2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <netinet/in.h>
63 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
66 #if defined(__APPLE__) || defined (BSD)
67 # define MDB_USE_POSIX_SEM 1
68 # define MDB_FDATASYNC fsync
69 #elif defined(ANDROID)
70 # define MDB_FDATASYNC fsync
75 #ifdef MDB_USE_POSIX_SEM
76 #include <semaphore.h>
81 #include <valgrind/memcheck.h>
82 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
83 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
84 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
85 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
86 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
88 #define VGMEMP_CREATE(h,r,z)
89 #define VGMEMP_ALLOC(h,a,s)
90 #define VGMEMP_FREE(h,a)
91 #define VGMEMP_DESTROY(h)
92 #define VGMEMP_DEFINED(a,s)
96 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
97 /* Solaris just defines one or the other */
98 # define LITTLE_ENDIAN 1234
99 # define BIG_ENDIAN 4321
100 # ifdef _LITTLE_ENDIAN
101 # define BYTE_ORDER LITTLE_ENDIAN
103 # define BYTE_ORDER BIG_ENDIAN
106 # define BYTE_ORDER __BYTE_ORDER
110 #ifndef LITTLE_ENDIAN
111 #define LITTLE_ENDIAN __LITTLE_ENDIAN
114 #define BIG_ENDIAN __BIG_ENDIAN
117 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
118 #define MISALIGNED_OK 1
124 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
125 # error "Unknown or unsupported endianness (BYTE_ORDER)"
126 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
127 # error "Two's complement, reasonably sized integer types, please"
130 /** @defgroup internal MDB Internals
133 /** @defgroup compat Windows Compatibility Macros
134 * A bunch of macros to minimize the amount of platform-specific ifdefs
135 * needed throughout the rest of the code. When the features this library
136 * needs are similar enough to POSIX to be hidden in a one-or-two line
137 * replacement, this macro approach is used.
141 #define pthread_t DWORD
142 #define pthread_mutex_t HANDLE
143 #define pthread_key_t DWORD
144 #define pthread_self() GetCurrentThreadId()
145 #define pthread_key_create(x,y) \
146 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
147 #define pthread_key_delete(x) TlsFree(x)
148 #define pthread_getspecific(x) TlsGetValue(x)
149 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
150 #define pthread_mutex_unlock(x) ReleaseMutex(x)
151 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
152 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
153 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
154 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
155 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
156 #define getpid() GetCurrentProcessId()
157 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
158 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
159 #define ErrCode() GetLastError()
160 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
161 #define close(fd) CloseHandle(fd)
162 #define munmap(ptr,len) UnmapViewOfFile(ptr)
165 #ifdef MDB_USE_POSIX_SEM
167 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
168 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
169 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
170 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
173 mdb_sem_wait(sem_t *sem)
176 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
181 /** Lock the reader mutex.
183 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
184 /** Unlock the reader mutex.
186 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
188 /** Lock the writer mutex.
189 * Only a single write transaction is allowed at a time. Other writers
190 * will block waiting for this mutex.
192 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
193 /** Unlock the writer mutex.
195 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
196 #endif /* MDB_USE_POSIX_SEM */
198 /** Get the error code for the last failed system function.
200 #define ErrCode() errno
202 /** An abstraction for a file handle.
203 * On POSIX systems file handles are small integers. On Windows
204 * they're opaque pointers.
208 /** A value for an invalid file handle.
209 * Mainly used to initialize file variables and signify that they are
212 #define INVALID_HANDLE_VALUE (-1)
214 /** Get the size of a memory page for the system.
215 * This is the basic size that the platform's memory manager uses, and is
216 * fundamental to the use of memory-mapped files.
218 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
221 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
224 #define MNAME_LEN (sizeof(pthread_mutex_t))
230 /** A flag for opening a file and requesting synchronous data writes.
231 * This is only used when writing a meta page. It's not strictly needed;
232 * we could just do a normal write and then immediately perform a flush.
233 * But if this flag is available it saves us an extra system call.
235 * @note If O_DSYNC is undefined but exists in /usr/include,
236 * preferably set some compiler flag to get the definition.
237 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
240 # define MDB_DSYNC O_DSYNC
244 /** Function for flushing the data of a file. Define this to fsync
245 * if fdatasync() is not supported.
247 #ifndef MDB_FDATASYNC
248 # define MDB_FDATASYNC fdatasync
252 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
263 /** A page number in the database.
264 * Note that 64 bit page numbers are overkill, since pages themselves
265 * already represent 12-13 bits of addressable memory, and the OS will
266 * always limit applications to a maximum of 63 bits of address space.
268 * @note In the #MDB_node structure, we only store 48 bits of this value,
269 * which thus limits us to only 60 bits of addressable data.
271 typedef MDB_ID pgno_t;
273 /** A transaction ID.
274 * See struct MDB_txn.mt_txnid for details.
276 typedef MDB_ID txnid_t;
278 /** @defgroup debug Debug Macros
282 /** Enable debug output.
283 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
284 * read from and written to the database (used for free space management).
289 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
292 # define DPRINTF (void) /* Vararg macros may be unsupported */
294 static int mdb_debug;
295 static txnid_t mdb_debug_start;
297 /** Print a debug message with printf formatting. */
298 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
299 ((void) ((mdb_debug) && \
300 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
302 # define DPRINTF(fmt, ...) ((void) 0)
303 # define MDB_DEBUG_SKIP
305 /** Print a debug string.
306 * The string is printed literally, with no format processing.
308 #define DPUTS(arg) DPRINTF("%s", arg)
311 /** A default memory page size.
312 * The actual size is platform-dependent, but we use this for
313 * boot-strapping. We probably should not be using this any more.
314 * The #GET_PAGESIZE() macro is used to get the actual size.
316 * Note that we don't currently support Huge pages. On Linux,
317 * regular data files cannot use Huge pages, and in general
318 * Huge pages aren't actually pageable. We rely on the OS
319 * demand-pager to read our data and page it out when memory
320 * pressure from other processes is high. So until OSs have
321 * actual paging support for Huge pages, they're not viable.
323 #define MDB_PAGESIZE 4096
325 /** The minimum number of keys required in a database page.
326 * Setting this to a larger value will place a smaller bound on the
327 * maximum size of a data item. Data items larger than this size will
328 * be pushed into overflow pages instead of being stored directly in
329 * the B-tree node. This value used to default to 4. With a page size
330 * of 4096 bytes that meant that any item larger than 1024 bytes would
331 * go into an overflow page. That also meant that on average 2-3KB of
332 * each overflow page was wasted space. The value cannot be lower than
333 * 2 because then there would no longer be a tree structure. With this
334 * value, items larger than 2KB will go into overflow pages, and on
335 * average only 1KB will be wasted.
337 #define MDB_MINKEYS 2
339 /** A stamp that identifies a file as an MDB file.
340 * There's nothing special about this value other than that it is easily
341 * recognizable, and it will reflect any byte order mismatches.
343 #define MDB_MAGIC 0xBEEFC0DE
345 /** The version number for a database's file format. */
346 #define MDB_VERSION 1
348 /** @brief The maximum size of a key in the database.
350 * We require that keys all fit onto a regular page. This limit
351 * could be raised a bit further if needed; to something just
352 * under #MDB_PAGESIZE / #MDB_MINKEYS.
354 * Note that data items in an #MDB_DUPSORT database are actually keys
355 * of a subDB, so they're also limited to this size.
357 #ifndef MDB_MAXKEYSIZE
358 #define MDB_MAXKEYSIZE 511
361 /** @brief The maximum size of a data item.
363 * We only store a 32 bit value for node sizes.
365 #define MAXDATASIZE 0xffffffffUL
370 * This is used for printing a hex dump of a key's contents.
372 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
373 /** Display a key in hex.
375 * Invoke a function to display a key in hex.
377 #define DKEY(x) mdb_dkey(x, kbuf)
379 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
383 /** An invalid page number.
384 * Mainly used to denote an empty tree.
386 #define P_INVALID (~(pgno_t)0)
388 /** Test if the flags \b f are set in a flag word \b w. */
389 #define F_ISSET(w, f) (((w) & (f)) == (f))
391 /** Used for offsets within a single page.
392 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
395 typedef uint16_t indx_t;
397 /** Default size of memory map.
398 * This is certainly too small for any actual applications. Apps should always set
399 * the size explicitly using #mdb_env_set_mapsize().
401 #define DEFAULT_MAPSIZE 1048576
403 /** @defgroup readers Reader Lock Table
404 * Readers don't acquire any locks for their data access. Instead, they
405 * simply record their transaction ID in the reader table. The reader
406 * mutex is needed just to find an empty slot in the reader table. The
407 * slot's address is saved in thread-specific data so that subsequent read
408 * transactions started by the same thread need no further locking to proceed.
410 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
412 * No reader table is used if the database is on a read-only filesystem.
414 * Since the database uses multi-version concurrency control, readers don't
415 * actually need any locking. This table is used to keep track of which
416 * readers are using data from which old transactions, so that we'll know
417 * when a particular old transaction is no longer in use. Old transactions
418 * that have discarded any data pages can then have those pages reclaimed
419 * for use by a later write transaction.
421 * The lock table is constructed such that reader slots are aligned with the
422 * processor's cache line size. Any slot is only ever used by one thread.
423 * This alignment guarantees that there will be no contention or cache
424 * thrashing as threads update their own slot info, and also eliminates
425 * any need for locking when accessing a slot.
427 * A writer thread will scan every slot in the table to determine the oldest
428 * outstanding reader transaction. Any freed pages older than this will be
429 * reclaimed by the writer. The writer doesn't use any locks when scanning
430 * this table. This means that there's no guarantee that the writer will
431 * see the most up-to-date reader info, but that's not required for correct
432 * operation - all we need is to know the upper bound on the oldest reader,
433 * we don't care at all about the newest reader. So the only consequence of
434 * reading stale information here is that old pages might hang around a
435 * while longer before being reclaimed. That's actually good anyway, because
436 * the longer we delay reclaiming old pages, the more likely it is that a
437 * string of contiguous pages can be found after coalescing old pages from
438 * many old transactions together.
441 /** Number of slots in the reader table.
442 * This value was chosen somewhat arbitrarily. 126 readers plus a
443 * couple mutexes fit exactly into 8KB on my development machine.
444 * Applications should set the table size using #mdb_env_set_maxreaders().
446 #define DEFAULT_READERS 126
448 /** The size of a CPU cache line in bytes. We want our lock structures
449 * aligned to this size to avoid false cache line sharing in the
451 * This value works for most CPUs. For Itanium this should be 128.
457 /** The information we store in a single slot of the reader table.
458 * In addition to a transaction ID, we also record the process and
459 * thread ID that owns a slot, so that we can detect stale information,
460 * e.g. threads or processes that went away without cleaning up.
461 * @note We currently don't check for stale records. We simply re-init
462 * the table when we know that we're the only process opening the
465 typedef struct MDB_rxbody {
466 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
467 * Multiple readers that start at the same time will probably have the
468 * same ID here. Again, it's not important to exclude them from
469 * anything; all we need to know is which version of the DB they
470 * started from so we can avoid overwriting any data used in that
471 * particular version.
474 /** The process ID of the process owning this reader txn. */
476 /** The thread ID of the thread owning this txn. */
480 /** The actual reader record, with cacheline padding. */
481 typedef struct MDB_reader {
484 /** shorthand for mrb_txnid */
485 #define mr_txnid mru.mrx.mrb_txnid
486 #define mr_pid mru.mrx.mrb_pid
487 #define mr_tid mru.mrx.mrb_tid
488 /** cache line alignment */
489 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
493 /** The header for the reader table.
494 * The table resides in a memory-mapped file. (This is a different file
495 * than is used for the main database.)
497 * For POSIX the actual mutexes reside in the shared memory of this
498 * mapped file. On Windows, mutexes are named objects allocated by the
499 * kernel; we store the mutex names in this mapped file so that other
500 * processes can grab them. This same approach is also used on
501 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
502 * process-shared POSIX mutexes. For these cases where a named object
503 * is used, the object name is derived from a 64 bit FNV hash of the
504 * environment pathname. As such, naming collisions are extremely
505 * unlikely. If a collision occurs, the results are unpredictable.
507 typedef struct MDB_txbody {
508 /** Stamp identifying this as an MDB file. It must be set
511 /** Version number of this lock file. Must be set to #MDB_VERSION. */
512 uint32_t mtb_version;
513 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
514 char mtb_rmname[MNAME_LEN];
516 /** Mutex protecting access to this table.
517 * This is the reader lock that #LOCK_MUTEX_R acquires.
519 pthread_mutex_t mtb_mutex;
521 /** The ID of the last transaction committed to the database.
522 * This is recorded here only for convenience; the value can always
523 * be determined by reading the main database meta pages.
526 /** The number of slots that have been used in the reader table.
527 * This always records the maximum count, it is not decremented
528 * when readers release their slots.
530 unsigned mtb_numreaders;
533 /** The actual reader table definition. */
534 typedef struct MDB_txninfo {
537 #define mti_magic mt1.mtb.mtb_magic
538 #define mti_version mt1.mtb.mtb_version
539 #define mti_mutex mt1.mtb.mtb_mutex
540 #define mti_rmname mt1.mtb.mtb_rmname
541 #define mti_txnid mt1.mtb.mtb_txnid
542 #define mti_numreaders mt1.mtb.mtb_numreaders
543 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
546 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
547 char mt2_wmname[MNAME_LEN];
548 #define mti_wmname mt2.mt2_wmname
550 pthread_mutex_t mt2_wmutex;
551 #define mti_wmutex mt2.mt2_wmutex
553 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
555 MDB_reader mti_readers[1];
559 /** Common header for all page types.
560 * Overflow records occupy a number of contiguous pages with no
561 * headers on any page after the first.
563 typedef struct MDB_page {
564 #define mp_pgno mp_p.p_pgno
565 #define mp_next mp_p.p_next
567 pgno_t p_pgno; /**< page number */
568 void * p_next; /**< for in-memory list of freed structs */
571 /** @defgroup mdb_page Page Flags
573 * Flags for the page headers.
576 #define P_BRANCH 0x01 /**< branch page */
577 #define P_LEAF 0x02 /**< leaf page */
578 #define P_OVERFLOW 0x04 /**< overflow page */
579 #define P_META 0x08 /**< meta page */
580 #define P_DIRTY 0x10 /**< dirty page */
581 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
582 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
584 uint16_t mp_flags; /**< @ref mdb_page */
585 #define mp_lower mp_pb.pb.pb_lower
586 #define mp_upper mp_pb.pb.pb_upper
587 #define mp_pages mp_pb.pb_pages
590 indx_t pb_lower; /**< lower bound of free space */
591 indx_t pb_upper; /**< upper bound of free space */
593 uint32_t pb_pages; /**< number of overflow pages */
595 indx_t mp_ptrs[1]; /**< dynamic size */
598 /** Size of the page header, excluding dynamic data at the end */
599 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
601 /** Address of first usable data byte in a page, after the header */
602 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
604 /** Number of nodes on a page */
605 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
607 /** The amount of space remaining in the page */
608 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
610 /** The percentage of space used in the page, in tenths of a percent. */
611 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
612 ((env)->me_psize - PAGEHDRSZ))
613 /** The minimum page fill factor, in tenths of a percent.
614 * Pages emptier than this are candidates for merging.
616 #define FILL_THRESHOLD 250
618 /** Test if a page is a leaf page */
619 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
620 /** Test if a page is a LEAF2 page */
621 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
622 /** Test if a page is a branch page */
623 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
624 /** Test if a page is an overflow page */
625 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
626 /** Test if a page is a sub page */
627 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
629 /** The number of overflow pages needed to store the given size. */
630 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
632 /** Header for a single key/data pair within a page.
633 * We guarantee 2-byte alignment for nodes.
635 typedef struct MDB_node {
636 /** lo and hi are used for data size on leaf nodes and for
637 * child pgno on branch nodes. On 64 bit platforms, flags
638 * is also used for pgno. (Branch nodes have no flags).
639 * They are in host byte order in case that lets some
640 * accesses be optimized into a 32-bit word access.
642 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
643 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
644 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
645 /** @defgroup mdb_node Node Flags
647 * Flags for node headers.
650 #define F_BIGDATA 0x01 /**< data put on overflow page */
651 #define F_SUBDATA 0x02 /**< data is a sub-database */
652 #define F_DUPDATA 0x04 /**< data has duplicates */
654 /** valid flags for #mdb_node_add() */
655 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
658 unsigned short mn_flags; /**< @ref mdb_node */
659 unsigned short mn_ksize; /**< key size */
660 char mn_data[1]; /**< key and data are appended here */
663 /** Size of the node header, excluding dynamic data at the end */
664 #define NODESIZE offsetof(MDB_node, mn_data)
666 /** Bit position of top word in page number, for shifting mn_flags */
667 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
669 /** Size of a node in a branch page with a given key.
670 * This is just the node header plus the key, there is no data.
672 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
674 /** Size of a node in a leaf page with a given key and data.
675 * This is node header plus key plus data size.
677 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
679 /** Address of node \b i in page \b p */
680 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
682 /** Address of the key for the node */
683 #define NODEKEY(node) (void *)((node)->mn_data)
685 /** Address of the data for a node */
686 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
688 /** Get the page number pointed to by a branch node */
689 #define NODEPGNO(node) \
690 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
691 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
692 /** Set the page number in a branch node */
693 #define SETPGNO(node,pgno) do { \
694 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
695 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
697 /** Get the size of the data in a leaf node */
698 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
699 /** Set the size of the data for a leaf node */
700 #define SETDSZ(node,size) do { \
701 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
702 /** The size of a key in a node */
703 #define NODEKSZ(node) ((node)->mn_ksize)
705 /** Copy a page number from src to dst */
707 #define COPY_PGNO(dst,src) dst = src
709 #if SIZE_MAX > 4294967295UL
710 #define COPY_PGNO(dst,src) do { \
711 unsigned short *s, *d; \
712 s = (unsigned short *)&(src); \
713 d = (unsigned short *)&(dst); \
720 #define COPY_PGNO(dst,src) do { \
721 unsigned short *s, *d; \
722 s = (unsigned short *)&(src); \
723 d = (unsigned short *)&(dst); \
729 /** The address of a key in a LEAF2 page.
730 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
731 * There are no node headers, keys are stored contiguously.
733 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
735 /** Set the \b node's key into \b key, if requested. */
736 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
737 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
739 /** Information about a single database in the environment. */
740 typedef struct MDB_db {
741 uint32_t md_pad; /**< also ksize for LEAF2 pages */
742 uint16_t md_flags; /**< @ref mdb_dbi_open */
743 uint16_t md_depth; /**< depth of this tree */
744 pgno_t md_branch_pages; /**< number of internal pages */
745 pgno_t md_leaf_pages; /**< number of leaf pages */
746 pgno_t md_overflow_pages; /**< number of overflow pages */
747 size_t md_entries; /**< number of data items */
748 pgno_t md_root; /**< the root page of this tree */
751 /** mdb_dbi_open flags */
752 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
753 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
754 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
755 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
757 /** Handle for the DB used to track free pages. */
759 /** Handle for the default DB. */
762 /** Meta page content. */
763 typedef struct MDB_meta {
764 /** Stamp identifying this as an MDB file. It must be set
767 /** Version number of this lock file. Must be set to #MDB_VERSION. */
769 void *mm_address; /**< address for fixed mapping */
770 size_t mm_mapsize; /**< size of mmap region */
771 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
772 /** The size of pages used in this DB */
773 #define mm_psize mm_dbs[0].md_pad
774 /** Any persistent environment flags. @ref mdb_env */
775 #define mm_flags mm_dbs[0].md_flags
776 pgno_t mm_last_pg; /**< last used page in file */
777 txnid_t mm_txnid; /**< txnid that committed this page */
780 /** Buffer for a stack-allocated dirty page.
781 * The members define size and alignment, and silence type
782 * aliasing warnings. They are not used directly; that could
783 * mean incorrectly using several union members in parallel.
785 typedef union MDB_pagebuf {
786 char mb_raw[MDB_PAGESIZE];
789 char mm_pad[PAGEHDRSZ];
794 /** Auxiliary DB info.
795 * The information here is mostly static/read-only. There is
796 * only a single copy of this record in the environment.
