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 /** Buffer for a stack-allocated dirty page.
768 * The members define size and alignment, and silence type
769 * aliasing warnings. They are not used directly; that could
770 * mean incorrectly using several union members in parallel.
772 typedef union MDB_pagebuf {
773 char mb_raw[MDB_PAGESIZE];
776 char mm_pad[PAGEHDRSZ];
781 /** Auxiliary DB info.
782 * The information here is mostly static/read-only. There is
783 * only a single copy of this record in the environment.
785 typedef struct MDB_dbx {
786 MDB_val md_name; /**< name of the database */
787 MDB_cmp_func *md_cmp; /**< function for comparing keys */
788 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
789 MDB_rel_func *md_rel; /**< user relocate function */
790 void *md_relctx; /**< user-provided context for md_rel */
793 /** A database transaction.
794 * Every operation requires a transaction handle.
797 MDB_txn *mt_parent; /**< parent of a nested txn */
798 MDB_txn *mt_child; /**< nested txn under this txn */
799 pgno_t mt_next_pgno; /**< next unallocated page */
800 /** The ID of this transaction. IDs are integers incrementing from 1.
801 * Only committed write transactions increment the ID. If a transaction
802 * aborts, the ID may be re-used by the next writer.
805 MDB_env *mt_env; /**< the DB environment */
806 /** The list of pages that became unused during this transaction.
810 ID2L dirty_list; /**< modified pages */
811 MDB_reader *reader; /**< this thread's slot in the reader table */
813 /** Array of records for each DB known in the environment. */
815 /** Array of MDB_db records for each known DB */
817 /** @defgroup mt_dbflag Transaction DB Flags
821 #define DB_DIRTY 0x01 /**< DB was written in this txn */
822 #define DB_STALE 0x02 /**< DB record is older than txnID */
824 /** Array of cursors for each DB */
825 MDB_cursor **mt_cursors;
826 /** Array of flags for each DB */
827 unsigned char *mt_dbflags;
828 /** Number of DB records in use. This number only ever increments;
829 * we don't decrement it when individual DB handles are closed.
833 /** @defgroup mdb_txn Transaction Flags
837 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
838 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
840 unsigned int mt_flags; /**< @ref mdb_txn */
841 /** Tracks which of the two meta pages was used at the start
842 * of this transaction.
844 unsigned int mt_toggle;
847 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
848 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
849 * raise this on a 64 bit machine.
851 #define CURSOR_STACK 32
855 /** Cursors are used for all DB operations */
857 /** Next cursor on this DB in this txn */
859 /** Original cursor if this is a shadow */
861 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
862 struct MDB_xcursor *mc_xcursor;
863 /** The transaction that owns this cursor */
865 /** The database handle this cursor operates on */
867 /** The database record for this cursor */
869 /** The database auxiliary record for this cursor */
871 /** The @ref mt_dbflag for this database */
872 unsigned char *mc_dbflag;
873 unsigned short mc_snum; /**< number of pushed pages */
874 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
875 /** @defgroup mdb_cursor Cursor Flags
877 * Cursor state flags.
880 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
881 #define C_EOF 0x02 /**< No more data */
882 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
883 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
884 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
886 unsigned int mc_flags; /**< @ref mdb_cursor */
887 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
888 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
891 /** Context for sorted-dup records.
892 * We could have gone to a fully recursive design, with arbitrarily
893 * deep nesting of sub-databases. But for now we only handle these
894 * levels - main DB, optional sub-DB, sorted-duplicate DB.
896 typedef struct MDB_xcursor {
897 /** A sub-cursor for traversing the Dup DB */
898 MDB_cursor mx_cursor;
899 /** The database record for this Dup DB */
901 /** The auxiliary DB record for this Dup DB */
903 /** The @ref mt_dbflag for this Dup DB */
904 unsigned char mx_dbflag;
907 /** A set of pages freed by an earlier transaction. */
908 typedef struct MDB_oldpages {
909 /** Usually we only read one record from the FREEDB at a time, but
910 * in case we read more, this will chain them together.
912 struct MDB_oldpages *mo_next;
913 /** The ID of the transaction in which these pages were freed. */
915 /** An #IDL of the pages */
916 pgno_t mo_pages[1]; /* dynamic */
919 /** The database environment. */
921 HANDLE me_fd; /**< The main data file */
922 HANDLE me_lfd; /**< The lock file */
923 HANDLE me_mfd; /**< just for writing the meta pages */
924 /** Failed to update the meta page. Probably an I/O error. */
925 #define MDB_FATAL_ERROR 0x80000000U
926 uint32_t me_flags; /**< @ref mdb_env */
927 uint32_t me_extrapad; /**< unused for now */
928 unsigned int me_maxreaders; /**< size of the reader table */
929 MDB_dbi me_numdbs; /**< number of DBs opened */
930 MDB_dbi me_maxdbs; /**< size of the DB table */
931 char *me_path; /**< path to the DB files */
932 char *me_map; /**< the memory map of the data file */
933 MDB_txninfo *me_txns; /**< the memory map of the lock file */
934 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
935 MDB_txn *me_txn; /**< current write transaction */
936 size_t me_mapsize; /**< size of the data memory map */
937 off_t me_size; /**< current file size */
938 pgno_t me_maxpg; /**< me_mapsize / me_psize */
939 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
940 unsigned int me_db_toggle; /**< which DB table is current */
941 txnid_t me_wtxnid; /**< ID of last txn we committed */
942 MDB_dbx *me_dbxs; /**< array of static DB info */
943 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
944 MDB_oldpages *me_pghead; /**< list of old page records */
945 pthread_key_t me_txkey; /**< thread-key for readers */
946 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
947 /** IDL of pages that became unused in a write txn */
949 /** ID2L of pages that were written during a write txn */
950 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
951 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
952 LAZY_RWLOCK_DEF(me_dblock)
954 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
958 sem_t *me_rmutex; /* Apple doesn't support shared mutexes */
962 /** max number of pages to commit in one writev() call */
963 #define MDB_COMMIT_PAGES 64
964 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
965 #undef MDB_COMMIT_PAGES
966 #define MDB_COMMIT_PAGES IOV_MAX
969 static MDB_page *mdb_page_alloc(MDB_cursor *mc, int num);
970 static MDB_page *mdb_page_new(MDB_cursor *mc, uint32_t flags, int num);
971 static int mdb_page_touch(MDB_cursor *mc);
973 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
974 static int mdb_page_search_root(MDB_cursor *mc,
975 MDB_val *key, int modify);
976 static int mdb_page_search(MDB_cursor *mc,
977 MDB_val *key, int modify);
978 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
979 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
980 pgno_t newpgno, unsigned int nflags);
982 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
983 static int mdb_env_read_meta(MDB_env *env, int *which);
984 static int mdb_env_write_meta(MDB_txn *txn);
986 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
987 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
988 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
989 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
990 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
991 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
992 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
993 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
994 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
996 static int mdb_rebalance(MDB_cursor *mc);
997 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
999 static void mdb_cursor_pop(MDB_cursor *mc);
1000 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1002 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1003 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1004 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1005 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1006 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1008 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1009 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1011 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1012 static void mdb_xcursor_init0(MDB_cursor *mc);
1013 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1015 static int mdb_drop0(MDB_cursor *mc, int subs);
1016 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1019 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1023 static SECURITY_DESCRIPTOR mdb_null_sd;
1024 static SECURITY_ATTRIBUTES mdb_all_sa;
1025 static int mdb_sec_inited;
1028 /** Return the library version info. */
1030 mdb_version(int *major, int *minor, int *patch)
1032 if (major) *major = MDB_VERSION_MAJOR;
1033 if (minor) *minor = MDB_VERSION_MINOR;
1034 if (patch) *patch = MDB_VERSION_PATCH;
1035 return MDB_VERSION_STRING;
1038 /** Table of descriptions for MDB @ref errors */
1039 static char *const mdb_errstr[] = {
1040 "MDB_KEYEXIST: Key/data pair already exists",
1041 "MDB_NOTFOUND: No matching key/data pair found",
1042 "MDB_PAGE_NOTFOUND: Requested page not found",
1043 "MDB_CORRUPTED: Located page was wrong type",
1044 "MDB_PANIC: Update of meta page failed",
1045 "MDB_VERSION_MISMATCH: Database environment version mismatch"
1049 mdb_strerror(int err)
1052 return ("Successful return: 0");
1054 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
1055 return mdb_errstr[err - MDB_KEYEXIST];
1057 return strerror(err);
1061 /** Display a key in hexadecimal and return the address of the result.
1062 * @param[in] key the key to display
1063 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1064 * @return The key in hexadecimal form.
1067 mdb_dkey(MDB_val *key, char *buf)
1070 unsigned char *c = key->mv_data;
1072 if (key->mv_size > MAXKEYSIZE)
1073 return "MAXKEYSIZE";
1074 /* may want to make this a dynamic check: if the key is mostly
1075 * printable characters, print it as-is instead of converting to hex.
1079 for (i=0; i<key->mv_size; i++)
1080 ptr += sprintf(ptr, "%02x", *c++);
1082 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1087 /** Display all the keys in the page. */
1089 mdb_page_keys(MDB_page *mp)
1092 unsigned int i, nkeys;
1096 nkeys = NUMKEYS(mp);
1097 DPRINTF("numkeys %d", nkeys);
1098 for (i=0; i<nkeys; i++) {
1099 node = NODEPTR(mp, i);
1100 key.mv_size = node->mn_ksize;
1101 key.mv_data = node->mn_data;
1102 DPRINTF("key %d: %s", i, DKEY(&key));
1108 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1110 return txn->mt_dbxs[dbi].md_cmp(a, b);
1114 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1116 if (txn->mt_dbxs[dbi].md_dcmp)
1117 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1119 return EINVAL; /* too bad you can't distinguish this from a valid result */
1122 /** Allocate a single page.
1123 * Re-use old malloc'd pages first, otherwise just malloc.
1126 mdb_page_malloc(MDB_cursor *mc) {
1128 size_t sz = mc->mc_txn->mt_env->me_psize;
1129 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1130 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1131 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1132 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1133 } else if ((ret = malloc(sz)) != NULL) {
1134 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1139 /** Allocate pages for writing.
1140 * If there are free pages available from older transactions, they
1141 * will be re-used first. Otherwise a new page will be allocated.
1142 * @param[in] mc cursor A cursor handle identifying the transaction and
1143 * database for which we are allocating.
1144 * @param[in] num the number of pages to allocate.
1145 * @return Address of the allocated page(s). Requests for multiple pages
1146 * will always be satisfied by a single contiguous chunk of memory.
1149 mdb_page_alloc(MDB_cursor *mc, int num)
1151 MDB_txn *txn = mc->mc_txn;
1153 pgno_t pgno = P_INVALID;
1156 if (txn->mt_txnid > 2) {
1158 if (!txn->mt_env->me_pghead && mc->mc_dbi != FREE_DBI &&
1159 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1160 /* See if there's anything in the free DB */
1163 txnid_t *kptr, oldest;
1165 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1166 mdb_page_search(&m2, NULL, 0);
1167 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1168 kptr = (txnid_t *)NODEKEY(leaf);
1172 oldest = txn->mt_txnid - 1;
1173 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1174 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1175 if (mr && mr < oldest)
1180 if (oldest > *kptr) {
1181 /* It's usable, grab it.
1187 mdb_node_read(txn, leaf, &data);
1188 idl = (ID *) data.mv_data;
1189 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1190 mop->mo_next = txn->mt_env->me_pghead;
1191 mop->mo_txnid = *kptr;
1192 txn->mt_env->me_pghead = mop;
1193 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1198 DPRINTF("IDL read txn %zu root %zu num %zu",
1199 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1200 for (i=0; i<idl[0]; i++) {
1201 DPRINTF("IDL %zu", idl[i+1]);
1205 /* drop this IDL from the DB */
1206 m2.mc_ki[m2.mc_top] = 0;
1207 m2.mc_flags = C_INITIALIZED;
1208 mdb_cursor_del(&m2, 0);
1211 if (txn->mt_env->me_pghead) {
1212 MDB_oldpages *mop = txn->mt_env->me_pghead;
1214 /* FIXME: For now, always use fresh pages. We
1215 * really ought to search the free list for a
1220 /* peel pages off tail, so we only have to truncate the list */
1221 pgno = MDB_IDL_LAST(mop->mo_pages);
1222 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1224 if (mop->mo_pages[2] > mop->mo_pages[1])
1225 mop->mo_pages[0] = 0;
1229 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1230 txn->mt_env->me_pghead = mop->mo_next;
1237 if (pgno == P_INVALID) {
1238 /* DB size is maxed out */
1239 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1240 DPUTS("DB size maxed out");
1244 if (txn->mt_env->me_dpages && num == 1) {
1245 np = txn->mt_env->me_dpages;
1246 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1247 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1248 txn->mt_env->me_dpages = np->mp_next;
1250 size_t sz = txn->mt_env->me_psize * num;
1251 if ((np = malloc(sz)) == NULL)
1253 VGMEMP_ALLOC(txn->mt_env, np, sz);
1255 if (pgno == P_INVALID) {
1256 np->mp_pgno = txn->mt_next_pgno;
1257 txn->mt_next_pgno += num;
1261 mid.mid = np->mp_pgno;
1263 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1268 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1269 * @param[in] mc cursor pointing to the page to be touched
1270 * @return 0 on success, non-zero on failure.
