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.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define pthread_t DWORD
144 #define pthread_mutex_t HANDLE
145 #define pthread_key_t DWORD
146 #define pthread_self() GetCurrentThreadId()
147 #define pthread_key_create(x,y) \
148 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
149 #define pthread_key_delete(x) TlsFree(x)
150 #define pthread_getspecific(x) TlsGetValue(x)
151 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
152 #define pthread_mutex_unlock(x) ReleaseMutex(x)
153 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
154 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
158 #define getpid() GetCurrentProcessId()
159 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
160 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
161 #define ErrCode() GetLastError()
162 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
163 #define close(fd) CloseHandle(fd)
164 #define munmap(ptr,len) UnmapViewOfFile(ptr)
167 #ifdef MDB_USE_POSIX_SEM
169 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
170 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
171 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
172 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
175 mdb_sem_wait(sem_t *sem)
178 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
183 /** Lock the reader mutex.
185 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
186 /** Unlock the reader mutex.
188 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
190 /** Lock the writer mutex.
191 * Only a single write transaction is allowed at a time. Other writers
192 * will block waiting for this mutex.
194 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
195 /** Unlock the writer mutex.
197 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
198 #endif /* MDB_USE_POSIX_SEM */
200 /** Get the error code for the last failed system function.
202 #define ErrCode() errno
204 /** An abstraction for a file handle.
205 * On POSIX systems file handles are small integers. On Windows
206 * they're opaque pointers.
210 /** A value for an invalid file handle.
211 * Mainly used to initialize file variables and signify that they are
214 #define INVALID_HANDLE_VALUE (-1)
216 /** Get the size of a memory page for the system.
217 * This is the basic size that the platform's memory manager uses, and is
218 * fundamental to the use of memory-mapped files.
220 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
223 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
226 #define MNAME_LEN (sizeof(pthread_mutex_t))
232 /** A flag for opening a file and requesting synchronous data writes.
233 * This is only used when writing a meta page. It's not strictly needed;
234 * we could just do a normal write and then immediately perform a flush.
235 * But if this flag is available it saves us an extra system call.
237 * @note If O_DSYNC is undefined but exists in /usr/include,
238 * preferably set some compiler flag to get the definition.
239 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
242 # define MDB_DSYNC O_DSYNC
246 /** Function for flushing the data of a file. Define this to fsync
247 * if fdatasync() is not supported.
249 #ifndef MDB_FDATASYNC
250 # define MDB_FDATASYNC fdatasync
254 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
265 /** A page number in the database.
266 * Note that 64 bit page numbers are overkill, since pages themselves
267 * already represent 12-13 bits of addressable memory, and the OS will
268 * always limit applications to a maximum of 63 bits of address space.
270 * @note In the #MDB_node structure, we only store 48 bits of this value,
271 * which thus limits us to only 60 bits of addressable data.
273 typedef MDB_ID pgno_t;
275 /** A transaction ID.
276 * See struct MDB_txn.mt_txnid for details.
278 typedef MDB_ID txnid_t;
280 /** @defgroup debug Debug Macros
284 /** Enable debug output.
285 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
286 * read from and written to the database (used for free space management).
291 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
294 # define DPRINTF (void) /* Vararg macros may be unsupported */
296 static int mdb_debug;
297 static txnid_t mdb_debug_start;
299 /** Print a debug message with printf formatting. */
300 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
301 ((void) ((mdb_debug) && \
302 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
304 # define DPRINTF(fmt, ...) ((void) 0)
305 # define MDB_DEBUG_SKIP
307 /** Print a debug string.
308 * The string is printed literally, with no format processing.
310 #define DPUTS(arg) DPRINTF("%s", arg)
313 /** A default memory page size.
314 * The actual size is platform-dependent, but we use this for
315 * boot-strapping. We probably should not be using this any more.
316 * The #GET_PAGESIZE() macro is used to get the actual size.
318 * Note that we don't currently support Huge pages. On Linux,
319 * regular data files cannot use Huge pages, and in general
320 * Huge pages aren't actually pageable. We rely on the OS
321 * demand-pager to read our data and page it out when memory
322 * pressure from other processes is high. So until OSs have
323 * actual paging support for Huge pages, they're not viable.
325 #define MDB_PAGESIZE 4096
327 /** The minimum number of keys required in a database page.
328 * Setting this to a larger value will place a smaller bound on the
329 * maximum size of a data item. Data items larger than this size will
330 * be pushed into overflow pages instead of being stored directly in
331 * the B-tree node. This value used to default to 4. With a page size
332 * of 4096 bytes that meant that any item larger than 1024 bytes would
333 * go into an overflow page. That also meant that on average 2-3KB of
334 * each overflow page was wasted space. The value cannot be lower than
335 * 2 because then there would no longer be a tree structure. With this
336 * value, items larger than 2KB will go into overflow pages, and on
337 * average only 1KB will be wasted.
339 #define MDB_MINKEYS 2
341 /** A stamp that identifies a file as an MDB file.
342 * There's nothing special about this value other than that it is easily
343 * recognizable, and it will reflect any byte order mismatches.
345 #define MDB_MAGIC 0xBEEFC0DE
347 /** The version number for a database's file format. */
348 #define MDB_VERSION 1
350 /** @brief The maximum size of a key in the database.
352 * We require that keys all fit onto a regular page. This limit
353 * could be raised a bit further if needed; to something just
354 * under #MDB_PAGESIZE / #MDB_MINKEYS.
356 * Note that data items in an #MDB_DUPSORT database are actually keys
357 * of a subDB, so they're also limited to this size.
359 #ifndef MDB_MAXKEYSIZE
360 #define MDB_MAXKEYSIZE 511
363 /** @brief The maximum size of a data item.
365 * We only store a 32 bit value for node sizes.
367 #define MAXDATASIZE 0xffffffffUL
372 * This is used for printing a hex dump of a key's contents.
374 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
375 /** Display a key in hex.
377 * Invoke a function to display a key in hex.
379 #define DKEY(x) mdb_dkey(x, kbuf)
381 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
385 /** An invalid page number.
386 * Mainly used to denote an empty tree.
388 #define P_INVALID (~(pgno_t)0)
390 /** Test if the flags \b f are set in a flag word \b w. */
391 #define F_ISSET(w, f) (((w) & (f)) == (f))
393 /** Used for offsets within a single page.
394 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
397 typedef uint16_t indx_t;
399 /** Default size of memory map.
400 * This is certainly too small for any actual applications. Apps should always set
401 * the size explicitly using #mdb_env_set_mapsize().
403 #define DEFAULT_MAPSIZE 1048576
405 /** @defgroup readers Reader Lock Table
406 * Readers don't acquire any locks for their data access. Instead, they
407 * simply record their transaction ID in the reader table. The reader
408 * mutex is needed just to find an empty slot in the reader table. The
409 * slot's address is saved in thread-specific data so that subsequent read
410 * transactions started by the same thread need no further locking to proceed.
412 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
414 * No reader table is used if the database is on a read-only filesystem.
416 * Since the database uses multi-version concurrency control, readers don't
417 * actually need any locking. This table is used to keep track of which
418 * readers are using data from which old transactions, so that we'll know
419 * when a particular old transaction is no longer in use. Old transactions
420 * that have discarded any data pages can then have those pages reclaimed
421 * for use by a later write transaction.
423 * The lock table is constructed such that reader slots are aligned with the
424 * processor's cache line size. Any slot is only ever used by one thread.
425 * This alignment guarantees that there will be no contention or cache
426 * thrashing as threads update their own slot info, and also eliminates
427 * any need for locking when accessing a slot.
429 * A writer thread will scan every slot in the table to determine the oldest
430 * outstanding reader transaction. Any freed pages older than this will be
431 * reclaimed by the writer. The writer doesn't use any locks when scanning
432 * this table. This means that there's no guarantee that the writer will
433 * see the most up-to-date reader info, but that's not required for correct
434 * operation - all we need is to know the upper bound on the oldest reader,
435 * we don't care at all about the newest reader. So the only consequence of
436 * reading stale information here is that old pages might hang around a
437 * while longer before being reclaimed. That's actually good anyway, because
438 * the longer we delay reclaiming old pages, the more likely it is that a
439 * string of contiguous pages can be found after coalescing old pages from
440 * many old transactions together.
443 /** Number of slots in the reader table.
444 * This value was chosen somewhat arbitrarily. 126 readers plus a
445 * couple mutexes fit exactly into 8KB on my development machine.
446 * Applications should set the table size using #mdb_env_set_maxreaders().
448 #define DEFAULT_READERS 126
450 /** The size of a CPU cache line in bytes. We want our lock structures
451 * aligned to this size to avoid false cache line sharing in the
453 * This value works for most CPUs. For Itanium this should be 128.
459 /** The information we store in a single slot of the reader table.
460 * In addition to a transaction ID, we also record the process and
461 * thread ID that owns a slot, so that we can detect stale information,
462 * e.g. threads or processes that went away without cleaning up.
463 * @note We currently don't check for stale records. We simply re-init
464 * the table when we know that we're the only process opening the
467 typedef struct MDB_rxbody {
468 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
469 * Multiple readers that start at the same time will probably have the
470 * same ID here. Again, it's not important to exclude them from
471 * anything; all we need to know is which version of the DB they
472 * started from so we can avoid overwriting any data used in that
473 * particular version.
476 /** The process ID of the process owning this reader txn. */
478 /** The thread ID of the thread owning this txn. */
482 /** The actual reader record, with cacheline padding. */
483 typedef struct MDB_reader {
486 /** shorthand for mrb_txnid */
487 #define mr_txnid mru.mrx.mrb_txnid
488 #define mr_pid mru.mrx.mrb_pid
489 #define mr_tid mru.mrx.mrb_tid
490 /** cache line alignment */
491 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
495 /** The header for the reader table.
496 * The table resides in a memory-mapped file. (This is a different file
497 * than is used for the main database.)
499 * For POSIX the actual mutexes reside in the shared memory of this
500 * mapped file. On Windows, mutexes are named objects allocated by the
501 * kernel; we store the mutex names in this mapped file so that other
502 * processes can grab them. This same approach is also used on
503 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
504 * process-shared POSIX mutexes. For these cases where a named object
505 * is used, the object name is derived from a 64 bit FNV hash of the
506 * environment pathname. As such, naming collisions are extremely
507 * unlikely. If a collision occurs, the results are unpredictable.
509 typedef struct MDB_txbody {
510 /** Stamp identifying this as an MDB file. It must be set
513 /** Version number of this lock file. Must be set to #MDB_VERSION. */
514 uint32_t mtb_version;
515 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
516 char mtb_rmname[MNAME_LEN];
518 /** Mutex protecting access to this table.
519 * This is the reader lock that #LOCK_MUTEX_R acquires.
521 pthread_mutex_t mtb_mutex;
523 /** The ID of the last transaction committed to the database.
524 * This is recorded here only for convenience; the value can always
525 * be determined by reading the main database meta pages.
528 /** The number of slots that have been used in the reader table.
529 * This always records the maximum count, it is not decremented
530 * when readers release their slots.
532 unsigned mtb_numreaders;
535 /** The actual reader table definition. */
536 typedef struct MDB_txninfo {
539 #define mti_magic mt1.mtb.mtb_magic
540 #define mti_version mt1.mtb.mtb_version
541 #define mti_mutex mt1.mtb.mtb_mutex
542 #define mti_rmname mt1.mtb.mtb_rmname
543 #define mti_txnid mt1.mtb.mtb_txnid
544 #define mti_numreaders mt1.mtb.mtb_numreaders
545 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
548 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
549 char mt2_wmname[MNAME_LEN];
550 #define mti_wmname mt2.mt2_wmname
552 pthread_mutex_t mt2_wmutex;
553 #define mti_wmutex mt2.mt2_wmutex
555 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
557 MDB_reader mti_readers[1];
561 /** Common header for all page types.
562 * Overflow records occupy a number of contiguous pages with no
563 * headers on any page after the first.
565 typedef struct MDB_page {
566 #define mp_pgno mp_p.p_pgno
567 #define mp_next mp_p.p_next
569 pgno_t p_pgno; /**< page number */
570 void * p_next; /**< for in-memory list of freed structs */
573 /** @defgroup mdb_page Page Flags
575 * Flags for the page headers.
578 #define P_BRANCH 0x01 /**< branch page */
579 #define P_LEAF 0x02 /**< leaf page */
580 #define P_OVERFLOW 0x04 /**< overflow page */
581 #define P_META 0x08 /**< meta page */
582 #define P_DIRTY 0x10 /**< dirty page */
583 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
584 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
586 uint16_t mp_flags; /**< @ref mdb_page */
587 #define mp_lower mp_pb.pb.pb_lower
588 #define mp_upper mp_pb.pb.pb_upper
589 #define mp_pages mp_pb.pb_pages
592 indx_t pb_lower; /**< lower bound of free space */
593 indx_t pb_upper; /**< upper bound of free space */
595 uint32_t pb_pages; /**< number of overflow pages */
597 indx_t mp_ptrs[1]; /**< dynamic size */
600 /** Size of the page header, excluding dynamic data at the end */
601 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
603 /** Address of first usable data byte in a page, after the header */
604 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
606 /** Number of nodes on a page */
607 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
609 /** The amount of space remaining in the page */
610 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
612 /** The percentage of space used in the page, in tenths of a percent. */
613 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
614 ((env)->me_psize - PAGEHDRSZ))
615 /** The minimum page fill factor, in tenths of a percent.
616 * Pages emptier than this are candidates for merging.
618 #define FILL_THRESHOLD 250
620 /** Test if a page is a leaf page */
621 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
622 /** Test if a page is a LEAF2 page */
623 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
624 /** Test if a page is a branch page */
625 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
626 /** Test if a page is an overflow page */
627 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
628 /** Test if a page is a sub page */
629 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
631 /** The number of overflow pages needed to store the given size. */
632 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
634 /** Header for a single key/data pair within a page.
635 * We guarantee 2-byte alignment for nodes.
637 typedef struct MDB_node {
638 /** lo and hi are used for data size on leaf nodes and for
639 * child pgno on branch nodes. On 64 bit platforms, flags
640 * is also used for pgno. (Branch nodes have no flags).
641 * They are in host byte order in case that lets some
642 * accesses be optimized into a 32-bit word access.
644 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
645 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
646 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
647 /** @defgroup mdb_node Node Flags
649 * Flags for node headers.
652 #define F_BIGDATA 0x01 /**< data put on overflow page */
653 #define F_SUBDATA 0x02 /**< data is a sub-database */
654 #define F_DUPDATA 0x04 /**< data has duplicates */
656 /** valid flags for #mdb_node_add() */
657 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
660 unsigned short mn_flags; /**< @ref mdb_node */
661 unsigned short mn_ksize; /**< key size */
662 char mn_data[1]; /**< key and data are appended here */
665 /** Size of the node header, excluding dynamic data at the end */
666 #define NODESIZE offsetof(MDB_node, mn_data)
668 /** Bit position of top word in page number, for shifting mn_flags */
669 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
671 /** Size of a node in a branch page with a given key.
672 * This is just the node header plus the key, there is no data.
674 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
676 /** Size of a node in a leaf page with a given key and data.
677 * This is node header plus key plus data size.
679 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
681 /** Address of node \b i in page \b p */
682 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
684 /** Address of the key for the node */
685 #define NODEKEY(node) (void *)((node)->mn_data)
687 /** Address of the data for a node */
688 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
690 /** Get the page number pointed to by a branch node */
691 #define NODEPGNO(node) \
692 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
693 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
694 /** Set the page number in a branch node */
695 #define SETPGNO(node,pgno) do { \
696 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
697 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
699 /** Get the size of the data in a leaf node */
700 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
701 /** Set the size of the data for a leaf node */
702 #define SETDSZ(node,size) do { \
703 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
704 /** The size of a key in a node */
705 #define NODEKSZ(node) ((node)->mn_ksize)
707 /** Copy a page number from src to dst */
709 #define COPY_PGNO(dst,src) dst = src
711 #if SIZE_MAX > 4294967295UL
712 #define COPY_PGNO(dst,src) do { \
713 unsigned short *s, *d; \
714 s = (unsigned short *)&(src); \
715 d = (unsigned short *)&(dst); \
722 #define COPY_PGNO(dst,src) do { \
723 unsigned short *s, *d; \
724 s = (unsigned short *)&(src); \
725 d = (unsigned short *)&(dst); \
731 /** The address of a key in a LEAF2 page.
732 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
733 * There are no node headers, keys are stored contiguously.
735 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
737 /** Set the \b node's key into \b key, if requested. */
738 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
739 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
741 /** Information about a single database in the environment. */
742 typedef struct MDB_db {
743 uint32_t md_pad; /**< also ksize for LEAF2 pages */
744 uint16_t md_flags; /**< @ref mdb_dbi_open */
745 uint16_t md_depth; /**< depth of this tree */
746 pgno_t md_branch_pages; /**< number of internal pages */
747 pgno_t md_leaf_pages; /**< number of leaf pages */
748 pgno_t md_overflow_pages; /**< number of overflow pages */
749 size_t md_entries; /**< number of data items */
750 pgno_t md_root; /**< the root page of this tree */
753 /** mdb_dbi_open flags */
754 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
755 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
756 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
757 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
759 /** Handle for the DB used to track free pages. */
761 /** Handle for the default DB. */
764 /** Meta page content. */
765 typedef struct MDB_meta {
766 /** Stamp identifying this as an MDB file. It must be set
769 /** Version number of this lock file. Must be set to #MDB_VERSION. */
771 void *mm_address; /**< address for fixed mapping */
772 size_t mm_mapsize; /**< size of mmap region */
773 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
774 /** The size of pages used in this DB */
775 #define mm_psize mm_dbs[0].md_pad
776 /** Any persistent environment flags. @ref mdb_env */
777 #define mm_flags mm_dbs[0].md_flags
778 pgno_t mm_last_pg; /**< last used page in file */
779 txnid_t mm_txnid; /**< txnid that committed this page */
782 /** Buffer for a stack-allocated dirty page.
783 * The members define size and alignment, and silence type
784 * aliasing warnings. They are not used directly; that could
785 * mean incorrectly using several union members in parallel.
787 typedef union MDB_pagebuf {
788 char mb_raw[MDB_PAGESIZE];
791 char mm_pad[PAGEHDRSZ];
796 /** Auxiliary DB info.
797 * The information here is mostly static/read-only. There is
798 * only a single copy of this record in the environment.
800 typedef struct MDB_dbx {
801 MDB_val md_name; /**< name of the database */
802 MDB_cmp_func *md_cmp; /**< function for comparing keys */
803 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
804 MDB_rel_func *md_rel; /**< user relocate function */
805 void *md_relctx; /**< user-provided context for md_rel */
808 /** A database transaction.
809 * Every operation requires a transaction handle.
812 MDB_txn *mt_parent; /**< parent of a nested txn */
813 MDB_txn *mt_child; /**< nested txn under this txn */
814 pgno_t mt_next_pgno; /**< next unallocated page */
815 /** The ID of this transaction. IDs are integers incrementing from 1.
816 * Only committed write transactions increment the ID. If a transaction
817 * aborts, the ID may be re-used by the next writer.
820 MDB_env *mt_env; /**< the DB environment */
821 /** The list of pages that became unused during this transaction.
825 MDB_ID2L dirty_list; /**< for write txns: modified pages */
826 MDB_reader *reader; /**< this thread's reader table slot or NULL */
828 /** Array of records for each DB known in the environment. */
830 /** Array of MDB_db records for each known DB */
832 /** @defgroup mt_dbflag Transaction DB Flags
836 #define DB_DIRTY 0x01 /**< DB was written in this txn */
837 #define DB_STALE 0x02 /**< DB record is older than txnID */
838 #define DB_NEW 0x04 /**< DB handle opened in this txn */
839 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
841 /** In write txns, array of cursors for each DB */
842 MDB_cursor **mt_cursors;
843 /** Array of flags for each DB */
844 unsigned char *mt_dbflags;
845 /** Number of DB records in use. This number only ever increments;
846 * we don't decrement it when individual DB handles are closed.