798 typedef struct MDB_dbx {
799 MDB_val md_name; /**< name of the database */
800 MDB_cmp_func *md_cmp; /**< function for comparing keys */
801 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
802 MDB_rel_func *md_rel; /**< user relocate function */
803 void *md_relctx; /**< user-provided context for md_rel */
806 /** A database transaction.
807 * Every operation requires a transaction handle.
810 MDB_txn *mt_parent; /**< parent of a nested txn */
811 MDB_txn *mt_child; /**< nested txn under this txn */
812 pgno_t mt_next_pgno; /**< next unallocated page */
813 /** The ID of this transaction. IDs are integers incrementing from 1.
814 * Only committed write transactions increment the ID. If a transaction
815 * aborts, the ID may be re-used by the next writer.
818 MDB_env *mt_env; /**< the DB environment */
819 /** The list of pages that became unused during this transaction.
823 MDB_ID2L dirty_list; /**< for write txns: modified pages */
824 MDB_reader *reader; /**< this thread's reader table slot or NULL */
826 /** Array of records for each DB known in the environment. */
828 /** Array of MDB_db records for each known DB */
830 /** @defgroup mt_dbflag Transaction DB Flags
834 #define DB_DIRTY 0x01 /**< DB was written in this txn */
835 #define DB_STALE 0x02 /**< DB record is older than txnID */
836 #define DB_NEW 0x04 /**< DB handle opened in this txn */
837 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
839 /** In write txns, array of cursors for each DB */
840 MDB_cursor **mt_cursors;
841 /** Array of flags for each DB */
842 unsigned char *mt_dbflags;
843 /** Number of DB records in use. This number only ever increments;
844 * we don't decrement it when individual DB handles are closed.
848 /** @defgroup mdb_txn Transaction Flags
852 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
853 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
854 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
856 unsigned int mt_flags; /**< @ref mdb_txn */
857 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
858 unsigned int mt_dirty_room;
859 /** Tracks which of the two meta pages was used at the start
860 * of this transaction.
862 unsigned int mt_toggle;
865 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
866 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
867 * raise this on a 64 bit machine.
869 #define CURSOR_STACK 32
873 /** Cursors are used for all DB operations */
875 /** Next cursor on this DB in this txn */
877 /** Original cursor if this is a shadow */
879 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
880 struct MDB_xcursor *mc_xcursor;
881 /** The transaction that owns this cursor */
883 /** The database handle this cursor operates on */
885 /** The database record for this cursor */
887 /** The database auxiliary record for this cursor */
889 /** The @ref mt_dbflag for this database */
890 unsigned char *mc_dbflag;
891 unsigned short mc_snum; /**< number of pushed pages */
892 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
893 /** @defgroup mdb_cursor Cursor Flags
895 * Cursor state flags.
898 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
899 #define C_EOF 0x02 /**< No more data */
900 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
901 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
902 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
903 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
904 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
906 unsigned int mc_flags; /**< @ref mdb_cursor */
907 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
908 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
911 /** Context for sorted-dup records.
912 * We could have gone to a fully recursive design, with arbitrarily
913 * deep nesting of sub-databases. But for now we only handle these
914 * levels - main DB, optional sub-DB, sorted-duplicate DB.
916 typedef struct MDB_xcursor {
917 /** A sub-cursor for traversing the Dup DB */
918 MDB_cursor mx_cursor;
919 /** The database record for this Dup DB */
921 /** The auxiliary DB record for this Dup DB */
923 /** The @ref mt_dbflag for this Dup DB */
924 unsigned char mx_dbflag;
927 /** State of FreeDB old pages, stored in the MDB_env */
928 typedef struct MDB_pgstate {
929 txnid_t mf_pglast; /**< ID of last old page record we used */
930 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
931 pgno_t *mf_pgfree; /**< memory to free when dropping me_pghead */
934 /** The database environment. */
936 HANDLE me_fd; /**< The main data file */
937 HANDLE me_lfd; /**< The lock file */
938 HANDLE me_mfd; /**< just for writing the meta pages */
939 /** Failed to update the meta page. Probably an I/O error. */
940 #define MDB_FATAL_ERROR 0x80000000U
941 /** Some fields are initialized. */
942 #define MDB_ENV_ACTIVE 0x20000000U
943 /** me_txkey is set */
944 #define MDB_ENV_TXKEY 0x10000000U
945 uint32_t me_flags; /**< @ref mdb_env */
946 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
947 unsigned int me_maxreaders; /**< size of the reader table */
948 unsigned int me_numreaders; /**< max numreaders set by this env */
949 MDB_dbi me_numdbs; /**< number of DBs opened */
950 MDB_dbi me_maxdbs; /**< size of the DB table */
951 pid_t me_pid; /**< process ID of this env */
952 char *me_path; /**< path to the DB files */
953 char *me_map; /**< the memory map of the data file */
954 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
955 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
956 MDB_txn *me_txn; /**< current write transaction */
957 size_t me_mapsize; /**< size of the data memory map */
958 off_t me_size; /**< current file size */
959 pgno_t me_maxpg; /**< me_mapsize / me_psize */
960 MDB_dbx *me_dbxs; /**< array of static DB info */
961 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
962 pthread_key_t me_txkey; /**< thread-key for readers */
963 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
964 # define me_pglast me_pgstate.mf_pglast
965 # define me_pghead me_pgstate.mf_pghead
966 # define me_pgfree me_pgstate.mf_pgfree
967 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
968 /** IDL of pages that became unused in a write txn */
970 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
971 MDB_ID2L me_dirty_list;
972 /** Max number of freelist items that can fit in a single overflow page */
973 unsigned int me_maxfree_1pg;
974 /** Max size of a node on a page */
975 unsigned int me_nodemax;
977 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
979 #elif defined(MDB_USE_POSIX_SEM)
980 sem_t *me_rmutex; /* Shared mutexes are not supported */
985 /** Nested transaction */
986 typedef struct MDB_ntxn {
987 MDB_txn mnt_txn; /* the transaction */
988 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
991 /** max number of pages to commit in one writev() call */
992 #define MDB_COMMIT_PAGES 64
993 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
994 #undef MDB_COMMIT_PAGES
995 #define MDB_COMMIT_PAGES IOV_MAX
998 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
999 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1000 static int mdb_page_touch(MDB_cursor *mc);
1002 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1003 static int mdb_page_search_root(MDB_cursor *mc,
1004 MDB_val *key, int modify);
1005 #define MDB_PS_MODIFY 1
1006 #define MDB_PS_ROOTONLY 2
1007 static int mdb_page_search(MDB_cursor *mc,
1008 MDB_val *key, int flags);
1009 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1011 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1012 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1013 pgno_t newpgno, unsigned int nflags);
1015 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1016 static int mdb_env_pick_meta(const MDB_env *env);
1017 static int mdb_env_write_meta(MDB_txn *txn);
1018 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1019 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1021 static void mdb_env_close0(MDB_env *env, int excl);
1023 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1024 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1025 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1026 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1027 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1028 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1029 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1030 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1031 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1033 static int mdb_rebalance(MDB_cursor *mc);
1034 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1036 static void mdb_cursor_pop(MDB_cursor *mc);
1037 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1039 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1040 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1041 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1042 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1043 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1045 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1046 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1048 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1049 static void mdb_xcursor_init0(MDB_cursor *mc);
1050 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1052 static int mdb_drop0(MDB_cursor *mc, int subs);
1053 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1056 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1060 static SECURITY_DESCRIPTOR mdb_null_sd;
1061 static SECURITY_ATTRIBUTES mdb_all_sa;
1062 static int mdb_sec_inited;
1065 /** Return the library version info. */
1067 mdb_version(int *major, int *minor, int *patch)
1069 if (major) *major = MDB_VERSION_MAJOR;
1070 if (minor) *minor = MDB_VERSION_MINOR;
1071 if (patch) *patch = MDB_VERSION_PATCH;
1072 return MDB_VERSION_STRING;
1075 /** Table of descriptions for MDB @ref errors */
1076 static char *const mdb_errstr[] = {
1077 "MDB_KEYEXIST: Key/data pair already exists",
1078 "MDB_NOTFOUND: No matching key/data pair found",
1079 "MDB_PAGE_NOTFOUND: Requested page not found",
1080 "MDB_CORRUPTED: Located page was wrong type",
1081 "MDB_PANIC: Update of meta page failed",
1082 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1083 "MDB_INVALID: File is not an MDB file",
1084 "MDB_MAP_FULL: Environment mapsize limit reached",
1085 "MDB_DBS_FULL: Environment maxdbs limit reached",
1086 "MDB_READERS_FULL: Environment maxreaders limit reached",
1087 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1088 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1089 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1090 "MDB_PAGE_FULL: Internal error - page has no more space",
1091 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1092 "MDB_INCOMPATIBLE: Database flags changed or would change",
1093 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1097 mdb_strerror(int err)
1101 return ("Successful return: 0");
1103 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1104 i = err - MDB_KEYEXIST;
1105 return mdb_errstr[i];
1108 return strerror(err);
1112 /** Display a key in hexadecimal and return the address of the result.
1113 * @param[in] key the key to display
1114 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1115 * @return The key in hexadecimal form.
1118 mdb_dkey(MDB_val *key, char *buf)
1121 unsigned char *c = key->mv_data;
1127 if (key->mv_size > MDB_MAXKEYSIZE)
1128 return "MDB_MAXKEYSIZE";
1129 /* may want to make this a dynamic check: if the key is mostly
1130 * printable characters, print it as-is instead of converting to hex.
1134 for (i=0; i<key->mv_size; i++)
1135 ptr += sprintf(ptr, "%02x", *c++);
1137 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1142 /** Display all the keys in the page. */
1144 mdb_page_list(MDB_page *mp)
1147 unsigned int i, nkeys, nsize;
1151 nkeys = NUMKEYS(mp);
1152 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1153 for (i=0; i<nkeys; i++) {
1154 node = NODEPTR(mp, i);
1155 key.mv_size = node->mn_ksize;
1156 key.mv_data = node->mn_data;
1157 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1158 if (IS_BRANCH(mp)) {
1159 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1162 if (F_ISSET(node->mn_flags, F_BIGDATA))
1163 nsize += sizeof(pgno_t);
1165 nsize += NODEDSZ(node);
1166 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1172 mdb_cursor_chk(MDB_cursor *mc)
1178 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1179 for (i=0; i<mc->mc_top; i++) {
1181 node = NODEPTR(mp, mc->mc_ki[i]);
1182 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1185 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1191 /** Count all the pages in each DB and in the freelist
1192 * and make sure it matches the actual number of pages
1195 static void mdb_audit(MDB_txn *txn)
1199 MDB_ID freecount, count;
1204 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1205 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1206 freecount += *(MDB_ID *)data.mv_data;
1209 for (i = 0; i<txn->mt_numdbs; i++) {
1211 mdb_cursor_init(&mc, txn, i, &mx);
1212 if (txn->mt_dbs[i].md_root == P_INVALID)
1214 count += txn->mt_dbs[i].md_branch_pages +
1215 txn->mt_dbs[i].md_leaf_pages +
1216 txn->mt_dbs[i].md_overflow_pages;
1217 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1218 mdb_page_search(&mc, NULL, 0);
1222 mp = mc.mc_pg[mc.mc_top];
1223 for (j=0; j<NUMKEYS(mp); j++) {
1224 MDB_node *leaf = NODEPTR(mp, j);
1225 if (leaf->mn_flags & F_SUBDATA) {
1227 memcpy(&db, NODEDATA(leaf), sizeof(db));
1228 count += db.md_branch_pages + db.md_leaf_pages +
1229 db.md_overflow_pages;
1233 while (mdb_cursor_sibling(&mc, 1) == 0);
1236 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1237 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1238 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1244 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1246 return txn->mt_dbxs[dbi].md_cmp(a, b);
1250 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1252 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1255 /** Allocate a page.
1256 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1259 mdb_page_malloc(MDB_cursor *mc, unsigned num)
1261 MDB_env *env = mc->mc_txn->mt_env;
1262 MDB_page *ret = env->me_dpages;
1263 size_t sz = env->me_psize;
1266 VGMEMP_ALLOC(env, ret, sz);
1267 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1268 env->me_dpages = ret->mp_next;
1274 if ((ret = malloc(sz)) != NULL) {
1275 VGMEMP_ALLOC(env, ret, sz);
1280 /** Free a single page.
1281 * Saves single pages to a list, for future reuse.
1282 * (This is not used for multi-page overflow pages.)
1285 mdb_page_free(MDB_env *env, MDB_page *mp)
1287 mp->mp_next = env->me_dpages;
1288 VGMEMP_FREE(env, mp);
1289 env->me_dpages = mp;
1292 /* Return all dirty pages to dpage list */
1294 mdb_dlist_free(MDB_txn *txn)
1296 MDB_env *env = txn->mt_env;
1297 MDB_ID2L dl = txn->mt_u.dirty_list;
1298 unsigned i, n = dl[0].mid;
1300 for (i = 1; i <= n; i++) {
1301 MDB_page *dp = dl[i].mptr;
1302 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1303 mdb_page_free(env, dp);
1305 /* large pages just get freed directly */
1306 VGMEMP_FREE(env, dp);
1313 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1315 mdb_find_oldest(MDB_txn *txn)
1318 txnid_t mr, oldest = txn->mt_txnid - 1;
1319 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1320 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1330 /** Allocate pages for writing.
1331 * If there are free pages available from older transactions, they
1332 * will be re-used first. Otherwise a new page will be allocated.
1333 * @param[in] mc cursor A cursor handle identifying the transaction and
1334 * database for which we are allocating.
1335 * @param[in] num the number of pages to allocate.
1336 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1337 * will always be satisfied by a single contiguous chunk of memory.
1338 * @return 0 on success, non-zero on failure.
1341 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1343 MDB_txn *txn = mc->mc_txn;
1345 pgno_t pgno = P_INVALID;
1347 txnid_t oldest = 0, last;
1352 /* If our dirty list is already full, we can't do anything */
1353 if (txn->mt_dirty_room == 0)
1354 return MDB_TXN_FULL;
1356 /* The free list won't have any content at all until txn 2 has
1357 * committed. The pages freed by txn 2 will be unreferenced
1358 * after txn 3 commits, and so will be safe to re-use in txn 4.
1360 if (txn->mt_txnid > 3) {
1361 if (!txn->mt_env->me_pghead &&
1362 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1363 /* See if there's anything in the free DB */
1369 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1370 if (!txn->mt_env->me_pglast) {
1371 mdb_page_search(&m2, NULL, 0);
1372 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1373 kptr = (txnid_t *)NODEKEY(leaf);
1378 last = txn->mt_env->me_pglast + 1;
1380 key.mv_data = &last;
1381 key.mv_size = sizeof(last);
1382 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1385 last = *(txnid_t *)key.mv_data;
1389 oldest = mdb_find_oldest(txn);
1391 if (oldest > last) {
1392 /* It's usable, grab it.
1396 if (!txn->mt_env->me_pglast) {
1397 mdb_node_read(txn, leaf, &data);
1399 idl = (MDB_ID *) data.mv_data;
1400 /* We might have a zero-length IDL due to freelist growth
1401 * during a prior commit
1404 txn->mt_env->me_pglast = last;
1407 mop = mdb_midl_alloc(idl[0]);
1410 txn->mt_env->me_pglast = last;
1411 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1412 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1417 DPRINTF("IDL read txn %zu root %zu num %zu",
1418 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1419 for (i=0; i<idl[0]; i++) {
1420 DPRINTF("IDL %zu", idl[i+1]);
1427 if (txn->mt_env->me_pghead) {
1428 pgno_t *mop = txn->mt_env->me_pghead;
1431 int retry = 1, readit = 0, n2 = num-1;
1432 unsigned int i, j, k;
1434 /* If current list is too short, must fetch more and coalesce */
1435 if (mop[0] < (unsigned)num)
1438 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1440 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1441 /* If on freelist, don't try to read more. If what we have
1442 * right now isn't enough just use new pages.
1443 * TODO: get all of this working. Many circular dependencies...
1445 if (mc->mc_dbi == FREE_DBI) {
1454 last = txn->mt_env->me_pglast + 1;
1456 /* We haven't hit the readers list yet? */
1458 oldest = mdb_find_oldest(txn);
1461 /* There's nothing we can use on the freelist */
1462 if (oldest - last < 1)
1465 key.mv_data = &last;
1466 key.mv_size = sizeof(last);
1467 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1469 if (rc == MDB_NOTFOUND)
1473 last = *(txnid_t*)key.mv_data;
1476 idl = (MDB_ID *) data.mv_data;
1477 mop2 = mdb_midl_alloc(idl[0] + mop[0]);
1480 /* merge in sorted order */
1481 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1483 while (i>0 || j>0) {
1484 if (i && idl[i] < mop[j])
1485 mop2[k--] = idl[i--];
1487 mop2[k--] = mop[j--];
1489 txn->mt_env->me_pglast = last;
1490 mdb_midl_free(txn->mt_env->me_pgfree);
1491 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1493 /* Keep trying to read until we have enough */
1494 if (mop[0] < (unsigned)num) {
1499 /* current list has enough pages, but are they contiguous? */
1500 for (i=mop[0]; i>=(unsigned)num; i--) {
1501 if (mop[i-n2] == mop[i] + n2) {
1504 /* move any stragglers down */
1505 for (j=i+num; j<=mop[0]; j++)
1512 /* Stop if we succeeded, or no retries */
1513 if (!retry || pgno != P_INVALID)
1519 /* peel pages off tail, so we only have to truncate the list */
1520 pgno = MDB_IDL_LAST(mop);
1523 if (MDB_IDL_IS_ZERO(mop)) {
1524 mdb_midl_free(txn->mt_env->me_pgfree);
1525 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1530 if (pgno == P_INVALID) {
1531 /* DB size is maxed out */
1532 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1533 DPUTS("DB size maxed out");
1534 return MDB_MAP_FULL;
1537 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1538 if (pgno == P_INVALID) {
1539 pgno = txn->mt_next_pgno;
1540 txn->mt_next_pgno += num;
1542 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1545 if (!(np = mdb_page_malloc(mc, num)))
1547 if (pgno == P_INVALID) {
1548 np->mp_pgno = txn->mt_next_pgno;
1549 txn->mt_next_pgno += num;
1554 mid.mid = np->mp_pgno;
1556 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1557 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1559 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1561 txn->mt_dirty_room--;
1567 /** Copy a page: avoid copying unused portions of the page.
1568 * @param[in] dst page to copy into
1569 * @param[in] src page to copy from
1570 * @param[in] psize size of a page
1573 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1575 dst->mp_flags = src->mp_flags | P_DIRTY;
1576 dst->mp_pages = src->mp_pages;
1578 if (IS_LEAF2(src)) {
1579 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1581 unsigned int i, nkeys = NUMKEYS(src);
1582 for (i=0; i<nkeys; i++)
1583 dst->mp_ptrs[i] = src->mp_ptrs[i];
1584 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1585 psize - src->mp_upper);
1589 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1590 * @param[in] mc cursor pointing to the page to be touched
1591 * @return 0 on success, non-zero on failure.