1273 mdb_page_touch(MDB_cursor *mc)
1275 MDB_page *mp = mc->mc_pg[mc->mc_top];
1278 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1280 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1282 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1283 assert(mp->mp_pgno != np->mp_pgno);
1284 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1286 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1289 mp->mp_flags |= P_DIRTY;
1292 /* Adjust other cursors pointing to mp */
1293 if (mc->mc_flags & C_SUB) {
1294 MDB_cursor *m2, *m3;
1295 MDB_dbi dbi = mc->mc_dbi-1;
1297 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1298 if (m2 == mc) continue;
1299 m3 = &m2->mc_xcursor->mx_cursor;
1300 if (m3->mc_snum < mc->mc_snum) continue;
1301 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1302 m3->mc_pg[mc->mc_top] = mp;
1308 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1309 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1310 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1311 m2->mc_pg[mc->mc_top] = mp;
1315 mc->mc_pg[mc->mc_top] = mp;
1316 /** If this page has a parent, update the parent to point to
1320 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1322 mc->mc_db->md_root = mp->mp_pgno;
1323 } else if (mc->mc_txn->mt_parent) {
1326 /* If txn has a parent, make sure the page is in our
1329 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1330 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1331 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1332 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1333 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1334 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1335 mc->mc_pg[mc->mc_top] = mp;
1341 np = mdb_page_malloc(mc);
1342 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1343 mid.mid = np->mp_pgno;
1345 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1353 mdb_env_sync(MDB_env *env, int force)
1356 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1357 if (MDB_FDATASYNC(env->me_fd))
1363 /** Make shadow copies of all of parent txn's cursors */
1365 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1367 MDB_cursor *mc, *m2;
1368 unsigned int i, j, size;
1370 for (i=0;i<src->mt_numdbs; i++) {
1371 if (src->mt_cursors[i]) {
1372 size = sizeof(MDB_cursor);
1373 if (src->mt_cursors[i]->mc_xcursor)
1374 size += sizeof(MDB_xcursor);
1375 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1382 mc->mc_db = &dst->mt_dbs[i];
1383 mc->mc_dbx = m2->mc_dbx;
1384 mc->mc_dbflag = &dst->mt_dbflags[i];
1385 mc->mc_snum = m2->mc_snum;
1386 mc->mc_top = m2->mc_top;
1387 mc->mc_flags = m2->mc_flags | C_SHADOW;
1388 for (j=0; j<mc->mc_snum; j++) {
1389 mc->mc_pg[j] = m2->mc_pg[j];
1390 mc->mc_ki[j] = m2->mc_ki[j];
1392 if (m2->mc_xcursor) {
1393 MDB_xcursor *mx, *mx2;
1394 mx = (MDB_xcursor *)(mc+1);
1395 mc->mc_xcursor = mx;
1396 mx2 = m2->mc_xcursor;
1397 mx->mx_db = mx2->mx_db;
1398 mx->mx_dbx = mx2->mx_dbx;
1399 mx->mx_dbflag = mx2->mx_dbflag;
1400 mx->mx_cursor.mc_txn = dst;
1401 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1402 mx->mx_cursor.mc_db = &mx->mx_db;
1403 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1404 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1405 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1406 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1407 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1408 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1409 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1410 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1413 mc->mc_xcursor = NULL;
1415 mc->mc_next = dst->mt_cursors[i];
1416 dst->mt_cursors[i] = mc;
1423 /** Merge shadow cursors back into parent's */
1425 mdb_cursor_merge(MDB_txn *txn)
1428 for (i=0; i<txn->mt_numdbs; i++) {
1429 if (txn->mt_cursors[i]) {
1431 while ((mc = txn->mt_cursors[i])) {
1432 txn->mt_cursors[i] = mc->mc_next;
1433 if (mc->mc_flags & C_SHADOW) {
1434 MDB_cursor *m2 = mc->mc_orig;
1436 m2->mc_snum = mc->mc_snum;
1437 m2->mc_top = mc->mc_top;
1438 for (j=0; j<mc->mc_snum; j++) {
1439 m2->mc_pg[j] = mc->mc_pg[j];
1440 m2->mc_ki[j] = mc->mc_ki[j];
1443 if (mc->mc_flags & C_ALLOCD)
1451 mdb_txn_reset0(MDB_txn *txn);
1453 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1454 * @param[in] txn the transaction handle to initialize
1455 * @return 0 on success, non-zero on failure. This can only
1456 * fail for read-only transactions, and then only if the
1457 * reader table is full.
1460 mdb_txn_renew0(MDB_txn *txn)
1462 MDB_env *env = txn->mt_env;
1465 if (txn->mt_flags & MDB_TXN_RDONLY) {
1466 MDB_reader *r = pthread_getspecific(env->me_txkey);
1469 pid_t pid = getpid();
1470 pthread_t tid = pthread_self();
1473 for (i=0; i<env->me_txns->mti_numreaders; i++)
1474 if (env->me_txns->mti_readers[i].mr_pid == 0)
1476 if (i == env->me_maxreaders) {
1477 UNLOCK_MUTEX_R(env);
1480 env->me_txns->mti_readers[i].mr_pid = pid;
1481 env->me_txns->mti_readers[i].mr_tid = tid;
1482 if (i >= env->me_txns->mti_numreaders)
1483 env->me_txns->mti_numreaders = i+1;
1484 UNLOCK_MUTEX_R(env);
1485 r = &env->me_txns->mti_readers[i];
1486 pthread_setspecific(env->me_txkey, r);
1488 txn->mt_toggle = env->me_txns->mti_me_toggle;
1489 txn->mt_txnid = env->me_txns->mti_txnid;
1490 /* This happens if a different process was the
1491 * last writer to the DB.
1493 if (env->me_wtxnid < txn->mt_txnid)
1494 mt_dbflag = DB_STALE;
1495 r->mr_txnid = txn->mt_txnid;
1496 txn->mt_u.reader = r;
1500 txn->mt_txnid = env->me_txns->mti_txnid;
1501 if (env->me_wtxnid < txn->mt_txnid)
1502 mt_dbflag = DB_STALE;
1504 txn->mt_toggle = env->me_txns->mti_me_toggle;
1505 txn->mt_u.dirty_list = env->me_dirty_list;
1506 txn->mt_u.dirty_list[0].mid = 0;
1507 txn->mt_free_pgs = env->me_free_pgs;
1508 txn->mt_free_pgs[0] = 0;
1509 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1513 /* Copy the DB arrays */
1514 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1515 txn->mt_numdbs = env->me_numdbs;
1516 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1517 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1518 if (txn->mt_numdbs > 2)
1519 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1520 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1521 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1523 memset(txn->mt_dbflags, mt_dbflag, env->me_numdbs);
1529 mdb_txn_renew(MDB_txn *txn)
1536 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1537 DPUTS("environment had fatal error, must shutdown!");
1541 rc = mdb_txn_renew0(txn);
1542 if (rc == MDB_SUCCESS) {
1543 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1544 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1545 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1551 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1556 if (env->me_flags & MDB_FATAL_ERROR) {
1557 DPUTS("environment had fatal error, must shutdown!");
1561 /* parent already has an active child txn */
1562 if (parent->mt_child) {
1566 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1567 if (!(flags & MDB_RDONLY))
1568 size += env->me_maxdbs * sizeof(MDB_cursor *);
1570 if ((txn = calloc(1, size)) == NULL) {
1571 DPRINTF("calloc: %s", strerror(ErrCode()));
1574 txn->mt_dbs = (MDB_db *)(txn+1);
1575 if (flags & MDB_RDONLY) {
1576 txn->mt_flags |= MDB_TXN_RDONLY;
1577 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1579 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1580 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1585 txn->mt_free_pgs = mdb_midl_alloc();
1586 if (!txn->mt_free_pgs) {
1590 txn->mt_u.dirty_list = malloc(sizeof(ID2)*MDB_IDL_UM_SIZE);
1591 if (!txn->mt_u.dirty_list) {
1592 free(txn->mt_free_pgs);
1596 txn->mt_txnid = parent->mt_txnid;
1597 txn->mt_toggle = parent->mt_toggle;
1598 txn->mt_u.dirty_list[0].mid = 0;
1599 txn->mt_free_pgs[0] = 0;
1600 txn->mt_next_pgno = parent->mt_next_pgno;
1601 parent->mt_child = txn;
1602 txn->mt_parent = parent;
1603 txn->mt_numdbs = parent->mt_numdbs;
1604 txn->mt_dbxs = parent->mt_dbxs;
1605 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1606 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1607 mdb_cursor_shadow(parent, txn);
1610 rc = mdb_txn_renew0(txn);
1616 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1617 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1618 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1624 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1625 * @param[in] txn the transaction handle to reset
1628 mdb_txn_reset0(MDB_txn *txn)
1630 MDB_env *env = txn->mt_env;
1632 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1633 txn->mt_u.reader->mr_txnid = 0;
1639 /* close(free) all cursors */
1640 for (i=0; i<txn->mt_numdbs; i++) {
1641 if (txn->mt_cursors[i]) {
1643 while ((mc = txn->mt_cursors[i])) {
1644 txn->mt_cursors[i] = mc->mc_next;
1645 if (mc->mc_flags & C_ALLOCD)
1651 /* return all dirty pages to dpage list */
1652 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1653 dp = txn->mt_u.dirty_list[i].mptr;
1654 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1655 dp->mp_next = txn->mt_env->me_dpages;
1656 VGMEMP_FREE(txn->mt_env, dp);
1657 txn->mt_env->me_dpages = dp;
1659 /* large pages just get freed directly */
1660 VGMEMP_FREE(txn->mt_env, dp);
1665 if (txn->mt_parent) {
1666 txn->mt_parent->mt_child = NULL;
1667 free(txn->mt_free_pgs);
1668 free(txn->mt_u.dirty_list);
1671 if (mdb_midl_shrink(&txn->mt_free_pgs))
1672 env->me_free_pgs = txn->mt_free_pgs;
1675 while ((mop = txn->mt_env->me_pghead)) {
1676 txn->mt_env->me_pghead = mop->mo_next;
1681 /* The writer mutex was locked in mdb_txn_begin. */
1682 UNLOCK_MUTEX_W(env);
1687 mdb_txn_reset(MDB_txn *txn)
1692 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1693 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1694 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1696 mdb_txn_reset0(txn);
1700 mdb_txn_abort(MDB_txn *txn)
1705 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1706 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1707 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1710 mdb_txn_abort(txn->mt_child);
1712 mdb_txn_reset0(txn);
1717 mdb_txn_commit(MDB_txn *txn)
1728 assert(txn != NULL);
1729 assert(txn->mt_env != NULL);
1731 if (txn->mt_child) {
1732 mdb_txn_commit(txn->mt_child);
1733 txn->mt_child = NULL;
1738 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1739 if (txn->mt_numdbs > env->me_numdbs) {
1740 /* update the DB tables */
1741 int toggle = !env->me_db_toggle;
1745 ip = &env->me_dbs[toggle][env->me_numdbs];
1746 jp = &txn->mt_dbs[env->me_numdbs];
1747 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1748 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
1752 env->me_db_toggle = toggle;
1753 env->me_numdbs = txn->mt_numdbs;
1754 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1760 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1761 DPUTS("error flag is set, can't commit");
1763 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1768 /* Merge (and close) our cursors with parent's */
1769 mdb_cursor_merge(txn);
1771 if (txn->mt_parent) {
1777 /* Update parent's DB table */
1778 ip = &txn->mt_parent->mt_dbs[2];
1779 jp = &txn->mt_dbs[2];
1780 for (i = 2; i < txn->mt_numdbs; i++) {
1781 if (ip->md_root != jp->md_root)
1785 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1787 /* Append our free list to parent's */
1788 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1790 mdb_midl_free(txn->mt_free_pgs);
1792 /* Merge our dirty list with parent's */
1793 dst = txn->mt_parent->mt_u.dirty_list;
1794 src = txn->mt_u.dirty_list;
1795 x = mdb_mid2l_search(dst, src[1].mid);
1796 for (y=1; y<=src[0].mid; y++) {
1797 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1801 dst[x].mptr = src[y].mptr;
1804 for (; y<=src[0].mid; y++) {
1805 if (++x >= MDB_IDL_UM_MAX) {
1812 free(txn->mt_u.dirty_list);
1813 txn->mt_parent->mt_child = NULL;
1818 if (txn != env->me_txn) {
1819 DPUTS("attempt to commit unknown transaction");
1824 if (!txn->mt_u.dirty_list[0].mid)
1827 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1828 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1830 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1832 /* should only be one record now */
1833 if (env->me_pghead) {
1834 /* make sure first page of freeDB is touched and on freelist */
1835 mdb_page_search(&mc, NULL, 1);
1837 /* save to free list */
1838 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1842 /* make sure last page of freeDB is touched and on freelist */
1843 key.mv_size = MAXKEYSIZE+1;
1845 mdb_page_search(&mc, &key, 1);
1847 mdb_midl_sort(txn->mt_free_pgs);
1851 ID *idl = txn->mt_free_pgs;
1852 DPRINTF("IDL write txn %zu root %zu num %zu",
1853 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1854 for (i=0; i<idl[0]; i++) {
1855 DPRINTF("IDL %zu", idl[i+1]);
1859 /* write to last page of freeDB */
1860 key.mv_size = sizeof(pgno_t);
1861 key.mv_data = &txn->mt_txnid;
1862 data.mv_data = txn->mt_free_pgs;
1863 /* The free list can still grow during this call,
1864 * despite the pre-emptive touches above. So check
1865 * and make sure the entire thing got written.
1868 i = txn->mt_free_pgs[0];
1869 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1870 rc = mdb_cursor_put(&mc, &key, &data, 0);
1875 } while (i != txn->mt_free_pgs[0]);
1876 if (mdb_midl_shrink(&txn->mt_free_pgs))
1877 env->me_free_pgs = txn->mt_free_pgs;
1879 /* should only be one record now */
1880 if (env->me_pghead) {
1884 mop = env->me_pghead;
1885 env->me_pghead = NULL;
1886 key.mv_size = sizeof(pgno_t);
1887 key.mv_data = &mop->mo_txnid;
1888 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
1889 data.mv_data = mop->mo_pages;
1890 mdb_cursor_put(&mc, &key, &data, 0);
1894 /* Update DB root pointers. Their pages have already been
1895 * touched so this is all in-place and cannot fail.
