free(env->me_dbs[1]);
2895 free(env->me_dbs[0]);
2899 LAZY_RWLOCK_DESTROY(&env->me_dblock);
2900 pthread_key_delete(env->me_txkey);
2903 munmap(env->me_map, env->me_mapsize);
2908 pid_t pid = getpid();
2910 for (i=0; i<env->me_txns->mti_numreaders; i++)
2911 if (env->me_txns->mti_readers[i].mr_pid == pid)
2912 env->me_txns->mti_readers[i].mr_pid = 0;
2913 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
2916 mdb_midl_free(env->me_free_pgs);
2920 /** Compare two items pointing at aligned size_t's */
2922 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
2924 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
2925 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
2928 /** Compare two items pointing at aligned int's */
2930 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
2932 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
2933 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
2936 /** Compare two items pointing at ints of unknown alignment.
2937 * Nodes and keys are guaranteed to be 2-byte aligned.
2940 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
2942 #if BYTE_ORDER == LITTLE_ENDIAN
2943 unsigned short *u, *c;
2946 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
2947 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
2950 } while(!x && u > (unsigned short *)a->mv_data);
2953 return memcmp(a->mv_data, b->mv_data, a->mv_size);
2957 /** Compare two items lexically */
2959 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
2966 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
2972 diff = memcmp(a->mv_data, b->mv_data, len);
2973 return diff ? diff : len_diff<0 ? -1 : len_diff;
2976 /** Compare two items in reverse byte order */
2978 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
2980 const unsigned char *p1, *p2, *p1_lim;
2984 p1_lim = (const unsigned char *)a->mv_data;
2985 p1 = (const unsigned char *)a->mv_data + a->mv_size;
2986 p2 = (const unsigned char *)b->mv_data + b->mv_size;
2988 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
2994 while (p1 > p1_lim) {
2995 diff = *--p1 - *--p2;
2999 return len_diff<0 ? -1 : len_diff;
3002 /** Search for key within a page, using binary search.
3003 * Returns the smallest entry larger or equal to the key.
3004 * If exactp is non-null, stores whether the found entry was an exact match
3005 * in *exactp (1 or 0).
3006 * Updates the cursor index with the index of the found entry.
3007 * If no entry larger or equal to the key is found, returns NULL.
3010 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3012 unsigned int i = 0, nkeys;
3015 MDB_page *mp = mc->mc_pg[mc->mc_top];
3016 MDB_node *node = NULL;
3021 nkeys = NUMKEYS(mp);
3026 COPY_PGNO(pgno, mp->mp_pgno);
3027 DPRINTF("searching %u keys in %s %spage %zu",
3028 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3035 low = IS_LEAF(mp) ? 0 : 1;
3037 cmp = mc->mc_dbx->md_cmp;
3039 /* Branch pages have no data, so if using integer keys,
3040 * alignment is guaranteed. Use faster mdb_cmp_int.
3042 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3043 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3050 nodekey.mv_size = mc->mc_db->md_pad;
3051 node = NODEPTR(mp, 0); /* fake */
3052 while (low <= high) {
3053 i = (low + high) >> 1;
3054 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3055 rc = cmp(key, &nodekey);
3056 DPRINTF("found leaf index %u [%s], rc = %i",
3057 i, DKEY(&nodekey), rc);
3066 while (low <= high) {
3067 i = (low + high) >> 1;
3069 node = NODEPTR(mp, i);
3070 nodekey.mv_size = NODEKSZ(node);
3071 nodekey.mv_data = NODEKEY(node);
3073 rc = cmp(key, &nodekey);
3076 DPRINTF("found leaf index %u [%s], rc = %i",
3077 i, DKEY(&nodekey), rc);
3079 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3080 i, DKEY(&nodekey), NODEPGNO(node), rc);
3091 if (rc > 0) { /* Found entry is less than the key. */
3092 i++; /* Skip to get the smallest entry larger than key. */
3094 node = NODEPTR(mp, i);
3097 *exactp = (rc == 0);
3098 /* store the key index */
3099 mc->mc_ki[mc->mc_top] = i;
3101 /* There is no entry larger or equal to the key. */
3104 /* nodeptr is fake for LEAF2 */
3110 mdb_cursor_adjust(MDB_cursor *mc, func)
3114 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3115 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3122 /** Pop a page off the top of the cursor's stack. */
3124 mdb_cursor_pop(MDB_cursor *mc)
3129 top = mc->mc_pg[mc->mc_top];
3134 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3135 mc->mc_dbi, (void *) mc);
3139 /** Push a page onto the top of the cursor's stack. */
3141 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3143 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3144 mc->mc_dbi, (void *) mc);
3146 if (mc->mc_snum >= CURSOR_STACK) {
3147 assert(mc->mc_snum < CURSOR_STACK);
3151 mc->mc_top = mc->mc_snum++;
3152 mc->mc_pg[mc->mc_top] = mp;
3153 mc->mc_ki[mc->mc_top] = 0;
3158 /** Find the address of the page corresponding to a given page number.
3159 * @param[in] txn the transaction for this access.
3160 * @param[in] pgno the page number for the page to retrieve.
3161 * @param[out] ret address of a pointer where the page's address will be stored.
3162 * @return 0 on success, non-zero on failure.
3165 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3169 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3171 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3172 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3173 p = txn->mt_u.dirty_list[x].mptr;
3177 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
3178 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3182 DPRINTF("page %zu not found", pgno);
3185 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3188 /** Search for the page a given key should be in.
3189 * Pushes parent pages on the cursor stack. This function continues a
3190 * search on a cursor that has already been initialized. (Usually by
3191 * #mdb_page_search() but also by #mdb_node_move().)
3192 * @param[in,out] mc the cursor for this operation.
3193 * @param[in] key the key to search for. If NULL, search for the lowest
3194 * page. (This is used by #mdb_cursor_first().)
3195 * @param[in] modify If true, visited pages are updated with new page numbers.
3196 * @return 0 on success, non-zero on failure.
3199 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3201 MDB_page *mp = mc->mc_pg[mc->mc_top];
3206 while (IS_BRANCH(mp)) {
3210 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3211 assert(NUMKEYS(mp) > 1);
3212 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3214 if (key == NULL) /* Initialize cursor to first page. */
3216 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3217 /* cursor to last page */
3221 node = mdb_node_search(mc, key, &exact);
3223 i = NUMKEYS(mp) - 1;
3225 i = mc->mc_ki[mc->mc_top];
3234 DPRINTF("following index %u for key [%s]",
3236 assert(i < NUMKEYS(mp));
3237 node = NODEPTR(mp, i);
3239 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3242 mc->mc_ki[mc->mc_top] = i;
3243 if ((rc = mdb_cursor_push(mc, mp)))
3247 if ((rc = mdb_page_touch(mc)) != 0)
3249 mp = mc->mc_pg[mc->mc_top];
3254 DPRINTF("internal error, index points to a %02X page!?",
3256 return MDB_CORRUPTED;
3259 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3260 key ? DKEY(key) : NULL);
3265 /** Search for the page a given key should be in.
3266 * Pushes parent pages on the cursor stack. This function just sets up
3267 * the search; it finds the root page for \b mc's database and sets this
3268 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3269 * called to complete the search.
3270 * @param[in,out] mc the cursor for this operation.
3271 * @param[in] key the key to search for. If NULL, search for the lowest
3272 * page. (This is used by #mdb_cursor_first().)
3273 * @param[in] modify If true, visited pages are updated with new page numbers.
3274 * @return 0 on success, non-zero on failure.
3277 mdb_page_search(MDB_cursor *mc, MDB_val *key, int modify)
3282 /* Make sure the txn is still viable, then find the root from
3283 * the txn's db table.
3285 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3286 DPUTS("transaction has failed, must abort");
3289 /* Make sure we're using an up-to-date root */
3290 if (mc->mc_dbi > MAIN_DBI) {
3291 if ((*mc->mc_dbflag & DB_STALE) ||
3292 (modify && !(*mc->mc_dbflag & DB_DIRTY))) {
3294 unsigned char dbflag = 0;
3295 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3296 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, modify);
3299 if (*mc->mc_dbflag & DB_STALE) {
3302 MDB_node *leaf = mdb_node_search(&mc2,
3303 &mc->mc_dbx->md_name, &exact);
3305 return MDB_NOTFOUND;
3306 mdb_node_read(mc->mc_txn, leaf, &data);
3307 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3311 *mc->mc_dbflag = dbflag;
3314 root = mc->mc_db->md_root;
3316 if (root == P_INVALID) { /* Tree is empty. */
3317 DPUTS("tree is empty");
3318 return MDB_NOTFOUND;
3323 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3329 DPRINTF("db %u root page %zu has flags 0x%X",
3330 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3333 if ((rc = mdb_page_touch(mc)))
3337 return mdb_page_search_root(mc, key, modify);
3340 /** Return the data associated with a given node.