1594 mdb_page_touch(MDB_cursor *mc)
1596 MDB_page *mp = mc->mc_pg[mc->mc_top];
1600 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1602 if ((rc = mdb_page_alloc(mc, 1, &np)))
1604 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1605 assert(mp->mp_pgno != np->mp_pgno);
1606 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1608 /* If page isn't full, just copy the used portion */
1609 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1612 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1614 np->mp_flags |= P_DIRTY;
1619 /* Adjust other cursors pointing to mp */
1620 if (mc->mc_flags & C_SUB) {
1621 MDB_cursor *m2, *m3;
1622 MDB_dbi dbi = mc->mc_dbi-1;
1624 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1625 if (m2 == mc) continue;
1626 m3 = &m2->mc_xcursor->mx_cursor;
1627 if (m3->mc_snum < mc->mc_snum) continue;
1628 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1629 m3->mc_pg[mc->mc_top] = mp;
1635 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1636 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1637 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1638 m2->mc_pg[mc->mc_top] = mp;
1639 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1640 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
1641 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
1642 if (!(leaf->mn_flags & F_SUBDATA)) {
1643 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1649 mc->mc_pg[mc->mc_top] = mp;
1650 /** If this page has a parent, update the parent to point to
1654 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1656 mc->mc_db->md_root = mp->mp_pgno;
1657 } else if (mc->mc_txn->mt_parent && !(mp->mp_flags & P_SUBP)) {
1659 MDB_ID2 mid, *dl = mc->mc_txn->mt_u.dirty_list;
1660 /* If txn has a parent, make sure the page is in our
1664 unsigned x = mdb_mid2l_search(dl, mp->mp_pgno);
1665 if (x <= dl[0].mid && dl[x].mid == mp->mp_pgno) {
1668 mc->mc_pg[mc->mc_top] = np;
1672 assert(dl[0].mid < MDB_IDL_UM_MAX);
1674 np = mdb_page_malloc(mc, 1);
1677 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1678 mid.mid = np->mp_pgno;
1680 mdb_mid2l_insert(dl, &mid);
1688 mdb_env_sync(MDB_env *env, int force)
1691 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1692 if (env->me_flags & MDB_WRITEMAP) {
1693 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1694 ? MS_ASYNC : MS_SYNC;
1695 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1698 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1702 if (MDB_FDATASYNC(env->me_fd))
1709 /** Make shadow copies of all of parent txn's cursors */
1711 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1713 MDB_cursor *mc, *m2;
1714 unsigned int i, j, size;
1716 for (i=0;i<src->mt_numdbs; i++) {
1717 if (src->mt_cursors[i]) {
1718 size = sizeof(MDB_cursor);
1719 if (src->mt_cursors[i]->mc_xcursor)
1720 size += sizeof(MDB_xcursor);
1721 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1728 mc->mc_db = &dst->mt_dbs[i];
1729 mc->mc_dbx = m2->mc_dbx;
1730 mc->mc_dbflag = &dst->mt_dbflags[i];
1731 mc->mc_snum = m2->mc_snum;
1732 mc->mc_top = m2->mc_top;
1733 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1734 for (j=0; j<mc->mc_snum; j++) {
1735 mc->mc_pg[j] = m2->mc_pg[j];
1736 mc->mc_ki[j] = m2->mc_ki[j];
1738 if (m2->mc_xcursor) {
1739 MDB_xcursor *mx, *mx2;
1740 mx = (MDB_xcursor *)(mc+1);
1741 mc->mc_xcursor = mx;
1742 mx2 = m2->mc_xcursor;
1743 mx->mx_db = mx2->mx_db;
1744 mx->mx_dbx = mx2->mx_dbx;
1745 mx->mx_dbflag = mx2->mx_dbflag;
1746 mx->mx_cursor.mc_txn = dst;
1747 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1748 mx->mx_cursor.mc_db = &mx->mx_db;
1749 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1750 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1751 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1752 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1753 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1754 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1755 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1756 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1759 mc->mc_xcursor = NULL;
1761 mc->mc_next = dst->mt_cursors[i];
1762 dst->mt_cursors[i] = mc;
1769 /** Close this write txn's cursors, after optionally merging its shadow
1770 * cursors back into parent's.
1771 * @param[in] txn the transaction handle.
1772 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1773 * @return 0 on success, non-zero on failure.
1776 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1778 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1781 for (i = txn->mt_numdbs; --i >= 0; ) {
1782 for (mc = cursors[i]; mc; mc = next) {
1784 if (mc->mc_flags & merge) {
1785 MDB_cursor *m2 = mc->mc_orig;
1786 m2->mc_snum = mc->mc_snum;
1787 m2->mc_top = mc->mc_top;
1788 for (j = mc->mc_snum; --j >= 0; ) {
1789 m2->mc_pg[j] = mc->mc_pg[j];
1790 m2->mc_ki[j] = mc->mc_ki[j];
1793 if (mc->mc_flags & C_ALLOCD)
1801 mdb_txn_reset0(MDB_txn *txn);
1803 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1804 * @param[in] txn the transaction handle to initialize
1805 * @return 0 on success, non-zero on failure.
1808 mdb_txn_renew0(MDB_txn *txn)
1810 MDB_env *env = txn->mt_env;
1813 int rc, new_notls = 0;
1816 txn->mt_numdbs = env->me_numdbs;
1817 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1819 if (txn->mt_flags & MDB_TXN_RDONLY) {
1820 if (!env->me_txns) {
1821 i = mdb_env_pick_meta(env);
1822 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1823 txn->mt_u.reader = NULL;
1825 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1826 pthread_getspecific(env->me_txkey);
1828 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1829 return MDB_BAD_RSLOT;
1831 pid_t pid = env->me_pid;
1832 pthread_t tid = pthread_self();
1835 for (i=0; i<env->me_txns->mti_numreaders; i++)
1836 if (env->me_txns->mti_readers[i].mr_pid == 0)
1838 if (i == env->me_maxreaders) {
1839 UNLOCK_MUTEX_R(env);
1840 return MDB_READERS_FULL;
1842 env->me_txns->mti_readers[i].mr_pid = pid;
1843 env->me_txns->mti_readers[i].mr_tid = tid;
1844 if (i >= env->me_txns->mti_numreaders)
1845 env->me_txns->mti_numreaders = i+1;
1846 /* Save numreaders for un-mutexed mdb_env_close() */
1847 env->me_numreaders = env->me_txns->mti_numreaders;
1848 UNLOCK_MUTEX_R(env);
1849 r = &env->me_txns->mti_readers[i];
1850 new_notls = (env->me_flags & MDB_NOTLS);
1851 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1856 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1857 txn->mt_u.reader = r;
1859 txn->mt_toggle = txn->mt_txnid & 1;
1860 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1864 txn->mt_txnid = env->me_txns->mti_txnid;
1865 txn->mt_toggle = txn->mt_txnid & 1;
1866 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1869 if (txn->mt_txnid == mdb_debug_start)
1872 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1873 txn->mt_u.dirty_list = env->me_dirty_list;
1874 txn->mt_u.dirty_list[0].mid = 0;
1875 txn->mt_free_pgs = env->me_free_pgs;
1876 txn->mt_free_pgs[0] = 0;
1880 /* Copy the DB info and flags */
1881 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1882 for (i=2; i<txn->mt_numdbs; i++) {
1883 x = env->me_dbflags[i];
1884 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1885 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1887 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1889 if (env->me_maxpg < txn->mt_next_pgno) {
1890 mdb_txn_reset0(txn);
1892 txn->mt_u.reader->mr_pid = 0;
1893 txn->mt_u.reader = NULL;
1895 return MDB_MAP_RESIZED;
1902 mdb_txn_renew(MDB_txn *txn)
1906 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1909 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1910 DPUTS("environment had fatal error, must shutdown!");
1914 rc = mdb_txn_renew0(txn);
1915 if (rc == MDB_SUCCESS) {
1916 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1917 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1918 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1924 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1928 int rc, size, tsize = sizeof(MDB_txn);
1930 if (env->me_flags & MDB_FATAL_ERROR) {
1931 DPUTS("environment had fatal error, must shutdown!");
1934 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1937 /* Nested transactions: Max 1 child, write txns only, no writemap */
1938 if (parent->mt_child ||
1939 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1940 (env->me_flags & MDB_WRITEMAP))
1944 tsize = sizeof(MDB_ntxn);
1946 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1947 if (!(flags & MDB_RDONLY))
1948 size += env->me_maxdbs * sizeof(MDB_cursor *);
1950 if ((txn = calloc(1, size)) == NULL) {
1951 DPRINTF("calloc: %s", strerror(ErrCode()));
1954 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1955 if (flags & MDB_RDONLY) {
1956 txn->mt_flags |= MDB_TXN_RDONLY;
1957 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1959 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1960 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1966 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1967 if (!txn->mt_u.dirty_list ||
1968 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1970 free(txn->mt_u.dirty_list);
1974 txn->mt_txnid = parent->mt_txnid;
1975 txn->mt_toggle = parent->mt_toggle;
1976 txn->mt_dirty_room = parent->mt_dirty_room;
1977 txn->mt_u.dirty_list[0].mid = 0;
1978 txn->mt_free_pgs[0] = 0;
1979 txn->mt_next_pgno = parent->mt_next_pgno;
1980 parent->mt_child = txn;
1981 txn->mt_parent = parent;
1982 txn->mt_numdbs = parent->mt_numdbs;
1983 txn->mt_dbxs = parent->mt_dbxs;
1984 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1985 /* Copy parent's mt_dbflags, but clear DB_NEW */
1986 for (i=0; i<txn->mt_numdbs; i++)
1987 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1989 ntxn = (MDB_ntxn *)txn;
1990 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1991 if (env->me_pghead) {
1992 size = MDB_IDL_SIZEOF(env->me_pghead);
1993 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1995 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1999 env->me_pgfree = env->me_pghead;
2001 rc = mdb_cursor_shadow(parent, txn);
2003 mdb_txn_reset0(txn);
2005 rc = mdb_txn_renew0(txn);
2011 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
2012 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2013 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2019 /** Export or close DBI handles opened in this txn. */
2021 mdb_dbis_update(MDB_txn *txn, int keep)
2024 MDB_dbi n = txn->mt_numdbs;
2025 MDB_env *env = txn->mt_env;
2026 unsigned char *tdbflags = txn->mt_dbflags;
2028 for (i = n; --i >= 2;) {
2029 if (tdbflags[i] & DB_NEW) {
2031 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2033 char *ptr = env->me_dbxs[i].md_name.mv_data;
2034 env->me_dbxs[i].md_name.mv_data = NULL;
2035 env->me_dbxs[i].md_name.mv_size = 0;
2036 env->me_dbflags[i] = 0;
2041 if (keep && env->me_numdbs < n)
2045 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2046 * May be called twice for readonly txns: First reset it, then abort.
2047 * @param[in] txn the transaction handle to reset
2050 mdb_txn_reset0(MDB_txn *txn)
2052 MDB_env *env = txn->mt_env;
2054 /* Close any DBI handles opened in this txn */
2055 mdb_dbis_update(txn, 0);
2057 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2058 if (txn->mt_u.reader) {
2059 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2060 if (!(env->me_flags & MDB_NOTLS))
2061 txn->mt_u.reader = NULL; /* txn does not own reader */
2063 txn->mt_numdbs = 0; /* close nothing if called again */
2064 txn->mt_dbxs = NULL; /* mark txn as reset */
2066 mdb_cursors_close(txn, 0);
2068 if (!(env->me_flags & MDB_WRITEMAP)) {
2069 mdb_dlist_free(txn);
2071 mdb_midl_free(env->me_pgfree);
2073 if (txn->mt_parent) {
2074 txn->mt_parent->mt_child = NULL;
2075 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2076 mdb_midl_free(txn->mt_free_pgs);
2077 free(txn->mt_u.dirty_list);
2080 if (mdb_midl_shrink(&txn->mt_free_pgs))
2081 env->me_free_pgs = txn->mt_free_pgs;
2084 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
2085 txn->mt_env->me_pglast = 0;
2088 /* The writer mutex was locked in mdb_txn_begin. */
2089 UNLOCK_MUTEX_W(env);
2094 mdb_txn_reset(MDB_txn *txn)
2099 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2100 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2101 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2103 /* This call is only valid for read-only txns */
2104 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2107 mdb_txn_reset0(txn);
2111 mdb_txn_abort(MDB_txn *txn)
2116 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2117 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2118 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2121 mdb_txn_abort(txn->mt_child);
2123 mdb_txn_reset0(txn);
2124 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2125 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2126 txn->mt_u.reader->mr_pid = 0;
2132 mdb_txn_commit(MDB_txn *txn)
2140 pgno_t next, freecnt;
2141 txnid_t oldpg_txnid, id;
2144 assert(txn != NULL);
2145 assert(txn->mt_env != NULL);
2147 if (txn->mt_child) {
2148 rc = mdb_txn_commit(txn->mt_child);
2149 txn->mt_child = NULL;
2158 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2159 mdb_dbis_update(txn, 1);
2160 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2165 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2166 DPUTS("error flag is set, can't commit");
2168 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2173 if (txn->mt_parent) {
2174 MDB_txn *parent = txn->mt_parent;
2178 /* Append our free list to parent's */
2179 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2183 mdb_midl_free(txn->mt_free_pgs);
2185 parent->mt_next_pgno = txn->mt_next_pgno;
2186 parent->mt_flags = txn->mt_flags;
2188 /* Merge our cursors into parent's and close them */
2189 mdb_cursors_close(txn, C_SHADOW);
2191 /* Update parent's DB table. */
2192 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2193 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2194 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2195 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2196 for (i=2; i<txn->mt_numdbs; i++) {
2197 /* preserve parent's DB_NEW status */
2198 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2199 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2202 dst = txn->mt_parent->mt_u.dirty_list;
2203 src = txn->mt_u.dirty_list;
2204 /* Find len = length of merging our dirty list with parent's */
2206 dst[0].mid = 0; /* simplify loops */
2207 if (parent->mt_parent) {
2208 len = x + src[0].mid;
2209 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2210 for (i = x; y && i; y--) {
2211 pgno_t yp = src[y].mid;
2212 while (yp < dst[i].mid)
2214 if (yp == dst[i].mid) {
2219 } else { /* Simplify the above for single-ancestor case */
2220 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2222 /* Merge our dirty list with parent's */
2224 for (i = len; y; dst[i--] = src[y--]) {
2225 pgno_t yp = src[y].mid;
2226 while (yp < dst[x].mid)
2227 dst[i--] = dst[x--];
2228 if (yp == dst[x].mid)
2229 free(dst[x--].mptr);
2233 free(txn->mt_u.dirty_list);
2234 parent->mt_dirty_room = txn->mt_dirty_room;
2236 txn->mt_parent->mt_child = NULL;
2237 free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pgfree);
2242 if (txn != env->me_txn) {
2243 DPUTS("attempt to commit unknown transaction");
2248 mdb_cursors_close(txn, 0);
2250 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2253 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2254 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2256 /* Update DB root pointers */
2257 if (txn->mt_numdbs > 2) {
2260 data.mv_size = sizeof(MDB_db);
2262 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2263 for (i = 2; i < txn->mt_numdbs; i++) {
2264 if (txn->mt_dbflags[i] & DB_DIRTY) {
2265 data.mv_data = &txn->mt_dbs[i];
2266 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2273 /* Save the freelist as of this transaction to the freeDB. This
2274 * can change the freelist, so keep trying until it stabilizes.
2276 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2277 * except the code below can decrease env->me_pglast to split pghead.
2278 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2279 * can only appear in env->me_pghead when env->me_pglast increases.
2280 * Until then, the me_pghead pointer won't move but can become NULL.
2283 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2284 oldpg_txnid = id = 0;
2287 /* should only be one record now */
2288 if (env->me_pghead || env->me_pglast) {
2289 /* make sure first page of freeDB is touched and on freelist */
2290 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2291 if (rc && rc != MDB_NOTFOUND) {
2298 /* Delete IDLs we used from the free list */
2299 if (env->me_pglast) {
2304 rc = mdb_cursor_first(&mc, &key, NULL);
2307 oldpg_txnid = *(txnid_t *)key.mv_data;
2309 assert(oldpg_txnid <= env->me_pglast);
2311 rc = mdb_cursor_del(&mc, 0);
2314 } while (oldpg_txnid < env->me_pglast);
2317 /* Save IDL of pages freed by this txn, to freeDB */
2319 if (freecnt != txn->mt_free_pgs[0]) {
2322 /* make sure last page of freeDB is touched and on freelist */
2323 key.mv_size = MDB_MAXKEYSIZE+1;
2325 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2326 if (rc && rc != MDB_NOTFOUND)
2332 MDB_IDL idl = txn->mt_free_pgs;
2333 mdb_midl_sort(txn->mt_free_pgs);
2334 DPRINTF("IDL write txn %zu root %zu num %zu",
2335 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2336 for (i=1; i<=idl[0]; i++) {
2337 DPRINTF("IDL %zu", idl[i]);
2341 /* write to last page of freeDB */
2342 key.mv_size = sizeof(pgno_t);
2343 key.mv_data = &txn->mt_txnid;
2344 /* The free list can still grow during this call,
2345 * despite the pre-emptive touches above. So retry
2346 * until the reserved space remains big enough.
2349 assert(freecnt < txn->mt_free_pgs[0]);
2350 freecnt = txn->mt_free_pgs[0];
2351 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2352 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2355 } while (freecnt != txn->mt_free_pgs[0]);
2356 mdb_midl_sort(txn->mt_free_pgs);
2357 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2358 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2359 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2362 /* Put back page numbers we took from freeDB but did not use */
2363 if (env->me_pghead) {
2368 mop = env->me_pghead;
2369 id = env->me_pglast;
2370 key.mv_size = sizeof(id);
2372 /* These steps may grow the freelist again
2373 * due to freed overflow pages...
2378 if (orig > env->me_maxfree_1pg && id > 4)
2379 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2380 data.mv_size = (orig + 1) * sizeof(pgno_t);
2381 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2384 assert(!env->me_pghead || env->me_pglast);
2385 /* mop could have been used again here */
2386 if (id != env->me_pglast || env->me_pghead == NULL)
2387 goto again; /* was completely used up */
2388 assert(mop == env->me_pghead);
2389 } while (mop[0] < orig && --i);
2390 memcpy(data.mv_data, mop, data.mv_size);
2393 *(pgno_t *)data.mv_data = orig;
2394 mop[orig] = mop[0] - orig;
2395 env->me_pghead = mop += orig;
2396 /* Save more oldpages at the previous txnid. */
2397 assert(env->me_pglast == id && id == oldpg_txnid);
2398 env->me_pglast = --oldpg_txnid;
2402 /* Check for growth of freelist again */
2403 if (freecnt != txn->mt_free_pgs[0])
2406 mdb_midl_free(env->me_pgfree);
2407 env->me_pghead = env->me_pgfree = NULL;
2409 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2410 if (mdb_midl_shrink(&txn->mt_free_pgs))
2411 env->me_free_pgs = txn->mt_free_pgs;
2418 if (env->me_flags & MDB_WRITEMAP) {
2419 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2420 dp = txn->mt_u.dirty_list[i].mptr;
2421 /* clear dirty flag */
2422 dp->mp_flags &= ~P_DIRTY;
2424 txn->mt_u.dirty_list[0].mid = 0;
2428 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2434 /* Windows actually supports scatter/gather I/O, but only on
2435 * unbuffered file handles. Since we're relying on the OS page
2436 * cache for all our data, that's self-defeating. So we just
2437 * write pages one at a time. We use the ov structure to set
2438 * the write offset, to at least save the overhead of a Seek
2442 memset(&ov, 0, sizeof(ov));
2443 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2445 dp = txn->mt_u.dirty_list[i].mptr;
2446 DPRINTF("committing page %zu", dp->mp_pgno);
2447 size = dp->mp_pgno * env->me_psize;
2448 ov.Offset = size & 0xffffffff;
2449 ov.OffsetHigh = size >> 16;
2450 ov.OffsetHigh >>= 16;
2451 /* clear dirty flag */
2452 dp->mp_flags &= ~P_DIRTY;
2453 wsize = env->me_psize;
2454 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2455 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2458 DPRINTF("WriteFile: %d", n);
2465 struct iovec iov[MDB_COMMIT_PAGES];
2469 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2470 dp = txn->mt_u.dirty_list[i].mptr;
2471 if (dp->mp_pgno != next) {
2473 rc = writev(env->me_fd, iov, n);
2477 DPUTS("short write, filesystem full?");
2479 DPRINTF("writev: %s", strerror(n));
2486 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2489 DPRINTF("committing page %zu", dp->mp_pgno);
2490 iov[n].iov_len = env->me_psize;
2491 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2492 iov[n].iov_base = (char *)dp;
2493 size += iov[n].iov_len;
2494 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2495 /* clear dirty flag */
2496 dp->mp_flags &= ~P_DIRTY;
2497 if (++n >= MDB_COMMIT_PAGES) {
2507 rc = writev(env->me_fd, iov, n);
2511 DPUTS("short write, filesystem full?");
2513 DPRINTF("writev: %s", strerror(n));
2520 mdb_dlist_free(txn);
2523 if ((n = mdb_env_sync(env, 0)) != 0 ||
2524 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2532 mdb_dbis_update(txn, 1);
2534 UNLOCK_MUTEX_W(env);
2540 /** Read the environment parameters of a DB environment before
2541 * mapping it into memory.