1900 data.mv_size = sizeof(MDB_db);
1902 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
1903 for (i = 2; i < txn->mt_numdbs; i++) {
1904 if (txn->mt_dbflags[i] & DB_DIRTY) {
1905 data.mv_data = &txn->mt_dbs[i];
1906 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1911 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
1917 /* Windows actually supports scatter/gather I/O, but only on
1918 * unbuffered file handles. Since we're relying on the OS page
1919 * cache for all our data, that's self-defeating. So we just
1920 * write pages one at a time. We use the ov structure to set
1921 * the write offset, to at least save the overhead of a Seek
1925 memset(&ov, 0, sizeof(ov));
1926 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1928 dp = txn->mt_u.dirty_list[i].mptr;
1929 DPRINTF("committing page %zu", dp->mp_pgno);
1930 size = dp->mp_pgno * env->me_psize;
1931 ov.Offset = size & 0xffffffff;
1932 ov.OffsetHigh = size >> 16;
1933 ov.OffsetHigh >>= 16;
1934 /* clear dirty flag */
1935 dp->mp_flags &= ~P_DIRTY;
1936 wsize = env->me_psize;
1937 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
1938 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
1941 DPRINTF("WriteFile: %d", n);
1948 struct iovec iov[MDB_COMMIT_PAGES];
1952 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
1953 dp = txn->mt_u.dirty_list[i].mptr;
1954 if (dp->mp_pgno != next) {
1956 DPRINTF("committing %u dirty pages", n);
1957 rc = writev(env->me_fd, iov, n);
1961 DPUTS("short write, filesystem full?");
1963 DPRINTF("writev: %s", strerror(n));
1970 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
1973 DPRINTF("committing page %zu", dp->mp_pgno);
1974 iov[n].iov_len = env->me_psize;
1975 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
1976 iov[n].iov_base = (char *)dp;
1977 size += iov[n].iov_len;
1978 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
1979 /* clear dirty flag */
1980 dp->mp_flags &= ~P_DIRTY;
1981 if (++n >= MDB_COMMIT_PAGES) {
1991 DPRINTF("committing %u dirty pages", n);
1992 rc = writev(env->me_fd, iov, n);
1996 DPUTS("short write, filesystem full?");
1998 DPRINTF("writev: %s", strerror(n));
2005 /* Drop the dirty pages.
2007 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2008 dp = txn->mt_u.dirty_list[i].mptr;
2009 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2010 dp->mp_next = txn->mt_env->me_dpages;
2011 VGMEMP_FREE(txn->mt_env, dp);
2012 txn->mt_env->me_dpages = dp;
2014 VGMEMP_FREE(txn->mt_env, dp);
2017 txn->mt_u.dirty_list[i].mid = 0;
2019 txn->mt_u.dirty_list[0].mid = 0;
2021 if ((n = mdb_env_sync(env, 0)) != 0 ||
2022 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2026 env->me_wtxnid = txn->mt_txnid;
2030 /* update the DB tables */
2032 int toggle = !env->me_db_toggle;
2036 ip = &env->me_dbs[toggle][2];
2037 jp = &txn->mt_dbs[2];
2038 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
2039 for (i = 2; i < txn->mt_numdbs; i++) {
2040 if (ip->md_root != jp->md_root)
2045 env->me_db_toggle = toggle;
2046 env->me_numdbs = txn->mt_numdbs;
2047 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
2050 UNLOCK_MUTEX_W(env);
2056 /** Read the environment parameters of a DB environment before
2057 * mapping it into memory.
2058 * @param[in] env the environment handle
2059 * @param[out] meta address of where to store the meta information
2060 * @return 0 on success, non-zero on failure.
2063 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2070 /* We don't know the page size yet, so use a minimum value.
2074 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2076 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2081 else if (rc != MDB_PAGESIZE) {
2085 DPRINTF("read: %s", strerror(err));
2089 p = (MDB_page *)&pbuf;
2091 if (!F_ISSET(p->mp_flags, P_META)) {
2092 DPRINTF("page %zu not a meta page", p->mp_pgno);
2097 if (m->mm_magic != MDB_MAGIC) {
2098 DPUTS("meta has invalid magic");
2102 if (m->mm_version != MDB_VERSION) {
2103 DPRINTF("database is version %u, expected version %u",
2104 m->mm_version, MDB_VERSION);
2105 return MDB_VERSION_MISMATCH;
2108 memcpy(meta, m, sizeof(*m));
2112 /** Write the environment parameters of a freshly created DB environment.
2113 * @param[in] env the environment handle
2114 * @param[out] meta address of where to store the meta information
2115 * @return 0 on success, non-zero on failure.
2118 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2125 DPUTS("writing new meta page");
2127 GET_PAGESIZE(psize);
2129 meta->mm_magic = MDB_MAGIC;
2130 meta->mm_version = MDB_VERSION;
2131 meta->mm_psize = psize;
2132 meta->mm_last_pg = 1;
2133 meta->mm_flags = env->me_flags & 0xffff;
2134 meta->mm_flags |= MDB_INTEGERKEY;
2135 meta->mm_dbs[0].md_root = P_INVALID;
2136 meta->mm_dbs[1].md_root = P_INVALID;
2138 p = calloc(2, psize);
2140 p->mp_flags = P_META;
2143 memcpy(m, meta, sizeof(*meta));
2145 q = (MDB_page *)((char *)p + psize);
2148 q->mp_flags = P_META;
2151 memcpy(m, meta, sizeof(*meta));
2156 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2157 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2160 rc = write(env->me_fd, p, psize * 2);
2161 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2167 /** Update the environment info to commit a transaction.
2168 * @param[in] txn the transaction that's being committed
2169 * @return 0 on success, non-zero on failure.
2172 mdb_env_write_meta(MDB_txn *txn)
2175 MDB_meta meta, metab;
2177 int rc, len, toggle;
2183 assert(txn != NULL);
2184 assert(txn->mt_env != NULL);
2186 toggle = !txn->mt_toggle;
2187 DPRINTF("writing meta page %d for root page %zu",
2188 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2192 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2193 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2195 ptr = (char *)&meta;
2196 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2197 len = sizeof(MDB_meta) - off;
2200 meta.mm_dbs[0] = txn->mt_dbs[0];
2201 meta.mm_dbs[1] = txn->mt_dbs[1];
2202 meta.mm_last_pg = txn->mt_next_pgno - 1;
2203 meta.mm_txnid = txn->mt_txnid;
2206 off += env->me_psize;
2209 /* Write to the SYNC fd */
2212 memset(&ov, 0, sizeof(ov));
2214 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2217 rc = pwrite(env->me_mfd, ptr, len, off);
2222 DPUTS("write failed, disk error?");
2223 /* On a failure, the pagecache still contains the new data.
2224 * Write some old data back, to prevent it from being used.
2225 * Use the non-SYNC fd; we know it will fail anyway.
2227 meta.mm_last_pg = metab.mm_last_pg;
2228 meta.mm_txnid = metab.mm_txnid;
2230 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2232 r2 = pwrite(env->me_fd, ptr, len, off);
2234 env->me_flags |= MDB_FATAL_ERROR;
2237 /* Memory ordering issues are irrelevant; since the entire writer
2238 * is wrapped by wmutex, all of these changes will become visible
2239 * after the wmutex is unlocked. Since the DB is multi-version,
2240 * readers will get consistent data regardless of how fresh or
2241 * how stale their view of these values is.
2243 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
2244 txn->mt_env->me_txns->mti_me_toggle = toggle;
2245 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2246 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
2251 /** Check both meta pages to see which one is newer.
2252 * @param[in] env the environment handle
2253 * @param[out] which address of where to store the meta toggle ID
2254 * @return 0 on success, non-zero on failure.
2257 mdb_env_read_meta(MDB_env *env, int *which)
2261 assert(env != NULL);
2263 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2266 DPRINTF("Using meta page %d", toggle);
2273 mdb_env_create(MDB_env **env)
2277 e = calloc(1, sizeof(MDB_env));
2281 e->me_free_pgs = mdb_midl_alloc();
2282 if (!e->me_free_pgs) {
2286 e->me_maxreaders = DEFAULT_READERS;
2288 e->me_fd = INVALID_HANDLE_VALUE;
2289 e->me_lfd = INVALID_HANDLE_VALUE;
2290 e->me_mfd = INVALID_HANDLE_VALUE;
2291 VGMEMP_CREATE(e,0,0);
2297 mdb_env_set_mapsize(MDB_env *env, size_t size)
2301 env->me_mapsize = size;
2303 env->me_maxpg = env->me_mapsize / env->me_psize;
2308 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2312 env->me_maxdbs = dbs;
2317 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2319 if (env->me_map || readers < 1)
2321 env->me_maxreaders = readers;
2326 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2328 if (!env || !readers)
2330 *readers = env->me_maxreaders;
2334 /** Further setup required for opening an MDB environment
2337 mdb_env_open2(MDB_env *env, unsigned int flags)
2339 int i, newenv = 0, toggle;
2343 env->me_flags = flags;
2345 memset(&meta, 0, sizeof(meta));
2347 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2350 DPUTS("new mdbenv");
2354 if (!env->me_mapsize) {
2355 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2361 LONG sizelo, sizehi;
2362 sizelo = env->me_mapsize & 0xffffffff;
2363 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2365 /* Windows won't create mappings for zero length files.
2366 * Just allocate the maxsize right now.
2369 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2370 if (!SetEndOfFile(env->me_fd))
2372 SetFilePointer(env->me_fd, 0, NULL, 0);
2374 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2375 sizehi, sizelo, NULL);
2378 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2386 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2388 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2390 if (env->me_map == MAP_FAILED) {
2397 meta.mm_mapsize = env->me_mapsize;
2398 if (flags & MDB_FIXEDMAP)
2399 meta.mm_address = env->me_map;
2400 i = mdb_env_init_meta(env, &meta);
2401 if (i != MDB_SUCCESS) {
2402 munmap(env->me_map, env->me_mapsize);
2406 env->me_psize = meta.mm_psize;
2408 env->me_maxpg = env->me_mapsize / env->me_psize;
2410 p = (MDB_page *)env->me_map;
2411 env->me_metas[0] = METADATA(p);
2412 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2414 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
2417 DPRINTF("opened database version %u, pagesize %u",
2418 env->me_metas[toggle]->mm_version, env->me_psize);
2419 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
2420 DPRINTF("entries: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
2421 DPRINTF("branch pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
2422 DPRINTF("leaf pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
2423 DPRINTF("overflow pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
2424 DPRINTF("root: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
2430 /** Release a reader thread's slot in the reader lock table.
2431 * This function is called automatically when a thread exits.
2432 * Windows doesn't support destructor callbacks for thread-specific storage,
2433 * so this function is not compiled there.
2434 * @param[in] ptr This points to the slot in the reader lock table.
2437 mdb_env_reader_dest(void *ptr)
2439 MDB_reader *reader = ptr;
2441 reader->mr_txnid = 0;
2447 /** Downgrade the exclusive lock on the region back to shared */
2449 mdb_env_share_locks(MDB_env *env)
2453 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2455 env->me_txns->mti_me_toggle = toggle;
2456 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2461 /* First acquire a shared lock. The Unlock will
2462 * then release the existing exclusive lock.
2464 memset(&ov, 0, sizeof(ov));
2465 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2466 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2470 struct flock lock_info;
2471 /* The shared lock replaces the existing lock */
2472 memset((void *)&lock_info, 0, sizeof(lock_info));
2473 lock_info.l_type = F_RDLCK;
2474 lock_info.l_whence = SEEK_SET;
2475 lock_info.l_start = 0;
2476 lock_info.l_len = 1;
2477 fcntl(env->me_lfd, F_SETLK, &lock_info);
2481 #if defined(_WIN32) || defined(__APPLE__)
2483 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2485 * @(#) $Revision: 5.1 $
2486 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2487 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2489 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2493 * Please do not copyright this code. This code is in the public domain.
2495 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2496 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2497 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2498 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2499 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2500 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2501 * PERFORMANCE OF THIS SOFTWARE.
2504 * chongo <Landon Curt Noll> /\oo/\
2505 * http://www.isthe.com/chongo/
2507 * Share and Enjoy! :-)
2510 typedef unsigned long long mdb_hash_t;
2511 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2513 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2514 * @param[in] str string to hash
2515 * @param[in] hval initial value for hash
2516 * @return 64 bit hash
2518 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2519 * hval arg on the first call.
2522 mdb_hash_str(char *str, mdb_hash_t hval)
2524 unsigned char *s = (unsigned char *)str; /* unsigned string */
2526 * FNV-1a hash each octet of the string
2529 /* xor the bottom with the current octet */
2530 hval ^= (mdb_hash_t)*s++;
2532 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2533 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2534 (hval << 7) + (hval << 8) + (hval << 40);
2536 /* return our new hash value */
2540 /** Hash the string and output the hash in hex.
2541 * @param[in] str string to hash
2542 * @param[out] hexbuf an array of 17 chars to hold the hash
2545 mdb_hash_hex(char *str, char *hexbuf)
2548 mdb_hash_t h = mdb_hash_str(str, MDB_HASH_INIT);
2549 for (i=0; i<8; i++) {
2550 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2556 /** Open and/or initialize the lock region for the environment.
2557 * @param[in] env The MDB environment.
2558 * @param[in] lpath The pathname of the file used for the lock region.
2559 * @param[in] mode The Unix permissions for the file, if we create it.
2560 * @param[out] excl Set to true if we got an exclusive lock on the region.
2561 * @return 0 on success, non-zero on failure.
2564 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2572 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2573 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2574 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2578 /* Try to get exclusive lock. If we succeed, then
2579 * nobody is using the lock region and we should initialize it.
2582 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2586 memset(&ov, 0, sizeof(ov));
2587 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2593 size = GetFileSize(env->me_lfd, NULL);
2595 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2599 /* Try to get exclusive lock. If we succeed, then
2600 * nobody is using the lock region and we should initialize it.
