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 */
933 /** The database environment. */
935 HANDLE me_fd; /**< The main data file */
936 HANDLE me_lfd; /**< The lock file */
937 HANDLE me_mfd; /**< just for writing the meta pages */
938 /** Failed to update the meta page. Probably an I/O error. */
939 #define MDB_FATAL_ERROR 0x80000000U
940 /** Some fields are initialized. */
941 #define MDB_ENV_ACTIVE 0x20000000U
942 /** me_txkey is set */
943 #define MDB_ENV_TXKEY 0x10000000U
944 uint32_t me_flags; /**< @ref mdb_env */
945 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
946 unsigned int me_maxreaders; /**< size of the reader table */
947 unsigned int me_numreaders; /**< max numreaders set by this env */
948 MDB_dbi me_numdbs; /**< number of DBs opened */
949 MDB_dbi me_maxdbs; /**< size of the DB table */
950 pid_t me_pid; /**< process ID of this env */
951 char *me_path; /**< path to the DB files */
952 char *me_map; /**< the memory map of the data file */
953 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
954 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
955 MDB_txn *me_txn; /**< current write transaction */
956 size_t me_mapsize; /**< size of the data memory map */
957 off_t me_size; /**< current file size */
958 pgno_t me_maxpg; /**< me_mapsize / me_psize */
959 MDB_dbx *me_dbxs; /**< array of static DB info */
960 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
961 pthread_key_t me_txkey; /**< thread-key for readers */
962 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
963 # define me_pglast me_pgstate.mf_pglast
964 # define me_pghead me_pgstate.mf_pghead
965 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
966 /** IDL of pages that became unused in a write txn */
968 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
969 MDB_ID2L me_dirty_list;
970 /** Max number of freelist items that can fit in a single overflow page */
972 /** Max size of a node on a page */
973 unsigned int me_nodemax;
975 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
977 #elif defined(MDB_USE_POSIX_SEM)
978 sem_t *me_rmutex; /* Shared mutexes are not supported */
983 /** Nested transaction */
984 typedef struct MDB_ntxn {
985 MDB_txn mnt_txn; /* the transaction */
986 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
989 /** max number of pages to commit in one writev() call */
990 #define MDB_COMMIT_PAGES 64
991 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
992 #undef MDB_COMMIT_PAGES
993 #define MDB_COMMIT_PAGES IOV_MAX
996 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
997 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
998 static int mdb_page_touch(MDB_cursor *mc);
1000 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1001 static int mdb_page_search_root(MDB_cursor *mc,
1002 MDB_val *key, int modify);
1003 #define MDB_PS_MODIFY 1
1004 #define MDB_PS_ROOTONLY 2
1005 static int mdb_page_search(MDB_cursor *mc,
1006 MDB_val *key, int flags);
1007 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1009 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1010 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1011 pgno_t newpgno, unsigned int nflags);
1013 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1014 static int mdb_env_pick_meta(const MDB_env *env);
1015 static int mdb_env_write_meta(MDB_txn *txn);
1016 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1017 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1019 static void mdb_env_close0(MDB_env *env, int excl);
1021 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1022 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1023 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1024 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1025 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1026 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1027 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1028 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1029 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1031 static int mdb_rebalance(MDB_cursor *mc);
1032 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1034 static void mdb_cursor_pop(MDB_cursor *mc);
1035 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1037 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1038 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1039 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1040 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1041 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1043 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1044 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1046 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1047 static void mdb_xcursor_init0(MDB_cursor *mc);
1048 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1050 static int mdb_drop0(MDB_cursor *mc, int subs);
1051 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1054 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1058 static SECURITY_DESCRIPTOR mdb_null_sd;
1059 static SECURITY_ATTRIBUTES mdb_all_sa;
1060 static int mdb_sec_inited;
1063 /** Return the library version info. */
1065 mdb_version(int *major, int *minor, int *patch)
1067 if (major) *major = MDB_VERSION_MAJOR;
1068 if (minor) *minor = MDB_VERSION_MINOR;
1069 if (patch) *patch = MDB_VERSION_PATCH;
1070 return MDB_VERSION_STRING;
1073 /** Table of descriptions for MDB @ref errors */
1074 static char *const mdb_errstr[] = {
1075 "MDB_KEYEXIST: Key/data pair already exists",
1076 "MDB_NOTFOUND: No matching key/data pair found",
1077 "MDB_PAGE_NOTFOUND: Requested page not found",
1078 "MDB_CORRUPTED: Located page was wrong type",
1079 "MDB_PANIC: Update of meta page failed",
1080 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1081 "MDB_INVALID: File is not an MDB file",
1082 "MDB_MAP_FULL: Environment mapsize limit reached",
1083 "MDB_DBS_FULL: Environment maxdbs limit reached",
1084 "MDB_READERS_FULL: Environment maxreaders limit reached",
1085 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1086 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1087 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1088 "MDB_PAGE_FULL: Internal error - page has no more space",
1089 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1090 "MDB_INCOMPATIBLE: Database flags changed or would change",
1091 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1095 mdb_strerror(int err)
1099 return ("Successful return: 0");
1101 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1102 i = err - MDB_KEYEXIST;
1103 return mdb_errstr[i];
1106 return strerror(err);
1110 /** Display a key in hexadecimal and return the address of the result.
1111 * @param[in] key the key to display
1112 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1113 * @return The key in hexadecimal form.
1116 mdb_dkey(MDB_val *key, char *buf)
1119 unsigned char *c = key->mv_data;
1125 if (key->mv_size > MDB_MAXKEYSIZE)
1126 return "MDB_MAXKEYSIZE";
1127 /* may want to make this a dynamic check: if the key is mostly
1128 * printable characters, print it as-is instead of converting to hex.
1132 for (i=0; i<key->mv_size; i++)
1133 ptr += sprintf(ptr, "%02x", *c++);
1135 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1140 /** Display all the keys in the page. */
1142 mdb_page_list(MDB_page *mp)
1145 unsigned int i, nkeys, nsize;
1149 nkeys = NUMKEYS(mp);
1150 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1151 for (i=0; i<nkeys; i++) {
1152 node = NODEPTR(mp, i);
1153 key.mv_size = node->mn_ksize;
1154 key.mv_data = node->mn_data;
1155 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1156 if (IS_BRANCH(mp)) {
1157 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1160 if (F_ISSET(node->mn_flags, F_BIGDATA))
1161 nsize += sizeof(pgno_t);
1163 nsize += NODEDSZ(node);
1164 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1170 mdb_cursor_chk(MDB_cursor *mc)
1176 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1177 for (i=0; i<mc->mc_top; i++) {
1179 node = NODEPTR(mp, mc->mc_ki[i]);
1180 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1183 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1189 /** Count all the pages in each DB and in the freelist
1190 * and make sure it matches the actual number of pages
1193 static void mdb_audit(MDB_txn *txn)
1197 MDB_ID freecount, count;
1202 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1203 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1204 freecount += *(MDB_ID *)data.mv_data;
1207 for (i = 0; i<txn->mt_numdbs; i++) {
1209 mdb_cursor_init(&mc, txn, i, &mx);
1210 if (txn->mt_dbs[i].md_root == P_INVALID)
1212 count += txn->mt_dbs[i].md_branch_pages +
1213 txn->mt_dbs[i].md_leaf_pages +
1214 txn->mt_dbs[i].md_overflow_pages;
1215 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1216 mdb_page_search(&mc, NULL, 0);
1220 mp = mc.mc_pg[mc.mc_top];
1221 for (j=0; j<NUMKEYS(mp); j++) {
1222 MDB_node *leaf = NODEPTR(mp, j);
1223 if (leaf->mn_flags & F_SUBDATA) {
1225 memcpy(&db, NODEDATA(leaf), sizeof(db));
1226 count += db.md_branch_pages + db.md_leaf_pages +
1227 db.md_overflow_pages;
1231 while (mdb_cursor_sibling(&mc, 1) == 0);
1234 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1235 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1236 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1242 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1244 return txn->mt_dbxs[dbi].md_cmp(a, b);
1248 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1250 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1253 /** Allocate a page.
1254 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1257 mdb_page_malloc(MDB_cursor *mc, unsigned num)
1259 MDB_env *env = mc->mc_txn->mt_env;
1260 MDB_page *ret = env->me_dpages;
1261 size_t sz = env->me_psize;
1264 VGMEMP_ALLOC(env, ret, sz);
1265 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1266 env->me_dpages = ret->mp_next;
1272 if ((ret = malloc(sz)) != NULL) {
1273 VGMEMP_ALLOC(env, ret, sz);
1278 /** Free a single page.
1279 * Saves single pages to a list, for future reuse.
1280 * (This is not used for multi-page overflow pages.)
1283 mdb_page_free(MDB_env *env, MDB_page *mp)
1285 mp->mp_next = env->me_dpages;
1286 VGMEMP_FREE(env, mp);
1287 env->me_dpages = mp;
1290 /* Return all dirty pages to dpage list */
1292 mdb_dlist_free(MDB_txn *txn)
1294 MDB_env *env = txn->mt_env;
1295 MDB_ID2L dl = txn->mt_u.dirty_list;
1296 unsigned i, n = dl[0].mid;
1298 for (i = 1; i <= n; i++) {
1299 MDB_page *dp = dl[i].mptr;
1300 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1301 mdb_page_free(env, dp);
1303 /* large pages just get freed directly */
1304 VGMEMP_FREE(env, dp);
1311 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1313 mdb_find_oldest(MDB_txn *txn)
1316 txnid_t mr, oldest = txn->mt_txnid - 1;
1317 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1318 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1328 /** Allocate pages for writing.
1329 * If there are free pages available from older transactions, they
1330 * will be re-used first. Otherwise a new page will be allocated.
1331 * @param[in] mc cursor A cursor handle identifying the transaction and
1332 * database for which we are allocating.
1333 * @param[in] num the number of pages to allocate.
1334 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1335 * will always be satisfied by a single contiguous chunk of memory.
1336 * @return 0 on success, non-zero on failure.
1339 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1341 MDB_txn *txn = mc->mc_txn;
1343 pgno_t pgno = P_INVALID;
1345 txnid_t oldest = 0, last;
1350 /* If our dirty list is already full, we can't do anything */
1351 if (txn->mt_dirty_room == 0)
1352 return MDB_TXN_FULL;
1354 /* The free list won't have any content at all until txn 2 has
1355 * committed. The pages freed by txn 2 will be unreferenced
1356 * after txn 3 commits, and so will be safe to re-use in txn 4.
1358 if (txn->mt_txnid > 3) {
1359 if (!txn->mt_env->me_pghead &&
1360 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1361 /* See if there's anything in the free DB */
1367 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1368 if (!txn->mt_env->me_pglast) {
1369 mdb_page_search(&m2, NULL, 0);
1370 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1371 kptr = (txnid_t *)NODEKEY(leaf);
1376 last = txn->mt_env->me_pglast + 1;
1378 key.mv_data = &last;
1379 key.mv_size = sizeof(last);
1380 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1383 last = *(txnid_t *)key.mv_data;
1387 oldest = mdb_find_oldest(txn);
1389 if (oldest > last) {
1390 /* It's usable, grab it.
1394 if (!txn->mt_env->me_pglast) {
1395 mdb_node_read(txn, leaf, &data);
1397 idl = (MDB_ID *) data.mv_data;
1398 /* We might have a zero-length IDL due to freelist growth
1399 * during a prior commit
1402 txn->mt_env->me_pglast = last;
1405 mop = mdb_midl_alloc(idl[0]);
1408 txn->mt_env->me_pglast = last;
1409 txn->mt_env->me_pghead = mop;
1410 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1415 DPRINTF("IDL read txn %zu root %zu num %zu",
1416 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1417 for (i = idl[0]; i; i--)
1418 DPRINTF("IDL %zu", idl[i]);
1424 if (txn->mt_env->me_pghead) {
1425 pgno_t *mop = txn->mt_env->me_pghead;
1428 int retry = 1, readit = 0, n2 = num-1;
1429 unsigned int i, j, k;
1431 /* If current list is too short, must fetch more and coalesce */
1432 if (mop[0] < (unsigned)num)
1435 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1437 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1438 /* If on freelist, don't try to read more. If what we have
1439 * right now isn't enough just use new pages.
1440 * TODO: get all of this working. Many circular dependencies...
1442 if (mc->mc_dbi == FREE_DBI) {
1451 last = txn->mt_env->me_pglast + 1;
1453 /* We haven't hit the readers list yet? */
1455 oldest = mdb_find_oldest(txn);
1458 /* There's nothing we can use on the freelist */
1459 if (oldest - last < 1)
1462 key.mv_data = &last;
1463 key.mv_size = sizeof(last);
1464 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1466 if (rc == MDB_NOTFOUND)
1470 last = *(txnid_t*)key.mv_data;
1473 idl = (MDB_ID *) data.mv_data;
1474 mop2 = mdb_midl_alloc(idl[0] + mop[0]);
1477 /* merge in sorted order */
1478 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1480 while (i>0 || j>0) {
1481 if (i && idl[i] < mop[j])
1482 mop2[k--] = idl[i--];
1484 mop2[k--] = mop[j--];
1486 txn->mt_env->me_pglast = last;
1487 mdb_midl_free(txn->mt_env->me_pghead);
1488 txn->mt_env->me_pghead = mop2;
1490 /* Keep trying to read until we have enough */
1491 if (mop[0] < (unsigned)num) {
1496 /* current list has enough pages, but are they contiguous? */
1497 for (i=mop[0]; i>=(unsigned)num; i--) {
1498 if (mop[i-n2] == mop[i] + n2) {
1501 /* move any stragglers down */
1502 for (j=i+num; j<=mop[0]; j++)
1509 /* Stop if we succeeded, or no retries */
1510 if (!retry || pgno != P_INVALID)
1516 /* peel pages off tail, so we only have to truncate the list */
1517 pgno = MDB_IDL_LAST(mop);
1520 if (MDB_IDL_IS_ZERO(mop)) {
1521 mdb_midl_free(txn->mt_env->me_pghead);
1522 txn->mt_env->me_pghead = NULL;
1527 if (pgno == P_INVALID) {
1528 /* DB size is maxed out */
1529 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1530 DPUTS("DB size maxed out");
1531 return MDB_MAP_FULL;
1534 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1535 if (pgno == P_INVALID) {
1536 pgno = txn->mt_next_pgno;
1537 txn->mt_next_pgno += num;
1539 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1542 if (!(np = mdb_page_malloc(mc, num)))
1544 if (pgno == P_INVALID) {
1545 np->mp_pgno = txn->mt_next_pgno;
1546 txn->mt_next_pgno += num;
1551 mid.mid = np->mp_pgno;
1553 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1554 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1556 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1558 txn->mt_dirty_room--;
1564 /** Copy a page: avoid copying unused portions of the page.
1565 * @param[in] dst page to copy into
1566 * @param[in] src page to copy from
1567 * @param[in] psize size of a page
1570 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1572 dst->mp_flags = src->mp_flags | P_DIRTY;
1573 dst->mp_pages = src->mp_pages;
1575 if (IS_LEAF2(src)) {
1576 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1578 unsigned int i, nkeys = NUMKEYS(src);
1579 for (i=0; i<nkeys; i++)
1580 dst->mp_ptrs[i] = src->mp_ptrs[i];
1581 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1582 psize - src->mp_upper);
1586 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1587 * @param[in] mc cursor pointing to the page to be touched
1588 * @return 0 on success, non-zero on failure.