850 /** @defgroup mdb_txn Transaction Flags
854 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
855 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
856 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
858 unsigned int mt_flags; /**< @ref mdb_txn */
859 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
860 unsigned int mt_dirty_room;
861 /** Tracks which of the two meta pages was used at the start
862 * of this transaction.
864 unsigned int mt_toggle;
867 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
868 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
869 * raise this on a 64 bit machine.
871 #define CURSOR_STACK 32
875 /** Cursors are used for all DB operations */
877 /** Next cursor on this DB in this txn */
879 /** Original cursor if this is a shadow */
881 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
882 struct MDB_xcursor *mc_xcursor;
883 /** The transaction that owns this cursor */
885 /** The database handle this cursor operates on */
887 /** The database record for this cursor */
889 /** The database auxiliary record for this cursor */
891 /** The @ref mt_dbflag for this database */
892 unsigned char *mc_dbflag;
893 unsigned short mc_snum; /**< number of pushed pages */
894 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
895 /** @defgroup mdb_cursor Cursor Flags
897 * Cursor state flags.
900 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
901 #define C_EOF 0x02 /**< No more data */
902 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
903 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
904 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
905 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
906 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
908 unsigned int mc_flags; /**< @ref mdb_cursor */
909 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
910 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
913 /** Context for sorted-dup records.
914 * We could have gone to a fully recursive design, with arbitrarily
915 * deep nesting of sub-databases. But for now we only handle these
916 * levels - main DB, optional sub-DB, sorted-duplicate DB.
918 typedef struct MDB_xcursor {
919 /** A sub-cursor for traversing the Dup DB */
920 MDB_cursor mx_cursor;
921 /** The database record for this Dup DB */
923 /** The auxiliary DB record for this Dup DB */
925 /** The @ref mt_dbflag for this Dup DB */
926 unsigned char mx_dbflag;
929 /** State of FreeDB old pages, stored in the MDB_env */
930 typedef struct MDB_pgstate {
931 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
932 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
935 /** The database environment. */
937 HANDLE me_fd; /**< The main data file */
938 HANDLE me_lfd; /**< The lock file */
939 HANDLE me_mfd; /**< just for writing the meta pages */
940 /** Failed to update the meta page. Probably an I/O error. */
941 #define MDB_FATAL_ERROR 0x80000000U
942 /** Some fields are initialized. */
943 #define MDB_ENV_ACTIVE 0x20000000U
944 /** me_txkey is set */
945 #define MDB_ENV_TXKEY 0x10000000U
946 uint32_t me_flags; /**< @ref mdb_env */
947 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
948 unsigned int me_maxreaders; /**< size of the reader table */
949 unsigned int me_numreaders; /**< max numreaders set by this env */
950 MDB_dbi me_numdbs; /**< number of DBs opened */
951 MDB_dbi me_maxdbs; /**< size of the DB table */
952 pid_t me_pid; /**< process ID of this env */
953 char *me_path; /**< path to the DB files */
954 char *me_map; /**< the memory map of the data file */
955 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
956 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
957 MDB_txn *me_txn; /**< current write transaction */
958 size_t me_mapsize; /**< size of the data memory map */
959 off_t me_size; /**< current file size */
960 pgno_t me_maxpg; /**< me_mapsize / me_psize */
961 MDB_dbx *me_dbxs; /**< array of static DB info */
962 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
963 pthread_key_t me_txkey; /**< thread-key for readers */
964 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
965 # define me_pglast me_pgstate.mf_pglast
966 # define me_pghead me_pgstate.mf_pghead
967 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
968 /** IDL of pages that became unused in a write txn */
970 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
971 MDB_ID2L me_dirty_list;
972 /** Max number of freelist items that can fit in a single overflow page */
974 /** Max size of a node on a page */
975 unsigned int me_nodemax;
977 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
979 #elif defined(MDB_USE_POSIX_SEM)
980 sem_t *me_rmutex; /* Shared mutexes are not supported */
985 /** Nested transaction */
986 typedef struct MDB_ntxn {
987 MDB_txn mnt_txn; /* the transaction */
988 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
991 /** max number of pages to commit in one writev() call */
992 #define MDB_COMMIT_PAGES 64
993 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
994 #undef MDB_COMMIT_PAGES
995 #define MDB_COMMIT_PAGES IOV_MAX
998 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
999 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1000 static int mdb_page_touch(MDB_cursor *mc);
1002 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1003 static int mdb_page_search_root(MDB_cursor *mc,
1004 MDB_val *key, int modify);
1005 #define MDB_PS_MODIFY 1
1006 #define MDB_PS_ROOTONLY 2
1007 static int mdb_page_search(MDB_cursor *mc,
1008 MDB_val *key, int flags);
1009 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1011 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1012 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1013 pgno_t newpgno, unsigned int nflags);
1015 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1016 static int mdb_env_pick_meta(const MDB_env *env);
1017 static int mdb_env_write_meta(MDB_txn *txn);
1018 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1019 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1021 static void mdb_env_close0(MDB_env *env, int excl);
1023 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1024 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1025 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1026 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1027 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1028 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1029 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1030 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1031 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1033 static int mdb_rebalance(MDB_cursor *mc);
1034 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1036 static void mdb_cursor_pop(MDB_cursor *mc);
1037 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1039 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1040 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1041 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1042 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1043 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1045 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1046 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1048 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1049 static void mdb_xcursor_init0(MDB_cursor *mc);
1050 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1052 static int mdb_drop0(MDB_cursor *mc, int subs);
1053 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1056 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1060 static SECURITY_DESCRIPTOR mdb_null_sd;
1061 static SECURITY_ATTRIBUTES mdb_all_sa;
1062 static int mdb_sec_inited;
1065 /** Return the library version info. */
1067 mdb_version(int *major, int *minor, int *patch)
1069 if (major) *major = MDB_VERSION_MAJOR;
1070 if (minor) *minor = MDB_VERSION_MINOR;
1071 if (patch) *patch = MDB_VERSION_PATCH;
1072 return MDB_VERSION_STRING;
1075 /** Table of descriptions for MDB @ref errors */
1076 static char *const mdb_errstr[] = {
1077 "MDB_KEYEXIST: Key/data pair already exists",
1078 "MDB_NOTFOUND: No matching key/data pair found",
1079 "MDB_PAGE_NOTFOUND: Requested page not found",
1080 "MDB_CORRUPTED: Located page was wrong type",
1081 "MDB_PANIC: Update of meta page failed",
1082 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1083 "MDB_INVALID: File is not an MDB file",
1084 "MDB_MAP_FULL: Environment mapsize limit reached",
1085 "MDB_DBS_FULL: Environment maxdbs limit reached",
1086 "MDB_READERS_FULL: Environment maxreaders limit reached",
1087 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1088 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1089 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1090 "MDB_PAGE_FULL: Internal error - page has no more space",
1091 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1092 "MDB_INCOMPATIBLE: Database flags changed or would change",
1093 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1097 mdb_strerror(int err)
1101 return ("Successful return: 0");
1103 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1104 i = err - MDB_KEYEXIST;
1105 return mdb_errstr[i];
1108 return strerror(err);
1112 /** Display a key in hexadecimal and return the address of the result.
1113 * @param[in] key the key to display
1114 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1115 * @return The key in hexadecimal form.
1118 mdb_dkey(MDB_val *key, char *buf)
1121 unsigned char *c = key->mv_data;
1127 if (key->mv_size > MDB_MAXKEYSIZE)
1128 return "MDB_MAXKEYSIZE";
1129 /* may want to make this a dynamic check: if the key is mostly
1130 * printable characters, print it as-is instead of converting to hex.
1134 for (i=0; i<key->mv_size; i++)
1135 ptr += sprintf(ptr, "%02x", *c++);
1137 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1142 /** Display all the keys in the page. */
1144 mdb_page_list(MDB_page *mp)
1147 unsigned int i, nkeys, nsize;
1151 nkeys = NUMKEYS(mp);
1152 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1153 for (i=0; i<nkeys; i++) {
1154 node = NODEPTR(mp, i);
1155 key.mv_size = node->mn_ksize;
1156 key.mv_data = node->mn_data;
1157 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1158 if (IS_BRANCH(mp)) {
1159 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1162 if (F_ISSET(node->mn_flags, F_BIGDATA))
1163 nsize += sizeof(pgno_t);
1165 nsize += NODEDSZ(node);
1166 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1172 mdb_cursor_chk(MDB_cursor *mc)
1178 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1179 for (i=0; i<mc->mc_top; i++) {
1181 node = NODEPTR(mp, mc->mc_ki[i]);
1182 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1185 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1191 /** Count all the pages in each DB and in the freelist
1192 * and make sure it matches the actual number of pages
1195 static void mdb_audit(MDB_txn *txn)
1199 MDB_ID freecount, count;
1204 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1205 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1206 freecount += *(MDB_ID *)data.mv_data;
1209 for (i = 0; i<txn->mt_numdbs; i++) {
1211 mdb_cursor_init(&mc, txn, i, &mx);
1212 if (txn->mt_dbs[i].md_root == P_INVALID)
1214 count += txn->mt_dbs[i].md_branch_pages +
1215 txn->mt_dbs[i].md_leaf_pages +
1216 txn->mt_dbs[i].md_overflow_pages;
1217 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1218 mdb_page_search(&mc, NULL, 0);
1222 mp = mc.mc_pg[mc.mc_top];
1223 for (j=0; j<NUMKEYS(mp); j++) {
1224 MDB_node *leaf = NODEPTR(mp, j);
1225 if (leaf->mn_flags & F_SUBDATA) {
1227 memcpy(&db, NODEDATA(leaf), sizeof(db));
1228 count += db.md_branch_pages + db.md_leaf_pages +
1229 db.md_overflow_pages;
1233 while (mdb_cursor_sibling(&mc, 1) == 0);
1236 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1237 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1238 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1244 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1246 return txn->mt_dbxs[dbi].md_cmp(a, b);
1250 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1252 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1255 /** Allocate a page.
1256 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1259 mdb_page_malloc(MDB_txn *txn, unsigned num)
1261 MDB_env *env = txn->mt_env;
1262 MDB_page *ret = env->me_dpages;
1263 size_t sz = env->me_psize;
1266 VGMEMP_ALLOC(env, ret, sz);
1267 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1268 env->me_dpages = ret->mp_next;
1274 if ((ret = malloc(sz)) != NULL) {
1275 VGMEMP_ALLOC(env, ret, sz);
1280 /** Free a single page.
1281 * Saves single pages to a list, for future reuse.
1282 * (This is not used for multi-page overflow pages.)
1285 mdb_page_free(MDB_env *env, MDB_page *mp)
1287 mp->mp_next = env->me_dpages;
1288 VGMEMP_FREE(env, mp);
1289 env->me_dpages = mp;
1292 /* Free a dirty page */
1294 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1296 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1297 mdb_page_free(env, dp);
1299 /* large pages just get freed directly */
1300 VGMEMP_FREE(env, dp);
1305 /* Return all dirty pages to dpage list */
1307 mdb_dlist_free(MDB_txn *txn)
1309 MDB_env *env = txn->mt_env;
1310 MDB_ID2L dl = txn->mt_u.dirty_list;
1311 unsigned i, n = dl[0].mid;
1313 for (i = 1; i <= n; i++) {
1314 mdb_dpage_free(env, dl[i].mptr);
1319 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1321 mdb_find_oldest(MDB_txn *txn)
1324 txnid_t mr, oldest = txn->mt_txnid - 1;
1325 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1326 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1336 /** Allocate pages for writing.
1337 * If there are free pages available from older transactions, they
1338 * will be re-used first. Otherwise a new page will be allocated.
1339 * @param[in] mc cursor A cursor handle identifying the transaction and
1340 * database for which we are allocating.
1341 * @param[in] num the number of pages to allocate.
1342 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1343 * will always be satisfied by a single contiguous chunk of memory.
1344 * @return 0 on success, non-zero on failure.
1347 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1349 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1350 /* Get at most <Max_retries> more freeDB records once me_pghead
1351 * has enough pages. If not enough, use new pages from the map.
1352 * If <Paranoid> and mc is updating the freeDB, only get new
1353 * records if me_pghead is empty. Then the freelist cannot play
1354 * catch-up with itself by growing while trying to save it.
1356 enum { Paranoid = 1, Max_retries = 500 };
1358 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1360 int rc, n2 = num-1, retry = Max_retries;
1361 MDB_txn *txn = mc->mc_txn;
1362 MDB_env *env = txn->mt_env;
1363 pgno_t pgno, *mop = env->me_pghead;
1364 unsigned mop_len = mop ? mop[0] : 0;
1367 txnid_t oldest = 0, last;
1373 /* If our dirty list is already full, we can't do anything */
1374 if (txn->mt_dirty_room == 0)
1375 return MDB_TXN_FULL;
1377 for (op = MDB_FIRST;; op = MDB_NEXT) {
1378 unsigned int i, j, k;
1381 pgno_t *idl, old_id, new_id;
1383 /* Seek a big enough contiguous page range. Prefer
1384 * pages at the tail, just truncating the list.
1386 if (mop_len >= (unsigned)num) {
1390 if (mop[i-n2] == pgno+n2) {
1391 mop[0] = mop_len -= num;
1392 /* Move any stragglers down */
1393 for (j = i-n2; j <= mop_len; )
1394 mop[j++] = mop[++i];
1397 } while (--i >= (unsigned)num);
1398 if (Max_retries < INT_MAX && --retry < 0)
1402 if (op == MDB_FIRST) { /* 1st iteration */
1403 /* Prepare to fetch more and coalesce */
1404 oldest = mdb_find_oldest(txn);
1405 last = env->me_pglast;
1406 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1409 key.mv_data = &last; /* will loop up last+1 */
1410 key.mv_size = sizeof(last);
1412 if (Paranoid && mc->mc_dbi == FREE_DBI)
1415 if (Paranoid && retry < 0 && mop_len)
1419 /* Do not fetch more if the record will be too recent */
1422 rc = mdb_cursor_get(&m2, &key, NULL, op);
1424 if (rc == MDB_NOTFOUND)
1428 last = *(txnid_t*)key.mv_data;
1431 np = m2.mc_pg[m2.mc_top];
1432 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1433 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1436 idl = (MDB_ID *) data.mv_data;
1439 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1441 } else if (mop_len+i > mop[-1]) {
1442 if ((rc = mdb_midl_grow(&env->me_pghead, i)) != 0)
1444 mop = env->me_pghead;
1446 env->me_pglast = last;
1448 DPRINTF("IDL read txn %zu root %zu num %u",
1449 last, txn->mt_dbs[FREE_DBI].md_root, i);
1451 DPRINTF("IDL %zu", idl[k]);
1453 /* Merge in descending sorted order */
1456 mop[0] = (pgno_t)-1;
1460 for (; old_id < new_id; old_id = mop[--j])
1467 /* Use new pages from the map when nothing suitable in the freeDB */
1469 if (txn->mt_next_pgno + num >= env->me_maxpg) {
1470 DPUTS("DB size maxed out");
1471 return MDB_MAP_FULL;
1475 if (env->me_flags & MDB_WRITEMAP) {
1476 if (pgno == P_INVALID) {
1477 pgno = txn->mt_next_pgno;
1478 txn->mt_next_pgno += num;
1480 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1482 if (!(np = mdb_page_malloc(txn, num)))
1484 if (pgno == P_INVALID) {
1485 pgno = txn->mt_next_pgno;
1486 txn->mt_next_pgno += num;
1489 mid.mid = np->mp_pgno = pgno;
1491 if (env->me_flags & MDB_WRITEMAP) {
1492 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1494 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1496 txn->mt_dirty_room--;
1502 /** Copy the used portions of a non-overflow page.
1503 * @param[in] dst page to copy into
1504 * @param[in] src page to copy from
1505 * @param[in] psize size of a page
1508 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1510 enum { Align = sizeof(pgno_t) };
1511 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1513 /* If page isn't full, just copy the used portion. Adjust
1514 * alignment so memcpy may copy words instead of bytes.
1516 if ((unused &= -Align) && !IS_LEAF2(src)) {
1518 memcpy(dst, src, (lower + (Align-1)) & -Align);
1519 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1522 memcpy(dst, src, psize - unused);
1526 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1527 * @param[in] mc cursor pointing to the page to be touched
1528 * @return 0 on success, non-zero on failure.
1531 mdb_page_touch(MDB_cursor *mc)
1533 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1534 MDB_txn *txn = mc->mc_txn;
1535 MDB_cursor *m2, *m3;
1540 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1541 if ((rc = mdb_page_alloc(mc, 1, &np)))
1544 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi,mp->mp_pgno,pgno);
1545 assert(mp->mp_pgno != pgno);
1546 mdb_midl_append(&txn->mt_free_pgs, mp->mp_pgno);
1547 /* Update the parent page, if any, to point to the new page */
1549 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1550 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1551 SETPGNO(node, pgno);
1553 mc->mc_db->md_root = pgno;
1555 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1556 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1558 /* If txn has a parent, make sure the page is in our
1562 unsigned x = mdb_mid2l_search(dl, pgno);
1563 if (x <= dl[0].mid && dl[x].mid == pgno) {
1566 mc->mc_pg[mc->mc_top] = np;
1570 assert(dl[0].mid < MDB_IDL_UM_MAX);
1572 np = mdb_page_malloc(txn, 1);
1577 mdb_mid2l_insert(dl, &mid);
1582 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1584 np->mp_flags |= P_DIRTY;
1586 /* Adjust cursors pointing to mp */
1587 mc->mc_pg[mc->mc_top] = np;
1589 if (mc->mc_flags & C_SUB) {
1591 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1592 m3 = &m2->mc_xcursor->mx_cursor;
1593 if (m3->mc_snum < mc->mc_snum) continue;
1594 if (m3->mc_pg[mc->mc_top] == mp)
1595 m3->mc_pg[mc->mc_top] = np;
1598 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1599 if (m2->mc_snum < mc->mc_snum) continue;
1600 if (m2->mc_pg[mc->mc_top] == mp) {
1601 m2->mc_pg[mc->mc_top] = np;
1602 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1603 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1605 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1606 if (!(leaf->mn_flags & F_SUBDATA))
1607 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1616 mdb_env_sync(MDB_env *env, int force)
1619 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1620 if (env->me_flags & MDB_WRITEMAP) {
1621 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1622 ? MS_ASYNC : MS_SYNC;
1623 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1626 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1630 if (MDB_FDATASYNC(env->me_fd))
1637 /** Make shadow copies of all of parent txn's cursors */
1639 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1641 MDB_cursor *mc, *m2;
1642 unsigned int i, j, size;
1644 for (i=0;i<src->mt_numdbs; i++) {
1645 if (src->mt_cursors[i]) {
1646 size = sizeof(MDB_cursor);
1647 if (src->mt_cursors[i]->mc_xcursor)
1648 size += sizeof(MDB_xcursor);
1649 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1656 mc->mc_db = &dst->mt_dbs[i];
1657 mc->mc_dbx = m2->mc_dbx;
1658 mc->mc_dbflag = &dst->mt_dbflags[i];
1659 mc->mc_snum = m2->mc_snum;
1660 mc->mc_top = m2->mc_top;
1661 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1662 for (j=0; j<mc->mc_snum; j++) {
1663 mc->mc_pg[j] = m2->mc_pg[j];
1664 mc->mc_ki[j] = m2->mc_ki[j];
1666 if (m2->mc_xcursor) {
1667 MDB_xcursor *mx, *mx2;
1668 mx = (MDB_xcursor *)(mc+1);
1669 mc->mc_xcursor = mx;
1670 mx2 = m2->mc_xcursor;
1671 mx->mx_db = mx2->mx_db;
1672 mx->mx_dbx = mx2->mx_dbx;
1673 mx->mx_dbflag = mx2->mx_dbflag;
1674 mx->mx_cursor.mc_txn = dst;
1675 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1676 mx->mx_cursor.mc_db = &mx->mx_db;
1677 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1678 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1679 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1680 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1681 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1682 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1683 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1684 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1687 mc->mc_xcursor = NULL;
1689 mc->mc_next = dst->mt_cursors[i];
1690 dst->mt_cursors[i] = mc;
1697 /** Close this write txn's cursors, after optionally merging its shadow
1698 * cursors back into parent's.
