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 txnid_t mf_pglast; /**< ID of last old page record we used */
932 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
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_cursor *mc, unsigned num)
1261 MDB_env *env = mc->mc_txn->mt_env;
1262 MDB_page *ret = env->me_dpages;
1263 size_t sz = env->me_psize;
1266 VGMEMP_ALLOC(env, ret, sz);
1267 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1268 env->me_dpages = ret->mp_next;
1274 if ((ret = malloc(sz)) != NULL) {
1275 VGMEMP_ALLOC(env, ret, sz);
1280 /** Free a single page.
1281 * Saves single pages to a list, for future reuse.
1282 * (This is not used for multi-page overflow pages.)
1285 mdb_page_free(MDB_env *env, MDB_page *mp)
1287 mp->mp_next = env->me_dpages;
1288 VGMEMP_FREE(env, mp);
1289 env->me_dpages = mp;
1292 /* Return all dirty pages to dpage list */
1294 mdb_dlist_free(MDB_txn *txn)
1296 MDB_env *env = txn->mt_env;
1297 MDB_ID2L dl = txn->mt_u.dirty_list;
1298 unsigned i, n = dl[0].mid;
1300 for (i = 1; i <= n; i++) {
1301 MDB_page *dp = dl[i].mptr;
1302 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1303 mdb_page_free(env, dp);
1305 /* large pages just get freed directly */
1306 VGMEMP_FREE(env, dp);
1313 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1315 mdb_find_oldest(MDB_txn *txn)
1318 txnid_t mr, oldest = txn->mt_txnid - 1;
1319 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1320 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1330 /** Allocate pages for writing.
1331 * If there are free pages available from older transactions, they
1332 * will be re-used first. Otherwise a new page will be allocated.
1333 * @param[in] mc cursor A cursor handle identifying the transaction and
1334 * database for which we are allocating.
1335 * @param[in] num the number of pages to allocate.
1336 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1337 * will always be satisfied by a single contiguous chunk of memory.
1338 * @return 0 on success, non-zero on failure.
1341 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1343 MDB_txn *txn = mc->mc_txn;
1345 pgno_t pgno = P_INVALID;
1347 txnid_t oldest = 0, last;
1352 /* If our dirty list is already full, we can't do anything */
1353 if (txn->mt_dirty_room == 0)
1354 return MDB_TXN_FULL;
1356 /* The free list won't have any content at all until txn 2 has
1357 * committed. The pages freed by txn 2 will be unreferenced
1358 * after txn 3 commits, and so will be safe to re-use in txn 4.
1360 if (txn->mt_txnid > 3) {
1361 if (!txn->mt_env->me_pghead &&
1362 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1363 /* See if there's anything in the free DB */
1369 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1370 if (!txn->mt_env->me_pglast) {
1371 mdb_page_search(&m2, NULL, 0);
1372 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1373 kptr = (txnid_t *)NODEKEY(leaf);
1378 last = txn->mt_env->me_pglast + 1;
1380 key.mv_data = &last;
1381 key.mv_size = sizeof(last);
1382 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1385 last = *(txnid_t *)key.mv_data;
1389 oldest = mdb_find_oldest(txn);
1391 if (oldest > last) {
1392 /* It's usable, grab it.
1396 if (!txn->mt_env->me_pglast) {
1397 mdb_node_read(txn, leaf, &data);
1399 idl = (MDB_ID *) data.mv_data;
1400 /* We might have a zero-length IDL due to freelist growth
1401 * during a prior commit
1404 txn->mt_env->me_pglast = last;
1407 mop = mdb_midl_alloc(idl[0]);
1410 txn->mt_env->me_pglast = last;
1411 txn->mt_env->me_pghead = mop;
1412 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1417 DPRINTF("IDL read txn %zu root %zu num %zu",
1418 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1419 for (i = idl[0]; i; i--)
1420 DPRINTF("IDL %zu", idl[i]);
1426 if (txn->mt_env->me_pghead) {
1427 pgno_t *mop = txn->mt_env->me_pghead;
1430 int retry = 1, readit = 0, n2 = num-1;
1431 unsigned int i, j, k;
1433 /* If current list is too short, must fetch more and coalesce */
1434 if (mop[0] < (unsigned)num)
1437 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1439 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1440 /* If on freelist, don't try to read more. If what we have
1441 * right now isn't enough just use new pages.
1442 * TODO: get all of this working. Many circular dependencies...
1444 if (mc->mc_dbi == FREE_DBI) {
1453 last = txn->mt_env->me_pglast + 1;
1455 /* We haven't hit the readers list yet? */
1457 oldest = mdb_find_oldest(txn);
1460 /* There's nothing we can use on the freelist */
1461 if (oldest - last < 1)
1464 key.mv_data = &last;
1465 key.mv_size = sizeof(last);
1466 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1468 if (rc == MDB_NOTFOUND)
1472 last = *(txnid_t*)key.mv_data;
1475 idl = (MDB_ID *) data.mv_data;
1476 mop2 = mdb_midl_alloc(idl[0] + mop[0]);
1479 /* merge in sorted order */
1480 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1482 while (i>0 || j>0) {
1483 if (i && idl[i] < mop[j])
1484 mop2[k--] = idl[i--];
1486 mop2[k--] = mop[j--];
1488 txn->mt_env->me_pglast = last;
1489 mdb_midl_free(txn->mt_env->me_pghead);
1490 txn->mt_env->me_pghead = mop2;
1492 /* Keep trying to read until we have enough */
1493 if (mop[0] < (unsigned)num) {
1498 /* current list has enough pages, but are they contiguous? */
1499 for (i=mop[0]; i>=(unsigned)num; i--) {
1500 if (mop[i-n2] == mop[i] + n2) {
1503 /* move any stragglers down */
1504 for (j=i+num; j<=mop[0]; j++)
1511 /* Stop if we succeeded, or no retries */
1512 if (!retry || pgno != P_INVALID)
1518 /* peel pages off tail, so we only have to truncate the list */
1519 pgno = MDB_IDL_LAST(mop);
1522 if (MDB_IDL_IS_ZERO(mop)) {
1523 mdb_midl_free(txn->mt_env->me_pghead);
1524 txn->mt_env->me_pghead = NULL;
1529 if (pgno == P_INVALID) {
1530 /* DB size is maxed out */
1531 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1532 DPUTS("DB size maxed out");
1533 return MDB_MAP_FULL;
1536 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1537 if (pgno == P_INVALID) {
1538 pgno = txn->mt_next_pgno;
1539 txn->mt_next_pgno += num;
1541 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1544 if (!(np = mdb_page_malloc(mc, num)))
1546 if (pgno == P_INVALID) {
1547 np->mp_pgno = txn->mt_next_pgno;
1548 txn->mt_next_pgno += num;
1553 mid.mid = np->mp_pgno;
1555 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1556 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1558 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1560 txn->mt_dirty_room--;
1566 /** Copy the used portions of a non-overflow page.
1567 * @param[in] dst page to copy into
1568 * @param[in] src page to copy from
1569 * @param[in] psize size of a page
1572 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1574 enum { Align = sizeof(pgno_t) };
1575 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1577 /* If page isn't full, just copy the used portion. Adjust
1578 * alignment so memcpy may copy words instead of bytes.
1580 if ((unused &= -Align) && !IS_LEAF2(src)) {
1582 memcpy(dst, src, (lower + (Align-1)) & -Align);
1583 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1586 memcpy(dst, src, psize - unused);
1590 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1591 * @param[in] mc cursor pointing to the page to be touched
1592 * @return 0 on success, non-zero on failure.
1595 mdb_page_touch(MDB_cursor *mc)
1597 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1598 MDB_cursor *m2, *m3;
1603 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1604 if ((rc = mdb_page_alloc(mc, 1, &np)))
1607 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi,mp->mp_pgno,pgno);
1608 assert(mp->mp_pgno != pgno);
1609 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1610 /* Update the parent page, if any, to point to the new page */
1612 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1613 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1614 SETPGNO(node, pgno);
1616 mc->mc_db->md_root = pgno;
1618 } else if (mc->mc_txn->mt_parent && !(mp->mp_flags & P_SUBP)) {
1619 MDB_ID2 mid, *dl = mc->mc_txn->mt_u.dirty_list;
1621 /* If txn has a parent, make sure the page is in our
1625 unsigned x = mdb_mid2l_search(dl, pgno);
1626 if (x <= dl[0].mid && dl[x].mid == pgno) {
1629 mc->mc_pg[mc->mc_top] = np;
1633 assert(dl[0].mid < MDB_IDL_UM_MAX);
1635 np = mdb_page_malloc(mc, 1);
1640 mdb_mid2l_insert(dl, &mid);
1645 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1647 np->mp_flags |= P_DIRTY;
1649 /* Adjust cursors pointing to mp */
1650 mc->mc_pg[mc->mc_top] = np;
1652 if (mc->mc_flags & C_SUB) {
1654 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1655 m3 = &m2->mc_xcursor->mx_cursor;
1656 if (m3->mc_snum < mc->mc_snum) continue;
1657 if (m3->mc_pg[mc->mc_top] == mp)
1658 m3->mc_pg[mc->mc_top] = np;
1661 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1662 if (m2->mc_snum < mc->mc_snum) continue;
1663 if (m2->mc_pg[mc->mc_top] == mp) {
1664 m2->mc_pg[mc->mc_top] = np;
1665 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1666 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1668 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1669 if (!(leaf->mn_flags & F_SUBDATA))
1670 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1679 mdb_env_sync(MDB_env *env, int force)
1682 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1683 if (env->me_flags & MDB_WRITEMAP) {
1684 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1685 ? MS_ASYNC : MS_SYNC;
1686 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1689 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1693 if (MDB_FDATASYNC(env->me_fd))
1700 /** Make shadow copies of all of parent txn's cursors */
1702 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1704 MDB_cursor *mc, *m2;
1705 unsigned int i, j, size;
1707 for (i=0;i<src->mt_numdbs; i++) {
1708 if (src->mt_cursors[i]) {
1709 size = sizeof(MDB_cursor);
1710 if (src->mt_cursors[i]->mc_xcursor)
1711 size += sizeof(MDB_xcursor);
1712 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1719 mc->mc_db = &dst->mt_dbs[i];
1720 mc->mc_dbx = m2->mc_dbx;
1721 mc->mc_dbflag = &dst->mt_dbflags[i];
1722 mc->mc_snum = m2->mc_snum;
1723 mc->mc_top = m2->mc_top;
1724 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1725 for (j=0; j<mc->mc_snum; j++) {
1726 mc->mc_pg[j] = m2->mc_pg[j];
1727 mc->mc_ki[j] = m2->mc_ki[j];
1729 if (m2->mc_xcursor) {
1730 MDB_xcursor *mx, *mx2;
1731 mx = (MDB_xcursor *)(mc+1);
1732 mc->mc_xcursor = mx;
1733 mx2 = m2->mc_xcursor;
1734 mx->mx_db = mx2->mx_db;
1735 mx->mx_dbx = mx2->mx_dbx;
1736 mx->mx_dbflag = mx2->mx_dbflag;
1737 mx->mx_cursor.mc_txn = dst;
1738 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1739 mx->mx_cursor.mc_db = &mx->mx_db;
1740 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1741 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1742 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1743 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1744 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1745 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1746 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1747 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1750 mc->mc_xcursor = NULL;
1752 mc->mc_next = dst->mt_cursors[i];
1753 dst->mt_cursors[i] = mc;
1760 /** Close this write txn's cursors, after optionally merging its shadow
1761 * cursors back into parent's.
1762 * @param[in] txn the transaction handle.
1763 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1764 * @return 0 on success, non-zero on failure.
1767 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1769 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1772 for (i = txn->mt_numdbs; --i >= 0; ) {
1773 for (mc = cursors[i]; mc; mc = next) {
1775 if (mc->mc_flags & merge) {
1776 MDB_cursor *m2 = mc->mc_orig;
1777 m2->mc_snum = mc->mc_snum;
1778 m2->mc_top = mc->mc_top;
1779 for (j = mc->mc_snum; --j >= 0; ) {
1780 m2->mc_pg[j] = mc->mc_pg[j];
1781 m2->mc_ki[j] = mc->mc_ki[j];
1784 if (mc->mc_flags & C_ALLOCD)
1792 mdb_txn_reset0(MDB_txn *txn);
1794 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1795 * @param[in] txn the transaction handle to initialize
1796 * @return 0 on success, non-zero on failure.