3341 * @param[in] txn The transaction for this operation.
3342 * @param[in] leaf The node being read.
3343 * @param[out] data Updated to point to the node's data.
3344 * @return 0 on success, non-zero on failure.
3347 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3349 MDB_page *omp; /* overflow page */
3353 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3354 data->mv_size = NODEDSZ(leaf);
3355 data->mv_data = NODEDATA(leaf);
3359 /* Read overflow data.
3361 data->mv_size = NODEDSZ(leaf);
3362 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3363 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3364 DPRINTF("read overflow page %zu failed", pgno);
3367 data->mv_data = METADATA(omp);
3373 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3374 MDB_val *key, MDB_val *data)
3383 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3385 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3388 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3392 mdb_cursor_init(&mc, txn, dbi, &mx);
3393 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3396 /** Find a sibling for a page.
3397 * Replaces the page at the top of the cursor's stack with the
3398 * specified sibling, if one exists.
3399 * @param[in] mc The cursor for this operation.
3400 * @param[in] move_right Non-zero if the right sibling is requested,
3401 * otherwise the left sibling.
3402 * @return 0 on success, non-zero on failure.
3405 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3411 if (mc->mc_snum < 2) {
3412 return MDB_NOTFOUND; /* root has no siblings */
3416 DPRINTF("parent page is page %zu, index %u",
3417 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3419 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3420 : (mc->mc_ki[mc->mc_top] == 0)) {
3421 DPRINTF("no more keys left, moving to %s sibling",
3422 move_right ? "right" : "left");
3423 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3427 mc->mc_ki[mc->mc_top]++;
3429 mc->mc_ki[mc->mc_top]--;
3430 DPRINTF("just moving to %s index key %u",
3431 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3433 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3435 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3436 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3439 mdb_cursor_push(mc, mp);
3444 /** Move the cursor to the next data item. */
3446 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3452 if (mc->mc_flags & C_EOF) {
3453 return MDB_NOTFOUND;
3456 assert(mc->mc_flags & C_INITIALIZED);
3458 mp = mc->mc_pg[mc->mc_top];
3460 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3461 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3462 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3463 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3464 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3465 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3469 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3470 if (op == MDB_NEXT_DUP)
3471 return MDB_NOTFOUND;
3475 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3477 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3478 DPUTS("=====> move to next sibling page");
3479 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3480 mc->mc_flags |= C_EOF;
3481 mc->mc_flags &= ~C_INITIALIZED;
3482 return MDB_NOTFOUND;
3484 mp = mc->mc_pg[mc->mc_top];
3485 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3487 mc->mc_ki[mc->mc_top]++;
3489 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3490 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3493 key->mv_size = mc->mc_db->md_pad;
3494 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3498 assert(IS_LEAF(mp));
3499 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3501 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3502 mdb_xcursor_init1(mc, leaf);
3505 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3508 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3509 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3510 if (rc != MDB_SUCCESS)
3515 MDB_SET_KEY(leaf, key);
3519 /** Move the cursor to the previous data item. */
3521 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3527 assert(mc->mc_flags & C_INITIALIZED);
3529 mp = mc->mc_pg[mc->mc_top];
3531 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3532 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3533 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3534 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3535 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3536 if (op != MDB_PREV || rc == MDB_SUCCESS)
3539 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3540 if (op == MDB_PREV_DUP)
3541 return MDB_NOTFOUND;
3546 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3548 if (mc->mc_ki[mc->mc_top] == 0) {
3549 DPUTS("=====> move to prev sibling page");
3550 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3551 mc->mc_flags &= ~C_INITIALIZED;
3552 return MDB_NOTFOUND;
3554 mp = mc->mc_pg[mc->mc_top];
3555 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3556 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3558 mc->mc_ki[mc->mc_top]--;
3560 mc->mc_flags &= ~C_EOF;
3562 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3563 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3566 key->mv_size = mc->mc_db->md_pad;
3567 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3571 assert(IS_LEAF(mp));
3572 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3574 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3575 mdb_xcursor_init1(mc, leaf);
3578 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3581 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3582 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3583 if (rc != MDB_SUCCESS)
3588 MDB_SET_KEY(leaf, key);
3592 /** Set the cursor on a specific data item. */
3594 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3595 MDB_cursor_op op, int *exactp)
3604 assert(key->mv_size > 0);
3606 /* See if we're already on the right page */
3607 if (mc->mc_flags & C_INITIALIZED) {
3610 mp = mc->mc_pg[mc->mc_top];
3612 mc->mc_ki[mc->mc_top] = 0;
3613 return MDB_NOTFOUND;
3615 if (mp->mp_flags & P_LEAF2) {
3616 nodekey.mv_size = mc->mc_db->md_pad;
3617 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3619 leaf = NODEPTR(mp, 0);
3620 MDB_SET_KEY(leaf, &nodekey);
3622 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3624 /* Probably happens rarely, but first node on the page
3625 * was the one we wanted.
3627 mc->mc_ki[mc->mc_top] = 0;
3628 leaf = NODEPTR(mp, 0);
3635 unsigned int nkeys = NUMKEYS(mp);
3637 if (mp->mp_flags & P_LEAF2) {
3638 nodekey.mv_data = LEAF2KEY(mp,
3639 nkeys-1, nodekey.mv_size);
3641 leaf = NODEPTR(mp, nkeys-1);
3642 MDB_SET_KEY(leaf, &nodekey);
3644 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3646 /* last node was the one we wanted */
3647 mc->mc_ki[mc->mc_top] = nkeys-1;
3648 leaf = NODEPTR(mp, nkeys-1);
3654 /* This is definitely the right page, skip search_page */
3659 /* If any parents have right-sibs, search.
3660 * Otherwise, there's nothing further.