2603 struct flock lock_info;
2604 memset((void *)&lock_info, 0, sizeof(lock_info));
2605 lock_info.l_type = F_WRLCK;
2606 lock_info.l_whence = SEEK_SET;
2607 lock_info.l_start = 0;
2608 lock_info.l_len = 1;
2609 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2613 lock_info.l_type = F_RDLCK;
2614 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2621 size = lseek(env->me_lfd, 0, SEEK_END);
2623 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2624 if (size < rsize && *excl) {
2626 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2627 if (!SetEndOfFile(env->me_lfd)) {
2632 if (ftruncate(env->me_lfd, rsize) != 0) {
2639 size = rsize - sizeof(MDB_txninfo);
2640 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2645 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2651 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2653 if (!env->me_txns) {
2658 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2660 if (m == MAP_FAILED) {
2661 env->me_txns = NULL;
2671 if (!mdb_sec_inited) {
2672 InitializeSecurityDescriptor(&mdb_null_sd,
2673 SECURITY_DESCRIPTOR_REVISION);
2674 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2675 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2676 mdb_all_sa.bInheritHandle = FALSE;
2677 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2680 mdb_hash_hex(lpath, hexbuf);
2681 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2682 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2683 if (!env->me_rmutex) {
2687 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2688 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2689 if (!env->me_wmutex) {
2696 mdb_hash_hex(lpath, hexbuf);
2697 sprintf(env->me_txns->mti_rmname, "MDBr%s", hexbuf);
2698 if (sem_unlink(env->me_txns->mti_rmname)) {
2700 if (rc != ENOENT && rc != EINVAL)
2703 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2704 if (!env->me_rmutex) {
2708 sprintf(env->me_txns->mti_wmname, "MDBw%s", hexbuf);
2709 if (sem_unlink(env->me_txns->mti_wmname)) {
2711 if (rc != ENOENT && rc != EINVAL)
2714 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2715 if (!env->me_wmutex) {
2719 #else /* __APPLE__ */
2720 pthread_mutexattr_t mattr;
2722 pthread_mutexattr_init(&mattr);
2723 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2727 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2728 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2729 #endif /* __APPLE__ */
2731 env->me_txns->mti_version = MDB_VERSION;
2732 env->me_txns->mti_magic = MDB_MAGIC;
2733 env->me_txns->mti_txnid = 0;
2734 env->me_txns->mti_numreaders = 0;
2735 env->me_txns->mti_me_toggle = 0;
2738 if (env->me_txns->mti_magic != MDB_MAGIC) {
2739 DPUTS("lock region has invalid magic");
2743 if (env->me_txns->mti_version != MDB_VERSION) {
2744 DPRINTF("lock region is version %u, expected version %u",
2745 env->me_txns->mti_version, MDB_VERSION);
2746 rc = MDB_VERSION_MISMATCH;
2750 if (rc != EACCES && rc != EAGAIN) {
2754 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2755 if (!env->me_rmutex) {
2759 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2760 if (!env->me_wmutex) {
2766 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2767 if (!env->me_rmutex) {
2771 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2772 if (!env->me_wmutex) {
2782 env->me_lfd = INVALID_HANDLE_VALUE;
2787 /** The name of the lock file in the DB environment */
2788 #define LOCKNAME "/lock.mdb"
2789 /** The name of the data file in the DB environment */
2790 #define DATANAME "/data.mdb"
2791 /** The suffix of the lock file when no subdir is used */
2792 #define LOCKSUFF "-lock"
2795 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2797 int oflags, rc, len, excl;
2798 char *lpath, *dpath;
2801 if (flags & MDB_NOSUBDIR) {
2802 rc = len + sizeof(LOCKSUFF) + len + 1;
2804 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
2809 if (flags & MDB_NOSUBDIR) {
2810 dpath = lpath + len + sizeof(LOCKSUFF);
2811 sprintf(lpath, "%s" LOCKSUFF, path);
2812 strcpy(dpath, path);
2814 dpath = lpath + len + sizeof(LOCKNAME);
2815 sprintf(lpath, "%s" LOCKNAME, path);
2816 sprintf(dpath, "%s" DATANAME, path);
2819 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2824 if (F_ISSET(flags, MDB_RDONLY)) {
2825 oflags = GENERIC_READ;
2826 len = OPEN_EXISTING;
2828 oflags = GENERIC_READ|GENERIC_WRITE;
2831 mode = FILE_ATTRIBUTE_NORMAL;
2832 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2833 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2838 if (F_ISSET(flags, MDB_RDONLY))
2841 oflags = O_RDWR | O_CREAT;
2843 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2849 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2850 /* synchronous fd for meta writes */
2852 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2853 mode |= FILE_FLAG_WRITE_THROUGH;
2854 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2855 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2860 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2861 oflags |= MDB_DSYNC;
2862 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
2867 env->me_path = strdup(path);
2868 DPRINTF("opened dbenv %p", (void *) env);
2869 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
2870 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
2872 mdb_env_share_locks(env);
2873 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
2874 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
2875 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
2881 if (env->me_fd != INVALID_HANDLE_VALUE) {
2883 env->me_fd = INVALID_HANDLE_VALUE;
2885 if (env->me_lfd != INVALID_HANDLE_VALUE) {
2887 env->me_lfd = INVALID_HANDLE_VALUE;
2895 mdb_env_close(MDB_env *env)
2902 VGMEMP_DESTROY(env);
2903 while (env->me_dpages) {
2904 dp = env->me_dpages;
2905 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
2906 env->me_dpages = dp->mp_next;
2910 free(env->me_dbs[1]);
2911 free(env->me_dbs[0]);
2915 LAZY_RWLOCK_DESTROY(&env->me_dblock);
2916 pthread_key_delete(env->me_txkey);
2919 munmap(env->me_map, env->me_mapsize);
2924 pid_t pid = getpid();
2926 for (i=0; i<env->me_txns->mti_numreaders; i++)
2927 if (env->me_txns->mti_readers[i].mr_pid == pid)
2928 env->me_txns->mti_readers[i].mr_pid = 0;
2929 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
2932 mdb_midl_free(env->me_free_pgs);
2936 /** Compare two items pointing at aligned size_t's */
2938 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
2940 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
2941 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
2944 /** Compare two items pointing at aligned int's */
2946 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
2948 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
2949 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
2952 /** Compare two items pointing at ints of unknown alignment.
2953 * Nodes and keys are guaranteed to be 2-byte aligned.
2956 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
2958 #if BYTE_ORDER == LITTLE_ENDIAN
2959 unsigned short *u, *c;
2962 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
2963 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
2966 } while(!x && u > (unsigned short *)a->mv_data);
2969 return memcmp(a->mv_data, b->mv_data, a->mv_size);
2973 /** Compare two items lexically */
2975 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
2982 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
2988 diff = memcmp(a->mv_data, b->mv_data, len);
2989 return diff ? diff : len_diff<0 ? -1 : len_diff;
2992 /** Compare two items in reverse byte order */
2994 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
2996 const unsigned char *p1, *p2, *p1_lim;
3000 p1_lim = (const unsigned char *)a->mv_data;
3001 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3002 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3004 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3010 while (p1 > p1_lim) {
3011 diff = *--p1 - *--p2;
3015 return len_diff<0 ? -1 : len_diff;
3018 /** Search for key within a page, using binary search.
3019 * Returns the smallest entry larger or equal to the key.
3020 * If exactp is non-null, stores whether the found entry was an exact match
3021 * in *exactp (1 or 0).
3022 * Updates the cursor index with the index of the found entry.
3023 * If no entry larger or equal to the key is found, returns NULL.
3026 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3028 unsigned int i = 0, nkeys;
3031 MDB_page *mp = mc->mc_pg[mc->mc_top];
3032 MDB_node *node = NULL;
3037 nkeys = NUMKEYS(mp);
3042 COPY_PGNO(pgno, mp->mp_pgno);
3043 DPRINTF("searching %u keys in %s %spage %zu",
3044 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3051 low = IS_LEAF(mp) ? 0 : 1;
3053 cmp = mc->mc_dbx->md_cmp;
3055 /* Branch pages have no data, so if using integer keys,
3056 * alignment is guaranteed. Use faster mdb_cmp_int.
3058 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3059 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3066 nodekey.mv_size = mc->mc_db->md_pad;
3067 node = NODEPTR(mp, 0); /* fake */
3068 while (low <= high) {
3069 i = (low + high) >> 1;
3070 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3071 rc = cmp(key, &nodekey);
3072 DPRINTF("found leaf index %u [%s], rc = %i",
3073 i, DKEY(&nodekey), rc);
3082 while (low <= high) {
3083 i = (low + high) >> 1;
3085 node = NODEPTR(mp, i);
3086 nodekey.mv_size = NODEKSZ(node);
3087 nodekey.mv_data = NODEKEY(node);
3089 rc = cmp(key, &nodekey);
3092 DPRINTF("found leaf index %u [%s], rc = %i",
3093 i, DKEY(&nodekey), rc);
3095 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3096 i, DKEY(&nodekey), NODEPGNO(node), rc);
3107 if (rc > 0) { /* Found entry is less than the key. */
3108 i++; /* Skip to get the smallest entry larger than key. */
3110 node = NODEPTR(mp, i);
3113 *exactp = (rc == 0);
3114 /* store the key index */
3115 mc->mc_ki[mc->mc_top] = i;
3117 /* There is no entry larger or equal to the key. */
3120 /* nodeptr is fake for LEAF2 */
3126 mdb_cursor_adjust(MDB_cursor *mc, func)
3130 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3131 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3138 /** Pop a page off the top of the cursor's stack. */
3140 mdb_cursor_pop(MDB_cursor *mc)
3145 top = mc->mc_pg[mc->mc_top];
3150 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3151 mc->mc_dbi, (void *) mc);
3155 /** Push a page onto the top of the cursor's stack. */
3157 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3159 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3160 mc->mc_dbi, (void *) mc);
3162 if (mc->mc_snum >= CURSOR_STACK) {
3163 assert(mc->mc_snum < CURSOR_STACK);
3167 mc->mc_top = mc->mc_snum++;
3168 mc->mc_pg[mc->mc_top] = mp;
3169 mc->mc_ki[mc->mc_top] = 0;
3174 /** Find the address of the page corresponding to a given page number.
3175 * @param[in] txn the transaction for this access.
3176 * @param[in] pgno the page number for the page to retrieve.
3177 * @param[out] ret address of a pointer where the page's address will be stored.
3178 * @return 0 on success, non-zero on failure.
3181 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3185 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3187 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3188 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3189 p = txn->mt_u.dirty_list[x].mptr;
3193 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
3194 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3198 DPRINTF("page %zu not found", pgno);
3201 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3204 /** Search for the page a given key should be in.
3205 * Pushes parent pages on the cursor stack. This function continues a
3206 * search on a cursor that has already been initialized. (Usually by
3207 * #mdb_page_search() but also by #mdb_node_move().)
3208 * @param[in,out] mc the cursor for this operation.
3209 * @param[in] key the key to search for. If NULL, search for the lowest
3210 * page. (This is used by #mdb_cursor_first().)
3211 * @param[in] modify If true, visited pages are updated with new page numbers.
3212 * @return 0 on success, non-zero on failure.
3215 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3217 MDB_page *mp = mc->mc_pg[mc->mc_top];
3222 while (IS_BRANCH(mp)) {
3226 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3227 assert(NUMKEYS(mp) > 1);
3228 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3230 if (key == NULL) /* Initialize cursor to first page. */
3232 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3233 /* cursor to last page */
3237 node = mdb_node_search(mc, key, &exact);
3239 i = NUMKEYS(mp) - 1;
3241 i = mc->mc_ki[mc->mc_top];
3250 DPRINTF("following index %u for key [%s]",
3252 assert(i < NUMKEYS(mp));
3253 node = NODEPTR(mp, i);
3255 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3258 mc->mc_ki[mc->mc_top] = i;
3259 if ((rc = mdb_cursor_push(mc, mp)))
3263 if ((rc = mdb_page_touch(mc)) != 0)
3265 mp = mc->mc_pg[mc->mc_top];
3270 DPRINTF("internal error, index points to a %02X page!?",
3272 return MDB_CORRUPTED;
3275 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3276 key ? DKEY(key) : NULL);
3281 /** Search for the page a given key should be in.
3282 * Pushes parent pages on the cursor stack. This function just sets up
3283 * the search; it finds the root page for \b mc's database and sets this
3284 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3285 * called to complete the search.
3286 * @param[in,out] mc the cursor for this operation.
3287 * @param[in] key the key to search for. If NULL, search for the lowest
3288 * page. (This is used by #mdb_cursor_first().)
3289 * @param[in] modify If true, visited pages are updated with new page numbers.
3290 * @return 0 on success, non-zero on failure.
3293 mdb_page_search(MDB_cursor *mc, MDB_val *key, int modify)
3298 /* Make sure the txn is still viable, then find the root from
3299 * the txn's db table.
3301 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3302 DPUTS("transaction has failed, must abort");
3305 /* Make sure we're using an up-to-date root */
3306 if (mc->mc_dbi > MAIN_DBI) {
3307 if ((*mc->mc_dbflag & DB_STALE) ||
3308 (modify && !(*mc->mc_dbflag & DB_DIRTY))) {
3310 unsigned char dbflag = 0;
3311 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3312 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, modify);
3315 if (*mc->mc_dbflag & DB_STALE) {
3318 MDB_node *leaf = mdb_node_search(&mc2,
3319 &mc->mc_dbx->md_name, &exact);
3321 return MDB_NOTFOUND;
3322 mdb_node_read(mc->mc_txn, leaf, &data);
3323 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3327 *mc->mc_dbflag = dbflag;
3330 root = mc->mc_db->md_root;
3332 if (root == P_INVALID) { /* Tree is empty. */
3333 DPUTS("tree is empty");
3334 return MDB_NOTFOUND;
3339 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3345 DPRINTF("db %u root page %zu has flags 0x%X",
3346 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3349 if ((rc = mdb_page_touch(mc)))
3353 return mdb_page_search_root(mc, key, modify);
3356 /** Return the data associated with a given node.
3357 * @param[in] txn The transaction for this operation.
3358 * @param[in] leaf The node being read.
3359 * @param[out] data Updated to point to the node's data.
3360 * @return 0 on success, non-zero on failure.
3363 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3365 MDB_page *omp; /* overflow page */
3369 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3370 data->mv_size = NODEDSZ(leaf);
3371 data->mv_data = NODEDATA(leaf);
3375 /* Read overflow data.
3377 data->mv_size = NODEDSZ(leaf);
3378 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3379 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3380 DPRINTF("read overflow page %zu failed", pgno);
3383 data->mv_data = METADATA(omp);
3389 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3390 MDB_val *key, MDB_val *data)
3399 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3401 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3404 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3408 mdb_cursor_init(&mc, txn, dbi, &mx);
3409 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3412 /** Find a sibling for a page.
3413 * Replaces the page at the top of the cursor's stack with the
3414 * specified sibling, if one exists.
3415 * @param[in] mc The cursor for this operation.
3416 * @param[in] move_right Non-zero if the right sibling is requested,
3417 * otherwise the left sibling.
3418 * @return 0 on success, non-zero on failure.