1799 mdb_txn_renew0(MDB_txn *txn)
1801 MDB_env *env = txn->mt_env;
1804 int rc, new_notls = 0;
1807 txn->mt_numdbs = env->me_numdbs;
1808 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1810 if (txn->mt_flags & MDB_TXN_RDONLY) {
1811 if (!env->me_txns) {
1812 i = mdb_env_pick_meta(env);
1813 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1814 txn->mt_u.reader = NULL;
1816 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1817 pthread_getspecific(env->me_txkey);
1819 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1820 return MDB_BAD_RSLOT;
1822 pid_t pid = env->me_pid;
1823 pthread_t tid = pthread_self();
1826 for (i=0; i<env->me_txns->mti_numreaders; i++)
1827 if (env->me_txns->mti_readers[i].mr_pid == 0)
1829 if (i == env->me_maxreaders) {
1830 UNLOCK_MUTEX_R(env);
1831 return MDB_READERS_FULL;
1833 env->me_txns->mti_readers[i].mr_pid = pid;
1834 env->me_txns->mti_readers[i].mr_tid = tid;
1835 if (i >= env->me_txns->mti_numreaders)
1836 env->me_txns->mti_numreaders = i+1;
1837 /* Save numreaders for un-mutexed mdb_env_close() */
1838 env->me_numreaders = env->me_txns->mti_numreaders;
1839 UNLOCK_MUTEX_R(env);
1840 r = &env->me_txns->mti_readers[i];
1841 new_notls = (env->me_flags & MDB_NOTLS);
1842 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1847 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1848 txn->mt_u.reader = r;
1850 txn->mt_toggle = txn->mt_txnid & 1;
1851 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1855 txn->mt_txnid = env->me_txns->mti_txnid;
1856 txn->mt_toggle = txn->mt_txnid & 1;
1857 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1860 if (txn->mt_txnid == mdb_debug_start)
1863 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1864 txn->mt_u.dirty_list = env->me_dirty_list;
1865 txn->mt_u.dirty_list[0].mid = 0;
1866 txn->mt_free_pgs = env->me_free_pgs;
1867 txn->mt_free_pgs[0] = 0;
1871 /* Copy the DB info and flags */
1872 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1873 for (i=2; i<txn->mt_numdbs; i++) {
1874 x = env->me_dbflags[i];
1875 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1876 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1878 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1880 if (env->me_maxpg < txn->mt_next_pgno) {
1881 mdb_txn_reset0(txn);
1883 txn->mt_u.reader->mr_pid = 0;
1884 txn->mt_u.reader = NULL;
1886 return MDB_MAP_RESIZED;
1893 mdb_txn_renew(MDB_txn *txn)
1897 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1900 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1901 DPUTS("environment had fatal error, must shutdown!");
1905 rc = mdb_txn_renew0(txn);
1906 if (rc == MDB_SUCCESS) {
1907 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1908 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1909 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1915 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1919 int rc, size, tsize = sizeof(MDB_txn);
1921 if (env->me_flags & MDB_FATAL_ERROR) {
1922 DPUTS("environment had fatal error, must shutdown!");
1925 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1928 /* Nested transactions: Max 1 child, write txns only, no writemap */
1929 if (parent->mt_child ||
1930 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1931 (env->me_flags & MDB_WRITEMAP))
1935 tsize = sizeof(MDB_ntxn);
1937 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1938 if (!(flags & MDB_RDONLY))
1939 size += env->me_maxdbs * sizeof(MDB_cursor *);
1941 if ((txn = calloc(1, size)) == NULL) {
1942 DPRINTF("calloc: %s", strerror(ErrCode()));
1945 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1946 if (flags & MDB_RDONLY) {
1947 txn->mt_flags |= MDB_TXN_RDONLY;
1948 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1950 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1951 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1957 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1958 if (!txn->mt_u.dirty_list ||
1959 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1961 free(txn->mt_u.dirty_list);
1965 txn->mt_txnid = parent->mt_txnid;
1966 txn->mt_toggle = parent->mt_toggle;
1967 txn->mt_dirty_room = parent->mt_dirty_room;
1968 txn->mt_u.dirty_list[0].mid = 0;
1969 txn->mt_free_pgs[0] = 0;
1970 txn->mt_next_pgno = parent->mt_next_pgno;
1971 parent->mt_child = txn;
1972 txn->mt_parent = parent;
1973 txn->mt_numdbs = parent->mt_numdbs;
1974 txn->mt_dbxs = parent->mt_dbxs;
1975 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1976 /* Copy parent's mt_dbflags, but clear DB_NEW */
1977 for (i=0; i<txn->mt_numdbs; i++)
1978 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1980 ntxn = (MDB_ntxn *)txn;
1981 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1982 if (env->me_pghead) {
1983 size = MDB_IDL_SIZEOF(env->me_pghead);
1984 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1986 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1991 rc = mdb_cursor_shadow(parent, txn);
1993 mdb_txn_reset0(txn);
1995 rc = mdb_txn_renew0(txn);
2001 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
2002 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2003 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2009 /** Export or close DBI handles opened in this txn. */
2011 mdb_dbis_update(MDB_txn *txn, int keep)
2014 MDB_dbi n = txn->mt_numdbs;
2015 MDB_env *env = txn->mt_env;
2016 unsigned char *tdbflags = txn->mt_dbflags;
2018 for (i = n; --i >= 2;) {
2019 if (tdbflags[i] & DB_NEW) {
2021 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2023 char *ptr = env->me_dbxs[i].md_name.mv_data;
2024 env->me_dbxs[i].md_name.mv_data = NULL;
2025 env->me_dbxs[i].md_name.mv_size = 0;
2026 env->me_dbflags[i] = 0;
2031 if (keep && env->me_numdbs < n)
2035 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2036 * May be called twice for readonly txns: First reset it, then abort.
2037 * @param[in] txn the transaction handle to reset
2040 mdb_txn_reset0(MDB_txn *txn)
2042 MDB_env *env = txn->mt_env;
2044 /* Close any DBI handles opened in this txn */
2045 mdb_dbis_update(txn, 0);
2047 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2048 if (txn->mt_u.reader) {
2049 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2050 if (!(env->me_flags & MDB_NOTLS))
2051 txn->mt_u.reader = NULL; /* txn does not own reader */
2053 txn->mt_numdbs = 0; /* close nothing if called again */
2054 txn->mt_dbxs = NULL; /* mark txn as reset */
2056 mdb_cursors_close(txn, 0);
2058 if (!(env->me_flags & MDB_WRITEMAP)) {
2059 mdb_dlist_free(txn);
2061 mdb_midl_free(env->me_pghead);
2063 if (txn->mt_parent) {
2064 txn->mt_parent->mt_child = NULL;
2065 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2066 mdb_midl_free(txn->mt_free_pgs);
2067 free(txn->mt_u.dirty_list);
2070 if (mdb_midl_shrink(&txn->mt_free_pgs))
2071 env->me_free_pgs = txn->mt_free_pgs;
2074 txn->mt_env->me_pghead = NULL;
2075 txn->mt_env->me_pglast = 0;
2078 /* The writer mutex was locked in mdb_txn_begin. */
2079 UNLOCK_MUTEX_W(env);
2084 mdb_txn_reset(MDB_txn *txn)
2089 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2090 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2091 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2093 /* This call is only valid for read-only txns */
2094 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2097 mdb_txn_reset0(txn);
2101 mdb_txn_abort(MDB_txn *txn)
2106 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2107 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2108 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2111 mdb_txn_abort(txn->mt_child);
2113 mdb_txn_reset0(txn);
2114 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2115 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2116 txn->mt_u.reader->mr_pid = 0;
2121 /** Save the freelist as of this transaction to the freeDB.
2122 * This changes the freelist. Keep trying until it stabilizes.
2125 mdb_freelist_save(MDB_txn *txn)
2127 /* env->me_pghead[] can grow and shrink during this call.
2128 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2129 * Page numbers cannot disappear from txn->mt_free_pgs[].
2132 MDB_env *env = txn->mt_env;
2133 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2134 txnid_t pglast = 0, head_id = 0;
2135 pgno_t freecnt = 0, *free_pgs, *mop;
2136 ssize_t head_room = 0, total_room = 0, mop_len;
2138 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2140 if (env->me_pghead || env->me_pglast) {
2141 /* Make sure first page of freeDB is touched and on freelist */
2142 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2143 if (rc && rc != MDB_NOTFOUND)
2148 /* Come back here after each Put() in case freelist changed */
2151 /* If using records from freeDB which we have not yet
2152 * deleted, delete them and any we reserved for me_pghead.
2154 while (pglast < env->me_pglast) {
2155 rc = mdb_cursor_first(&mc, &key, NULL);
2158 pglast = head_id = *(txnid_t *)key.mv_data;
2159 total_room = head_room = 0;
2160 assert(pglast <= env->me_pglast);
2161 rc = mdb_cursor_del(&mc, 0);
2166 /* Save the IDL of pages freed by this txn, to a single record */
2167 if (freecnt < txn->mt_free_pgs[0]) {
2169 /* Make sure last page of freeDB is touched and on freelist */
2170 key.mv_size = MDB_MAXKEYSIZE+1;
2172 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2173 if (rc && rc != MDB_NOTFOUND)
2176 free_pgs = txn->mt_free_pgs;
2177 /* Write to last page of freeDB */
2178 key.mv_size = sizeof(txn->mt_txnid);
2179 key.mv_data = &txn->mt_txnid;
2181 freecnt = free_pgs[0];
2182 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2183 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2186 /* Retry if mt_free_pgs[] grew during the Put() */
2187 free_pgs = txn->mt_free_pgs;
2188 } while (freecnt < free_pgs[0]);
2189 mdb_midl_sort(free_pgs);
2190 memcpy(data.mv_data, free_pgs, data.mv_size);
2193 unsigned int i = free_pgs[0];
2194 DPRINTF("IDL write txn %zu root %zu num %u",
2195 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2197 DPRINTF("IDL %zu", free_pgs[i]);
2203 mop = env->me_pghead;
2204 mop_len = mop ? mop[0] : 0;
2206 /* Reserve records for me_pghead[]. Split it if multi-page,
2207 * to avoid searching freeDB for a page range. Use keys in
2208 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2210 if (total_room >= mop_len) {
2211 if (total_room == mop_len || --more < 0)
2213 } else if (head_room >= maxfree_1pg && head_id > 1) {
2214 /* Keep current record (overflow page), add a new one */
2218 /* (Re)write {key = head_id, IDL length = head_room} */
2219 total_room -= head_room;
2220 head_room = mop_len - total_room;
2221 if (head_room > maxfree_1pg && head_id > 1) {
2222 /* Overflow multi-page for part of me_pghead */
2223 head_room /= head_id; /* amortize page sizes */
2224 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2225 } else if (head_room < 0) {
2226 /* Rare case, not bothering to delete this record */
2229 key.mv_size = sizeof(head_id);
2230 key.mv_data = &head_id;
2231 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2232 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2235 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2236 total_room += head_room;
2239 /* Fill in the reserved, touched me_pghead records. Avoid write ops
2240 * so they cannot rearrange anything, just read the destinations.