3662 for (i=0; i<mc->mc_top; i++)
3664 NUMKEYS(mc->mc_pg[i])-1)
3666 if (i == mc->mc_top) {
3667 /* There are no other pages */
3668 mc->mc_ki[mc->mc_top] = nkeys;
3669 return MDB_NOTFOUND;
3673 /* There are no other pages */
3674 mc->mc_ki[mc->mc_top] = 0;
3675 return MDB_NOTFOUND;
3679 rc = mdb_page_search(mc, key, 0);
3680 if (rc != MDB_SUCCESS)
3683 mp = mc->mc_pg[mc->mc_top];
3684 assert(IS_LEAF(mp));
3687 leaf = mdb_node_search(mc, key, exactp);
3688 if (exactp != NULL && !*exactp) {
3689 /* MDB_SET specified and not an exact match. */
3690 return MDB_NOTFOUND;
3694 DPUTS("===> inexact leaf not found, goto sibling");
3695 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3696 return rc; /* no entries matched */
3697 mp = mc->mc_pg[mc->mc_top];
3698 assert(IS_LEAF(mp));
3699 leaf = NODEPTR(mp, 0);
3703 mc->mc_flags |= C_INITIALIZED;
3704 mc->mc_flags &= ~C_EOF;
3707 key->mv_size = mc->mc_db->md_pad;
3708 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3712 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3713 mdb_xcursor_init1(mc, leaf);
3716 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3717 if (op == MDB_SET || op == MDB_SET_RANGE) {
3718 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3721 if (op == MDB_GET_BOTH) {
3727 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3728 if (rc != MDB_SUCCESS)
3731 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3733 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3735 rc = mc->mc_dbx->md_dcmp(data, &d2);
3737 if (op == MDB_GET_BOTH || rc > 0)
3738 return MDB_NOTFOUND;
3743 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3744 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3749 /* The key already matches in all other cases */
3750 if (op == MDB_SET_RANGE)
3751 MDB_SET_KEY(leaf, key);
3752 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3757 /** Move the cursor to the first item in the database. */
3759 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3764 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3765 rc = mdb_page_search(mc, NULL, 0);
3766 if (rc != MDB_SUCCESS)
3769 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3771 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3772 mc->mc_flags |= C_INITIALIZED;
3773 mc->mc_flags &= ~C_EOF;
3775 mc->mc_ki[mc->mc_top] = 0;
3777 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3778 key->mv_size = mc->mc_db->md_pad;
3779 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3784 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3785 mdb_xcursor_init1(mc, leaf);
3786 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3791 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3792 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3796 MDB_SET_KEY(leaf, key);
3800 /** Move the cursor to the last item in the database. */
3802 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3808 lkey.mv_size = MAXKEYSIZE+1;
3809 lkey.mv_data = NULL;
3811 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3812 rc = mdb_page_search(mc, &lkey, 0);
3813 if (rc != MDB_SUCCESS)
3816 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3818 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3819 mc->mc_flags |= C_INITIALIZED;
3820 mc->mc_flags &= ~C_EOF;
3822 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3824 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3825 key->mv_size = mc->mc_db->md_pad;
3826 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3831 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3832 mdb_xcursor_init1(mc, leaf);
3833 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3838 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3839 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3844 MDB_SET_KEY(leaf, key);
3849 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3859 case MDB_GET_BOTH_RANGE:
3860 if (data == NULL || mc->mc_xcursor == NULL) {
3867 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3869 } else if (op == MDB_SET_RANGE)
3870 rc = mdb_cursor_set(mc, key, data, op, NULL);
3872 rc = mdb_cursor_set(mc, key, data, op, &exact);
3874 case MDB_GET_MULTIPLE:
3876 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
3877 !(mc->mc_flags & C_INITIALIZED)) {
3882 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
3883 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
3886 case MDB_NEXT_MULTIPLE:
3888 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
3892 if (!(mc->mc_flags & C_INITIALIZED))
3893 rc = mdb_cursor_first(mc, key, data);
3895 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
3896 if (rc == MDB_SUCCESS) {
3897 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
3900 mx = &mc->mc_xcursor->mx_cursor;
3901 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
3903 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
3904 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
3912 case MDB_NEXT_NODUP:
3913 if (!(mc->mc_flags & C_INITIALIZED))
3914 rc = mdb_cursor_first(mc, key, data);
3916 rc = mdb_cursor_next(mc, key, data, op);
3920 case MDB_PREV_NODUP:
3921 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
3922 rc = mdb_cursor_last(mc, key, data);
3924 rc = mdb_cursor_prev(mc, key, data, op);
3927 rc = mdb_cursor_first(mc, key, data);
3931 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
3932 !(mc->mc_flags & C_INITIALIZED) ||
3933 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3937 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3940 rc = mdb_cursor_last(mc, key, data);
3944 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
3945 !(mc->mc_flags & C_INITIALIZED) ||
3946 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3950 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3953 DPRINTF("unhandled/unimplemented cursor operation %u", op);
3961 /** Touch all the pages in the cursor stack.
3962 * Makes sure all the pages are writable, before attempting a write operation.
3963 * @param[in] mc The cursor to operate on.
3966 mdb_cursor_touch(MDB_cursor *mc)
3970 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
3972 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3973 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 1);
3976 *mc->mc_dbflag = DB_DIRTY;
3978 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
3979 rc = mdb_page_touch(mc);
3983 mc->mc_top = mc->mc_snum-1;
3988 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3991 MDB_node *leaf = NULL;
3992 MDB_val xdata, *rdata, dkey;
3996 unsigned int mcount = 0;
3999 char pbuf[MDB_PAGESIZE];
4000 char dbuf[MAXKEYSIZE+1];
4001 unsigned int nflags;
4004 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4007 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4008 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4012 if (flags == MDB_CURRENT) {
4013 if (!(mc->mc_flags & C_INITIALIZED))
4016 } else if (mc->mc_db->md_root == P_INVALID) {
4018 /* new database, write a root leaf page */
4019 DPUTS("allocating new root leaf page");
4020 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4024 mdb_cursor_push(mc, np);
4025 mc->mc_db->md_root = np->mp_pgno;
4026 mc->mc_db->md_depth++;
4027 *mc->mc_dbflag = DB_DIRTY;
4028 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4030 np->mp_flags |= P_LEAF2;
4031 mc->mc_flags |= C_INITIALIZED;
4037 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4038 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4039 DPRINTF("duplicate key [%s]", DKEY(key));
4041 return MDB_KEYEXIST;
4043 if (rc && rc != MDB_NOTFOUND)
4047 /* Cursor is positioned, now make sure all pages are writable */
4048 rc2 = mdb_cursor_touch(mc);
4053 /* The key already exists */
4054 if (rc == MDB_SUCCESS) {
4055 /* there's only a key anyway, so this is a no-op */
4056 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4057 unsigned int ksize = mc->mc_db->md_pad;
4058 if (key->mv_size != ksize)
4060 if (flags == MDB_CURRENT) {
4061 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4062 memcpy(ptr, key->mv_data, ksize);
4067 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4070 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4071 /* Was a single item before, must convert now */
4073 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4074 /* Just overwrite the current item */
4075 if (flags == MDB_CURRENT)
4078 dkey.mv_size = NODEDSZ(leaf);
4079 dkey.mv_data = NODEDATA(leaf);
4080 #if UINT_MAX < SIZE_MAX
4081 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4082 #ifdef MISALIGNED_OK
4083 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4085 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4088 /* if data matches, ignore it */
4089 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4090 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4092 /* create a fake page for the dup items */
4093 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4094 dkey.mv_data = dbuf;
4095 fp = (MDB_page *)pbuf;
4096 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4097 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4098 fp->mp_lower = PAGEHDRSZ;
4099 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4100 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4101 fp->mp_flags |= P_LEAF2;
4102 fp->mp_pad = data->mv_size;
4104 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4105 (dkey.mv_size & 1) + (data->mv_size & 1);
4107 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4110 xdata.mv_size = fp->mp_upper;
4111 xdata.mv_data = pbuf;
4115 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4116 /* See if we need to convert from fake page to subDB */
4118 unsigned int offset;
4121 fp = NODEDATA(leaf);
4122 if (flags == MDB_CURRENT) {
4123 fp->mp_flags |= P_DIRTY;
4124 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4125 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4129 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4130 offset = fp->mp_pad;
4132 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4134 offset += offset & 1;
4135 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4136 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4138 /* yes, convert it */
4140 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4141 dummy.md_pad = fp->mp_pad;
4142 dummy.md_flags = MDB_DUPFIXED;
4143 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4144 dummy.md_flags |= MDB_INTEGERKEY;
4147 dummy.md_branch_pages = 0;
4148 dummy.md_leaf_pages = 1;
4149 dummy.md_overflow_pages = 0;
4150 dummy.md_entries = NUMKEYS(fp);
4152 xdata.mv_size = sizeof(MDB_db);
4153 xdata.mv_data = &dummy;
4154 mp = mdb_page_alloc(mc, 1);
4157 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4158 flags |= F_DUPDATA|F_SUBDATA;
4159 dummy.md_root = mp->mp_pgno;
4161 /* no, just grow it */
4163 xdata.mv_size = NODEDSZ(leaf) + offset;
4164 xdata.mv_data = pbuf;
4165 mp = (MDB_page *)pbuf;
4166 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4169 mp->mp_flags = fp->mp_flags | P_DIRTY;
4170 mp->mp_pad = fp->mp_pad;
4171 mp->mp_lower = fp->mp_lower;
4172 mp->mp_upper = fp->mp_upper + offset;
4174 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4176 nsize = NODEDSZ(leaf) - fp->mp_upper;
4177 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4178 for (i=0; i<NUMKEYS(fp); i++)
4179 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4181 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4185 /* data is on sub-DB, just store it */
4186 flags |= F_DUPDATA|F_SUBDATA;
4190 /* same size, just replace it */
4191 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
4192 NODEDSZ(leaf) == data->mv_size) {
4193 if (F_ISSET(flags, MDB_RESERVE))
4194 data->mv_data = NODEDATA(leaf);
4196 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4199 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4201 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4207 nflags = flags & NODE_ADD_FLAGS;
4208 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4209 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4210 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4211 nflags &= ~MDB_APPEND;
4212 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4214 /* There is room already in this leaf page. */
4215 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4216 if (rc == 0 && !do_sub) {
4217 /* Adjust other cursors pointing to mp */
4218 MDB_cursor *m2, *m3;
4219 MDB_dbi dbi = mc->mc_dbi;
4220 unsigned i = mc->mc_top;
4221 MDB_page *mp = mc->mc_pg[i];
4223 if (mc->mc_flags & C_SUB)
4226 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4227 if (mc->mc_flags & C_SUB)
4228 m3 = &m2->mc_xcursor->mx_cursor;
4231 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4232 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4239 if (rc != MDB_SUCCESS)
4240 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4242 /* Now store the actual data in the child DB. Note that we're
4243 * storing the user data in the keys field, so there are strict
4244 * size limits on dupdata. The actual data fields of the child
4245 * DB are all zero size.