1591 mdb_page_touch(MDB_cursor *mc)
1593 MDB_page *mp = mc->mc_pg[mc->mc_top];
1597 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1599 if ((rc = mdb_page_alloc(mc, 1, &np)))
1601 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1602 assert(mp->mp_pgno != np->mp_pgno);
1603 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1605 /* If page isn't full, just copy the used portion */
1606 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1609 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1611 np->mp_flags |= P_DIRTY;
1616 /* Adjust other cursors pointing to mp */
1617 if (mc->mc_flags & C_SUB) {
1618 MDB_cursor *m2, *m3;
1619 MDB_dbi dbi = mc->mc_dbi-1;
1621 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1622 if (m2 == mc) continue;
1623 m3 = &m2->mc_xcursor->mx_cursor;
1624 if (m3->mc_snum < mc->mc_snum) continue;
1625 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1626 m3->mc_pg[mc->mc_top] = mp;
1632 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1633 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1634 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1635 m2->mc_pg[mc->mc_top] = mp;
1636 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1637 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
1638 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
1639 if (!(leaf->mn_flags & F_SUBDATA)) {
1640 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1646 mc->mc_pg[mc->mc_top] = mp;
1647 /** If this page has a parent, update the parent to point to
1651 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1653 mc->mc_db->md_root = mp->mp_pgno;
1654 } else if (mc->mc_txn->mt_parent && !(mp->mp_flags & P_SUBP)) {
1656 MDB_ID2 mid, *dl = mc->mc_txn->mt_u.dirty_list;
1657 /* If txn has a parent, make sure the page is in our
1661 unsigned x = mdb_mid2l_search(dl, mp->mp_pgno);
1662 if (x <= dl[0].mid && dl[x].mid == mp->mp_pgno) {
1665 mc->mc_pg[mc->mc_top] = np;
1669 assert(dl[0].mid < MDB_IDL_UM_MAX);
1671 np = mdb_page_malloc(mc, 1);
1674 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1675 mid.mid = np->mp_pgno;
1677 mdb_mid2l_insert(dl, &mid);
1685 mdb_env_sync(MDB_env *env, int force)
1688 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1689 if (env->me_flags & MDB_WRITEMAP) {
1690 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1691 ? MS_ASYNC : MS_SYNC;
1692 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1695 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1699 if (MDB_FDATASYNC(env->me_fd))
1706 /** Make shadow copies of all of parent txn's cursors */
1708 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1710 MDB_cursor *mc, *m2;
1711 unsigned int i, j, size;
1713 for (i=0;i<src->mt_numdbs; i++) {
1714 if (src->mt_cursors[i]) {
1715 size = sizeof(MDB_cursor);
1716 if (src->mt_cursors[i]->mc_xcursor)
1717 size += sizeof(MDB_xcursor);
1718 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1725 mc->mc_db = &dst->mt_dbs[i];
1726 mc->mc_dbx = m2->mc_dbx;
1727 mc->mc_dbflag = &dst->mt_dbflags[i];
1728 mc->mc_snum = m2->mc_snum;
1729 mc->mc_top = m2->mc_top;
1730 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1731 for (j=0; j<mc->mc_snum; j++) {
1732 mc->mc_pg[j] = m2->mc_pg[j];
1733 mc->mc_ki[j] = m2->mc_ki[j];
1735 if (m2->mc_xcursor) {
1736 MDB_xcursor *mx, *mx2;
1737 mx = (MDB_xcursor *)(mc+1);
1738 mc->mc_xcursor = mx;
1739 mx2 = m2->mc_xcursor;
1740 mx->mx_db = mx2->mx_db;
1741 mx->mx_dbx = mx2->mx_dbx;
1742 mx->mx_dbflag = mx2->mx_dbflag;
1743 mx->mx_cursor.mc_txn = dst;
1744 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1745 mx->mx_cursor.mc_db = &mx->mx_db;
1746 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1747 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1748 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1749 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1750 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1751 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1752 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1753 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1756 mc->mc_xcursor = NULL;
1758 mc->mc_next = dst->mt_cursors[i];
1759 dst->mt_cursors[i] = mc;
1766 /** Close this write txn's cursors, after optionally merging its shadow
1767 * cursors back into parent's.
1768 * @param[in] txn the transaction handle.
1769 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1770 * @return 0 on success, non-zero on failure.
1773 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1775 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1778 for (i = txn->mt_numdbs; --i >= 0; ) {
1779 for (mc = cursors[i]; mc; mc = next) {
1781 if (mc->mc_flags & merge) {
1782 MDB_cursor *m2 = mc->mc_orig;
1783 m2->mc_snum = mc->mc_snum;
1784 m2->mc_top = mc->mc_top;
1785 for (j = mc->mc_snum; --j >= 0; ) {
1786 m2->mc_pg[j] = mc->mc_pg[j];
1787 m2->mc_ki[j] = mc->mc_ki[j];
1790 if (mc->mc_flags & C_ALLOCD)
1798 mdb_txn_reset0(MDB_txn *txn);
1800 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1801 * @param[in] txn the transaction handle to initialize
1802 * @return 0 on success, non-zero on failure.
1805 mdb_txn_renew0(MDB_txn *txn)
1807 MDB_env *env = txn->mt_env;
1810 int rc, new_notls = 0;
1813 txn->mt_numdbs = env->me_numdbs;
1814 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1816 if (txn->mt_flags & MDB_TXN_RDONLY) {
1817 if (!env->me_txns) {
1818 i = mdb_env_pick_meta(env);
1819 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1820 txn->mt_u.reader = NULL;
1822 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1823 pthread_getspecific(env->me_txkey);
1825 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1826 return MDB_BAD_RSLOT;
1828 pid_t pid = env->me_pid;
1829 pthread_t tid = pthread_self();
1832 for (i=0; i<env->me_txns->mti_numreaders; i++)
1833 if (env->me_txns->mti_readers[i].mr_pid == 0)
1835 if (i == env->me_maxreaders) {
1836 UNLOCK_MUTEX_R(env);
1837 return MDB_READERS_FULL;
1839 env->me_txns->mti_readers[i].mr_pid = pid;
1840 env->me_txns->mti_readers[i].mr_tid = tid;
1841 if (i >= env->me_txns->mti_numreaders)
1842 env->me_txns->mti_numreaders = i+1;
1843 /* Save numreaders for un-mutexed mdb_env_close() */
1844 env->me_numreaders = env->me_txns->mti_numreaders;
1845 UNLOCK_MUTEX_R(env);
1846 r = &env->me_txns->mti_readers[i];
1847 new_notls = (env->me_flags & MDB_NOTLS);
1848 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1853 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1854 txn->mt_u.reader = r;
1856 txn->mt_toggle = txn->mt_txnid & 1;
1857 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1861 txn->mt_txnid = env->me_txns->mti_txnid;
1862 txn->mt_toggle = txn->mt_txnid & 1;
1863 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1866 if (txn->mt_txnid == mdb_debug_start)
1869 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1870 txn->mt_u.dirty_list = env->me_dirty_list;
1871 txn->mt_u.dirty_list[0].mid = 0;
1872 txn->mt_free_pgs = env->me_free_pgs;
1873 txn->mt_free_pgs[0] = 0;
1877 /* Copy the DB info and flags */
1878 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1879 for (i=2; i<txn->mt_numdbs; i++) {
1880 x = env->me_dbflags[i];
1881 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1882 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1884 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1886 if (env->me_maxpg < txn->mt_next_pgno) {
1887 mdb_txn_reset0(txn);
1889 txn->mt_u.reader->mr_pid = 0;
1890 txn->mt_u.reader = NULL;
1892 return MDB_MAP_RESIZED;
1899 mdb_txn_renew(MDB_txn *txn)
1903 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1906 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1907 DPUTS("environment had fatal error, must shutdown!");
1911 rc = mdb_txn_renew0(txn);
1912 if (rc == MDB_SUCCESS) {
1913 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1914 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1915 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1921 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1925 int rc, size, tsize = sizeof(MDB_txn);
1927 if (env->me_flags & MDB_FATAL_ERROR) {
1928 DPUTS("environment had fatal error, must shutdown!");
1931 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1934 /* Nested transactions: Max 1 child, write txns only, no writemap */
1935 if (parent->mt_child ||
1936 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1937 (env->me_flags & MDB_WRITEMAP))
1941 tsize = sizeof(MDB_ntxn);
1943 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1944 if (!(flags & MDB_RDONLY))
1945 size += env->me_maxdbs * sizeof(MDB_cursor *);
1947 if ((txn = calloc(1, size)) == NULL) {
1948 DPRINTF("calloc: %s", strerror(ErrCode()));
1951 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1952 if (flags & MDB_RDONLY) {
1953 txn->mt_flags |= MDB_TXN_RDONLY;
1954 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1956 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1957 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1963 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1964 if (!txn->mt_u.dirty_list ||
1965 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1967 free(txn->mt_u.dirty_list);
1971 txn->mt_txnid = parent->mt_txnid;
1972 txn->mt_toggle = parent->mt_toggle;
1973 txn->mt_dirty_room = parent->mt_dirty_room;
1974 txn->mt_u.dirty_list[0].mid = 0;
1975 txn->mt_free_pgs[0] = 0;
1976 txn->mt_next_pgno = parent->mt_next_pgno;
1977 parent->mt_child = txn;
1978 txn->mt_parent = parent;
1979 txn->mt_numdbs = parent->mt_numdbs;
1980 txn->mt_dbxs = parent->mt_dbxs;
1981 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1982 /* Copy parent's mt_dbflags, but clear DB_NEW */
1983 for (i=0; i<txn->mt_numdbs; i++)
1984 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1986 ntxn = (MDB_ntxn *)txn;
1987 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1988 if (env->me_pghead) {
1989 size = MDB_IDL_SIZEOF(env->me_pghead);
1990 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1992 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1997 rc = mdb_cursor_shadow(parent, txn);
1999 mdb_txn_reset0(txn);
2001 rc = mdb_txn_renew0(txn);
2007 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
2008 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2009 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2015 /** Export or close DBI handles opened in this txn. */
2017 mdb_dbis_update(MDB_txn *txn, int keep)
2020 MDB_dbi n = txn->mt_numdbs;
2021 MDB_env *env = txn->mt_env;
2022 unsigned char *tdbflags = txn->mt_dbflags;
2024 for (i = n; --i >= 2;) {
2025 if (tdbflags[i] & DB_NEW) {
2027 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2029 char *ptr = env->me_dbxs[i].md_name.mv_data;
2030 env->me_dbxs[i].md_name.mv_data = NULL;
2031 env->me_dbxs[i].md_name.mv_size = 0;
2032 env->me_dbflags[i] = 0;
2037 if (keep && env->me_numdbs < n)
2041 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2042 * May be called twice for readonly txns: First reset it, then abort.
2043 * @param[in] txn the transaction handle to reset
2046 mdb_txn_reset0(MDB_txn *txn)
2048 MDB_env *env = txn->mt_env;
2050 /* Close any DBI handles opened in this txn */
2051 mdb_dbis_update(txn, 0);
2053 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2054 if (txn->mt_u.reader) {
2055 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2056 if (!(env->me_flags & MDB_NOTLS))
2057 txn->mt_u.reader = NULL; /* txn does not own reader */
2059 txn->mt_numdbs = 0; /* close nothing if called again */
2060 txn->mt_dbxs = NULL; /* mark txn as reset */
2062 mdb_cursors_close(txn, 0);
2064 if (!(env->me_flags & MDB_WRITEMAP)) {
2065 mdb_dlist_free(txn);
2067 mdb_midl_free(env->me_pghead);
2069 if (txn->mt_parent) {
2070 txn->mt_parent->mt_child = NULL;
2071 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2072 mdb_midl_free(txn->mt_free_pgs);
2073 free(txn->mt_u.dirty_list);
2076 if (mdb_midl_shrink(&txn->mt_free_pgs))
2077 env->me_free_pgs = txn->mt_free_pgs;
2080 txn->mt_env->me_pghead = NULL;
2081 txn->mt_env->me_pglast = 0;
2084 /* The writer mutex was locked in mdb_txn_begin. */
2085 UNLOCK_MUTEX_W(env);
2090 mdb_txn_reset(MDB_txn *txn)
2095 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2096 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2097 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2099 /* This call is only valid for read-only txns */
2100 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2103 mdb_txn_reset0(txn);
2107 mdb_txn_abort(MDB_txn *txn)
2112 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2113 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2114 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2117 mdb_txn_abort(txn->mt_child);
2119 mdb_txn_reset0(txn);
2120 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2121 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2122 txn->mt_u.reader->mr_pid = 0;
2127 /** Save the freelist as of this transaction to the freeDB.
2128 * This changes the freelist. Keep trying until it stabilizes.
2131 mdb_freelist_save(MDB_txn *txn)
2133 /* env->me_pghead[] can grow and shrink during this call.
2134 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2135 * Page numbers cannot disappear from txn->mt_free_pgs[].
2138 MDB_env *env = txn->mt_env;
2139 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2140 txnid_t pglast = 0, head_id = 0;
2141 pgno_t freecnt = 0, *free_pgs, *mop;
2142 ssize_t head_room = 0, total_room = 0, mop_len;
2144 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2146 if (env->me_pghead || env->me_pglast) {
2147 /* Make sure first page of freeDB is touched and on freelist */
2148 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2149 if (rc && rc != MDB_NOTFOUND)
2154 /* Come back here after each Put() in case freelist changed */
2157 /* If using records from freeDB which we have not yet
2158 * deleted, delete them and any we reserved for me_pghead.
2160 while (pglast < env->me_pglast) {
2161 rc = mdb_cursor_first(&mc, &key, NULL);
2164 pglast = head_id = *(txnid_t *)key.mv_data;
2165 total_room = head_room = 0;
2166 assert(pglast <= env->me_pglast);
2167 rc = mdb_cursor_del(&mc, 0);
2172 /* Save the IDL of pages freed by this txn, to a single record */
2173 if (freecnt < txn->mt_free_pgs[0]) {
2175 /* Make sure last page of freeDB is touched and on freelist */
2176 key.mv_size = MDB_MAXKEYSIZE+1;
2178 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2179 if (rc && rc != MDB_NOTFOUND)
2182 free_pgs = txn->mt_free_pgs;
2183 /* Write to last page of freeDB */
2184 key.mv_size = sizeof(txn->mt_txnid);
2185 key.mv_data = &txn->mt_txnid;
2187 freecnt = free_pgs[0];
2188 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2189 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2192 /* Retry if mt_free_pgs[] grew during the Put() */
2193 free_pgs = txn->mt_free_pgs;
2194 } while (freecnt < free_pgs[0]);
2195 mdb_midl_sort(free_pgs);
2196 memcpy(data.mv_data, free_pgs, data.mv_size);
2199 unsigned int i = free_pgs[0];
2200 DPRINTF("IDL write txn %zu root %zu num %u",
2201 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2203 DPRINTF("IDL %zu", free_pgs[i]);
2209 mop = env->me_pghead;
2210 mop_len = mop ? mop[0] : 0;
2212 /* Reserve records for me_pghead[]. Split it if multi-page,
2213 * to avoid searching freeDB for a page range. Use keys in
2214 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2216 if (total_room >= mop_len) {
2217 if (total_room == mop_len || --more < 0)
2219 } else if (head_room >= maxfree_1pg && head_id > 1) {
2220 /* Keep current record (overflow page), add a new one */
2224 /* (Re)write {key = head_id, IDL length = head_room} */
2225 total_room -= head_room;
2226 head_room = mop_len - total_room;
2227 if (head_room > maxfree_1pg && head_id > 1) {
2228 /* Overflow multi-page for part of me_pghead */
2229 head_room /= head_id; /* amortize page sizes */
2230 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2231 } else if (head_room < 0) {
2232 /* Rare case, not bothering to delete this record */
2235 key.mv_size = sizeof(head_id);
2236 key.mv_data = &head_id;
2237 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2238 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2241 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2242 total_room += head_room;
2245 /* Fill in the reserved, touched me_pghead records. Avoid write ops
2246 * so they cannot rearrange anything, just read the destinations.
2253 rc = mdb_cursor_first(&mc, &key, &data);
2254 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2255 MDB_IDL dest = data.mv_data;
2256 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2258 assert(len >= 0 && *(txnid_t*)key.mv_data <= env->me_pglast);
2262 memcpy(dest, mop -= len, len * sizeof(MDB_ID));
2263 if (! (mop_len -= len))
2271 mdb_txn_commit(MDB_txn *txn)
2281 assert(txn != NULL);
2282 assert(txn->mt_env != NULL);
2284 if (txn->mt_child) {
2285 rc = mdb_txn_commit(txn->mt_child);
2286 txn->mt_child = NULL;
2295 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2296 mdb_dbis_update(txn, 1);
2297 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2302 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2303 DPUTS("error flag is set, can't commit");
2305 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2310 if (txn->mt_parent) {
2311 MDB_txn *parent = txn->mt_parent;
2315 /* Append our free list to parent's */
2316 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2320 mdb_midl_free(txn->mt_free_pgs);
2322 parent->mt_next_pgno = txn->mt_next_pgno;
2323 parent->mt_flags = txn->mt_flags;
2325 /* Merge our cursors into parent's and close them */
2326 mdb_cursors_close(txn, C_SHADOW);
2328 /* Update parent's DB table. */
2329 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2330 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2331 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2332 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2333 for (i=2; i<txn->mt_numdbs; i++) {
2334 /* preserve parent's DB_NEW status */
2335 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2336 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2339 dst = txn->mt_parent->mt_u.dirty_list;
2340 src = txn->mt_u.dirty_list;
2341 /* Find len = length of merging our dirty list with parent's */
2343 dst[0].mid = 0; /* simplify loops */
2344 if (parent->mt_parent) {
2345 len = x + src[0].mid;
2346 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2347 for (i = x; y && i; y--) {
2348 pgno_t yp = src[y].mid;
2349 while (yp < dst[i].mid)
2351 if (yp == dst[i].mid) {
2356 } else { /* Simplify the above for single-ancestor case */
2357 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2359 /* Merge our dirty list with parent's */
2361 for (i = len; y; dst[i--] = src[y--]) {
2362 pgno_t yp = src[y].mid;
2363 while (yp < dst[x].mid)
2364 dst[i--] = dst[x--];
2365 if (yp == dst[x].mid)
2366 free(dst[x--].mptr);
2370 free(txn->mt_u.dirty_list);
2371 parent->mt_dirty_room = txn->mt_dirty_room;
2373 txn->mt_parent->mt_child = NULL;
2374 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2379 if (txn != env->me_txn) {
2380 DPUTS("attempt to commit unknown transaction");
2385 mdb_cursors_close(txn, 0);
2387 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2390 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2391 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2393 /* Update DB root pointers */
2394 if (txn->mt_numdbs > 2) {
2398 data.mv_size = sizeof(MDB_db);
2400 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2401 for (i = 2; i < txn->mt_numdbs; i++) {
2402 if (txn->mt_dbflags[i] & DB_DIRTY) {
2403 data.mv_data = &txn->mt_dbs[i];
2404 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2411 rc = mdb_freelist_save(txn);
2415 mdb_midl_free(env->me_pghead);
2416 env->me_pghead = NULL;
2417 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2418 if (mdb_midl_shrink(&txn->mt_free_pgs))
2419 env->me_free_pgs = txn->mt_free_pgs;
2426 if (env->me_flags & MDB_WRITEMAP) {
2427 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2428 dp = txn->mt_u.dirty_list[i].mptr;
2429 /* clear dirty flag */
2430 dp->mp_flags &= ~P_DIRTY;
2432 txn->mt_u.dirty_list[0].mid = 0;
2436 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2442 /* Windows actually supports scatter/gather I/O, but only on
2443 * unbuffered file handles. Since we're relying on the OS page
2444 * cache for all our data, that's self-defeating. So we just
2445 * write pages one at a time. We use the ov structure to set
2446 * the write offset, to at least save the overhead of a Seek
2450 memset(&ov, 0, sizeof(ov));
2451 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2453 dp = txn->mt_u.dirty_list[i].mptr;
2454 DPRINTF("committing page %zu", dp->mp_pgno);
2455 size = dp->mp_pgno * env->me_psize;
2456 ov.Offset = size & 0xffffffff;
2457 ov.OffsetHigh = size >> 16;
2458 ov.OffsetHigh >>= 16;
2459 /* clear dirty flag */
2460 dp->mp_flags &= ~P_DIRTY;
2461 wsize = env->me_psize;
2462 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2463 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2466 DPRINTF("WriteFile: %d", n);
2473 struct iovec iov[MDB_COMMIT_PAGES];
2477 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2478 dp = txn->mt_u.dirty_list[i].mptr;
2479 if (dp->mp_pgno != next) {
2481 rc = writev(env->me_fd, iov, n);
2485 DPUTS("short write, filesystem full?");
2487 DPRINTF("writev: %s", strerror(n));
2494 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2497 DPRINTF("committing page %zu", dp->mp_pgno);
2498 iov[n].iov_len = env->me_psize;
2499 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2500 iov[n].iov_base = (char *)dp;
2501 size += iov[n].iov_len;
2502 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2503 /* clear dirty flag */
2504 dp->mp_flags &= ~P_DIRTY;
2505 if (++n >= MDB_COMMIT_PAGES) {
2515 rc = writev(env->me_fd, iov, n);
2519 DPUTS("short write, filesystem full?");
2521 DPRINTF("writev: %s", strerror(n));
2528 mdb_dlist_free(txn);
2531 if ((n = mdb_env_sync(env, 0)) != 0 ||
2532 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2540 mdb_dbis_update(txn, 1);
2542 UNLOCK_MUTEX_W(env);
2552 /** Read the environment parameters of a DB environment before
2553 * mapping it into memory.