3421 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3427 if (mc->mc_snum < 2) {
3428 return MDB_NOTFOUND; /* root has no siblings */
3432 DPRINTF("parent page is page %zu, index %u",
3433 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3435 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3436 : (mc->mc_ki[mc->mc_top] == 0)) {
3437 DPRINTF("no more keys left, moving to %s sibling",
3438 move_right ? "right" : "left");
3439 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3443 mc->mc_ki[mc->mc_top]++;
3445 mc->mc_ki[mc->mc_top]--;
3446 DPRINTF("just moving to %s index key %u",
3447 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3449 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3451 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3452 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3455 mdb_cursor_push(mc, mp);
3460 /** Move the cursor to the next data item. */
3462 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3468 if (mc->mc_flags & C_EOF) {
3469 return MDB_NOTFOUND;
3472 assert(mc->mc_flags & C_INITIALIZED);
3474 mp = mc->mc_pg[mc->mc_top];
3476 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3477 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3478 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3479 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3480 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3481 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3485 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3486 if (op == MDB_NEXT_DUP)
3487 return MDB_NOTFOUND;
3491 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3493 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3494 DPUTS("=====> move to next sibling page");
3495 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3496 mc->mc_flags |= C_EOF;
3497 mc->mc_flags &= ~C_INITIALIZED;
3498 return MDB_NOTFOUND;
3500 mp = mc->mc_pg[mc->mc_top];
3501 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3503 mc->mc_ki[mc->mc_top]++;
3505 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3506 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3509 key->mv_size = mc->mc_db->md_pad;
3510 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3514 assert(IS_LEAF(mp));
3515 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3517 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3518 mdb_xcursor_init1(mc, leaf);
3521 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3524 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3525 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3526 if (rc != MDB_SUCCESS)
3531 MDB_SET_KEY(leaf, key);
3535 /** Move the cursor to the previous data item. */
3537 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3543 assert(mc->mc_flags & C_INITIALIZED);
3545 mp = mc->mc_pg[mc->mc_top];
3547 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3548 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3549 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3550 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3551 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3552 if (op != MDB_PREV || rc == MDB_SUCCESS)
3555 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3556 if (op == MDB_PREV_DUP)
3557 return MDB_NOTFOUND;
3562 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3564 if (mc->mc_ki[mc->mc_top] == 0) {
3565 DPUTS("=====> move to prev sibling page");
3566 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3567 mc->mc_flags &= ~C_INITIALIZED;
3568 return MDB_NOTFOUND;
3570 mp = mc->mc_pg[mc->mc_top];
3571 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3572 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3574 mc->mc_ki[mc->mc_top]--;
3576 mc->mc_flags &= ~C_EOF;
3578 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3579 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3582 key->mv_size = mc->mc_db->md_pad;
3583 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3587 assert(IS_LEAF(mp));
3588 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3590 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3591 mdb_xcursor_init1(mc, leaf);
3594 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3597 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3598 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3599 if (rc != MDB_SUCCESS)
3604 MDB_SET_KEY(leaf, key);
3608 /** Set the cursor on a specific data item. */
3610 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3611 MDB_cursor_op op, int *exactp)
3620 assert(key->mv_size > 0);
3622 /* See if we're already on the right page */
3623 if (mc->mc_flags & C_INITIALIZED) {
3626 mp = mc->mc_pg[mc->mc_top];
3628 mc->mc_ki[mc->mc_top] = 0;
3629 return MDB_NOTFOUND;
3631 if (mp->mp_flags & P_LEAF2) {
3632 nodekey.mv_size = mc->mc_db->md_pad;
3633 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3635 leaf = NODEPTR(mp, 0);
3636 MDB_SET_KEY(leaf, &nodekey);
3638 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3640 /* Probably happens rarely, but first node on the page
3641 * was the one we wanted.
3643 mc->mc_ki[mc->mc_top] = 0;
3644 leaf = NODEPTR(mp, 0);
3651 unsigned int nkeys = NUMKEYS(mp);
3653 if (mp->mp_flags & P_LEAF2) {
3654 nodekey.mv_data = LEAF2KEY(mp,
3655 nkeys-1, nodekey.mv_size);
3657 leaf = NODEPTR(mp, nkeys-1);
3658 MDB_SET_KEY(leaf, &nodekey);
3660 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3662 /* last node was the one we wanted */
3663 mc->mc_ki[mc->mc_top] = nkeys-1;
3664 leaf = NODEPTR(mp, nkeys-1);
3670 /* This is definitely the right page, skip search_page */
3675 /* If any parents have right-sibs, search.
3676 * Otherwise, there's nothing further.
3678 for (i=0; i<mc->mc_top; i++)
3680 NUMKEYS(mc->mc_pg[i])-1)
3682 if (i == mc->mc_top) {
3683 /* There are no other pages */
3684 mc->mc_ki[mc->mc_top] = nkeys;
3685 return MDB_NOTFOUND;
3689 /* There are no other pages */
3690 mc->mc_ki[mc->mc_top] = 0;
3691 return MDB_NOTFOUND;
3695 rc = mdb_page_search(mc, key, 0);
3696 if (rc != MDB_SUCCESS)
3699 mp = mc->mc_pg[mc->mc_top];
3700 assert(IS_LEAF(mp));
3703 leaf = mdb_node_search(mc, key, exactp);
3704 if (exactp != NULL && !*exactp) {
3705 /* MDB_SET specified and not an exact match. */
3706 return MDB_NOTFOUND;
3710 DPUTS("===> inexact leaf not found, goto sibling");
3711 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3712 return rc; /* no entries matched */
3713 mp = mc->mc_pg[mc->mc_top];
3714 assert(IS_LEAF(mp));
3715 leaf = NODEPTR(mp, 0);
3719 mc->mc_flags |= C_INITIALIZED;
3720 mc->mc_flags &= ~C_EOF;
3723 key->mv_size = mc->mc_db->md_pad;
3724 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3728 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3729 mdb_xcursor_init1(mc, leaf);
3732 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3733 if (op == MDB_SET || op == MDB_SET_RANGE) {
3734 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3737 if (op == MDB_GET_BOTH) {
3743 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3744 if (rc != MDB_SUCCESS)
3747 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3749 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3751 rc = mc->mc_dbx->md_dcmp(data, &d2);
3753 if (op == MDB_GET_BOTH || rc > 0)
3754 return MDB_NOTFOUND;
3759 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3760 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3765 /* The key already matches in all other cases */
3766 if (op == MDB_SET_RANGE)
3767 MDB_SET_KEY(leaf, key);
3768 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3773 /** Move the cursor to the first item in the database. */
3775 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3780 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3781 rc = mdb_page_search(mc, NULL, 0);
3782 if (rc != MDB_SUCCESS)
3785 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3787 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3788 mc->mc_flags |= C_INITIALIZED;
3789 mc->mc_flags &= ~C_EOF;
3791 mc->mc_ki[mc->mc_top] = 0;
3793 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3794 key->mv_size = mc->mc_db->md_pad;
3795 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3800 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3801 mdb_xcursor_init1(mc, leaf);
3802 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3807 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3808 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3812 MDB_SET_KEY(leaf, key);
3816 /** Move the cursor to the last item in the database. */
3818 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3824 lkey.mv_size = MAXKEYSIZE+1;
3825 lkey.mv_data = NULL;
3827 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3828 rc = mdb_page_search(mc, &lkey, 0);
3829 if (rc != MDB_SUCCESS)
3832 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3834 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3835 mc->mc_flags |= C_INITIALIZED;
3836 mc->mc_flags &= ~C_EOF;
3838 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3840 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3841 key->mv_size = mc->mc_db->md_pad;
3842 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3847 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3848 mdb_xcursor_init1(mc, leaf);
3849 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3854 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3855 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3860 MDB_SET_KEY(leaf, key);
3865 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3875 case MDB_GET_BOTH_RANGE:
3876 if (data == NULL || mc->mc_xcursor == NULL) {
3883 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3885 } else if (op == MDB_SET_RANGE)
3886 rc = mdb_cursor_set(mc, key, data, op, NULL);
3888 rc = mdb_cursor_set(mc, key, data, op, &exact);
3890 case MDB_GET_MULTIPLE:
3892 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
3893 !(mc->mc_flags & C_INITIALIZED)) {
3898 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
3899 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
3902 case MDB_NEXT_MULTIPLE:
3904 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
3908 if (!(mc->mc_flags & C_INITIALIZED))
3909 rc = mdb_cursor_first(mc, key, data);
3911 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
3912 if (rc == MDB_SUCCESS) {
3913 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
3916 mx = &mc->mc_xcursor->mx_cursor;
3917 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
3919 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
3920 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
3928 case MDB_NEXT_NODUP:
3929 if (!(mc->mc_flags & C_INITIALIZED))
3930 rc = mdb_cursor_first(mc, key, data);
3932 rc = mdb_cursor_next(mc, key, data, op);
3936 case MDB_PREV_NODUP:
3937 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
3938 rc = mdb_cursor_last(mc, key, data);
3940 rc = mdb_cursor_prev(mc, key, data, op);
3943 rc = mdb_cursor_first(mc, key, data);
3947 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
3948 !(mc->mc_flags & C_INITIALIZED) ||
3949 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3953 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3956 rc = mdb_cursor_last(mc, key, data);
3960 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
3961 !(mc->mc_flags & C_INITIALIZED) ||
3962 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
3966 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3969 DPRINTF("unhandled/unimplemented cursor operation %u", op);
3977 /** Touch all the pages in the cursor stack.
3978 * Makes sure all the pages are writable, before attempting a write operation.
3979 * @param[in] mc The cursor to operate on.
3982 mdb_cursor_touch(MDB_cursor *mc)
3986 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
3988 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3989 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 1);
3992 *mc->mc_dbflag = DB_DIRTY;
3994 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
3995 rc = mdb_page_touch(mc);
3999 mc->mc_top = mc->mc_snum-1;
4004 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4007 MDB_node *leaf = NULL;
4008 MDB_val xdata, *rdata, dkey;
4012 unsigned int mcount = 0;
4016 char dbuf[MAXKEYSIZE+1];
4017 unsigned int nflags;
4020 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4023 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4024 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4028 if (flags == MDB_CURRENT) {
4029 if (!(mc->mc_flags & C_INITIALIZED))
4032 } else if (mc->mc_db->md_root == P_INVALID) {
4034 /* new database, write a root leaf page */
4035 DPUTS("allocating new root leaf page");
4036 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4040 mdb_cursor_push(mc, np);
4041 mc->mc_db->md_root = np->mp_pgno;
4042 mc->mc_db->md_depth++;
4043 *mc->mc_dbflag = DB_DIRTY;
4044 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4046 np->mp_flags |= P_LEAF2;
4047 mc->mc_flags |= C_INITIALIZED;
4053 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4054 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4055 DPRINTF("duplicate key [%s]", DKEY(key));
4057 return MDB_KEYEXIST;
4059 if (rc && rc != MDB_NOTFOUND)
4063 /* Cursor is positioned, now make sure all pages are writable */
4064 rc2 = mdb_cursor_touch(mc);
4069 /* The key already exists */
4070 if (rc == MDB_SUCCESS) {
4071 /* there's only a key anyway, so this is a no-op */
4072 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4073 unsigned int ksize = mc->mc_db->md_pad;
4074 if (key->mv_size != ksize)
4076 if (flags == MDB_CURRENT) {
4077 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4078 memcpy(ptr, key->mv_data, ksize);
4083 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4086 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4087 /* Was a single item before, must convert now */
4089 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4090 /* Just overwrite the current item */
4091 if (flags == MDB_CURRENT)
4094 dkey.mv_size = NODEDSZ(leaf);
4095 dkey.mv_data = NODEDATA(leaf);
4096 #if UINT_MAX < SIZE_MAX
4097 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4098 #ifdef MISALIGNED_OK
4099 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4101 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4104 /* if data matches, ignore it */
4105 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4106 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4108 /* create a fake page for the dup items */
4109 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4110 dkey.mv_data = dbuf;
4111 fp = (MDB_page *)&pbuf;
4112 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4113 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4114 fp->mp_lower = PAGEHDRSZ;
4115 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4116 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4117 fp->mp_flags |= P_LEAF2;
4118 fp->mp_pad = data->mv_size;
4120 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4121 (dkey.mv_size & 1) + (data->mv_size & 1);
4123 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4126 xdata.mv_size = fp->mp_upper;
4131 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4132 /* See if we need to convert from fake page to subDB */
4134 unsigned int offset;
4137 fp = NODEDATA(leaf);
4138 if (flags == MDB_CURRENT) {
4139 fp->mp_flags |= P_DIRTY;
4140 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4141 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4145 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4146 offset = fp->mp_pad;
4148 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4150 offset += offset & 1;
4151 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4152 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4154 /* yes, convert it */
4156 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4157 dummy.md_pad = fp->mp_pad;
4158 dummy.md_flags = MDB_DUPFIXED;
4159 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4160 dummy.md_flags |= MDB_INTEGERKEY;
4163 dummy.md_branch_pages = 0;
4164 dummy.md_leaf_pages = 1;
4165 dummy.md_overflow_pages = 0;
4166 dummy.md_entries = NUMKEYS(fp);
4168 xdata.mv_size = sizeof(MDB_db);
4169 xdata.mv_data = &dummy;
4170 mp = mdb_page_alloc(mc, 1);
4173 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4174 flags |= F_DUPDATA|F_SUBDATA;
4175 dummy.md_root = mp->mp_pgno;
4177 /* no, just grow it */
4179 xdata.mv_size = NODEDSZ(leaf) + offset;
4180 xdata.mv_data = &pbuf;
4181 mp = (MDB_page *)&pbuf;
4182 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4185 mp->mp_flags = fp->mp_flags | P_DIRTY;
4186 mp->mp_pad = fp->mp_pad;
4187 mp->mp_lower = fp->mp_lower;
4188 mp->mp_upper = fp->mp_upper + offset;
4190 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4192 nsize = NODEDSZ(leaf) - fp->mp_upper;
4193 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4194 for (i=0; i<NUMKEYS(fp); i++)
4195 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4197 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4201 /* data is on sub-DB, just store it */
4202 flags |= F_DUPDATA|F_SUBDATA;
4206 /* same size, just replace it */
4207 if (!F_ISSET(leaf->mn_flags, F_BIGDATA) &&
4208 NODEDSZ(leaf) == data->mv_size) {
4209 if (F_ISSET(flags, MDB_RESERVE))
4210 data->mv_data = NODEDATA(leaf);
4212 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4215 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4217 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4223 nflags = flags & NODE_ADD_FLAGS;
4224 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4225 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4226 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4227 nflags &= ~MDB_APPEND;
4228 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4230 /* There is room already in this leaf page. */
4231 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4232 if (rc == 0 && !do_sub) {
4233 /* Adjust other cursors pointing to mp */
4234 MDB_cursor *m2, *m3;
4235 MDB_dbi dbi = mc->mc_dbi;
4236 unsigned i = mc->mc_top;
4237 MDB_page *mp = mc->mc_pg[i];
4239 if (mc->mc_flags & C_SUB)
4242 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4243 if (mc->mc_flags & C_SUB)
4244 m3 = &m2->mc_xcursor->mx_cursor;
4247 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4248 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4255 if (rc != MDB_SUCCESS)
4256 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4258 /* Now store the actual data in the child DB. Note that we're
4259 * storing the user data in the keys field, so there are strict
4260 * size limits on dupdata. The actual data fields of the child
4261 * DB are all zero size.