2247 rc = mdb_cursor_first(&mc, &key, &data);
2248 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2249 MDB_IDL dest = data.mv_data;
2250 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2252 assert(len >= 0 && *(txnid_t*)key.mv_data <= env->me_pglast);
2256 memcpy(dest, mop -= len, len * sizeof(MDB_ID));
2257 if (! (mop_len -= len))
2265 mdb_txn_commit(MDB_txn *txn)
2275 assert(txn != NULL);
2276 assert(txn->mt_env != NULL);
2278 if (txn->mt_child) {
2279 rc = mdb_txn_commit(txn->mt_child);
2280 txn->mt_child = NULL;
2289 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2290 mdb_dbis_update(txn, 1);
2291 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2296 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2297 DPUTS("error flag is set, can't commit");
2299 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2304 if (txn->mt_parent) {
2305 MDB_txn *parent = txn->mt_parent;
2309 /* Append our free list to parent's */
2310 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2314 mdb_midl_free(txn->mt_free_pgs);
2316 parent->mt_next_pgno = txn->mt_next_pgno;
2317 parent->mt_flags = txn->mt_flags;
2319 /* Merge our cursors into parent's and close them */
2320 mdb_cursors_close(txn, C_SHADOW);
2322 /* Update parent's DB table. */
2323 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2324 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2325 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2326 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2327 for (i=2; i<txn->mt_numdbs; i++) {
2328 /* preserve parent's DB_NEW status */
2329 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2330 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2333 dst = txn->mt_parent->mt_u.dirty_list;
2334 src = txn->mt_u.dirty_list;
2335 /* Find len = length of merging our dirty list with parent's */
2337 dst[0].mid = 0; /* simplify loops */
2338 if (parent->mt_parent) {
2339 len = x + src[0].mid;
2340 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2341 for (i = x; y && i; y--) {
2342 pgno_t yp = src[y].mid;
2343 while (yp < dst[i].mid)
2345 if (yp == dst[i].mid) {
2350 } else { /* Simplify the above for single-ancestor case */
2351 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2353 /* Merge our dirty list with parent's */
2355 for (i = len; y; dst[i--] = src[y--]) {
2356 pgno_t yp = src[y].mid;
2357 while (yp < dst[x].mid)
2358 dst[i--] = dst[x--];
2359 if (yp == dst[x].mid)
2360 free(dst[x--].mptr);
2364 free(txn->mt_u.dirty_list);
2365 parent->mt_dirty_room = txn->mt_dirty_room;
2367 txn->mt_parent->mt_child = NULL;
2368 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2373 if (txn != env->me_txn) {
2374 DPUTS("attempt to commit unknown transaction");
2379 mdb_cursors_close(txn, 0);
2381 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2384 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2385 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2387 /* Update DB root pointers */
2388 if (txn->mt_numdbs > 2) {
2392 data.mv_size = sizeof(MDB_db);
2394 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2395 for (i = 2; i < txn->mt_numdbs; i++) {
2396 if (txn->mt_dbflags[i] & DB_DIRTY) {
2397 data.mv_data = &txn->mt_dbs[i];
2398 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2405 rc = mdb_freelist_save(txn);
2409 mdb_midl_free(env->me_pghead);
2410 env->me_pghead = NULL;
2411 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2412 if (mdb_midl_shrink(&txn->mt_free_pgs))
2413 env->me_free_pgs = txn->mt_free_pgs;
2420 if (env->me_flags & MDB_WRITEMAP) {
2421 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2422 dp = txn->mt_u.dirty_list[i].mptr;
2423 /* clear dirty flag */
2424 dp->mp_flags &= ~P_DIRTY;
2426 txn->mt_u.dirty_list[0].mid = 0;
2430 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2436 /* Windows actually supports scatter/gather I/O, but only on
2437 * unbuffered file handles. Since we're relying on the OS page
2438 * cache for all our data, that's self-defeating. So we just
2439 * write pages one at a time. We use the ov structure to set
2440 * the write offset, to at least save the overhead of a Seek
2444 memset(&ov, 0, sizeof(ov));
2445 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2447 dp = txn->mt_u.dirty_list[i].mptr;
2448 DPRINTF("committing page %zu", dp->mp_pgno);
2449 size = dp->mp_pgno * env->me_psize;
2450 ov.Offset = size & 0xffffffff;
2451 ov.OffsetHigh = size >> 16;
2452 ov.OffsetHigh >>= 16;
2453 /* clear dirty flag */
2454 dp->mp_flags &= ~P_DIRTY;
2455 wsize = env->me_psize;
2456 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2457 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2460 DPRINTF("WriteFile: %d", n);
2467 struct iovec iov[MDB_COMMIT_PAGES];
2471 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2472 dp = txn->mt_u.dirty_list[i].mptr;
2473 if (dp->mp_pgno != next) {
2475 rc = writev(env->me_fd, iov, n);
2479 DPUTS("short write, filesystem full?");
2481 DPRINTF("writev: %s", strerror(n));
2488 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2491 DPRINTF("committing page %zu", dp->mp_pgno);
2492 iov[n].iov_len = env->me_psize;
2493 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2494 iov[n].iov_base = (char *)dp;
2495 size += iov[n].iov_len;
2496 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2497 /* clear dirty flag */
2498 dp->mp_flags &= ~P_DIRTY;
2499 if (++n >= MDB_COMMIT_PAGES) {
2509 rc = writev(env->me_fd, iov, n);
2513 DPUTS("short write, filesystem full?");
2515 DPRINTF("writev: %s", strerror(n));
2522 mdb_dlist_free(txn);
2525 if ((n = mdb_env_sync(env, 0)) != 0 ||
2526 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2534 mdb_dbis_update(txn, 1);
2536 UNLOCK_MUTEX_W(env);
2546 /** Read the environment parameters of a DB environment before
2547 * mapping it into memory.
2548 * @param[in] env the environment handle
2549 * @param[out] meta address of where to store the meta information
2550 * @return 0 on success, non-zero on failure.
2553 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2560 /* We don't know the page size yet, so use a minimum value.
2561 * Read both meta pages so we can use the latest one.
2564 for (i=0; i<2; i++) {
2566 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2568 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2573 else if (rc != MDB_PAGESIZE) {
2577 DPRINTF("read: %s", strerror(err));
2581 p = (MDB_page *)&pbuf;
2583 if (!F_ISSET(p->mp_flags, P_META)) {
2584 DPRINTF("page %zu not a meta page", p->mp_pgno);
2589 if (m->mm_magic != MDB_MAGIC) {
2590 DPUTS("meta has invalid magic");
2594 if (m->mm_version != MDB_VERSION) {
2595 DPRINTF("database is version %u, expected version %u",
2596 m->mm_version, MDB_VERSION);
2597 return MDB_VERSION_MISMATCH;
2601 if (m->mm_txnid > meta->mm_txnid)
2602 memcpy(meta, m, sizeof(*m));
2604 memcpy(meta, m, sizeof(*m));
2606 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2608 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2616 /** Write the environment parameters of a freshly created DB environment.
2617 * @param[in] env the environment handle
2618 * @param[out] meta address of where to store the meta information
2619 * @return 0 on success, non-zero on failure.
2622 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2629 DPUTS("writing new meta page");
2631 GET_PAGESIZE(psize);
2633 meta->mm_magic = MDB_MAGIC;
2634 meta->mm_version = MDB_VERSION;
2635 meta->mm_mapsize = env->me_mapsize;
2636 meta->mm_psize = psize;
2637 meta->mm_last_pg = 1;
2638 meta->mm_flags = env->me_flags & 0xffff;
2639 meta->mm_flags |= MDB_INTEGERKEY;
2640 meta->mm_dbs[0].md_root = P_INVALID;
2641 meta->mm_dbs[1].md_root = P_INVALID;
2643 p = calloc(2, psize);
2645 p->mp_flags = P_META;
2648 memcpy(m, meta, sizeof(*meta));
2650 q = (MDB_page *)((char *)p + psize);
2653 q->mp_flags = P_META;
2656 memcpy(m, meta, sizeof(*meta));
2661 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2662 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2663 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2666 lseek(env->me_fd, 0, SEEK_SET);
2667 rc = write(env->me_fd, p, psize * 2);
2668 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2674 /** Update the environment info to commit a transaction.
2675 * @param[in] txn the transaction that's being committed
2676 * @return 0 on success, non-zero on failure.
2679 mdb_env_write_meta(MDB_txn *txn)
2682 MDB_meta meta, metab, *mp;
2684 int rc, len, toggle;
2691 assert(txn != NULL);
2692 assert(txn->mt_env != NULL);
2694 toggle = !txn->mt_toggle;
2695 DPRINTF("writing meta page %d for root page %zu",
2696 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2699 mp = env->me_metas[toggle];
2701 if (env->me_flags & MDB_WRITEMAP) {
2702 /* Persist any increases of mapsize config */
2703 if (env->me_mapsize > mp->mm_mapsize)
2704 mp->mm_mapsize = env->me_mapsize;
2705 mp->mm_dbs[0] = txn->mt_dbs[0];
2706 mp->mm_dbs[1] = txn->mt_dbs[1];
2707 mp->mm_last_pg = txn->mt_next_pgno - 1;
2708 mp->mm_txnid = txn->mt_txnid;
2709 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2710 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2713 ptr += env->me_psize;
2714 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2721 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2722 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2724 ptr = (char *)&meta;
2725 if (env->me_mapsize > mp->mm_mapsize) {
2726 /* Persist any increases of mapsize config */
2727 meta.mm_mapsize = env->me_mapsize;
2728 off = offsetof(MDB_meta, mm_mapsize);
2730 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2732 len = sizeof(MDB_meta) - off;
2735 meta.mm_dbs[0] = txn->mt_dbs[0];
2736 meta.mm_dbs[1] = txn->mt_dbs[1];
2737 meta.mm_last_pg = txn->mt_next_pgno - 1;
2738 meta.mm_txnid = txn->mt_txnid;
2741 off += env->me_psize;
2744 /* Write to the SYNC fd */
2745 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2746 env->me_fd : env->me_mfd;
2749 memset(&ov, 0, sizeof(ov));
2751 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2754 rc = pwrite(mfd, ptr, len, off);
2759 DPUTS("write failed, disk error?");
2760 /* On a failure, the pagecache still contains the new data.
2761 * Write some old data back, to prevent it from being used.
2762 * Use the non-SYNC fd; we know it will fail anyway.
2764 meta.mm_last_pg = metab.mm_last_pg;
2765 meta.mm_txnid = metab.mm_txnid;
2767 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2769 r2 = pwrite(env->me_fd, ptr, len, off);
2772 env->me_flags |= MDB_FATAL_ERROR;
2776 /* Memory ordering issues are irrelevant; since the entire writer
2777 * is wrapped by wmutex, all of these changes will become visible
2778 * after the wmutex is unlocked. Since the DB is multi-version,
2779 * readers will get consistent data regardless of how fresh or
2780 * how stale their view of these values is.
2782 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2787 /** Check both meta pages to see which one is newer.
2788 * @param[in] env the environment handle
2789 * @return meta toggle (0 or 1).
2792 mdb_env_pick_meta(const MDB_env *env)
2794 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2798 mdb_env_create(MDB_env **env)
2802 e = calloc(1, sizeof(MDB_env));
2806 e->me_maxreaders = DEFAULT_READERS;
2807 e->me_maxdbs = e->me_numdbs = 2;
2808 e->me_fd = INVALID_HANDLE_VALUE;
2809 e->me_lfd = INVALID_HANDLE_VALUE;
2810 e->me_mfd = INVALID_HANDLE_VALUE;
2811 #ifdef MDB_USE_POSIX_SEM
2812 e->me_rmutex = SEM_FAILED;
2813 e->me_wmutex = SEM_FAILED;
2815 e->me_pid = getpid();
2816 VGMEMP_CREATE(e,0,0);
2822 mdb_env_set_mapsize(MDB_env *env, size_t size)
2826 env->me_mapsize = size;
2828 env->me_maxpg = env->me_mapsize / env->me_psize;
2833 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2837 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2842 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2844 if (env->me_map || readers < 1)
2846 env->me_maxreaders = readers;
2851 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2853 if (!env || !readers)
2855 *readers = env->me_maxreaders;
2859 /** Further setup required for opening an MDB environment
2862 mdb_env_open2(MDB_env *env)
2864 unsigned int flags = env->me_flags;
2865 int i, newenv = 0, prot;
2869 memset(&meta, 0, sizeof(meta));
2871 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2874 DPUTS("new mdbenv");
2878 /* Was a mapsize configured? */
2879 if (!env->me_mapsize) {
2880 /* If this is a new environment, take the default,
2881 * else use the size recorded in the existing env.
2883 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2884 } else if (env->me_mapsize < meta.mm_mapsize) {
2885 /* If the configured size is smaller, make sure it's
2886 * still big enough. Silently round up to minimum if not.
2888 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2889 if (env->me_mapsize < minsize)
2890 env->me_mapsize = minsize;
2896 LONG sizelo, sizehi;
2897 sizelo = env->me_mapsize & 0xffffffff;
2898 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2900 /* Windows won't create mappings for zero length files.
2901 * Just allocate the maxsize right now.
2904 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2905 if (!SetEndOfFile(env->me_fd))
2907 SetFilePointer(env->me_fd, 0, NULL, 0);
2909 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2910 PAGE_READWRITE : PAGE_READONLY,
2911 sizehi, sizelo, NULL);
2914 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2915 FILE_MAP_WRITE : FILE_MAP_READ,
2916 0, 0, env->me_mapsize, meta.mm_address);
2924 if (flags & MDB_WRITEMAP) {
2926 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2929 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2931 if (env->me_map == MAP_FAILED) {
2935 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2937 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2939 #ifdef POSIX_MADV_RANDOM
2940 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2941 #endif /* POSIX_MADV_RANDOM */
2942 #endif /* MADV_RANDOM */
2946 if (flags & MDB_FIXEDMAP)
2947 meta.mm_address = env->me_map;
2948 i = mdb_env_init_meta(env, &meta);
2949 if (i != MDB_SUCCESS) {
2952 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2953 /* Can happen because the address argument to mmap() is just a
2954 * hint. mmap() can pick another, e.g. if the range is in use.
2955 * The MAP_FIXED flag would prevent that, but then mmap could
2956 * instead unmap existing pages to make room for the new map.
2958 return EBUSY; /* TODO: Make a new MDB_* error code? */
2960 env->me_psize = meta.mm_psize;
2961 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2962 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2964 env->me_maxpg = env->me_mapsize / env->me_psize;
2966 p = (MDB_page *)env->me_map;
2967 env->me_metas[0] = METADATA(p);
2968 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2972 int toggle = mdb_env_pick_meta(env);
2973 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2975 DPRINTF("opened database version %u, pagesize %u",
2976 env->me_metas[0]->mm_version, env->me_psize);
2977 DPRINTF("using meta page %d", toggle);
2978 DPRINTF("depth: %u", db->md_depth);
2979 DPRINTF("entries: %zu", db->md_entries);
2980 DPRINTF("branch pages: %zu", db->md_branch_pages);
2981 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2982 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2983 DPRINTF("root: %zu", db->md_root);
2991 /** Release a reader thread's slot in the reader lock table.
2992 * This function is called automatically when a thread exits.
2993 * @param[in] ptr This points to the slot in the reader lock table.
2996 mdb_env_reader_dest(void *ptr)
2998 MDB_reader *reader = ptr;
3004 /** Junk for arranging thread-specific callbacks on Windows. This is
3005 * necessarily platform and compiler-specific. Windows supports up
3006 * to 1088 keys. Let's assume nobody opens more than 64 environments
3007 * in a single process, for now. They can override this if needed.