2554 * @param[in] env the environment handle
2555 * @param[out] meta address of where to store the meta information
2556 * @return 0 on success, non-zero on failure.
2559 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2566 /* We don't know the page size yet, so use a minimum value.
2567 * Read both meta pages so we can use the latest one.
2570 for (i=0; i<2; i++) {
2572 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2574 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2579 else if (rc != MDB_PAGESIZE) {
2583 DPRINTF("read: %s", strerror(err));
2587 p = (MDB_page *)&pbuf;
2589 if (!F_ISSET(p->mp_flags, P_META)) {
2590 DPRINTF("page %zu not a meta page", p->mp_pgno);
2595 if (m->mm_magic != MDB_MAGIC) {
2596 DPUTS("meta has invalid magic");
2600 if (m->mm_version != MDB_VERSION) {
2601 DPRINTF("database is version %u, expected version %u",
2602 m->mm_version, MDB_VERSION);
2603 return MDB_VERSION_MISMATCH;
2607 if (m->mm_txnid > meta->mm_txnid)
2608 memcpy(meta, m, sizeof(*m));
2610 memcpy(meta, m, sizeof(*m));
2612 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2614 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2622 /** Write the environment parameters of a freshly created DB environment.
2623 * @param[in] env the environment handle
2624 * @param[out] meta address of where to store the meta information
2625 * @return 0 on success, non-zero on failure.
2628 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2635 DPUTS("writing new meta page");
2637 GET_PAGESIZE(psize);
2639 meta->mm_magic = MDB_MAGIC;
2640 meta->mm_version = MDB_VERSION;
2641 meta->mm_mapsize = env->me_mapsize;
2642 meta->mm_psize = psize;
2643 meta->mm_last_pg = 1;
2644 meta->mm_flags = env->me_flags & 0xffff;
2645 meta->mm_flags |= MDB_INTEGERKEY;
2646 meta->mm_dbs[0].md_root = P_INVALID;
2647 meta->mm_dbs[1].md_root = P_INVALID;
2649 p = calloc(2, psize);
2651 p->mp_flags = P_META;
2654 memcpy(m, meta, sizeof(*meta));
2656 q = (MDB_page *)((char *)p + psize);
2659 q->mp_flags = P_META;
2662 memcpy(m, meta, sizeof(*meta));
2667 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2668 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2669 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2672 lseek(env->me_fd, 0, SEEK_SET);
2673 rc = write(env->me_fd, p, psize * 2);
2674 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2680 /** Update the environment info to commit a transaction.
2681 * @param[in] txn the transaction that's being committed
2682 * @return 0 on success, non-zero on failure.
2685 mdb_env_write_meta(MDB_txn *txn)
2688 MDB_meta meta, metab, *mp;
2690 int rc, len, toggle;
2697 assert(txn != NULL);
2698 assert(txn->mt_env != NULL);
2700 toggle = !txn->mt_toggle;
2701 DPRINTF("writing meta page %d for root page %zu",
2702 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2705 mp = env->me_metas[toggle];
2707 if (env->me_flags & MDB_WRITEMAP) {
2708 /* Persist any increases of mapsize config */
2709 if (env->me_mapsize > mp->mm_mapsize)
2710 mp->mm_mapsize = env->me_mapsize;
2711 mp->mm_dbs[0] = txn->mt_dbs[0];
2712 mp->mm_dbs[1] = txn->mt_dbs[1];
2713 mp->mm_last_pg = txn->mt_next_pgno - 1;
2714 mp->mm_txnid = txn->mt_txnid;
2715 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2716 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2719 ptr += env->me_psize;
2720 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2727 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2728 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2730 ptr = (char *)&meta;
2731 if (env->me_mapsize > mp->mm_mapsize) {
2732 /* Persist any increases of mapsize config */
2733 meta.mm_mapsize = env->me_mapsize;
2734 off = offsetof(MDB_meta, mm_mapsize);
2736 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2738 len = sizeof(MDB_meta) - off;
2741 meta.mm_dbs[0] = txn->mt_dbs[0];
2742 meta.mm_dbs[1] = txn->mt_dbs[1];
2743 meta.mm_last_pg = txn->mt_next_pgno - 1;
2744 meta.mm_txnid = txn->mt_txnid;
2747 off += env->me_psize;
2750 /* Write to the SYNC fd */
2751 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2752 env->me_fd : env->me_mfd;
2755 memset(&ov, 0, sizeof(ov));
2757 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2760 rc = pwrite(mfd, ptr, len, off);
2765 DPUTS("write failed, disk error?");
2766 /* On a failure, the pagecache still contains the new data.
2767 * Write some old data back, to prevent it from being used.
2768 * Use the non-SYNC fd; we know it will fail anyway.
2770 meta.mm_last_pg = metab.mm_last_pg;
2771 meta.mm_txnid = metab.mm_txnid;
2773 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2775 r2 = pwrite(env->me_fd, ptr, len, off);
2778 env->me_flags |= MDB_FATAL_ERROR;
2782 /* Memory ordering issues are irrelevant; since the entire writer
2783 * is wrapped by wmutex, all of these changes will become visible
2784 * after the wmutex is unlocked. Since the DB is multi-version,
2785 * readers will get consistent data regardless of how fresh or
2786 * how stale their view of these values is.
2788 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2793 /** Check both meta pages to see which one is newer.
2794 * @param[in] env the environment handle
2795 * @return meta toggle (0 or 1).
2798 mdb_env_pick_meta(const MDB_env *env)
2800 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2804 mdb_env_create(MDB_env **env)
2808 e = calloc(1, sizeof(MDB_env));
2812 e->me_maxreaders = DEFAULT_READERS;
2813 e->me_maxdbs = e->me_numdbs = 2;
2814 e->me_fd = INVALID_HANDLE_VALUE;
2815 e->me_lfd = INVALID_HANDLE_VALUE;
2816 e->me_mfd = INVALID_HANDLE_VALUE;
2817 #ifdef MDB_USE_POSIX_SEM
2818 e->me_rmutex = SEM_FAILED;
2819 e->me_wmutex = SEM_FAILED;
2821 e->me_pid = getpid();
2822 VGMEMP_CREATE(e,0,0);
2828 mdb_env_set_mapsize(MDB_env *env, size_t size)
2832 env->me_mapsize = size;
2834 env->me_maxpg = env->me_mapsize / env->me_psize;
2839 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2843 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2848 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2850 if (env->me_map || readers < 1)
2852 env->me_maxreaders = readers;
2857 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2859 if (!env || !readers)
2861 *readers = env->me_maxreaders;
2865 /** Further setup required for opening an MDB environment
2868 mdb_env_open2(MDB_env *env)
2870 unsigned int flags = env->me_flags;
2871 int i, newenv = 0, prot;
2875 memset(&meta, 0, sizeof(meta));
2877 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2880 DPUTS("new mdbenv");
2884 /* Was a mapsize configured? */
2885 if (!env->me_mapsize) {
2886 /* If this is a new environment, take the default,
2887 * else use the size recorded in the existing env.
2889 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2890 } else if (env->me_mapsize < meta.mm_mapsize) {
2891 /* If the configured size is smaller, make sure it's
2892 * still big enough. Silently round up to minimum if not.
2894 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2895 if (env->me_mapsize < minsize)
2896 env->me_mapsize = minsize;
2902 LONG sizelo, sizehi;
2903 sizelo = env->me_mapsize & 0xffffffff;
2904 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2906 /* Windows won't create mappings for zero length files.
2907 * Just allocate the maxsize right now.
2910 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2911 if (!SetEndOfFile(env->me_fd))
2913 SetFilePointer(env->me_fd, 0, NULL, 0);
2915 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2916 PAGE_READWRITE : PAGE_READONLY,
2917 sizehi, sizelo, NULL);
2920 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2921 FILE_MAP_WRITE : FILE_MAP_READ,
2922 0, 0, env->me_mapsize, meta.mm_address);
2930 if (flags & MDB_WRITEMAP) {
2932 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2935 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2937 if (env->me_map == MAP_FAILED) {
2941 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2943 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2945 #ifdef POSIX_MADV_RANDOM
2946 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2947 #endif /* POSIX_MADV_RANDOM */
2948 #endif /* MADV_RANDOM */
2952 if (flags & MDB_FIXEDMAP)
2953 meta.mm_address = env->me_map;
2954 i = mdb_env_init_meta(env, &meta);
2955 if (i != MDB_SUCCESS) {
2958 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2959 /* Can happen because the address argument to mmap() is just a
2960 * hint. mmap() can pick another, e.g. if the range is in use.
2961 * The MAP_FIXED flag would prevent that, but then mmap could
2962 * instead unmap existing pages to make room for the new map.
2964 return EBUSY; /* TODO: Make a new MDB_* error code? */
2966 env->me_psize = meta.mm_psize;
2967 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2968 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2970 env->me_maxpg = env->me_mapsize / env->me_psize;
2972 p = (MDB_page *)env->me_map;
2973 env->me_metas[0] = METADATA(p);
2974 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2978 int toggle = mdb_env_pick_meta(env);
2979 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2981 DPRINTF("opened database version %u, pagesize %u",
2982 env->me_metas[0]->mm_version, env->me_psize);
2983 DPRINTF("using meta page %d", toggle);
2984 DPRINTF("depth: %u", db->md_depth);
2985 DPRINTF("entries: %zu", db->md_entries);
2986 DPRINTF("branch pages: %zu", db->md_branch_pages);
2987 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2988 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2989 DPRINTF("root: %zu", db->md_root);
2997 /** Release a reader thread's slot in the reader lock table.
2998 * This function is called automatically when a thread exits.
2999 * @param[in] ptr This points to the slot in the reader lock table.
3002 mdb_env_reader_dest(void *ptr)
3004 MDB_reader *reader = ptr;
3010 /** Junk for arranging thread-specific callbacks on Windows. This is
3011 * necessarily platform and compiler-specific. Windows supports up
3012 * to 1088 keys. Let's assume nobody opens more than 64 environments
3013 * in a single process, for now. They can override this if needed.
3015 #ifndef MAX_TLS_KEYS
3016 #define MAX_TLS_KEYS 64
3018 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3019 static int mdb_tls_nkeys;
3021 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3025 case DLL_PROCESS_ATTACH: break;
3026 case DLL_THREAD_ATTACH: break;
3027 case DLL_THREAD_DETACH:
3028 for (i=0; i<mdb_tls_nkeys; i++) {
3029 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3030 mdb_env_reader_dest(r);
3033 case DLL_PROCESS_DETACH: break;
3038 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3040 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3044 /* Force some symbol references.
3045 * _tls_used forces the linker to create the TLS directory if not already done
3046 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3048 #pragma comment(linker, "/INCLUDE:_tls_used")
3049 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3050 #pragma const_seg(".CRT$XLB")
3051 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3052 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3055 #pragma comment(linker, "/INCLUDE:__tls_used")
3056 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3057 #pragma data_seg(".CRT$XLB")
3058 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3060 #endif /* WIN 32/64 */
3061 #endif /* !__GNUC__ */
3064 /** Downgrade the exclusive lock on the region back to shared */
3066 mdb_env_share_locks(MDB_env *env, int *excl)
3068 int rc = 0, toggle = mdb_env_pick_meta(env);
3070 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3075 /* First acquire a shared lock. The Unlock will
3076 * then release the existing exclusive lock.
3078 memset(&ov, 0, sizeof(ov));
3079 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3082 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3088 struct flock lock_info;
3089 /* The shared lock replaces the existing lock */
3090 memset((void *)&lock_info, 0, sizeof(lock_info));
3091 lock_info.l_type = F_RDLCK;
3092 lock_info.l_whence = SEEK_SET;
3093 lock_info.l_start = 0;
3094 lock_info.l_len = 1;
3095 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3096 (rc = ErrCode()) == EINTR) ;
3097 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3104 /** Try to get exlusive lock, otherwise shared.
3105 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3108 mdb_env_excl_lock(MDB_env *env, int *excl)
3112 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3116 memset(&ov, 0, sizeof(ov));
3117 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3124 struct flock lock_info;
3125 memset((void *)&lock_info, 0, sizeof(lock_info));
3126 lock_info.l_type = F_WRLCK;
3127 lock_info.l_whence = SEEK_SET;
3128 lock_info.l_start = 0;
3129 lock_info.l_len = 1;
3130 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3131 (rc = ErrCode()) == EINTR) ;
3135 # ifdef MDB_USE_POSIX_SEM
3136 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3139 lock_info.l_type = F_RDLCK;
3140 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3141 (rc = ErrCode()) == EINTR) ;
3149 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3151 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3153 * @(#) $Revision: 5.1 $
3154 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3155 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3157 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3161 * Please do not copyright this code. This code is in the public domain.
3163 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3164 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3165 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3166 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3167 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3168 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3169 * PERFORMANCE OF THIS SOFTWARE.
3172 * chongo <Landon Curt Noll> /\oo/\
3173 * http://www.isthe.com/chongo/
3175 * Share and Enjoy! :-)
3178 typedef unsigned long long mdb_hash_t;
3179 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3181 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3182 * @param[in] str string to hash
3183 * @param[in] hval initial value for hash
3184 * @return 64 bit hash
3186 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3187 * hval arg on the first call.
3190 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3192 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3193 unsigned char *end = s + val->mv_size;
3195 * FNV-1a hash each octet of the string
3198 /* xor the bottom with the current octet */
3199 hval ^= (mdb_hash_t)*s++;
3201 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3202 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3203 (hval << 7) + (hval << 8) + (hval << 40);
3205 /* return our new hash value */
3209 /** Hash the string and output the hash in hex.
3210 * @param[in] str string to hash
3211 * @param[out] hexbuf an array of 17 chars to hold the hash
3214 mdb_hash_hex(MDB_val *val, char *hexbuf)
3217 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3218 for (i=0; i<8; i++) {
3219 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3225 /** Open and/or initialize the lock region for the environment.
3226 * @param[in] env The MDB environment.
3227 * @param[in] lpath The pathname of the file used for the lock region.
3228 * @param[in] mode The Unix permissions for the file, if we create it.
3229 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3230 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3231 * @return 0 on success, non-zero on failure.
3234 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3237 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3239 # define MDB_ERRCODE_ROFS EROFS
3240 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3241 # define MDB_CLOEXEC O_CLOEXEC
3244 # define MDB_CLOEXEC 0
3251 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3252 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3253 FILE_ATTRIBUTE_NORMAL, NULL);
3255 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3257 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3259 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3264 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3265 /* Lose record locks when exec*() */
3266 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3267 fcntl(env->me_lfd, F_SETFD, fdflags);
3270 if (!(env->me_flags & MDB_NOTLS)) {
3271 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3274 env->me_flags |= MDB_ENV_TXKEY;
3276 /* Windows TLS callbacks need help finding their TLS info. */
3277 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3281 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3285 /* Try to get exclusive lock. If we succeed, then
3286 * nobody is using the lock region and we should initialize it.