4268 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4269 if (flags & MDB_CURRENT) {
4270 xflags = MDB_CURRENT;
4272 mdb_xcursor_init1(mc, leaf);
4273 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4275 /* converted, write the original data first */
4277 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4281 /* Adjust other cursors pointing to mp */
4283 unsigned i = mc->mc_top;
4284 MDB_page *mp = mc->mc_pg[i];
4286 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4287 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4288 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4289 mdb_xcursor_init1(m2, leaf);
4294 xflags |= (flags & MDB_APPEND);
4295 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4296 if (flags & F_SUBDATA) {
4297 void *db = NODEDATA(leaf);
4298 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4301 /* sub-writes might have failed so check rc again.
4302 * Don't increment count if we just replaced an existing item.
4304 if (!rc && !(flags & MDB_CURRENT))
4305 mc->mc_db->md_entries++;
4306 if (flags & MDB_MULTIPLE) {
4308 if (mcount < data[1].mv_size) {
4309 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4310 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4320 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4325 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4328 if (!mc->mc_flags & C_INITIALIZED)
4331 rc = mdb_cursor_touch(mc);
4335 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4337 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4338 if (flags != MDB_NODUPDATA) {
4339 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4340 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4342 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4343 /* If sub-DB still has entries, we're done */
4344 if (mc->mc_xcursor->mx_db.md_entries) {
4345 if (leaf->mn_flags & F_SUBDATA) {
4346 /* update subDB info */
4347 void *db = NODEDATA(leaf);
4348 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4350 /* shrink fake page */
4351 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4353 mc->mc_db->md_entries--;
4356 /* otherwise fall thru and delete the sub-DB */
4359 if (leaf->mn_flags & F_SUBDATA) {
4360 /* add all the child DB's pages to the free list */
4361 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4362 if (rc == MDB_SUCCESS) {
4363 mc->mc_db->md_entries -=
4364 mc->mc_xcursor->mx_db.md_entries;
4369 return mdb_cursor_del0(mc, leaf);
4372 /** Allocate and initialize new pages for a database.
4373 * @param[in] mc a cursor on the database being added to.
4374 * @param[in] flags flags defining what type of page is being allocated.
4375 * @param[in] num the number of pages to allocate. This is usually 1,
4376 * unless allocating overflow pages for a large record.
4377 * @return Address of a page, or NULL on failure.
4380 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4384 if ((np = mdb_page_alloc(mc, num)) == NULL)
4386 DPRINTF("allocated new mpage %zu, page size %u",
4387 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4388 np->mp_flags = flags | P_DIRTY;
4389 np->mp_lower = PAGEHDRSZ;
4390 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4393 mc->mc_db->md_branch_pages++;
4394 else if (IS_LEAF(np))
4395 mc->mc_db->md_leaf_pages++;
4396 else if (IS_OVERFLOW(np)) {
4397 mc->mc_db->md_overflow_pages += num;
4404 /** Calculate the size of a leaf node.
4405 * The size depends on the environment's page size; if a data item
4406 * is too large it will be put onto an overflow page and the node
4407 * size will only include the key and not the data. Sizes are always
4408 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4409 * of the #MDB_node headers.
4410 * @param[in] env The environment handle.
4411 * @param[in] key The key for the node.
4412 * @param[in] data The data for the node.
4413 * @return The number of bytes needed to store the node.
4416 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4420 sz = LEAFSIZE(key, data);
4421 if (sz >= env->me_psize / MDB_MINKEYS) {
4422 /* put on overflow page */
4423 sz -= data->mv_size - sizeof(pgno_t);
4427 return sz + sizeof(indx_t);
4430 /** Calculate the size of a branch node.
4431 * The size should depend on the environment's page size but since
4432 * we currently don't support spilling large keys onto overflow
4433 * pages, it's simply the size of the #MDB_node header plus the
4434 * size of the key. Sizes are always rounded up to an even number
4435 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4436 * @param[in] env The environment handle.
4437 * @param[in] key The key for the node.
4438 * @return The number of bytes needed to store the node.
4441 mdb_branch_size(MDB_env *env, MDB_val *key)
4446 if (sz >= env->me_psize / MDB_MINKEYS) {
4447 /* put on overflow page */
4448 /* not implemented */
4449 /* sz -= key->size - sizeof(pgno_t); */
4452 return sz + sizeof(indx_t);
4455 /** Add a node to the page pointed to by the cursor.
4456 * @param[in] mc The cursor for this operation.
4457 * @param[in] indx The index on the page where the new node should be added.
4458 * @param[in] key The key for the new node.
4459 * @param[in] data The data for the new node, if any.
4460 * @param[in] pgno The page number, if adding a branch node.
4461 * @param[in] flags Flags for the node.
4462 * @return 0 on success, non-zero on failure. Possible errors are:
4464 * <li>ENOMEM - failed to allocate overflow pages for the node.
4465 * <li>ENOSPC - there is insufficient room in the page. This error
4466 * should never happen since all callers already calculate the
4467 * page's free space before calling this function.
4471 mdb_node_add(MDB_cursor *mc, indx_t indx,
4472 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4475 size_t node_size = NODESIZE;
4478 MDB_page *mp = mc->mc_pg[mc->mc_top];
4479 MDB_page *ofp = NULL; /* overflow page */
4482 assert(mp->mp_upper >= mp->mp_lower);
4484 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4485 IS_LEAF(mp) ? "leaf" : "branch",
4486 IS_SUBP(mp) ? "sub-" : "",
4487 mp->mp_pgno, indx, data ? data->mv_size : 0,
4488 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4491 /* Move higher keys up one slot. */
4492 int ksize = mc->mc_db->md_pad, dif;
4493 char *ptr = LEAF2KEY(mp, indx, ksize);
4494 dif = NUMKEYS(mp) - indx;
4496 memmove(ptr+ksize, ptr, dif*ksize);
4497 /* insert new key */
4498 memcpy(ptr, key->mv_data, ksize);
4500 /* Just using these for counting */
4501 mp->mp_lower += sizeof(indx_t);
4502 mp->mp_upper -= ksize - sizeof(indx_t);
4507 node_size += key->mv_size;
4511 if (F_ISSET(flags, F_BIGDATA)) {
4512 /* Data already on overflow page. */
4513 node_size += sizeof(pgno_t);
4514 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4515 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4516 /* Put data on overflow page. */
4517 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4518 data->mv_size, node_size+data->mv_size);
4519 node_size += sizeof(pgno_t);
4520 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4522 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4525 node_size += data->mv_size;
4528 node_size += node_size & 1;
4530 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4531 DPRINTF("not enough room in page %zu, got %u ptrs",
4532 mp->mp_pgno, NUMKEYS(mp));
4533 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4534 mp->mp_upper - mp->mp_lower);
4535 DPRINTF("node size = %zu", node_size);
4539 /* Move higher pointers up one slot. */
4540 for (i = NUMKEYS(mp); i > indx; i--)
4541 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4543 /* Adjust free space offsets. */
4544 ofs = mp->mp_upper - node_size;
4545 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4546 mp->mp_ptrs[indx] = ofs;
4548 mp->mp_lower += sizeof(indx_t);
4550 /* Write the node data. */
4551 node = NODEPTR(mp, indx);
4552 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4553 node->mn_flags = flags;
4555 SETDSZ(node,data->mv_size);
4560 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4565 if (F_ISSET(flags, F_BIGDATA))
4566 memcpy(node->mn_data + key->mv_size, data->mv_data,
4568 else if (F_ISSET(flags, MDB_RESERVE))
4569 data->mv_data = node->mn_data + key->mv_size;
4571 memcpy(node->mn_data + key->mv_size, data->mv_data,
4574 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4576 if (F_ISSET(flags, MDB_RESERVE))
4577 data->mv_data = METADATA(ofp);
4579 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4586 /** Delete the specified node from a page.
4587 * @param[in] mp The page to operate on.
4588 * @param[in] indx The index of the node to delete.
4589 * @param[in] ksize The size of a node. Only used if the page is
4590 * part of a #MDB_DUPFIXED database.
4593 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4596 indx_t i, j, numkeys, ptr;
4603 COPY_PGNO(pgno, mp->mp_pgno);
4604 DPRINTF("delete node %u on %s page %zu", indx,
4605 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4608 assert(indx < NUMKEYS(mp));
4611 int x = NUMKEYS(mp) - 1 - indx;
4612 base = LEAF2KEY(mp, indx, ksize);
4614 memmove(base, base + ksize, x * ksize);
4615 mp->mp_lower -= sizeof(indx_t);
4616 mp->mp_upper += ksize - sizeof(indx_t);
4620 node = NODEPTR(mp, indx);
4621 sz = NODESIZE + node->mn_ksize;
4623 if (F_ISSET(node->mn_flags, F_BIGDATA))
4624 sz += sizeof(pgno_t);
4626 sz += NODEDSZ(node);
4630 ptr = mp->mp_ptrs[indx];
4631 numkeys = NUMKEYS(mp);
4632 for (i = j = 0; i < numkeys; i++) {
4634 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4635 if (mp->mp_ptrs[i] < ptr)
4636 mp->mp_ptrs[j] += sz;
4641 base = (char *)mp + mp->mp_upper;
4642 memmove(base + sz, base, ptr - mp->mp_upper);
4644 mp->mp_lower -= sizeof(indx_t);
4648 /** Compact the main page after deleting a node on a subpage.
4649 * @param[in] mp The main page to operate on.
4650 * @param[in] indx The index of the subpage on the main page.
4653 mdb_node_shrink(MDB_page *mp, indx_t indx)
4660 indx_t i, numkeys, ptr;
4662 node = NODEPTR(mp, indx);
4663 sp = (MDB_page *)NODEDATA(node);
4664 osize = NODEDSZ(node);
4666 delta = sp->mp_upper - sp->mp_lower;
4667 SETDSZ(node, osize - delta);
4668 xp = (MDB_page *)((char *)sp + delta);
4670 /* shift subpage upward */
4672 nsize = NUMKEYS(sp) * sp->mp_pad;
4673 memmove(METADATA(xp), METADATA(sp), nsize);
4676 nsize = osize - sp->mp_upper;
4677 numkeys = NUMKEYS(sp);
4678 for (i=numkeys-1; i>=0; i--)
4679 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4681 xp->mp_upper = sp->mp_lower;
4682 xp->mp_lower = sp->mp_lower;
4683 xp->mp_flags = sp->mp_flags;
4684 xp->mp_pad = sp->mp_pad;
4685 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4687 /* shift lower nodes upward */
4688 ptr = mp->mp_ptrs[indx];
4689 numkeys = NUMKEYS(mp);
4690 for (i = 0; i < numkeys; i++) {
4691 if (mp->mp_ptrs[i] <= ptr)
4692 mp->mp_ptrs[i] += delta;
4695 base = (char *)mp + mp->mp_upper;
4696 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4697 mp->mp_upper += delta;
4700 /** Initial setup of a sorted-dups cursor.
4701 * Sorted duplicates are implemented as a sub-database for the given key.
4702 * The duplicate data items are actually keys of the sub-database.
4703 * Operations on the duplicate data items are performed using a sub-cursor
4704 * initialized when the sub-database is first accessed. This function does
4705 * the preliminary setup of the sub-cursor, filling in the fields that
4706 * depend only on the parent DB.
4707 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4710 mdb_xcursor_init0(MDB_cursor *mc)
4712 MDB_xcursor *mx = mc->mc_xcursor;
4714 mx->mx_cursor.mc_xcursor = NULL;
4715 mx->mx_cursor.mc_txn = mc->mc_txn;
4716 mx->mx_cursor.mc_db = &mx->mx_db;
4717 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
4718 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
4719 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
4720 mx->mx_cursor.mc_snum = 0;
4721 mx->mx_cursor.mc_flags = C_SUB;
4722 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
4723 mx->mx_dbx.md_dcmp = NULL;
4724 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
4727 /** Final setup of a sorted-dups cursor.
4728 * Sets up the fields that depend on the data from the main cursor.
4729 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4730 * @param[in] node The data containing the #MDB_db record for the
4731 * sorted-dup database.