3009 #ifndef MAX_TLS_KEYS
3010 #define MAX_TLS_KEYS 64
3012 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3013 static int mdb_tls_nkeys;
3015 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3019 case DLL_PROCESS_ATTACH: break;
3020 case DLL_THREAD_ATTACH: break;
3021 case DLL_THREAD_DETACH:
3022 for (i=0; i<mdb_tls_nkeys; i++) {
3023 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3024 mdb_env_reader_dest(r);
3027 case DLL_PROCESS_DETACH: break;
3032 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3034 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3038 /* Force some symbol references.
3039 * _tls_used forces the linker to create the TLS directory if not already done
3040 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3042 #pragma comment(linker, "/INCLUDE:_tls_used")
3043 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3044 #pragma const_seg(".CRT$XLB")
3045 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3046 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3049 #pragma comment(linker, "/INCLUDE:__tls_used")
3050 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3051 #pragma data_seg(".CRT$XLB")
3052 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3054 #endif /* WIN 32/64 */
3055 #endif /* !__GNUC__ */
3058 /** Downgrade the exclusive lock on the region back to shared */
3060 mdb_env_share_locks(MDB_env *env, int *excl)
3062 int rc = 0, toggle = mdb_env_pick_meta(env);
3064 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3069 /* First acquire a shared lock. The Unlock will
3070 * then release the existing exclusive lock.
3072 memset(&ov, 0, sizeof(ov));
3073 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3076 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3082 struct flock lock_info;
3083 /* The shared lock replaces the existing lock */
3084 memset((void *)&lock_info, 0, sizeof(lock_info));
3085 lock_info.l_type = F_RDLCK;
3086 lock_info.l_whence = SEEK_SET;
3087 lock_info.l_start = 0;
3088 lock_info.l_len = 1;
3089 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3090 (rc = ErrCode()) == EINTR) ;
3091 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3098 /** Try to get exlusive lock, otherwise shared.
3099 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3102 mdb_env_excl_lock(MDB_env *env, int *excl)
3106 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3110 memset(&ov, 0, sizeof(ov));
3111 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3118 struct flock lock_info;
3119 memset((void *)&lock_info, 0, sizeof(lock_info));
3120 lock_info.l_type = F_WRLCK;
3121 lock_info.l_whence = SEEK_SET;
3122 lock_info.l_start = 0;
3123 lock_info.l_len = 1;
3124 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3125 (rc = ErrCode()) == EINTR) ;
3129 # ifdef MDB_USE_POSIX_SEM
3130 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3133 lock_info.l_type = F_RDLCK;
3134 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3135 (rc = ErrCode()) == EINTR) ;
3143 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3145 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3147 * @(#) $Revision: 5.1 $
3148 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3149 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3151 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3155 * Please do not copyright this code. This code is in the public domain.
3157 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3158 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3159 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3160 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3161 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3162 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3163 * PERFORMANCE OF THIS SOFTWARE.
3166 * chongo <Landon Curt Noll> /\oo/\
3167 * http://www.isthe.com/chongo/
3169 * Share and Enjoy! :-)
3172 typedef unsigned long long mdb_hash_t;
3173 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3175 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3176 * @param[in] str string to hash
3177 * @param[in] hval initial value for hash
3178 * @return 64 bit hash
3180 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3181 * hval arg on the first call.
3184 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3186 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3187 unsigned char *end = s + val->mv_size;
3189 * FNV-1a hash each octet of the string
3192 /* xor the bottom with the current octet */
3193 hval ^= (mdb_hash_t)*s++;
3195 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3196 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3197 (hval << 7) + (hval << 8) + (hval << 40);
3199 /* return our new hash value */
3203 /** Hash the string and output the hash in hex.
3204 * @param[in] str string to hash
3205 * @param[out] hexbuf an array of 17 chars to hold the hash
3208 mdb_hash_hex(MDB_val *val, char *hexbuf)
3211 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3212 for (i=0; i<8; i++) {
3213 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3219 /** Open and/or initialize the lock region for the environment.
3220 * @param[in] env The MDB environment.
3221 * @param[in] lpath The pathname of the file used for the lock region.
3222 * @param[in] mode The Unix permissions for the file, if we create it.
3223 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3224 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3225 * @return 0 on success, non-zero on failure.
3228 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3231 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3233 # define MDB_ERRCODE_ROFS EROFS
3234 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3235 # define MDB_CLOEXEC O_CLOEXEC
3238 # define MDB_CLOEXEC 0
3245 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3246 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3247 FILE_ATTRIBUTE_NORMAL, NULL);
3249 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3251 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3253 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3258 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3259 /* Lose record locks when exec*() */
3260 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3261 fcntl(env->me_lfd, F_SETFD, fdflags);
3264 if (!(env->me_flags & MDB_NOTLS)) {
3265 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3268 env->me_flags |= MDB_ENV_TXKEY;
3270 /* Windows TLS callbacks need help finding their TLS info. */
3271 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3275 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3279 /* Try to get exclusive lock. If we succeed, then
3280 * nobody is using the lock region and we should initialize it.
3282 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3285 size = GetFileSize(env->me_lfd, NULL);
3287 size = lseek(env->me_lfd, 0, SEEK_END);
3289 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3290 if (size < rsize && *excl > 0) {
3292 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3293 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3295 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3299 size = rsize - sizeof(MDB_txninfo);
3300 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3305 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3307 if (!mh) goto fail_errno;
3308 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3310 if (!env->me_txns) goto fail_errno;
3312 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3314 if (m == MAP_FAILED) goto fail_errno;
3320 BY_HANDLE_FILE_INFORMATION stbuf;
3329 if (!mdb_sec_inited) {
3330 InitializeSecurityDescriptor(&mdb_null_sd,
3331 SECURITY_DESCRIPTOR_REVISION);
3332 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3333 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3334 mdb_all_sa.bInheritHandle = FALSE;
3335 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3338 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3339 idbuf.volume = stbuf.dwVolumeSerialNumber;
3340 idbuf.nhigh = stbuf.nFileIndexHigh;
3341 idbuf.nlow = stbuf.nFileIndexLow;
3342 val.mv_data = &idbuf;
3343 val.mv_size = sizeof(idbuf);
3344 mdb_hash_hex(&val, hexbuf);
3345 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3346 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3347 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3348 if (!env->me_rmutex) goto fail_errno;
3349 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3350 if (!env->me_wmutex) goto fail_errno;
3351 #elif defined(MDB_USE_POSIX_SEM)
3360 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3361 idbuf.dev = stbuf.st_dev;
3362 idbuf.ino = stbuf.st_ino;
3363 val.mv_data = &idbuf;
3364 val.mv_size = sizeof(idbuf);
3365 mdb_hash_hex(&val, hexbuf);
3366 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3367 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3368 /* Clean up after a previous run, if needed: Try to
3369 * remove both semaphores before doing anything else.
3371 sem_unlink(env->me_txns->mti_rmname);
3372 sem_unlink(env->me_txns->mti_wmname);
3373 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3374 O_CREAT|O_EXCL, mode, 1);
3375 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3376 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3377 O_CREAT|O_EXCL, mode, 1);
3378 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3379 #else /* MDB_USE_POSIX_SEM */
3380 pthread_mutexattr_t mattr;
3382 if ((rc = pthread_mutexattr_init(&mattr))
3383 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3384 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3385 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3387 pthread_mutexattr_destroy(&mattr);
3388 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3390 env->me_txns->mti_version = MDB_VERSION;
3391 env->me_txns->mti_magic = MDB_MAGIC;
3392 env->me_txns->mti_txnid = 0;
3393 env->me_txns->mti_numreaders = 0;
3396 if (env->me_txns->mti_magic != MDB_MAGIC) {
3397 DPUTS("lock region has invalid magic");
3401 if (env->me_txns->mti_version != MDB_VERSION) {
3402 DPRINTF("lock region is version %u, expected version %u",
3403 env->me_txns->mti_version, MDB_VERSION);
3404 rc = MDB_VERSION_MISMATCH;
3408 if (rc != EACCES && rc != EAGAIN) {
3412 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3413 if (!env->me_rmutex) goto fail_errno;
3414 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3415 if (!env->me_wmutex) goto fail_errno;
3416 #elif defined(MDB_USE_POSIX_SEM)
3417 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3418 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3419 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3420 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3431 /** The name of the lock file in the DB environment */
3432 #define LOCKNAME "/lock.mdb"
3433 /** The name of the data file in the DB environment */
3434 #define DATANAME "/data.mdb"
3435 /** The suffix of the lock file when no subdir is used */
3436 #define LOCKSUFF "-lock"
3437 /** Only a subset of the @ref mdb_env flags can be changed
3438 * at runtime. Changing other flags requires closing the
3439 * environment and re-opening it with the new flags.
3441 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3442 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3445 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3447 int oflags, rc, len, excl = -1;
3448 char *lpath, *dpath;
3450 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3454 if (flags & MDB_NOSUBDIR) {
3455 rc = len + sizeof(LOCKSUFF) + len + 1;
3457 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3462 if (flags & MDB_NOSUBDIR) {
3463 dpath = lpath + len + sizeof(LOCKSUFF);
3464 sprintf(lpath, "%s" LOCKSUFF, path);
3465 strcpy(dpath, path);
3467 dpath = lpath + len + sizeof(LOCKNAME);
3468 sprintf(lpath, "%s" LOCKNAME, path);
3469 sprintf(dpath, "%s" DATANAME, path);
3473 flags |= env->me_flags;
3474 if (flags & MDB_RDONLY) {
3475 /* silently ignore WRITEMAP when we're only getting read access */
3476 flags &= ~MDB_WRITEMAP;
3478 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3479 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3482 env->me_flags = flags |= MDB_ENV_ACTIVE;
3486 env->me_path = strdup(path);
3487 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3488 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3489 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3494 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3499 if (F_ISSET(flags, MDB_RDONLY)) {
3500 oflags = GENERIC_READ;
3501 len = OPEN_EXISTING;
3503 oflags = GENERIC_READ|GENERIC_WRITE;
3506 mode = FILE_ATTRIBUTE_NORMAL;
3507 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3508 NULL, len, mode, NULL);
3510 if (F_ISSET(flags, MDB_RDONLY))
3513 oflags = O_RDWR | O_CREAT;
3515 env->me_fd = open(dpath, oflags, mode);
3517 if (env->me_fd == INVALID_HANDLE_VALUE) {
3522 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3523 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3524 env->me_mfd = env->me_fd;
3526 /* Synchronous fd for meta writes. Needed even with
3527 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3530 env->me_mfd = CreateFile(dpath, oflags,
3531 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3532 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3534 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3536 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3541 DPRINTF("opened dbenv %p", (void *) env);
3543 rc = mdb_env_share_locks(env, &excl);
3549 mdb_env_close0(env, excl);
3555 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3557 mdb_env_close0(MDB_env *env, int excl)
3561 if (!(env->me_flags & MDB_ENV_ACTIVE))
3564 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3565 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3566 free(env->me_dbxs[i].md_name.mv_data);
3568 free(env->me_dbflags);
3571 free(env->me_dirty_list);
3572 if (env->me_free_pgs)
3573 mdb_midl_free(env->me_free_pgs);
3575 if (env->me_flags & MDB_ENV_TXKEY) {
3576 pthread_key_delete(env->me_txkey);
3578 /* Delete our key from the global list */
3579 for (i=0; i<mdb_tls_nkeys; i++)
3580 if (mdb_tls_keys[i] == env->me_txkey) {
3581 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3589 munmap(env->me_map, env->me_mapsize);
3591 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3593 if (env->me_fd != INVALID_HANDLE_VALUE)
3596 pid_t pid = env->me_pid;
3597 /* Clearing readers is done in this function because
3598 * me_txkey with its destructor must be disabled first.
3600 for (i = env->me_numreaders; --i >= 0; )
3601 if (env->me_txns->mti_readers[i].mr_pid == pid)
3602 env->me_txns->mti_readers[i].mr_pid = 0;
3604 if (env->me_rmutex) {
3605 CloseHandle(env->me_rmutex);
3606 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3608 /* Windows automatically destroys the mutexes when
3609 * the last handle closes.
3611 #elif defined(MDB_USE_POSIX_SEM)
3612 if (env->me_rmutex != SEM_FAILED) {
3613 sem_close(env->me_rmutex);
3614 if (env->me_wmutex != SEM_FAILED)
3615 sem_close(env->me_wmutex);
3616 /* If we have the filelock: If we are the
3617 * only remaining user, clean up semaphores.
3620 mdb_env_excl_lock(env, &excl);
3622 sem_unlink(env->me_txns->mti_rmname);
3623 sem_unlink(env->me_txns->mti_wmname);
3627 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3629 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3632 /* Unlock the lockfile. Windows would have unlocked it
3633 * after closing anyway, but not necessarily at once.
3635 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3641 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3645 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3647 MDB_txn *txn = NULL;
3652 /* Do the lock/unlock of the reader mutex before starting the
3653 * write txn. Otherwise other read txns could block writers.