3288 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3291 size = GetFileSize(env->me_lfd, NULL);
3293 size = lseek(env->me_lfd, 0, SEEK_END);
3295 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3296 if (size < rsize && *excl > 0) {
3298 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3299 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3301 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3305 size = rsize - sizeof(MDB_txninfo);
3306 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3311 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3313 if (!mh) goto fail_errno;
3314 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3316 if (!env->me_txns) goto fail_errno;
3318 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3320 if (m == MAP_FAILED) goto fail_errno;
3326 BY_HANDLE_FILE_INFORMATION stbuf;
3335 if (!mdb_sec_inited) {
3336 InitializeSecurityDescriptor(&mdb_null_sd,
3337 SECURITY_DESCRIPTOR_REVISION);
3338 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3339 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3340 mdb_all_sa.bInheritHandle = FALSE;
3341 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3344 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3345 idbuf.volume = stbuf.dwVolumeSerialNumber;
3346 idbuf.nhigh = stbuf.nFileIndexHigh;
3347 idbuf.nlow = stbuf.nFileIndexLow;
3348 val.mv_data = &idbuf;
3349 val.mv_size = sizeof(idbuf);
3350 mdb_hash_hex(&val, hexbuf);
3351 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3352 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3353 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3354 if (!env->me_rmutex) goto fail_errno;
3355 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3356 if (!env->me_wmutex) goto fail_errno;
3357 #elif defined(MDB_USE_POSIX_SEM)
3366 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3367 idbuf.dev = stbuf.st_dev;
3368 idbuf.ino = stbuf.st_ino;
3369 val.mv_data = &idbuf;
3370 val.mv_size = sizeof(idbuf);
3371 mdb_hash_hex(&val, hexbuf);
3372 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3373 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3374 /* Clean up after a previous run, if needed: Try to
3375 * remove both semaphores before doing anything else.
3377 sem_unlink(env->me_txns->mti_rmname);
3378 sem_unlink(env->me_txns->mti_wmname);
3379 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3380 O_CREAT|O_EXCL, mode, 1);
3381 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3382 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3383 O_CREAT|O_EXCL, mode, 1);
3384 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3385 #else /* MDB_USE_POSIX_SEM */
3386 pthread_mutexattr_t mattr;
3388 if ((rc = pthread_mutexattr_init(&mattr))
3389 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3390 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3391 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3393 pthread_mutexattr_destroy(&mattr);
3394 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3396 env->me_txns->mti_version = MDB_VERSION;
3397 env->me_txns->mti_magic = MDB_MAGIC;
3398 env->me_txns->mti_txnid = 0;
3399 env->me_txns->mti_numreaders = 0;
3402 if (env->me_txns->mti_magic != MDB_MAGIC) {
3403 DPUTS("lock region has invalid magic");
3407 if (env->me_txns->mti_version != MDB_VERSION) {
3408 DPRINTF("lock region is version %u, expected version %u",
3409 env->me_txns->mti_version, MDB_VERSION);
3410 rc = MDB_VERSION_MISMATCH;
3414 if (rc != EACCES && rc != EAGAIN) {
3418 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3419 if (!env->me_rmutex) goto fail_errno;
3420 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3421 if (!env->me_wmutex) goto fail_errno;
3422 #elif defined(MDB_USE_POSIX_SEM)
3423 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3424 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3425 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3426 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3437 /** The name of the lock file in the DB environment */
3438 #define LOCKNAME "/lock.mdb"
3439 /** The name of the data file in the DB environment */
3440 #define DATANAME "/data.mdb"
3441 /** The suffix of the lock file when no subdir is used */
3442 #define LOCKSUFF "-lock"
3443 /** Only a subset of the @ref mdb_env flags can be changed
3444 * at runtime. Changing other flags requires closing the
3445 * environment and re-opening it with the new flags.
3447 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3448 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3451 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3453 int oflags, rc, len, excl = -1;
3454 char *lpath, *dpath;
3456 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3460 if (flags & MDB_NOSUBDIR) {
3461 rc = len + sizeof(LOCKSUFF) + len + 1;
3463 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3468 if (flags & MDB_NOSUBDIR) {
3469 dpath = lpath + len + sizeof(LOCKSUFF);
3470 sprintf(lpath, "%s" LOCKSUFF, path);
3471 strcpy(dpath, path);
3473 dpath = lpath + len + sizeof(LOCKNAME);
3474 sprintf(lpath, "%s" LOCKNAME, path);
3475 sprintf(dpath, "%s" DATANAME, path);
3479 flags |= env->me_flags;
3480 if (flags & MDB_RDONLY) {
3481 /* silently ignore WRITEMAP when we're only getting read access */
3482 flags &= ~MDB_WRITEMAP;
3484 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3485 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3488 env->me_flags = flags |= MDB_ENV_ACTIVE;
3492 env->me_path = strdup(path);
3493 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3494 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3495 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3500 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3505 if (F_ISSET(flags, MDB_RDONLY)) {
3506 oflags = GENERIC_READ;
3507 len = OPEN_EXISTING;
3509 oflags = GENERIC_READ|GENERIC_WRITE;
3512 mode = FILE_ATTRIBUTE_NORMAL;
3513 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3514 NULL, len, mode, NULL);
3516 if (F_ISSET(flags, MDB_RDONLY))
3519 oflags = O_RDWR | O_CREAT;
3521 env->me_fd = open(dpath, oflags, mode);
3523 if (env->me_fd == INVALID_HANDLE_VALUE) {
3528 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3529 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3530 env->me_mfd = env->me_fd;
3532 /* Synchronous fd for meta writes. Needed even with
3533 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3536 env->me_mfd = CreateFile(dpath, oflags,
3537 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3538 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3540 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3542 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3547 DPRINTF("opened dbenv %p", (void *) env);
3549 rc = mdb_env_share_locks(env, &excl);
3555 mdb_env_close0(env, excl);
3561 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3563 mdb_env_close0(MDB_env *env, int excl)
3567 if (!(env->me_flags & MDB_ENV_ACTIVE))
3570 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3571 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3572 free(env->me_dbxs[i].md_name.mv_data);
3574 free(env->me_dbflags);
3577 free(env->me_dirty_list);
3578 if (env->me_free_pgs)
3579 mdb_midl_free(env->me_free_pgs);
3581 if (env->me_flags & MDB_ENV_TXKEY) {
3582 pthread_key_delete(env->me_txkey);
3584 /* Delete our key from the global list */
3585 for (i=0; i<mdb_tls_nkeys; i++)
3586 if (mdb_tls_keys[i] == env->me_txkey) {
3587 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3595 munmap(env->me_map, env->me_mapsize);
3597 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3599 if (env->me_fd != INVALID_HANDLE_VALUE)
3602 pid_t pid = env->me_pid;
3603 /* Clearing readers is done in this function because
3604 * me_txkey with its destructor must be disabled first.
3606 for (i = env->me_numreaders; --i >= 0; )
3607 if (env->me_txns->mti_readers[i].mr_pid == pid)
3608 env->me_txns->mti_readers[i].mr_pid = 0;
3610 if (env->me_rmutex) {
3611 CloseHandle(env->me_rmutex);
3612 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3614 /* Windows automatically destroys the mutexes when
3615 * the last handle closes.
3617 #elif defined(MDB_USE_POSIX_SEM)
3618 if (env->me_rmutex != SEM_FAILED) {
3619 sem_close(env->me_rmutex);
3620 if (env->me_wmutex != SEM_FAILED)
3621 sem_close(env->me_wmutex);
3622 /* If we have the filelock: If we are the
3623 * only remaining user, clean up semaphores.
3626 mdb_env_excl_lock(env, &excl);
3628 sem_unlink(env->me_txns->mti_rmname);
3629 sem_unlink(env->me_txns->mti_wmname);
3633 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3635 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3638 /* Unlock the lockfile. Windows would have unlocked it
3639 * after closing anyway, but not necessarily at once.
3641 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3647 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3651 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3653 MDB_txn *txn = NULL;
3658 /* Do the lock/unlock of the reader mutex before starting the
3659 * write txn. Otherwise other read txns could block writers.
3661 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3666 /* We must start the actual read txn after blocking writers */
3667 mdb_txn_reset0(txn);
3669 /* Temporarily block writers until we snapshot the meta pages */
3672 rc = mdb_txn_renew0(txn);
3674 UNLOCK_MUTEX_W(env);
3679 wsize = env->me_psize * 2;
3683 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3684 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3687 rc = write(fd, env->me_map, wsize);
3688 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3691 UNLOCK_MUTEX_W(env);
3696 ptr = env->me_map + wsize;
3697 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3698 #define MAX_WRITE 2147483648U
3702 if (wsize > MAX_WRITE)
3706 rc = WriteFile(fd, ptr, w2, &len, NULL);
3707 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3716 if (wsize > MAX_WRITE)
3720 wres = write(fd, ptr, w2);
3721 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3734 mdb_env_copy(MDB_env *env, const char *path)
3738 HANDLE newfd = INVALID_HANDLE_VALUE;
3740 if (env->me_flags & MDB_NOSUBDIR) {
3741 lpath = (char *)path;
3744 len += sizeof(DATANAME);
3745 lpath = malloc(len);
3748 sprintf(lpath, "%s" DATANAME, path);
3751 /* The destination path must exist, but the destination file must not.
3752 * We don't want the OS to cache the writes, since the source data is
3753 * already in the OS cache.
3756 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3757 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3759 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3765 if (!(env->me_flags & MDB_NOSUBDIR))
3767 if (newfd == INVALID_HANDLE_VALUE) {
3772 #ifdef F_NOCACHE /* __APPLE__ */
3773 rc = fcntl(newfd, F_NOCACHE, 1);
3780 rc = mdb_env_copyfd(env, newfd);
3783 if (newfd != INVALID_HANDLE_VALUE)
3790 mdb_env_close(MDB_env *env)
3797 VGMEMP_DESTROY(env);
3798 while ((dp = env->me_dpages) != NULL) {
3799 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3800 env->me_dpages = dp->mp_next;
3804 mdb_env_close0(env, 0);
3808 /** Compare two items pointing at aligned size_t's */
3810 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3812 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3813 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3816 /** Compare two items pointing at aligned int's */
3818 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3820 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3821 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3824 /** Compare two items pointing at ints of unknown alignment.
3825 * Nodes and keys are guaranteed to be 2-byte aligned.
3828 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3830 #if BYTE_ORDER == LITTLE_ENDIAN
3831 unsigned short *u, *c;
3834 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3835 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3838 } while(!x && u > (unsigned short *)a->mv_data);
3841 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3845 /** Compare two items lexically */
3847 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3854 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3860 diff = memcmp(a->mv_data, b->mv_data, len);
3861 return diff ? diff : len_diff<0 ? -1 : len_diff;
3864 /** Compare two items in reverse byte order */
3866 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3868 const unsigned char *p1, *p2, *p1_lim;
3872 p1_lim = (const unsigned char *)a->mv_data;
3873 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3874 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3876 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3882 while (p1 > p1_lim) {
3883 diff = *--p1 - *--p2;
3887 return len_diff<0 ? -1 : len_diff;
3890 /** Search for key within a page, using binary search.
3891 * Returns the smallest entry larger or equal to the key.
3892 * If exactp is non-null, stores whether the found entry was an exact match
3893 * in *exactp (1 or 0).
3894 * Updates the cursor index with the index of the found entry.
3895 * If no entry larger or equal to the key is found, returns NULL.
3898 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3900 unsigned int i = 0, nkeys;
3903 MDB_page *mp = mc->mc_pg[mc->mc_top];
3904 MDB_node *node = NULL;
3909 nkeys = NUMKEYS(mp);
3914 COPY_PGNO(pgno, mp->mp_pgno);
3915 DPRINTF("searching %u keys in %s %spage %zu",
3916 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3923 low = IS_LEAF(mp) ? 0 : 1;
3925 cmp = mc->mc_dbx->md_cmp;
3927 /* Branch pages have no data, so if using integer keys,
3928 * alignment is guaranteed. Use faster mdb_cmp_int.
3930 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3931 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3938 nodekey.mv_size = mc->mc_db->md_pad;
3939 node = NODEPTR(mp, 0); /* fake */
3940 while (low <= high) {
3941 i = (low + high) >> 1;
3942 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3943 rc = cmp(key, &nodekey);
3944 DPRINTF("found leaf index %u [%s], rc = %i",
3945 i, DKEY(&nodekey), rc);
3954 while (low <= high) {
3955 i = (low + high) >> 1;
3957 node = NODEPTR(mp, i);
3958 nodekey.mv_size = NODEKSZ(node);
3959 nodekey.mv_data = NODEKEY(node);
3961 rc = cmp(key, &nodekey);
3964 DPRINTF("found leaf index %u [%s], rc = %i",
3965 i, DKEY(&nodekey), rc);
3967 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3968 i, DKEY(&nodekey), NODEPGNO(node), rc);
3979 if (rc > 0) { /* Found entry is less than the key. */
3980 i++; /* Skip to get the smallest entry larger than key. */
3982 node = NODEPTR(mp, i);
3985 *exactp = (rc == 0);
3986 /* store the key index */
3987 mc->mc_ki[mc->mc_top] = i;
3989 /* There is no entry larger or equal to the key. */
3992 /* nodeptr is fake for LEAF2 */
3998 mdb_cursor_adjust(MDB_cursor *mc, func)
4002 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4003 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4010 /** Pop a page off the top of the cursor's stack. */
4012 mdb_cursor_pop(MDB_cursor *mc)
4015 #ifndef MDB_DEBUG_SKIP
4016 MDB_page *top = mc->mc_pg[mc->mc_top];
4022 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
4023 mc->mc_dbi, (void *) mc);
4027 /** Push a page onto the top of the cursor's stack. */
4029 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4031 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4032 mc->mc_dbi, (void *) mc);
4034 if (mc->mc_snum >= CURSOR_STACK) {
4035 assert(mc->mc_snum < CURSOR_STACK);
4036 return MDB_CURSOR_FULL;
4039 mc->mc_top = mc->mc_snum++;
4040 mc->mc_pg[mc->mc_top] = mp;
4041 mc->mc_ki[mc->mc_top] = 0;
4046 /** Find the address of the page corresponding to a given page number.
4047 * @param[in] txn the transaction for this access.
4048 * @param[in] pgno the page number for the page to retrieve.
4049 * @param[out] ret address of a pointer where the page's address will be stored.
4050 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4051 * @return 0 on success, non-zero on failure.
4054 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4059 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4060 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4065 MDB_ID2L dl = tx2->mt_u.dirty_list;
4067 unsigned x = mdb_mid2l_search(dl, pgno);
4068 if (x <= dl[0].mid && dl[x].mid == pgno) {
4074 } while ((tx2 = tx2->mt_parent) != NULL);
4077 if (pgno < txn->mt_next_pgno) {
4079 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4081 DPRINTF("page %zu not found", pgno);
4083 return MDB_PAGE_NOTFOUND;
4093 /** Search for the page a given key should be in.
4094 * Pushes parent pages on the cursor stack. This function continues a
4095 * search on a cursor that has already been initialized. (Usually by
4096 * #mdb_page_search() but also by #mdb_node_move().)
4097 * @param[in,out] mc the cursor for this operation.
4098 * @param[in] key the key to search for. If NULL, search for the lowest
4099 * page. (This is used by #mdb_cursor_first().)
4100 * @param[in] modify If true, visited pages are updated with new page numbers.
4101 * @return 0 on success, non-zero on failure.
4104 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4106 MDB_page *mp = mc->mc_pg[mc->mc_top];
4111 while (IS_BRANCH(mp)) {
4115 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4116 assert(NUMKEYS(mp) > 1);
4117 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4119 if (key == NULL) /* Initialize cursor to first page. */
4121 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4122 /* cursor to last page */
4126 node = mdb_node_search(mc, key, &exact);
4128 i = NUMKEYS(mp) - 1;
4130 i = mc->mc_ki[mc->mc_top];
4139 DPRINTF("following index %u for key [%s]",
4141 assert(i < NUMKEYS(mp));
4142 node = NODEPTR(mp, i);
4144 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4147 mc->mc_ki[mc->mc_top] = i;
4148 if ((rc = mdb_cursor_push(mc, mp)))
4152 if ((rc = mdb_page_touch(mc)) != 0)
4154 mp = mc->mc_pg[mc->mc_top];
4159 DPRINTF("internal error, index points to a %02X page!?",
4161 return MDB_CORRUPTED;
4164 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4165 key ? DKEY(key) : NULL);
4170 /** Search for the lowest key under the current branch page.
4171 * This just bypasses a NUMKEYS check in the current page
4172 * before calling mdb_page_search_root(), because the callers
4173 * are all in situations where the current page is known to
4177 mdb_page_search_lowest(MDB_cursor *mc)
4179 MDB_page *mp = mc->mc_pg[mc->mc_top];
4180 MDB_node *node = NODEPTR(mp, 0);
4183 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4186 mc->mc_ki[mc->mc_top] = 0;
4187 if ((rc = mdb_cursor_push(mc, mp)))
4189 return mdb_page_search_root(mc, NULL, 0);
4192 /** Search for the page a given key should be in.