4253 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4254 if (flags & MDB_CURRENT) {
4255 xflags = MDB_CURRENT;
4257 mdb_xcursor_init1(mc, leaf);
4258 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4260 /* converted, write the original data first */
4262 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4266 /* Adjust other cursors pointing to mp */
4268 unsigned i = mc->mc_top;
4269 MDB_page *mp = mc->mc_pg[i];
4271 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4272 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4273 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4274 mdb_xcursor_init1(m2, leaf);
4279 xflags |= (flags & MDB_APPEND);
4280 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4281 if (flags & F_SUBDATA) {
4282 db = NODEDATA(leaf);
4283 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4286 /* sub-writes might have failed so check rc again.
4287 * Don't increment count if we just replaced an existing item.
4289 if (!rc && !(flags & MDB_CURRENT))
4290 mc->mc_db->md_entries++;
4291 if (flags & MDB_MULTIPLE) {
4293 if (mcount < data[1].mv_size) {
4294 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4295 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4305 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4310 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4313 if (!mc->mc_flags & C_INITIALIZED)
4316 rc = mdb_cursor_touch(mc);
4320 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4322 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4323 if (flags != MDB_NODUPDATA) {
4324 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4325 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4327 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4328 /* If sub-DB still has entries, we're done */
4329 if (mc->mc_xcursor->mx_db.md_entries) {
4330 if (leaf->mn_flags & F_SUBDATA) {
4331 /* update subDB info */
4332 MDB_db *db = NODEDATA(leaf);
4333 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4335 /* shrink fake page */
4336 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4338 mc->mc_db->md_entries--;
4341 /* otherwise fall thru and delete the sub-DB */
4344 if (leaf->mn_flags & F_SUBDATA) {
4345 /* add all the child DB's pages to the free list */
4346 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4347 if (rc == MDB_SUCCESS) {
4348 mc->mc_db->md_entries -=
4349 mc->mc_xcursor->mx_db.md_entries;
4354 return mdb_cursor_del0(mc, leaf);
4357 /** Allocate and initialize new pages for a database.
4358 * @param[in] mc a cursor on the database being added to.
4359 * @param[in] flags flags defining what type of page is being allocated.
4360 * @param[in] num the number of pages to allocate. This is usually 1,
4361 * unless allocating overflow pages for a large record.
4362 * @return Address of a page, or NULL on failure.
4365 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4369 if ((np = mdb_page_alloc(mc, num)) == NULL)
4371 DPRINTF("allocated new mpage %zu, page size %u",
4372 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4373 np->mp_flags = flags | P_DIRTY;
4374 np->mp_lower = PAGEHDRSZ;
4375 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4378 mc->mc_db->md_branch_pages++;
4379 else if (IS_LEAF(np))
4380 mc->mc_db->md_leaf_pages++;
4381 else if (IS_OVERFLOW(np)) {
4382 mc->mc_db->md_overflow_pages += num;
4389 /** Calculate the size of a leaf node.
4390 * The size depends on the environment's page size; if a data item
4391 * is too large it will be put onto an overflow page and the node
4392 * size will only include the key and not the data. Sizes are always
4393 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4394 * of the #MDB_node headers.
4395 * @param[in] env The environment handle.
4396 * @param[in] key The key for the node.
4397 * @param[in] data The data for the node.
4398 * @return The number of bytes needed to store the node.
4401 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4405 sz = LEAFSIZE(key, data);
4406 if (sz >= env->me_psize / MDB_MINKEYS) {
4407 /* put on overflow page */
4408 sz -= data->mv_size - sizeof(pgno_t);
4412 return sz + sizeof(indx_t);
4415 /** Calculate the size of a branch node.
4416 * The size should depend on the environment's page size but since
4417 * we currently don't support spilling large keys onto overflow
4418 * pages, it's simply the size of the #MDB_node header plus the
4419 * size of the key. Sizes are always rounded up to an even number
4420 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4421 * @param[in] env The environment handle.
4422 * @param[in] key The key for the node.
4423 * @return The number of bytes needed to store the node.
4426 mdb_branch_size(MDB_env *env, MDB_val *key)
4431 if (sz >= env->me_psize / MDB_MINKEYS) {
4432 /* put on overflow page */
4433 /* not implemented */
4434 /* sz -= key->size - sizeof(pgno_t); */
4437 return sz + sizeof(indx_t);
4440 /** Add a node to the page pointed to by the cursor.
4441 * @param[in] mc The cursor for this operation.
4442 * @param[in] indx The index on the page where the new node should be added.
4443 * @param[in] key The key for the new node.
4444 * @param[in] data The data for the new node, if any.
4445 * @param[in] pgno The page number, if adding a branch node.
4446 * @param[in] flags Flags for the node.
4447 * @return 0 on success, non-zero on failure. Possible errors are:
4449 * <li>ENOMEM - failed to allocate overflow pages for the node.
4450 * <li>ENOSPC - there is insufficient room in the page. This error
4451 * should never happen since all callers already calculate the
4452 * page's free space before calling this function.
4456 mdb_node_add(MDB_cursor *mc, indx_t indx,
4457 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4460 size_t node_size = NODESIZE;
4463 MDB_page *mp = mc->mc_pg[mc->mc_top];
4464 MDB_page *ofp = NULL; /* overflow page */
4467 assert(mp->mp_upper >= mp->mp_lower);
4469 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4470 IS_LEAF(mp) ? "leaf" : "branch",
4471 IS_SUBP(mp) ? "sub-" : "",
4472 mp->mp_pgno, indx, data ? data->mv_size : 0,
4473 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4476 /* Move higher keys up one slot. */
4477 int ksize = mc->mc_db->md_pad, dif;
4478 char *ptr = LEAF2KEY(mp, indx, ksize);
4479 dif = NUMKEYS(mp) - indx;
4481 memmove(ptr+ksize, ptr, dif*ksize);
4482 /* insert new key */
4483 memcpy(ptr, key->mv_data, ksize);
4485 /* Just using these for counting */
4486 mp->mp_lower += sizeof(indx_t);
4487 mp->mp_upper -= ksize - sizeof(indx_t);
4492 node_size += key->mv_size;
4496 if (F_ISSET(flags, F_BIGDATA)) {
4497 /* Data already on overflow page. */
4498 node_size += sizeof(pgno_t);
4499 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4500 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4501 /* Put data on overflow page. */
4502 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4503 data->mv_size, node_size+data->mv_size);
4504 node_size += sizeof(pgno_t);
4505 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4507 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4510 node_size += data->mv_size;
4513 node_size += node_size & 1;
4515 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4516 DPRINTF("not enough room in page %zu, got %u ptrs",
4517 mp->mp_pgno, NUMKEYS(mp));
4518 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4519 mp->mp_upper - mp->mp_lower);
4520 DPRINTF("node size = %zu", node_size);
4524 /* Move higher pointers up one slot. */
4525 for (i = NUMKEYS(mp); i > indx; i--)
4526 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4528 /* Adjust free space offsets. */
4529 ofs = mp->mp_upper - node_size;
4530 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4531 mp->mp_ptrs[indx] = ofs;
4533 mp->mp_lower += sizeof(indx_t);
4535 /* Write the node data. */
4536 node = NODEPTR(mp, indx);
4537 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4538 node->mn_flags = flags;
4540 SETDSZ(node,data->mv_size);
4545 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4550 if (F_ISSET(flags, F_BIGDATA))
4551 memcpy(node->mn_data + key->mv_size, data->mv_data,
4553 else if (F_ISSET(flags, MDB_RESERVE))
4554 data->mv_data = node->mn_data + key->mv_size;
4556 memcpy(node->mn_data + key->mv_size, data->mv_data,
4559 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4561 if (F_ISSET(flags, MDB_RESERVE))
4562 data->mv_data = METADATA(ofp);
4564 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4571 /** Delete the specified node from a page.
4572 * @param[in] mp The page to operate on.
4573 * @param[in] indx The index of the node to delete.
4574 * @param[in] ksize The size of a node. Only used if the page is
4575 * part of a #MDB_DUPFIXED database.