4734 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
4736 MDB_xcursor *mx = mc->mc_xcursor;
4738 if (node->mn_flags & F_SUBDATA) {
4739 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
4740 mx->mx_cursor.mc_snum = 0;
4741 mx->mx_cursor.mc_flags = C_SUB;
4743 MDB_page *fp = NODEDATA(node);
4744 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
4745 mx->mx_db.md_flags = 0;
4746 mx->mx_db.md_depth = 1;
4747 mx->mx_db.md_branch_pages = 0;
4748 mx->mx_db.md_leaf_pages = 1;
4749 mx->mx_db.md_overflow_pages = 0;
4750 mx->mx_db.md_entries = NUMKEYS(fp);
4751 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
4752 mx->mx_cursor.mc_snum = 1;
4753 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
4754 mx->mx_cursor.mc_top = 0;
4755 mx->mx_cursor.mc_pg[0] = fp;
4756 mx->mx_cursor.mc_ki[0] = 0;
4757 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4758 mx->mx_db.md_flags = MDB_DUPFIXED;
4759 mx->mx_db.md_pad = fp->mp_pad;
4760 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4761 mx->mx_db.md_flags |= MDB_INTEGERKEY;
4764 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
4766 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
4768 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
4769 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
4770 #if UINT_MAX < SIZE_MAX
4771 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
4772 #ifdef MISALIGNED_OK
4773 mx->mx_dbx.md_cmp = mdb_cmp_long;
4775 mx->mx_dbx.md_cmp = mdb_cmp_cint;
4780 /** Initialize a cursor for a given transaction and database. */
4782 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
4787 mc->mc_db = &txn->mt_dbs[dbi];
4788 mc->mc_dbx = &txn->mt_dbxs[dbi];
4789 mc->mc_dbflag = &txn->mt_dbflags[dbi];
4793 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4795 mc->mc_xcursor = mx;
4796 mdb_xcursor_init0(mc);
4798 mc->mc_xcursor = NULL;
4803 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
4806 MDB_xcursor *mx = NULL;
4807 size_t size = sizeof(MDB_cursor);
4809 if (txn == NULL || ret == NULL || !dbi || dbi >= txn->mt_numdbs)
4812 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
4813 size += sizeof(MDB_xcursor);
4815 if ((mc = malloc(size)) != NULL) {
4816 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4817 mx = (MDB_xcursor *)(mc + 1);
4819 mdb_cursor_init(mc, txn, dbi, mx);
4820 if (txn->mt_cursors) {
4821 mc->mc_next = txn->mt_cursors[dbi];
4822 txn->mt_cursors[dbi] = mc;
4824 mc->mc_flags |= C_ALLOCD;
4834 /* Return the count of duplicate data items for the current key */
4836 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
4840 if (mc == NULL || countp == NULL)
4843 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
4846 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4847 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4850 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
4853 *countp = mc->mc_xcursor->mx_db.md_entries;
4859 mdb_cursor_close(MDB_cursor *mc)
4862 /* remove from txn, if tracked */
4863 if (mc->mc_txn->mt_cursors) {
4864 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
4865 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
4867 *prev = mc->mc_next;
4869 if (mc->mc_flags & C_ALLOCD)
4875 mdb_cursor_txn(MDB_cursor *mc)
4877 if (!mc) return NULL;
4882 mdb_cursor_dbi(MDB_cursor *mc)
4888 /** Replace the key for a node with a new key.
4889 * @param[in] mp The page containing the node to operate on.
4890 * @param[in] indx The index of the node to operate on.
4891 * @param[in] key The new key to use.
4892 * @return 0 on success, non-zero on failure.
4895 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
4897 indx_t ptr, i, numkeys;
4904 node = NODEPTR(mp, indx);
4905 ptr = mp->mp_ptrs[indx];
4909 char kbuf2[(MAXKEYSIZE*2+1)];
4910 k2.mv_data = NODEKEY(node);
4911 k2.mv_size = node->mn_ksize;
4912 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
4914 mdb_dkey(&k2, kbuf2),
4920 delta = key->mv_size - node->mn_ksize;
4922 if (delta > 0 && SIZELEFT(mp) < delta) {
4923 DPRINTF("OUCH! Not enough room, delta = %d", delta);
4927 numkeys = NUMKEYS(mp);
4928 for (i = 0; i < numkeys; i++) {
4929 if (mp->mp_ptrs[i] <= ptr)
4930 mp->mp_ptrs[i] -= delta;
4933 base = (char *)mp + mp->mp_upper;
4934 len = ptr - mp->mp_upper + NODESIZE;
4935 memmove(base - delta, base, len);
4936 mp->mp_upper -= delta;
4938 node = NODEPTR(mp, indx);
4939 node->mn_ksize = key->mv_size;
4943 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4948 /** Move a node from csrc to cdst.
4951 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
4957 unsigned short flags;
4961 /* Mark src and dst as dirty. */
4962 if ((rc = mdb_page_touch(csrc)) ||
4963 (rc = mdb_page_touch(cdst)))
4966 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
4967 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
4968 key.mv_size = csrc->mc_db->md_pad;
4969 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
4971 data.mv_data = NULL;
4975 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
4976 assert(!((long)srcnode&1));
4977 srcpg = NODEPGNO(srcnode);
4978 flags = srcnode->mn_flags;
4979 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
4980 unsigned int snum = csrc->mc_snum;
4982 /* must find the lowest key below src */
4983 mdb_page_search_root(csrc, NULL, 0);
4984 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
4985 key.mv_size = NODEKSZ(s2);
4986 key.mv_data = NODEKEY(s2);
4987 csrc->mc_snum = snum--;
4988 csrc->mc_top = snum;
4990 key.mv_size = NODEKSZ(srcnode);
4991 key.mv_data = NODEKEY(srcnode);
4993 data.mv_size = NODEDSZ(srcnode);
4994 data.mv_data = NODEDATA(srcnode);
4996 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
4997 unsigned int snum = cdst->mc_snum;
5000 /* must find the lowest key below dst */
5001 mdb_page_search_root(cdst, NULL, 0);
5002 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5003 bkey.mv_size = NODEKSZ(s2);
5004 bkey.mv_data = NODEKEY(s2);
5005 cdst->mc_snum = snum--;
5006 cdst->mc_top = snum;
5007 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5010 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5011 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5012 csrc->mc_ki[csrc->mc_top],
5014 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5015 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5017 /* Add the node to the destination page.
5019 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5020 if (rc != MDB_SUCCESS)
5023 /* Delete the node from the source page.
5025 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5028 /* Adjust other cursors pointing to mp */
5029 MDB_cursor *m2, *m3;
5030 MDB_dbi dbi = csrc->mc_dbi;
5031 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5033 if (csrc->mc_flags & C_SUB)
5036 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5037 if (m2 == csrc) continue;
5038 if (csrc->mc_flags & C_SUB)
5039 m3 = &m2->mc_xcursor->mx_cursor;
5042 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5043 csrc->mc_ki[csrc->mc_top]) {
5044 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5045 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5050 /* Update the parent separators.
5052 if (csrc->mc_ki[csrc->mc_top] == 0) {
5053 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5054 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5055 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5057 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5058 key.mv_size = NODEKSZ(srcnode);
5059 key.mv_data = NODEKEY(srcnode);
5061 DPRINTF("update separator for source page %zu to [%s]",
5062 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5063 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5064 &key)) != MDB_SUCCESS)
5067 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5069 nullkey.mv_size = 0;
5070 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5071 assert(rc == MDB_SUCCESS);
5075 if (cdst->mc_ki[cdst->mc_top] == 0) {
5076 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5077 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5078 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5080 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5081 key.mv_size = NODEKSZ(srcnode);
5082 key.mv_data = NODEKEY(srcnode);
5084 DPRINTF("update separator for destination page %zu to [%s]",
5085 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5086 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5087 &key)) != MDB_SUCCESS)
5090 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5092 nullkey.mv_size = 0;
5093 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5094 assert(rc == MDB_SUCCESS);
5101 /** Merge one page into another.
5102 * The nodes from the page pointed to by \b csrc will
5103 * be copied to the page pointed to by \b cdst and then
5104 * the \b csrc page will be freed.
5105 * @param[in] csrc Cursor pointing to the source page.
5106 * @param[in] cdst Cursor pointing to the destination page.
5109 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5117 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5118 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5120 assert(csrc->mc_snum > 1); /* can't merge root page */
5121 assert(cdst->mc_snum > 1);
5123 /* Mark dst as dirty. */
5124 if ((rc = mdb_page_touch(cdst)))
5127 /* Move all nodes from src to dst.
5129 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5130 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5131 key.mv_size = csrc->mc_db->md_pad;
5132 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5133 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5134 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5135 if (rc != MDB_SUCCESS)
5137 key.mv_data = (char *)key.mv_data + key.mv_size;
5140 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5141 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5142 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5143 unsigned int snum = csrc->mc_snum;
5145 /* must find the lowest key below src */
5146 mdb_page_search_root(csrc, NULL, 0);
5147 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5148 key.mv_size = NODEKSZ(s2);
5149 key.mv_data = NODEKEY(s2);
5150 csrc->mc_snum = snum--;
5151 csrc->mc_top = snum;
5153 key.mv_size = srcnode->mn_ksize;
5154 key.mv_data = NODEKEY(srcnode);
5157 data.mv_size = NODEDSZ(srcnode);
5158 data.mv_data = NODEDATA(srcnode);
5159 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5160 if (rc != MDB_SUCCESS)
5165 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5166 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);
5168 /* Unlink the src page from parent and add to free list.
5170 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5171 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5173 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5177 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5178 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5179 csrc->mc_db->md_leaf_pages--;
5181 csrc->mc_db->md_branch_pages--;
5183 /* Adjust other cursors pointing to mp */
5184 MDB_cursor *m2, *m3;
5185 MDB_dbi dbi = csrc->mc_dbi;
5186 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5188 if (csrc->mc_flags & C_SUB)
5191 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5192 if (csrc->mc_flags & C_SUB)
5193 m3 = &m2->mc_xcursor->mx_cursor;
5196 if (m3 == csrc) continue;
5197 if (m3->mc_snum < csrc->mc_snum) continue;
5198 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5199 m3->mc_pg[csrc->mc_top] = mp;
5200 m3->mc_ki[csrc->mc_top] += nkeys;
5204 mdb_cursor_pop(csrc);
5206 return mdb_rebalance(csrc);
5209 /** Copy the contents of a cursor.
5210 * @param[in] csrc The cursor to copy from.
5211 * @param[out] cdst The cursor to copy to.
5214 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5218 cdst->mc_txn = csrc->mc_txn;
5219 cdst->mc_dbi = csrc->mc_dbi;
5220 cdst->mc_db = csrc->mc_db;
5221 cdst->mc_dbx = csrc->mc_dbx;
5222 cdst->mc_snum = csrc->mc_snum;
5223 cdst->mc_top = csrc->mc_top;
5224 cdst->mc_flags = csrc->mc_flags;
5226 for (i=0; i<csrc->mc_snum; i++) {
5227 cdst->mc_pg[i] = csrc->mc_pg[i];
5228 cdst->mc_ki[i] = csrc->mc_ki[i];
5232 /** Rebalance the tree after a delete operation.
5233 * @param[in] mc Cursor pointing to the page where rebalancing
5235 * @return 0 on success, non-zero on failure.
5238 mdb_rebalance(MDB_cursor *mc)
5248 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5249 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5250 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5251 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5255 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5258 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5259 DPRINTF("no need to rebalance page %zu, above fill threshold",
5265 if (mc->mc_snum < 2) {
5266 MDB_page *mp = mc->mc_pg[0];
5267 if (NUMKEYS(mp) == 0) {
5268 DPUTS("tree is completely empty");
5269 mc->mc_db->md_root = P_INVALID;
5270 mc->mc_db->md_depth = 0;
5271 mc->mc_db->md_leaf_pages = 0;
5272 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5276 /* Adjust other cursors pointing to mp */
5277 MDB_cursor *m2, *m3;
5278 MDB_dbi dbi = mc->mc_dbi;
5280 if (mc->mc_flags & C_SUB)
5283 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5284 if (m2 == mc) continue;
5285 if (mc->mc_flags & C_SUB)
5286 m3 = &m2->mc_xcursor->mx_cursor;
5289 if (m3->mc_snum < mc->mc_snum) continue;
5290 if (m3->mc_pg[0] == mp) {
5296 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5297 DPUTS("collapsing root page!");
5298 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5299 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5300 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5303 mc->mc_db->md_depth--;
5304 mc->mc_db->md_branch_pages--;
5306 /* Adjust other cursors pointing to mp */
5307 MDB_cursor *m2, *m3;
5308 MDB_dbi dbi = mc->mc_dbi;
5310 if (mc->mc_flags & C_SUB)
5313 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5314 if (m2 == mc) continue;
5315 if (mc->mc_flags & C_SUB)
5316 m3 = &m2->mc_xcursor->mx_cursor;
5319 if (m3->mc_snum < mc->mc_snum) continue;
5320 if (m3->mc_pg[0] == mp) {
5321 m3->mc_pg[0] = mc->mc_pg[0];
5326 DPUTS("root page doesn't need rebalancing");
5330 /* The parent (branch page) must have at least 2 pointers,
5331 * otherwise the tree is invalid.
5333 ptop = mc->mc_top-1;
5334 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5336 /* Leaf page fill factor is below the threshold.
5337 * Try to move keys from left or right neighbor, or
5338 * merge with a neighbor page.
5343 mdb_cursor_copy(mc, &mn);
5344 mn.mc_xcursor = NULL;
5346 if (mc->mc_ki[ptop] == 0) {
5347 /* We're the leftmost leaf in our parent.
5349 DPUTS("reading right neighbor");
5351 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5352 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5354 mn.mc_ki[mn.mc_top] = 0;
5355 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5357 /* There is at least one neighbor to the left.
5359 DPUTS("reading left neighbor");
5361 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5362 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5364 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5365 mc->mc_ki[mc->mc_top] = 0;
5368 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5369 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);
5371 /* If the neighbor page is above threshold and has at least two
5372 * keys, move one key from it.
5374 * Otherwise we should try to merge them.
5376 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5377 return mdb_node_move(&mn, mc);
5378 else { /* FIXME: if (has_enough_room()) */
5379 mc->mc_flags &= ~C_INITIALIZED;
5380 if (mc->mc_ki[ptop] == 0)
5381 return mdb_page_merge(&mn, mc);
5383 return mdb_page_merge(mc, &mn);
5387 /** Complete a delete operation started by #mdb_cursor_del(). */
5389 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5393 /* add overflow pages to free list */
5394 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5398 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5399 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5400 for (i=0; i<ovpages; i++) {
5401 DPRINTF("freed ov page %zu", pg);
5402 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5406 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5407 mc->mc_db->md_entries--;
5408 rc = mdb_rebalance(mc);
5409 if (rc != MDB_SUCCESS)
5410 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5416 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5417 MDB_val *key, MDB_val *data)
5422 MDB_val rdata, *xdata;
5426 assert(key != NULL);
5428 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5430 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5433 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5437 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5441 mdb_cursor_init(&mc, txn, dbi, &mx);
5452 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5454 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5458 /** Split a page and insert a new node.
5459 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5460 * The cursor will be updated to point to the actual page and index where
5461 * the node got inserted after the split.
5462 * @param[in] newkey The key for the newly inserted node.