1699 * @param[in] txn the transaction handle.
1700 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1701 * @return 0 on success, non-zero on failure.
1704 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1706 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1709 for (i = txn->mt_numdbs; --i >= 0; ) {
1710 for (mc = cursors[i]; mc; mc = next) {
1712 if (mc->mc_flags & merge) {
1713 MDB_cursor *m2 = mc->mc_orig;
1714 m2->mc_snum = mc->mc_snum;
1715 m2->mc_top = mc->mc_top;
1716 for (j = mc->mc_snum; --j >= 0; ) {
1717 m2->mc_pg[j] = mc->mc_pg[j];
1718 m2->mc_ki[j] = mc->mc_ki[j];
1721 if (mc->mc_flags & C_ALLOCD)
1729 mdb_txn_reset0(MDB_txn *txn);
1731 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1732 * @param[in] txn the transaction handle to initialize
1733 * @return 0 on success, non-zero on failure.
1736 mdb_txn_renew0(MDB_txn *txn)
1738 MDB_env *env = txn->mt_env;
1741 int rc, new_notls = 0;
1744 txn->mt_numdbs = env->me_numdbs;
1745 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1747 if (txn->mt_flags & MDB_TXN_RDONLY) {
1748 if (!env->me_txns) {
1749 i = mdb_env_pick_meta(env);
1750 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1751 txn->mt_u.reader = NULL;
1753 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1754 pthread_getspecific(env->me_txkey);
1756 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1757 return MDB_BAD_RSLOT;
1759 pid_t pid = env->me_pid;
1760 pthread_t tid = pthread_self();
1763 for (i=0; i<env->me_txns->mti_numreaders; i++)
1764 if (env->me_txns->mti_readers[i].mr_pid == 0)
1766 if (i == env->me_maxreaders) {
1767 UNLOCK_MUTEX_R(env);
1768 return MDB_READERS_FULL;
1770 env->me_txns->mti_readers[i].mr_pid = pid;
1771 env->me_txns->mti_readers[i].mr_tid = tid;
1772 if (i >= env->me_txns->mti_numreaders)
1773 env->me_txns->mti_numreaders = i+1;
1774 /* Save numreaders for un-mutexed mdb_env_close() */
1775 env->me_numreaders = env->me_txns->mti_numreaders;
1776 UNLOCK_MUTEX_R(env);
1777 r = &env->me_txns->mti_readers[i];
1778 new_notls = (env->me_flags & MDB_NOTLS);
1779 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1784 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1785 txn->mt_u.reader = r;
1787 txn->mt_toggle = txn->mt_txnid & 1;
1788 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1792 txn->mt_txnid = env->me_txns->mti_txnid;
1793 txn->mt_toggle = txn->mt_txnid & 1;
1794 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1797 if (txn->mt_txnid == mdb_debug_start)
1800 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1801 txn->mt_u.dirty_list = env->me_dirty_list;
1802 txn->mt_u.dirty_list[0].mid = 0;
1803 txn->mt_free_pgs = env->me_free_pgs;
1804 txn->mt_free_pgs[0] = 0;
1808 /* Copy the DB info and flags */
1809 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1810 for (i=2; i<txn->mt_numdbs; i++) {
1811 x = env->me_dbflags[i];
1812 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1813 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1815 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1817 if (env->me_maxpg < txn->mt_next_pgno) {
1818 mdb_txn_reset0(txn);
1820 txn->mt_u.reader->mr_pid = 0;
1821 txn->mt_u.reader = NULL;
1823 return MDB_MAP_RESIZED;
1830 mdb_txn_renew(MDB_txn *txn)
1834 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1837 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1838 DPUTS("environment had fatal error, must shutdown!");
1842 rc = mdb_txn_renew0(txn);
1843 if (rc == MDB_SUCCESS) {
1844 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1845 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1846 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1852 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1856 int rc, size, tsize = sizeof(MDB_txn);
1858 if (env->me_flags & MDB_FATAL_ERROR) {
1859 DPUTS("environment had fatal error, must shutdown!");
1862 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1865 /* Nested transactions: Max 1 child, write txns only, no writemap */
1866 if (parent->mt_child ||
1867 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1868 (env->me_flags & MDB_WRITEMAP))
1872 tsize = sizeof(MDB_ntxn);
1874 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1875 if (!(flags & MDB_RDONLY))
1876 size += env->me_maxdbs * sizeof(MDB_cursor *);
1878 if ((txn = calloc(1, size)) == NULL) {
1879 DPRINTF("calloc: %s", strerror(ErrCode()));
1882 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1883 if (flags & MDB_RDONLY) {
1884 txn->mt_flags |= MDB_TXN_RDONLY;
1885 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1887 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1888 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1894 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1895 if (!txn->mt_u.dirty_list ||
1896 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1898 free(txn->mt_u.dirty_list);
1902 txn->mt_txnid = parent->mt_txnid;
1903 txn->mt_toggle = parent->mt_toggle;
1904 txn->mt_dirty_room = parent->mt_dirty_room;
1905 txn->mt_u.dirty_list[0].mid = 0;
1906 txn->mt_next_pgno = parent->mt_next_pgno;
1907 parent->mt_child = txn;
1908 txn->mt_parent = parent;
1909 txn->mt_numdbs = parent->mt_numdbs;
1910 txn->mt_flags = parent->mt_flags;
1911 txn->mt_dbxs = parent->mt_dbxs;
1912 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1913 /* Copy parent's mt_dbflags, but clear DB_NEW */
1914 for (i=0; i<txn->mt_numdbs; i++)
1915 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1917 ntxn = (MDB_ntxn *)txn;
1918 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1919 if (env->me_pghead) {
1920 size = MDB_IDL_SIZEOF(env->me_pghead);
1921 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1923 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1928 rc = mdb_cursor_shadow(parent, txn);
1930 mdb_txn_reset0(txn);
1932 rc = mdb_txn_renew0(txn);
1938 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1939 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1940 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1946 /** Export or close DBI handles opened in this txn. */
1948 mdb_dbis_update(MDB_txn *txn, int keep)
1951 MDB_dbi n = txn->mt_numdbs;
1952 MDB_env *env = txn->mt_env;
1953 unsigned char *tdbflags = txn->mt_dbflags;
1955 for (i = n; --i >= 2;) {
1956 if (tdbflags[i] & DB_NEW) {
1958 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
1960 char *ptr = env->me_dbxs[i].md_name.mv_data;
1961 env->me_dbxs[i].md_name.mv_data = NULL;
1962 env->me_dbxs[i].md_name.mv_size = 0;
1963 env->me_dbflags[i] = 0;
1968 if (keep && env->me_numdbs < n)
1972 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1973 * May be called twice for readonly txns: First reset it, then abort.
1974 * @param[in] txn the transaction handle to reset
1977 mdb_txn_reset0(MDB_txn *txn)
1979 MDB_env *env = txn->mt_env;
1981 /* Close any DBI handles opened in this txn */
1982 mdb_dbis_update(txn, 0);
1984 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1985 if (txn->mt_u.reader) {
1986 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1987 if (!(env->me_flags & MDB_NOTLS))
1988 txn->mt_u.reader = NULL; /* txn does not own reader */
1990 txn->mt_numdbs = 0; /* close nothing if called again */
1991 txn->mt_dbxs = NULL; /* mark txn as reset */
1993 mdb_cursors_close(txn, 0);
1995 if (!(env->me_flags & MDB_WRITEMAP)) {
1996 mdb_dlist_free(txn);
1998 mdb_midl_free(env->me_pghead);
2000 if (txn->mt_parent) {
2001 txn->mt_parent->mt_child = NULL;
2002 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2003 mdb_midl_free(txn->mt_free_pgs);
2004 free(txn->mt_u.dirty_list);
2008 if (mdb_midl_shrink(&txn->mt_free_pgs))
2009 env->me_free_pgs = txn->mt_free_pgs;
2010 env->me_pghead = NULL;
2014 /* The writer mutex was locked in mdb_txn_begin. */
2015 UNLOCK_MUTEX_W(env);
2020 mdb_txn_reset(MDB_txn *txn)
2025 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2026 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2027 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2029 /* This call is only valid for read-only txns */
2030 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2033 mdb_txn_reset0(txn);
2037 mdb_txn_abort(MDB_txn *txn)
2042 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2043 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2044 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2047 mdb_txn_abort(txn->mt_child);
2049 mdb_txn_reset0(txn);
2050 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2051 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2052 txn->mt_u.reader->mr_pid = 0;
2057 /** Save the freelist as of this transaction to the freeDB.
2058 * This changes the freelist. Keep trying until it stabilizes.
2061 mdb_freelist_save(MDB_txn *txn)
2063 /* env->me_pghead[] can grow and shrink during this call.
2064 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2065 * Page numbers cannot disappear from txn->mt_free_pgs[].
2068 MDB_env *env = txn->mt_env;
2069 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2070 txnid_t pglast = 0, head_id = 0;
2071 pgno_t freecnt = 0, *free_pgs, *mop;
2072 ssize_t head_room = 0, total_room = 0, mop_len;
2074 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2076 if (env->me_pghead) {
2077 /* Make sure first page of freeDB is touched and on freelist */
2078 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2079 if (rc && rc != MDB_NOTFOUND)
2084 /* Come back here after each Put() in case freelist changed */
2087 /* If using records from freeDB which we have not yet
2088 * deleted, delete them and any we reserved for me_pghead.
2090 while (pglast < env->me_pglast) {
2091 rc = mdb_cursor_first(&mc, &key, NULL);
2094 pglast = head_id = *(txnid_t *)key.mv_data;
2095 total_room = head_room = 0;
2096 assert(pglast <= env->me_pglast);
2097 rc = mdb_cursor_del(&mc, 0);
2102 /* Save the IDL of pages freed by this txn, to a single record */
2103 if (freecnt < txn->mt_free_pgs[0]) {
2105 /* Make sure last page of freeDB is touched and on freelist */
2106 key.mv_size = MDB_MAXKEYSIZE+1;
2108 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2109 if (rc && rc != MDB_NOTFOUND)
2112 free_pgs = txn->mt_free_pgs;
2113 /* Write to last page of freeDB */
2114 key.mv_size = sizeof(txn->mt_txnid);
2115 key.mv_data = &txn->mt_txnid;
2117 freecnt = free_pgs[0];
2118 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2119 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2122 /* Retry if mt_free_pgs[] grew during the Put() */
2123 free_pgs = txn->mt_free_pgs;
2124 } while (freecnt < free_pgs[0]);
2125 mdb_midl_sort(free_pgs);
2126 memcpy(data.mv_data, free_pgs, data.mv_size);
2129 unsigned int i = free_pgs[0];
2130 DPRINTF("IDL write txn %zu root %zu num %u",
2131 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2133 DPRINTF("IDL %zu", free_pgs[i]);
2139 mop = env->me_pghead;
2140 mop_len = mop ? mop[0] : 0;
2142 /* Reserve records for me_pghead[]. Split it if multi-page,
2143 * to avoid searching freeDB for a page range. Use keys in
2144 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2146 if (total_room >= mop_len) {
2147 if (total_room == mop_len || --more < 0)
2149 } else if (head_room >= maxfree_1pg && head_id > 1) {
2150 /* Keep current record (overflow page), add a new one */
2154 /* (Re)write {key = head_id, IDL length = head_room} */
2155 total_room -= head_room;
2156 head_room = mop_len - total_room;
2157 if (head_room > maxfree_1pg && head_id > 1) {
2158 /* Overflow multi-page for part of me_pghead */
2159 head_room /= head_id; /* amortize page sizes */
2160 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2161 } else if (head_room < 0) {
2162 /* Rare case, not bothering to delete this record */
2165 key.mv_size = sizeof(head_id);
2166 key.mv_data = &head_id;
2167 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2168 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2171 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2172 total_room += head_room;
2175 /* Fill in the reserved, touched me_pghead records. Avoid write ops
2176 * so they cannot rearrange anything, just read the destinations.
2183 rc = mdb_cursor_first(&mc, &key, &data);
2184 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2185 MDB_IDL dest = data.mv_data;
2186 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2188 assert(len >= 0 && *(txnid_t*)key.mv_data <= env->me_pglast);
2192 memcpy(dest, mop -= len, len * sizeof(MDB_ID));
2193 if (! (mop_len -= len))
2200 /** Flush dirty pages to the map, after clearing their dirty flag.
2203 mdb_page_flush(MDB_txn *txn)
2205 MDB_env *env = txn->mt_env;
2206 MDB_ID2L dl = txn->mt_u.dirty_list;
2207 unsigned psize = env->me_psize;
2208 int i, pagecount = dl[0].mid, rc;
2209 size_t size, pos = 0;
2214 memset(&ov, 0, sizeof(ov));
2216 struct iovec iov[MDB_COMMIT_PAGES];
2217 ssize_t wpos, wsize, wres;
2218 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2222 if (env->me_flags & MDB_WRITEMAP) {
2223 /* Clear dirty flags */
2224 for (i = pagecount; i; i--) {
2226 dp->mp_flags &= ~P_DIRTY;
2232 /* Write the pages */
2234 if (i <= pagecount) {
2237 /* clear dirty flag */
2238 dp->mp_flags &= ~P_DIRTY;
2241 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2246 /* Windows actually supports scatter/gather I/O, but only on
2247 * unbuffered file handles. Since we're relying on the OS page
2248 * cache for all our data, that's self-defeating. So we just
2249 * write pages one at a time. We use the ov structure to set
2250 * the write offset, to at least save the overhead of a Seek
2253 DPRINTF("committing page %zu", pgno);
2254 ov.Offset = pos & 0xffffffff;
2255 ov.OffsetHigh = pos >> 16;
2256 ov.OffsetHigh >>= 16;
2257 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2259 DPRINTF("WriteFile: %d", rc);
2263 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2264 if (pos != next_pos || n == MDB_COMMIT_PAGES) {
2266 /* Write previous page(s) */
2267 lseek(env->me_fd, wpos, SEEK_SET);
2268 wres = writev(env->me_fd, iov, n);
2269 if (wres != wsize) {
2272 DPRINTF("writev: %s", strerror(rc));
2274 DPUTS("short write, filesystem full?");
2285 DPRINTF("committing page %zu", pgno);
2286 next_pos = pos + size;
2287 iov[n].iov_len = size;
2288 iov[n].iov_base = (char *)dp;
2294 mdb_dlist_free(txn);
2300 mdb_txn_commit(MDB_txn *txn)
2306 assert(txn != NULL);
2307 assert(txn->mt_env != NULL);
2309 if (txn->mt_child) {
2310 rc = mdb_txn_commit(txn->mt_child);
2311 txn->mt_child = NULL;
2318 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2319 mdb_dbis_update(txn, 1);
2320 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2325 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2326 DPUTS("error flag is set, can't commit");
2328 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2333 if (txn->mt_parent) {
2334 MDB_txn *parent = txn->mt_parent;
2338 /* Append our free list to parent's */
2339 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2342 mdb_midl_free(txn->mt_free_pgs);
2344 parent->mt_next_pgno = txn->mt_next_pgno;
2345 parent->mt_flags = txn->mt_flags;
2347 /* Merge our cursors into parent's and close them */
2348 mdb_cursors_close(txn, C_SHADOW);
2350 /* Update parent's DB table. */
2351 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2352 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2353 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2354 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2355 for (i=2; i<txn->mt_numdbs; i++) {
2356 /* preserve parent's DB_NEW status */
2357 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2358 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2361 dst = txn->mt_parent->mt_u.dirty_list;
2362 src = txn->mt_u.dirty_list;
2363 /* Find len = length of merging our dirty list with parent's */
2365 dst[0].mid = 0; /* simplify loops */
2366 if (parent->mt_parent) {
2367 len = x + src[0].mid;
2368 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2369 for (i = x; y && i; y--) {
2370 pgno_t yp = src[y].mid;
2371 while (yp < dst[i].mid)
2373 if (yp == dst[i].mid) {
2378 } else { /* Simplify the above for single-ancestor case */
2379 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2381 /* Merge our dirty list with parent's */
2383 for (i = len; y; dst[i--] = src[y--]) {
2384 pgno_t yp = src[y].mid;
2385 while (yp < dst[x].mid)
2386 dst[i--] = dst[x--];
2387 if (yp == dst[x].mid)
2388 free(dst[x--].mptr);
2392 free(txn->mt_u.dirty_list);
2393 parent->mt_dirty_room = txn->mt_dirty_room;
2395 txn->mt_parent->mt_child = NULL;
2396 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2401 if (txn != env->me_txn) {
2402 DPUTS("attempt to commit unknown transaction");
2407 mdb_cursors_close(txn, 0);
2409 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2412 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2413 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2415 /* Update DB root pointers */
2416 if (txn->mt_numdbs > 2) {
2420 data.mv_size = sizeof(MDB_db);
2422 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2423 for (i = 2; i < txn->mt_numdbs; i++) {
2424 if (txn->mt_dbflags[i] & DB_DIRTY) {
2425 data.mv_data = &txn->mt_dbs[i];
2426 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2433 rc = mdb_freelist_save(txn);
2437 mdb_midl_free(env->me_pghead);
2438 env->me_pghead = NULL;
2439 if (mdb_midl_shrink(&txn->mt_free_pgs))
2440 env->me_free_pgs = txn->mt_free_pgs;
2446 if ((rc = mdb_page_flush(txn)) ||
2447 (rc = mdb_env_sync(env, 0)) ||
2448 (rc = mdb_env_write_meta(txn)))
2454 mdb_dbis_update(txn, 1);
2456 UNLOCK_MUTEX_W(env);
2466 /** Read the environment parameters of a DB environment before
2467 * mapping it into memory.
2468 * @param[in] env the environment handle
2469 * @param[out] meta address of where to store the meta information
2470 * @return 0 on success, non-zero on failure.
2473 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2480 /* We don't know the page size yet, so use a minimum value.
2481 * Read both meta pages so we can use the latest one.
2484 for (i=0; i<2; i++) {
2486 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2488 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2493 else if (rc != MDB_PAGESIZE) {
2497 DPRINTF("read: %s", strerror(err));
2501 p = (MDB_page *)&pbuf;
2503 if (!F_ISSET(p->mp_flags, P_META)) {
2504 DPRINTF("page %zu not a meta page", p->mp_pgno);
2509 if (m->mm_magic != MDB_MAGIC) {
2510 DPUTS("meta has invalid magic");
2514 if (m->mm_version != MDB_VERSION) {
2515 DPRINTF("database is version %u, expected version %u",
2516 m->mm_version, MDB_VERSION);
2517 return MDB_VERSION_MISMATCH;
2521 if (m->mm_txnid > meta->mm_txnid)
2522 memcpy(meta, m, sizeof(*m));
2524 memcpy(meta, m, sizeof(*m));
2526 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2528 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2536 /** Write the environment parameters of a freshly created DB environment.
2537 * @param[in] env the environment handle
2538 * @param[out] meta address of where to store the meta information
2539 * @return 0 on success, non-zero on failure.
2542 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2549 DPUTS("writing new meta page");
2551 GET_PAGESIZE(psize);
2553 meta->mm_magic = MDB_MAGIC;
2554 meta->mm_version = MDB_VERSION;
2555 meta->mm_mapsize = env->me_mapsize;
2556 meta->mm_psize = psize;
2557 meta->mm_last_pg = 1;
2558 meta->mm_flags = env->me_flags & 0xffff;
2559 meta->mm_flags |= MDB_INTEGERKEY;
2560 meta->mm_dbs[0].md_root = P_INVALID;
2561 meta->mm_dbs[1].md_root = P_INVALID;
2563 p = calloc(2, psize);
2565 p->mp_flags = P_META;
2568 memcpy(m, meta, sizeof(*meta));
2570 q = (MDB_page *)((char *)p + psize);
2573 q->mp_flags = P_META;
2576 memcpy(m, meta, sizeof(*meta));
2581 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2582 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2583 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2586 lseek(env->me_fd, 0, SEEK_SET);
2587 rc = write(env->me_fd, p, psize * 2);
2588 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2594 /** Update the environment info to commit a transaction.