4578 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4581 indx_t i, j, numkeys, ptr;
4588 COPY_PGNO(pgno, mp->mp_pgno);
4589 DPRINTF("delete node %u on %s page %zu", indx,
4590 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4593 assert(indx < NUMKEYS(mp));
4596 int x = NUMKEYS(mp) - 1 - indx;
4597 base = LEAF2KEY(mp, indx, ksize);
4599 memmove(base, base + ksize, x * ksize);
4600 mp->mp_lower -= sizeof(indx_t);
4601 mp->mp_upper += ksize - sizeof(indx_t);
4605 node = NODEPTR(mp, indx);
4606 sz = NODESIZE + node->mn_ksize;
4608 if (F_ISSET(node->mn_flags, F_BIGDATA))
4609 sz += sizeof(pgno_t);
4611 sz += NODEDSZ(node);
4615 ptr = mp->mp_ptrs[indx];
4616 numkeys = NUMKEYS(mp);
4617 for (i = j = 0; i < numkeys; i++) {
4619 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4620 if (mp->mp_ptrs[i] < ptr)
4621 mp->mp_ptrs[j] += sz;
4626 base = (char *)mp + mp->mp_upper;
4627 memmove(base + sz, base, ptr - mp->mp_upper);
4629 mp->mp_lower -= sizeof(indx_t);
4633 /** Compact the main page after deleting a node on a subpage.
4634 * @param[in] mp The main page to operate on.
4635 * @param[in] indx The index of the subpage on the main page.
4638 mdb_node_shrink(MDB_page *mp, indx_t indx)
4645 indx_t i, numkeys, ptr;
4647 node = NODEPTR(mp, indx);
4648 sp = (MDB_page *)NODEDATA(node);
4649 osize = NODEDSZ(node);
4651 delta = sp->mp_upper - sp->mp_lower;
4652 SETDSZ(node, osize - delta);
4653 xp = (MDB_page *)((char *)sp + delta);
4655 /* shift subpage upward */
4657 nsize = NUMKEYS(sp) * sp->mp_pad;
4658 memmove(METADATA(xp), METADATA(sp), nsize);
4661 nsize = osize - sp->mp_upper;
4662 numkeys = NUMKEYS(sp);
4663 for (i=numkeys-1; i>=0; i--)
4664 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4666 xp->mp_upper = sp->mp_lower;
4667 xp->mp_lower = sp->mp_lower;
4668 xp->mp_flags = sp->mp_flags;
4669 xp->mp_pad = sp->mp_pad;
4670 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4672 /* shift lower nodes upward */
4673 ptr = mp->mp_ptrs[indx];
4674 numkeys = NUMKEYS(mp);
4675 for (i = 0; i < numkeys; i++) {
4676 if (mp->mp_ptrs[i] <= ptr)
4677 mp->mp_ptrs[i] += delta;
4680 base = (char *)mp + mp->mp_upper;
4681 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4682 mp->mp_upper += delta;
4685 /** Initial setup of a sorted-dups cursor.
4686 * Sorted duplicates are implemented as a sub-database for the given key.
4687 * The duplicate data items are actually keys of the sub-database.
4688 * Operations on the duplicate data items are performed using a sub-cursor
4689 * initialized when the sub-database is first accessed. This function does
4690 * the preliminary setup of the sub-cursor, filling in the fields that
4691 * depend only on the parent DB.
4692 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4695 mdb_xcursor_init0(MDB_cursor *mc)
4697 MDB_xcursor *mx = mc->mc_xcursor;
4699 mx->mx_cursor.mc_xcursor = NULL;
4700 mx->mx_cursor.mc_txn = mc->mc_txn;
4701 mx->mx_cursor.mc_db = &mx->mx_db;
4702 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
4703 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
4704 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
4705 mx->mx_cursor.mc_snum = 0;
4706 mx->mx_cursor.mc_flags = C_SUB;
4707 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
4708 mx->mx_dbx.md_dcmp = NULL;
4709 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
4712 /** Final setup of a sorted-dups cursor.
4713 * Sets up the fields that depend on the data from the main cursor.
4714 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4715 * @param[in] node The data containing the #MDB_db record for the
4716 * sorted-dup database.
4719 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
4721 MDB_xcursor *mx = mc->mc_xcursor;
4723 if (node->mn_flags & F_SUBDATA) {
4724 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
4725 mx->mx_cursor.mc_snum = 0;
4726 mx->mx_cursor.mc_flags = C_SUB;
4728 MDB_page *fp = NODEDATA(node);
4729 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
4730 mx->mx_db.md_flags = 0;
4731 mx->mx_db.md_depth = 1;
4732 mx->mx_db.md_branch_pages = 0;
4733 mx->mx_db.md_leaf_pages = 1;
4734 mx->mx_db.md_overflow_pages = 0;
4735 mx->mx_db.md_entries = NUMKEYS(fp);
4736 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
4737 mx->mx_cursor.mc_snum = 1;
4738 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
4739 mx->mx_cursor.mc_top = 0;
4740 mx->mx_cursor.mc_pg[0] = fp;
4741 mx->mx_cursor.mc_ki[0] = 0;
4742 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4743 mx->mx_db.md_flags = MDB_DUPFIXED;
4744 mx->mx_db.md_pad = fp->mp_pad;
4745 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4746 mx->mx_db.md_flags |= MDB_INTEGERKEY;
4749 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
4751 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
4753 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
4754 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
4755 #if UINT_MAX < SIZE_MAX
4756 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
4757 #ifdef MISALIGNED_OK
4758 mx->mx_dbx.md_cmp = mdb_cmp_long;
4760 mx->mx_dbx.md_cmp = mdb_cmp_cint;
4765 /** Initialize a cursor for a given transaction and database. */
4767 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
4772 mc->mc_db = &txn->mt_dbs[dbi];
4773 mc->mc_dbx = &txn->mt_dbxs[dbi];
4774 mc->mc_dbflag = &txn->mt_dbflags[dbi];
4778 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4780 mc->mc_xcursor = mx;
4781 mdb_xcursor_init0(mc);
4783 mc->mc_xcursor = NULL;
4788 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
4791 MDB_xcursor *mx = NULL;
4792 size_t size = sizeof(MDB_cursor);
4794 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
4797 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
4798 size += sizeof(MDB_xcursor);
4800 if ((mc = malloc(size)) != NULL) {
4801 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4802 mx = (MDB_xcursor *)(mc + 1);
4804 mdb_cursor_init(mc, txn, dbi, mx);
4805 if (txn->mt_cursors) {
4806 mc->mc_next = txn->mt_cursors[dbi];
4807 txn->mt_cursors[dbi] = mc;
4809 mc->mc_flags |= C_ALLOCD;
4819 /* Return the count of duplicate data items for the current key */
4821 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
4825 if (mc == NULL || countp == NULL)
4828 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
4831 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4832 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4835 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
4838 *countp = mc->mc_xcursor->mx_db.md_entries;
4844 mdb_cursor_close(MDB_cursor *mc)
4847 /* remove from txn, if tracked */
4848 if (mc->mc_txn->mt_cursors) {
4849 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
4850 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
4852 *prev = mc->mc_next;
4854 if (mc->mc_flags & C_ALLOCD)
4860 mdb_cursor_txn(MDB_cursor *mc)
4862 if (!mc) return NULL;
4867 mdb_cursor_dbi(MDB_cursor *mc)
4873 /** Replace the key for a node with a new key.
4874 * @param[in] mp The page containing the node to operate on.
4875 * @param[in] indx The index of the node to operate on.
4876 * @param[in] key The new key to use.
4877 * @return 0 on success, non-zero on failure.
4880 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
4882 indx_t ptr, i, numkeys;
4889 node = NODEPTR(mp, indx);
4890 ptr = mp->mp_ptrs[indx];
4894 char kbuf2[(MAXKEYSIZE*2+1)];
4895 k2.mv_data = NODEKEY(node);
4896 k2.mv_size = node->mn_ksize;
4897 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
4899 mdb_dkey(&k2, kbuf2),
4905 delta = key->mv_size - node->mn_ksize;
4907 if (delta > 0 && SIZELEFT(mp) < delta) {
4908 DPRINTF("OUCH! Not enough room, delta = %d", delta);
4912 numkeys = NUMKEYS(mp);
4913 for (i = 0; i < numkeys; i++) {
4914 if (mp->mp_ptrs[i] <= ptr)
4915 mp->mp_ptrs[i] -= delta;
4918 base = (char *)mp + mp->mp_upper;
4919 len = ptr - mp->mp_upper + NODESIZE;
4920 memmove(base - delta, base, len);
4921 mp->mp_upper -= delta;
4923 node = NODEPTR(mp, indx);
4924 node->mn_ksize = key->mv_size;
4928 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4933 /** Move a node from csrc to cdst.