4193 * Pushes parent pages on the cursor stack. This function just sets up
4194 * the search; it finds the root page for \b mc's database and sets this
4195 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4196 * called to complete the search.
4197 * @param[in,out] mc the cursor for this operation.
4198 * @param[in] key the key to search for. If NULL, search for the lowest
4199 * page. (This is used by #mdb_cursor_first().)
4200 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4201 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4202 * @return 0 on success, non-zero on failure.
4205 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4210 /* Make sure the txn is still viable, then find the root from
4211 * the txn's db table.
4213 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4214 DPUTS("transaction has failed, must abort");
4217 /* Make sure we're using an up-to-date root */
4218 if (mc->mc_dbi > MAIN_DBI) {
4219 if ((*mc->mc_dbflag & DB_STALE) ||
4220 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4222 unsigned char dbflag = 0;
4223 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4224 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4227 if (*mc->mc_dbflag & DB_STALE) {
4231 MDB_node *leaf = mdb_node_search(&mc2,
4232 &mc->mc_dbx->md_name, &exact);
4234 return MDB_NOTFOUND;
4235 mdb_node_read(mc->mc_txn, leaf, &data);
4236 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4238 /* The txn may not know this DBI, or another process may
4239 * have dropped and recreated the DB with other flags.
4241 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4242 return MDB_INCOMPATIBLE;
4243 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4245 if (flags & MDB_PS_MODIFY)
4247 *mc->mc_dbflag &= ~DB_STALE;
4248 *mc->mc_dbflag |= dbflag;
4251 root = mc->mc_db->md_root;
4253 if (root == P_INVALID) { /* Tree is empty. */
4254 DPUTS("tree is empty");
4255 return MDB_NOTFOUND;
4260 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4261 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4267 DPRINTF("db %u root page %zu has flags 0x%X",
4268 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4270 if (flags & MDB_PS_MODIFY) {
4271 if ((rc = mdb_page_touch(mc)))
4275 if (flags & MDB_PS_ROOTONLY)
4278 return mdb_page_search_root(mc, key, flags);
4282 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4284 MDB_txn *txn = mc->mc_txn;
4285 pgno_t pg = mp->mp_pgno;
4286 unsigned i, ovpages = mp->mp_pages;
4289 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4290 /* If the page is dirty we just acquired it, so we should
4291 * give it back to our current free list, if any.
4292 * Not currently supported in nested txns.
4293 * Otherwise put it onto the list of pages we freed in this txn.
4295 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && txn->mt_env->me_pghead) {
4297 pgno_t *mop = txn->mt_env->me_pghead;
4298 MDB_ID2 *dl, ix, iy;
4299 /* Prepare to insert pg */
4300 j = mop[0] + ovpages;
4302 rc = mdb_midl_grow(&mop, ovpages);
4305 txn->mt_env->me_pghead = mop;
4307 /* Remove from dirty list */
4308 dl = txn->mt_u.dirty_list;
4310 for (ix = dl[x]; ix.mid != pg; ix = iy) {
4317 return MDB_CORRUPTED;
4320 /* Insert in me_pghead */
4321 for (i = mop[0]; i && mop[i] < pg; i--)
4327 for (i=0; i<ovpages; i++) {
4328 mdb_midl_append(&txn->mt_free_pgs, pg);
4332 mc->mc_db->md_overflow_pages -= ovpages;
4336 /** Return the data associated with a given node.
4337 * @param[in] txn The transaction for this operation.
4338 * @param[in] leaf The node being read.
4339 * @param[out] data Updated to point to the node's data.
4340 * @return 0 on success, non-zero on failure.
4343 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4345 MDB_page *omp; /* overflow page */
4349 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4350 data->mv_size = NODEDSZ(leaf);
4351 data->mv_data = NODEDATA(leaf);
4355 /* Read overflow data.
4357 data->mv_size = NODEDSZ(leaf);
4358 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4359 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4360 DPRINTF("read overflow page %zu failed", pgno);
4363 data->mv_data = METADATA(omp);
4369 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4370 MDB_val *key, MDB_val *data)
4379 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4381 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4384 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4388 mdb_cursor_init(&mc, txn, dbi, &mx);
4389 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4392 /** Find a sibling for a page.
4393 * Replaces the page at the top of the cursor's stack with the
4394 * specified sibling, if one exists.
4395 * @param[in] mc The cursor for this operation.
4396 * @param[in] move_right Non-zero if the right sibling is requested,
4397 * otherwise the left sibling.
4398 * @return 0 on success, non-zero on failure.
4401 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4407 if (mc->mc_snum < 2) {
4408 return MDB_NOTFOUND; /* root has no siblings */
4412 DPRINTF("parent page is page %zu, index %u",
4413 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4415 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4416 : (mc->mc_ki[mc->mc_top] == 0)) {
4417 DPRINTF("no more keys left, moving to %s sibling",
4418 move_right ? "right" : "left");
4419 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4420 /* undo cursor_pop before returning */
4427 mc->mc_ki[mc->mc_top]++;
4429 mc->mc_ki[mc->mc_top]--;
4430 DPRINTF("just moving to %s index key %u",
4431 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4433 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4435 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4436 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4439 mdb_cursor_push(mc, mp);
4441 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4446 /** Move the cursor to the next data item. */
4448 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4454 if (mc->mc_flags & C_EOF) {
4455 return MDB_NOTFOUND;
4458 assert(mc->mc_flags & C_INITIALIZED);
4460 mp = mc->mc_pg[mc->mc_top];
4462 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4463 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4464 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4465 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4466 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4467 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4471 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4472 if (op == MDB_NEXT_DUP)
4473 return MDB_NOTFOUND;
4477 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4479 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4480 DPUTS("=====> move to next sibling page");
4481 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4482 mc->mc_flags |= C_EOF;
4483 mc->mc_flags &= ~C_INITIALIZED;
4484 return MDB_NOTFOUND;
4486 mp = mc->mc_pg[mc->mc_top];
4487 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4489 mc->mc_ki[mc->mc_top]++;
4491 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4492 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4495 key->mv_size = mc->mc_db->md_pad;
4496 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4500 assert(IS_LEAF(mp));
4501 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4503 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4504 mdb_xcursor_init1(mc, leaf);
4507 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4510 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4511 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4512 if (rc != MDB_SUCCESS)
4517 MDB_GET_KEY(leaf, key);
4521 /** Move the cursor to the previous data item. */
4523 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4529 assert(mc->mc_flags & C_INITIALIZED);
4531 mp = mc->mc_pg[mc->mc_top];
4533 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4534 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4535 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4536 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4537 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4538 if (op != MDB_PREV || rc == MDB_SUCCESS)
4541 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4542 if (op == MDB_PREV_DUP)
4543 return MDB_NOTFOUND;
4548 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4550 if (mc->mc_ki[mc->mc_top] == 0) {
4551 DPUTS("=====> move to prev sibling page");
4552 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4553 mc->mc_flags &= ~C_INITIALIZED;
4554 return MDB_NOTFOUND;
4556 mp = mc->mc_pg[mc->mc_top];
4557 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4558 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4560 mc->mc_ki[mc->mc_top]--;
4562 mc->mc_flags &= ~C_EOF;
4564 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4565 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4568 key->mv_size = mc->mc_db->md_pad;
4569 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4573 assert(IS_LEAF(mp));
4574 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4576 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4577 mdb_xcursor_init1(mc, leaf);
4580 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4583 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4584 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4585 if (rc != MDB_SUCCESS)
4590 MDB_GET_KEY(leaf, key);
4594 /** Set the cursor on a specific data item. */
4596 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4597 MDB_cursor_op op, int *exactp)
4601 MDB_node *leaf = NULL;
4606 assert(key->mv_size > 0);
4608 /* See if we're already on the right page */
4609 if (mc->mc_flags & C_INITIALIZED) {
4612 mp = mc->mc_pg[mc->mc_top];
4614 mc->mc_ki[mc->mc_top] = 0;
4615 return MDB_NOTFOUND;
4617 if (mp->mp_flags & P_LEAF2) {
4618 nodekey.mv_size = mc->mc_db->md_pad;
4619 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4621 leaf = NODEPTR(mp, 0);
4622 MDB_GET_KEY(leaf, &nodekey);
4624 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4626 /* Probably happens rarely, but first node on the page
4627 * was the one we wanted.
4629 mc->mc_ki[mc->mc_top] = 0;
4636 unsigned int nkeys = NUMKEYS(mp);
4638 if (mp->mp_flags & P_LEAF2) {
4639 nodekey.mv_data = LEAF2KEY(mp,
4640 nkeys-1, nodekey.mv_size);
4642 leaf = NODEPTR(mp, nkeys-1);
4643 MDB_GET_KEY(leaf, &nodekey);
4645 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4647 /* last node was the one we wanted */
4648 mc->mc_ki[mc->mc_top] = nkeys-1;
4654 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4655 /* This is definitely the right page, skip search_page */
4656 if (mp->mp_flags & P_LEAF2) {
4657 nodekey.mv_data = LEAF2KEY(mp,
4658 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4660 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4661 MDB_GET_KEY(leaf, &nodekey);
4663 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4665 /* current node was the one we wanted */
4675 /* If any parents have right-sibs, search.
4676 * Otherwise, there's nothing further.
4678 for (i=0; i<mc->mc_top; i++)
4680 NUMKEYS(mc->mc_pg[i])-1)
4682 if (i == mc->mc_top) {
4683 /* There are no other pages */
4684 mc->mc_ki[mc->mc_top] = nkeys;
4685 return MDB_NOTFOUND;
4689 /* There are no other pages */
4690 mc->mc_ki[mc->mc_top] = 0;
4691 return MDB_NOTFOUND;
4695 rc = mdb_page_search(mc, key, 0);
4696 if (rc != MDB_SUCCESS)
4699 mp = mc->mc_pg[mc->mc_top];
4700 assert(IS_LEAF(mp));
4703 leaf = mdb_node_search(mc, key, exactp);
4704 if (exactp != NULL && !*exactp) {
4705 /* MDB_SET specified and not an exact match. */
4706 return MDB_NOTFOUND;
4710 DPUTS("===> inexact leaf not found, goto sibling");
4711 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4712 return rc; /* no entries matched */
4713 mp = mc->mc_pg[mc->mc_top];
4714 assert(IS_LEAF(mp));
4715 leaf = NODEPTR(mp, 0);
4719 mc->mc_flags |= C_INITIALIZED;
4720 mc->mc_flags &= ~C_EOF;
4723 key->mv_size = mc->mc_db->md_pad;
4724 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4728 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4729 mdb_xcursor_init1(mc, leaf);
4732 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4733 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4734 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4737 if (op == MDB_GET_BOTH) {
4743 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4744 if (rc != MDB_SUCCESS)
4747 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4749 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4751 rc = mc->mc_dbx->md_dcmp(data, &d2);
4753 if (op == MDB_GET_BOTH || rc > 0)
4754 return MDB_NOTFOUND;
4759 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4760 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4765 /* The key already matches in all other cases */
4766 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4767 MDB_GET_KEY(leaf, key);
4768 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4773 /** Move the cursor to the first item in the database. */
4775 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4780 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4781 rc = mdb_page_search(mc, NULL, 0);
4782 if (rc != MDB_SUCCESS)
4785 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4787 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4788 mc->mc_flags |= C_INITIALIZED;
4789 mc->mc_flags &= ~C_EOF;
4791 mc->mc_ki[mc->mc_top] = 0;
4793 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4794 key->mv_size = mc->mc_db->md_pad;
4795 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4800 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4801 mdb_xcursor_init1(mc, leaf);
4802 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4807 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4808 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4812 MDB_GET_KEY(leaf, key);
4816 /** Move the cursor to the last item in the database. */
4818 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4823 if (!(mc->mc_flags & C_EOF)) {
4825 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4828 lkey.mv_size = MDB_MAXKEYSIZE+1;
4829 lkey.mv_data = NULL;
4830 rc = mdb_page_search(mc, &lkey, 0);
4831 if (rc != MDB_SUCCESS)
4834 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4837 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4838 mc->mc_flags |= C_INITIALIZED|C_EOF;
4839 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4841 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4842 key->mv_size = mc->mc_db->md_pad;
4843 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4848 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4849 mdb_xcursor_init1(mc, leaf);
4850 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4855 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4856 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4861 MDB_GET_KEY(leaf, key);
4866 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4875 case MDB_GET_CURRENT:
4876 if (!(mc->mc_flags & C_INITIALIZED)) {
4879 MDB_page *mp = mc->mc_pg[mc->mc_top];
4881 mc->mc_ki[mc->mc_top] = 0;
4887 key->mv_size = mc->mc_db->md_pad;
4888 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4890 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4891 MDB_GET_KEY(leaf, key);
4893 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4894 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4896 rc = mdb_node_read(mc->mc_txn, leaf, data);
4903 case MDB_GET_BOTH_RANGE:
4904 if (data == NULL || mc->mc_xcursor == NULL) {
4912 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4914 } else if (op == MDB_SET_RANGE)
4915 rc = mdb_cursor_set(mc, key, data, op, NULL);
4917 rc = mdb_cursor_set(mc, key, data, op, &exact);
4919 case MDB_GET_MULTIPLE:
4921 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4922 !(mc->mc_flags & C_INITIALIZED)) {
4927 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4928 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4931 case MDB_NEXT_MULTIPLE:
4933 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4937 if (!(mc->mc_flags & C_INITIALIZED))
4938 rc = mdb_cursor_first(mc, key, data);
4940 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4941 if (rc == MDB_SUCCESS) {
4942 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4945 mx = &mc->mc_xcursor->mx_cursor;
4946 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4948 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4949 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4957 case MDB_NEXT_NODUP:
4958 if (!(mc->mc_flags & C_INITIALIZED))
4959 rc = mdb_cursor_first(mc, key, data);
4961 rc = mdb_cursor_next(mc, key, data, op);
4965 case MDB_PREV_NODUP:
4966 if (!(mc->mc_flags & C_INITIALIZED)) {
4967 rc = mdb_cursor_last(mc, key, data);
4970 mc->mc_flags |= C_INITIALIZED;
4971 mc->mc_ki[mc->mc_top]++;
4973 rc = mdb_cursor_prev(mc, key, data, op);
4976 rc = mdb_cursor_first(mc, key, data);
4980 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4981 !(mc->mc_flags & C_INITIALIZED) ||
4982 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4986 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4989 rc = mdb_cursor_last(mc, key, data);
4993 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4994 !(mc->mc_flags & C_INITIALIZED) ||
4995 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4999 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5002 DPRINTF("unhandled/unimplemented cursor operation %u", op);
5010 /** Touch all the pages in the cursor stack.
5011 * Makes sure all the pages are writable, before attempting a write operation.
5012 * @param[in] mc The cursor to operate on.