2595 * @param[in] txn the transaction that's being committed
2596 * @return 0 on success, non-zero on failure.
2599 mdb_env_write_meta(MDB_txn *txn)
2602 MDB_meta meta, metab, *mp;
2604 int rc, len, toggle;
2611 assert(txn != NULL);
2612 assert(txn->mt_env != NULL);
2614 toggle = !txn->mt_toggle;
2615 DPRINTF("writing meta page %d for root page %zu",
2616 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2619 mp = env->me_metas[toggle];
2621 if (env->me_flags & MDB_WRITEMAP) {
2622 /* Persist any increases of mapsize config */
2623 if (env->me_mapsize > mp->mm_mapsize)
2624 mp->mm_mapsize = env->me_mapsize;
2625 mp->mm_dbs[0] = txn->mt_dbs[0];
2626 mp->mm_dbs[1] = txn->mt_dbs[1];
2627 mp->mm_last_pg = txn->mt_next_pgno - 1;
2628 mp->mm_txnid = txn->mt_txnid;
2629 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2630 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2633 ptr += env->me_psize;
2634 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2641 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2642 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2644 ptr = (char *)&meta;
2645 if (env->me_mapsize > mp->mm_mapsize) {
2646 /* Persist any increases of mapsize config */
2647 meta.mm_mapsize = env->me_mapsize;
2648 off = offsetof(MDB_meta, mm_mapsize);
2650 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2652 len = sizeof(MDB_meta) - off;
2655 meta.mm_dbs[0] = txn->mt_dbs[0];
2656 meta.mm_dbs[1] = txn->mt_dbs[1];
2657 meta.mm_last_pg = txn->mt_next_pgno - 1;
2658 meta.mm_txnid = txn->mt_txnid;
2661 off += env->me_psize;
2664 /* Write to the SYNC fd */
2665 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2666 env->me_fd : env->me_mfd;
2669 memset(&ov, 0, sizeof(ov));
2671 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2674 rc = pwrite(mfd, ptr, len, off);
2679 DPUTS("write failed, disk error?");
2680 /* On a failure, the pagecache still contains the new data.
2681 * Write some old data back, to prevent it from being used.
2682 * Use the non-SYNC fd; we know it will fail anyway.
2684 meta.mm_last_pg = metab.mm_last_pg;
2685 meta.mm_txnid = metab.mm_txnid;
2687 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2689 r2 = pwrite(env->me_fd, ptr, len, off);
2692 env->me_flags |= MDB_FATAL_ERROR;
2696 /* Memory ordering issues are irrelevant; since the entire writer
2697 * is wrapped by wmutex, all of these changes will become visible
2698 * after the wmutex is unlocked. Since the DB is multi-version,
2699 * readers will get consistent data regardless of how fresh or
2700 * how stale their view of these values is.
2702 env->me_txns->mti_txnid = txn->mt_txnid;
2707 /** Check both meta pages to see which one is newer.
2708 * @param[in] env the environment handle
2709 * @return meta toggle (0 or 1).
2712 mdb_env_pick_meta(const MDB_env *env)
2714 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2718 mdb_env_create(MDB_env **env)
2722 e = calloc(1, sizeof(MDB_env));
2726 e->me_maxreaders = DEFAULT_READERS;
2727 e->me_maxdbs = e->me_numdbs = 2;
2728 e->me_fd = INVALID_HANDLE_VALUE;
2729 e->me_lfd = INVALID_HANDLE_VALUE;
2730 e->me_mfd = INVALID_HANDLE_VALUE;
2731 #ifdef MDB_USE_POSIX_SEM
2732 e->me_rmutex = SEM_FAILED;
2733 e->me_wmutex = SEM_FAILED;
2735 e->me_pid = getpid();
2736 VGMEMP_CREATE(e,0,0);
2742 mdb_env_set_mapsize(MDB_env *env, size_t size)
2746 env->me_mapsize = size;
2748 env->me_maxpg = env->me_mapsize / env->me_psize;
2753 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2757 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2762 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2764 if (env->me_map || readers < 1)
2766 env->me_maxreaders = readers;
2771 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2773 if (!env || !readers)
2775 *readers = env->me_maxreaders;
2779 /** Further setup required for opening an MDB environment
2782 mdb_env_open2(MDB_env *env)
2784 unsigned int flags = env->me_flags;
2785 int i, newenv = 0, prot;
2789 memset(&meta, 0, sizeof(meta));
2791 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2794 DPUTS("new mdbenv");
2798 /* Was a mapsize configured? */
2799 if (!env->me_mapsize) {
2800 /* If this is a new environment, take the default,
2801 * else use the size recorded in the existing env.
2803 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2804 } else if (env->me_mapsize < meta.mm_mapsize) {
2805 /* If the configured size is smaller, make sure it's
2806 * still big enough. Silently round up to minimum if not.
2808 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2809 if (env->me_mapsize < minsize)
2810 env->me_mapsize = minsize;
2816 LONG sizelo, sizehi;
2817 sizelo = env->me_mapsize & 0xffffffff;
2818 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2820 /* Windows won't create mappings for zero length files.
2821 * Just allocate the maxsize right now.
2824 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2825 if (!SetEndOfFile(env->me_fd))
2827 SetFilePointer(env->me_fd, 0, NULL, 0);
2829 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2830 PAGE_READWRITE : PAGE_READONLY,
2831 sizehi, sizelo, NULL);
2834 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2835 FILE_MAP_WRITE : FILE_MAP_READ,
2836 0, 0, env->me_mapsize, meta.mm_address);
2844 if (flags & MDB_WRITEMAP) {
2846 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2849 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2851 if (env->me_map == MAP_FAILED) {
2855 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2857 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2859 #ifdef POSIX_MADV_RANDOM
2860 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2861 #endif /* POSIX_MADV_RANDOM */
2862 #endif /* MADV_RANDOM */
2866 if (flags & MDB_FIXEDMAP)
2867 meta.mm_address = env->me_map;
2868 i = mdb_env_init_meta(env, &meta);
2869 if (i != MDB_SUCCESS) {
2872 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2873 /* Can happen because the address argument to mmap() is just a
2874 * hint. mmap() can pick another, e.g. if the range is in use.
2875 * The MAP_FIXED flag would prevent that, but then mmap could
2876 * instead unmap existing pages to make room for the new map.
2878 return EBUSY; /* TODO: Make a new MDB_* error code? */
2880 env->me_psize = meta.mm_psize;
2881 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2882 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2884 env->me_maxpg = env->me_mapsize / env->me_psize;
2886 p = (MDB_page *)env->me_map;
2887 env->me_metas[0] = METADATA(p);
2888 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2892 int toggle = mdb_env_pick_meta(env);
2893 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2895 DPRINTF("opened database version %u, pagesize %u",
2896 env->me_metas[0]->mm_version, env->me_psize);
2897 DPRINTF("using meta page %d", toggle);
2898 DPRINTF("depth: %u", db->md_depth);
2899 DPRINTF("entries: %zu", db->md_entries);
2900 DPRINTF("branch pages: %zu", db->md_branch_pages);
2901 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2902 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2903 DPRINTF("root: %zu", db->md_root);
2911 /** Release a reader thread's slot in the reader lock table.
2912 * This function is called automatically when a thread exits.
2913 * @param[in] ptr This points to the slot in the reader lock table.
2916 mdb_env_reader_dest(void *ptr)
2918 MDB_reader *reader = ptr;
2924 /** Junk for arranging thread-specific callbacks on Windows. This is
2925 * necessarily platform and compiler-specific. Windows supports up
2926 * to 1088 keys. Let's assume nobody opens more than 64 environments
2927 * in a single process, for now. They can override this if needed.
2929 #ifndef MAX_TLS_KEYS
2930 #define MAX_TLS_KEYS 64
2932 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2933 static int mdb_tls_nkeys;
2935 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2939 case DLL_PROCESS_ATTACH: break;
2940 case DLL_THREAD_ATTACH: break;
2941 case DLL_THREAD_DETACH:
2942 for (i=0; i<mdb_tls_nkeys; i++) {
2943 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2944 mdb_env_reader_dest(r);
2947 case DLL_PROCESS_DETACH: break;
2952 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2954 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2958 /* Force some symbol references.
2959 * _tls_used forces the linker to create the TLS directory if not already done
2960 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2962 #pragma comment(linker, "/INCLUDE:_tls_used")
2963 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2964 #pragma const_seg(".CRT$XLB")
2965 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2966 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2969 #pragma comment(linker, "/INCLUDE:__tls_used")
2970 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2971 #pragma data_seg(".CRT$XLB")
2972 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2974 #endif /* WIN 32/64 */
2975 #endif /* !__GNUC__ */
2978 /** Downgrade the exclusive lock on the region back to shared */
2980 mdb_env_share_locks(MDB_env *env, int *excl)
2982 int rc = 0, toggle = mdb_env_pick_meta(env);
2984 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2989 /* First acquire a shared lock. The Unlock will
2990 * then release the existing exclusive lock.
2992 memset(&ov, 0, sizeof(ov));
2993 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2996 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3002 struct flock lock_info;
3003 /* The shared lock replaces the existing lock */
3004 memset((void *)&lock_info, 0, sizeof(lock_info));
3005 lock_info.l_type = F_RDLCK;
3006 lock_info.l_whence = SEEK_SET;
3007 lock_info.l_start = 0;
3008 lock_info.l_len = 1;
3009 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3010 (rc = ErrCode()) == EINTR) ;
3011 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3018 /** Try to get exlusive lock, otherwise shared.
3019 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3022 mdb_env_excl_lock(MDB_env *env, int *excl)
3026 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3030 memset(&ov, 0, sizeof(ov));
3031 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3038 struct flock lock_info;
3039 memset((void *)&lock_info, 0, sizeof(lock_info));
3040 lock_info.l_type = F_WRLCK;
3041 lock_info.l_whence = SEEK_SET;
3042 lock_info.l_start = 0;
3043 lock_info.l_len = 1;
3044 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3045 (rc = ErrCode()) == EINTR) ;
3049 # ifdef MDB_USE_POSIX_SEM
3050 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3053 lock_info.l_type = F_RDLCK;
3054 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3055 (rc = ErrCode()) == EINTR) ;
3063 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3065 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3067 * @(#) $Revision: 5.1 $
3068 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3069 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3071 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3075 * Please do not copyright this code. This code is in the public domain.
3077 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3078 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3079 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3080 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3081 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3082 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3083 * PERFORMANCE OF THIS SOFTWARE.
3086 * chongo <Landon Curt Noll> /\oo/\
3087 * http://www.isthe.com/chongo/
3089 * Share and Enjoy! :-)
3092 typedef unsigned long long mdb_hash_t;
3093 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3095 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3096 * @param[in] str string to hash
3097 * @param[in] hval initial value for hash
3098 * @return 64 bit hash
3100 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3101 * hval arg on the first call.
3104 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3106 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3107 unsigned char *end = s + val->mv_size;
3109 * FNV-1a hash each octet of the string
3112 /* xor the bottom with the current octet */
3113 hval ^= (mdb_hash_t)*s++;
3115 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3116 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3117 (hval << 7) + (hval << 8) + (hval << 40);
3119 /* return our new hash value */
3123 /** Hash the string and output the hash in hex.
3124 * @param[in] str string to hash
3125 * @param[out] hexbuf an array of 17 chars to hold the hash
3128 mdb_hash_hex(MDB_val *val, char *hexbuf)
3131 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3132 for (i=0; i<8; i++) {
3133 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3139 /** Open and/or initialize the lock region for the environment.
3140 * @param[in] env The MDB environment.
3141 * @param[in] lpath The pathname of the file used for the lock region.
3142 * @param[in] mode The Unix permissions for the file, if we create it.
3143 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3144 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3145 * @return 0 on success, non-zero on failure.
3148 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3151 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3153 # define MDB_ERRCODE_ROFS EROFS
3154 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3155 # define MDB_CLOEXEC O_CLOEXEC
3158 # define MDB_CLOEXEC 0
3165 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3166 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3167 FILE_ATTRIBUTE_NORMAL, NULL);
3169 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3171 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3173 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3178 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3179 /* Lose record locks when exec*() */
3180 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3181 fcntl(env->me_lfd, F_SETFD, fdflags);
3184 if (!(env->me_flags & MDB_NOTLS)) {
3185 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3188 env->me_flags |= MDB_ENV_TXKEY;
3190 /* Windows TLS callbacks need help finding their TLS info. */
3191 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3195 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3199 /* Try to get exclusive lock. If we succeed, then
3200 * nobody is using the lock region and we should initialize it.
3202 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3205 size = GetFileSize(env->me_lfd, NULL);
3207 size = lseek(env->me_lfd, 0, SEEK_END);
3209 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3210 if (size < rsize && *excl > 0) {
3212 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3213 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3215 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3219 size = rsize - sizeof(MDB_txninfo);
3220 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3225 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3227 if (!mh) goto fail_errno;
3228 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3230 if (!env->me_txns) goto fail_errno;
3232 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3234 if (m == MAP_FAILED) goto fail_errno;
3240 BY_HANDLE_FILE_INFORMATION stbuf;
3249 if (!mdb_sec_inited) {
3250 InitializeSecurityDescriptor(&mdb_null_sd,
3251 SECURITY_DESCRIPTOR_REVISION);
3252 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3253 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3254 mdb_all_sa.bInheritHandle = FALSE;
3255 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3258 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3259 idbuf.volume = stbuf.dwVolumeSerialNumber;
3260 idbuf.nhigh = stbuf.nFileIndexHigh;
3261 idbuf.nlow = stbuf.nFileIndexLow;
3262 val.mv_data = &idbuf;
3263 val.mv_size = sizeof(idbuf);
3264 mdb_hash_hex(&val, hexbuf);
3265 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3266 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3267 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3268 if (!env->me_rmutex) goto fail_errno;
3269 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3270 if (!env->me_wmutex) goto fail_errno;
3271 #elif defined(MDB_USE_POSIX_SEM)
3280 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3281 idbuf.dev = stbuf.st_dev;
3282 idbuf.ino = stbuf.st_ino;
3283 val.mv_data = &idbuf;
3284 val.mv_size = sizeof(idbuf);
3285 mdb_hash_hex(&val, hexbuf);
3286 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3287 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3288 /* Clean up after a previous run, if needed: Try to
3289 * remove both semaphores before doing anything else.
3291 sem_unlink(env->me_txns->mti_rmname);
3292 sem_unlink(env->me_txns->mti_wmname);
3293 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3294 O_CREAT|O_EXCL, mode, 1);
3295 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3296 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3297 O_CREAT|O_EXCL, mode, 1);
3298 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3299 #else /* MDB_USE_POSIX_SEM */
3300 pthread_mutexattr_t mattr;
3302 if ((rc = pthread_mutexattr_init(&mattr))
3303 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3304 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3305 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3307 pthread_mutexattr_destroy(&mattr);
3308 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3310 env->me_txns->mti_version = MDB_VERSION;
3311 env->me_txns->mti_magic = MDB_MAGIC;
3312 env->me_txns->mti_txnid = 0;
3313 env->me_txns->mti_numreaders = 0;
3316 if (env->me_txns->mti_magic != MDB_MAGIC) {
3317 DPUTS("lock region has invalid magic");
3321 if (env->me_txns->mti_version != MDB_VERSION) {
3322 DPRINTF("lock region is version %u, expected version %u",
3323 env->me_txns->mti_version, MDB_VERSION);
3324 rc = MDB_VERSION_MISMATCH;
3328 if (rc != EACCES && rc != EAGAIN) {
3332 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3333 if (!env->me_rmutex) goto fail_errno;
3334 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3335 if (!env->me_wmutex) goto fail_errno;
3336 #elif defined(MDB_USE_POSIX_SEM)
3337 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3338 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3339 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3340 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3351 /** The name of the lock file in the DB environment */
3352 #define LOCKNAME "/lock.mdb"
3353 /** The name of the data file in the DB environment */
3354 #define DATANAME "/data.mdb"
3355 /** The suffix of the lock file when no subdir is used */
3356 #define LOCKSUFF "-lock"
3357 /** Only a subset of the @ref mdb_env flags can be changed
3358 * at runtime. Changing other flags requires closing the
3359 * environment and re-opening it with the new flags.
3361 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3362 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3365 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3367 int oflags, rc, len, excl = -1;
3368 char *lpath, *dpath;
3370 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3374 if (flags & MDB_NOSUBDIR) {
3375 rc = len + sizeof(LOCKSUFF) + len + 1;
3377 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3382 if (flags & MDB_NOSUBDIR) {
3383 dpath = lpath + len + sizeof(LOCKSUFF);
3384 sprintf(lpath, "%s" LOCKSUFF, path);
3385 strcpy(dpath, path);
3387 dpath = lpath + len + sizeof(LOCKNAME);
3388 sprintf(lpath, "%s" LOCKNAME, path);
3389 sprintf(dpath, "%s" DATANAME, path);
3393 flags |= env->me_flags;
3394 if (flags & MDB_RDONLY) {
3395 /* silently ignore WRITEMAP when we're only getting read access */
3396 flags &= ~MDB_WRITEMAP;
3398 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3399 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3402 env->me_flags = flags |= MDB_ENV_ACTIVE;
3406 env->me_path = strdup(path);
3407 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3408 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3409 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3414 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3419 if (F_ISSET(flags, MDB_RDONLY)) {
3420 oflags = GENERIC_READ;
3421 len = OPEN_EXISTING;
3423 oflags = GENERIC_READ|GENERIC_WRITE;
3426 mode = FILE_ATTRIBUTE_NORMAL;
3427 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3428 NULL, len, mode, NULL);
3430 if (F_ISSET(flags, MDB_RDONLY))
3433 oflags = O_RDWR | O_CREAT;
3435 env->me_fd = open(dpath, oflags, mode);
3437 if (env->me_fd == INVALID_HANDLE_VALUE) {
3442 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3443 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3444 env->me_mfd = env->me_fd;
3446 /* Synchronous fd for meta writes. Needed even with
3447 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3450 env->me_mfd = CreateFile(dpath, oflags,
3451 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3452 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3454 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3456 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3461 DPRINTF("opened dbenv %p", (void *) env);
3463 rc = mdb_env_share_locks(env, &excl);
3469 mdb_env_close0(env, excl);
3475 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3477 mdb_env_close0(MDB_env *env, int excl)
3481 if (!(env->me_flags & MDB_ENV_ACTIVE))
3484 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3485 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3486 free(env->me_dbxs[i].md_name.mv_data);
3488 free(env->me_dbflags);
3491 free(env->me_dirty_list);
3492 mdb_midl_free(env->me_free_pgs);
3494 if (env->me_flags & MDB_ENV_TXKEY) {
3495 pthread_key_delete(env->me_txkey);
3497 /* Delete our key from the global list */
3498 for (i=0; i<mdb_tls_nkeys; i++)
3499 if (mdb_tls_keys[i] == env->me_txkey) {
3500 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3508 munmap(env->me_map, env->me_mapsize);
3510 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3512 if (env->me_fd != INVALID_HANDLE_VALUE)
3515 pid_t pid = env->me_pid;
3516 /* Clearing readers is done in this function because
3517 * me_txkey with its destructor must be disabled first.
3519 for (i = env->me_numreaders; --i >= 0; )
3520 if (env->me_txns->mti_readers[i].mr_pid == pid)
3521 env->me_txns->mti_readers[i].mr_pid = 0;
3523 if (env->me_rmutex) {
3524 CloseHandle(env->me_rmutex);
3525 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3527 /* Windows automatically destroys the mutexes when
3528 * the last handle closes.