2542 * @param[in] env the environment handle
2543 * @param[out] meta address of where to store the meta information
2544 * @return 0 on success, non-zero on failure.
2547 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2554 /* We don't know the page size yet, so use a minimum value.
2555 * Read both meta pages so we can use the latest one.
2558 for (i=0; i<2; i++) {
2560 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2562 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2567 else if (rc != MDB_PAGESIZE) {
2571 DPRINTF("read: %s", strerror(err));
2575 p = (MDB_page *)&pbuf;
2577 if (!F_ISSET(p->mp_flags, P_META)) {
2578 DPRINTF("page %zu not a meta page", p->mp_pgno);
2583 if (m->mm_magic != MDB_MAGIC) {
2584 DPUTS("meta has invalid magic");
2588 if (m->mm_version != MDB_VERSION) {
2589 DPRINTF("database is version %u, expected version %u",
2590 m->mm_version, MDB_VERSION);
2591 return MDB_VERSION_MISMATCH;
2595 if (m->mm_txnid > meta->mm_txnid)
2596 memcpy(meta, m, sizeof(*m));
2598 memcpy(meta, m, sizeof(*m));
2600 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2602 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2610 /** Write the environment parameters of a freshly created DB environment.
2611 * @param[in] env the environment handle
2612 * @param[out] meta address of where to store the meta information
2613 * @return 0 on success, non-zero on failure.
2616 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2623 DPUTS("writing new meta page");
2625 GET_PAGESIZE(psize);
2627 meta->mm_magic = MDB_MAGIC;
2628 meta->mm_version = MDB_VERSION;
2629 meta->mm_mapsize = env->me_mapsize;
2630 meta->mm_psize = psize;
2631 meta->mm_last_pg = 1;
2632 meta->mm_flags = env->me_flags & 0xffff;
2633 meta->mm_flags |= MDB_INTEGERKEY;
2634 meta->mm_dbs[0].md_root = P_INVALID;
2635 meta->mm_dbs[1].md_root = P_INVALID;
2637 p = calloc(2, psize);
2639 p->mp_flags = P_META;
2642 memcpy(m, meta, sizeof(*meta));
2644 q = (MDB_page *)((char *)p + psize);
2647 q->mp_flags = P_META;
2650 memcpy(m, meta, sizeof(*meta));
2655 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2656 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2657 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2660 lseek(env->me_fd, 0, SEEK_SET);
2661 rc = write(env->me_fd, p, psize * 2);
2662 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2668 /** Update the environment info to commit a transaction.
2669 * @param[in] txn the transaction that's being committed
2670 * @return 0 on success, non-zero on failure.
2673 mdb_env_write_meta(MDB_txn *txn)
2676 MDB_meta meta, metab, *mp;
2678 int rc, len, toggle;
2685 assert(txn != NULL);
2686 assert(txn->mt_env != NULL);
2688 toggle = !txn->mt_toggle;
2689 DPRINTF("writing meta page %d for root page %zu",
2690 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2693 mp = env->me_metas[toggle];
2695 if (env->me_flags & MDB_WRITEMAP) {
2696 /* Persist any increases of mapsize config */
2697 if (env->me_mapsize > mp->mm_mapsize)
2698 mp->mm_mapsize = env->me_mapsize;
2699 mp->mm_dbs[0] = txn->mt_dbs[0];
2700 mp->mm_dbs[1] = txn->mt_dbs[1];
2701 mp->mm_last_pg = txn->mt_next_pgno - 1;
2702 mp->mm_txnid = txn->mt_txnid;
2703 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2704 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2707 ptr += env->me_psize;
2708 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2715 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2716 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2718 ptr = (char *)&meta;
2719 if (env->me_mapsize > mp->mm_mapsize) {
2720 /* Persist any increases of mapsize config */
2721 meta.mm_mapsize = env->me_mapsize;
2722 off = offsetof(MDB_meta, mm_mapsize);
2724 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2726 len = sizeof(MDB_meta) - off;
2729 meta.mm_dbs[0] = txn->mt_dbs[0];
2730 meta.mm_dbs[1] = txn->mt_dbs[1];
2731 meta.mm_last_pg = txn->mt_next_pgno - 1;
2732 meta.mm_txnid = txn->mt_txnid;
2735 off += env->me_psize;
2738 /* Write to the SYNC fd */
2739 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2740 env->me_fd : env->me_mfd;
2743 memset(&ov, 0, sizeof(ov));
2745 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2748 rc = pwrite(mfd, ptr, len, off);
2753 DPUTS("write failed, disk error?");
2754 /* On a failure, the pagecache still contains the new data.
2755 * Write some old data back, to prevent it from being used.
2756 * Use the non-SYNC fd; we know it will fail anyway.
2758 meta.mm_last_pg = metab.mm_last_pg;
2759 meta.mm_txnid = metab.mm_txnid;
2761 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2763 r2 = pwrite(env->me_fd, ptr, len, off);
2766 env->me_flags |= MDB_FATAL_ERROR;
2770 /* Memory ordering issues are irrelevant; since the entire writer
2771 * is wrapped by wmutex, all of these changes will become visible
2772 * after the wmutex is unlocked. Since the DB is multi-version,
2773 * readers will get consistent data regardless of how fresh or
2774 * how stale their view of these values is.
2776 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2781 /** Check both meta pages to see which one is newer.
2782 * @param[in] env the environment handle
2783 * @return meta toggle (0 or 1).
2786 mdb_env_pick_meta(const MDB_env *env)
2788 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2792 mdb_env_create(MDB_env **env)
2796 e = calloc(1, sizeof(MDB_env));
2800 e->me_maxreaders = DEFAULT_READERS;
2801 e->me_maxdbs = e->me_numdbs = 2;
2802 e->me_fd = INVALID_HANDLE_VALUE;
2803 e->me_lfd = INVALID_HANDLE_VALUE;
2804 e->me_mfd = INVALID_HANDLE_VALUE;
2805 #ifdef MDB_USE_POSIX_SEM
2806 e->me_rmutex = SEM_FAILED;
2807 e->me_wmutex = SEM_FAILED;
2809 e->me_pid = getpid();
2810 VGMEMP_CREATE(e,0,0);
2816 mdb_env_set_mapsize(MDB_env *env, size_t size)
2820 env->me_mapsize = size;
2822 env->me_maxpg = env->me_mapsize / env->me_psize;
2827 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2831 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2836 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2838 if (env->me_map || readers < 1)
2840 env->me_maxreaders = readers;
2845 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2847 if (!env || !readers)
2849 *readers = env->me_maxreaders;
2853 /** Further setup required for opening an MDB environment
2856 mdb_env_open2(MDB_env *env)
2858 unsigned int flags = env->me_flags;
2859 int i, newenv = 0, prot;
2863 memset(&meta, 0, sizeof(meta));
2865 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2868 DPUTS("new mdbenv");
2872 /* Was a mapsize configured? */
2873 if (!env->me_mapsize) {
2874 /* If this is a new environment, take the default,
2875 * else use the size recorded in the existing env.
2877 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2878 } else if (env->me_mapsize < meta.mm_mapsize) {
2879 /* If the configured size is smaller, make sure it's
2880 * still big enough. Silently round up to minimum if not.
2882 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2883 if (env->me_mapsize < minsize)
2884 env->me_mapsize = minsize;
2890 LONG sizelo, sizehi;
2891 sizelo = env->me_mapsize & 0xffffffff;
2892 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2894 /* Windows won't create mappings for zero length files.
2895 * Just allocate the maxsize right now.
2898 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2899 if (!SetEndOfFile(env->me_fd))
2901 SetFilePointer(env->me_fd, 0, NULL, 0);
2903 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2904 PAGE_READWRITE : PAGE_READONLY,
2905 sizehi, sizelo, NULL);
2908 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2909 FILE_MAP_WRITE : FILE_MAP_READ,
2910 0, 0, env->me_mapsize, meta.mm_address);
2918 if (flags & MDB_WRITEMAP) {
2920 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2923 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2925 if (env->me_map == MAP_FAILED) {
2929 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2931 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2933 #ifdef POSIX_MADV_RANDOM
2934 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2935 #endif /* POSIX_MADV_RANDOM */
2936 #endif /* MADV_RANDOM */
2940 if (flags & MDB_FIXEDMAP)
2941 meta.mm_address = env->me_map;
2942 i = mdb_env_init_meta(env, &meta);
2943 if (i != MDB_SUCCESS) {
2946 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2947 /* Can happen because the address argument to mmap() is just a
2948 * hint. mmap() can pick another, e.g. if the range is in use.
2949 * The MAP_FIXED flag would prevent that, but then mmap could
2950 * instead unmap existing pages to make room for the new map.
2952 return EBUSY; /* TODO: Make a new MDB_* error code? */
2954 env->me_psize = meta.mm_psize;
2955 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2956 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2958 env->me_maxpg = env->me_mapsize / env->me_psize;
2960 p = (MDB_page *)env->me_map;
2961 env->me_metas[0] = METADATA(p);
2962 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2966 int toggle = mdb_env_pick_meta(env);
2967 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2969 DPRINTF("opened database version %u, pagesize %u",
2970 env->me_metas[0]->mm_version, env->me_psize);
2971 DPRINTF("using meta page %d", toggle);
2972 DPRINTF("depth: %u", db->md_depth);
2973 DPRINTF("entries: %zu", db->md_entries);
2974 DPRINTF("branch pages: %zu", db->md_branch_pages);
2975 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2976 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2977 DPRINTF("root: %zu", db->md_root);
2985 /** Release a reader thread's slot in the reader lock table.
2986 * This function is called automatically when a thread exits.
2987 * @param[in] ptr This points to the slot in the reader lock table.
2990 mdb_env_reader_dest(void *ptr)
2992 MDB_reader *reader = ptr;
2998 /** Junk for arranging thread-specific callbacks on Windows. This is
2999 * necessarily platform and compiler-specific. Windows supports up
3000 * to 1088 keys. Let's assume nobody opens more than 64 environments
3001 * in a single process, for now. They can override this if needed.
3003 #ifndef MAX_TLS_KEYS
3004 #define MAX_TLS_KEYS 64
3006 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3007 static int mdb_tls_nkeys;
3009 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3013 case DLL_PROCESS_ATTACH: break;
3014 case DLL_THREAD_ATTACH: break;
3015 case DLL_THREAD_DETACH:
3016 for (i=0; i<mdb_tls_nkeys; i++) {
3017 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3018 mdb_env_reader_dest(r);
3021 case DLL_PROCESS_DETACH: break;
3026 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3028 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3032 /* Force some symbol references.
3033 * _tls_used forces the linker to create the TLS directory if not already done
3034 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3036 #pragma comment(linker, "/INCLUDE:_tls_used")
3037 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3038 #pragma const_seg(".CRT$XLB")
3039 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3040 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3043 #pragma comment(linker, "/INCLUDE:__tls_used")
3044 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3045 #pragma data_seg(".CRT$XLB")
3046 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3048 #endif /* WIN 32/64 */
3049 #endif /* !__GNUC__ */
3052 /** Downgrade the exclusive lock on the region back to shared */
3054 mdb_env_share_locks(MDB_env *env, int *excl)
3056 int rc = 0, toggle = mdb_env_pick_meta(env);
3058 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3063 /* First acquire a shared lock. The Unlock will
3064 * then release the existing exclusive lock.
3066 memset(&ov, 0, sizeof(ov));
3067 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3070 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3076 struct flock lock_info;
3077 /* The shared lock replaces the existing lock */
3078 memset((void *)&lock_info, 0, sizeof(lock_info));
3079 lock_info.l_type = F_RDLCK;
3080 lock_info.l_whence = SEEK_SET;
3081 lock_info.l_start = 0;
3082 lock_info.l_len = 1;
3083 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3084 (rc = ErrCode()) == EINTR) ;
3085 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3092 /** Try to get exlusive lock, otherwise shared.
3093 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3096 mdb_env_excl_lock(MDB_env *env, int *excl)
3100 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3104 memset(&ov, 0, sizeof(ov));
3105 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3112 struct flock lock_info;
3113 memset((void *)&lock_info, 0, sizeof(lock_info));
3114 lock_info.l_type = F_WRLCK;
3115 lock_info.l_whence = SEEK_SET;
3116 lock_info.l_start = 0;
3117 lock_info.l_len = 1;
3118 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3119 (rc = ErrCode()) == EINTR) ;
3123 # ifdef MDB_USE_POSIX_SEM
3124 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3127 lock_info.l_type = F_RDLCK;
3128 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3129 (rc = ErrCode()) == EINTR) ;
3137 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3139 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3141 * @(#) $Revision: 5.1 $
3142 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3143 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3145 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3149 * Please do not copyright this code. This code is in the public domain.
3151 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3152 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3153 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3154 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3155 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3156 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3157 * PERFORMANCE OF THIS SOFTWARE.
3160 * chongo <Landon Curt Noll> /\oo/\
3161 * http://www.isthe.com/chongo/
3163 * Share and Enjoy! :-)
3166 typedef unsigned long long mdb_hash_t;
3167 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3169 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3170 * @param[in] str string to hash
3171 * @param[in] hval initial value for hash
3172 * @return 64 bit hash
3174 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3175 * hval arg on the first call.
3178 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3180 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3181 unsigned char *end = s + val->mv_size;
3183 * FNV-1a hash each octet of the string
3186 /* xor the bottom with the current octet */
3187 hval ^= (mdb_hash_t)*s++;
3189 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3190 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3191 (hval << 7) + (hval << 8) + (hval << 40);
3193 /* return our new hash value */
3197 /** Hash the string and output the hash in hex.
3198 * @param[in] str string to hash
3199 * @param[out] hexbuf an array of 17 chars to hold the hash
3202 mdb_hash_hex(MDB_val *val, char *hexbuf)
3205 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3206 for (i=0; i<8; i++) {
3207 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3213 /** Open and/or initialize the lock region for the environment.
3214 * @param[in] env The MDB environment.
3215 * @param[in] lpath The pathname of the file used for the lock region.
3216 * @param[in] mode The Unix permissions for the file, if we create it.
3217 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3218 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3219 * @return 0 on success, non-zero on failure.
3222 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3225 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3227 # define MDB_ERRCODE_ROFS EROFS
3228 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3229 # define MDB_CLOEXEC O_CLOEXEC
3232 # define MDB_CLOEXEC 0
3239 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3240 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3241 FILE_ATTRIBUTE_NORMAL, NULL);
3243 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3245 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3247 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3252 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3253 /* Lose record locks when exec*() */
3254 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3255 fcntl(env->me_lfd, F_SETFD, fdflags);
3258 if (!(env->me_flags & MDB_NOTLS)) {
3259 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3262 env->me_flags |= MDB_ENV_TXKEY;
3264 /* Windows TLS callbacks need help finding their TLS info. */
3265 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3269 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3273 /* Try to get exclusive lock. If we succeed, then
3274 * nobody is using the lock region and we should initialize it.
3276 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3279 size = GetFileSize(env->me_lfd, NULL);
3281 size = lseek(env->me_lfd, 0, SEEK_END);
3283 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3284 if (size < rsize && *excl > 0) {
3286 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3287 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3289 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3293 size = rsize - sizeof(MDB_txninfo);
3294 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3299 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3301 if (!mh) goto fail_errno;
3302 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3304 if (!env->me_txns) goto fail_errno;
3306 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3308 if (m == MAP_FAILED) goto fail_errno;
3314 BY_HANDLE_FILE_INFORMATION stbuf;
3323 if (!mdb_sec_inited) {
3324 InitializeSecurityDescriptor(&mdb_null_sd,
3325 SECURITY_DESCRIPTOR_REVISION);
3326 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3327 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3328 mdb_all_sa.bInheritHandle = FALSE;
3329 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3332 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3333 idbuf.volume = stbuf.dwVolumeSerialNumber;
3334 idbuf.nhigh = stbuf.nFileIndexHigh;
3335 idbuf.nlow = stbuf.nFileIndexLow;
3336 val.mv_data = &idbuf;
3337 val.mv_size = sizeof(idbuf);
3338 mdb_hash_hex(&val, hexbuf);
3339 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3340 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3341 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3342 if (!env->me_rmutex) goto fail_errno;
3343 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3344 if (!env->me_wmutex) goto fail_errno;
3345 #elif defined(MDB_USE_POSIX_SEM)
3354 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3355 idbuf.dev = stbuf.st_dev;
3356 idbuf.ino = stbuf.st_ino;
3357 val.mv_data = &idbuf;
3358 val.mv_size = sizeof(idbuf);
3359 mdb_hash_hex(&val, hexbuf);
3360 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3361 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3362 /* Clean up after a previous run, if needed: Try to
3363 * remove both semaphores before doing anything else.
3365 sem_unlink(env->me_txns->mti_rmname);
3366 sem_unlink(env->me_txns->mti_wmname);
3367 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3368 O_CREAT|O_EXCL, mode, 1);
3369 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3370 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3371 O_CREAT|O_EXCL, mode, 1);
3372 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3373 #else /* MDB_USE_POSIX_SEM */
3374 pthread_mutexattr_t mattr;
3376 if ((rc = pthread_mutexattr_init(&mattr))
3377 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3378 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3379 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3381 pthread_mutexattr_destroy(&mattr);
3382 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3384 env->me_txns->mti_version = MDB_VERSION;
3385 env->me_txns->mti_magic = MDB_MAGIC;
3386 env->me_txns->mti_txnid = 0;
3387 env->me_txns->mti_numreaders = 0;
3390 if (env->me_txns->mti_magic != MDB_MAGIC) {
3391 DPUTS("lock region has invalid magic");
3395 if (env->me_txns->mti_version != MDB_VERSION) {
3396 DPRINTF("lock region is version %u, expected version %u",
3397 env->me_txns->mti_version, MDB_VERSION);
3398 rc = MDB_VERSION_MISMATCH;
3402 if (rc != EACCES && rc != EAGAIN) {
3406 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3407 if (!env->me_rmutex) goto fail_errno;
3408 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3409 if (!env->me_wmutex) goto fail_errno;
3410 #elif defined(MDB_USE_POSIX_SEM)
3411 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3412 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3413 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3414 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3425 /** The name of the lock file in the DB environment */
3426 #define LOCKNAME "/lock.mdb"
3427 /** The name of the data file in the DB environment */
3428 #define DATANAME "/data.mdb"
3429 /** The suffix of the lock file when no subdir is used */
3430 #define LOCKSUFF "-lock"
3431 /** Only a subset of the @ref mdb_env flags can be changed
3432 * at runtime. Changing other flags requires closing the
3433 * environment and re-opening it with the new flags.
3435 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3436 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3439 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3441 int oflags, rc, len, excl = -1;
3442 char *lpath, *dpath;
3444 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3448 if (flags & MDB_NOSUBDIR) {
3449 rc = len + sizeof(LOCKSUFF) + len + 1;
3451 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3456 if (flags & MDB_NOSUBDIR) {
3457 dpath = lpath + len + sizeof(LOCKSUFF);
3458 sprintf(lpath, "%s" LOCKSUFF, path);
3459 strcpy(dpath, path);
3461 dpath = lpath + len + sizeof(LOCKNAME);
3462 sprintf(lpath, "%s" LOCKNAME, path);
3463 sprintf(dpath, "%s" DATANAME, path);
3467 flags |= env->me_flags;
3468 if (flags & MDB_RDONLY) {
3469 /* silently ignore WRITEMAP when we're only getting read access */
3470 flags &= ~MDB_WRITEMAP;
3472 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3473 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3476 env->me_flags = flags |= MDB_ENV_ACTIVE;
3480 env->me_path = strdup(path);
3481 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3482 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3483 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3488 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3493 if (F_ISSET(flags, MDB_RDONLY)) {
3494 oflags = GENERIC_READ;
3495 len = OPEN_EXISTING;
3497 oflags = GENERIC_READ|GENERIC_WRITE;
3500 mode = FILE_ATTRIBUTE_NORMAL;
3501 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3502 NULL, len, mode, NULL);
3504 if (F_ISSET(flags, MDB_RDONLY))
3507 oflags = O_RDWR | O_CREAT;
3509 env->me_fd = open(dpath, oflags, mode);
3511 if (env->me_fd == INVALID_HANDLE_VALUE) {
3516 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3517 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3518 env->me_mfd = env->me_fd;
3520 /* Synchronous fd for meta writes. Needed even with
3521 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3524 env->me_mfd = CreateFile(dpath, oflags,
3525 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3526 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3528 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3530 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3535 DPRINTF("opened dbenv %p", (void *) env);
3537 rc = mdb_env_share_locks(env, &excl);
3543 mdb_env_close0(env, excl);
3549 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3551 mdb_env_close0(MDB_env *env, int excl)
3555 if (!(env->me_flags & MDB_ENV_ACTIVE))
3558 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3559 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3560 free(env->me_dbxs[i].md_name.mv_data);
3562 free(env->me_dbflags);
3565 free(env->me_dirty_list);
3566 if (env->me_free_pgs)
3567 mdb_midl_free(env->me_free_pgs);
3569 if (env->me_flags & MDB_ENV_TXKEY) {
3570 pthread_key_delete(env->me_txkey);
3572 /* Delete our key from the global list */
3573 for (i=0; i<mdb_tls_nkeys; i++)
3574 if (mdb_tls_keys[i] == env->me_txkey) {
3575 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3583 munmap(env->me_map, env->me_mapsize);
3585 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3587 if (env->me_fd != INVALID_HANDLE_VALUE)
3590 pid_t pid = env->me_pid;
3591 /* Clearing readers is done in this function because
3592 * me_txkey with its destructor must be disabled first.