4936 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
4942 unsigned short flags;
4946 /* Mark src and dst as dirty. */
4947 if ((rc = mdb_page_touch(csrc)) ||
4948 (rc = mdb_page_touch(cdst)))
4951 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
4952 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
4953 key.mv_size = csrc->mc_db->md_pad;
4954 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
4956 data.mv_data = NULL;
4960 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
4961 assert(!((long)srcnode&1));
4962 srcpg = NODEPGNO(srcnode);
4963 flags = srcnode->mn_flags;
4964 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
4965 unsigned int snum = csrc->mc_snum;
4967 /* must find the lowest key below src */
4968 mdb_page_search_root(csrc, NULL, 0);
4969 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
4970 key.mv_size = NODEKSZ(s2);
4971 key.mv_data = NODEKEY(s2);
4972 csrc->mc_snum = snum--;
4973 csrc->mc_top = snum;
4975 key.mv_size = NODEKSZ(srcnode);
4976 key.mv_data = NODEKEY(srcnode);
4978 data.mv_size = NODEDSZ(srcnode);
4979 data.mv_data = NODEDATA(srcnode);
4981 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
4982 unsigned int snum = cdst->mc_snum;
4985 /* must find the lowest key below dst */
4986 mdb_page_search_root(cdst, NULL, 0);
4987 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
4988 bkey.mv_size = NODEKSZ(s2);
4989 bkey.mv_data = NODEKEY(s2);
4990 cdst->mc_snum = snum--;
4991 cdst->mc_top = snum;
4992 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
4995 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
4996 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
4997 csrc->mc_ki[csrc->mc_top],
4999 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5000 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5002 /* Add the node to the destination page.
5004 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5005 if (rc != MDB_SUCCESS)
5008 /* Delete the node from the source page.
5010 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5013 /* Adjust other cursors pointing to mp */
5014 MDB_cursor *m2, *m3;
5015 MDB_dbi dbi = csrc->mc_dbi;
5016 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5018 if (csrc->mc_flags & C_SUB)
5021 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5022 if (m2 == csrc) continue;
5023 if (csrc->mc_flags & C_SUB)
5024 m3 = &m2->mc_xcursor->mx_cursor;
5027 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5028 csrc->mc_ki[csrc->mc_top]) {
5029 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5030 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5035 /* Update the parent separators.
5037 if (csrc->mc_ki[csrc->mc_top] == 0) {
5038 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5039 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5040 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5042 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5043 key.mv_size = NODEKSZ(srcnode);
5044 key.mv_data = NODEKEY(srcnode);
5046 DPRINTF("update separator for source page %zu to [%s]",
5047 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5048 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5049 &key)) != MDB_SUCCESS)
5052 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5054 nullkey.mv_size = 0;
5055 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5056 assert(rc == MDB_SUCCESS);
5060 if (cdst->mc_ki[cdst->mc_top] == 0) {
5061 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5062 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5063 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5065 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5066 key.mv_size = NODEKSZ(srcnode);
5067 key.mv_data = NODEKEY(srcnode);
5069 DPRINTF("update separator for destination page %zu to [%s]",
5070 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5071 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5072 &key)) != MDB_SUCCESS)
5075 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5077 nullkey.mv_size = 0;
5078 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5079 assert(rc == MDB_SUCCESS);
5086 /** Merge one page into another.
5087 * The nodes from the page pointed to by \b csrc will
5088 * be copied to the page pointed to by \b cdst and then
5089 * the \b csrc page will be freed.
5090 * @param[in] csrc Cursor pointing to the source page.
5091 * @param[in] cdst Cursor pointing to the destination page.
5094 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5102 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5103 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5105 assert(csrc->mc_snum > 1); /* can't merge root page */
5106 assert(cdst->mc_snum > 1);
5108 /* Mark dst as dirty. */
5109 if ((rc = mdb_page_touch(cdst)))
5112 /* Move all nodes from src to dst.
5114 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5115 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5116 key.mv_size = csrc->mc_db->md_pad;
5117 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5118 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5119 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5120 if (rc != MDB_SUCCESS)
5122 key.mv_data = (char *)key.mv_data + key.mv_size;
5125 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5126 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5127 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5128 unsigned int snum = csrc->mc_snum;
5130 /* must find the lowest key below src */
5131 mdb_page_search_root(csrc, NULL, 0);
5132 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5133 key.mv_size = NODEKSZ(s2);
5134 key.mv_data = NODEKEY(s2);
5135 csrc->mc_snum = snum--;
5136 csrc->mc_top = snum;
5138 key.mv_size = srcnode->mn_ksize;
5139 key.mv_data = NODEKEY(srcnode);
5142 data.mv_size = NODEDSZ(srcnode);
5143 data.mv_data = NODEDATA(srcnode);
5144 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5145 if (rc != MDB_SUCCESS)
5150 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5151 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);
5153 /* Unlink the src page from parent and add to free list.
5155 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5156 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5158 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5162 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5163 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5164 csrc->mc_db->md_leaf_pages--;
5166 csrc->mc_db->md_branch_pages--;
5168 /* Adjust other cursors pointing to mp */
5169 MDB_cursor *m2, *m3;
5170 MDB_dbi dbi = csrc->mc_dbi;
5171 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5173 if (csrc->mc_flags & C_SUB)
5176 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5177 if (csrc->mc_flags & C_SUB)
5178 m3 = &m2->mc_xcursor->mx_cursor;
5181 if (m3 == csrc) continue;
5182 if (m3->mc_snum < csrc->mc_snum) continue;
5183 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5184 m3->mc_pg[csrc->mc_top] = mp;
5185 m3->mc_ki[csrc->mc_top] += nkeys;
5189 mdb_cursor_pop(csrc);
5191 return mdb_rebalance(csrc);
5194 /** Copy the contents of a cursor.
5195 * @param[in] csrc The cursor to copy from.
5196 * @param[out] cdst The cursor to copy to.
5199 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5203 cdst->mc_txn = csrc->mc_txn;
5204 cdst->mc_dbi = csrc->mc_dbi;
5205 cdst->mc_db = csrc->mc_db;
5206 cdst->mc_dbx = csrc->mc_dbx;
5207 cdst->mc_snum = csrc->mc_snum;
5208 cdst->mc_top = csrc->mc_top;
5209 cdst->mc_flags = csrc->mc_flags;
5211 for (i=0; i<csrc->mc_snum; i++) {
5212 cdst->mc_pg[i] = csrc->mc_pg[i];
5213 cdst->mc_ki[i] = csrc->mc_ki[i];
5217 /** Rebalance the tree after a delete operation.
5218 * @param[in] mc Cursor pointing to the page where rebalancing
5220 * @return 0 on success, non-zero on failure.
5223 mdb_rebalance(MDB_cursor *mc)
5233 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5234 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5235 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5236 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5240 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5243 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5244 DPRINTF("no need to rebalance page %zu, above fill threshold",
5250 if (mc->mc_snum < 2) {
5251 MDB_page *mp = mc->mc_pg[0];
5252 if (NUMKEYS(mp) == 0) {
5253 DPUTS("tree is completely empty");
5254 mc->mc_db->md_root = P_INVALID;
5255 mc->mc_db->md_depth = 0;
5256 mc->mc_db->md_leaf_pages = 0;
5257 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5261 /* Adjust other cursors pointing to mp */
5262 MDB_cursor *m2, *m3;
5263 MDB_dbi dbi = mc->mc_dbi;
5265 if (mc->mc_flags & C_SUB)
5268 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5269 if (m2 == mc) continue;
5270 if (mc->mc_flags & C_SUB)
5271 m3 = &m2->mc_xcursor->mx_cursor;
5274 if (m3->mc_snum < mc->mc_snum) continue;
5275 if (m3->mc_pg[0] == mp) {
5281 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5282 DPUTS("collapsing root page!");
5283 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5284 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5285 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5288 mc->mc_db->md_depth--;
5289 mc->mc_db->md_branch_pages--;
5291 /* Adjust other cursors pointing to mp */
5292 MDB_cursor *m2, *m3;
5293 MDB_dbi dbi = mc->mc_dbi;
5295 if (mc->mc_flags & C_SUB)
5298 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5299 if (m2 == mc) continue;
5300 if (mc->mc_flags & C_SUB)
5301 m3 = &m2->mc_xcursor->mx_cursor;
5304 if (m3->mc_snum < mc->mc_snum) continue;
5305 if (m3->mc_pg[0] == mp) {
5306 m3->mc_pg[0] = mc->mc_pg[0];
5311 DPUTS("root page doesn't need rebalancing");
5315 /* The parent (branch page) must have at least 2 pointers,
5316 * otherwise the tree is invalid.