3655 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3660 /* We must start the actual read txn after blocking writers */
3661 mdb_txn_reset0(txn);
3663 /* Temporarily block writers until we snapshot the meta pages */
3666 rc = mdb_txn_renew0(txn);
3668 UNLOCK_MUTEX_W(env);
3673 wsize = env->me_psize * 2;
3677 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3678 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3681 rc = write(fd, env->me_map, wsize);
3682 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3685 UNLOCK_MUTEX_W(env);
3690 ptr = env->me_map + wsize;
3691 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3692 #define MAX_WRITE 2147483648U
3696 if (wsize > MAX_WRITE)
3700 rc = WriteFile(fd, ptr, w2, &len, NULL);
3701 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3710 if (wsize > MAX_WRITE)
3714 wres = write(fd, ptr, w2);
3715 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3728 mdb_env_copy(MDB_env *env, const char *path)
3732 HANDLE newfd = INVALID_HANDLE_VALUE;
3734 if (env->me_flags & MDB_NOSUBDIR) {
3735 lpath = (char *)path;
3738 len += sizeof(DATANAME);
3739 lpath = malloc(len);
3742 sprintf(lpath, "%s" DATANAME, path);
3745 /* The destination path must exist, but the destination file must not.
3746 * We don't want the OS to cache the writes, since the source data is
3747 * already in the OS cache.
3750 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3751 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3753 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3759 if (!(env->me_flags & MDB_NOSUBDIR))
3761 if (newfd == INVALID_HANDLE_VALUE) {
3766 #ifdef F_NOCACHE /* __APPLE__ */
3767 rc = fcntl(newfd, F_NOCACHE, 1);
3774 rc = mdb_env_copyfd(env, newfd);
3777 if (newfd != INVALID_HANDLE_VALUE)
3784 mdb_env_close(MDB_env *env)
3791 VGMEMP_DESTROY(env);
3792 while ((dp = env->me_dpages) != NULL) {
3793 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3794 env->me_dpages = dp->mp_next;
3798 mdb_env_close0(env, 0);
3802 /** Compare two items pointing at aligned size_t's */
3804 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3806 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3807 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3810 /** Compare two items pointing at aligned int's */
3812 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3814 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3815 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3818 /** Compare two items pointing at ints of unknown alignment.
3819 * Nodes and keys are guaranteed to be 2-byte aligned.
3822 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3824 #if BYTE_ORDER == LITTLE_ENDIAN
3825 unsigned short *u, *c;
3828 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3829 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3832 } while(!x && u > (unsigned short *)a->mv_data);
3835 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3839 /** Compare two items lexically */
3841 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3848 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3854 diff = memcmp(a->mv_data, b->mv_data, len);
3855 return diff ? diff : len_diff<0 ? -1 : len_diff;
3858 /** Compare two items in reverse byte order */
3860 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3862 const unsigned char *p1, *p2, *p1_lim;
3866 p1_lim = (const unsigned char *)a->mv_data;
3867 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3868 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3870 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3876 while (p1 > p1_lim) {
3877 diff = *--p1 - *--p2;
3881 return len_diff<0 ? -1 : len_diff;
3884 /** Search for key within a page, using binary search.
3885 * Returns the smallest entry larger or equal to the key.
3886 * If exactp is non-null, stores whether the found entry was an exact match
3887 * in *exactp (1 or 0).
3888 * Updates the cursor index with the index of the found entry.
3889 * If no entry larger or equal to the key is found, returns NULL.
3892 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3894 unsigned int i = 0, nkeys;
3897 MDB_page *mp = mc->mc_pg[mc->mc_top];
3898 MDB_node *node = NULL;
3903 nkeys = NUMKEYS(mp);
3908 COPY_PGNO(pgno, mp->mp_pgno);
3909 DPRINTF("searching %u keys in %s %spage %zu",
3910 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3917 low = IS_LEAF(mp) ? 0 : 1;
3919 cmp = mc->mc_dbx->md_cmp;
3921 /* Branch pages have no data, so if using integer keys,
3922 * alignment is guaranteed. Use faster mdb_cmp_int.
3924 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3925 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3932 nodekey.mv_size = mc->mc_db->md_pad;
3933 node = NODEPTR(mp, 0); /* fake */
3934 while (low <= high) {
3935 i = (low + high) >> 1;
3936 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3937 rc = cmp(key, &nodekey);
3938 DPRINTF("found leaf index %u [%s], rc = %i",
3939 i, DKEY(&nodekey), rc);
3948 while (low <= high) {
3949 i = (low + high) >> 1;
3951 node = NODEPTR(mp, i);
3952 nodekey.mv_size = NODEKSZ(node);
3953 nodekey.mv_data = NODEKEY(node);
3955 rc = cmp(key, &nodekey);
3958 DPRINTF("found leaf index %u [%s], rc = %i",
3959 i, DKEY(&nodekey), rc);
3961 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3962 i, DKEY(&nodekey), NODEPGNO(node), rc);
3973 if (rc > 0) { /* Found entry is less than the key. */
3974 i++; /* Skip to get the smallest entry larger than key. */
3976 node = NODEPTR(mp, i);
3979 *exactp = (rc == 0);
3980 /* store the key index */
3981 mc->mc_ki[mc->mc_top] = i;
3983 /* There is no entry larger or equal to the key. */
3986 /* nodeptr is fake for LEAF2 */
3992 mdb_cursor_adjust(MDB_cursor *mc, func)
3996 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3997 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4004 /** Pop a page off the top of the cursor's stack. */
4006 mdb_cursor_pop(MDB_cursor *mc)
4009 #ifndef MDB_DEBUG_SKIP
4010 MDB_page *top = mc->mc_pg[mc->mc_top];
4016 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
4017 mc->mc_dbi, (void *) mc);
4021 /** Push a page onto the top of the cursor's stack. */
4023 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4025 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4026 mc->mc_dbi, (void *) mc);
4028 if (mc->mc_snum >= CURSOR_STACK) {
4029 assert(mc->mc_snum < CURSOR_STACK);
4030 return MDB_CURSOR_FULL;
4033 mc->mc_top = mc->mc_snum++;
4034 mc->mc_pg[mc->mc_top] = mp;
4035 mc->mc_ki[mc->mc_top] = 0;
4040 /** Find the address of the page corresponding to a given page number.
4041 * @param[in] txn the transaction for this access.
4042 * @param[in] pgno the page number for the page to retrieve.
4043 * @param[out] ret address of a pointer where the page's address will be stored.
4044 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4045 * @return 0 on success, non-zero on failure.
4048 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4053 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4054 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4059 MDB_ID2L dl = tx2->mt_u.dirty_list;
4061 unsigned x = mdb_mid2l_search(dl, pgno);
4062 if (x <= dl[0].mid && dl[x].mid == pgno) {
4068 } while ((tx2 = tx2->mt_parent) != NULL);
4071 if (pgno < txn->mt_next_pgno) {
4073 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4075 DPRINTF("page %zu not found", pgno);
4077 return MDB_PAGE_NOTFOUND;
4087 /** Search for the page a given key should be in.
4088 * Pushes parent pages on the cursor stack. This function continues a
4089 * search on a cursor that has already been initialized. (Usually by
4090 * #mdb_page_search() but also by #mdb_node_move().)
4091 * @param[in,out] mc the cursor for this operation.
4092 * @param[in] key the key to search for. If NULL, search for the lowest
4093 * page. (This is used by #mdb_cursor_first().)
4094 * @param[in] modify If true, visited pages are updated with new page numbers.
4095 * @return 0 on success, non-zero on failure.
4098 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4100 MDB_page *mp = mc->mc_pg[mc->mc_top];
4105 while (IS_BRANCH(mp)) {
4109 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4110 assert(NUMKEYS(mp) > 1);
4111 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4113 if (key == NULL) /* Initialize cursor to first page. */
4115 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4116 /* cursor to last page */
4120 node = mdb_node_search(mc, key, &exact);
4122 i = NUMKEYS(mp) - 1;
4124 i = mc->mc_ki[mc->mc_top];
4133 DPRINTF("following index %u for key [%s]",
4135 assert(i < NUMKEYS(mp));
4136 node = NODEPTR(mp, i);
4138 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4141 mc->mc_ki[mc->mc_top] = i;
4142 if ((rc = mdb_cursor_push(mc, mp)))
4146 if ((rc = mdb_page_touch(mc)) != 0)
4148 mp = mc->mc_pg[mc->mc_top];
4153 DPRINTF("internal error, index points to a %02X page!?",
4155 return MDB_CORRUPTED;
4158 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4159 key ? DKEY(key) : NULL);
4164 /** Search for the lowest key under the current branch page.
4165 * This just bypasses a NUMKEYS check in the current page
4166 * before calling mdb_page_search_root(), because the callers
4167 * are all in situations where the current page is known to
4171 mdb_page_search_lowest(MDB_cursor *mc)
4173 MDB_page *mp = mc->mc_pg[mc->mc_top];
4174 MDB_node *node = NODEPTR(mp, 0);
4177 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4180 mc->mc_ki[mc->mc_top] = 0;
4181 if ((rc = mdb_cursor_push(mc, mp)))
4183 return mdb_page_search_root(mc, NULL, 0);
4186 /** Search for the page a given key should be in.
4187 * Pushes parent pages on the cursor stack. This function just sets up
4188 * the search; it finds the root page for \b mc's database and sets this
4189 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4190 * called to complete the search.
4191 * @param[in,out] mc the cursor for this operation.
4192 * @param[in] key the key to search for. If NULL, search for the lowest
4193 * page. (This is used by #mdb_cursor_first().)
4194 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4195 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4196 * @return 0 on success, non-zero on failure.
4199 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4204 /* Make sure the txn is still viable, then find the root from
4205 * the txn's db table.
4207 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4208 DPUTS("transaction has failed, must abort");
4211 /* Make sure we're using an up-to-date root */
4212 if (mc->mc_dbi > MAIN_DBI) {
4213 if ((*mc->mc_dbflag & DB_STALE) ||
4214 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4216 unsigned char dbflag = 0;
4217 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4218 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4221 if (*mc->mc_dbflag & DB_STALE) {
4225 MDB_node *leaf = mdb_node_search(&mc2,
4226 &mc->mc_dbx->md_name, &exact);
4228 return MDB_NOTFOUND;
4229 mdb_node_read(mc->mc_txn, leaf, &data);
4230 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4232 /* The txn may not know this DBI, or another process may
4233 * have dropped and recreated the DB with other flags.
4235 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4236 return MDB_INCOMPATIBLE;
4237 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4239 if (flags & MDB_PS_MODIFY)
4241 *mc->mc_dbflag &= ~DB_STALE;
4242 *mc->mc_dbflag |= dbflag;
4245 root = mc->mc_db->md_root;
4247 if (root == P_INVALID) { /* Tree is empty. */
4248 DPUTS("tree is empty");
4249 return MDB_NOTFOUND;
4254 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4255 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4261 DPRINTF("db %u root page %zu has flags 0x%X",
4262 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4264 if (flags & MDB_PS_MODIFY) {
4265 if ((rc = mdb_page_touch(mc)))
4269 if (flags & MDB_PS_ROOTONLY)
4272 return mdb_page_search_root(mc, key, flags);
4276 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4278 MDB_txn *txn = mc->mc_txn;
4279 pgno_t pg = mp->mp_pgno;
4280 unsigned i, ovpages = mp->mp_pages;
4283 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4284 /* If the page is dirty we just acquired it, so we should
4285 * give it back to our current free list, if any.
4286 * Not currently supported in nested txns.
4287 * Otherwise put it onto the list of pages we freed in this txn.
4289 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && txn->mt_env->me_pghead) {
4291 pgno_t *mop = txn->mt_env->me_pghead;
4292 MDB_ID2 *dl, ix, iy;
4293 /* Prepare to insert pg */
4294 j = mop[0] + ovpages;
4296 rc = mdb_midl_grow(&mop, ovpages);
4299 txn->mt_env->me_pghead = mop;
4301 /* Remove from dirty list */
4302 dl = txn->mt_u.dirty_list;
4304 for (ix = dl[x]; ix.mid != pg; ix = iy) {
4311 return MDB_CORRUPTED;
4314 /* Insert in me_pghead */
4315 for (i = mop[0]; i && mop[i] < pg; i--)
4321 for (i=0; i<ovpages; i++) {
4322 mdb_midl_append(&txn->mt_free_pgs, pg);
4326 mc->mc_db->md_overflow_pages -= ovpages;
4330 /** Return the data associated with a given node.
4331 * @param[in] txn The transaction for this operation.
4332 * @param[in] leaf The node being read.
4333 * @param[out] data Updated to point to the node's data.
4334 * @return 0 on success, non-zero on failure.
4337 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4339 MDB_page *omp; /* overflow page */
4343 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4344 data->mv_size = NODEDSZ(leaf);
4345 data->mv_data = NODEDATA(leaf);
4349 /* Read overflow data.
4351 data->mv_size = NODEDSZ(leaf);
4352 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4353 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4354 DPRINTF("read overflow page %zu failed", pgno);
4357 data->mv_data = METADATA(omp);
4363 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4364 MDB_val *key, MDB_val *data)
4373 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4375 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4378 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4382 mdb_cursor_init(&mc, txn, dbi, &mx);
4383 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4386 /** Find a sibling for a page.