3530 #elif defined(MDB_USE_POSIX_SEM)
3531 if (env->me_rmutex != SEM_FAILED) {
3532 sem_close(env->me_rmutex);
3533 if (env->me_wmutex != SEM_FAILED)
3534 sem_close(env->me_wmutex);
3535 /* If we have the filelock: If we are the
3536 * only remaining user, clean up semaphores.
3539 mdb_env_excl_lock(env, &excl);
3541 sem_unlink(env->me_txns->mti_rmname);
3542 sem_unlink(env->me_txns->mti_wmname);
3546 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3548 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3551 /* Unlock the lockfile. Windows would have unlocked it
3552 * after closing anyway, but not necessarily at once.
3554 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3560 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3564 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3566 MDB_txn *txn = NULL;
3571 /* Do the lock/unlock of the reader mutex before starting the
3572 * write txn. Otherwise other read txns could block writers.
3574 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3579 /* We must start the actual read txn after blocking writers */
3580 mdb_txn_reset0(txn);
3582 /* Temporarily block writers until we snapshot the meta pages */
3585 rc = mdb_txn_renew0(txn);
3587 UNLOCK_MUTEX_W(env);
3592 wsize = env->me_psize * 2;
3596 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3597 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3600 rc = write(fd, env->me_map, wsize);
3601 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3604 UNLOCK_MUTEX_W(env);
3609 ptr = env->me_map + wsize;
3610 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3611 #define MAX_WRITE 2147483648U
3615 if (wsize > MAX_WRITE)
3619 rc = WriteFile(fd, ptr, w2, &len, NULL);
3620 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3629 if (wsize > MAX_WRITE)
3633 wres = write(fd, ptr, w2);
3634 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3647 mdb_env_copy(MDB_env *env, const char *path)
3651 HANDLE newfd = INVALID_HANDLE_VALUE;
3653 if (env->me_flags & MDB_NOSUBDIR) {
3654 lpath = (char *)path;
3657 len += sizeof(DATANAME);
3658 lpath = malloc(len);
3661 sprintf(lpath, "%s" DATANAME, path);
3664 /* The destination path must exist, but the destination file must not.
3665 * We don't want the OS to cache the writes, since the source data is
3666 * already in the OS cache.
3669 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3670 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3672 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3678 if (!(env->me_flags & MDB_NOSUBDIR))
3680 if (newfd == INVALID_HANDLE_VALUE) {
3685 #ifdef F_NOCACHE /* __APPLE__ */
3686 rc = fcntl(newfd, F_NOCACHE, 1);
3693 rc = mdb_env_copyfd(env, newfd);
3696 if (newfd != INVALID_HANDLE_VALUE)
3703 mdb_env_close(MDB_env *env)
3710 VGMEMP_DESTROY(env);
3711 while ((dp = env->me_dpages) != NULL) {
3712 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3713 env->me_dpages = dp->mp_next;
3717 mdb_env_close0(env, 0);
3721 /** Compare two items pointing at aligned size_t's */
3723 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3725 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3726 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3729 /** Compare two items pointing at aligned int's */
3731 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3733 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3734 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3737 /** Compare two items pointing at ints of unknown alignment.
3738 * Nodes and keys are guaranteed to be 2-byte aligned.
3741 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3743 #if BYTE_ORDER == LITTLE_ENDIAN
3744 unsigned short *u, *c;
3747 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3748 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3751 } while(!x && u > (unsigned short *)a->mv_data);
3754 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3758 /** Compare two items lexically */
3760 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3767 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3773 diff = memcmp(a->mv_data, b->mv_data, len);
3774 return diff ? diff : len_diff<0 ? -1 : len_diff;
3777 /** Compare two items in reverse byte order */
3779 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3781 const unsigned char *p1, *p2, *p1_lim;
3785 p1_lim = (const unsigned char *)a->mv_data;
3786 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3787 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3789 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3795 while (p1 > p1_lim) {
3796 diff = *--p1 - *--p2;
3800 return len_diff<0 ? -1 : len_diff;
3803 /** Search for key within a page, using binary search.
3804 * Returns the smallest entry larger or equal to the key.
3805 * If exactp is non-null, stores whether the found entry was an exact match
3806 * in *exactp (1 or 0).
3807 * Updates the cursor index with the index of the found entry.
3808 * If no entry larger or equal to the key is found, returns NULL.
3811 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3813 unsigned int i = 0, nkeys;
3816 MDB_page *mp = mc->mc_pg[mc->mc_top];
3817 MDB_node *node = NULL;
3822 nkeys = NUMKEYS(mp);
3827 COPY_PGNO(pgno, mp->mp_pgno);
3828 DPRINTF("searching %u keys in %s %spage %zu",
3829 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3836 low = IS_LEAF(mp) ? 0 : 1;
3838 cmp = mc->mc_dbx->md_cmp;
3840 /* Branch pages have no data, so if using integer keys,
3841 * alignment is guaranteed. Use faster mdb_cmp_int.
3843 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3844 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3851 nodekey.mv_size = mc->mc_db->md_pad;
3852 node = NODEPTR(mp, 0); /* fake */
3853 while (low <= high) {
3854 i = (low + high) >> 1;
3855 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3856 rc = cmp(key, &nodekey);
3857 DPRINTF("found leaf index %u [%s], rc = %i",
3858 i, DKEY(&nodekey), rc);
3867 while (low <= high) {
3868 i = (low + high) >> 1;
3870 node = NODEPTR(mp, i);
3871 nodekey.mv_size = NODEKSZ(node);
3872 nodekey.mv_data = NODEKEY(node);
3874 rc = cmp(key, &nodekey);
3877 DPRINTF("found leaf index %u [%s], rc = %i",
3878 i, DKEY(&nodekey), rc);
3880 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3881 i, DKEY(&nodekey), NODEPGNO(node), rc);
3892 if (rc > 0) { /* Found entry is less than the key. */
3893 i++; /* Skip to get the smallest entry larger than key. */
3895 node = NODEPTR(mp, i);
3898 *exactp = (rc == 0);
3899 /* store the key index */
3900 mc->mc_ki[mc->mc_top] = i;
3902 /* There is no entry larger or equal to the key. */
3905 /* nodeptr is fake for LEAF2 */
3911 mdb_cursor_adjust(MDB_cursor *mc, func)
3915 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3916 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3923 /** Pop a page off the top of the cursor's stack. */
3925 mdb_cursor_pop(MDB_cursor *mc)
3928 #ifndef MDB_DEBUG_SKIP
3929 MDB_page *top = mc->mc_pg[mc->mc_top];
3935 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3936 mc->mc_dbi, (void *) mc);
3940 /** Push a page onto the top of the cursor's stack. */
3942 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3944 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3945 mc->mc_dbi, (void *) mc);
3947 if (mc->mc_snum >= CURSOR_STACK) {
3948 assert(mc->mc_snum < CURSOR_STACK);
3949 return MDB_CURSOR_FULL;
3952 mc->mc_top = mc->mc_snum++;
3953 mc->mc_pg[mc->mc_top] = mp;
3954 mc->mc_ki[mc->mc_top] = 0;
3959 /** Find the address of the page corresponding to a given page number.
3960 * @param[in] txn the transaction for this access.
3961 * @param[in] pgno the page number for the page to retrieve.
3962 * @param[out] ret address of a pointer where the page's address will be stored.
3963 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
3964 * @return 0 on success, non-zero on failure.
3967 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
3972 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3973 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3978 MDB_ID2L dl = tx2->mt_u.dirty_list;
3980 unsigned x = mdb_mid2l_search(dl, pgno);
3981 if (x <= dl[0].mid && dl[x].mid == pgno) {
3987 } while ((tx2 = tx2->mt_parent) != NULL);
3990 if (pgno < txn->mt_next_pgno) {
3992 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3994 DPRINTF("page %zu not found", pgno);
3996 return MDB_PAGE_NOTFOUND;
4006 /** Search for the page a given key should be in.
4007 * Pushes parent pages on the cursor stack. This function continues a
4008 * search on a cursor that has already been initialized. (Usually by
4009 * #mdb_page_search() but also by #mdb_node_move().)
4010 * @param[in,out] mc the cursor for this operation.
4011 * @param[in] key the key to search for. If NULL, search for the lowest
4012 * page. (This is used by #mdb_cursor_first().)
4013 * @param[in] modify If true, visited pages are updated with new page numbers.
4014 * @return 0 on success, non-zero on failure.
4017 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4019 MDB_page *mp = mc->mc_pg[mc->mc_top];
4024 while (IS_BRANCH(mp)) {
4028 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4029 assert(NUMKEYS(mp) > 1);
4030 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4032 if (key == NULL) /* Initialize cursor to first page. */
4034 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4035 /* cursor to last page */
4039 node = mdb_node_search(mc, key, &exact);
4041 i = NUMKEYS(mp) - 1;
4043 i = mc->mc_ki[mc->mc_top];
4052 DPRINTF("following index %u for key [%s]",
4054 assert(i < NUMKEYS(mp));
4055 node = NODEPTR(mp, i);
4057 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4060 mc->mc_ki[mc->mc_top] = i;
4061 if ((rc = mdb_cursor_push(mc, mp)))
4065 if ((rc = mdb_page_touch(mc)) != 0)
4067 mp = mc->mc_pg[mc->mc_top];
4072 DPRINTF("internal error, index points to a %02X page!?",
4074 return MDB_CORRUPTED;
4077 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4078 key ? DKEY(key) : NULL);
4083 /** Search for the lowest key under the current branch page.
4084 * This just bypasses a NUMKEYS check in the current page
4085 * before calling mdb_page_search_root(), because the callers
4086 * are all in situations where the current page is known to
4090 mdb_page_search_lowest(MDB_cursor *mc)
4092 MDB_page *mp = mc->mc_pg[mc->mc_top];
4093 MDB_node *node = NODEPTR(mp, 0);
4096 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4099 mc->mc_ki[mc->mc_top] = 0;
4100 if ((rc = mdb_cursor_push(mc, mp)))
4102 return mdb_page_search_root(mc, NULL, 0);
4105 /** Search for the page a given key should be in.
4106 * Pushes parent pages on the cursor stack. This function just sets up
4107 * the search; it finds the root page for \b mc's database and sets this
4108 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4109 * called to complete the search.
4110 * @param[in,out] mc the cursor for this operation.
4111 * @param[in] key the key to search for. If NULL, search for the lowest
4112 * page. (This is used by #mdb_cursor_first().)
4113 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4114 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4115 * @return 0 on success, non-zero on failure.
4118 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4123 /* Make sure the txn is still viable, then find the root from
4124 * the txn's db table.
4126 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4127 DPUTS("transaction has failed, must abort");
4130 /* Make sure we're using an up-to-date root */
4131 if (mc->mc_dbi > MAIN_DBI) {
4132 if ((*mc->mc_dbflag & DB_STALE) ||
4133 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4135 unsigned char dbflag = 0;
4136 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4137 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4140 if (*mc->mc_dbflag & DB_STALE) {
4144 MDB_node *leaf = mdb_node_search(&mc2,
4145 &mc->mc_dbx->md_name, &exact);
4147 return MDB_NOTFOUND;
4148 mdb_node_read(mc->mc_txn, leaf, &data);
4149 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4151 /* The txn may not know this DBI, or another process may
4152 * have dropped and recreated the DB with other flags.
4154 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4155 return MDB_INCOMPATIBLE;
4156 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4158 if (flags & MDB_PS_MODIFY)
4160 *mc->mc_dbflag &= ~DB_STALE;
4161 *mc->mc_dbflag |= dbflag;
4164 root = mc->mc_db->md_root;
4166 if (root == P_INVALID) { /* Tree is empty. */
4167 DPUTS("tree is empty");
4168 return MDB_NOTFOUND;
4173 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4174 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4180 DPRINTF("db %u root page %zu has flags 0x%X",
4181 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4183 if (flags & MDB_PS_MODIFY) {
4184 if ((rc = mdb_page_touch(mc)))
4188 if (flags & MDB_PS_ROOTONLY)
4191 return mdb_page_search_root(mc, key, flags);
4195 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4197 MDB_txn *txn = mc->mc_txn;
4198 pgno_t pg = mp->mp_pgno;
4199 unsigned i, ovpages = mp->mp_pages;
4202 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4203 /* If the page is dirty we just acquired it, so we should
4204 * give it back to our current free list, if any.
4205 * Not currently supported in nested txns.
4206 * Otherwise put it onto the list of pages we freed in this txn.
4208 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && txn->mt_env->me_pghead) {
4210 pgno_t *mop = txn->mt_env->me_pghead;
4211 MDB_ID2 *dl, ix, iy;
4212 /* Prepare to insert pg */
4213 j = mop[0] + ovpages;
4215 rc = mdb_midl_grow(&mop, ovpages);
4218 txn->mt_env->me_pghead = mop;
4220 /* Remove from dirty list */
4221 dl = txn->mt_u.dirty_list;
4223 for (ix = dl[x]; ix.mid != pg; ix = iy) {
4230 return MDB_CORRUPTED;
4233 /* Insert in me_pghead */
4234 for (i = mop[0]; i && mop[i] < pg; i--)
4240 for (i=0; i<ovpages; i++) {
4241 mdb_midl_append(&txn->mt_free_pgs, pg);
4245 mc->mc_db->md_overflow_pages -= ovpages;
4249 /** Return the data associated with a given node.
4250 * @param[in] txn The transaction for this operation.
4251 * @param[in] leaf The node being read.
4252 * @param[out] data Updated to point to the node's data.
4253 * @return 0 on success, non-zero on failure.
4256 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4258 MDB_page *omp; /* overflow page */
4262 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4263 data->mv_size = NODEDSZ(leaf);
4264 data->mv_data = NODEDATA(leaf);
4268 /* Read overflow data.
4270 data->mv_size = NODEDSZ(leaf);
4271 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4272 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4273 DPRINTF("read overflow page %zu failed", pgno);
4276 data->mv_data = METADATA(omp);
4282 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4283 MDB_val *key, MDB_val *data)
4292 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4294 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4297 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4301 mdb_cursor_init(&mc, txn, dbi, &mx);
4302 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4305 /** Find a sibling for a page.
4306 * Replaces the page at the top of the cursor's stack with the
4307 * specified sibling, if one exists.
4308 * @param[in] mc The cursor for this operation.
4309 * @param[in] move_right Non-zero if the right sibling is requested,
4310 * otherwise the left sibling.
4311 * @return 0 on success, non-zero on failure.
4314 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4320 if (mc->mc_snum < 2) {
4321 return MDB_NOTFOUND; /* root has no siblings */
4325 DPRINTF("parent page is page %zu, index %u",
4326 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4328 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4329 : (mc->mc_ki[mc->mc_top] == 0)) {
4330 DPRINTF("no more keys left, moving to %s sibling",
4331 move_right ? "right" : "left");
4332 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4333 /* undo cursor_pop before returning */
4340 mc->mc_ki[mc->mc_top]++;
4342 mc->mc_ki[mc->mc_top]--;
4343 DPRINTF("just moving to %s index key %u",
4344 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4346 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4348 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4349 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4352 mdb_cursor_push(mc, mp);
4354 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4359 /** Move the cursor to the next data item. */
4361 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4367 if (mc->mc_flags & C_EOF) {
4368 return MDB_NOTFOUND;
4371 assert(mc->mc_flags & C_INITIALIZED);
4373 mp = mc->mc_pg[mc->mc_top];
4375 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4376 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4377 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4378 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4379 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4380 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4384 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4385 if (op == MDB_NEXT_DUP)
4386 return MDB_NOTFOUND;
4390 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4392 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4393 DPUTS("=====> move to next sibling page");
4394 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4395 mc->mc_flags |= C_EOF;
4396 mc->mc_flags &= ~C_INITIALIZED;
4397 return MDB_NOTFOUND;
4399 mp = mc->mc_pg[mc->mc_top];
4400 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4402 mc->mc_ki[mc->mc_top]++;
4404 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4405 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4408 key->mv_size = mc->mc_db->md_pad;
4409 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4413 assert(IS_LEAF(mp));
4414 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4416 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4417 mdb_xcursor_init1(mc, leaf);
4420 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4423 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4424 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4425 if (rc != MDB_SUCCESS)
4430 MDB_GET_KEY(leaf, key);
4434 /** Move the cursor to the previous data item. */
4436 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4442 assert(mc->mc_flags & C_INITIALIZED);
4444 mp = mc->mc_pg[mc->mc_top];
4446 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4447 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4448 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4449 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4450 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4451 if (op != MDB_PREV || rc == MDB_SUCCESS)
4454 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4455 if (op == MDB_PREV_DUP)
4456 return MDB_NOTFOUND;
4461 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4463 if (mc->mc_ki[mc->mc_top] == 0) {
4464 DPUTS("=====> move to prev sibling page");
4465 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4466 mc->mc_flags &= ~C_INITIALIZED;
4467 return MDB_NOTFOUND;
4469 mp = mc->mc_pg[mc->mc_top];
4470 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4471 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4473 mc->mc_ki[mc->mc_top]--;
4475 mc->mc_flags &= ~C_EOF;
4477 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4478 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4481 key->mv_size = mc->mc_db->md_pad;
4482 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4486 assert(IS_LEAF(mp));
4487 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4489 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4490 mdb_xcursor_init1(mc, leaf);
4493 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4496 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4497 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4498 if (rc != MDB_SUCCESS)
4503 MDB_GET_KEY(leaf, key);
4507 /** Set the cursor on a specific data item. */
4509 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4510 MDB_cursor_op op, int *exactp)
4514 MDB_node *leaf = NULL;
4519 assert(key->mv_size > 0);
4521 /* See if we're already on the right page */
4522 if (mc->mc_flags & C_INITIALIZED) {
4525 mp = mc->mc_pg[mc->mc_top];
4527 mc->mc_ki[mc->mc_top] = 0;
4528 return MDB_NOTFOUND;
4530 if (mp->mp_flags & P_LEAF2) {
4531 nodekey.mv_size = mc->mc_db->md_pad;
4532 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4534 leaf = NODEPTR(mp, 0);
4535 MDB_GET_KEY(leaf, &nodekey);
4537 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4539 /* Probably happens rarely, but first node on the page
4540 * was the one we wanted.
4542 mc->mc_ki[mc->mc_top] = 0;
4549 unsigned int nkeys = NUMKEYS(mp);
4551 if (mp->mp_flags & P_LEAF2) {
4552 nodekey.mv_data = LEAF2KEY(mp,
4553 nkeys-1, nodekey.mv_size);
4555 leaf = NODEPTR(mp, nkeys-1);
4556 MDB_GET_KEY(leaf, &nodekey);
4558 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4560 /* last node was the one we wanted */
4561 mc->mc_ki[mc->mc_top] = nkeys-1;
4567 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4568 /* This is definitely the right page, skip search_page */
4569 if (mp->mp_flags & P_LEAF2) {
4570 nodekey.mv_data = LEAF2KEY(mp,
4571 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4573 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4574 MDB_GET_KEY(leaf, &nodekey);
4576 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4578 /* current node was the one we wanted */
4588 /* If any parents have right-sibs, search.
4589 * Otherwise, there's nothing further.