3594 for (i = env->me_numreaders; --i >= 0; )
3595 if (env->me_txns->mti_readers[i].mr_pid == pid)
3596 env->me_txns->mti_readers[i].mr_pid = 0;
3598 if (env->me_rmutex) {
3599 CloseHandle(env->me_rmutex);
3600 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3602 /* Windows automatically destroys the mutexes when
3603 * the last handle closes.
3605 #elif defined(MDB_USE_POSIX_SEM)
3606 if (env->me_rmutex != SEM_FAILED) {
3607 sem_close(env->me_rmutex);
3608 if (env->me_wmutex != SEM_FAILED)
3609 sem_close(env->me_wmutex);
3610 /* If we have the filelock: If we are the
3611 * only remaining user, clean up semaphores.
3614 mdb_env_excl_lock(env, &excl);
3616 sem_unlink(env->me_txns->mti_rmname);
3617 sem_unlink(env->me_txns->mti_wmname);
3621 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3623 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3626 /* Unlock the lockfile. Windows would have unlocked it
3627 * after closing anyway, but not necessarily at once.
3629 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3635 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3639 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3641 MDB_txn *txn = NULL;
3646 /* Do the lock/unlock of the reader mutex before starting the
3647 * write txn. Otherwise other read txns could block writers.
3649 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3654 /* We must start the actual read txn after blocking writers */
3655 mdb_txn_reset0(txn);
3657 /* Temporarily block writers until we snapshot the meta pages */
3660 rc = mdb_txn_renew0(txn);
3662 UNLOCK_MUTEX_W(env);
3667 wsize = env->me_psize * 2;
3671 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3672 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3675 rc = write(fd, env->me_map, wsize);
3676 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3679 UNLOCK_MUTEX_W(env);
3684 ptr = env->me_map + wsize;
3685 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3686 #define MAX_WRITE 2147483648U
3690 if (wsize > MAX_WRITE)
3694 rc = WriteFile(fd, ptr, w2, &len, NULL);
3695 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3704 if (wsize > MAX_WRITE)
3708 wres = write(fd, ptr, w2);
3709 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3722 mdb_env_copy(MDB_env *env, const char *path)
3726 HANDLE newfd = INVALID_HANDLE_VALUE;
3728 if (env->me_flags & MDB_NOSUBDIR) {
3729 lpath = (char *)path;
3732 len += sizeof(DATANAME);
3733 lpath = malloc(len);
3736 sprintf(lpath, "%s" DATANAME, path);
3739 /* The destination path must exist, but the destination file must not.
3740 * We don't want the OS to cache the writes, since the source data is
3741 * already in the OS cache.
3744 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3745 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3747 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3753 if (!(env->me_flags & MDB_NOSUBDIR))
3755 if (newfd == INVALID_HANDLE_VALUE) {
3760 #ifdef F_NOCACHE /* __APPLE__ */
3761 rc = fcntl(newfd, F_NOCACHE, 1);
3768 rc = mdb_env_copyfd(env, newfd);
3771 if (newfd != INVALID_HANDLE_VALUE)
3778 mdb_env_close(MDB_env *env)
3785 VGMEMP_DESTROY(env);
3786 while ((dp = env->me_dpages) != NULL) {
3787 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3788 env->me_dpages = dp->mp_next;
3792 mdb_env_close0(env, 0);
3796 /** Compare two items pointing at aligned size_t's */
3798 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3800 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3801 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3804 /** Compare two items pointing at aligned int's */
3806 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3808 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3809 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3812 /** Compare two items pointing at ints of unknown alignment.
3813 * Nodes and keys are guaranteed to be 2-byte aligned.
3816 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3818 #if BYTE_ORDER == LITTLE_ENDIAN
3819 unsigned short *u, *c;
3822 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3823 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3826 } while(!x && u > (unsigned short *)a->mv_data);
3829 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3833 /** Compare two items lexically */
3835 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3842 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3848 diff = memcmp(a->mv_data, b->mv_data, len);
3849 return diff ? diff : len_diff<0 ? -1 : len_diff;
3852 /** Compare two items in reverse byte order */
3854 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3856 const unsigned char *p1, *p2, *p1_lim;
3860 p1_lim = (const unsigned char *)a->mv_data;
3861 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3862 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3864 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3870 while (p1 > p1_lim) {
3871 diff = *--p1 - *--p2;
3875 return len_diff<0 ? -1 : len_diff;
3878 /** Search for key within a page, using binary search.
3879 * Returns the smallest entry larger or equal to the key.
3880 * If exactp is non-null, stores whether the found entry was an exact match
3881 * in *exactp (1 or 0).
3882 * Updates the cursor index with the index of the found entry.
3883 * If no entry larger or equal to the key is found, returns NULL.
3886 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3888 unsigned int i = 0, nkeys;
3891 MDB_page *mp = mc->mc_pg[mc->mc_top];
3892 MDB_node *node = NULL;
3897 nkeys = NUMKEYS(mp);
3902 COPY_PGNO(pgno, mp->mp_pgno);
3903 DPRINTF("searching %u keys in %s %spage %zu",
3904 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3911 low = IS_LEAF(mp) ? 0 : 1;
3913 cmp = mc->mc_dbx->md_cmp;
3915 /* Branch pages have no data, so if using integer keys,
3916 * alignment is guaranteed. Use faster mdb_cmp_int.
3918 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3919 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3926 nodekey.mv_size = mc->mc_db->md_pad;
3927 node = NODEPTR(mp, 0); /* fake */
3928 while (low <= high) {
3929 i = (low + high) >> 1;
3930 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3931 rc = cmp(key, &nodekey);
3932 DPRINTF("found leaf index %u [%s], rc = %i",
3933 i, DKEY(&nodekey), rc);
3942 while (low <= high) {
3943 i = (low + high) >> 1;
3945 node = NODEPTR(mp, i);
3946 nodekey.mv_size = NODEKSZ(node);
3947 nodekey.mv_data = NODEKEY(node);
3949 rc = cmp(key, &nodekey);
3952 DPRINTF("found leaf index %u [%s], rc = %i",
3953 i, DKEY(&nodekey), rc);
3955 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3956 i, DKEY(&nodekey), NODEPGNO(node), rc);
3967 if (rc > 0) { /* Found entry is less than the key. */
3968 i++; /* Skip to get the smallest entry larger than key. */
3970 node = NODEPTR(mp, i);
3973 *exactp = (rc == 0);
3974 /* store the key index */
3975 mc->mc_ki[mc->mc_top] = i;
3977 /* There is no entry larger or equal to the key. */
3980 /* nodeptr is fake for LEAF2 */
3986 mdb_cursor_adjust(MDB_cursor *mc, func)
3990 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3991 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3998 /** Pop a page off the top of the cursor's stack. */
4000 mdb_cursor_pop(MDB_cursor *mc)
4003 #ifndef MDB_DEBUG_SKIP
4004 MDB_page *top = mc->mc_pg[mc->mc_top];
4010 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
4011 mc->mc_dbi, (void *) mc);
4015 /** Push a page onto the top of the cursor's stack. */
4017 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4019 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4020 mc->mc_dbi, (void *) mc);
4022 if (mc->mc_snum >= CURSOR_STACK) {
4023 assert(mc->mc_snum < CURSOR_STACK);
4024 return MDB_CURSOR_FULL;
4027 mc->mc_top = mc->mc_snum++;
4028 mc->mc_pg[mc->mc_top] = mp;
4029 mc->mc_ki[mc->mc_top] = 0;
4034 /** Find the address of the page corresponding to a given page number.
4035 * @param[in] txn the transaction for this access.
4036 * @param[in] pgno the page number for the page to retrieve.
4037 * @param[out] ret address of a pointer where the page's address will be stored.
4038 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4039 * @return 0 on success, non-zero on failure.
4042 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4047 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4048 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4053 MDB_ID2L dl = tx2->mt_u.dirty_list;
4055 unsigned x = mdb_mid2l_search(dl, pgno);
4056 if (x <= dl[0].mid && dl[x].mid == pgno) {
4062 } while ((tx2 = tx2->mt_parent) != NULL);
4065 if (pgno < txn->mt_next_pgno) {
4067 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4069 DPRINTF("page %zu not found", pgno);
4071 return MDB_PAGE_NOTFOUND;
4081 /** Search for the page a given key should be in.
4082 * Pushes parent pages on the cursor stack. This function continues a
4083 * search on a cursor that has already been initialized. (Usually by
4084 * #mdb_page_search() but also by #mdb_node_move().)
4085 * @param[in,out] mc the cursor for this operation.
4086 * @param[in] key the key to search for. If NULL, search for the lowest
4087 * page. (This is used by #mdb_cursor_first().)
4088 * @param[in] modify If true, visited pages are updated with new page numbers.
4089 * @return 0 on success, non-zero on failure.
4092 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4094 MDB_page *mp = mc->mc_pg[mc->mc_top];
4099 while (IS_BRANCH(mp)) {
4103 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4104 assert(NUMKEYS(mp) > 1);
4105 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4107 if (key == NULL) /* Initialize cursor to first page. */
4109 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4110 /* cursor to last page */
4114 node = mdb_node_search(mc, key, &exact);
4116 i = NUMKEYS(mp) - 1;
4118 i = mc->mc_ki[mc->mc_top];
4127 DPRINTF("following index %u for key [%s]",
4129 assert(i < NUMKEYS(mp));
4130 node = NODEPTR(mp, i);
4132 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4135 mc->mc_ki[mc->mc_top] = i;
4136 if ((rc = mdb_cursor_push(mc, mp)))
4140 if ((rc = mdb_page_touch(mc)) != 0)
4142 mp = mc->mc_pg[mc->mc_top];
4147 DPRINTF("internal error, index points to a %02X page!?",
4149 return MDB_CORRUPTED;
4152 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4153 key ? DKEY(key) : NULL);
4158 /** Search for the lowest key under the current branch page.
4159 * This just bypasses a NUMKEYS check in the current page
4160 * before calling mdb_page_search_root(), because the callers
4161 * are all in situations where the current page is known to
4165 mdb_page_search_lowest(MDB_cursor *mc)
4167 MDB_page *mp = mc->mc_pg[mc->mc_top];
4168 MDB_node *node = NODEPTR(mp, 0);
4171 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4174 mc->mc_ki[mc->mc_top] = 0;
4175 if ((rc = mdb_cursor_push(mc, mp)))
4177 return mdb_page_search_root(mc, NULL, 0);
4180 /** Search for the page a given key should be in.
4181 * Pushes parent pages on the cursor stack. This function just sets up
4182 * the search; it finds the root page for \b mc's database and sets this
4183 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4184 * called to complete the search.
4185 * @param[in,out] mc the cursor for this operation.
4186 * @param[in] key the key to search for. If NULL, search for the lowest
4187 * page. (This is used by #mdb_cursor_first().)
4188 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4189 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4190 * @return 0 on success, non-zero on failure.
4193 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4198 /* Make sure the txn is still viable, then find the root from
4199 * the txn's db table.
4201 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4202 DPUTS("transaction has failed, must abort");
4205 /* Make sure we're using an up-to-date root */
4206 if (mc->mc_dbi > MAIN_DBI) {
4207 if ((*mc->mc_dbflag & DB_STALE) ||
4208 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4210 unsigned char dbflag = 0;
4211 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4212 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4215 if (*mc->mc_dbflag & DB_STALE) {
4219 MDB_node *leaf = mdb_node_search(&mc2,
4220 &mc->mc_dbx->md_name, &exact);
4222 return MDB_NOTFOUND;
4223 mdb_node_read(mc->mc_txn, leaf, &data);
4224 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4226 /* The txn may not know this DBI, or another process may
4227 * have dropped and recreated the DB with other flags.
4229 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4230 return MDB_INCOMPATIBLE;
4231 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4233 if (flags & MDB_PS_MODIFY)
4235 *mc->mc_dbflag &= ~DB_STALE;
4236 *mc->mc_dbflag |= dbflag;
4239 root = mc->mc_db->md_root;
4241 if (root == P_INVALID) { /* Tree is empty. */
4242 DPUTS("tree is empty");
4243 return MDB_NOTFOUND;
4248 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4249 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4255 DPRINTF("db %u root page %zu has flags 0x%X",
4256 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4258 if (flags & MDB_PS_MODIFY) {
4259 if ((rc = mdb_page_touch(mc)))
4263 if (flags & MDB_PS_ROOTONLY)
4266 return mdb_page_search_root(mc, key, flags);
4270 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4272 pgno_t pg = mp->mp_pgno;
4273 unsigned i, ovpages = mp->mp_pages;
4276 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4277 mc->mc_db->md_overflow_pages -= ovpages;
4278 /* If the page is dirty we just acquired it, so we should
4279 * give it back to our current free list, if any.
4280 * Otherwise put it onto the list of pages we freed in this txn.
4282 if ((mp->mp_flags & P_DIRTY) && mc->mc_txn->mt_env->me_pghead) {
4284 pgno_t *mop = mc->mc_txn->mt_env->me_pghead;
4285 /* Remove from dirty list */
4286 x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, pg);
4287 for (; x < mc->mc_txn->mt_u.dirty_list[0].mid; x++)
4288 mc->mc_txn->mt_u.dirty_list[x] = mc->mc_txn->mt_u.dirty_list[x+1];
4289 mc->mc_txn->mt_u.dirty_list[0].mid--;
4290 /* Make room to insert pg */
4291 j = mop[0] + ovpages;
4293 rc = mdb_midl_grow(&mop, ovpages);
4296 mc->mc_txn->mt_env->me_pghead = mc->mc_txn->mt_env->me_pgfree = mop;
4298 for (i = mop[0]; i>0; i--) {
4308 for (i=0; i<ovpages; i++) {
4309 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4316 /** Return the data associated with a given node.
4317 * @param[in] txn The transaction for this operation.
4318 * @param[in] leaf The node being read.
4319 * @param[out] data Updated to point to the node's data.
4320 * @return 0 on success, non-zero on failure.
4323 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4325 MDB_page *omp; /* overflow page */
4329 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4330 data->mv_size = NODEDSZ(leaf);
4331 data->mv_data = NODEDATA(leaf);
4335 /* Read overflow data.
4337 data->mv_size = NODEDSZ(leaf);
4338 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4339 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4340 DPRINTF("read overflow page %zu failed", pgno);
4343 data->mv_data = METADATA(omp);
4349 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4350 MDB_val *key, MDB_val *data)
4359 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4361 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4364 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4368 mdb_cursor_init(&mc, txn, dbi, &mx);
4369 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4372 /** Find a sibling for a page.
4373 * Replaces the page at the top of the cursor's stack with the
4374 * specified sibling, if one exists.
4375 * @param[in] mc The cursor for this operation.
4376 * @param[in] move_right Non-zero if the right sibling is requested,
4377 * otherwise the left sibling.
4378 * @return 0 on success, non-zero on failure.
4381 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4387 if (mc->mc_snum < 2) {
4388 return MDB_NOTFOUND; /* root has no siblings */
4392 DPRINTF("parent page is page %zu, index %u",
4393 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4395 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4396 : (mc->mc_ki[mc->mc_top] == 0)) {
4397 DPRINTF("no more keys left, moving to %s sibling",
4398 move_right ? "right" : "left");
4399 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4400 /* undo cursor_pop before returning */
4407 mc->mc_ki[mc->mc_top]++;
4409 mc->mc_ki[mc->mc_top]--;
4410 DPRINTF("just moving to %s index key %u",
4411 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4413 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4415 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4416 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4419 mdb_cursor_push(mc, mp);
4421 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4426 /** Move the cursor to the next data item. */
4428 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4434 if (mc->mc_flags & C_EOF) {
4435 return MDB_NOTFOUND;
4438 assert(mc->mc_flags & C_INITIALIZED);
4440 mp = mc->mc_pg[mc->mc_top];
4442 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4443 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4444 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4445 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4446 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4447 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4451 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4452 if (op == MDB_NEXT_DUP)
4453 return MDB_NOTFOUND;
4457 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4459 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4460 DPUTS("=====> move to next sibling page");
4461 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4462 mc->mc_flags |= C_EOF;
4463 mc->mc_flags &= ~C_INITIALIZED;
4464 return MDB_NOTFOUND;
4466 mp = mc->mc_pg[mc->mc_top];
4467 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4469 mc->mc_ki[mc->mc_top]++;
4471 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4472 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4475 key->mv_size = mc->mc_db->md_pad;
4476 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4480 assert(IS_LEAF(mp));
4481 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4483 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4484 mdb_xcursor_init1(mc, leaf);
4487 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4490 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4491 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4492 if (rc != MDB_SUCCESS)
4497 MDB_GET_KEY(leaf, key);
4501 /** Move the cursor to the previous data item. */
4503 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4509 assert(mc->mc_flags & C_INITIALIZED);
4511 mp = mc->mc_pg[mc->mc_top];
4513 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4514 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4515 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4516 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4517 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4518 if (op != MDB_PREV || rc == MDB_SUCCESS)
4521 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4522 if (op == MDB_PREV_DUP)
4523 return MDB_NOTFOUND;
4528 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4530 if (mc->mc_ki[mc->mc_top] == 0) {
4531 DPUTS("=====> move to prev sibling page");
4532 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4533 mc->mc_flags &= ~C_INITIALIZED;
4534 return MDB_NOTFOUND;
4536 mp = mc->mc_pg[mc->mc_top];
4537 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4538 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4540 mc->mc_ki[mc->mc_top]--;
4542 mc->mc_flags &= ~C_EOF;
4544 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4545 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4548 key->mv_size = mc->mc_db->md_pad;
4549 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4553 assert(IS_LEAF(mp));
4554 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4556 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4557 mdb_xcursor_init1(mc, leaf);
4560 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4563 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4564 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4565 if (rc != MDB_SUCCESS)
4570 MDB_GET_KEY(leaf, key);
4574 /** Set the cursor on a specific data item. */
4576 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4577 MDB_cursor_op op, int *exactp)
4581 MDB_node *leaf = NULL;
4586 assert(key->mv_size > 0);
4588 /* See if we're already on the right page */
4589 if (mc->mc_flags & C_INITIALIZED) {
4592 mp = mc->mc_pg[mc->mc_top];
4594 mc->mc_ki[mc->mc_top] = 0;
4595 return MDB_NOTFOUND;
4597 if (mp->mp_flags & P_LEAF2) {
4598 nodekey.mv_size = mc->mc_db->md_pad;
4599 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4601 leaf = NODEPTR(mp, 0);
4602 MDB_GET_KEY(leaf, &nodekey);
4604 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4606 /* Probably happens rarely, but first node on the page
4607 * was the one we wanted.