5015 mdb_cursor_touch(MDB_cursor *mc)
5019 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5022 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5023 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5026 *mc->mc_dbflag |= DB_DIRTY;
5028 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5029 rc = mdb_page_touch(mc);
5033 mc->mc_top = mc->mc_snum-1;
5038 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5041 MDB_node *leaf = NULL;
5042 MDB_val xdata, *rdata, dkey;
5045 int do_sub = 0, insert = 0;
5046 unsigned int mcount = 0;
5050 char dbuf[MDB_MAXKEYSIZE+1];
5051 unsigned int nflags;
5054 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5057 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5060 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5063 #if SIZE_MAX > MAXDATASIZE
5064 if (data->mv_size > MAXDATASIZE)
5068 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
5069 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5073 if (flags == MDB_CURRENT) {
5074 if (!(mc->mc_flags & C_INITIALIZED))
5077 } else if (mc->mc_db->md_root == P_INVALID) {
5079 /* new database, write a root leaf page */
5080 DPUTS("allocating new root leaf page");
5081 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5085 mdb_cursor_push(mc, np);
5086 mc->mc_db->md_root = np->mp_pgno;
5087 mc->mc_db->md_depth++;
5088 *mc->mc_dbflag |= DB_DIRTY;
5089 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5091 np->mp_flags |= P_LEAF2;
5092 mc->mc_flags |= C_INITIALIZED;
5098 if (flags & MDB_APPEND) {
5100 rc = mdb_cursor_last(mc, &k2, &d2);
5102 rc = mc->mc_dbx->md_cmp(key, &k2);
5105 mc->mc_ki[mc->mc_top]++;
5107 /* new key is <= last key */
5112 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5114 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5115 DPRINTF("duplicate key [%s]", DKEY(key));
5117 return MDB_KEYEXIST;
5119 if (rc && rc != MDB_NOTFOUND)
5123 /* Cursor is positioned, now make sure all pages are writable */
5124 rc2 = mdb_cursor_touch(mc);
5129 /* The key already exists */
5130 if (rc == MDB_SUCCESS) {
5131 /* there's only a key anyway, so this is a no-op */
5132 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5133 unsigned int ksize = mc->mc_db->md_pad;
5134 if (key->mv_size != ksize)
5136 if (flags == MDB_CURRENT) {
5137 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5138 memcpy(ptr, key->mv_data, ksize);
5143 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5146 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5147 /* Was a single item before, must convert now */
5149 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5150 /* Just overwrite the current item */
5151 if (flags == MDB_CURRENT)
5154 dkey.mv_size = NODEDSZ(leaf);
5155 dkey.mv_data = NODEDATA(leaf);
5156 #if UINT_MAX < SIZE_MAX
5157 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5158 #ifdef MISALIGNED_OK
5159 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5161 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5164 /* if data matches, ignore it */
5165 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5166 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5168 /* create a fake page for the dup items */
5169 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5170 dkey.mv_data = dbuf;
5171 fp = (MDB_page *)&pbuf;
5172 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5173 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5174 fp->mp_lower = PAGEHDRSZ;
5175 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5176 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5177 fp->mp_flags |= P_LEAF2;
5178 fp->mp_pad = data->mv_size;
5179 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5181 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5182 (dkey.mv_size & 1) + (data->mv_size & 1);
5184 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5187 xdata.mv_size = fp->mp_upper;
5192 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5193 /* See if we need to convert from fake page to subDB */
5195 unsigned int offset;
5198 fp = NODEDATA(leaf);
5199 if (flags == MDB_CURRENT) {
5201 fp->mp_flags |= P_DIRTY;
5202 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5203 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5207 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5208 offset = fp->mp_pad;
5209 if (SIZELEFT(fp) >= offset)
5211 offset *= 4; /* space for 4 more */
5213 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5215 offset += offset & 1;
5216 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5217 offset >= mc->mc_txn->mt_env->me_nodemax) {
5218 /* yes, convert it */
5220 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5221 dummy.md_pad = fp->mp_pad;
5222 dummy.md_flags = MDB_DUPFIXED;
5223 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5224 dummy.md_flags |= MDB_INTEGERKEY;
5227 dummy.md_branch_pages = 0;
5228 dummy.md_leaf_pages = 1;
5229 dummy.md_overflow_pages = 0;
5230 dummy.md_entries = NUMKEYS(fp);
5232 xdata.mv_size = sizeof(MDB_db);
5233 xdata.mv_data = &dummy;
5234 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5236 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5237 flags |= F_DUPDATA|F_SUBDATA;
5238 dummy.md_root = mp->mp_pgno;
5240 /* no, just grow it */
5242 xdata.mv_size = NODEDSZ(leaf) + offset;
5243 xdata.mv_data = &pbuf;
5244 mp = (MDB_page *)&pbuf;
5245 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5248 mp->mp_flags = fp->mp_flags | P_DIRTY;
5249 mp->mp_pad = fp->mp_pad;
5250 mp->mp_lower = fp->mp_lower;
5251 mp->mp_upper = fp->mp_upper + offset;
5253 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5255 nsize = NODEDSZ(leaf) - fp->mp_upper;
5256 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5257 for (i=0; i<NUMKEYS(fp); i++)
5258 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5260 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5264 /* data is on sub-DB, just store it */
5265 flags |= F_DUPDATA|F_SUBDATA;
5269 /* overflow page overwrites need special handling */
5270 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5273 unsigned psize = mc->mc_txn->mt_env->me_psize;
5274 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5276 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5277 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5279 ovpages = omp->mp_pages;
5281 /* Is the ov page writable and large enough? */
5282 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5283 /* yes, overwrite it. Note in this case we don't
5284 * bother to try shrinking the page if the new data
5285 * is smaller than the overflow threshold.
5288 /* It is writable only in a parent txn */
5289 size_t sz = (size_t) psize * ovpages, off;
5290 MDB_page *np = mdb_page_malloc(mc, ovpages);
5296 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5297 if (!(flags & MDB_RESERVE)) {
5298 /* Copy end of page, adjusting alignment so
5299 * compiler may copy words instead of bytes.
5301 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5302 memcpy((size_t *)((char *)np + off),
5303 (size_t *)((char *)omp + off), sz - off);
5306 memcpy(np, omp, sz); /* Copy beginning of page */
5309 SETDSZ(leaf, data->mv_size);
5310 if (F_ISSET(flags, MDB_RESERVE))
5311 data->mv_data = METADATA(omp);
5313 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5316 mdb_ovpage_free(mc, omp);
5318 } else if (NODEDSZ(leaf) == data->mv_size) {
5319 /* same size, just replace it. Note that we could
5320 * also reuse this node if the new data is smaller,
5321 * but instead we opt to shrink the node in that case.
5323 if (F_ISSET(flags, MDB_RESERVE))
5324 data->mv_data = NODEDATA(leaf);
5325 else if (data->mv_size)
5326 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5328 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5331 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5332 mc->mc_db->md_entries--;
5334 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5341 nflags = flags & NODE_ADD_FLAGS;
5342 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5343 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5344 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5345 nflags &= ~MDB_APPEND;
5347 nflags |= MDB_SPLIT_REPLACE;
5348 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5350 /* There is room already in this leaf page. */
5351 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5352 if (rc == 0 && !do_sub && insert) {
5353 /* Adjust other cursors pointing to mp */
5354 MDB_cursor *m2, *m3;
5355 MDB_dbi dbi = mc->mc_dbi;
5356 unsigned i = mc->mc_top;
5357 MDB_page *mp = mc->mc_pg[i];
5359 if (mc->mc_flags & C_SUB)
5362 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5363 if (mc->mc_flags & C_SUB)
5364 m3 = &m2->mc_xcursor->mx_cursor;
5367 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5368 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5375 if (rc != MDB_SUCCESS)
5376 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5378 /* Now store the actual data in the child DB. Note that we're
5379 * storing the user data in the keys field, so there are strict
5380 * size limits on dupdata. The actual data fields of the child
5381 * DB are all zero size.
5388 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5389 if (flags & MDB_CURRENT) {
5390 xflags = MDB_CURRENT;
5392 mdb_xcursor_init1(mc, leaf);
5393 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5395 /* converted, write the original data first */
5397 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5401 /* Adjust other cursors pointing to mp */
5403 unsigned i = mc->mc_top;
5404 MDB_page *mp = mc->mc_pg[i];
5406 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5407 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5408 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5409 mdb_xcursor_init1(m2, leaf);
5413 /* we've done our job */
5416 if (flags & MDB_APPENDDUP)
5417 xflags |= MDB_APPEND;
5418 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5419 if (flags & F_SUBDATA) {
5420 void *db = NODEDATA(leaf);
5421 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5424 /* sub-writes might have failed so check rc again.
5425 * Don't increment count if we just replaced an existing item.
5427 if (!rc && !(flags & MDB_CURRENT))
5428 mc->mc_db->md_entries++;
5429 if (flags & MDB_MULTIPLE) {
5431 if (mcount < data[1].mv_size) {
5432 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5433 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5439 /* If we succeeded and the key didn't exist before, make sure
5440 * the cursor is marked valid.
5443 mc->mc_flags |= C_INITIALIZED;
5448 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5453 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5456 if (!(mc->mc_flags & C_INITIALIZED))
5459 rc = mdb_cursor_touch(mc);
5463 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5465 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5466 if (flags != MDB_NODUPDATA) {
5467 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5468 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5470 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5471 /* If sub-DB still has entries, we're done */
5472 if (mc->mc_xcursor->mx_db.md_entries) {
5473 if (leaf->mn_flags & F_SUBDATA) {
5474 /* update subDB info */
5475 void *db = NODEDATA(leaf);
5476 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5479 /* shrink fake page */
5480 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5481 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5482 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5483 /* fix other sub-DB cursors pointed at this fake page */
5484 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5485 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5486 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5487 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5488 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5491 mc->mc_db->md_entries--;
5494 /* otherwise fall thru and delete the sub-DB */
5497 if (leaf->mn_flags & F_SUBDATA) {
5498 /* add all the child DB's pages to the free list */
5499 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5500 if (rc == MDB_SUCCESS) {
5501 mc->mc_db->md_entries -=
5502 mc->mc_xcursor->mx_db.md_entries;
5507 return mdb_cursor_del0(mc, leaf);
5510 /** Allocate and initialize new pages for a database.
5511 * @param[in] mc a cursor on the database being added to.
5512 * @param[in] flags flags defining what type of page is being allocated.
5513 * @param[in] num the number of pages to allocate. This is usually 1,
5514 * unless allocating overflow pages for a large record.
5515 * @param[out] mp Address of a page, or NULL on failure.
5516 * @return 0 on success, non-zero on failure.
5519 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5524 if ((rc = mdb_page_alloc(mc, num, &np)))
5526 DPRINTF("allocated new mpage %zu, page size %u",
5527 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5528 np->mp_flags = flags | P_DIRTY;
5529 np->mp_lower = PAGEHDRSZ;
5530 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5533 mc->mc_db->md_branch_pages++;
5534 else if (IS_LEAF(np))
5535 mc->mc_db->md_leaf_pages++;
5536 else if (IS_OVERFLOW(np)) {
5537 mc->mc_db->md_overflow_pages += num;
5545 /** Calculate the size of a leaf node.
5546 * The size depends on the environment's page size; if a data item
5547 * is too large it will be put onto an overflow page and the node
5548 * size will only include the key and not the data. Sizes are always
5549 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5550 * of the #MDB_node headers.
5551 * @param[in] env The environment handle.
5552 * @param[in] key The key for the node.
5553 * @param[in] data The data for the node.
5554 * @return The number of bytes needed to store the node.
5557 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5561 sz = LEAFSIZE(key, data);
5562 if (sz >= env->me_nodemax) {
5563 /* put on overflow page */
5564 sz -= data->mv_size - sizeof(pgno_t);
5568 return sz + sizeof(indx_t);
5571 /** Calculate the size of a branch node.
5572 * The size should depend on the environment's page size but since
5573 * we currently don't support spilling large keys onto overflow
5574 * pages, it's simply the size of the #MDB_node header plus the
5575 * size of the key. Sizes are always rounded up to an even number
5576 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5577 * @param[in] env The environment handle.
5578 * @param[in] key The key for the node.
5579 * @return The number of bytes needed to store the node.
5582 mdb_branch_size(MDB_env *env, MDB_val *key)
5587 if (sz >= env->me_nodemax) {
5588 /* put on overflow page */
5589 /* not implemented */
5590 /* sz -= key->size - sizeof(pgno_t); */
5593 return sz + sizeof(indx_t);
5596 /** Add a node to the page pointed to by the cursor.
5597 * @param[in] mc The cursor for this operation.
5598 * @param[in] indx The index on the page where the new node should be added.
5599 * @param[in] key The key for the new node.
5600 * @param[in] data The data for the new node, if any.
5601 * @param[in] pgno The page number, if adding a branch node.
5602 * @param[in] flags Flags for the node.
5603 * @return 0 on success, non-zero on failure. Possible errors are:
5605 * <li>ENOMEM - failed to allocate overflow pages for the node.
5606 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5607 * should never happen since all callers already calculate the
5608 * page's free space before calling this function.
5612 mdb_node_add(MDB_cursor *mc, indx_t indx,
5613 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5616 size_t node_size = NODESIZE;
5619 MDB_page *mp = mc->mc_pg[mc->mc_top];
5620 MDB_page *ofp = NULL; /* overflow page */
5623 assert(mp->mp_upper >= mp->mp_lower);
5625 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5626 IS_LEAF(mp) ? "leaf" : "branch",
5627 IS_SUBP(mp) ? "sub-" : "",
5628 mp->mp_pgno, indx, data ? data->mv_size : 0,
5629 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5632 /* Move higher keys up one slot. */
5633 int ksize = mc->mc_db->md_pad, dif;
5634 char *ptr = LEAF2KEY(mp, indx, ksize);
5635 dif = NUMKEYS(mp) - indx;
5637 memmove(ptr+ksize, ptr, dif*ksize);
5638 /* insert new key */
5639 memcpy(ptr, key->mv_data, ksize);
5641 /* Just using these for counting */
5642 mp->mp_lower += sizeof(indx_t);
5643 mp->mp_upper -= ksize - sizeof(indx_t);
5648 node_size += key->mv_size;
5652 if (F_ISSET(flags, F_BIGDATA)) {
5653 /* Data already on overflow page. */
5654 node_size += sizeof(pgno_t);
5655 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5656 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5658 /* Put data on overflow page. */
5659 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5660 data->mv_size, node_size+data->mv_size);
5661 node_size += sizeof(pgno_t);
5662 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5664 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5667 node_size += data->mv_size;
5670 node_size += node_size & 1;
5672 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5673 DPRINTF("not enough room in page %zu, got %u ptrs",
5674 mp->mp_pgno, NUMKEYS(mp));
5675 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5676 mp->mp_upper - mp->mp_lower);
5677 DPRINTF("node size = %zu", node_size);
5678 return MDB_PAGE_FULL;
5681 /* Move higher pointers up one slot. */
5682 for (i = NUMKEYS(mp); i > indx; i--)
5683 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5685 /* Adjust free space offsets. */
5686 ofs = mp->mp_upper - node_size;
5687 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5688 mp->mp_ptrs[indx] = ofs;
5690 mp->mp_lower += sizeof(indx_t);
5692 /* Write the node data. */
5693 node = NODEPTR(mp, indx);
5694 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5695 node->mn_flags = flags;
5697 SETDSZ(node,data->mv_size);
5702 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5707 if (F_ISSET(flags, F_BIGDATA))
5708 memcpy(node->mn_data + key->mv_size, data->mv_data,
5710 else if (F_ISSET(flags, MDB_RESERVE))
5711 data->mv_data = node->mn_data + key->mv_size;
5713 memcpy(node->mn_data + key->mv_size, data->mv_data,
5716 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5718 if (F_ISSET(flags, MDB_RESERVE))
5719 data->mv_data = METADATA(ofp);
5721 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5728 /** Delete the specified node from a page.
5729 * @param[in] mp The page to operate on.
5730 * @param[in] indx The index of the node to delete.
5731 * @param[in] ksize The size of a node. Only used if the page is
5732 * part of a #MDB_DUPFIXED database.
5735 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5738 indx_t i, j, numkeys, ptr;
5745 COPY_PGNO(pgno, mp->mp_pgno);
5746 DPRINTF("delete node %u on %s page %zu", indx,
5747 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5750 assert(indx < NUMKEYS(mp));
5753 int x = NUMKEYS(mp) - 1 - indx;
5754 base = LEAF2KEY(mp, indx, ksize);
5756 memmove(base, base + ksize, x * ksize);
5757 mp->mp_lower -= sizeof(indx_t);
5758 mp->mp_upper += ksize - sizeof(indx_t);
5762 node = NODEPTR(mp, indx);
5763 sz = NODESIZE + node->mn_ksize;
5765 if (F_ISSET(node->mn_flags, F_BIGDATA))
5766 sz += sizeof(pgno_t);
5768 sz += NODEDSZ(node);
5772 ptr = mp->mp_ptrs[indx];
5773 numkeys = NUMKEYS(mp);
5774 for (i = j = 0; i < numkeys; i++) {
5776 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5777 if (mp->mp_ptrs[i] < ptr)
5778 mp->mp_ptrs[j] += sz;
5783 base = (char *)mp + mp->mp_upper;
5784 memmove(base + sz, base, ptr - mp->mp_upper);
5786 mp->mp_lower -= sizeof(indx_t);
5790 /** Compact the main page after deleting a node on a subpage.
5791 * @param[in] mp The main page to operate on.
5792 * @param[in] indx The index of the subpage on the main page.
5795 mdb_node_shrink(MDB_page *mp, indx_t indx)
5802 indx_t i, numkeys, ptr;
5804 node = NODEPTR(mp, indx);
5805 sp = (MDB_page *)NODEDATA(node);
5806 osize = NODEDSZ(node);
5808 delta = sp->mp_upper - sp->mp_lower;
5809 SETDSZ(node, osize - delta);
5810 xp = (MDB_page *)((char *)sp + delta);
5812 /* shift subpage upward */
5814 nsize = NUMKEYS(sp) * sp->mp_pad;
5815 memmove(METADATA(xp), METADATA(sp), nsize);
5818 nsize = osize - sp->mp_upper;
5819 numkeys = NUMKEYS(sp);
5820 for (i=numkeys-1; i>=0; i--)
5821 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5823 xp->mp_upper = sp->mp_lower;
5824 xp->mp_lower = sp->mp_lower;
5825 xp->mp_flags = sp->mp_flags;
5826 xp->mp_pad = sp->mp_pad;
5827 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5829 /* shift lower nodes upward */
5830 ptr = mp->mp_ptrs[indx];
5831 numkeys = NUMKEYS(mp);
5832 for (i = 0; i < numkeys; i++) {
5833 if (mp->mp_ptrs[i] <= ptr)
5834 mp->mp_ptrs[i] += delta;
5837 base = (char *)mp + mp->mp_upper;
5838 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5839 mp->mp_upper += delta;
5842 /** Initial setup of a sorted-dups cursor.
5843 * Sorted duplicates are implemented as a sub-database for the given key.
5844 * The duplicate data items are actually keys of the sub-database.
5845 * Operations on the duplicate data items are performed using a sub-cursor
5846 * initialized when the sub-database is first accessed. This function does
5847 * the preliminary setup of the sub-cursor, filling in the fields that
5848 * depend only on the parent DB.
5849 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5852 mdb_xcursor_init0(MDB_cursor *mc)
5854 MDB_xcursor *mx = mc->mc_xcursor;
5856 mx->mx_cursor.mc_xcursor = NULL;
5857 mx->mx_cursor.mc_txn = mc->mc_txn;
5858 mx->mx_cursor.mc_db = &mx->mx_db;
5859 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5860 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5861 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5862 mx->mx_cursor.mc_snum = 0;
5863 mx->mx_cursor.mc_top = 0;
5864 mx->mx_cursor.mc_flags = C_SUB;
5865 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5866 mx->mx_dbx.md_dcmp = NULL;
5867 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5870 /** Final setup of a sorted-dups cursor.