4591 for (i=0; i<mc->mc_top; i++)
4593 NUMKEYS(mc->mc_pg[i])-1)
4595 if (i == mc->mc_top) {
4596 /* There are no other pages */
4597 mc->mc_ki[mc->mc_top] = nkeys;
4598 return MDB_NOTFOUND;
4602 /* There are no other pages */
4603 mc->mc_ki[mc->mc_top] = 0;
4604 return MDB_NOTFOUND;
4608 rc = mdb_page_search(mc, key, 0);
4609 if (rc != MDB_SUCCESS)
4612 mp = mc->mc_pg[mc->mc_top];
4613 assert(IS_LEAF(mp));
4616 leaf = mdb_node_search(mc, key, exactp);
4617 if (exactp != NULL && !*exactp) {
4618 /* MDB_SET specified and not an exact match. */
4619 return MDB_NOTFOUND;
4623 DPUTS("===> inexact leaf not found, goto sibling");
4624 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4625 return rc; /* no entries matched */
4626 mp = mc->mc_pg[mc->mc_top];
4627 assert(IS_LEAF(mp));
4628 leaf = NODEPTR(mp, 0);
4632 mc->mc_flags |= C_INITIALIZED;
4633 mc->mc_flags &= ~C_EOF;
4636 key->mv_size = mc->mc_db->md_pad;
4637 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4641 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4642 mdb_xcursor_init1(mc, leaf);
4645 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4646 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4647 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4650 if (op == MDB_GET_BOTH) {
4656 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4657 if (rc != MDB_SUCCESS)
4660 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4662 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4664 rc = mc->mc_dbx->md_dcmp(data, &d2);
4666 if (op == MDB_GET_BOTH || rc > 0)
4667 return MDB_NOTFOUND;
4672 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4673 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4678 /* The key already matches in all other cases */
4679 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4680 MDB_GET_KEY(leaf, key);
4681 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4686 /** Move the cursor to the first item in the database. */
4688 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4693 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4694 rc = mdb_page_search(mc, NULL, 0);
4695 if (rc != MDB_SUCCESS)
4698 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4700 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4701 mc->mc_flags |= C_INITIALIZED;
4702 mc->mc_flags &= ~C_EOF;
4704 mc->mc_ki[mc->mc_top] = 0;
4706 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4707 key->mv_size = mc->mc_db->md_pad;
4708 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4713 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4714 mdb_xcursor_init1(mc, leaf);
4715 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4720 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4721 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4725 MDB_GET_KEY(leaf, key);
4729 /** Move the cursor to the last item in the database. */
4731 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4736 if (!(mc->mc_flags & C_EOF)) {
4738 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4741 lkey.mv_size = MDB_MAXKEYSIZE+1;
4742 lkey.mv_data = NULL;
4743 rc = mdb_page_search(mc, &lkey, 0);
4744 if (rc != MDB_SUCCESS)
4747 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4750 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4751 mc->mc_flags |= C_INITIALIZED|C_EOF;
4752 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4754 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4755 key->mv_size = mc->mc_db->md_pad;
4756 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4761 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4762 mdb_xcursor_init1(mc, leaf);
4763 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4768 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4769 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4774 MDB_GET_KEY(leaf, key);
4779 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4788 case MDB_GET_CURRENT:
4789 if (!(mc->mc_flags & C_INITIALIZED)) {
4792 MDB_page *mp = mc->mc_pg[mc->mc_top];
4794 mc->mc_ki[mc->mc_top] = 0;
4800 key->mv_size = mc->mc_db->md_pad;
4801 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4803 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4804 MDB_GET_KEY(leaf, key);
4806 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4807 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4809 rc = mdb_node_read(mc->mc_txn, leaf, data);
4816 case MDB_GET_BOTH_RANGE:
4817 if (data == NULL || mc->mc_xcursor == NULL) {
4825 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4827 } else if (op == MDB_SET_RANGE)
4828 rc = mdb_cursor_set(mc, key, data, op, NULL);
4830 rc = mdb_cursor_set(mc, key, data, op, &exact);
4832 case MDB_GET_MULTIPLE:
4834 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4835 !(mc->mc_flags & C_INITIALIZED)) {
4840 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4841 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4844 case MDB_NEXT_MULTIPLE:
4846 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4850 if (!(mc->mc_flags & C_INITIALIZED))
4851 rc = mdb_cursor_first(mc, key, data);
4853 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4854 if (rc == MDB_SUCCESS) {
4855 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4858 mx = &mc->mc_xcursor->mx_cursor;
4859 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4861 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4862 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4870 case MDB_NEXT_NODUP:
4871 if (!(mc->mc_flags & C_INITIALIZED))
4872 rc = mdb_cursor_first(mc, key, data);
4874 rc = mdb_cursor_next(mc, key, data, op);
4878 case MDB_PREV_NODUP:
4879 if (!(mc->mc_flags & C_INITIALIZED)) {
4880 rc = mdb_cursor_last(mc, key, data);
4883 mc->mc_flags |= C_INITIALIZED;
4884 mc->mc_ki[mc->mc_top]++;
4886 rc = mdb_cursor_prev(mc, key, data, op);
4889 rc = mdb_cursor_first(mc, key, data);
4893 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4894 !(mc->mc_flags & C_INITIALIZED) ||
4895 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4899 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4902 rc = mdb_cursor_last(mc, key, data);
4906 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4907 !(mc->mc_flags & C_INITIALIZED) ||
4908 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4912 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4915 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4923 /** Touch all the pages in the cursor stack.
4924 * Makes sure all the pages are writable, before attempting a write operation.
4925 * @param[in] mc The cursor to operate on.
4928 mdb_cursor_touch(MDB_cursor *mc)
4932 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4935 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
4936 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4939 *mc->mc_dbflag |= DB_DIRTY;
4941 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4942 rc = mdb_page_touch(mc);
4946 mc->mc_top = mc->mc_snum-1;
4951 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4954 MDB_node *leaf = NULL;
4955 MDB_val xdata, *rdata, dkey;
4958 int do_sub = 0, insert = 0;
4959 unsigned int mcount = 0;
4963 char dbuf[MDB_MAXKEYSIZE+1];
4964 unsigned int nflags;
4967 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4970 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4973 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4976 #if SIZE_MAX > MAXDATASIZE
4977 if (data->mv_size > MAXDATASIZE)
4981 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4982 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4986 if (flags == MDB_CURRENT) {
4987 if (!(mc->mc_flags & C_INITIALIZED))
4990 } else if (mc->mc_db->md_root == P_INVALID) {
4992 /* new database, write a root leaf page */
4993 DPUTS("allocating new root leaf page");
4994 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4998 mdb_cursor_push(mc, np);
4999 mc->mc_db->md_root = np->mp_pgno;
5000 mc->mc_db->md_depth++;
5001 *mc->mc_dbflag |= DB_DIRTY;
5002 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5004 np->mp_flags |= P_LEAF2;
5005 mc->mc_flags |= C_INITIALIZED;
5011 if (flags & MDB_APPEND) {
5013 rc = mdb_cursor_last(mc, &k2, &d2);
5015 rc = mc->mc_dbx->md_cmp(key, &k2);
5018 mc->mc_ki[mc->mc_top]++;
5020 /* new key is <= last key */
5025 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5027 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5028 DPRINTF("duplicate key [%s]", DKEY(key));
5030 return MDB_KEYEXIST;
5032 if (rc && rc != MDB_NOTFOUND)
5036 /* Cursor is positioned, now make sure all pages are writable */
5037 rc2 = mdb_cursor_touch(mc);
5042 /* The key already exists */
5043 if (rc == MDB_SUCCESS) {
5044 /* there's only a key anyway, so this is a no-op */
5045 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5046 unsigned int ksize = mc->mc_db->md_pad;
5047 if (key->mv_size != ksize)
5049 if (flags == MDB_CURRENT) {
5050 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5051 memcpy(ptr, key->mv_data, ksize);
5056 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5059 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5060 /* Was a single item before, must convert now */
5062 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5063 /* Just overwrite the current item */
5064 if (flags == MDB_CURRENT)
5067 dkey.mv_size = NODEDSZ(leaf);
5068 dkey.mv_data = NODEDATA(leaf);
5069 #if UINT_MAX < SIZE_MAX
5070 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5071 #ifdef MISALIGNED_OK
5072 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5074 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5077 /* if data matches, ignore it */
5078 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5079 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5081 /* create a fake page for the dup items */
5082 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5083 dkey.mv_data = dbuf;
5084 fp = (MDB_page *)&pbuf;
5085 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5086 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5087 fp->mp_lower = PAGEHDRSZ;
5088 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5089 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5090 fp->mp_flags |= P_LEAF2;
5091 fp->mp_pad = data->mv_size;
5092 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5094 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5095 (dkey.mv_size & 1) + (data->mv_size & 1);
5097 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5100 xdata.mv_size = fp->mp_upper;
5105 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5106 /* See if we need to convert from fake page to subDB */
5108 unsigned int offset;
5112 fp = NODEDATA(leaf);
5113 if (flags == MDB_CURRENT) {
5115 fp->mp_flags |= P_DIRTY;
5116 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5117 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5121 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5122 offset = fp->mp_pad;
5123 if (SIZELEFT(fp) >= offset)
5125 offset *= 4; /* space for 4 more */
5127 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5129 offset += offset & 1;
5130 fp_flags = fp->mp_flags;
5131 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5132 offset >= mc->mc_txn->mt_env->me_nodemax) {
5133 /* yes, convert it */
5135 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5136 dummy.md_pad = fp->mp_pad;
5137 dummy.md_flags = MDB_DUPFIXED;
5138 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5139 dummy.md_flags |= MDB_INTEGERKEY;
5142 dummy.md_branch_pages = 0;
5143 dummy.md_leaf_pages = 1;
5144 dummy.md_overflow_pages = 0;
5145 dummy.md_entries = NUMKEYS(fp);
5147 xdata.mv_size = sizeof(MDB_db);
5148 xdata.mv_data = &dummy;
5149 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5151 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5152 flags |= F_DUPDATA|F_SUBDATA;
5153 dummy.md_root = mp->mp_pgno;
5154 fp_flags &= ~P_SUBP;
5156 /* no, just grow it */
5158 xdata.mv_size = NODEDSZ(leaf) + offset;
5159 xdata.mv_data = &pbuf;
5160 mp = (MDB_page *)&pbuf;
5161 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5164 mp->mp_flags = fp_flags | P_DIRTY;
5165 mp->mp_pad = fp->mp_pad;
5166 mp->mp_lower = fp->mp_lower;
5167 mp->mp_upper = fp->mp_upper + offset;
5169 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5171 nsize = NODEDSZ(leaf) - fp->mp_upper;
5172 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5173 for (i=0; i<NUMKEYS(fp); i++)
5174 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5176 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5180 /* data is on sub-DB, just store it */
5181 flags |= F_DUPDATA|F_SUBDATA;
5185 /* overflow page overwrites need special handling */
5186 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5189 unsigned psize = mc->mc_txn->mt_env->me_psize;
5190 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5192 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5193 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5195 ovpages = omp->mp_pages;
5197 /* Is the ov page writable and large enough? */
5198 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5199 /* yes, overwrite it. Note in this case we don't
5200 * bother to try shrinking the page if the new data
5201 * is smaller than the overflow threshold.
5204 /* It is writable only in a parent txn */
5205 size_t sz = (size_t) psize * ovpages, off;
5206 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5212 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5213 if (!(flags & MDB_RESERVE)) {
5214 /* Copy end of page, adjusting alignment so
5215 * compiler may copy words instead of bytes.
5217 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5218 memcpy((size_t *)((char *)np + off),
5219 (size_t *)((char *)omp + off), sz - off);
5222 memcpy(np, omp, sz); /* Copy beginning of page */
5225 SETDSZ(leaf, data->mv_size);
5226 if (F_ISSET(flags, MDB_RESERVE))
5227 data->mv_data = METADATA(omp);
5229 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5232 mdb_ovpage_free(mc, omp);
5234 } else if (NODEDSZ(leaf) == data->mv_size) {
5235 /* same size, just replace it. Note that we could
5236 * also reuse this node if the new data is smaller,
5237 * but instead we opt to shrink the node in that case.
5239 if (F_ISSET(flags, MDB_RESERVE))
5240 data->mv_data = NODEDATA(leaf);
5241 else if (data->mv_size)
5242 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5244 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5247 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5248 mc->mc_db->md_entries--;
5250 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5257 nflags = flags & NODE_ADD_FLAGS;
5258 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5259 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5260 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5261 nflags &= ~MDB_APPEND;
5263 nflags |= MDB_SPLIT_REPLACE;
5264 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5266 /* There is room already in this leaf page. */
5267 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5268 if (rc == 0 && !do_sub && insert) {
5269 /* Adjust other cursors pointing to mp */
5270 MDB_cursor *m2, *m3;
5271 MDB_dbi dbi = mc->mc_dbi;
5272 unsigned i = mc->mc_top;
5273 MDB_page *mp = mc->mc_pg[i];
5275 if (mc->mc_flags & C_SUB)
5278 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5279 if (mc->mc_flags & C_SUB)
5280 m3 = &m2->mc_xcursor->mx_cursor;
5283 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5284 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5291 if (rc != MDB_SUCCESS)
5292 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5294 /* Now store the actual data in the child DB. Note that we're
5295 * storing the user data in the keys field, so there are strict
5296 * size limits on dupdata. The actual data fields of the child
5297 * DB are all zero size.
5304 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5305 if (flags & MDB_CURRENT) {
5306 xflags = MDB_CURRENT;
5308 mdb_xcursor_init1(mc, leaf);
5309 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5311 /* converted, write the original data first */
5313 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5317 /* Adjust other cursors pointing to mp */
5319 unsigned i = mc->mc_top;
5320 MDB_page *mp = mc->mc_pg[i];
5322 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5323 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5324 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5325 mdb_xcursor_init1(m2, leaf);
5329 /* we've done our job */
5332 if (flags & MDB_APPENDDUP)
5333 xflags |= MDB_APPEND;
5334 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5335 if (flags & F_SUBDATA) {
5336 void *db = NODEDATA(leaf);
5337 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5340 /* sub-writes might have failed so check rc again.
5341 * Don't increment count if we just replaced an existing item.
5343 if (!rc && !(flags & MDB_CURRENT))
5344 mc->mc_db->md_entries++;
5345 if (flags & MDB_MULTIPLE) {
5347 if (mcount < data[1].mv_size) {
5348 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5349 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5355 /* If we succeeded and the key didn't exist before, make sure
5356 * the cursor is marked valid.
5359 mc->mc_flags |= C_INITIALIZED;
5364 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5369 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5372 if (!(mc->mc_flags & C_INITIALIZED))
5375 rc = mdb_cursor_touch(mc);
5379 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5381 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5382 if (flags != MDB_NODUPDATA) {
5383 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5384 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5386 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5387 /* If sub-DB still has entries, we're done */
5388 if (mc->mc_xcursor->mx_db.md_entries) {
5389 if (leaf->mn_flags & F_SUBDATA) {
5390 /* update subDB info */
5391 void *db = NODEDATA(leaf);
5392 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5395 /* shrink fake page */
5396 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5397 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5398 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5399 /* fix other sub-DB cursors pointed at this fake page */
5400 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5401 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5402 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5403 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5404 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5407 mc->mc_db->md_entries--;
5410 /* otherwise fall thru and delete the sub-DB */
5413 if (leaf->mn_flags & F_SUBDATA) {
5414 /* add all the child DB's pages to the free list */
5415 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5416 if (rc == MDB_SUCCESS) {
5417 mc->mc_db->md_entries -=
5418 mc->mc_xcursor->mx_db.md_entries;
5423 return mdb_cursor_del0(mc, leaf);
5426 /** Allocate and initialize new pages for a database.
5427 * @param[in] mc a cursor on the database being added to.
5428 * @param[in] flags flags defining what type of page is being allocated.
5429 * @param[in] num the number of pages to allocate. This is usually 1,
5430 * unless allocating overflow pages for a large record.
5431 * @param[out] mp Address of a page, or NULL on failure.
5432 * @return 0 on success, non-zero on failure.
5435 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5440 if ((rc = mdb_page_alloc(mc, num, &np)))
5442 DPRINTF("allocated new mpage %zu, page size %u",
5443 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5444 np->mp_flags = flags | P_DIRTY;
5445 np->mp_lower = PAGEHDRSZ;
5446 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5449 mc->mc_db->md_branch_pages++;
5450 else if (IS_LEAF(np))
5451 mc->mc_db->md_leaf_pages++;
5452 else if (IS_OVERFLOW(np)) {
5453 mc->mc_db->md_overflow_pages += num;
5461 /** Calculate the size of a leaf node.
5462 * The size depends on the environment's page size; if a data item
5463 * is too large it will be put onto an overflow page and the node
5464 * size will only include the key and not the data. Sizes are always
5465 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5466 * of the #MDB_node headers.
5467 * @param[in] env The environment handle.
5468 * @param[in] key The key for the node.
5469 * @param[in] data The data for the node.
5470 * @return The number of bytes needed to store the node.
5473 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5477 sz = LEAFSIZE(key, data);
5478 if (sz >= env->me_nodemax) {
5479 /* put on overflow page */
5480 sz -= data->mv_size - sizeof(pgno_t);
5484 return sz + sizeof(indx_t);
5487 /** Calculate the size of a branch node.
5488 * The size should depend on the environment's page size but since
5489 * we currently don't support spilling large keys onto overflow
5490 * pages, it's simply the size of the #MDB_node header plus the
5491 * size of the key. Sizes are always rounded up to an even number
5492 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5493 * @param[in] env The environment handle.
5494 * @param[in] key The key for the node.
5495 * @return The number of bytes needed to store the node.
5498 mdb_branch_size(MDB_env *env, MDB_val *key)
5503 if (sz >= env->me_nodemax) {
5504 /* put on overflow page */
5505 /* not implemented */
5506 /* sz -= key->size - sizeof(pgno_t); */
5509 return sz + sizeof(indx_t);
5512 /** Add a node to the page pointed to by the cursor.
5513 * @param[in] mc The cursor for this operation.
5514 * @param[in] indx The index on the page where the new node should be added.
5515 * @param[in] key The key for the new node.
5516 * @param[in] data The data for the new node, if any.
5517 * @param[in] pgno The page number, if adding a branch node.
5518 * @param[in] flags Flags for the node.
5519 * @return 0 on success, non-zero on failure. Possible errors are:
5521 * <li>ENOMEM - failed to allocate overflow pages for the node.
5522 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5523 * should never happen since all callers already calculate the
5524 * page's free space before calling this function.
5528 mdb_node_add(MDB_cursor *mc, indx_t indx,
5529 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5532 size_t node_size = NODESIZE;
5535 MDB_page *mp = mc->mc_pg[mc->mc_top];
5536 MDB_page *ofp = NULL; /* overflow page */
5539 assert(mp->mp_upper >= mp->mp_lower);
5541 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5542 IS_LEAF(mp) ? "leaf" : "branch",
5543 IS_SUBP(mp) ? "sub-" : "",
5544 mp->mp_pgno, indx, data ? data->mv_size : 0,
5545 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5548 /* Move higher keys up one slot. */
5549 int ksize = mc->mc_db->md_pad, dif;
5550 char *ptr = LEAF2KEY(mp, indx, ksize);
5551 dif = NUMKEYS(mp) - indx;
5553 memmove(ptr+ksize, ptr, dif*ksize);
5554 /* insert new key */
5555 memcpy(ptr, key->mv_data, ksize);
5557 /* Just using these for counting */
5558 mp->mp_lower += sizeof(indx_t);
5559 mp->mp_upper -= ksize - sizeof(indx_t);
5564 node_size += key->mv_size;
5568 if (F_ISSET(flags, F_BIGDATA)) {
5569 /* Data already on overflow page. */
5570 node_size += sizeof(pgno_t);
5571 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5572 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5574 /* Put data on overflow page. */
5575 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5576 data->mv_size, node_size+data->mv_size);
5577 node_size += sizeof(pgno_t);
5578 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5580 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5583 node_size += data->mv_size;
5586 node_size += node_size & 1;
5588 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5589 DPRINTF("not enough room in page %zu, got %u ptrs",
5590 mp->mp_pgno, NUMKEYS(mp));
5591 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5592 mp->mp_upper - mp->mp_lower);
5593 DPRINTF("node size = %zu", node_size);
5594 return MDB_PAGE_FULL;
5597 /* Move higher pointers up one slot. */
5598 for (i = NUMKEYS(mp); i > indx; i--)
5599 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5601 /* Adjust free space offsets. */
5602 ofs = mp->mp_upper - node_size;
5603 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5604 mp->mp_ptrs[indx] = ofs;
5606 mp->mp_lower += sizeof(indx_t);
5608 /* Write the node data. */
5609 node = NODEPTR(mp, indx);
5610 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5611 node->mn_flags = flags;
5613 SETDSZ(node,data->mv_size);
5618 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5623 if (F_ISSET(flags, F_BIGDATA))
5624 memcpy(node->mn_data + key->mv_size, data->mv_data,
5626 else if (F_ISSET(flags, MDB_RESERVE))
5627 data->mv_data = node->mn_data + key->mv_size;
5629 memcpy(node->mn_data + key->mv_size, data->mv_data,
5632 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5634 if (F_ISSET(flags, MDB_RESERVE))
5635 data->mv_data = METADATA(ofp);
5637 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5644 /** Delete the specified node from a page.