5463 * @param[in] newdata The data for the newly inserted node.
5464 * @param[in] newpgno The page number, if the new node is a branch node.
5465 * @return 0 on success, non-zero on failure.
5468 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5469 unsigned int nflags)
5472 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0;
5475 unsigned int i, j, split_indx, nkeys, pmax;
5477 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5479 MDB_page *mp, *rp, *pp;
5484 mp = mc->mc_pg[mc->mc_top];
5485 newindx = mc->mc_ki[mc->mc_top];
5487 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5488 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5489 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5491 if (mc->mc_snum < 2) {
5492 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5494 /* shift current top to make room for new parent */
5495 mc->mc_pg[1] = mc->mc_pg[0];
5496 mc->mc_ki[1] = mc->mc_ki[0];
5499 mc->mc_db->md_root = pp->mp_pgno;
5500 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5501 mc->mc_db->md_depth++;
5504 /* Add left (implicit) pointer. */
5505 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5506 /* undo the pre-push */
5507 mc->mc_pg[0] = mc->mc_pg[1];
5508 mc->mc_ki[0] = mc->mc_ki[1];
5509 mc->mc_db->md_root = mp->mp_pgno;
5510 mc->mc_db->md_depth--;
5517 ptop = mc->mc_top-1;
5518 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5521 /* Create a right sibling. */
5522 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5524 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5526 mdb_cursor_copy(mc, &mn);
5527 mn.mc_pg[mn.mc_top] = rp;
5528 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5530 if (nflags & MDB_APPEND) {
5531 mn.mc_ki[mn.mc_top] = 0;
5538 nkeys = NUMKEYS(mp);
5539 split_indx = nkeys / 2 + 1;
5544 unsigned int lsize, rsize, ksize;
5545 /* Move half of the keys to the right sibling */
5547 x = mc->mc_ki[mc->mc_top] - split_indx;
5548 ksize = mc->mc_db->md_pad;
5549 split = LEAF2KEY(mp, split_indx, ksize);
5550 rsize = (nkeys - split_indx) * ksize;
5551 lsize = (nkeys - split_indx) * sizeof(indx_t);
5552 mp->mp_lower -= lsize;
5553 rp->mp_lower += lsize;
5554 mp->mp_upper += rsize - lsize;
5555 rp->mp_upper -= rsize - lsize;
5556 sepkey.mv_size = ksize;
5557 if (newindx == split_indx) {
5558 sepkey.mv_data = newkey->mv_data;
5560 sepkey.mv_data = split;
5563 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5564 memcpy(rp->mp_ptrs, split, rsize);
5565 sepkey.mv_data = rp->mp_ptrs;
5566 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5567 memcpy(ins, newkey->mv_data, ksize);
5568 mp->mp_lower += sizeof(indx_t);
5569 mp->mp_upper -= ksize - sizeof(indx_t);
5572 memcpy(rp->mp_ptrs, split, x * ksize);
5573 ins = LEAF2KEY(rp, x, ksize);
5574 memcpy(ins, newkey->mv_data, ksize);
5575 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5576 rp->mp_lower += sizeof(indx_t);
5577 rp->mp_upper -= ksize - sizeof(indx_t);
5578 mc->mc_ki[mc->mc_top] = x;
5579 mc->mc_pg[mc->mc_top] = rp;
5584 /* For leaf pages, check the split point based on what
5585 * fits where, since otherwise add_node can fail.
5588 unsigned int psize, nsize;
5589 /* Maximum free space in an empty page */
5590 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5591 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5592 if (newindx < split_indx) {
5594 for (i=0; 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);
5609 for (i=nkeys-1; i>=split_indx; i--) {
5610 node = NODEPTR(mp, i);
5611 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5612 if (F_ISSET(node->mn_flags, F_BIGDATA))
5613 psize += sizeof(pgno_t);
5615 psize += NODEDSZ(node);
5625 /* First find the separating key between the split pages.
5627 if (newindx == split_indx) {
5628 sepkey.mv_size = newkey->mv_size;
5629 sepkey.mv_data = newkey->mv_data;
5631 node = NODEPTR(mp, split_indx);
5632 sepkey.mv_size = node->mn_ksize;
5633 sepkey.mv_data = NODEKEY(node);
5637 DPRINTF("separator is [%s]", DKEY(&sepkey));
5639 /* Copy separator key to the parent.
5641 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5644 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5646 /* Right page might now have changed parent.
5647 * Check if left page also changed parent.
5649 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
5650 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
5651 mc->mc_pg[ptop] = mn.mc_pg[ptop];
5652 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
5656 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
5659 if (rc != MDB_SUCCESS) {
5662 if (nflags & MDB_APPEND) {
5663 mc->mc_pg[mc->mc_top] = rp;
5664 mc->mc_ki[mc->mc_top] = 0;
5665 return mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
5671 /* Move half of the keys to the right sibling. */
5673 /* grab a page to hold a temporary copy */
5674 copy = mdb_page_malloc(mc);
5678 copy->mp_pgno = mp->mp_pgno;
5679 copy->mp_flags = mp->mp_flags;
5680 copy->mp_lower = PAGEHDRSZ;
5681 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
5682 mc->mc_pg[mc->mc_top] = copy;
5683 for (i = j = 0; i <= nkeys; j++) {
5684 if (i == split_indx) {
5685 /* Insert in right sibling. */
5686 /* Reset insert index for right sibling. */
5687 j = (i == newindx && ins_new);
5688 mc->mc_pg[mc->mc_top] = rp;
5691 if (i == newindx && !ins_new) {
5692 /* Insert the original entry that caused the split. */
5693 rkey.mv_data = newkey->mv_data;
5694 rkey.mv_size = newkey->mv_size;
5703 /* Update page and index for the new key. */
5705 mc->mc_pg[mc->mc_top] = copy;
5706 mc->mc_ki[mc->mc_top] = j;
5707 } else if (i == nkeys) {
5710 node = NODEPTR(mp, i);
5711 rkey.mv_data = NODEKEY(node);
5712 rkey.mv_size = node->mn_ksize;
5714 xdata.mv_data = NODEDATA(node);
5715 xdata.mv_size = NODEDSZ(node);
5718 pgno = NODEPGNO(node);
5719 flags = node->mn_flags;
5724 if (!IS_LEAF(mp) && j == 0) {
5725 /* First branch index doesn't need key data. */
5729 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
5732 nkeys = NUMKEYS(copy);
5733 for (i=0; i<nkeys; i++)
5734 mp->mp_ptrs[i] = copy->mp_ptrs[i];
5735 mp->mp_lower = copy->mp_lower;
5736 mp->mp_upper = copy->mp_upper;
5737 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
5738 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
5740 /* reset back to original page */
5741 if (!newindx || (newindx < split_indx)) {
5742 mc->mc_pg[mc->mc_top] = mp;
5743 if (nflags & MDB_RESERVE) {
5744 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5745 if (!(node->mn_flags & F_BIGDATA))
5746 newdata->mv_data = NODEDATA(node);
5750 /* return tmp page to freelist */
5751 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
5752 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
5753 mc->mc_txn->mt_env->me_dpages = copy;
5756 /* Adjust other cursors pointing to mp */
5757 MDB_cursor *m2, *m3;
5758 MDB_dbi dbi = mc->mc_dbi;
5760 if (mc->mc_flags & C_SUB)
5763 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5764 if (m2 == mc) continue;
5765 if (mc->mc_flags & C_SUB)
5766 m3 = &m2->mc_xcursor->mx_cursor;
5769 if (!(m3->mc_flags & C_INITIALIZED))
5773 for (i=m3->mc_top; i>0; i--) {
5774 m3->mc_ki[i+1] = m3->mc_ki[i];
5775 m3->mc_pg[i+1] = m3->mc_pg[i];
5777 m3->mc_ki[0] = mc->mc_ki[0];
5778 m3->mc_pg[0] = mc->mc_pg[0];
5782 if (m3->mc_pg[mc->mc_top] == mp) {
5783 if (m3->mc_ki[m3->mc_top] >= split_indx) {
5784 m3->mc_pg[m3->mc_top] = rp;
5785 m3->mc_ki[m3->mc_top] -= split_indx;
5794 mdb_put(MDB_txn *txn, MDB_dbi dbi,
5795 MDB_val *key, MDB_val *data, unsigned int flags)
5800 assert(key != NULL);
5801 assert(data != NULL);
5803 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5806 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5810 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5814 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
5817 mdb_cursor_init(&mc, txn, dbi, &mx);
5818 return mdb_cursor_put(&mc, key, data, flags);
5821 /** Only a subset of the @ref mdb_env flags can be changed
5822 * at runtime. Changing other flags requires closing the environment
5823 * and re-opening it with the new flags.
5825 #define CHANGEABLE (MDB_NOSYNC)
5827 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
5829 if ((flag & CHANGEABLE) != flag)
5832 env->me_flags |= flag;
5834 env->me_flags &= ~flag;
5839 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
5844 *arg = env->me_flags;
5849 mdb_env_get_path(MDB_env *env, const char **arg)
5854 *arg = env->me_path;
5858 /** Common code for #mdb_stat() and #mdb_env_stat().
5859 * @param[in] env the environment to operate in.
5860 * @param[in] db the #MDB_db record containing the stats to return.
5861 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
5862 * @return 0, this function always succeeds.
5865 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
5867 arg->ms_psize = env->me_psize;
5868 arg->ms_depth = db->md_depth;
5869 arg->ms_branch_pages = db->md_branch_pages;
5870 arg->ms_leaf_pages = db->md_leaf_pages;
5871 arg->ms_overflow_pages = db->md_overflow_pages;
5872 arg->ms_entries = db->md_entries;
5877 mdb_env_stat(MDB_env *env, MDB_stat *arg)
5881 if (env == NULL || arg == NULL)
5884 mdb_env_read_meta(env, &toggle);
5886 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
5889 /** Set the default comparison functions for a database.
5890 * Called immediately after a database is opened to set the defaults.
5891 * The user can then override them with #mdb_set_compare() or
5892 * #mdb_set_dupsort().
5893 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
5894 * @param[in] dbi A database handle returned by #mdb_open()
5897 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
5899 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
5900 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memnr;
5901 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
5902 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_cint;
5904 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memn;
5906 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5907 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
5908 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
5909 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_int;
5911 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_cint;
5912 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
5913 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memnr;
5915 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memn;
5918 txn->mt_dbxs[dbi].md_dcmp = NULL;
5922 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
5927 int rc, dbflag, exact;
5930 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
5931 mdb_default_cmp(txn, FREE_DBI);
5937 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
5938 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
5939 mdb_default_cmp(txn, MAIN_DBI);
5943 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
5944 mdb_default_cmp(txn, MAIN_DBI);
5947 /* Is the DB already open? */
5949 for (i=2; i<txn->mt_numdbs; i++) {
5950 if (len == txn->mt_dbxs[i].md_name.mv_size &&
5951 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
5957 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
5960 /* Find the DB info */
5964 key.mv_data = (void *)name;
5965 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
5966 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
5967 if (rc == MDB_SUCCESS) {
5968 /* make sure this is actually a DB */
5969 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
5970 if (!(node->mn_flags & F_SUBDATA))
5972 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
5973 /* Create if requested */
5975 data.mv_size = sizeof(MDB_db);
5976 data.mv_data = &dummy;
5977 memset(&dummy, 0, sizeof(dummy));
5978 dummy.md_root = P_INVALID;
5979 dummy.md_flags = flags & 0xffff;
5980 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
5984 /* OK, got info, add to table */
5985 if (rc == MDB_SUCCESS) {
5986 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
5987 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
5988 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
5989 txn->mt_dbflags[txn->mt_numdbs] = dbflag;
5990 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
5991 *dbi = txn->mt_numdbs;
5992 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
5993 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
5994 mdb_default_cmp(txn, txn->mt_numdbs);
6001 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6003 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6006 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6009 void mdb_close(MDB_env *env, MDB_dbi dbi)
6012 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6014 ptr = env->me_dbxs[dbi].md_name.mv_data;
6015 env->me_dbxs[dbi].md_name.mv_data = NULL;
6016 env->me_dbxs[dbi].md_name.mv_size = 0;
6020 /** Add all the DB's pages to the free list.
6021 * @param[in] mc Cursor on the DB to free.
6022 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6023 * @return 0 on success, non-zero on failure.
6026 mdb_drop0(MDB_cursor *mc, int subs)
6030 rc = mdb_page_search(mc, NULL, 0);
6031 if (rc == MDB_SUCCESS) {
6036 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6037 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6040 mdb_cursor_copy(mc, &mx);
6041 while (mc->mc_snum > 0) {
6042 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6043 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6044 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6045 if (ni->mn_flags & F_SUBDATA) {
6046 mdb_xcursor_init1(mc, ni);
6047 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6053 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6055 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6058 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6063 rc = mdb_cursor_sibling(mc, 1);
6065 /* no more siblings, go back to beginning
6066 * of previous level. (stack was already popped
6067 * by mdb_cursor_sibling)
6069 for (i=1; i<mc->mc_top; i++)
6070 mc->mc_pg[i] = mx.mc_pg[i];
6074 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6075 mc->mc_db->md_root);
6080 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6085 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6088 rc = mdb_cursor_open(txn, dbi, &mc);
6092 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6096 /* Can't delete the main DB */
6097 if (del && dbi > MAIN_DBI) {
6098 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6100 mdb_close(txn->mt_env, dbi);
6102 txn->mt_dbflags[dbi] |= DB_DIRTY;
6103 txn->mt_dbs[dbi].md_depth = 0;
6104 txn->mt_dbs[dbi].md_branch_pages = 0;
6105 txn->mt_dbs[dbi].md_leaf_pages = 0;
6106 txn->mt_dbs[dbi].md_overflow_pages = 0;
6107 txn->mt_dbs[dbi].md_entries = 0;
6108 txn->mt_dbs[dbi].md_root = P_INVALID;
6111 mdb_cursor_close(mc);
6115 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6117 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6120 txn->mt_dbxs[dbi].md_cmp = cmp;
6124 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6126 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6129 txn->mt_dbxs[dbi].md_dcmp = cmp;
6133 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6135 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6138 txn->mt_dbxs[dbi].md_rel = rel;
6142 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6144 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6147 txn->mt_dbxs[dbi].md_relctx = ctx;