4609 mc->mc_ki[mc->mc_top] = 0;
4616 unsigned int nkeys = NUMKEYS(mp);
4618 if (mp->mp_flags & P_LEAF2) {
4619 nodekey.mv_data = LEAF2KEY(mp,
4620 nkeys-1, nodekey.mv_size);
4622 leaf = NODEPTR(mp, nkeys-1);
4623 MDB_GET_KEY(leaf, &nodekey);
4625 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4627 /* last node was the one we wanted */
4628 mc->mc_ki[mc->mc_top] = nkeys-1;
4634 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4635 /* This is definitely the right page, skip search_page */
4636 if (mp->mp_flags & P_LEAF2) {
4637 nodekey.mv_data = LEAF2KEY(mp,
4638 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4640 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4641 MDB_GET_KEY(leaf, &nodekey);
4643 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4645 /* current node was the one we wanted */
4655 /* If any parents have right-sibs, search.
4656 * Otherwise, there's nothing further.
4658 for (i=0; i<mc->mc_top; i++)
4660 NUMKEYS(mc->mc_pg[i])-1)
4662 if (i == mc->mc_top) {
4663 /* There are no other pages */
4664 mc->mc_ki[mc->mc_top] = nkeys;
4665 return MDB_NOTFOUND;
4669 /* There are no other pages */
4670 mc->mc_ki[mc->mc_top] = 0;
4671 return MDB_NOTFOUND;
4675 rc = mdb_page_search(mc, key, 0);
4676 if (rc != MDB_SUCCESS)
4679 mp = mc->mc_pg[mc->mc_top];
4680 assert(IS_LEAF(mp));
4683 leaf = mdb_node_search(mc, key, exactp);
4684 if (exactp != NULL && !*exactp) {
4685 /* MDB_SET specified and not an exact match. */
4686 return MDB_NOTFOUND;
4690 DPUTS("===> inexact leaf not found, goto sibling");
4691 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4692 return rc; /* no entries matched */
4693 mp = mc->mc_pg[mc->mc_top];
4694 assert(IS_LEAF(mp));
4695 leaf = NODEPTR(mp, 0);
4699 mc->mc_flags |= C_INITIALIZED;
4700 mc->mc_flags &= ~C_EOF;
4703 key->mv_size = mc->mc_db->md_pad;
4704 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4708 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4709 mdb_xcursor_init1(mc, leaf);
4712 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4713 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4714 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4717 if (op == MDB_GET_BOTH) {
4723 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4724 if (rc != MDB_SUCCESS)
4727 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4729 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4731 rc = mc->mc_dbx->md_dcmp(data, &d2);
4733 if (op == MDB_GET_BOTH || rc > 0)
4734 return MDB_NOTFOUND;
4739 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4740 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4745 /* The key already matches in all other cases */
4746 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4747 MDB_GET_KEY(leaf, key);
4748 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4753 /** Move the cursor to the first item in the database. */
4755 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4760 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4761 rc = mdb_page_search(mc, NULL, 0);
4762 if (rc != MDB_SUCCESS)
4765 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4767 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4768 mc->mc_flags |= C_INITIALIZED;
4769 mc->mc_flags &= ~C_EOF;
4771 mc->mc_ki[mc->mc_top] = 0;
4773 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4774 key->mv_size = mc->mc_db->md_pad;
4775 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4780 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4781 mdb_xcursor_init1(mc, leaf);
4782 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4787 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4788 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4792 MDB_GET_KEY(leaf, key);
4796 /** Move the cursor to the last item in the database. */
4798 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4803 if (!(mc->mc_flags & C_EOF)) {
4805 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4808 lkey.mv_size = MDB_MAXKEYSIZE+1;
4809 lkey.mv_data = NULL;
4810 rc = mdb_page_search(mc, &lkey, 0);
4811 if (rc != MDB_SUCCESS)
4814 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4817 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4818 mc->mc_flags |= C_INITIALIZED|C_EOF;
4819 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4821 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4822 key->mv_size = mc->mc_db->md_pad;
4823 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4828 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4829 mdb_xcursor_init1(mc, leaf);
4830 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4835 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4836 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4841 MDB_GET_KEY(leaf, key);
4846 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4855 case MDB_GET_CURRENT:
4856 if (!(mc->mc_flags & C_INITIALIZED)) {
4859 MDB_page *mp = mc->mc_pg[mc->mc_top];
4861 mc->mc_ki[mc->mc_top] = 0;
4867 key->mv_size = mc->mc_db->md_pad;
4868 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4870 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4871 MDB_GET_KEY(leaf, key);
4873 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4874 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4876 rc = mdb_node_read(mc->mc_txn, leaf, data);
4883 case MDB_GET_BOTH_RANGE:
4884 if (data == NULL || mc->mc_xcursor == NULL) {
4892 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4894 } else if (op == MDB_SET_RANGE)
4895 rc = mdb_cursor_set(mc, key, data, op, NULL);
4897 rc = mdb_cursor_set(mc, key, data, op, &exact);
4899 case MDB_GET_MULTIPLE:
4901 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4902 !(mc->mc_flags & C_INITIALIZED)) {
4907 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4908 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4911 case MDB_NEXT_MULTIPLE:
4913 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4917 if (!(mc->mc_flags & C_INITIALIZED))
4918 rc = mdb_cursor_first(mc, key, data);
4920 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4921 if (rc == MDB_SUCCESS) {
4922 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4925 mx = &mc->mc_xcursor->mx_cursor;
4926 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4928 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4929 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4937 case MDB_NEXT_NODUP:
4938 if (!(mc->mc_flags & C_INITIALIZED))
4939 rc = mdb_cursor_first(mc, key, data);
4941 rc = mdb_cursor_next(mc, key, data, op);
4945 case MDB_PREV_NODUP:
4946 if (!(mc->mc_flags & C_INITIALIZED)) {
4947 rc = mdb_cursor_last(mc, key, data);
4948 mc->mc_flags |= C_INITIALIZED;
4949 mc->mc_ki[mc->mc_top]++;
4951 rc = mdb_cursor_prev(mc, key, data, op);
4954 rc = mdb_cursor_first(mc, key, data);
4958 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4959 !(mc->mc_flags & C_INITIALIZED) ||
4960 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4964 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4967 rc = mdb_cursor_last(mc, key, data);
4971 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4972 !(mc->mc_flags & C_INITIALIZED) ||
4973 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4977 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4980 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4988 /** Touch all the pages in the cursor stack.
4989 * Makes sure all the pages are writable, before attempting a write operation.
4990 * @param[in] mc The cursor to operate on.
4993 mdb_cursor_touch(MDB_cursor *mc)
4997 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5000 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5001 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5004 *mc->mc_dbflag |= DB_DIRTY;
5006 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5007 rc = mdb_page_touch(mc);
5011 mc->mc_top = mc->mc_snum-1;
5016 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5019 MDB_node *leaf = NULL;
5020 MDB_val xdata, *rdata, dkey;
5023 int do_sub = 0, insert = 0;
5024 unsigned int mcount = 0;
5028 char dbuf[MDB_MAXKEYSIZE+1];
5029 unsigned int nflags;
5032 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5035 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5038 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5041 #if SIZE_MAX > MAXDATASIZE
5042 if (data->mv_size > MAXDATASIZE)
5046 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
5047 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5051 if (flags == MDB_CURRENT) {
5052 if (!(mc->mc_flags & C_INITIALIZED))
5055 } else if (mc->mc_db->md_root == P_INVALID) {
5057 /* new database, write a root leaf page */
5058 DPUTS("allocating new root leaf page");
5059 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5063 mdb_cursor_push(mc, np);
5064 mc->mc_db->md_root = np->mp_pgno;
5065 mc->mc_db->md_depth++;
5066 *mc->mc_dbflag |= DB_DIRTY;
5067 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5069 np->mp_flags |= P_LEAF2;
5070 mc->mc_flags |= C_INITIALIZED;
5076 if (flags & MDB_APPEND) {
5078 rc = mdb_cursor_last(mc, &k2, &d2);
5080 rc = mc->mc_dbx->md_cmp(key, &k2);
5083 mc->mc_ki[mc->mc_top]++;
5085 /* new key is <= last key */
5090 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5092 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5093 DPRINTF("duplicate key [%s]", DKEY(key));
5095 return MDB_KEYEXIST;
5097 if (rc && rc != MDB_NOTFOUND)
5101 /* Cursor is positioned, now make sure all pages are writable */
5102 rc2 = mdb_cursor_touch(mc);
5107 /* The key already exists */
5108 if (rc == MDB_SUCCESS) {
5109 /* there's only a key anyway, so this is a no-op */
5110 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5111 unsigned int ksize = mc->mc_db->md_pad;
5112 if (key->mv_size != ksize)
5114 if (flags == MDB_CURRENT) {
5115 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5116 memcpy(ptr, key->mv_data, ksize);
5121 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5124 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5125 /* Was a single item before, must convert now */
5127 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5128 /* Just overwrite the current item */
5129 if (flags == MDB_CURRENT)
5132 dkey.mv_size = NODEDSZ(leaf);
5133 dkey.mv_data = NODEDATA(leaf);
5134 #if UINT_MAX < SIZE_MAX
5135 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5136 #ifdef MISALIGNED_OK
5137 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5139 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5142 /* if data matches, ignore it */
5143 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5144 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5146 /* create a fake page for the dup items */
5147 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5148 dkey.mv_data = dbuf;
5149 fp = (MDB_page *)&pbuf;
5150 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5151 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5152 fp->mp_lower = PAGEHDRSZ;
5153 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5154 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5155 fp->mp_flags |= P_LEAF2;
5156 fp->mp_pad = data->mv_size;
5157 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5159 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5160 (dkey.mv_size & 1) + (data->mv_size & 1);
5162 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5165 xdata.mv_size = fp->mp_upper;
5170 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5171 /* See if we need to convert from fake page to subDB */
5173 unsigned int offset;
5176 fp = NODEDATA(leaf);
5177 if (flags == MDB_CURRENT) {
5179 fp->mp_flags |= P_DIRTY;
5180 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5181 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5185 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5186 offset = fp->mp_pad;
5187 if (SIZELEFT(fp) >= offset)
5189 offset *= 4; /* space for 4 more */
5191 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5193 offset += offset & 1;
5194 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5195 offset >= mc->mc_txn->mt_env->me_nodemax) {
5196 /* yes, convert it */
5198 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5199 dummy.md_pad = fp->mp_pad;
5200 dummy.md_flags = MDB_DUPFIXED;
5201 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5202 dummy.md_flags |= MDB_INTEGERKEY;
5205 dummy.md_branch_pages = 0;
5206 dummy.md_leaf_pages = 1;
5207 dummy.md_overflow_pages = 0;
5208 dummy.md_entries = NUMKEYS(fp);
5210 xdata.mv_size = sizeof(MDB_db);
5211 xdata.mv_data = &dummy;
5212 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5214 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5215 flags |= F_DUPDATA|F_SUBDATA;
5216 dummy.md_root = mp->mp_pgno;
5218 /* no, just grow it */
5220 xdata.mv_size = NODEDSZ(leaf) + offset;
5221 xdata.mv_data = &pbuf;
5222 mp = (MDB_page *)&pbuf;
5223 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5226 mp->mp_flags = fp->mp_flags | P_DIRTY;
5227 mp->mp_pad = fp->mp_pad;
5228 mp->mp_lower = fp->mp_lower;
5229 mp->mp_upper = fp->mp_upper + offset;
5231 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5233 nsize = NODEDSZ(leaf) - fp->mp_upper;
5234 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5235 for (i=0; i<NUMKEYS(fp); i++)
5236 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5238 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5242 /* data is on sub-DB, just store it */
5243 flags |= F_DUPDATA|F_SUBDATA;
5247 /* overflow page overwrites need special handling */
5248 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5251 unsigned psize = mc->mc_txn->mt_env->me_psize;
5252 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5254 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5255 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5257 ovpages = omp->mp_pages;
5259 /* Is the ov page writable and large enough? */
5260 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5261 /* yes, overwrite it. Note in this case we don't
5262 * bother to try shrinking the page if the new data
5263 * is smaller than the overflow threshold.
5266 /* It is writable only in a parent txn */
5267 size_t sz = (size_t) psize * ovpages, off;
5268 MDB_page *np = mdb_page_malloc(mc, ovpages);
5274 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5275 if (!(flags & MDB_RESERVE)) {
5276 /* Copy end of page, adjusting alignment so
5277 * compiler may copy words instead of bytes.
5279 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5280 memcpy((size_t *)((char *)np + off),
5281 (size_t *)((char *)omp + off), sz - off);
5284 memcpy(np, omp, sz); /* Copy beginning of page */
5287 SETDSZ(leaf, data->mv_size);
5288 if (F_ISSET(flags, MDB_RESERVE))
5289 data->mv_data = METADATA(omp);
5291 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5294 mdb_ovpage_free(mc, omp);
5296 } else if (NODEDSZ(leaf) == data->mv_size) {
5297 /* same size, just replace it. Note that we could
5298 * also reuse this node if the new data is smaller,
5299 * but instead we opt to shrink the node in that case.
5301 if (F_ISSET(flags, MDB_RESERVE))
5302 data->mv_data = NODEDATA(leaf);
5303 else if (data->mv_size)
5304 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5306 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5309 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5310 mc->mc_db->md_entries--;
5312 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5319 nflags = flags & NODE_ADD_FLAGS;
5320 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5321 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5322 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5323 nflags &= ~MDB_APPEND;
5325 nflags |= MDB_SPLIT_REPLACE;
5326 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5328 /* There is room already in this leaf page. */
5329 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5330 if (rc == 0 && !do_sub && insert) {
5331 /* Adjust other cursors pointing to mp */
5332 MDB_cursor *m2, *m3;
5333 MDB_dbi dbi = mc->mc_dbi;
5334 unsigned i = mc->mc_top;
5335 MDB_page *mp = mc->mc_pg[i];
5337 if (mc->mc_flags & C_SUB)
5340 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5341 if (mc->mc_flags & C_SUB)
5342 m3 = &m2->mc_xcursor->mx_cursor;
5345 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5346 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5353 if (rc != MDB_SUCCESS)
5354 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5356 /* Now store the actual data in the child DB. Note that we're
5357 * storing the user data in the keys field, so there are strict
5358 * size limits on dupdata. The actual data fields of the child
5359 * DB are all zero size.
5366 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5367 if (flags & MDB_CURRENT) {
5368 xflags = MDB_CURRENT;
5370 mdb_xcursor_init1(mc, leaf);
5371 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5373 /* converted, write the original data first */
5375 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5379 /* Adjust other cursors pointing to mp */
5381 unsigned i = mc->mc_top;
5382 MDB_page *mp = mc->mc_pg[i];
5384 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5385 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5386 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5387 mdb_xcursor_init1(m2, leaf);
5391 /* we've done our job */
5394 if (flags & MDB_APPENDDUP)
5395 xflags |= MDB_APPEND;
5396 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5397 if (flags & F_SUBDATA) {
5398 void *db = NODEDATA(leaf);
5399 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5402 /* sub-writes might have failed so check rc again.
5403 * Don't increment count if we just replaced an existing item.
5405 if (!rc && !(flags & MDB_CURRENT))
5406 mc->mc_db->md_entries++;
5407 if (flags & MDB_MULTIPLE) {
5409 if (mcount < data[1].mv_size) {
5410 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5411 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5417 /* If we succeeded and the key didn't exist before, make sure
5418 * the cursor is marked valid.
5421 mc->mc_flags |= C_INITIALIZED;
5426 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5431 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5434 if (!(mc->mc_flags & C_INITIALIZED))
5437 rc = mdb_cursor_touch(mc);
5441 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5443 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5444 if (flags != MDB_NODUPDATA) {
5445 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5446 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5448 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5449 /* If sub-DB still has entries, we're done */
5450 if (mc->mc_xcursor->mx_db.md_entries) {
5451 if (leaf->mn_flags & F_SUBDATA) {
5452 /* update subDB info */
5453 void *db = NODEDATA(leaf);
5454 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5457 /* shrink fake page */
5458 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5459 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5460 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5461 /* fix other sub-DB cursors pointed at this fake page */
5462 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5463 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5464 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5465 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5466 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5469 mc->mc_db->md_entries--;
5472 /* otherwise fall thru and delete the sub-DB */
5475 if (leaf->mn_flags & F_SUBDATA) {
5476 /* add all the child DB's pages to the free list */
5477 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5478 if (rc == MDB_SUCCESS) {
5479 mc->mc_db->md_entries -=
5480 mc->mc_xcursor->mx_db.md_entries;
5485 return mdb_cursor_del0(mc, leaf);
5488 /** Allocate and initialize new pages for a database.
5489 * @param[in] mc a cursor on the database being added to.
5490 * @param[in] flags flags defining what type of page is being allocated.
5491 * @param[in] num the number of pages to allocate. This is usually 1,
5492 * unless allocating overflow pages for a large record.
5493 * @param[out] mp Address of a page, or NULL on failure.
5494 * @return 0 on success, non-zero on failure.
5497 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5502 if ((rc = mdb_page_alloc(mc, num, &np)))
5504 DPRINTF("allocated new mpage %zu, page size %u",
5505 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5506 np->mp_flags = flags | P_DIRTY;
5507 np->mp_lower = PAGEHDRSZ;
5508 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5511 mc->mc_db->md_branch_pages++;
5512 else if (IS_LEAF(np))
5513 mc->mc_db->md_leaf_pages++;
5514 else if (IS_OVERFLOW(np)) {
5515 mc->mc_db->md_overflow_pages += num;
5523 /** Calculate the size of a leaf node.
5524 * The size depends on the environment's page size; if a data item
5525 * is too large it will be put onto an overflow page and the node
5526 * size will only include the key and not the data. Sizes are always
5527 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5528 * of the #MDB_node headers.
5529 * @param[in] env The environment handle.
5530 * @param[in] key The key for the node.
5531 * @param[in] data The data for the node.
5532 * @return The number of bytes needed to store the node.
5535 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5539 sz = LEAFSIZE(key, data);
5540 if (sz >= env->me_nodemax) {
5541 /* put on overflow page */
5542 sz -= data->mv_size - sizeof(pgno_t);
5546 return sz + sizeof(indx_t);
5549 /** Calculate the size of a branch node.
5550 * The size should depend on the environment's page size but since
5551 * we currently don't support spilling large keys onto overflow
5552 * pages, it's simply the size of the #MDB_node header plus the
5553 * size of the key. Sizes are always rounded up to an even number
5554 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5555 * @param[in] env The environment handle.
5556 * @param[in] key The key for the node.
5557 * @return The number of bytes needed to store the node.
5560 mdb_branch_size(MDB_env *env, MDB_val *key)
5565 if (sz >= env->me_nodemax) {
5566 /* put on overflow page */
5567 /* not implemented */
5568 /* sz -= key->size - sizeof(pgno_t); */
5571 return sz + sizeof(indx_t);
5574 /** Add a node to the page pointed to by the cursor.
5575 * @param[in] mc The cursor for this operation.
5576 * @param[in] indx The index on the page where the new node should be added.
5577 * @param[in] key The key for the new node.
5578 * @param[in] data The data for the new node, if any.
5579 * @param[in] pgno The page number, if adding a branch node.
5580 * @param[in] flags Flags for the node.
5581 * @return 0 on success, non-zero on failure. Possible errors are:
5583 * <li>ENOMEM - failed to allocate overflow pages for the node.
5584 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5585 * should never happen since all callers already calculate the
5586 * page's free space before calling this function.