5645 * @param[in] mp The page to operate on.
5646 * @param[in] indx The index of the node to delete.
5647 * @param[in] ksize The size of a node. Only used if the page is
5648 * part of a #MDB_DUPFIXED database.
5651 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5654 indx_t i, j, numkeys, ptr;
5661 COPY_PGNO(pgno, mp->mp_pgno);
5662 DPRINTF("delete node %u on %s page %zu", indx,
5663 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5666 assert(indx < NUMKEYS(mp));
5669 int x = NUMKEYS(mp) - 1 - indx;
5670 base = LEAF2KEY(mp, indx, ksize);
5672 memmove(base, base + ksize, x * ksize);
5673 mp->mp_lower -= sizeof(indx_t);
5674 mp->mp_upper += ksize - sizeof(indx_t);
5678 node = NODEPTR(mp, indx);
5679 sz = NODESIZE + node->mn_ksize;
5681 if (F_ISSET(node->mn_flags, F_BIGDATA))
5682 sz += sizeof(pgno_t);
5684 sz += NODEDSZ(node);
5688 ptr = mp->mp_ptrs[indx];
5689 numkeys = NUMKEYS(mp);
5690 for (i = j = 0; i < numkeys; i++) {
5692 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5693 if (mp->mp_ptrs[i] < ptr)
5694 mp->mp_ptrs[j] += sz;
5699 base = (char *)mp + mp->mp_upper;
5700 memmove(base + sz, base, ptr - mp->mp_upper);
5702 mp->mp_lower -= sizeof(indx_t);
5706 /** Compact the main page after deleting a node on a subpage.
5707 * @param[in] mp The main page to operate on.
5708 * @param[in] indx The index of the subpage on the main page.
5711 mdb_node_shrink(MDB_page *mp, indx_t indx)
5718 indx_t i, numkeys, ptr;
5720 node = NODEPTR(mp, indx);
5721 sp = (MDB_page *)NODEDATA(node);
5722 osize = NODEDSZ(node);
5724 delta = sp->mp_upper - sp->mp_lower;
5725 SETDSZ(node, osize - delta);
5726 xp = (MDB_page *)((char *)sp + delta);
5728 /* shift subpage upward */
5730 nsize = NUMKEYS(sp) * sp->mp_pad;
5731 memmove(METADATA(xp), METADATA(sp), nsize);
5734 nsize = osize - sp->mp_upper;
5735 numkeys = NUMKEYS(sp);
5736 for (i=numkeys-1; i>=0; i--)
5737 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5739 xp->mp_upper = sp->mp_lower;
5740 xp->mp_lower = sp->mp_lower;
5741 xp->mp_flags = sp->mp_flags;
5742 xp->mp_pad = sp->mp_pad;
5743 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5745 /* shift lower nodes upward */
5746 ptr = mp->mp_ptrs[indx];
5747 numkeys = NUMKEYS(mp);
5748 for (i = 0; i < numkeys; i++) {
5749 if (mp->mp_ptrs[i] <= ptr)
5750 mp->mp_ptrs[i] += delta;
5753 base = (char *)mp + mp->mp_upper;
5754 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5755 mp->mp_upper += delta;
5758 /** Initial setup of a sorted-dups cursor.
5759 * Sorted duplicates are implemented as a sub-database for the given key.
5760 * The duplicate data items are actually keys of the sub-database.
5761 * Operations on the duplicate data items are performed using a sub-cursor
5762 * initialized when the sub-database is first accessed. This function does
5763 * the preliminary setup of the sub-cursor, filling in the fields that
5764 * depend only on the parent DB.
5765 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5768 mdb_xcursor_init0(MDB_cursor *mc)
5770 MDB_xcursor *mx = mc->mc_xcursor;
5772 mx->mx_cursor.mc_xcursor = NULL;
5773 mx->mx_cursor.mc_txn = mc->mc_txn;
5774 mx->mx_cursor.mc_db = &mx->mx_db;
5775 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5776 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5777 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5778 mx->mx_cursor.mc_snum = 0;
5779 mx->mx_cursor.mc_top = 0;
5780 mx->mx_cursor.mc_flags = C_SUB;
5781 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5782 mx->mx_dbx.md_dcmp = NULL;
5783 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5786 /** Final setup of a sorted-dups cursor.
5787 * Sets up the fields that depend on the data from the main cursor.
5788 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5789 * @param[in] node The data containing the #MDB_db record for the
5790 * sorted-dup database.
5793 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5795 MDB_xcursor *mx = mc->mc_xcursor;
5797 if (node->mn_flags & F_SUBDATA) {
5798 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5799 mx->mx_cursor.mc_pg[0] = 0;
5800 mx->mx_cursor.mc_snum = 0;
5801 mx->mx_cursor.mc_flags = C_SUB;
5803 MDB_page *fp = NODEDATA(node);
5804 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5805 mx->mx_db.md_flags = 0;
5806 mx->mx_db.md_depth = 1;
5807 mx->mx_db.md_branch_pages = 0;
5808 mx->mx_db.md_leaf_pages = 1;
5809 mx->mx_db.md_overflow_pages = 0;
5810 mx->mx_db.md_entries = NUMKEYS(fp);
5811 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5812 mx->mx_cursor.mc_snum = 1;
5813 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5814 mx->mx_cursor.mc_top = 0;
5815 mx->mx_cursor.mc_pg[0] = fp;
5816 mx->mx_cursor.mc_ki[0] = 0;
5817 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5818 mx->mx_db.md_flags = MDB_DUPFIXED;
5819 mx->mx_db.md_pad = fp->mp_pad;
5820 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5821 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5824 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5826 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5828 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5829 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5830 #if UINT_MAX < SIZE_MAX
5831 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5832 #ifdef MISALIGNED_OK
5833 mx->mx_dbx.md_cmp = mdb_cmp_long;
5835 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5840 /** Initialize a cursor for a given transaction and database. */
5842 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5847 mc->mc_db = &txn->mt_dbs[dbi];
5848 mc->mc_dbx = &txn->mt_dbxs[dbi];
5849 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5854 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5856 mc->mc_xcursor = mx;
5857 mdb_xcursor_init0(mc);
5859 mc->mc_xcursor = NULL;
5861 if (*mc->mc_dbflag & DB_STALE) {
5862 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5867 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5870 size_t size = sizeof(MDB_cursor);
5872 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5875 /* Allow read access to the freelist */
5876 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5879 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5880 size += sizeof(MDB_xcursor);
5882 if ((mc = malloc(size)) != NULL) {
5883 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5884 if (txn->mt_cursors) {
5885 mc->mc_next = txn->mt_cursors[dbi];
5886 txn->mt_cursors[dbi] = mc;
5887 mc->mc_flags |= C_UNTRACK;
5889 mc->mc_flags |= C_ALLOCD;
5900 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5904 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5907 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5910 flags = mc->mc_flags;
5912 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5914 mc->mc_flags |= (flags & C_ALLOCD);
5918 /* Return the count of duplicate data items for the current key */
5920 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5924 if (mc == NULL || countp == NULL)
5927 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5930 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5931 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5934 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5937 *countp = mc->mc_xcursor->mx_db.md_entries;
5943 mdb_cursor_close(MDB_cursor *mc)
5946 /* remove from txn, if tracked */
5947 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
5948 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5949 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5951 *prev = mc->mc_next;
5953 if (mc->mc_flags & C_ALLOCD)
5959 mdb_cursor_txn(MDB_cursor *mc)
5961 if (!mc) return NULL;
5966 mdb_cursor_dbi(MDB_cursor *mc)
5972 /** Replace the key for a node with a new key.
5973 * @param[in] mc Cursor pointing to the node to operate on.
5974 * @param[in] key The new key to use.
5975 * @return 0 on success, non-zero on failure.
5978 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5985 indx_t ptr, i, numkeys, indx;
5988 indx = mc->mc_ki[mc->mc_top];
5989 mp = mc->mc_pg[mc->mc_top];
5990 node = NODEPTR(mp, indx);
5991 ptr = mp->mp_ptrs[indx];
5995 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5996 k2.mv_data = NODEKEY(node);
5997 k2.mv_size = node->mn_ksize;
5998 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6000 mdb_dkey(&k2, kbuf2),
6006 delta0 = delta = key->mv_size - node->mn_ksize;
6008 /* Must be 2-byte aligned. If new key is
6009 * shorter by 1, the shift will be skipped.
6011 delta += (delta & 1);
6013 if (delta > 0 && SIZELEFT(mp) < delta) {
6015 /* not enough space left, do a delete and split */
6016 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6017 pgno = NODEPGNO(node);
6018 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6019 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6022 numkeys = NUMKEYS(mp);
6023 for (i = 0; i < numkeys; i++) {
6024 if (mp->mp_ptrs[i] <= ptr)
6025 mp->mp_ptrs[i] -= delta;
6028 base = (char *)mp + mp->mp_upper;
6029 len = ptr - mp->mp_upper + NODESIZE;
6030 memmove(base - delta, base, len);
6031 mp->mp_upper -= delta;
6033 node = NODEPTR(mp, indx);
6036 /* But even if no shift was needed, update ksize */
6038 node->mn_ksize = key->mv_size;
6041 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6047 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6049 /** Move a node from csrc to cdst.
6052 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6059 unsigned short flags;
6063 /* Mark src and dst as dirty. */
6064 if ((rc = mdb_page_touch(csrc)) ||
6065 (rc = mdb_page_touch(cdst)))
6068 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6069 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6070 key.mv_size = csrc->mc_db->md_pad;
6071 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6073 data.mv_data = NULL;
6077 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6078 assert(!((long)srcnode&1));
6079 srcpg = NODEPGNO(srcnode);
6080 flags = srcnode->mn_flags;
6081 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6082 unsigned int snum = csrc->mc_snum;
6084 /* must find the lowest key below src */
6085 mdb_page_search_lowest(csrc);
6086 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6087 key.mv_size = csrc->mc_db->md_pad;
6088 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6090 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6091 key.mv_size = NODEKSZ(s2);
6092 key.mv_data = NODEKEY(s2);
6094 csrc->mc_snum = snum--;
6095 csrc->mc_top = snum;
6097 key.mv_size = NODEKSZ(srcnode);
6098 key.mv_data = NODEKEY(srcnode);
6100 data.mv_size = NODEDSZ(srcnode);
6101 data.mv_data = NODEDATA(srcnode);
6103 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6104 unsigned int snum = cdst->mc_snum;
6107 /* must find the lowest key below dst */
6108 mdb_page_search_lowest(cdst);
6109 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6110 bkey.mv_size = cdst->mc_db->md_pad;
6111 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6113 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6114 bkey.mv_size = NODEKSZ(s2);
6115 bkey.mv_data = NODEKEY(s2);
6117 cdst->mc_snum = snum--;
6118 cdst->mc_top = snum;
6119 mdb_cursor_copy(cdst, &mn);
6121 rc = mdb_update_key(&mn, &bkey);
6126 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6127 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6128 csrc->mc_ki[csrc->mc_top],
6130 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6131 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6133 /* Add the node to the destination page.
6135 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6136 if (rc != MDB_SUCCESS)
6139 /* Delete the node from the source page.
6141 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6144 /* Adjust other cursors pointing to mp */
6145 MDB_cursor *m2, *m3;
6146 MDB_dbi dbi = csrc->mc_dbi;
6147 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6149 if (csrc->mc_flags & C_SUB)
6152 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6153 if (csrc->mc_flags & C_SUB)
6154 m3 = &m2->mc_xcursor->mx_cursor;
6157 if (m3 == csrc) continue;
6158 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6159 csrc->mc_ki[csrc->mc_top]) {
6160 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6161 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6166 /* Update the parent separators.
6168 if (csrc->mc_ki[csrc->mc_top] == 0) {
6169 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6170 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6171 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6173 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6174 key.mv_size = NODEKSZ(srcnode);
6175 key.mv_data = NODEKEY(srcnode);
6177 DPRINTF("update separator for source page %zu to [%s]",
6178 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6179 mdb_cursor_copy(csrc, &mn);
6182 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6185 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6187 indx_t ix = csrc->mc_ki[csrc->mc_top];
6188 nullkey.mv_size = 0;
6189 csrc->mc_ki[csrc->mc_top] = 0;
6190 rc = mdb_update_key(csrc, &nullkey);
6191 csrc->mc_ki[csrc->mc_top] = ix;
6192 assert(rc == MDB_SUCCESS);
6196 if (cdst->mc_ki[cdst->mc_top] == 0) {
6197 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6198 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6199 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6201 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6202 key.mv_size = NODEKSZ(srcnode);
6203 key.mv_data = NODEKEY(srcnode);
6205 DPRINTF("update separator for destination page %zu to [%s]",
6206 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6207 mdb_cursor_copy(cdst, &mn);
6210 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6213 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6215 indx_t ix = cdst->mc_ki[cdst->mc_top];
6216 nullkey.mv_size = 0;
6217 cdst->mc_ki[cdst->mc_top] = 0;
6218 rc = mdb_update_key(cdst, &nullkey);
6219 cdst->mc_ki[cdst->mc_top] = ix;
6220 assert(rc == MDB_SUCCESS);
6227 /** Merge one page into another.
6228 * The nodes from the page pointed to by \b csrc will
6229 * be copied to the page pointed to by \b cdst and then
6230 * the \b csrc page will be freed.
6231 * @param[in] csrc Cursor pointing to the source page.
6232 * @param[in] cdst Cursor pointing to the destination page.
6235 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6243 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6244 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6246 assert(csrc->mc_snum > 1); /* can't merge root page */
6247 assert(cdst->mc_snum > 1);
6249 /* Mark dst as dirty. */
6250 if ((rc = mdb_page_touch(cdst)))
6253 /* Move all nodes from src to dst.
6255 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6256 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6257 key.mv_size = csrc->mc_db->md_pad;
6258 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6259 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6260 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6261 if (rc != MDB_SUCCESS)
6263 key.mv_data = (char *)key.mv_data + key.mv_size;
6266 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6267 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6268 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6269 unsigned int snum = csrc->mc_snum;
6271 /* must find the lowest key below src */
6272 mdb_page_search_lowest(csrc);
6273 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6274 key.mv_size = csrc->mc_db->md_pad;
6275 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6277 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6278 key.mv_size = NODEKSZ(s2);
6279 key.mv_data = NODEKEY(s2);
6281 csrc->mc_snum = snum--;
6282 csrc->mc_top = snum;
6284 key.mv_size = srcnode->mn_ksize;
6285 key.mv_data = NODEKEY(srcnode);
6288 data.mv_size = NODEDSZ(srcnode);
6289 data.mv_data = NODEDATA(srcnode);
6290 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6291 if (rc != MDB_SUCCESS)
6296 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6297 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);
6299 /* Unlink the src page from parent and add to free list.
6301 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6302 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6305 rc = mdb_update_key(csrc, &key);
6311 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6312 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6313 csrc->mc_db->md_leaf_pages--;
6315 csrc->mc_db->md_branch_pages--;
6317 /* Adjust other cursors pointing to mp */
6318 MDB_cursor *m2, *m3;
6319 MDB_dbi dbi = csrc->mc_dbi;
6320 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6322 if (csrc->mc_flags & C_SUB)
6325 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6326 if (csrc->mc_flags & C_SUB)
6327 m3 = &m2->mc_xcursor->mx_cursor;
6330 if (m3 == csrc) continue;
6331 if (m3->mc_snum < csrc->mc_snum) continue;
6332 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6333 m3->mc_pg[csrc->mc_top] = mp;
6334 m3->mc_ki[csrc->mc_top] += nkeys;
6338 mdb_cursor_pop(csrc);
6340 return mdb_rebalance(csrc);
6343 /** Copy the contents of a cursor.
6344 * @param[in] csrc The cursor to copy from.
6345 * @param[out] cdst The cursor to copy to.
6348 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6352 cdst->mc_txn = csrc->mc_txn;
6353 cdst->mc_dbi = csrc->mc_dbi;
6354 cdst->mc_db = csrc->mc_db;
6355 cdst->mc_dbx = csrc->mc_dbx;
6356 cdst->mc_snum = csrc->mc_snum;
6357 cdst->mc_top = csrc->mc_top;
6358 cdst->mc_flags = csrc->mc_flags;
6360 for (i=0; i<csrc->mc_snum; i++) {
6361 cdst->mc_pg[i] = csrc->mc_pg[i];
6362 cdst->mc_ki[i] = csrc->mc_ki[i];
6366 /** Rebalance the tree after a delete operation.
6367 * @param[in] mc Cursor pointing to the page where rebalancing
6369 * @return 0 on success, non-zero on failure.
6372 mdb_rebalance(MDB_cursor *mc)
6376 unsigned int ptop, minkeys;
6379 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6383 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6384 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6385 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6386 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6390 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6391 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6394 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6395 DPRINTF("no need to rebalance page %zu, above fill threshold",
6401 if (mc->mc_snum < 2) {
6402 MDB_page *mp = mc->mc_pg[0];
6404 DPUTS("Can't rebalance a subpage, ignoring");
6407 if (NUMKEYS(mp) == 0) {
6408 DPUTS("tree is completely empty");
6409 mc->mc_db->md_root = P_INVALID;
6410 mc->mc_db->md_depth = 0;
6411 mc->mc_db->md_leaf_pages = 0;
6412 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6413 /* Adjust cursors pointing to mp */
6417 MDB_cursor *m2, *m3;
6418 MDB_dbi dbi = mc->mc_dbi;
6420 if (mc->mc_flags & C_SUB)
6423 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6424 if (mc->mc_flags & C_SUB)
6425 m3 = &m2->mc_xcursor->mx_cursor;
6428 if (m3->mc_snum < mc->mc_snum) continue;
6429 if (m3->mc_pg[0] == mp) {
6435 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6436 DPUTS("collapsing root page!");
6437 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6438 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6439 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6442 mc->mc_db->md_depth--;
6443 mc->mc_db->md_branch_pages--;
6445 /* Adjust other cursors pointing to mp */
6446 MDB_cursor *m2, *m3;
6447 MDB_dbi dbi = mc->mc_dbi;
6449 if (mc->mc_flags & C_SUB)
6452 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6453 if (mc->mc_flags & C_SUB)
6454 m3 = &m2->mc_xcursor->mx_cursor;
6457 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6458 if (m3->mc_pg[0] == mp) {
6459 m3->mc_pg[0] = mc->mc_pg[0];
6466 DPUTS("root page doesn't need rebalancing");
6470 /* The parent (branch page) must have at least 2 pointers,
6471 * otherwise the tree is invalid.
6473 ptop = mc->mc_top-1;
6474 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6476 /* Leaf page fill factor is below the threshold.
6477 * Try to move keys from left or right neighbor, or
6478 * merge with a neighbor page.
6483 mdb_cursor_copy(mc, &mn);
6484 mn.mc_xcursor = NULL;
6486 if (mc->mc_ki[ptop] == 0) {
6487 /* We're the leftmost leaf in our parent.
6489 DPUTS("reading right neighbor");
6491 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6492 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6495 mn.mc_ki[mn.mc_top] = 0;
6496 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6498 /* There is at least one neighbor to the left.