4387 * Replaces the page at the top of the cursor's stack with the
4388 * specified sibling, if one exists.
4389 * @param[in] mc The cursor for this operation.
4390 * @param[in] move_right Non-zero if the right sibling is requested,
4391 * otherwise the left sibling.
4392 * @return 0 on success, non-zero on failure.
4395 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4401 if (mc->mc_snum < 2) {
4402 return MDB_NOTFOUND; /* root has no siblings */
4406 DPRINTF("parent page is page %zu, index %u",
4407 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4409 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4410 : (mc->mc_ki[mc->mc_top] == 0)) {
4411 DPRINTF("no more keys left, moving to %s sibling",
4412 move_right ? "right" : "left");
4413 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4414 /* undo cursor_pop before returning */
4421 mc->mc_ki[mc->mc_top]++;
4423 mc->mc_ki[mc->mc_top]--;
4424 DPRINTF("just moving to %s index key %u",
4425 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4427 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4429 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4430 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4433 mdb_cursor_push(mc, mp);
4435 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4440 /** Move the cursor to the next data item. */
4442 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4448 if (mc->mc_flags & C_EOF) {
4449 return MDB_NOTFOUND;
4452 assert(mc->mc_flags & C_INITIALIZED);
4454 mp = mc->mc_pg[mc->mc_top];
4456 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4457 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4458 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4459 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4460 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4461 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4465 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4466 if (op == MDB_NEXT_DUP)
4467 return MDB_NOTFOUND;
4471 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4473 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4474 DPUTS("=====> move to next sibling page");
4475 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4476 mc->mc_flags |= C_EOF;
4477 mc->mc_flags &= ~C_INITIALIZED;
4478 return MDB_NOTFOUND;
4480 mp = mc->mc_pg[mc->mc_top];
4481 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4483 mc->mc_ki[mc->mc_top]++;
4485 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4486 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4489 key->mv_size = mc->mc_db->md_pad;
4490 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4494 assert(IS_LEAF(mp));
4495 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4497 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4498 mdb_xcursor_init1(mc, leaf);
4501 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4504 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4505 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4506 if (rc != MDB_SUCCESS)
4511 MDB_GET_KEY(leaf, key);
4515 /** Move the cursor to the previous data item. */
4517 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4523 assert(mc->mc_flags & C_INITIALIZED);
4525 mp = mc->mc_pg[mc->mc_top];
4527 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4528 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4529 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4530 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4531 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4532 if (op != MDB_PREV || rc == MDB_SUCCESS)
4535 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4536 if (op == MDB_PREV_DUP)
4537 return MDB_NOTFOUND;
4542 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4544 if (mc->mc_ki[mc->mc_top] == 0) {
4545 DPUTS("=====> move to prev sibling page");
4546 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4547 mc->mc_flags &= ~C_INITIALIZED;
4548 return MDB_NOTFOUND;
4550 mp = mc->mc_pg[mc->mc_top];
4551 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4552 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4554 mc->mc_ki[mc->mc_top]--;
4556 mc->mc_flags &= ~C_EOF;
4558 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4559 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4562 key->mv_size = mc->mc_db->md_pad;
4563 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4567 assert(IS_LEAF(mp));
4568 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4570 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4571 mdb_xcursor_init1(mc, leaf);
4574 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4577 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4578 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4579 if (rc != MDB_SUCCESS)
4584 MDB_GET_KEY(leaf, key);
4588 /** Set the cursor on a specific data item. */
4590 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4591 MDB_cursor_op op, int *exactp)
4595 MDB_node *leaf = NULL;
4600 assert(key->mv_size > 0);
4602 /* See if we're already on the right page */
4603 if (mc->mc_flags & C_INITIALIZED) {
4606 mp = mc->mc_pg[mc->mc_top];
4608 mc->mc_ki[mc->mc_top] = 0;
4609 return MDB_NOTFOUND;
4611 if (mp->mp_flags & P_LEAF2) {
4612 nodekey.mv_size = mc->mc_db->md_pad;
4613 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4615 leaf = NODEPTR(mp, 0);
4616 MDB_GET_KEY(leaf, &nodekey);
4618 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4620 /* Probably happens rarely, but first node on the page
4621 * was the one we wanted.
4623 mc->mc_ki[mc->mc_top] = 0;
4630 unsigned int nkeys = NUMKEYS(mp);
4632 if (mp->mp_flags & P_LEAF2) {
4633 nodekey.mv_data = LEAF2KEY(mp,
4634 nkeys-1, nodekey.mv_size);
4636 leaf = NODEPTR(mp, nkeys-1);
4637 MDB_GET_KEY(leaf, &nodekey);
4639 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4641 /* last node was the one we wanted */
4642 mc->mc_ki[mc->mc_top] = nkeys-1;
4648 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4649 /* This is definitely the right page, skip search_page */
4650 if (mp->mp_flags & P_LEAF2) {
4651 nodekey.mv_data = LEAF2KEY(mp,
4652 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4654 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4655 MDB_GET_KEY(leaf, &nodekey);
4657 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4659 /* current node was the one we wanted */
4669 /* If any parents have right-sibs, search.
4670 * Otherwise, there's nothing further.
4672 for (i=0; i<mc->mc_top; i++)
4674 NUMKEYS(mc->mc_pg[i])-1)
4676 if (i == mc->mc_top) {
4677 /* There are no other pages */
4678 mc->mc_ki[mc->mc_top] = nkeys;
4679 return MDB_NOTFOUND;
4683 /* There are no other pages */
4684 mc->mc_ki[mc->mc_top] = 0;
4685 return MDB_NOTFOUND;
4689 rc = mdb_page_search(mc, key, 0);
4690 if (rc != MDB_SUCCESS)
4693 mp = mc->mc_pg[mc->mc_top];
4694 assert(IS_LEAF(mp));
4697 leaf = mdb_node_search(mc, key, exactp);
4698 if (exactp != NULL && !*exactp) {
4699 /* MDB_SET specified and not an exact match. */
4700 return MDB_NOTFOUND;
4704 DPUTS("===> inexact leaf not found, goto sibling");
4705 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4706 return rc; /* no entries matched */
4707 mp = mc->mc_pg[mc->mc_top];
4708 assert(IS_LEAF(mp));
4709 leaf = NODEPTR(mp, 0);
4713 mc->mc_flags |= C_INITIALIZED;
4714 mc->mc_flags &= ~C_EOF;
4717 key->mv_size = mc->mc_db->md_pad;
4718 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4722 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4723 mdb_xcursor_init1(mc, leaf);
4726 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4727 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4728 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4731 if (op == MDB_GET_BOTH) {
4737 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4738 if (rc != MDB_SUCCESS)
4741 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4743 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4745 rc = mc->mc_dbx->md_dcmp(data, &d2);
4747 if (op == MDB_GET_BOTH || rc > 0)
4748 return MDB_NOTFOUND;
4753 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4754 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4759 /* The key already matches in all other cases */
4760 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4761 MDB_GET_KEY(leaf, key);
4762 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4767 /** Move the cursor to the first item in the database. */
4769 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4774 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4775 rc = mdb_page_search(mc, NULL, 0);
4776 if (rc != MDB_SUCCESS)
4779 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4781 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4782 mc->mc_flags |= C_INITIALIZED;
4783 mc->mc_flags &= ~C_EOF;
4785 mc->mc_ki[mc->mc_top] = 0;
4787 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4788 key->mv_size = mc->mc_db->md_pad;
4789 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4794 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4795 mdb_xcursor_init1(mc, leaf);
4796 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4801 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4802 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4806 MDB_GET_KEY(leaf, key);
4810 /** Move the cursor to the last item in the database. */
4812 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4817 if (!(mc->mc_flags & C_EOF)) {
4819 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4822 lkey.mv_size = MDB_MAXKEYSIZE+1;
4823 lkey.mv_data = NULL;
4824 rc = mdb_page_search(mc, &lkey, 0);
4825 if (rc != MDB_SUCCESS)
4828 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4831 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4832 mc->mc_flags |= C_INITIALIZED|C_EOF;
4833 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4835 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4836 key->mv_size = mc->mc_db->md_pad;
4837 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4842 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4843 mdb_xcursor_init1(mc, leaf);
4844 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4849 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4850 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4855 MDB_GET_KEY(leaf, key);
4860 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4869 case MDB_GET_CURRENT:
4870 if (!(mc->mc_flags & C_INITIALIZED)) {
4873 MDB_page *mp = mc->mc_pg[mc->mc_top];
4875 mc->mc_ki[mc->mc_top] = 0;
4881 key->mv_size = mc->mc_db->md_pad;
4882 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4884 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4885 MDB_GET_KEY(leaf, key);
4887 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4888 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4890 rc = mdb_node_read(mc->mc_txn, leaf, data);
4897 case MDB_GET_BOTH_RANGE:
4898 if (data == NULL || mc->mc_xcursor == NULL) {
4906 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4908 } else if (op == MDB_SET_RANGE)
4909 rc = mdb_cursor_set(mc, key, data, op, NULL);
4911 rc = mdb_cursor_set(mc, key, data, op, &exact);
4913 case MDB_GET_MULTIPLE:
4915 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4916 !(mc->mc_flags & C_INITIALIZED)) {
4921 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4922 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4925 case MDB_NEXT_MULTIPLE:
4927 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4931 if (!(mc->mc_flags & C_INITIALIZED))
4932 rc = mdb_cursor_first(mc, key, data);
4934 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4935 if (rc == MDB_SUCCESS) {
4936 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4939 mx = &mc->mc_xcursor->mx_cursor;
4940 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4942 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4943 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4951 case MDB_NEXT_NODUP:
4952 if (!(mc->mc_flags & C_INITIALIZED))
4953 rc = mdb_cursor_first(mc, key, data);
4955 rc = mdb_cursor_next(mc, key, data, op);
4959 case MDB_PREV_NODUP:
4960 if (!(mc->mc_flags & C_INITIALIZED)) {
4961 rc = mdb_cursor_last(mc, key, data);
4964 mc->mc_flags |= C_INITIALIZED;
4965 mc->mc_ki[mc->mc_top]++;
4967 rc = mdb_cursor_prev(mc, key, data, op);
4970 rc = mdb_cursor_first(mc, key, data);
4974 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4975 !(mc->mc_flags & C_INITIALIZED) ||
4976 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4980 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4983 rc = mdb_cursor_last(mc, key, data);
4987 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4988 !(mc->mc_flags & C_INITIALIZED) ||
4989 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4993 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4996 DPRINTF("unhandled/unimplemented cursor operation %u", op);
5004 /** Touch all the pages in the cursor stack.
5005 * Makes sure all the pages are writable, before attempting a write operation.
5006 * @param[in] mc The cursor to operate on.