5590 mdb_node_add(MDB_cursor *mc, indx_t indx,
5591 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5594 size_t node_size = NODESIZE;
5597 MDB_page *mp = mc->mc_pg[mc->mc_top];
5598 MDB_page *ofp = NULL; /* overflow page */
5601 assert(mp->mp_upper >= mp->mp_lower);
5603 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5604 IS_LEAF(mp) ? "leaf" : "branch",
5605 IS_SUBP(mp) ? "sub-" : "",
5606 mp->mp_pgno, indx, data ? data->mv_size : 0,
5607 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5610 /* Move higher keys up one slot. */
5611 int ksize = mc->mc_db->md_pad, dif;
5612 char *ptr = LEAF2KEY(mp, indx, ksize);
5613 dif = NUMKEYS(mp) - indx;
5615 memmove(ptr+ksize, ptr, dif*ksize);
5616 /* insert new key */
5617 memcpy(ptr, key->mv_data, ksize);
5619 /* Just using these for counting */
5620 mp->mp_lower += sizeof(indx_t);
5621 mp->mp_upper -= ksize - sizeof(indx_t);
5626 node_size += key->mv_size;
5630 if (F_ISSET(flags, F_BIGDATA)) {
5631 /* Data already on overflow page. */
5632 node_size += sizeof(pgno_t);
5633 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5634 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5636 /* Put data on overflow page. */
5637 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5638 data->mv_size, node_size+data->mv_size);
5639 node_size += sizeof(pgno_t);
5640 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5642 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5645 node_size += data->mv_size;
5648 node_size += node_size & 1;
5650 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5651 DPRINTF("not enough room in page %zu, got %u ptrs",
5652 mp->mp_pgno, NUMKEYS(mp));
5653 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5654 mp->mp_upper - mp->mp_lower);
5655 DPRINTF("node size = %zu", node_size);
5656 return MDB_PAGE_FULL;
5659 /* Move higher pointers up one slot. */
5660 for (i = NUMKEYS(mp); i > indx; i--)
5661 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5663 /* Adjust free space offsets. */
5664 ofs = mp->mp_upper - node_size;
5665 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5666 mp->mp_ptrs[indx] = ofs;
5668 mp->mp_lower += sizeof(indx_t);
5670 /* Write the node data. */
5671 node = NODEPTR(mp, indx);
5672 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5673 node->mn_flags = flags;
5675 SETDSZ(node,data->mv_size);
5680 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5685 if (F_ISSET(flags, F_BIGDATA))
5686 memcpy(node->mn_data + key->mv_size, data->mv_data,
5688 else if (F_ISSET(flags, MDB_RESERVE))
5689 data->mv_data = node->mn_data + key->mv_size;
5691 memcpy(node->mn_data + key->mv_size, data->mv_data,
5694 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5696 if (F_ISSET(flags, MDB_RESERVE))
5697 data->mv_data = METADATA(ofp);
5699 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5706 /** Delete the specified node from a page.
5707 * @param[in] mp The page to operate on.
5708 * @param[in] indx The index of the node to delete.
5709 * @param[in] ksize The size of a node. Only used if the page is
5710 * part of a #MDB_DUPFIXED database.
5713 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5716 indx_t i, j, numkeys, ptr;
5723 COPY_PGNO(pgno, mp->mp_pgno);
5724 DPRINTF("delete node %u on %s page %zu", indx,
5725 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5728 assert(indx < NUMKEYS(mp));
5731 int x = NUMKEYS(mp) - 1 - indx;
5732 base = LEAF2KEY(mp, indx, ksize);
5734 memmove(base, base + ksize, x * ksize);
5735 mp->mp_lower -= sizeof(indx_t);
5736 mp->mp_upper += ksize - sizeof(indx_t);
5740 node = NODEPTR(mp, indx);
5741 sz = NODESIZE + node->mn_ksize;
5743 if (F_ISSET(node->mn_flags, F_BIGDATA))
5744 sz += sizeof(pgno_t);
5746 sz += NODEDSZ(node);
5750 ptr = mp->mp_ptrs[indx];
5751 numkeys = NUMKEYS(mp);
5752 for (i = j = 0; i < numkeys; i++) {
5754 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5755 if (mp->mp_ptrs[i] < ptr)
5756 mp->mp_ptrs[j] += sz;
5761 base = (char *)mp + mp->mp_upper;
5762 memmove(base + sz, base, ptr - mp->mp_upper);
5764 mp->mp_lower -= sizeof(indx_t);
5768 /** Compact the main page after deleting a node on a subpage.
5769 * @param[in] mp The main page to operate on.
5770 * @param[in] indx The index of the subpage on the main page.
5773 mdb_node_shrink(MDB_page *mp, indx_t indx)
5780 indx_t i, numkeys, ptr;
5782 node = NODEPTR(mp, indx);
5783 sp = (MDB_page *)NODEDATA(node);
5784 osize = NODEDSZ(node);
5786 delta = sp->mp_upper - sp->mp_lower;
5787 SETDSZ(node, osize - delta);
5788 xp = (MDB_page *)((char *)sp + delta);
5790 /* shift subpage upward */
5792 nsize = NUMKEYS(sp) * sp->mp_pad;
5793 memmove(METADATA(xp), METADATA(sp), nsize);
5796 nsize = osize - sp->mp_upper;
5797 numkeys = NUMKEYS(sp);
5798 for (i=numkeys-1; i>=0; i--)
5799 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5801 xp->mp_upper = sp->mp_lower;
5802 xp->mp_lower = sp->mp_lower;
5803 xp->mp_flags = sp->mp_flags;
5804 xp->mp_pad = sp->mp_pad;
5805 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5807 /* shift lower nodes upward */
5808 ptr = mp->mp_ptrs[indx];
5809 numkeys = NUMKEYS(mp);
5810 for (i = 0; i < numkeys; i++) {
5811 if (mp->mp_ptrs[i] <= ptr)
5812 mp->mp_ptrs[i] += delta;
5815 base = (char *)mp + mp->mp_upper;
5816 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5817 mp->mp_upper += delta;
5820 /** Initial setup of a sorted-dups cursor.
5821 * Sorted duplicates are implemented as a sub-database for the given key.
5822 * The duplicate data items are actually keys of the sub-database.
5823 * Operations on the duplicate data items are performed using a sub-cursor
5824 * initialized when the sub-database is first accessed. This function does
5825 * the preliminary setup of the sub-cursor, filling in the fields that
5826 * depend only on the parent DB.
5827 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5830 mdb_xcursor_init0(MDB_cursor *mc)
5832 MDB_xcursor *mx = mc->mc_xcursor;
5834 mx->mx_cursor.mc_xcursor = NULL;
5835 mx->mx_cursor.mc_txn = mc->mc_txn;
5836 mx->mx_cursor.mc_db = &mx->mx_db;
5837 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5838 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5839 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5840 mx->mx_cursor.mc_snum = 0;
5841 mx->mx_cursor.mc_top = 0;
5842 mx->mx_cursor.mc_flags = C_SUB;
5843 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5844 mx->mx_dbx.md_dcmp = NULL;
5845 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5848 /** Final setup of a sorted-dups cursor.
5849 * Sets up the fields that depend on the data from the main cursor.
5850 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5851 * @param[in] node The data containing the #MDB_db record for the
5852 * sorted-dup database.
5855 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5857 MDB_xcursor *mx = mc->mc_xcursor;
5859 if (node->mn_flags & F_SUBDATA) {
5860 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5861 mx->mx_cursor.mc_pg[0] = 0;
5862 mx->mx_cursor.mc_snum = 0;
5863 mx->mx_cursor.mc_flags = C_SUB;
5865 MDB_page *fp = NODEDATA(node);
5866 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5867 mx->mx_db.md_flags = 0;
5868 mx->mx_db.md_depth = 1;
5869 mx->mx_db.md_branch_pages = 0;
5870 mx->mx_db.md_leaf_pages = 1;
5871 mx->mx_db.md_overflow_pages = 0;
5872 mx->mx_db.md_entries = NUMKEYS(fp);
5873 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5874 mx->mx_cursor.mc_snum = 1;
5875 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5876 mx->mx_cursor.mc_top = 0;
5877 mx->mx_cursor.mc_pg[0] = fp;
5878 mx->mx_cursor.mc_ki[0] = 0;
5879 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5880 mx->mx_db.md_flags = MDB_DUPFIXED;
5881 mx->mx_db.md_pad = fp->mp_pad;
5882 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5883 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5886 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5888 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5890 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5891 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5892 #if UINT_MAX < SIZE_MAX
5893 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5894 #ifdef MISALIGNED_OK
5895 mx->mx_dbx.md_cmp = mdb_cmp_long;
5897 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5902 /** Initialize a cursor for a given transaction and database. */
5904 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5909 mc->mc_db = &txn->mt_dbs[dbi];
5910 mc->mc_dbx = &txn->mt_dbxs[dbi];
5911 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5916 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5918 mc->mc_xcursor = mx;
5919 mdb_xcursor_init0(mc);
5921 mc->mc_xcursor = NULL;
5923 if (*mc->mc_dbflag & DB_STALE) {
5924 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5929 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5932 size_t size = sizeof(MDB_cursor);
5934 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5937 /* Allow read access to the freelist */
5938 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5941 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5942 size += sizeof(MDB_xcursor);
5944 if ((mc = malloc(size)) != NULL) {
5945 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5946 if (txn->mt_cursors) {
5947 mc->mc_next = txn->mt_cursors[dbi];
5948 txn->mt_cursors[dbi] = mc;
5949 mc->mc_flags |= C_UNTRACK;
5951 mc->mc_flags |= C_ALLOCD;
5962 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5966 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5969 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5972 flags = mc->mc_flags;
5974 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5976 mc->mc_flags |= (flags & C_ALLOCD);
5980 /* Return the count of duplicate data items for the current key */
5982 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5986 if (mc == NULL || countp == NULL)
5989 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5992 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5993 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5996 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5999 *countp = mc->mc_xcursor->mx_db.md_entries;
6005 mdb_cursor_close(MDB_cursor *mc)
6008 /* remove from txn, if tracked */
6009 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6010 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6011 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6013 *prev = mc->mc_next;
6015 if (mc->mc_flags & C_ALLOCD)
6021 mdb_cursor_txn(MDB_cursor *mc)
6023 if (!mc) return NULL;
6028 mdb_cursor_dbi(MDB_cursor *mc)
6034 /** Replace the key for a node with a new key.
6035 * @param[in] mc Cursor pointing to the node to operate on.
6036 * @param[in] key The new key to use.
6037 * @return 0 on success, non-zero on failure.
6040 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6047 indx_t ptr, i, numkeys, indx;
6050 indx = mc->mc_ki[mc->mc_top];
6051 mp = mc->mc_pg[mc->mc_top];
6052 node = NODEPTR(mp, indx);
6053 ptr = mp->mp_ptrs[indx];
6057 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6058 k2.mv_data = NODEKEY(node);
6059 k2.mv_size = node->mn_ksize;
6060 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6062 mdb_dkey(&k2, kbuf2),
6068 delta0 = delta = key->mv_size - node->mn_ksize;
6070 /* Must be 2-byte aligned. If new key is
6071 * shorter by 1, the shift will be skipped.
6073 delta += (delta & 1);
6075 if (delta > 0 && SIZELEFT(mp) < delta) {
6077 /* not enough space left, do a delete and split */
6078 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6079 pgno = NODEPGNO(node);
6080 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6081 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6084 numkeys = NUMKEYS(mp);
6085 for (i = 0; i < numkeys; i++) {
6086 if (mp->mp_ptrs[i] <= ptr)
6087 mp->mp_ptrs[i] -= delta;
6090 base = (char *)mp + mp->mp_upper;
6091 len = ptr - mp->mp_upper + NODESIZE;
6092 memmove(base - delta, base, len);
6093 mp->mp_upper -= delta;
6095 node = NODEPTR(mp, indx);
6098 /* But even if no shift was needed, update ksize */
6100 node->mn_ksize = key->mv_size;
6103 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6109 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6111 /** Move a node from csrc to cdst.
6114 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6121 unsigned short flags;
6125 /* Mark src and dst as dirty. */
6126 if ((rc = mdb_page_touch(csrc)) ||
6127 (rc = mdb_page_touch(cdst)))
6130 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6131 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6132 key.mv_size = csrc->mc_db->md_pad;
6133 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6135 data.mv_data = NULL;
6139 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6140 assert(!((long)srcnode&1));
6141 srcpg = NODEPGNO(srcnode);
6142 flags = srcnode->mn_flags;
6143 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6144 unsigned int snum = csrc->mc_snum;
6146 /* must find the lowest key below src */
6147 mdb_page_search_lowest(csrc);
6148 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6149 key.mv_size = csrc->mc_db->md_pad;
6150 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6152 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6153 key.mv_size = NODEKSZ(s2);
6154 key.mv_data = NODEKEY(s2);
6156 csrc->mc_snum = snum--;
6157 csrc->mc_top = snum;
6159 key.mv_size = NODEKSZ(srcnode);
6160 key.mv_data = NODEKEY(srcnode);
6162 data.mv_size = NODEDSZ(srcnode);
6163 data.mv_data = NODEDATA(srcnode);
6165 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6166 unsigned int snum = cdst->mc_snum;
6169 /* must find the lowest key below dst */
6170 mdb_page_search_lowest(cdst);
6171 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6172 bkey.mv_size = cdst->mc_db->md_pad;
6173 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6175 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6176 bkey.mv_size = NODEKSZ(s2);
6177 bkey.mv_data = NODEKEY(s2);
6179 cdst->mc_snum = snum--;
6180 cdst->mc_top = snum;
6181 mdb_cursor_copy(cdst, &mn);
6183 rc = mdb_update_key(&mn, &bkey);
6188 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6189 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6190 csrc->mc_ki[csrc->mc_top],
6192 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6193 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6195 /* Add the node to the destination page.
6197 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6198 if (rc != MDB_SUCCESS)
6201 /* Delete the node from the source page.
6203 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6206 /* Adjust other cursors pointing to mp */
6207 MDB_cursor *m2, *m3;
6208 MDB_dbi dbi = csrc->mc_dbi;
6209 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6211 if (csrc->mc_flags & C_SUB)
6214 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6215 if (m2 == csrc) continue;
6216 if (csrc->mc_flags & C_SUB)
6217 m3 = &m2->mc_xcursor->mx_cursor;
6220 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6221 csrc->mc_ki[csrc->mc_top]) {
6222 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6223 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6228 /* Update the parent separators.
6230 if (csrc->mc_ki[csrc->mc_top] == 0) {
6231 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6232 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6233 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6235 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6236 key.mv_size = NODEKSZ(srcnode);
6237 key.mv_data = NODEKEY(srcnode);
6239 DPRINTF("update separator for source page %zu to [%s]",
6240 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6241 mdb_cursor_copy(csrc, &mn);
6244 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6247 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6249 indx_t ix = csrc->mc_ki[csrc->mc_top];
6250 nullkey.mv_size = 0;
6251 csrc->mc_ki[csrc->mc_top] = 0;
6252 rc = mdb_update_key(csrc, &nullkey);
6253 csrc->mc_ki[csrc->mc_top] = ix;
6254 assert(rc == MDB_SUCCESS);
6258 if (cdst->mc_ki[cdst->mc_top] == 0) {
6259 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6260 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6261 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6263 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6264 key.mv_size = NODEKSZ(srcnode);
6265 key.mv_data = NODEKEY(srcnode);
6267 DPRINTF("update separator for destination page %zu to [%s]",
6268 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6269 mdb_cursor_copy(cdst, &mn);
6272 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6275 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6277 indx_t ix = cdst->mc_ki[cdst->mc_top];
6278 nullkey.mv_size = 0;
6279 cdst->mc_ki[cdst->mc_top] = 0;
6280 rc = mdb_update_key(cdst, &nullkey);
6281 cdst->mc_ki[cdst->mc_top] = ix;
6282 assert(rc == MDB_SUCCESS);
6289 /** Merge one page into another.
6290 * The nodes from the page pointed to by \b csrc will
6291 * be copied to the page pointed to by \b cdst and then
6292 * the \b csrc page will be freed.
6293 * @param[in] csrc Cursor pointing to the source page.
6294 * @param[in] cdst Cursor pointing to the destination page.
6297 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6305 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6306 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6308 assert(csrc->mc_snum > 1); /* can't merge root page */
6309 assert(cdst->mc_snum > 1);
6311 /* Mark dst as dirty. */
6312 if ((rc = mdb_page_touch(cdst)))
6315 /* Move all nodes from src to dst.
6317 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6318 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6319 key.mv_size = csrc->mc_db->md_pad;
6320 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6321 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6322 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6323 if (rc != MDB_SUCCESS)
6325 key.mv_data = (char *)key.mv_data + key.mv_size;
6328 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6329 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6330 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6331 unsigned int snum = csrc->mc_snum;
6333 /* must find the lowest key below src */
6334 mdb_page_search_lowest(csrc);
6335 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6336 key.mv_size = csrc->mc_db->md_pad;
6337 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6339 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6340 key.mv_size = NODEKSZ(s2);
6341 key.mv_data = NODEKEY(s2);
6343 csrc->mc_snum = snum--;
6344 csrc->mc_top = snum;
6346 key.mv_size = srcnode->mn_ksize;
6347 key.mv_data = NODEKEY(srcnode);
6350 data.mv_size = NODEDSZ(srcnode);
6351 data.mv_data = NODEDATA(srcnode);
6352 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6353 if (rc != MDB_SUCCESS)
6358 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6359 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);
6361 /* Unlink the src page from parent and add to free list.
6363 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6364 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6367 rc = mdb_update_key(csrc, &key);
6373 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6374 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6375 csrc->mc_db->md_leaf_pages--;
6377 csrc->mc_db->md_branch_pages--;
6379 /* Adjust other cursors pointing to mp */
6380 MDB_cursor *m2, *m3;
6381 MDB_dbi dbi = csrc->mc_dbi;
6382 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6384 if (csrc->mc_flags & C_SUB)
6387 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6388 if (csrc->mc_flags & C_SUB)
6389 m3 = &m2->mc_xcursor->mx_cursor;
6392 if (m3 == csrc) continue;
6393 if (m3->mc_snum < csrc->mc_snum) continue;
6394 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6395 m3->mc_pg[csrc->mc_top] = mp;
6396 m3->mc_ki[csrc->mc_top] += nkeys;
6400 mdb_cursor_pop(csrc);
6402 return mdb_rebalance(csrc);
6405 /** Copy the contents of a cursor.
6406 * @param[in] csrc The cursor to copy from.
6407 * @param[out] cdst The cursor to copy to.
6410 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6414 cdst->mc_txn = csrc->mc_txn;
6415 cdst->mc_dbi = csrc->mc_dbi;
6416 cdst->mc_db = csrc->mc_db;
6417 cdst->mc_dbx = csrc->mc_dbx;
6418 cdst->mc_snum = csrc->mc_snum;
6419 cdst->mc_top = csrc->mc_top;
6420 cdst->mc_flags = csrc->mc_flags;
6422 for (i=0; i<csrc->mc_snum; i++) {
6423 cdst->mc_pg[i] = csrc->mc_pg[i];
6424 cdst->mc_ki[i] = csrc->mc_ki[i];
6428 /** Rebalance the tree after a delete operation.
6429 * @param[in] mc Cursor pointing to the page where rebalancing
6431 * @return 0 on success, non-zero on failure.
6434 mdb_rebalance(MDB_cursor *mc)
6438 unsigned int ptop, minkeys;
6441 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6445 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6446 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6447 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6448 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6452 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6453 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6456 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6457 DPRINTF("no need to rebalance page %zu, above fill threshold",
6463 if (mc->mc_snum < 2) {
6464 MDB_page *mp = mc->mc_pg[0];
6466 DPUTS("Can't rebalance a subpage, ignoring");
6469 if (NUMKEYS(mp) == 0) {
6470 DPUTS("tree is completely empty");
6471 mc->mc_db->md_root = P_INVALID;
6472 mc->mc_db->md_depth = 0;
6473 mc->mc_db->md_leaf_pages = 0;
6474 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6478 /* Adjust other cursors pointing to mp */
6479 MDB_cursor *m2, *m3;
6480 MDB_dbi dbi = mc->mc_dbi;
6482 if (mc->mc_flags & C_SUB)
6485 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6486 if (m2 == mc) continue;
6487 if (mc->mc_flags & C_SUB)
6488 m3 = &m2->mc_xcursor->mx_cursor;
6491 if (m3->mc_snum < mc->mc_snum) continue;
6492 if (m3->mc_pg[0] == mp) {
6498 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6499 DPUTS("collapsing root page!");
6500 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6501 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6502 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6505 mc->mc_db->md_depth--;
6506 mc->mc_db->md_branch_pages--;
6508 /* Adjust other cursors pointing to mp */
6509 MDB_cursor *m2, *m3;
6510 MDB_dbi dbi = mc->mc_dbi;
6512 if (mc->mc_flags & C_SUB)
6515 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6516 if (m2 == mc) continue;
6517 if (mc->mc_flags & C_SUB)
6518 m3 = &m2->mc_xcursor->mx_cursor;
6521 if (m3->mc_snum < mc->mc_snum) continue;
6522 if (m3->mc_pg[0] == mp) {
6523 m3->mc_pg[0] = mc->mc_pg[0];
6530 DPUTS("root page doesn't need rebalancing");
6534 /* The parent (branch page) must have at least 2 pointers,
6535 * otherwise the tree is invalid.