5871 * Sets up the fields that depend on the data from the main cursor.
5872 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5873 * @param[in] node The data containing the #MDB_db record for the
5874 * sorted-dup database.
5877 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5879 MDB_xcursor *mx = mc->mc_xcursor;
5881 if (node->mn_flags & F_SUBDATA) {
5882 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5883 mx->mx_cursor.mc_pg[0] = 0;
5884 mx->mx_cursor.mc_snum = 0;
5885 mx->mx_cursor.mc_flags = C_SUB;
5887 MDB_page *fp = NODEDATA(node);
5888 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5889 mx->mx_db.md_flags = 0;
5890 mx->mx_db.md_depth = 1;
5891 mx->mx_db.md_branch_pages = 0;
5892 mx->mx_db.md_leaf_pages = 1;
5893 mx->mx_db.md_overflow_pages = 0;
5894 mx->mx_db.md_entries = NUMKEYS(fp);
5895 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5896 mx->mx_cursor.mc_snum = 1;
5897 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5898 mx->mx_cursor.mc_top = 0;
5899 mx->mx_cursor.mc_pg[0] = fp;
5900 mx->mx_cursor.mc_ki[0] = 0;
5901 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5902 mx->mx_db.md_flags = MDB_DUPFIXED;
5903 mx->mx_db.md_pad = fp->mp_pad;
5904 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5905 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5908 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5910 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5912 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5913 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5914 #if UINT_MAX < SIZE_MAX
5915 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5916 #ifdef MISALIGNED_OK
5917 mx->mx_dbx.md_cmp = mdb_cmp_long;
5919 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5924 /** Initialize a cursor for a given transaction and database. */
5926 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5931 mc->mc_db = &txn->mt_dbs[dbi];
5932 mc->mc_dbx = &txn->mt_dbxs[dbi];
5933 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5938 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5940 mc->mc_xcursor = mx;
5941 mdb_xcursor_init0(mc);
5943 mc->mc_xcursor = NULL;
5945 if (*mc->mc_dbflag & DB_STALE) {
5946 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5951 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5954 size_t size = sizeof(MDB_cursor);
5956 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5959 /* Allow read access to the freelist */
5960 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5963 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5964 size += sizeof(MDB_xcursor);
5966 if ((mc = malloc(size)) != NULL) {
5967 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5968 if (txn->mt_cursors) {
5969 mc->mc_next = txn->mt_cursors[dbi];
5970 txn->mt_cursors[dbi] = mc;
5971 mc->mc_flags |= C_UNTRACK;
5973 mc->mc_flags |= C_ALLOCD;
5984 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5988 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5991 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5994 flags = mc->mc_flags;
5996 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5998 mc->mc_flags |= (flags & C_ALLOCD);
6002 /* Return the count of duplicate data items for the current key */
6004 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6008 if (mc == NULL || countp == NULL)
6011 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
6014 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6015 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6018 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6021 *countp = mc->mc_xcursor->mx_db.md_entries;
6027 mdb_cursor_close(MDB_cursor *mc)
6030 /* remove from txn, if tracked */
6031 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6032 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6033 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6035 *prev = mc->mc_next;
6037 if (mc->mc_flags & C_ALLOCD)
6043 mdb_cursor_txn(MDB_cursor *mc)
6045 if (!mc) return NULL;
6050 mdb_cursor_dbi(MDB_cursor *mc)
6056 /** Replace the key for a node with a new key.
6057 * @param[in] mc Cursor pointing to the node to operate on.
6058 * @param[in] key The new key to use.
6059 * @return 0 on success, non-zero on failure.
6062 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6069 indx_t ptr, i, numkeys, indx;
6072 indx = mc->mc_ki[mc->mc_top];
6073 mp = mc->mc_pg[mc->mc_top];
6074 node = NODEPTR(mp, indx);
6075 ptr = mp->mp_ptrs[indx];
6079 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6080 k2.mv_data = NODEKEY(node);
6081 k2.mv_size = node->mn_ksize;
6082 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6084 mdb_dkey(&k2, kbuf2),
6090 delta0 = delta = key->mv_size - node->mn_ksize;
6092 /* Must be 2-byte aligned. If new key is
6093 * shorter by 1, the shift will be skipped.
6095 delta += (delta & 1);
6097 if (delta > 0 && SIZELEFT(mp) < delta) {
6099 /* not enough space left, do a delete and split */
6100 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6101 pgno = NODEPGNO(node);
6102 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6103 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6106 numkeys = NUMKEYS(mp);
6107 for (i = 0; i < numkeys; i++) {
6108 if (mp->mp_ptrs[i] <= ptr)
6109 mp->mp_ptrs[i] -= delta;
6112 base = (char *)mp + mp->mp_upper;
6113 len = ptr - mp->mp_upper + NODESIZE;
6114 memmove(base - delta, base, len);
6115 mp->mp_upper -= delta;
6117 node = NODEPTR(mp, indx);
6120 /* But even if no shift was needed, update ksize */
6122 node->mn_ksize = key->mv_size;
6125 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6131 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6133 /** Move a node from csrc to cdst.
6136 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6143 unsigned short flags;
6147 /* Mark src and dst as dirty. */
6148 if ((rc = mdb_page_touch(csrc)) ||
6149 (rc = mdb_page_touch(cdst)))
6152 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6153 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6154 key.mv_size = csrc->mc_db->md_pad;
6155 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6157 data.mv_data = NULL;
6161 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6162 assert(!((long)srcnode&1));
6163 srcpg = NODEPGNO(srcnode);
6164 flags = srcnode->mn_flags;
6165 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6166 unsigned int snum = csrc->mc_snum;
6168 /* must find the lowest key below src */
6169 mdb_page_search_lowest(csrc);
6170 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6171 key.mv_size = csrc->mc_db->md_pad;
6172 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6174 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6175 key.mv_size = NODEKSZ(s2);
6176 key.mv_data = NODEKEY(s2);
6178 csrc->mc_snum = snum--;
6179 csrc->mc_top = snum;
6181 key.mv_size = NODEKSZ(srcnode);
6182 key.mv_data = NODEKEY(srcnode);
6184 data.mv_size = NODEDSZ(srcnode);
6185 data.mv_data = NODEDATA(srcnode);
6187 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6188 unsigned int snum = cdst->mc_snum;
6191 /* must find the lowest key below dst */
6192 mdb_page_search_lowest(cdst);
6193 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6194 bkey.mv_size = cdst->mc_db->md_pad;
6195 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6197 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6198 bkey.mv_size = NODEKSZ(s2);
6199 bkey.mv_data = NODEKEY(s2);
6201 cdst->mc_snum = snum--;
6202 cdst->mc_top = snum;
6203 mdb_cursor_copy(cdst, &mn);
6205 rc = mdb_update_key(&mn, &bkey);
6210 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6211 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6212 csrc->mc_ki[csrc->mc_top],
6214 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6215 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6217 /* Add the node to the destination page.
6219 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6220 if (rc != MDB_SUCCESS)
6223 /* Delete the node from the source page.
6225 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6228 /* Adjust other cursors pointing to mp */
6229 MDB_cursor *m2, *m3;
6230 MDB_dbi dbi = csrc->mc_dbi;
6231 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6233 if (csrc->mc_flags & C_SUB)
6236 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6237 if (m2 == csrc) continue;
6238 if (csrc->mc_flags & C_SUB)
6239 m3 = &m2->mc_xcursor->mx_cursor;
6242 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6243 csrc->mc_ki[csrc->mc_top]) {
6244 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6245 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6250 /* Update the parent separators.
6252 if (csrc->mc_ki[csrc->mc_top] == 0) {
6253 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6254 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6255 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6257 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6258 key.mv_size = NODEKSZ(srcnode);
6259 key.mv_data = NODEKEY(srcnode);
6261 DPRINTF("update separator for source page %zu to [%s]",
6262 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6263 mdb_cursor_copy(csrc, &mn);
6266 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6269 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6271 indx_t ix = csrc->mc_ki[csrc->mc_top];
6272 nullkey.mv_size = 0;
6273 csrc->mc_ki[csrc->mc_top] = 0;
6274 rc = mdb_update_key(csrc, &nullkey);
6275 csrc->mc_ki[csrc->mc_top] = ix;
6276 assert(rc == MDB_SUCCESS);
6280 if (cdst->mc_ki[cdst->mc_top] == 0) {
6281 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6282 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6283 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6285 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6286 key.mv_size = NODEKSZ(srcnode);
6287 key.mv_data = NODEKEY(srcnode);
6289 DPRINTF("update separator for destination page %zu to [%s]",
6290 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6291 mdb_cursor_copy(cdst, &mn);
6294 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6297 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6299 indx_t ix = cdst->mc_ki[cdst->mc_top];
6300 nullkey.mv_size = 0;
6301 cdst->mc_ki[cdst->mc_top] = 0;
6302 rc = mdb_update_key(cdst, &nullkey);
6303 cdst->mc_ki[cdst->mc_top] = ix;
6304 assert(rc == MDB_SUCCESS);
6311 /** Merge one page into another.
6312 * The nodes from the page pointed to by \b csrc will
6313 * be copied to the page pointed to by \b cdst and then
6314 * the \b csrc page will be freed.
6315 * @param[in] csrc Cursor pointing to the source page.
6316 * @param[in] cdst Cursor pointing to the destination page.
6319 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6327 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6328 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6330 assert(csrc->mc_snum > 1); /* can't merge root page */
6331 assert(cdst->mc_snum > 1);
6333 /* Mark dst as dirty. */
6334 if ((rc = mdb_page_touch(cdst)))
6337 /* Move all nodes from src to dst.
6339 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6340 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6341 key.mv_size = csrc->mc_db->md_pad;
6342 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6343 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6344 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6345 if (rc != MDB_SUCCESS)
6347 key.mv_data = (char *)key.mv_data + key.mv_size;
6350 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6351 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6352 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6353 unsigned int snum = csrc->mc_snum;
6355 /* must find the lowest key below src */
6356 mdb_page_search_lowest(csrc);
6357 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6358 key.mv_size = csrc->mc_db->md_pad;
6359 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6361 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6362 key.mv_size = NODEKSZ(s2);
6363 key.mv_data = NODEKEY(s2);
6365 csrc->mc_snum = snum--;
6366 csrc->mc_top = snum;
6368 key.mv_size = srcnode->mn_ksize;
6369 key.mv_data = NODEKEY(srcnode);
6372 data.mv_size = NODEDSZ(srcnode);
6373 data.mv_data = NODEDATA(srcnode);
6374 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6375 if (rc != MDB_SUCCESS)
6380 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6381 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);
6383 /* Unlink the src page from parent and add to free list.
6385 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6386 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6389 rc = mdb_update_key(csrc, &key);
6395 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6396 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6397 csrc->mc_db->md_leaf_pages--;
6399 csrc->mc_db->md_branch_pages--;
6401 /* Adjust other cursors pointing to mp */
6402 MDB_cursor *m2, *m3;
6403 MDB_dbi dbi = csrc->mc_dbi;
6404 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6406 if (csrc->mc_flags & C_SUB)
6409 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6410 if (csrc->mc_flags & C_SUB)
6411 m3 = &m2->mc_xcursor->mx_cursor;
6414 if (m3 == csrc) continue;
6415 if (m3->mc_snum < csrc->mc_snum) continue;
6416 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6417 m3->mc_pg[csrc->mc_top] = mp;
6418 m3->mc_ki[csrc->mc_top] += nkeys;
6422 mdb_cursor_pop(csrc);
6424 return mdb_rebalance(csrc);
6427 /** Copy the contents of a cursor.
6428 * @param[in] csrc The cursor to copy from.
6429 * @param[out] cdst The cursor to copy to.
6432 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6436 cdst->mc_txn = csrc->mc_txn;
6437 cdst->mc_dbi = csrc->mc_dbi;
6438 cdst->mc_db = csrc->mc_db;
6439 cdst->mc_dbx = csrc->mc_dbx;
6440 cdst->mc_snum = csrc->mc_snum;
6441 cdst->mc_top = csrc->mc_top;
6442 cdst->mc_flags = csrc->mc_flags;
6444 for (i=0; i<csrc->mc_snum; i++) {
6445 cdst->mc_pg[i] = csrc->mc_pg[i];
6446 cdst->mc_ki[i] = csrc->mc_ki[i];
6450 /** Rebalance the tree after a delete operation.
6451 * @param[in] mc Cursor pointing to the page where rebalancing
6453 * @return 0 on success, non-zero on failure.
6456 mdb_rebalance(MDB_cursor *mc)
6460 unsigned int ptop, minkeys;
6463 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6467 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6468 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6469 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6470 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6474 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6475 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6478 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6479 DPRINTF("no need to rebalance page %zu, above fill threshold",
6485 if (mc->mc_snum < 2) {
6486 MDB_page *mp = mc->mc_pg[0];
6488 DPUTS("Can't rebalance a subpage, ignoring");
6491 if (NUMKEYS(mp) == 0) {
6492 DPUTS("tree is completely empty");
6493 mc->mc_db->md_root = P_INVALID;
6494 mc->mc_db->md_depth = 0;
6495 mc->mc_db->md_leaf_pages = 0;
6496 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6500 /* Adjust other cursors pointing to mp */
6501 MDB_cursor *m2, *m3;
6502 MDB_dbi dbi = mc->mc_dbi;
6504 if (mc->mc_flags & C_SUB)
6507 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6508 if (m2 == mc) continue;
6509 if (mc->mc_flags & C_SUB)
6510 m3 = &m2->mc_xcursor->mx_cursor;
6513 if (m3->mc_snum < mc->mc_snum) continue;
6514 if (m3->mc_pg[0] == mp) {
6520 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6521 DPUTS("collapsing root page!");
6522 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6523 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6524 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6527 mc->mc_db->md_depth--;
6528 mc->mc_db->md_branch_pages--;
6530 /* Adjust other cursors pointing to mp */
6531 MDB_cursor *m2, *m3;
6532 MDB_dbi dbi = mc->mc_dbi;
6534 if (mc->mc_flags & C_SUB)
6537 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6538 if (m2 == mc) continue;
6539 if (mc->mc_flags & C_SUB)
6540 m3 = &m2->mc_xcursor->mx_cursor;
6543 if (m3->mc_snum < mc->mc_snum) continue;
6544 if (m3->mc_pg[0] == mp) {
6545 m3->mc_pg[0] = mc->mc_pg[0];
6552 DPUTS("root page doesn't need rebalancing");
6556 /* The parent (branch page) must have at least 2 pointers,
6557 * otherwise the tree is invalid.
6559 ptop = mc->mc_top-1;
6560 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6562 /* Leaf page fill factor is below the threshold.
6563 * Try to move keys from left or right neighbor, or
6564 * merge with a neighbor page.
6569 mdb_cursor_copy(mc, &mn);
6570 mn.mc_xcursor = NULL;
6572 if (mc->mc_ki[ptop] == 0) {
6573 /* We're the leftmost leaf in our parent.
6575 DPUTS("reading right neighbor");
6577 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6578 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6581 mn.mc_ki[mn.mc_top] = 0;
6582 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6584 /* There is at least one neighbor to the left.
6586 DPUTS("reading left neighbor");
6588 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6589 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6592 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6593 mc->mc_ki[mc->mc_top] = 0;
6596 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6597 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);
6599 /* If the neighbor page is above threshold and has enough keys,
6600 * move one key from it. Otherwise we should try to merge them.
6601 * (A branch page must never have less than 2 keys.)
6603 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6604 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6605 return mdb_node_move(&mn, mc);
6607 if (mc->mc_ki[ptop] == 0)
6608 rc = mdb_page_merge(&mn, mc);
6610 rc = mdb_page_merge(mc, &mn);
6611 mc->mc_flags &= ~C_INITIALIZED;
6616 /** Complete a delete operation started by #mdb_cursor_del(). */
6618 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6622 /* add overflow pages to free list */
6623 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6627 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6628 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
6630 assert(IS_OVERFLOW(omp));
6631 mdb_ovpage_free(mc, omp);
6633 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6634 mc->mc_db->md_entries--;
6635 rc = mdb_rebalance(mc);
6636 if (rc != MDB_SUCCESS)
6637 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6638 /* if mc points past last node in page, invalidate */
6639 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6640 mc->mc_flags &= ~C_INITIALIZED;
6646 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6647 MDB_val *key, MDB_val *data)
6652 MDB_val rdata, *xdata;
6656 assert(key != NULL);
6658 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6660 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6663 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6667 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6671 mdb_cursor_init(&mc, txn, dbi, &mx);
6682 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6684 /* let mdb_page_split know about this cursor if needed:
6685 * delete will trigger a rebalance; if it needs to move
6686 * a node from one page to another, it will have to
6687 * update the parent's separator key(s). If the new sepkey
6688 * is larger than the current one, the parent page may
6689 * run out of space, triggering a split. We need this
6690 * cursor to be consistent until the end of the rebalance.
6692 mc.mc_next = txn->mt_cursors[dbi];
6693 txn->mt_cursors[dbi] = &mc;
6694 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6695 txn->mt_cursors[dbi] = mc.mc_next;
6700 /** Split a page and insert a new node.