5009 mdb_cursor_touch(MDB_cursor *mc)
5013 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5016 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5017 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5020 *mc->mc_dbflag |= DB_DIRTY;
5022 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5023 rc = mdb_page_touch(mc);
5027 mc->mc_top = mc->mc_snum-1;
5032 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5035 MDB_node *leaf = NULL;
5036 MDB_val xdata, *rdata, dkey;
5039 int do_sub = 0, insert = 0;
5040 unsigned int mcount = 0;
5044 char dbuf[MDB_MAXKEYSIZE+1];
5045 unsigned int nflags;
5048 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5051 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5054 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5057 #if SIZE_MAX > MAXDATASIZE
5058 if (data->mv_size > MAXDATASIZE)
5062 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
5063 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5067 if (flags == MDB_CURRENT) {
5068 if (!(mc->mc_flags & C_INITIALIZED))
5071 } else if (mc->mc_db->md_root == P_INVALID) {
5073 /* new database, write a root leaf page */
5074 DPUTS("allocating new root leaf page");
5075 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5079 mdb_cursor_push(mc, np);
5080 mc->mc_db->md_root = np->mp_pgno;
5081 mc->mc_db->md_depth++;
5082 *mc->mc_dbflag |= DB_DIRTY;
5083 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5085 np->mp_flags |= P_LEAF2;
5086 mc->mc_flags |= C_INITIALIZED;
5092 if (flags & MDB_APPEND) {
5094 rc = mdb_cursor_last(mc, &k2, &d2);
5096 rc = mc->mc_dbx->md_cmp(key, &k2);
5099 mc->mc_ki[mc->mc_top]++;
5101 /* new key is <= last key */
5106 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5108 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5109 DPRINTF("duplicate key [%s]", DKEY(key));
5111 return MDB_KEYEXIST;
5113 if (rc && rc != MDB_NOTFOUND)
5117 /* Cursor is positioned, now make sure all pages are writable */
5118 rc2 = mdb_cursor_touch(mc);
5123 /* The key already exists */
5124 if (rc == MDB_SUCCESS) {
5125 /* there's only a key anyway, so this is a no-op */
5126 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5127 unsigned int ksize = mc->mc_db->md_pad;
5128 if (key->mv_size != ksize)
5130 if (flags == MDB_CURRENT) {
5131 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5132 memcpy(ptr, key->mv_data, ksize);
5137 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5140 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5141 /* Was a single item before, must convert now */
5143 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5144 /* Just overwrite the current item */
5145 if (flags == MDB_CURRENT)
5148 dkey.mv_size = NODEDSZ(leaf);
5149 dkey.mv_data = NODEDATA(leaf);
5150 #if UINT_MAX < SIZE_MAX
5151 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5152 #ifdef MISALIGNED_OK
5153 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5155 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5158 /* if data matches, ignore it */
5159 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5160 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5162 /* create a fake page for the dup items */
5163 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5164 dkey.mv_data = dbuf;
5165 fp = (MDB_page *)&pbuf;
5166 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5167 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5168 fp->mp_lower = PAGEHDRSZ;
5169 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5170 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5171 fp->mp_flags |= P_LEAF2;
5172 fp->mp_pad = data->mv_size;
5173 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5175 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5176 (dkey.mv_size & 1) + (data->mv_size & 1);
5178 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5181 xdata.mv_size = fp->mp_upper;
5186 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5187 /* See if we need to convert from fake page to subDB */
5189 unsigned int offset;
5192 fp = NODEDATA(leaf);
5193 if (flags == MDB_CURRENT) {
5195 fp->mp_flags |= P_DIRTY;
5196 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5197 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5201 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5202 offset = fp->mp_pad;
5203 if (SIZELEFT(fp) >= offset)
5205 offset *= 4; /* space for 4 more */
5207 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5209 offset += offset & 1;
5210 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5211 offset >= mc->mc_txn->mt_env->me_nodemax) {
5212 /* yes, convert it */
5214 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5215 dummy.md_pad = fp->mp_pad;
5216 dummy.md_flags = MDB_DUPFIXED;
5217 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5218 dummy.md_flags |= MDB_INTEGERKEY;
5221 dummy.md_branch_pages = 0;
5222 dummy.md_leaf_pages = 1;
5223 dummy.md_overflow_pages = 0;
5224 dummy.md_entries = NUMKEYS(fp);
5226 xdata.mv_size = sizeof(MDB_db);
5227 xdata.mv_data = &dummy;
5228 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5230 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5231 flags |= F_DUPDATA|F_SUBDATA;
5232 dummy.md_root = mp->mp_pgno;
5234 /* no, just grow it */
5236 xdata.mv_size = NODEDSZ(leaf) + offset;
5237 xdata.mv_data = &pbuf;
5238 mp = (MDB_page *)&pbuf;
5239 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5242 mp->mp_flags = fp->mp_flags | P_DIRTY;
5243 mp->mp_pad = fp->mp_pad;
5244 mp->mp_lower = fp->mp_lower;
5245 mp->mp_upper = fp->mp_upper + offset;
5247 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5249 nsize = NODEDSZ(leaf) - fp->mp_upper;
5250 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5251 for (i=0; i<NUMKEYS(fp); i++)
5252 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5254 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5258 /* data is on sub-DB, just store it */
5259 flags |= F_DUPDATA|F_SUBDATA;
5263 /* overflow page overwrites need special handling */
5264 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5267 unsigned psize = mc->mc_txn->mt_env->me_psize;
5268 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5270 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5271 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5273 ovpages = omp->mp_pages;
5275 /* Is the ov page writable and large enough? */
5276 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5277 /* yes, overwrite it. Note in this case we don't
5278 * bother to try shrinking the page if the new data
5279 * is smaller than the overflow threshold.
5282 /* It is writable only in a parent txn */
5283 size_t sz = (size_t) psize * ovpages, off;
5284 MDB_page *np = mdb_page_malloc(mc, ovpages);
5290 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5291 if (!(flags & MDB_RESERVE)) {
5292 /* Copy end of page, adjusting alignment so
5293 * compiler may copy words instead of bytes.
5295 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5296 memcpy((size_t *)((char *)np + off),
5297 (size_t *)((char *)omp + off), sz - off);
5300 memcpy(np, omp, sz); /* Copy beginning of page */
5303 SETDSZ(leaf, data->mv_size);
5304 if (F_ISSET(flags, MDB_RESERVE))
5305 data->mv_data = METADATA(omp);
5307 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5310 mdb_ovpage_free(mc, omp);
5312 } else if (NODEDSZ(leaf) == data->mv_size) {
5313 /* same size, just replace it. Note that we could
5314 * also reuse this node if the new data is smaller,
5315 * but instead we opt to shrink the node in that case.
5317 if (F_ISSET(flags, MDB_RESERVE))
5318 data->mv_data = NODEDATA(leaf);
5319 else if (data->mv_size)
5320 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5322 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5325 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5326 mc->mc_db->md_entries--;
5328 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5335 nflags = flags & NODE_ADD_FLAGS;
5336 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5337 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5338 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5339 nflags &= ~MDB_APPEND;
5341 nflags |= MDB_SPLIT_REPLACE;
5342 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5344 /* There is room already in this leaf page. */
5345 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5346 if (rc == 0 && !do_sub && insert) {
5347 /* Adjust other cursors pointing to mp */
5348 MDB_cursor *m2, *m3;
5349 MDB_dbi dbi = mc->mc_dbi;
5350 unsigned i = mc->mc_top;
5351 MDB_page *mp = mc->mc_pg[i];
5353 if (mc->mc_flags & C_SUB)
5356 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5357 if (mc->mc_flags & C_SUB)
5358 m3 = &m2->mc_xcursor->mx_cursor;
5361 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5362 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5369 if (rc != MDB_SUCCESS)
5370 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5372 /* Now store the actual data in the child DB. Note that we're
5373 * storing the user data in the keys field, so there are strict
5374 * size limits on dupdata. The actual data fields of the child
5375 * DB are all zero size.
5382 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5383 if (flags & MDB_CURRENT) {
5384 xflags = MDB_CURRENT;
5386 mdb_xcursor_init1(mc, leaf);
5387 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5389 /* converted, write the original data first */
5391 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5395 /* Adjust other cursors pointing to mp */
5397 unsigned i = mc->mc_top;
5398 MDB_page *mp = mc->mc_pg[i];
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[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5403 mdb_xcursor_init1(m2, leaf);
5407 /* we've done our job */
5410 if (flags & MDB_APPENDDUP)
5411 xflags |= MDB_APPEND;
5412 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5413 if (flags & F_SUBDATA) {
5414 void *db = NODEDATA(leaf);
5415 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5418 /* sub-writes might have failed so check rc again.
5419 * Don't increment count if we just replaced an existing item.
5421 if (!rc && !(flags & MDB_CURRENT))
5422 mc->mc_db->md_entries++;
5423 if (flags & MDB_MULTIPLE) {
5425 if (mcount < data[1].mv_size) {
5426 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5427 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5433 /* If we succeeded and the key didn't exist before, make sure
5434 * the cursor is marked valid.
5437 mc->mc_flags |= C_INITIALIZED;
5442 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5447 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5450 if (!(mc->mc_flags & C_INITIALIZED))
5453 rc = mdb_cursor_touch(mc);
5457 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5459 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5460 if (flags != MDB_NODUPDATA) {
5461 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5462 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5464 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5465 /* If sub-DB still has entries, we're done */
5466 if (mc->mc_xcursor->mx_db.md_entries) {
5467 if (leaf->mn_flags & F_SUBDATA) {
5468 /* update subDB info */
5469 void *db = NODEDATA(leaf);
5470 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5473 /* shrink fake page */
5474 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5475 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5476 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5477 /* fix other sub-DB cursors pointed at this fake page */
5478 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5479 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5480 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5481 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5482 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5485 mc->mc_db->md_entries--;
5488 /* otherwise fall thru and delete the sub-DB */
5491 if (leaf->mn_flags & F_SUBDATA) {
5492 /* add all the child DB's pages to the free list */
5493 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5494 if (rc == MDB_SUCCESS) {
5495 mc->mc_db->md_entries -=
5496 mc->mc_xcursor->mx_db.md_entries;
5501 return mdb_cursor_del0(mc, leaf);
5504 /** Allocate and initialize new pages for a database.
5505 * @param[in] mc a cursor on the database being added to.
5506 * @param[in] flags flags defining what type of page is being allocated.
5507 * @param[in] num the number of pages to allocate. This is usually 1,
5508 * unless allocating overflow pages for a large record.
5509 * @param[out] mp Address of a page, or NULL on failure.
5510 * @return 0 on success, non-zero on failure.
5513 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5518 if ((rc = mdb_page_alloc(mc, num, &np)))
5520 DPRINTF("allocated new mpage %zu, page size %u",
5521 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5522 np->mp_flags = flags | P_DIRTY;
5523 np->mp_lower = PAGEHDRSZ;
5524 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5527 mc->mc_db->md_branch_pages++;
5528 else if (IS_LEAF(np))
5529 mc->mc_db->md_leaf_pages++;
5530 else if (IS_OVERFLOW(np)) {
5531 mc->mc_db->md_overflow_pages += num;
5539 /** Calculate the size of a leaf node.
5540 * The size depends on the environment's page size; if a data item
5541 * is too large it will be put onto an overflow page and the node
5542 * size will only include the key and not the data. Sizes are always
5543 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5544 * of the #MDB_node headers.
5545 * @param[in] env The environment handle.
5546 * @param[in] key The key for the node.
5547 * @param[in] data The data for the node.
5548 * @return The number of bytes needed to store the node.
5551 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5555 sz = LEAFSIZE(key, data);
5556 if (sz >= env->me_nodemax) {
5557 /* put on overflow page */
5558 sz -= data->mv_size - sizeof(pgno_t);
5562 return sz + sizeof(indx_t);
5565 /** Calculate the size of a branch node.
5566 * The size should depend on the environment's page size but since
5567 * we currently don't support spilling large keys onto overflow
5568 * pages, it's simply the size of the #MDB_node header plus the
5569 * size of the key. Sizes are always rounded up to an even number
5570 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5571 * @param[in] env The environment handle.
5572 * @param[in] key The key for the node.
5573 * @return The number of bytes needed to store the node.
5576 mdb_branch_size(MDB_env *env, MDB_val *key)
5581 if (sz >= env->me_nodemax) {
5582 /* put on overflow page */
5583 /* not implemented */
5584 /* sz -= key->size - sizeof(pgno_t); */
5587 return sz + sizeof(indx_t);
5590 /** Add a node to the page pointed to by the cursor.
5591 * @param[in] mc The cursor for this operation.
5592 * @param[in] indx The index on the page where the new node should be added.
5593 * @param[in] key The key for the new node.
5594 * @param[in] data The data for the new node, if any.
5595 * @param[in] pgno The page number, if adding a branch node.
5596 * @param[in] flags Flags for the node.
5597 * @return 0 on success, non-zero on failure. Possible errors are:
5599 * <li>ENOMEM - failed to allocate overflow pages for the node.
5600 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5601 * should never happen since all callers already calculate the
5602 * page's free space before calling this function.
5606 mdb_node_add(MDB_cursor *mc, indx_t indx,
5607 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5610 size_t node_size = NODESIZE;
5613 MDB_page *mp = mc->mc_pg[mc->mc_top];
5614 MDB_page *ofp = NULL; /* overflow page */
5617 assert(mp->mp_upper >= mp->mp_lower);
5619 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5620 IS_LEAF(mp) ? "leaf" : "branch",
5621 IS_SUBP(mp) ? "sub-" : "",
5622 mp->mp_pgno, indx, data ? data->mv_size : 0,
5623 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5626 /* Move higher keys up one slot. */
5627 int ksize = mc->mc_db->md_pad, dif;
5628 char *ptr = LEAF2KEY(mp, indx, ksize);
5629 dif = NUMKEYS(mp) - indx;
5631 memmove(ptr+ksize, ptr, dif*ksize);
5632 /* insert new key */
5633 memcpy(ptr, key->mv_data, ksize);
5635 /* Just using these for counting */
5636 mp->mp_lower += sizeof(indx_t);
5637 mp->mp_upper -= ksize - sizeof(indx_t);
5642 node_size += key->mv_size;
5646 if (F_ISSET(flags, F_BIGDATA)) {
5647 /* Data already on overflow page. */
5648 node_size += sizeof(pgno_t);
5649 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5650 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5652 /* Put data on overflow page. */
5653 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5654 data->mv_size, node_size+data->mv_size);
5655 node_size += sizeof(pgno_t);
5656 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5658 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5661 node_size += data->mv_size;
5664 node_size += node_size & 1;
5666 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5667 DPRINTF("not enough room in page %zu, got %u ptrs",
5668 mp->mp_pgno, NUMKEYS(mp));
5669 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5670 mp->mp_upper - mp->mp_lower);
5671 DPRINTF("node size = %zu", node_size);
5672 return MDB_PAGE_FULL;
5675 /* Move higher pointers up one slot. */
5676 for (i = NUMKEYS(mp); i > indx; i--)
5677 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5679 /* Adjust free space offsets. */
5680 ofs = mp->mp_upper - node_size;
5681 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5682 mp->mp_ptrs[indx] = ofs;
5684 mp->mp_lower += sizeof(indx_t);
5686 /* Write the node data. */
5687 node = NODEPTR(mp, indx);
5688 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5689 node->mn_flags = flags;
5691 SETDSZ(node,data->mv_size);
5696 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5701 if (F_ISSET(flags, F_BIGDATA))
5702 memcpy(node->mn_data + key->mv_size, data->mv_data,
5704 else if (F_ISSET(flags, MDB_RESERVE))
5705 data->mv_data = node->mn_data + key->mv_size;
5707 memcpy(node->mn_data + key->mv_size, data->mv_data,
5710 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5712 if (F_ISSET(flags, MDB_RESERVE))
5713 data->mv_data = METADATA(ofp);
5715 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5722 /** Delete the specified node from a page.