5318 ptop = mc->mc_top-1;
5319 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5321 /* Leaf page fill factor is below the threshold.
5322 * Try to move keys from left or right neighbor, or
5323 * merge with a neighbor page.
5328 mdb_cursor_copy(mc, &mn);
5329 mn.mc_xcursor = NULL;
5331 if (mc->mc_ki[ptop] == 0) {
5332 /* We're the leftmost leaf in our parent.
5334 DPUTS("reading right neighbor");
5336 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5337 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5339 mn.mc_ki[mn.mc_top] = 0;
5340 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5342 /* There is at least one neighbor to the left.
5344 DPUTS("reading left neighbor");
5346 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5347 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5349 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5350 mc->mc_ki[mc->mc_top] = 0;
5353 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5354 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);
5356 /* If the neighbor page is above threshold and has at least two
5357 * keys, move one key from it.
5359 * Otherwise we should try to merge them.
5361 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5362 return mdb_node_move(&mn, mc);
5363 else { /* FIXME: if (has_enough_room()) */
5364 mc->mc_flags &= ~C_INITIALIZED;
5365 if (mc->mc_ki[ptop] == 0)
5366 return mdb_page_merge(&mn, mc);
5368 return mdb_page_merge(mc, &mn);
5372 /** Complete a delete operation started by #mdb_cursor_del(). */
5374 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5378 /* add overflow pages to free list */
5379 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5383 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5384 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5385 for (i=0; i<ovpages; i++) {
5386 DPRINTF("freed ov page %zu", pg);
5387 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5391 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5392 mc->mc_db->md_entries--;
5393 rc = mdb_rebalance(mc);
5394 if (rc != MDB_SUCCESS)
5395 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5401 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5402 MDB_val *key, MDB_val *data)
5407 MDB_val rdata, *xdata;
5411 assert(key != NULL);
5413 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5415 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5418 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5422 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5426 mdb_cursor_init(&mc, txn, dbi, &mx);
5437 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5439 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5443 /** Split a page and insert a new node.
5444 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5445 * The cursor will be updated to point to the actual page and index where
5446 * the node got inserted after the split.
5447 * @param[in] newkey The key for the newly inserted node.
5448 * @param[in] newdata The data for the newly inserted node.
5449 * @param[in] newpgno The page number, if the new node is a branch node.
5450 * @return 0 on success, non-zero on failure.
5453 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5454 unsigned int nflags)
5457 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0;
5460 unsigned int i, j, split_indx, nkeys, pmax;
5462 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5464 MDB_page *mp, *rp, *pp;
5469 mp = mc->mc_pg[mc->mc_top];
5470 newindx = mc->mc_ki[mc->mc_top];
5472 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5473 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5474 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5476 if (mc->mc_snum < 2) {
5477 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5479 /* shift current top to make room for new parent */
5480 mc->mc_pg[1] = mc->mc_pg[0];
5481 mc->mc_ki[1] = mc->mc_ki[0];
5484 mc->mc_db->md_root = pp->mp_pgno;
5485 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5486 mc->mc_db->md_depth++;
5489 /* Add left (implicit) pointer. */
5490 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5491 /* undo the pre-push */
5492 mc->mc_pg[0] = mc->mc_pg[1];
5493 mc->mc_ki[0] = mc->mc_ki[1];
5494 mc->mc_db->md_root = mp->mp_pgno;
5495 mc->mc_db->md_depth--;
5502 ptop = mc->mc_top-1;
5503 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5506 /* Create a right sibling. */
5507 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5509 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5511 mdb_cursor_copy(mc, &mn);
5512 mn.mc_pg[mn.mc_top] = rp;
5513 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5515 if (nflags & MDB_APPEND) {
5516 mn.mc_ki[mn.mc_top] = 0;
5523 nkeys = NUMKEYS(mp);
5524 split_indx = nkeys / 2 + 1;
5529 unsigned int lsize, rsize, ksize;
5530 /* Move half of the keys to the right sibling */
5532 x = mc->mc_ki[mc->mc_top] - split_indx;
5533 ksize = mc->mc_db->md_pad;
5534 split = LEAF2KEY(mp, split_indx, ksize);
5535 rsize = (nkeys - split_indx) * ksize;
5536 lsize = (nkeys - split_indx) * sizeof(indx_t);
5537 mp->mp_lower -= lsize;
5538 rp->mp_lower += lsize;
5539 mp->mp_upper += rsize - lsize;
5540 rp->mp_upper -= rsize - lsize;
5541 sepkey.mv_size = ksize;
5542 if (newindx == split_indx) {
5543 sepkey.mv_data = newkey->mv_data;
5545 sepkey.mv_data = split;
5548 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5549 memcpy(rp->mp_ptrs, split, rsize);
5550 sepkey.mv_data = rp->mp_ptrs;
5551 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5552 memcpy(ins, newkey->mv_data, ksize);
5553 mp->mp_lower += sizeof(indx_t);
5554 mp->mp_upper -= ksize - sizeof(indx_t);
5557 memcpy(rp->mp_ptrs, split, x * ksize);
5558 ins = LEAF2KEY(rp, x, ksize);
5559 memcpy(ins, newkey->mv_data, ksize);
5560 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5561 rp->mp_lower += sizeof(indx_t);
5562 rp->mp_upper -= ksize - sizeof(indx_t);
5563 mc->mc_ki[mc->mc_top] = x;
5564 mc->mc_pg[mc->mc_top] = rp;
5569 /* For leaf pages, check the split point based on what
5570 * fits where, since otherwise add_node can fail.
5573 unsigned int psize, nsize;
5574 /* Maximum free space in an empty page */
5575 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5576 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5577 if (newindx < split_indx) {
5579 for (i=0; i<split_indx; i++) {
5580 node = NODEPTR(mp, i);
5581 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5582 if (F_ISSET(node->mn_flags, F_BIGDATA))
5583 psize += sizeof(pgno_t);
5585 psize += NODEDSZ(node);
5594 for (i=nkeys-1; i>=split_indx; i--) {
5595 node = NODEPTR(mp, i);
5596 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5597 if (F_ISSET(node->mn_flags, F_BIGDATA))
5598 psize += sizeof(pgno_t);
5600 psize += NODEDSZ(node);
5610 /* First find the separating key between the split pages.
5612 if (newindx == split_indx) {
5613 sepkey.mv_size = newkey->mv_size;
5614 sepkey.mv_data = newkey->mv_data;
5616 node = NODEPTR(mp, split_indx);
5617 sepkey.mv_size = node->mn_ksize;
5618 sepkey.mv_data = NODEKEY(node);
5622 DPRINTF("separator is [%s]", DKEY(&sepkey));
5624 /* Copy separator key to the parent.
5626 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5629 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5631 /* Right page might now have changed parent.
5632 * Check if left page also changed parent.