6537 ptop = mc->mc_top-1;
6538 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6540 /* Leaf page fill factor is below the threshold.
6541 * Try to move keys from left or right neighbor, or
6542 * merge with a neighbor page.
6547 mdb_cursor_copy(mc, &mn);
6548 mn.mc_xcursor = NULL;
6550 if (mc->mc_ki[ptop] == 0) {
6551 /* We're the leftmost leaf in our parent.
6553 DPUTS("reading right neighbor");
6555 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6556 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6559 mn.mc_ki[mn.mc_top] = 0;
6560 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6562 /* There is at least one neighbor to the left.
6564 DPUTS("reading left neighbor");
6566 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6567 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6570 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6571 mc->mc_ki[mc->mc_top] = 0;
6574 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6575 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);
6577 /* If the neighbor page is above threshold and has enough keys,
6578 * move one key from it. Otherwise we should try to merge them.
6579 * (A branch page must never have less than 2 keys.)
6581 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6582 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6583 return mdb_node_move(&mn, mc);
6585 if (mc->mc_ki[ptop] == 0)
6586 rc = mdb_page_merge(&mn, mc);
6588 rc = mdb_page_merge(mc, &mn);
6589 mc->mc_flags &= ~C_INITIALIZED;
6594 /** Complete a delete operation started by #mdb_cursor_del(). */
6596 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6600 /* add overflow pages to free list */
6601 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6605 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6606 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
6608 assert(IS_OVERFLOW(omp));
6609 mdb_ovpage_free(mc, omp);
6611 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6612 mc->mc_db->md_entries--;
6613 rc = mdb_rebalance(mc);
6614 if (rc != MDB_SUCCESS)
6615 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6616 /* if mc points past last node in page, invalidate */
6617 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6618 mc->mc_flags &= ~C_INITIALIZED;
6624 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6625 MDB_val *key, MDB_val *data)
6630 MDB_val rdata, *xdata;
6634 assert(key != NULL);
6636 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6638 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6641 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6645 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6649 mdb_cursor_init(&mc, txn, dbi, &mx);
6660 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6662 /* let mdb_page_split know about this cursor if needed:
6663 * delete will trigger a rebalance; if it needs to move
6664 * a node from one page to another, it will have to
6665 * update the parent's separator key(s). If the new sepkey
6666 * is larger than the current one, the parent page may
6667 * run out of space, triggering a split. We need this
6668 * cursor to be consistent until the end of the rebalance.
6670 mc.mc_next = txn->mt_cursors[dbi];
6671 txn->mt_cursors[dbi] = &mc;
6672 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6673 txn->mt_cursors[dbi] = mc.mc_next;
6678 /** Split a page and insert a new node.
6679 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6680 * The cursor will be updated to point to the actual page and index where
6681 * the node got inserted after the split.
6682 * @param[in] newkey The key for the newly inserted node.
6683 * @param[in] newdata The data for the newly inserted node.
6684 * @param[in] newpgno The page number, if the new node is a branch node.
6685 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6686 * @return 0 on success, non-zero on failure.
6689 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6690 unsigned int nflags)
6693 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6696 unsigned int i, j, split_indx, nkeys, pmax;
6698 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6700 MDB_page *mp, *rp, *pp;
6705 mp = mc->mc_pg[mc->mc_top];
6706 newindx = mc->mc_ki[mc->mc_top];
6708 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6709 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6710 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6712 /* Create a right sibling. */
6713 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6715 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6717 if (mc->mc_snum < 2) {
6718 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6720 /* shift current top to make room for new parent */
6721 mc->mc_pg[1] = mc->mc_pg[0];
6722 mc->mc_ki[1] = mc->mc_ki[0];
6725 mc->mc_db->md_root = pp->mp_pgno;
6726 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6727 mc->mc_db->md_depth++;
6730 /* Add left (implicit) pointer. */
6731 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6732 /* undo the pre-push */
6733 mc->mc_pg[0] = mc->mc_pg[1];
6734 mc->mc_ki[0] = mc->mc_ki[1];
6735 mc->mc_db->md_root = mp->mp_pgno;
6736 mc->mc_db->md_depth--;
6743 ptop = mc->mc_top-1;
6744 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6747 mc->mc_flags |= C_SPLITTING;
6748 mdb_cursor_copy(mc, &mn);
6749 mn.mc_pg[mn.mc_top] = rp;
6750 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6752 if (nflags & MDB_APPEND) {
6753 mn.mc_ki[mn.mc_top] = 0;
6755 split_indx = newindx;
6760 nkeys = NUMKEYS(mp);
6761 split_indx = nkeys / 2;
6762 if (newindx < split_indx)
6768 unsigned int lsize, rsize, ksize;
6769 /* Move half of the keys to the right sibling */
6771 x = mc->mc_ki[mc->mc_top] - split_indx;
6772 ksize = mc->mc_db->md_pad;
6773 split = LEAF2KEY(mp, split_indx, ksize);
6774 rsize = (nkeys - split_indx) * ksize;
6775 lsize = (nkeys - split_indx) * sizeof(indx_t);
6776 mp->mp_lower -= lsize;
6777 rp->mp_lower += lsize;
6778 mp->mp_upper += rsize - lsize;
6779 rp->mp_upper -= rsize - lsize;
6780 sepkey.mv_size = ksize;
6781 if (newindx == split_indx) {
6782 sepkey.mv_data = newkey->mv_data;
6784 sepkey.mv_data = split;
6787 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6788 memcpy(rp->mp_ptrs, split, rsize);
6789 sepkey.mv_data = rp->mp_ptrs;
6790 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6791 memcpy(ins, newkey->mv_data, ksize);
6792 mp->mp_lower += sizeof(indx_t);
6793 mp->mp_upper -= ksize - sizeof(indx_t);
6796 memcpy(rp->mp_ptrs, split, x * ksize);
6797 ins = LEAF2KEY(rp, x, ksize);
6798 memcpy(ins, newkey->mv_data, ksize);
6799 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6800 rp->mp_lower += sizeof(indx_t);
6801 rp->mp_upper -= ksize - sizeof(indx_t);
6802 mc->mc_ki[mc->mc_top] = x;
6803 mc->mc_pg[mc->mc_top] = rp;
6808 /* For leaf pages, check the split point based on what
6809 * fits where, since otherwise mdb_node_add can fail.
6811 * This check is only needed when the data items are
6812 * relatively large, such that being off by one will
6813 * make the difference between success or failure.
6815 * It's also relevant if a page happens to be laid out
6816 * such that one half of its nodes are all "small" and
6817 * the other half of its nodes are "large." If the new
6818 * item is also "large" and falls on the half with
6819 * "large" nodes, it also may not fit.
6822 unsigned int psize, nsize;
6823 /* Maximum free space in an empty page */
6824 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6825 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6826 if ((nkeys < 20) || (nsize > pmax/16)) {
6827 if (newindx <= split_indx) {
6830 for (i=0; i<split_indx; i++) {
6831 node = NODEPTR(mp, i);
6832 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6833 if (F_ISSET(node->mn_flags, F_BIGDATA))
6834 psize += sizeof(pgno_t);
6836 psize += NODEDSZ(node);
6840 split_indx = newindx;
6851 for (i=nkeys-1; i>=split_indx; i--) {
6852 node = NODEPTR(mp, i);
6853 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6854 if (F_ISSET(node->mn_flags, F_BIGDATA))
6855 psize += sizeof(pgno_t);
6857 psize += NODEDSZ(node);
6861 split_indx = newindx;
6872 /* First find the separating key between the split pages.
6873 * The case where newindx == split_indx is ambiguous; the
6874 * new item could go to the new page or stay on the original
6875 * page. If newpos == 1 it goes to the new page.
6877 if (newindx == split_indx && newpos) {
6878 sepkey.mv_size = newkey->mv_size;
6879 sepkey.mv_data = newkey->mv_data;
6881 node = NODEPTR(mp, split_indx);
6882 sepkey.mv_size = node->mn_ksize;
6883 sepkey.mv_data = NODEKEY(node);
6887 DPRINTF("separator is [%s]", DKEY(&sepkey));
6889 /* Copy separator key to the parent.
6891 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6895 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6898 if (mn.mc_snum == mc->mc_snum) {
6899 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6900 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6901 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6902 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6907 /* Right page might now have changed parent.
6908 * Check if left page also changed parent.
6910 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6911 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6912 for (i=0; i<ptop; i++) {
6913 mc->mc_pg[i] = mn.mc_pg[i];
6914 mc->mc_ki[i] = mn.mc_ki[i];
6916 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6917 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6921 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6924 mc->mc_flags ^= C_SPLITTING;
6925 if (rc != MDB_SUCCESS) {
6928 if (nflags & MDB_APPEND) {
6929 mc->mc_pg[mc->mc_top] = rp;
6930 mc->mc_ki[mc->mc_top] = 0;
6931 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6934 for (i=0; i<mc->mc_top; i++)
6935 mc->mc_ki[i] = mn.mc_ki[i];
6942 /* Move half of the keys to the right sibling. */
6944 /* grab a page to hold a temporary copy */
6945 copy = mdb_page_malloc(mc, 1);
6949 copy->mp_pgno = mp->mp_pgno;
6950 copy->mp_flags = mp->mp_flags;
6951 copy->mp_lower = PAGEHDRSZ;
6952 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6953 mc->mc_pg[mc->mc_top] = copy;
6954 for (i = j = 0; i <= nkeys; j++) {
6955 if (i == split_indx) {
6956 /* Insert in right sibling. */
6957 /* Reset insert index for right sibling. */
6958 if (i != newindx || (newpos ^ ins_new)) {
6960 mc->mc_pg[mc->mc_top] = rp;
6964 if (i == newindx && !ins_new) {
6965 /* Insert the original entry that caused the split. */
6966 rkey.mv_data = newkey->mv_data;
6967 rkey.mv_size = newkey->mv_size;
6976 /* Update index for the new key. */
6977 mc->mc_ki[mc->mc_top] = j;
6978 } else if (i == nkeys) {
6981 node = NODEPTR(mp, i);
6982 rkey.mv_data = NODEKEY(node);
6983 rkey.mv_size = node->mn_ksize;
6985 xdata.mv_data = NODEDATA(node);
6986 xdata.mv_size = NODEDSZ(node);
6989 pgno = NODEPGNO(node);
6990 flags = node->mn_flags;
6995 if (!IS_LEAF(mp) && j == 0) {
6996 /* First branch index doesn't need key data. */
7000 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7004 nkeys = NUMKEYS(copy);
7005 for (i=0; i<nkeys; i++)
7006 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7007 mp->mp_lower = copy->mp_lower;
7008 mp->mp_upper = copy->mp_upper;
7009 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7010 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7012 /* reset back to original page */
7013 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7014 mc->mc_pg[mc->mc_top] = mp;
7015 if (nflags & MDB_RESERVE) {
7016 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7017 if (!(node->mn_flags & F_BIGDATA))
7018 newdata->mv_data = NODEDATA(node);
7022 /* Make sure mc_ki is still valid.
7024 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7025 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7026 for (i=0; i<ptop; i++) {
7027 mc->mc_pg[i] = mn.mc_pg[i];
7028 mc->mc_ki[i] = mn.mc_ki[i];
7030 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7031 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7035 /* return tmp page to freelist */
7036 mdb_page_free(mc->mc_txn->mt_env, copy);
7039 /* Adjust other cursors pointing to mp */
7040 MDB_cursor *m2, *m3;
7041 MDB_dbi dbi = mc->mc_dbi;
7042 int fixup = NUMKEYS(mp);
7044 if (mc->mc_flags & C_SUB)
7047 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7048 if (m2 == mc) continue;
7049 if (mc->mc_flags & C_SUB)
7050 m3 = &m2->mc_xcursor->mx_cursor;
7053 if (!(m3->mc_flags & C_INITIALIZED))
7055 if (m3->mc_flags & C_SPLITTING)
7060 for (k=m3->mc_top; k>=0; k--) {
7061 m3->mc_ki[k+1] = m3->mc_ki[k];
7062 m3->mc_pg[k+1] = m3->mc_pg[k];
7064 if (m3->mc_ki[0] >= split_indx) {
7069 m3->mc_pg[0] = mc->mc_pg[0];
7073 if (m3->mc_pg[mc->mc_top] == mp) {
7074 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7075 m3->mc_ki[mc->mc_top]++;
7076 if (m3->mc_ki[mc->mc_top] >= fixup) {
7077 m3->mc_pg[mc->mc_top] = rp;
7078 m3->mc_ki[mc->mc_top] -= fixup;
7079 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7081 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7082 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7091 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7092 MDB_val *key, MDB_val *data, unsigned int flags)
7097 assert(key != NULL);
7098 assert(data != NULL);
7100 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7103 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7107 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7111 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7114 mdb_cursor_init(&mc, txn, dbi, &mx);
7115 return mdb_cursor_put(&mc, key, data, flags);
7119 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7121 if ((flag & CHANGEABLE) != flag)
7124 env->me_flags |= flag;
7126 env->me_flags &= ~flag;
7131 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7136 *arg = env->me_flags;
7141 mdb_env_get_path(MDB_env *env, const char **arg)
7146 *arg = env->me_path;
7150 /** Common code for #mdb_stat() and #mdb_env_stat().
7151 * @param[in] env the environment to operate in.
7152 * @param[in] db the #MDB_db record containing the stats to return.
7153 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7154 * @return 0, this function always succeeds.
7157 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7159 arg->ms_psize = env->me_psize;
7160 arg->ms_depth = db->md_depth;
7161 arg->ms_branch_pages = db->md_branch_pages;
7162 arg->ms_leaf_pages = db->md_leaf_pages;
7163 arg->ms_overflow_pages = db->md_overflow_pages;
7164 arg->ms_entries = db->md_entries;
7169 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7173 if (env == NULL || arg == NULL)
7176 toggle = mdb_env_pick_meta(env);
7178 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7182 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7186 if (env == NULL || arg == NULL)
7189 toggle = mdb_env_pick_meta(env);
7190 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7191 arg->me_mapsize = env->me_mapsize;
7192 arg->me_maxreaders = env->me_maxreaders;
7193 arg->me_numreaders = env->me_numreaders;
7194 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7195 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7199 /** Set the default comparison functions for a database.
7200 * Called immediately after a database is opened to set the defaults.
7201 * The user can then override them with #mdb_set_compare() or
7202 * #mdb_set_dupsort().
7203 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7204 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7207 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7209 uint16_t f = txn->mt_dbs[dbi].md_flags;
7211 txn->mt_dbxs[dbi].md_cmp =
7212 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7213 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7215 txn->mt_dbxs[dbi].md_dcmp =
7216 !(f & MDB_DUPSORT) ? 0 :
7217 ((f & MDB_INTEGERDUP)
7218 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7219 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7222 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7227 int rc, dbflag, exact;
7228 unsigned int unused = 0;
7231 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7232 mdb_default_cmp(txn, FREE_DBI);
7235 if ((flags & VALID_FLAGS) != flags)
7241 if (flags & PERSISTENT_FLAGS) {
7242 uint16_t f2 = flags & PERSISTENT_FLAGS;
7243 /* make sure flag changes get committed */
7244 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7245 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7246 txn->mt_flags |= MDB_TXN_DIRTY;
7249 mdb_default_cmp(txn, MAIN_DBI);
7253 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7254 mdb_default_cmp(txn, MAIN_DBI);
7257 /* Is the DB already open? */
7259 for (i=2; i<txn->mt_numdbs; i++) {
7260 if (!txn->mt_dbxs[i].md_name.mv_size) {
7261 /* Remember this free slot */
7262 if (!unused) unused = i;
7265 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7266 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7272 /* If no free slot and max hit, fail */
7273 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7274 return MDB_DBS_FULL;
7276 /* Cannot mix named databases with some mainDB flags */
7277 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7278 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7280 /* Find the DB info */
7281 dbflag = DB_NEW|DB_VALID;
7284 key.mv_data = (void *)name;
7285 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7286 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7287 if (rc == MDB_SUCCESS) {
7288 /* make sure this is actually a DB */
7289 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7290 if (!(node->mn_flags & F_SUBDATA))
7292 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7293 /* Create if requested */
7295 data.mv_size = sizeof(MDB_db);
7296 data.mv_data = &dummy;
7297 memset(&dummy, 0, sizeof(dummy));
7298 dummy.md_root = P_INVALID;
7299 dummy.md_flags = flags & PERSISTENT_FLAGS;
7300 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7304 /* OK, got info, add to table */
7305 if (rc == MDB_SUCCESS) {
7306 unsigned int slot = unused ? unused : txn->mt_numdbs;
7307 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7308 txn->mt_dbxs[slot].md_name.mv_size = len;
7309 txn->mt_dbxs[slot].md_rel = NULL;
7310 txn->mt_dbflags[slot] = dbflag;
7311 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7313 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7314 mdb_default_cmp(txn, slot);
7323 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7325 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7328 if (txn->mt_dbflags[dbi] & DB_STALE) {
7331 /* Stale, must read the DB's root. cursor_init does it for us. */
7332 mdb_cursor_init(&mc, txn, dbi, &mx);
7334 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7337 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7340 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7342 ptr = env->me_dbxs[dbi].md_name.mv_data;
7343 env->me_dbxs[dbi].md_name.mv_data = NULL;
7344 env->me_dbxs[dbi].md_name.mv_size = 0;
7345 env->me_dbflags[dbi] = 0;
7349 /** Add all the DB's pages to the free list.
7350 * @param[in] mc Cursor on the DB to free.
7351 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7352 * @return 0 on success, non-zero on failure.
7355 mdb_drop0(MDB_cursor *mc, int subs)
7359 rc = mdb_page_search(mc, NULL, 0);
7360 if (rc == MDB_SUCCESS) {
7365 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7366 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7369 mdb_cursor_copy(mc, &mx);
7370 while (mc->mc_snum > 0) {
7371 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7372 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7373 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7374 if (ni->mn_flags & F_BIGDATA) {
7378 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7379 rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL);
7382 assert(IS_OVERFLOW(omp));
7383 ovpages = omp->mp_pages;
7384 for (j=0; j<ovpages; j++) {
7385 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7388 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7389 mdb_xcursor_init1(mc, ni);
7390 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7396 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7398 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7401 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7406 mc->mc_ki[mc->mc_top] = i;
7407 rc = mdb_cursor_sibling(mc, 1);
7409 /* no more siblings, go back to beginning
7410 * of previous level.
7414 for (i=1; i<mc->mc_snum; i++) {
7416 mc->mc_pg[i] = mx.mc_pg[i];
7421 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7422 mc->mc_db->md_root);
7427 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7432 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7435 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7438 rc = mdb_cursor_open(txn, dbi, &mc);
7442 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7446 /* Can't delete the main DB */
7447 if (del && dbi > MAIN_DBI) {
7448 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7450 txn->mt_dbflags[dbi] = DB_STALE;
7451 mdb_dbi_close(txn->mt_env, dbi);
7454 /* reset the DB record, mark it dirty */
7455 txn->mt_dbflags[dbi] |= DB_DIRTY;
7456 txn->mt_dbs[dbi].md_depth = 0;
7457 txn->mt_dbs[dbi].md_branch_pages = 0;
7458 txn->mt_dbs[dbi].md_leaf_pages = 0;
7459 txn->mt_dbs[dbi].md_overflow_pages = 0;
7460 txn->mt_dbs[dbi].md_entries = 0;
7461 txn->mt_dbs[dbi].md_root = P_INVALID;
7463 txn->mt_flags |= MDB_TXN_DIRTY;
7466 mdb_cursor_close(mc);
7470 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7472 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7475 txn->mt_dbxs[dbi].md_cmp = cmp;
7479 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7481 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7484 txn->mt_dbxs[dbi].md_dcmp = cmp;
7488 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7490 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7493 txn->mt_dbxs[dbi].md_rel = rel;
7497 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7499 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7502 txn->mt_dbxs[dbi].md_relctx = ctx;