6500 DPUTS("reading left neighbor");
6502 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6503 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6506 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6507 mc->mc_ki[mc->mc_top] = 0;
6510 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6511 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);
6513 /* If the neighbor page is above threshold and has enough keys,
6514 * move one key from it. Otherwise we should try to merge them.
6515 * (A branch page must never have less than 2 keys.)
6517 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6518 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6519 return mdb_node_move(&mn, mc);
6521 if (mc->mc_ki[ptop] == 0)
6522 rc = mdb_page_merge(&mn, mc);
6524 rc = mdb_page_merge(mc, &mn);
6525 mc->mc_flags &= ~C_INITIALIZED;
6530 /** Complete a delete operation started by #mdb_cursor_del(). */
6532 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6536 /* add overflow pages to free list */
6537 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6541 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6542 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
6544 assert(IS_OVERFLOW(omp));
6545 mdb_ovpage_free(mc, omp);
6547 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6548 mc->mc_db->md_entries--;
6549 rc = mdb_rebalance(mc);
6550 if (rc != MDB_SUCCESS)
6551 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6552 /* if mc points past last node in page, invalidate */
6553 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6554 mc->mc_flags &= ~C_INITIALIZED;
6560 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6561 MDB_val *key, MDB_val *data)
6566 MDB_val rdata, *xdata;
6570 assert(key != NULL);
6572 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6574 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6577 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6581 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6585 mdb_cursor_init(&mc, txn, dbi, &mx);
6596 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6598 /* let mdb_page_split know about this cursor if needed:
6599 * delete will trigger a rebalance; if it needs to move
6600 * a node from one page to another, it will have to
6601 * update the parent's separator key(s). If the new sepkey
6602 * is larger than the current one, the parent page may
6603 * run out of space, triggering a split. We need this
6604 * cursor to be consistent until the end of the rebalance.
6606 mc.mc_next = txn->mt_cursors[dbi];
6607 txn->mt_cursors[dbi] = &mc;
6608 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6609 txn->mt_cursors[dbi] = mc.mc_next;
6614 /** Split a page and insert a new node.
6615 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6616 * The cursor will be updated to point to the actual page and index where
6617 * the node got inserted after the split.
6618 * @param[in] newkey The key for the newly inserted node.
6619 * @param[in] newdata The data for the newly inserted node.
6620 * @param[in] newpgno The page number, if the new node is a branch node.
6621 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6622 * @return 0 on success, non-zero on failure.
6625 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6626 unsigned int nflags)
6629 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6632 unsigned int i, j, split_indx, nkeys, pmax;
6634 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6636 MDB_page *mp, *rp, *pp;
6641 mp = mc->mc_pg[mc->mc_top];
6642 newindx = mc->mc_ki[mc->mc_top];
6644 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6645 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6646 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6648 /* Create a right sibling. */
6649 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6651 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6653 if (mc->mc_snum < 2) {
6654 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6656 /* shift current top to make room for new parent */
6657 mc->mc_pg[1] = mc->mc_pg[0];
6658 mc->mc_ki[1] = mc->mc_ki[0];
6661 mc->mc_db->md_root = pp->mp_pgno;
6662 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6663 mc->mc_db->md_depth++;
6666 /* Add left (implicit) pointer. */
6667 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6668 /* undo the pre-push */
6669 mc->mc_pg[0] = mc->mc_pg[1];
6670 mc->mc_ki[0] = mc->mc_ki[1];
6671 mc->mc_db->md_root = mp->mp_pgno;
6672 mc->mc_db->md_depth--;
6679 ptop = mc->mc_top-1;
6680 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6683 mc->mc_flags |= C_SPLITTING;
6684 mdb_cursor_copy(mc, &mn);
6685 mn.mc_pg[mn.mc_top] = rp;
6686 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6688 if (nflags & MDB_APPEND) {
6689 mn.mc_ki[mn.mc_top] = 0;
6691 split_indx = newindx;
6696 nkeys = NUMKEYS(mp);
6697 split_indx = nkeys / 2;
6698 if (newindx < split_indx)
6704 unsigned int lsize, rsize, ksize;
6705 /* Move half of the keys to the right sibling */
6707 x = mc->mc_ki[mc->mc_top] - split_indx;
6708 ksize = mc->mc_db->md_pad;
6709 split = LEAF2KEY(mp, split_indx, ksize);
6710 rsize = (nkeys - split_indx) * ksize;
6711 lsize = (nkeys - split_indx) * sizeof(indx_t);
6712 mp->mp_lower -= lsize;
6713 rp->mp_lower += lsize;
6714 mp->mp_upper += rsize - lsize;
6715 rp->mp_upper -= rsize - lsize;
6716 sepkey.mv_size = ksize;
6717 if (newindx == split_indx) {
6718 sepkey.mv_data = newkey->mv_data;
6720 sepkey.mv_data = split;
6723 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6724 memcpy(rp->mp_ptrs, split, rsize);
6725 sepkey.mv_data = rp->mp_ptrs;
6726 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6727 memcpy(ins, newkey->mv_data, ksize);
6728 mp->mp_lower += sizeof(indx_t);
6729 mp->mp_upper -= ksize - sizeof(indx_t);
6732 memcpy(rp->mp_ptrs, split, x * ksize);
6733 ins = LEAF2KEY(rp, x, ksize);
6734 memcpy(ins, newkey->mv_data, ksize);
6735 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6736 rp->mp_lower += sizeof(indx_t);
6737 rp->mp_upper -= ksize - sizeof(indx_t);
6738 mc->mc_ki[mc->mc_top] = x;
6739 mc->mc_pg[mc->mc_top] = rp;
6744 /* For leaf pages, check the split point based on what
6745 * fits where, since otherwise mdb_node_add can fail.
6747 * This check is only needed when the data items are
6748 * relatively large, such that being off by one will
6749 * make the difference between success or failure.
6751 * It's also relevant if a page happens to be laid out
6752 * such that one half of its nodes are all "small" and
6753 * the other half of its nodes are "large." If the new
6754 * item is also "large" and falls on the half with
6755 * "large" nodes, it also may not fit.
6758 unsigned int psize, nsize;
6759 /* Maximum free space in an empty page */
6760 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6761 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6762 if ((nkeys < 20) || (nsize > pmax/16)) {
6763 if (newindx <= split_indx) {
6766 for (i=0; i<split_indx; i++) {
6767 node = NODEPTR(mp, i);
6768 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6769 if (F_ISSET(node->mn_flags, F_BIGDATA))
6770 psize += sizeof(pgno_t);
6772 psize += NODEDSZ(node);
6776 split_indx = newindx;
6787 for (i=nkeys-1; i>=split_indx; i--) {
6788 node = NODEPTR(mp, i);
6789 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6790 if (F_ISSET(node->mn_flags, F_BIGDATA))
6791 psize += sizeof(pgno_t);
6793 psize += NODEDSZ(node);
6797 split_indx = newindx;
6808 /* First find the separating key between the split pages.
6809 * The case where newindx == split_indx is ambiguous; the
6810 * new item could go to the new page or stay on the original
6811 * page. If newpos == 1 it goes to the new page.
6813 if (newindx == split_indx && newpos) {
6814 sepkey.mv_size = newkey->mv_size;
6815 sepkey.mv_data = newkey->mv_data;
6817 node = NODEPTR(mp, split_indx);
6818 sepkey.mv_size = node->mn_ksize;
6819 sepkey.mv_data = NODEKEY(node);
6823 DPRINTF("separator is [%s]", DKEY(&sepkey));
6825 /* Copy separator key to the parent.
6827 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6831 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6834 if (mn.mc_snum == mc->mc_snum) {
6835 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6836 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6837 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6838 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6843 /* Right page might now have changed parent.
6844 * Check if left page also changed parent.
6846 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6847 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6848 for (i=0; i<ptop; i++) {
6849 mc->mc_pg[i] = mn.mc_pg[i];
6850 mc->mc_ki[i] = mn.mc_ki[i];
6852 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6853 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6857 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6860 mc->mc_flags ^= C_SPLITTING;
6861 if (rc != MDB_SUCCESS) {
6864 if (nflags & MDB_APPEND) {
6865 mc->mc_pg[mc->mc_top] = rp;
6866 mc->mc_ki[mc->mc_top] = 0;
6867 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6870 for (i=0; i<mc->mc_top; i++)
6871 mc->mc_ki[i] = mn.mc_ki[i];
6878 /* Move half of the keys to the right sibling. */
6880 /* grab a page to hold a temporary copy */
6881 copy = mdb_page_malloc(mc->mc_txn, 1);
6885 copy->mp_pgno = mp->mp_pgno;
6886 copy->mp_flags = mp->mp_flags;
6887 copy->mp_lower = PAGEHDRSZ;
6888 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6889 mc->mc_pg[mc->mc_top] = copy;
6890 for (i = j = 0; i <= nkeys; j++) {
6891 if (i == split_indx) {
6892 /* Insert in right sibling. */
6893 /* Reset insert index for right sibling. */
6894 if (i != newindx || (newpos ^ ins_new)) {
6896 mc->mc_pg[mc->mc_top] = rp;
6900 if (i == newindx && !ins_new) {
6901 /* Insert the original entry that caused the split. */
6902 rkey.mv_data = newkey->mv_data;
6903 rkey.mv_size = newkey->mv_size;
6912 /* Update index for the new key. */
6913 mc->mc_ki[mc->mc_top] = j;
6914 } else if (i == nkeys) {
6917 node = NODEPTR(mp, i);
6918 rkey.mv_data = NODEKEY(node);
6919 rkey.mv_size = node->mn_ksize;
6921 xdata.mv_data = NODEDATA(node);
6922 xdata.mv_size = NODEDSZ(node);
6925 pgno = NODEPGNO(node);
6926 flags = node->mn_flags;
6931 if (!IS_LEAF(mp) && j == 0) {
6932 /* First branch index doesn't need key data. */
6936 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6940 nkeys = NUMKEYS(copy);
6941 for (i=0; i<nkeys; i++)
6942 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6943 mp->mp_lower = copy->mp_lower;
6944 mp->mp_upper = copy->mp_upper;
6945 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6946 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6948 /* reset back to original page */
6949 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6950 mc->mc_pg[mc->mc_top] = mp;
6951 if (nflags & MDB_RESERVE) {
6952 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6953 if (!(node->mn_flags & F_BIGDATA))
6954 newdata->mv_data = NODEDATA(node);
6958 /* Make sure mc_ki is still valid.
6960 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6961 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6962 for (i=0; i<ptop; i++) {
6963 mc->mc_pg[i] = mn.mc_pg[i];
6964 mc->mc_ki[i] = mn.mc_ki[i];
6966 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6967 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6971 /* return tmp page to freelist */
6972 mdb_page_free(mc->mc_txn->mt_env, copy);
6975 /* Adjust other cursors pointing to mp */
6976 MDB_cursor *m2, *m3;
6977 MDB_dbi dbi = mc->mc_dbi;
6978 int fixup = NUMKEYS(mp);
6980 if (mc->mc_flags & C_SUB)
6983 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6984 if (mc->mc_flags & C_SUB)
6985 m3 = &m2->mc_xcursor->mx_cursor;
6990 if (!(m3->mc_flags & C_INITIALIZED))
6992 if (m3->mc_flags & C_SPLITTING)
6997 for (k=m3->mc_top; k>=0; k--) {
6998 m3->mc_ki[k+1] = m3->mc_ki[k];
6999 m3->mc_pg[k+1] = m3->mc_pg[k];
7001 if (m3->mc_ki[0] >= split_indx) {
7006 m3->mc_pg[0] = mc->mc_pg[0];
7010 if (m3->mc_pg[mc->mc_top] == mp) {
7011 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7012 m3->mc_ki[mc->mc_top]++;
7013 if (m3->mc_ki[mc->mc_top] >= fixup) {
7014 m3->mc_pg[mc->mc_top] = rp;
7015 m3->mc_ki[mc->mc_top] -= fixup;
7016 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7018 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7019 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7028 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7029 MDB_val *key, MDB_val *data, unsigned int flags)
7034 assert(key != NULL);
7035 assert(data != NULL);
7037 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7040 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7044 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7048 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7051 mdb_cursor_init(&mc, txn, dbi, &mx);
7052 return mdb_cursor_put(&mc, key, data, flags);
7056 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7058 if ((flag & CHANGEABLE) != flag)
7061 env->me_flags |= flag;
7063 env->me_flags &= ~flag;
7068 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7073 *arg = env->me_flags;
7078 mdb_env_get_path(MDB_env *env, const char **arg)
7083 *arg = env->me_path;
7087 /** Common code for #mdb_stat() and #mdb_env_stat().
7088 * @param[in] env the environment to operate in.
7089 * @param[in] db the #MDB_db record containing the stats to return.
7090 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7091 * @return 0, this function always succeeds.
7094 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7096 arg->ms_psize = env->me_psize;
7097 arg->ms_depth = db->md_depth;
7098 arg->ms_branch_pages = db->md_branch_pages;
7099 arg->ms_leaf_pages = db->md_leaf_pages;
7100 arg->ms_overflow_pages = db->md_overflow_pages;
7101 arg->ms_entries = db->md_entries;
7106 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7110 if (env == NULL || arg == NULL)
7113 toggle = mdb_env_pick_meta(env);
7115 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7119 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7123 if (env == NULL || arg == NULL)
7126 toggle = mdb_env_pick_meta(env);
7127 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7128 arg->me_mapsize = env->me_mapsize;
7129 arg->me_maxreaders = env->me_maxreaders;
7130 arg->me_numreaders = env->me_numreaders;
7131 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7132 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7136 /** Set the default comparison functions for a database.
7137 * Called immediately after a database is opened to set the defaults.
7138 * The user can then override them with #mdb_set_compare() or
7139 * #mdb_set_dupsort().
7140 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7141 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7144 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7146 uint16_t f = txn->mt_dbs[dbi].md_flags;
7148 txn->mt_dbxs[dbi].md_cmp =
7149 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7150 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7152 txn->mt_dbxs[dbi].md_dcmp =
7153 !(f & MDB_DUPSORT) ? 0 :
7154 ((f & MDB_INTEGERDUP)
7155 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7156 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7159 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7164 int rc, dbflag, exact;
7165 unsigned int unused = 0;
7168 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7169 mdb_default_cmp(txn, FREE_DBI);
7172 if ((flags & VALID_FLAGS) != flags)
7178 if (flags & PERSISTENT_FLAGS) {
7179 uint16_t f2 = flags & PERSISTENT_FLAGS;
7180 /* make sure flag changes get committed */
7181 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7182 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7183 txn->mt_flags |= MDB_TXN_DIRTY;
7186 mdb_default_cmp(txn, MAIN_DBI);
7190 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7191 mdb_default_cmp(txn, MAIN_DBI);
7194 /* Is the DB already open? */
7196 for (i=2; i<txn->mt_numdbs; i++) {
7197 if (!txn->mt_dbxs[i].md_name.mv_size) {
7198 /* Remember this free slot */
7199 if (!unused) unused = i;
7202 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7203 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7209 /* If no free slot and max hit, fail */
7210 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7211 return MDB_DBS_FULL;
7213 /* Cannot mix named databases with some mainDB flags */
7214 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7215 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7217 /* Find the DB info */
7218 dbflag = DB_NEW|DB_VALID;
7221 key.mv_data = (void *)name;
7222 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7223 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7224 if (rc == MDB_SUCCESS) {
7225 /* make sure this is actually a DB */
7226 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7227 if (!(node->mn_flags & F_SUBDATA))
7229 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7230 /* Create if requested */
7232 data.mv_size = sizeof(MDB_db);
7233 data.mv_data = &dummy;
7234 memset(&dummy, 0, sizeof(dummy));
7235 dummy.md_root = P_INVALID;
7236 dummy.md_flags = flags & PERSISTENT_FLAGS;
7237 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7241 /* OK, got info, add to table */
7242 if (rc == MDB_SUCCESS) {
7243 unsigned int slot = unused ? unused : txn->mt_numdbs;
7244 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7245 txn->mt_dbxs[slot].md_name.mv_size = len;
7246 txn->mt_dbxs[slot].md_rel = NULL;
7247 txn->mt_dbflags[slot] = dbflag;
7248 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7250 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7251 mdb_default_cmp(txn, slot);
7260 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7262 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7265 if (txn->mt_dbflags[dbi] & DB_STALE) {
7268 /* Stale, must read the DB's root. cursor_init does it for us. */
7269 mdb_cursor_init(&mc, txn, dbi, &mx);
7271 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7274 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7277 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7279 ptr = env->me_dbxs[dbi].md_name.mv_data;
7280 env->me_dbxs[dbi].md_name.mv_data = NULL;
7281 env->me_dbxs[dbi].md_name.mv_size = 0;
7282 env->me_dbflags[dbi] = 0;
7286 /** Add all the DB's pages to the free list.
7287 * @param[in] mc Cursor on the DB to free.
7288 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7289 * @return 0 on success, non-zero on failure.
7292 mdb_drop0(MDB_cursor *mc, int subs)
7296 rc = mdb_page_search(mc, NULL, 0);
7297 if (rc == MDB_SUCCESS) {
7298 MDB_txn *txn = mc->mc_txn;
7303 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7304 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7307 mdb_cursor_copy(mc, &mx);
7308 while (mc->mc_snum > 0) {
7309 MDB_page *mp = mc->mc_pg[mc->mc_top];
7310 unsigned n = NUMKEYS(mp);
7312 for (i=0; i<n; i++) {
7313 ni = NODEPTR(mp, i);
7314 if (ni->mn_flags & F_BIGDATA) {
7318 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7319 rc = mdb_page_get(txn, pg, &omp, NULL);
7322 assert(IS_OVERFLOW(omp));
7323 ovpages = omp->mp_pages;
7324 for (j=0; j<ovpages; j++) {
7325 mdb_midl_append(&txn->mt_free_pgs, pg);
7328 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7329 mdb_xcursor_init1(mc, ni);
7330 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7336 for (i=0; i<n; i++) {
7338 ni = NODEPTR(mp, i);
7341 mdb_midl_append(&txn->mt_free_pgs, pg);
7346 mc->mc_ki[mc->mc_top] = i;
7347 rc = mdb_cursor_sibling(mc, 1);
7349 /* no more siblings, go back to beginning
7350 * of previous level.
7354 for (i=1; i<mc->mc_snum; i++) {
7356 mc->mc_pg[i] = mx.mc_pg[i];
7361 mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
7366 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7368 MDB_cursor *mc, *m2;
7371 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7374 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7377 rc = mdb_cursor_open(txn, dbi, &mc);
7381 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7382 /* Invalidate the dropped DB's cursors */
7383 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7384 m2->mc_flags &= ~C_INITIALIZED;
7388 /* Can't delete the main DB */
7389 if (del && dbi > MAIN_DBI) {
7390 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7392 txn->mt_dbflags[dbi] = DB_STALE;
7393 mdb_dbi_close(txn->mt_env, dbi);
7396 /* reset the DB record, mark it dirty */
7397 txn->mt_dbflags[dbi] |= DB_DIRTY;
7398 txn->mt_dbs[dbi].md_depth = 0;
7399 txn->mt_dbs[dbi].md_branch_pages = 0;
7400 txn->mt_dbs[dbi].md_leaf_pages = 0;
7401 txn->mt_dbs[dbi].md_overflow_pages = 0;
7402 txn->mt_dbs[dbi].md_entries = 0;
7403 txn->mt_dbs[dbi].md_root = P_INVALID;
7405 txn->mt_flags |= MDB_TXN_DIRTY;
7408 mdb_cursor_close(mc);
7412 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7414 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7417 txn->mt_dbxs[dbi].md_cmp = cmp;
7421 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7423 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7426 txn->mt_dbxs[dbi].md_dcmp = cmp;
7430 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7432 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7435 txn->mt_dbxs[dbi].md_rel = rel;
7439 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7441 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7444 txn->mt_dbxs[dbi].md_relctx = ctx;