5634 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
5635 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
5636 mc->mc_pg[ptop] = mn.mc_pg[ptop];
5637 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
5641 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
5644 if (rc != MDB_SUCCESS) {
5647 if (nflags & MDB_APPEND) {
5648 mc->mc_pg[mc->mc_top] = rp;
5649 mc->mc_ki[mc->mc_top] = 0;
5650 return mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
5656 /* Move half of the keys to the right sibling. */
5658 /* grab a page to hold a temporary copy */
5659 copy = mdb_page_malloc(mc);
5663 copy->mp_pgno = mp->mp_pgno;
5664 copy->mp_flags = mp->mp_flags;
5665 copy->mp_lower = PAGEHDRSZ;
5666 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
5667 mc->mc_pg[mc->mc_top] = copy;
5668 for (i = j = 0; i <= nkeys; j++) {
5669 if (i == split_indx) {
5670 /* Insert in right sibling. */
5671 /* Reset insert index for right sibling. */
5672 j = (i == newindx && ins_new);
5673 mc->mc_pg[mc->mc_top] = rp;
5676 if (i == newindx && !ins_new) {
5677 /* Insert the original entry that caused the split. */
5678 rkey.mv_data = newkey->mv_data;
5679 rkey.mv_size = newkey->mv_size;
5688 /* Update page and index for the new key. */
5690 mc->mc_pg[mc->mc_top] = copy;
5691 mc->mc_ki[mc->mc_top] = j;
5692 } else if (i == nkeys) {
5695 node = NODEPTR(mp, i);
5696 rkey.mv_data = NODEKEY(node);
5697 rkey.mv_size = node->mn_ksize;
5699 xdata.mv_data = NODEDATA(node);
5700 xdata.mv_size = NODEDSZ(node);
5703 pgno = NODEPGNO(node);
5704 flags = node->mn_flags;
5709 if (!IS_LEAF(mp) && j == 0) {
5710 /* First branch index doesn't need key data. */
5714 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
5717 nkeys = NUMKEYS(copy);
5718 for (i=0; i<nkeys; i++)
5719 mp->mp_ptrs[i] = copy->mp_ptrs[i];
5720 mp->mp_lower = copy->mp_lower;
5721 mp->mp_upper = copy->mp_upper;
5722 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
5723 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
5725 /* reset back to original page */
5726 if (!newindx || (newindx < split_indx)) {
5727 mc->mc_pg[mc->mc_top] = mp;
5728 if (nflags & MDB_RESERVE) {
5729 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5730 if (!(node->mn_flags & F_BIGDATA))
5731 newdata->mv_data = NODEDATA(node);
5735 /* return tmp page to freelist */
5736 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
5737 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
5738 mc->mc_txn->mt_env->me_dpages = copy;
5741 /* Adjust other cursors pointing to mp */
5742 MDB_cursor *m2, *m3;
5743 MDB_dbi dbi = mc->mc_dbi;
5745 if (mc->mc_flags & C_SUB)
5748 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5749 if (m2 == mc) continue;
5750 if (mc->mc_flags & C_SUB)
5751 m3 = &m2->mc_xcursor->mx_cursor;
5754 if (!(m3->mc_flags & C_INITIALIZED))
5758 for (i=m3->mc_top; i>0; i--) {
5759 m3->mc_ki[i+1] = m3->mc_ki[i];
5760 m3->mc_pg[i+1] = m3->mc_pg[i];
5762 m3->mc_ki[0] = mc->mc_ki[0];
5763 m3->mc_pg[0] = mc->mc_pg[0];
5767 if (m3->mc_pg[mc->mc_top] == mp) {
5768 if (m3->mc_ki[m3->mc_top] >= split_indx) {
5769 m3->mc_pg[m3->mc_top] = rp;
5770 m3->mc_ki[m3->mc_top] -= split_indx;
5779 mdb_put(MDB_txn *txn, MDB_dbi dbi,
5780 MDB_val *key, MDB_val *data, unsigned int flags)
5785 assert(key != NULL);
5786 assert(data != NULL);
5788 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5791 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5795 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5799 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
5802 mdb_cursor_init(&mc, txn, dbi, &mx);
5803 return mdb_cursor_put(&mc, key, data, flags);
5806 /** Only a subset of the @ref mdb_env flags can be changed
5807 * at runtime. Changing other flags requires closing the environment
5808 * and re-opening it with the new flags.
5810 #define CHANGEABLE (MDB_NOSYNC)
5812 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
5814 if ((flag & CHANGEABLE) != flag)
5817 env->me_flags |= flag;
5819 env->me_flags &= ~flag;
5824 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
5829 *arg = env->me_flags;
5834 mdb_env_get_path(MDB_env *env, const char **arg)
5839 *arg = env->me_path;
5843 /** Common code for #mdb_stat() and #mdb_env_stat().
5844 * @param[in] env the environment to operate in.
5845 * @param[in] db the #MDB_db record containing the stats to return.
5846 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
5847 * @return 0, this function always succeeds.
5850 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
5852 arg->ms_psize = env->me_psize;
5853 arg->ms_depth = db->md_depth;
5854 arg->ms_branch_pages = db->md_branch_pages;
5855 arg->ms_leaf_pages = db->md_leaf_pages;
5856 arg->ms_overflow_pages = db->md_overflow_pages;
5857 arg->ms_entries = db->md_entries;
5862 mdb_env_stat(MDB_env *env, MDB_stat *arg)
5866 if (env == NULL || arg == NULL)
5869 mdb_env_read_meta(env, &toggle);
5871 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
5874 /** Set the default comparison functions for a database.
5875 * Called immediately after a database is opened to set the defaults.
5876 * The user can then override them with #mdb_set_compare() or
5877 * #mdb_set_dupsort().
5878 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
5879 * @param[in] dbi A database handle returned by #mdb_open()
5882 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
5884 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
5885 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memnr;
5886 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
5887 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_cint;
5889 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memn;
5891 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5892 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
5893 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
5894 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_int;
5896 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_cint;
5897 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
5898 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memnr;
5900 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memn;
5903 txn->mt_dbxs[dbi].md_dcmp = NULL;
5907 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
5912 int rc, dbflag, exact;
5915 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
5916 mdb_default_cmp(txn, FREE_DBI);
5922 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
5923 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
5924 mdb_default_cmp(txn, MAIN_DBI);
5928 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
5929 mdb_default_cmp(txn, MAIN_DBI);
5932 /* Is the DB already open? */
5934 for (i=2; i<txn->mt_numdbs; i++) {
5935 if (len == txn->mt_dbxs[i].md_name.mv_size &&
5936 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
5942 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
5945 /* Find the DB info */
5949 key.mv_data = (void *)name;
5950 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
5951 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
5952 if (rc == MDB_SUCCESS) {
5953 /* make sure this is actually a DB */
5954 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
5955 if (!(node->mn_flags & F_SUBDATA))
5957 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
5958 /* Create if requested */
5960 data.mv_size = sizeof(MDB_db);
5961 data.mv_data = &dummy;
5962 memset(&dummy, 0, sizeof(dummy));
5963 dummy.md_root = P_INVALID;
5964 dummy.md_flags = flags & 0xffff;
5965 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
5969 /* OK, got info, add to table */
5970 if (rc == MDB_SUCCESS) {
5971 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
5972 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
5973 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
5974 txn->mt_dbflags[txn->mt_numdbs] = dbflag;
5975 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
5976 *dbi = txn->mt_numdbs;
5977 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
5978 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
5979 mdb_default_cmp(txn, txn->mt_numdbs);
5986 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
5988 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
5991 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
5994 void mdb_close(MDB_env *env, MDB_dbi dbi)
5997 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
5999 ptr = env->me_dbxs[dbi].md_name.mv_data;
6000 env->me_dbxs[dbi].md_name.mv_data = NULL;
6001 env->me_dbxs[dbi].md_name.mv_size = 0;
6005 /** Add all the DB's pages to the free list.
6006 * @param[in] mc Cursor on the DB to free.
6007 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6008 * @return 0 on success, non-zero on failure.
6011 mdb_drop0(MDB_cursor *mc, int subs)
6015 rc = mdb_page_search(mc, NULL, 0);
6016 if (rc == MDB_SUCCESS) {
6021 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6022 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6025 mdb_cursor_copy(mc, &mx);
6026 while (mc->mc_snum > 0) {
6027 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6028 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6029 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6030 if (ni->mn_flags & F_SUBDATA) {
6031 mdb_xcursor_init1(mc, ni);
6032 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6038 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6040 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6043 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6048 rc = mdb_cursor_sibling(mc, 1);
6050 /* no more siblings, go back to beginning
6051 * of previous level. (stack was already popped
6052 * by mdb_cursor_sibling)
6054 for (i=1; i<mc->mc_top; i++)
6055 mc->mc_pg[i] = mx.mc_pg[i];
6059 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6060 mc->mc_db->md_root);
6065 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6070 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6073 rc = mdb_cursor_open(txn, dbi, &mc);
6077 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6081 /* Can't delete the main DB */
6082 if (del && dbi > MAIN_DBI) {
6083 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6085 mdb_close(txn->mt_env, dbi);
6087 txn->mt_dbflags[dbi] |= DB_DIRTY;
6088 txn->mt_dbs[dbi].md_depth = 0;
6089 txn->mt_dbs[dbi].md_branch_pages = 0;
6090 txn->mt_dbs[dbi].md_leaf_pages = 0;
6091 txn->mt_dbs[dbi].md_overflow_pages = 0;
6092 txn->mt_dbs[dbi].md_entries = 0;
6093 txn->mt_dbs[dbi].md_root = P_INVALID;
6096 mdb_cursor_close(mc);
6100 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6102 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6105 txn->mt_dbxs[dbi].md_cmp = cmp;
6109 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6111 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6114 txn->mt_dbxs[dbi].md_dcmp = cmp;
6118 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6120 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6123 txn->mt_dbxs[dbi].md_rel = rel;
6127 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6129 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6132 txn->mt_dbxs[dbi].md_relctx = ctx;