6701 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6702 * The cursor will be updated to point to the actual page and index where
6703 * the node got inserted after the split.
6704 * @param[in] newkey The key for the newly inserted node.
6705 * @param[in] newdata The data for the newly inserted node.
6706 * @param[in] newpgno The page number, if the new node is a branch node.
6707 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6708 * @return 0 on success, non-zero on failure.
6711 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6712 unsigned int nflags)
6715 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6718 unsigned int i, j, split_indx, nkeys, pmax;
6720 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6722 MDB_page *mp, *rp, *pp;
6727 mp = mc->mc_pg[mc->mc_top];
6728 newindx = mc->mc_ki[mc->mc_top];
6730 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6731 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6732 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6734 /* Create a right sibling. */
6735 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6737 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6739 if (mc->mc_snum < 2) {
6740 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6742 /* shift current top to make room for new parent */
6743 mc->mc_pg[1] = mc->mc_pg[0];
6744 mc->mc_ki[1] = mc->mc_ki[0];
6747 mc->mc_db->md_root = pp->mp_pgno;
6748 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6749 mc->mc_db->md_depth++;
6752 /* Add left (implicit) pointer. */
6753 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6754 /* undo the pre-push */
6755 mc->mc_pg[0] = mc->mc_pg[1];
6756 mc->mc_ki[0] = mc->mc_ki[1];
6757 mc->mc_db->md_root = mp->mp_pgno;
6758 mc->mc_db->md_depth--;
6765 ptop = mc->mc_top-1;
6766 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6769 mc->mc_flags |= C_SPLITTING;
6770 mdb_cursor_copy(mc, &mn);
6771 mn.mc_pg[mn.mc_top] = rp;
6772 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6774 if (nflags & MDB_APPEND) {
6775 mn.mc_ki[mn.mc_top] = 0;
6777 split_indx = newindx;
6782 nkeys = NUMKEYS(mp);
6783 split_indx = nkeys / 2;
6784 if (newindx < split_indx)
6790 unsigned int lsize, rsize, ksize;
6791 /* Move half of the keys to the right sibling */
6793 x = mc->mc_ki[mc->mc_top] - split_indx;
6794 ksize = mc->mc_db->md_pad;
6795 split = LEAF2KEY(mp, split_indx, ksize);
6796 rsize = (nkeys - split_indx) * ksize;
6797 lsize = (nkeys - split_indx) * sizeof(indx_t);
6798 mp->mp_lower -= lsize;
6799 rp->mp_lower += lsize;
6800 mp->mp_upper += rsize - lsize;
6801 rp->mp_upper -= rsize - lsize;
6802 sepkey.mv_size = ksize;
6803 if (newindx == split_indx) {
6804 sepkey.mv_data = newkey->mv_data;
6806 sepkey.mv_data = split;
6809 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6810 memcpy(rp->mp_ptrs, split, rsize);
6811 sepkey.mv_data = rp->mp_ptrs;
6812 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6813 memcpy(ins, newkey->mv_data, ksize);
6814 mp->mp_lower += sizeof(indx_t);
6815 mp->mp_upper -= ksize - sizeof(indx_t);
6818 memcpy(rp->mp_ptrs, split, x * ksize);
6819 ins = LEAF2KEY(rp, x, ksize);
6820 memcpy(ins, newkey->mv_data, ksize);
6821 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6822 rp->mp_lower += sizeof(indx_t);
6823 rp->mp_upper -= ksize - sizeof(indx_t);
6824 mc->mc_ki[mc->mc_top] = x;
6825 mc->mc_pg[mc->mc_top] = rp;
6830 /* For leaf pages, check the split point based on what
6831 * fits where, since otherwise mdb_node_add can fail.
6833 * This check is only needed when the data items are
6834 * relatively large, such that being off by one will
6835 * make the difference between success or failure.
6837 * It's also relevant if a page happens to be laid out
6838 * such that one half of its nodes are all "small" and
6839 * the other half of its nodes are "large." If the new
6840 * item is also "large" and falls on the half with
6841 * "large" nodes, it also may not fit.
6844 unsigned int psize, nsize;
6845 /* Maximum free space in an empty page */
6846 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6847 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6848 if ((nkeys < 20) || (nsize > pmax/16)) {
6849 if (newindx <= split_indx) {
6852 for (i=0; i<split_indx; i++) {
6853 node = NODEPTR(mp, i);
6854 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6855 if (F_ISSET(node->mn_flags, F_BIGDATA))
6856 psize += sizeof(pgno_t);
6858 psize += NODEDSZ(node);
6862 split_indx = newindx;
6873 for (i=nkeys-1; i>=split_indx; i--) {
6874 node = NODEPTR(mp, i);
6875 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6876 if (F_ISSET(node->mn_flags, F_BIGDATA))
6877 psize += sizeof(pgno_t);
6879 psize += NODEDSZ(node);
6883 split_indx = newindx;
6894 /* First find the separating key between the split pages.
6895 * The case where newindx == split_indx is ambiguous; the
6896 * new item could go to the new page or stay on the original
6897 * page. If newpos == 1 it goes to the new page.
6899 if (newindx == split_indx && newpos) {
6900 sepkey.mv_size = newkey->mv_size;
6901 sepkey.mv_data = newkey->mv_data;
6903 node = NODEPTR(mp, split_indx);
6904 sepkey.mv_size = node->mn_ksize;
6905 sepkey.mv_data = NODEKEY(node);
6909 DPRINTF("separator is [%s]", DKEY(&sepkey));
6911 /* Copy separator key to the parent.
6913 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6917 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6920 if (mn.mc_snum == mc->mc_snum) {
6921 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6922 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6923 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6924 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6929 /* Right page might now have changed parent.
6930 * Check if left page also changed parent.
6932 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6933 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6934 for (i=0; i<ptop; i++) {
6935 mc->mc_pg[i] = mn.mc_pg[i];
6936 mc->mc_ki[i] = mn.mc_ki[i];
6938 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6939 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6943 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6946 mc->mc_flags ^= C_SPLITTING;
6947 if (rc != MDB_SUCCESS) {
6950 if (nflags & MDB_APPEND) {
6951 mc->mc_pg[mc->mc_top] = rp;
6952 mc->mc_ki[mc->mc_top] = 0;
6953 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6956 for (i=0; i<mc->mc_top; i++)
6957 mc->mc_ki[i] = mn.mc_ki[i];
6964 /* Move half of the keys to the right sibling. */
6966 /* grab a page to hold a temporary copy */
6967 copy = mdb_page_malloc(mc, 1);
6971 copy->mp_pgno = mp->mp_pgno;
6972 copy->mp_flags = mp->mp_flags;
6973 copy->mp_lower = PAGEHDRSZ;
6974 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6975 mc->mc_pg[mc->mc_top] = copy;
6976 for (i = j = 0; i <= nkeys; j++) {
6977 if (i == split_indx) {
6978 /* Insert in right sibling. */
6979 /* Reset insert index for right sibling. */
6980 if (i != newindx || (newpos ^ ins_new)) {
6982 mc->mc_pg[mc->mc_top] = rp;
6986 if (i == newindx && !ins_new) {
6987 /* Insert the original entry that caused the split. */
6988 rkey.mv_data = newkey->mv_data;
6989 rkey.mv_size = newkey->mv_size;
6998 /* Update index for the new key. */
6999 mc->mc_ki[mc->mc_top] = j;
7000 } else if (i == nkeys) {
7003 node = NODEPTR(mp, i);
7004 rkey.mv_data = NODEKEY(node);
7005 rkey.mv_size = node->mn_ksize;
7007 xdata.mv_data = NODEDATA(node);
7008 xdata.mv_size = NODEDSZ(node);
7011 pgno = NODEPGNO(node);
7012 flags = node->mn_flags;
7017 if (!IS_LEAF(mp) && j == 0) {
7018 /* First branch index doesn't need key data. */
7022 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7026 nkeys = NUMKEYS(copy);
7027 for (i=0; i<nkeys; i++)
7028 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7029 mp->mp_lower = copy->mp_lower;
7030 mp->mp_upper = copy->mp_upper;
7031 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7032 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7034 /* reset back to original page */
7035 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7036 mc->mc_pg[mc->mc_top] = mp;
7037 if (nflags & MDB_RESERVE) {
7038 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7039 if (!(node->mn_flags & F_BIGDATA))
7040 newdata->mv_data = NODEDATA(node);
7044 /* Make sure mc_ki is still valid.
7046 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7047 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7048 for (i=0; i<ptop; i++) {
7049 mc->mc_pg[i] = mn.mc_pg[i];
7050 mc->mc_ki[i] = mn.mc_ki[i];
7052 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7053 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7057 /* return tmp page to freelist */
7058 mdb_page_free(mc->mc_txn->mt_env, copy);
7061 /* Adjust other cursors pointing to mp */
7062 MDB_cursor *m2, *m3;
7063 MDB_dbi dbi = mc->mc_dbi;
7064 int fixup = NUMKEYS(mp);
7066 if (mc->mc_flags & C_SUB)
7069 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7070 if (m2 == mc) continue;
7071 if (mc->mc_flags & C_SUB)
7072 m3 = &m2->mc_xcursor->mx_cursor;
7075 if (!(m3->mc_flags & C_INITIALIZED))
7077 if (m3->mc_flags & C_SPLITTING)
7082 for (k=m3->mc_top; k>=0; k--) {
7083 m3->mc_ki[k+1] = m3->mc_ki[k];
7084 m3->mc_pg[k+1] = m3->mc_pg[k];
7086 if (m3->mc_ki[0] >= split_indx) {
7091 m3->mc_pg[0] = mc->mc_pg[0];
7095 if (m3->mc_pg[mc->mc_top] == mp) {
7096 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7097 m3->mc_ki[mc->mc_top]++;
7098 if (m3->mc_ki[mc->mc_top] >= fixup) {
7099 m3->mc_pg[mc->mc_top] = rp;
7100 m3->mc_ki[mc->mc_top] -= fixup;
7101 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7103 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7104 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7113 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7114 MDB_val *key, MDB_val *data, unsigned int flags)
7119 assert(key != NULL);
7120 assert(data != NULL);
7122 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7125 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7129 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7133 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7136 mdb_cursor_init(&mc, txn, dbi, &mx);
7137 return mdb_cursor_put(&mc, key, data, flags);
7141 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7143 if ((flag & CHANGEABLE) != flag)
7146 env->me_flags |= flag;
7148 env->me_flags &= ~flag;
7153 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7158 *arg = env->me_flags;
7163 mdb_env_get_path(MDB_env *env, const char **arg)
7168 *arg = env->me_path;
7172 /** Common code for #mdb_stat() and #mdb_env_stat().
7173 * @param[in] env the environment to operate in.
7174 * @param[in] db the #MDB_db record containing the stats to return.
7175 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7176 * @return 0, this function always succeeds.
7179 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7181 arg->ms_psize = env->me_psize;
7182 arg->ms_depth = db->md_depth;
7183 arg->ms_branch_pages = db->md_branch_pages;
7184 arg->ms_leaf_pages = db->md_leaf_pages;
7185 arg->ms_overflow_pages = db->md_overflow_pages;
7186 arg->ms_entries = db->md_entries;
7191 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7195 if (env == NULL || arg == NULL)
7198 toggle = mdb_env_pick_meta(env);
7200 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7204 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7208 if (env == NULL || arg == NULL)
7211 toggle = mdb_env_pick_meta(env);
7212 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7213 arg->me_mapsize = env->me_mapsize;
7214 arg->me_maxreaders = env->me_maxreaders;
7215 arg->me_numreaders = env->me_numreaders;
7216 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7217 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7221 /** Set the default comparison functions for a database.
7222 * Called immediately after a database is opened to set the defaults.
7223 * The user can then override them with #mdb_set_compare() or
7224 * #mdb_set_dupsort().
7225 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7226 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7229 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7231 uint16_t f = txn->mt_dbs[dbi].md_flags;
7233 txn->mt_dbxs[dbi].md_cmp =
7234 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7235 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7237 txn->mt_dbxs[dbi].md_dcmp =
7238 !(f & MDB_DUPSORT) ? 0 :
7239 ((f & MDB_INTEGERDUP)
7240 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7241 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7244 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7249 int rc, dbflag, exact;
7250 unsigned int unused = 0;
7253 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7254 mdb_default_cmp(txn, FREE_DBI);
7257 if ((flags & VALID_FLAGS) != flags)
7263 if (flags & PERSISTENT_FLAGS) {
7264 uint16_t f2 = flags & PERSISTENT_FLAGS;
7265 /* make sure flag changes get committed */
7266 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7267 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7268 txn->mt_flags |= MDB_TXN_DIRTY;
7271 mdb_default_cmp(txn, MAIN_DBI);
7275 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7276 mdb_default_cmp(txn, MAIN_DBI);
7279 /* Is the DB already open? */
7281 for (i=2; i<txn->mt_numdbs; i++) {
7282 if (!txn->mt_dbxs[i].md_name.mv_size) {
7283 /* Remember this free slot */
7284 if (!unused) unused = i;
7287 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7288 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7294 /* If no free slot and max hit, fail */
7295 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7296 return MDB_DBS_FULL;
7298 /* Cannot mix named databases with some mainDB flags */
7299 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7300 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7302 /* Find the DB info */
7303 dbflag = DB_NEW|DB_VALID;
7306 key.mv_data = (void *)name;
7307 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7308 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7309 if (rc == MDB_SUCCESS) {
7310 /* make sure this is actually a DB */
7311 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7312 if (!(node->mn_flags & F_SUBDATA))
7314 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7315 /* Create if requested */
7317 data.mv_size = sizeof(MDB_db);
7318 data.mv_data = &dummy;
7319 memset(&dummy, 0, sizeof(dummy));
7320 dummy.md_root = P_INVALID;
7321 dummy.md_flags = flags & PERSISTENT_FLAGS;
7322 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7326 /* OK, got info, add to table */
7327 if (rc == MDB_SUCCESS) {
7328 unsigned int slot = unused ? unused : txn->mt_numdbs;
7329 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7330 txn->mt_dbxs[slot].md_name.mv_size = len;
7331 txn->mt_dbxs[slot].md_rel = NULL;
7332 txn->mt_dbflags[slot] = dbflag;
7333 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7335 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7336 mdb_default_cmp(txn, slot);
7345 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7347 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7350 if (txn->mt_dbflags[dbi] & DB_STALE) {
7353 /* Stale, must read the DB's root. cursor_init does it for us. */
7354 mdb_cursor_init(&mc, txn, dbi, &mx);
7356 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7359 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7362 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7364 ptr = env->me_dbxs[dbi].md_name.mv_data;
7365 env->me_dbxs[dbi].md_name.mv_data = NULL;
7366 env->me_dbxs[dbi].md_name.mv_size = 0;
7367 env->me_dbflags[dbi] = 0;
7371 /** Add all the DB's pages to the free list.
7372 * @param[in] mc Cursor on the DB to free.
7373 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7374 * @return 0 on success, non-zero on failure.
7377 mdb_drop0(MDB_cursor *mc, int subs)
7381 rc = mdb_page_search(mc, NULL, 0);
7382 if (rc == MDB_SUCCESS) {
7387 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7388 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7391 mdb_cursor_copy(mc, &mx);
7392 while (mc->mc_snum > 0) {
7393 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7394 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7395 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7396 if (ni->mn_flags & F_BIGDATA) {
7400 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7401 rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL);
7404 assert(IS_OVERFLOW(omp));
7405 ovpages = omp->mp_pages;
7406 for (j=0; j<ovpages; j++) {
7407 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7410 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7411 mdb_xcursor_init1(mc, ni);
7412 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7418 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7420 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7423 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7428 mc->mc_ki[mc->mc_top] = i;
7429 rc = mdb_cursor_sibling(mc, 1);
7431 /* no more siblings, go back to beginning
7432 * of previous level.
7436 for (i=1; i<mc->mc_snum; i++) {
7438 mc->mc_pg[i] = mx.mc_pg[i];
7443 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7444 mc->mc_db->md_root);
7449 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7454 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7457 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7460 rc = mdb_cursor_open(txn, dbi, &mc);
7464 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7468 /* Can't delete the main DB */
7469 if (del && dbi > MAIN_DBI) {
7470 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7472 txn->mt_dbflags[dbi] = DB_STALE;
7473 mdb_dbi_close(txn->mt_env, dbi);
7476 /* reset the DB record, mark it dirty */
7477 txn->mt_dbflags[dbi] |= DB_DIRTY;
7478 txn->mt_dbs[dbi].md_depth = 0;
7479 txn->mt_dbs[dbi].md_branch_pages = 0;
7480 txn->mt_dbs[dbi].md_leaf_pages = 0;
7481 txn->mt_dbs[dbi].md_overflow_pages = 0;
7482 txn->mt_dbs[dbi].md_entries = 0;
7483 txn->mt_dbs[dbi].md_root = P_INVALID;
7485 txn->mt_flags |= MDB_TXN_DIRTY;
7488 mdb_cursor_close(mc);
7492 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7494 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7497 txn->mt_dbxs[dbi].md_cmp = cmp;
7501 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7503 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7506 txn->mt_dbxs[dbi].md_dcmp = cmp;
7510 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7512 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7515 txn->mt_dbxs[dbi].md_rel = rel;
7519 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7521 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7524 txn->mt_dbxs[dbi].md_relctx = ctx;