5723 * @param[in] mp The page to operate on.
5724 * @param[in] indx The index of the node to delete.
5725 * @param[in] ksize The size of a node. Only used if the page is
5726 * part of a #MDB_DUPFIXED database.
5729 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5732 indx_t i, j, numkeys, ptr;
5739 COPY_PGNO(pgno, mp->mp_pgno);
5740 DPRINTF("delete node %u on %s page %zu", indx,
5741 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5744 assert(indx < NUMKEYS(mp));
5747 int x = NUMKEYS(mp) - 1 - indx;
5748 base = LEAF2KEY(mp, indx, ksize);
5750 memmove(base, base + ksize, x * ksize);
5751 mp->mp_lower -= sizeof(indx_t);
5752 mp->mp_upper += ksize - sizeof(indx_t);
5756 node = NODEPTR(mp, indx);
5757 sz = NODESIZE + node->mn_ksize;
5759 if (F_ISSET(node->mn_flags, F_BIGDATA))
5760 sz += sizeof(pgno_t);
5762 sz += NODEDSZ(node);
5766 ptr = mp->mp_ptrs[indx];
5767 numkeys = NUMKEYS(mp);
5768 for (i = j = 0; i < numkeys; i++) {
5770 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5771 if (mp->mp_ptrs[i] < ptr)
5772 mp->mp_ptrs[j] += sz;
5777 base = (char *)mp + mp->mp_upper;
5778 memmove(base + sz, base, ptr - mp->mp_upper);
5780 mp->mp_lower -= sizeof(indx_t);
5784 /** Compact the main page after deleting a node on a subpage.
5785 * @param[in] mp The main page to operate on.
5786 * @param[in] indx The index of the subpage on the main page.
5789 mdb_node_shrink(MDB_page *mp, indx_t indx)
5796 indx_t i, numkeys, ptr;
5798 node = NODEPTR(mp, indx);
5799 sp = (MDB_page *)NODEDATA(node);
5800 osize = NODEDSZ(node);
5802 delta = sp->mp_upper - sp->mp_lower;
5803 SETDSZ(node, osize - delta);
5804 xp = (MDB_page *)((char *)sp + delta);
5806 /* shift subpage upward */
5808 nsize = NUMKEYS(sp) * sp->mp_pad;
5809 memmove(METADATA(xp), METADATA(sp), nsize);
5812 nsize = osize - sp->mp_upper;
5813 numkeys = NUMKEYS(sp);
5814 for (i=numkeys-1; i>=0; i--)
5815 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5817 xp->mp_upper = sp->mp_lower;
5818 xp->mp_lower = sp->mp_lower;
5819 xp->mp_flags = sp->mp_flags;
5820 xp->mp_pad = sp->mp_pad;
5821 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5823 /* shift lower nodes upward */
5824 ptr = mp->mp_ptrs[indx];
5825 numkeys = NUMKEYS(mp);
5826 for (i = 0; i < numkeys; i++) {
5827 if (mp->mp_ptrs[i] <= ptr)
5828 mp->mp_ptrs[i] += delta;
5831 base = (char *)mp + mp->mp_upper;
5832 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5833 mp->mp_upper += delta;
5836 /** Initial setup of a sorted-dups cursor.
5837 * Sorted duplicates are implemented as a sub-database for the given key.
5838 * The duplicate data items are actually keys of the sub-database.
5839 * Operations on the duplicate data items are performed using a sub-cursor
5840 * initialized when the sub-database is first accessed. This function does
5841 * the preliminary setup of the sub-cursor, filling in the fields that
5842 * depend only on the parent DB.
5843 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5846 mdb_xcursor_init0(MDB_cursor *mc)
5848 MDB_xcursor *mx = mc->mc_xcursor;
5850 mx->mx_cursor.mc_xcursor = NULL;
5851 mx->mx_cursor.mc_txn = mc->mc_txn;
5852 mx->mx_cursor.mc_db = &mx->mx_db;
5853 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5854 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5855 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5856 mx->mx_cursor.mc_snum = 0;
5857 mx->mx_cursor.mc_top = 0;
5858 mx->mx_cursor.mc_flags = C_SUB;
5859 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5860 mx->mx_dbx.md_dcmp = NULL;
5861 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5864 /** Final setup of a sorted-dups cursor.
5865 * Sets up the fields that depend on the data from the main cursor.
5866 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5867 * @param[in] node The data containing the #MDB_db record for the
5868 * sorted-dup database.
5871 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5873 MDB_xcursor *mx = mc->mc_xcursor;
5875 if (node->mn_flags & F_SUBDATA) {
5876 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5877 mx->mx_cursor.mc_pg[0] = 0;
5878 mx->mx_cursor.mc_snum = 0;
5879 mx->mx_cursor.mc_flags = C_SUB;
5881 MDB_page *fp = NODEDATA(node);
5882 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5883 mx->mx_db.md_flags = 0;
5884 mx->mx_db.md_depth = 1;
5885 mx->mx_db.md_branch_pages = 0;
5886 mx->mx_db.md_leaf_pages = 1;
5887 mx->mx_db.md_overflow_pages = 0;
5888 mx->mx_db.md_entries = NUMKEYS(fp);
5889 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5890 mx->mx_cursor.mc_snum = 1;
5891 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5892 mx->mx_cursor.mc_top = 0;
5893 mx->mx_cursor.mc_pg[0] = fp;
5894 mx->mx_cursor.mc_ki[0] = 0;
5895 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5896 mx->mx_db.md_flags = MDB_DUPFIXED;
5897 mx->mx_db.md_pad = fp->mp_pad;
5898 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5899 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5902 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5904 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5906 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5907 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5908 #if UINT_MAX < SIZE_MAX
5909 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5910 #ifdef MISALIGNED_OK
5911 mx->mx_dbx.md_cmp = mdb_cmp_long;
5913 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5918 /** Initialize a cursor for a given transaction and database. */
5920 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5925 mc->mc_db = &txn->mt_dbs[dbi];
5926 mc->mc_dbx = &txn->mt_dbxs[dbi];
5927 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5932 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5934 mc->mc_xcursor = mx;
5935 mdb_xcursor_init0(mc);
5937 mc->mc_xcursor = NULL;
5939 if (*mc->mc_dbflag & DB_STALE) {
5940 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5945 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5948 size_t size = sizeof(MDB_cursor);
5950 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5953 /* Allow read access to the freelist */
5954 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5957 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5958 size += sizeof(MDB_xcursor);
5960 if ((mc = malloc(size)) != NULL) {
5961 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5962 if (txn->mt_cursors) {
5963 mc->mc_next = txn->mt_cursors[dbi];
5964 txn->mt_cursors[dbi] = mc;
5965 mc->mc_flags |= C_UNTRACK;
5967 mc->mc_flags |= C_ALLOCD;
5978 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5982 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5985 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5988 flags = mc->mc_flags;
5990 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5992 mc->mc_flags |= (flags & C_ALLOCD);
5996 /* Return the count of duplicate data items for the current key */
5998 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6002 if (mc == NULL || countp == NULL)
6005 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
6008 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6009 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6012 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6015 *countp = mc->mc_xcursor->mx_db.md_entries;
6021 mdb_cursor_close(MDB_cursor *mc)
6024 /* remove from txn, if tracked */
6025 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6026 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6027 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6029 *prev = mc->mc_next;
6031 if (mc->mc_flags & C_ALLOCD)
6037 mdb_cursor_txn(MDB_cursor *mc)
6039 if (!mc) return NULL;
6044 mdb_cursor_dbi(MDB_cursor *mc)
6050 /** Replace the key for a node with a new key.
6051 * @param[in] mc Cursor pointing to the node to operate on.
6052 * @param[in] key The new key to use.
6053 * @return 0 on success, non-zero on failure.
6056 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6063 indx_t ptr, i, numkeys, indx;
6066 indx = mc->mc_ki[mc->mc_top];
6067 mp = mc->mc_pg[mc->mc_top];
6068 node = NODEPTR(mp, indx);
6069 ptr = mp->mp_ptrs[indx];
6073 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6074 k2.mv_data = NODEKEY(node);
6075 k2.mv_size = node->mn_ksize;
6076 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6078 mdb_dkey(&k2, kbuf2),
6084 delta0 = delta = key->mv_size - node->mn_ksize;
6086 /* Must be 2-byte aligned. If new key is
6087 * shorter by 1, the shift will be skipped.
6089 delta += (delta & 1);
6091 if (delta > 0 && SIZELEFT(mp) < delta) {
6093 /* not enough space left, do a delete and split */
6094 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6095 pgno = NODEPGNO(node);
6096 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6097 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6100 numkeys = NUMKEYS(mp);
6101 for (i = 0; i < numkeys; i++) {
6102 if (mp->mp_ptrs[i] <= ptr)
6103 mp->mp_ptrs[i] -= delta;
6106 base = (char *)mp + mp->mp_upper;
6107 len = ptr - mp->mp_upper + NODESIZE;
6108 memmove(base - delta, base, len);
6109 mp->mp_upper -= delta;
6111 node = NODEPTR(mp, indx);
6114 /* But even if no shift was needed, update ksize */
6116 node->mn_ksize = key->mv_size;
6119 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6125 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6127 /** Move a node from csrc to cdst.
6130 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6137 unsigned short flags;
6141 /* Mark src and dst as dirty. */
6142 if ((rc = mdb_page_touch(csrc)) ||
6143 (rc = mdb_page_touch(cdst)))
6146 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6147 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6148 key.mv_size = csrc->mc_db->md_pad;
6149 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6151 data.mv_data = NULL;
6155 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6156 assert(!((long)srcnode&1));
6157 srcpg = NODEPGNO(srcnode);
6158 flags = srcnode->mn_flags;
6159 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6160 unsigned int snum = csrc->mc_snum;
6162 /* must find the lowest key below src */
6163 mdb_page_search_lowest(csrc);
6164 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6165 key.mv_size = csrc->mc_db->md_pad;
6166 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6168 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6169 key.mv_size = NODEKSZ(s2);
6170 key.mv_data = NODEKEY(s2);
6172 csrc->mc_snum = snum--;
6173 csrc->mc_top = snum;
6175 key.mv_size = NODEKSZ(srcnode);
6176 key.mv_data = NODEKEY(srcnode);
6178 data.mv_size = NODEDSZ(srcnode);
6179 data.mv_data = NODEDATA(srcnode);
6181 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6182 unsigned int snum = cdst->mc_snum;
6185 /* must find the lowest key below dst */
6186 mdb_page_search_lowest(cdst);
6187 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6188 bkey.mv_size = cdst->mc_db->md_pad;
6189 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6191 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6192 bkey.mv_size = NODEKSZ(s2);
6193 bkey.mv_data = NODEKEY(s2);
6195 cdst->mc_snum = snum--;
6196 cdst->mc_top = snum;
6197 mdb_cursor_copy(cdst, &mn);
6199 rc = mdb_update_key(&mn, &bkey);
6204 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6205 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6206 csrc->mc_ki[csrc->mc_top],
6208 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6209 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6211 /* Add the node to the destination page.
6213 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6214 if (rc != MDB_SUCCESS)
6217 /* Delete the node from the source page.
6219 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6222 /* Adjust other cursors pointing to mp */
6223 MDB_cursor *m2, *m3;
6224 MDB_dbi dbi = csrc->mc_dbi;
6225 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6227 if (csrc->mc_flags & C_SUB)
6230 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6231 if (csrc->mc_flags & C_SUB)
6232 m3 = &m2->mc_xcursor->mx_cursor;
6235 if (m3 == csrc) continue;
6236 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6237 csrc->mc_ki[csrc->mc_top]) {
6238 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6239 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6244 /* Update the parent separators.
6246 if (csrc->mc_ki[csrc->mc_top] == 0) {
6247 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6248 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6249 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6251 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6252 key.mv_size = NODEKSZ(srcnode);
6253 key.mv_data = NODEKEY(srcnode);
6255 DPRINTF("update separator for source page %zu to [%s]",
6256 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6257 mdb_cursor_copy(csrc, &mn);
6260 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6263 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6265 indx_t ix = csrc->mc_ki[csrc->mc_top];
6266 nullkey.mv_size = 0;
6267 csrc->mc_ki[csrc->mc_top] = 0;
6268 rc = mdb_update_key(csrc, &nullkey);
6269 csrc->mc_ki[csrc->mc_top] = ix;
6270 assert(rc == MDB_SUCCESS);
6274 if (cdst->mc_ki[cdst->mc_top] == 0) {
6275 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6276 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6277 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6279 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6280 key.mv_size = NODEKSZ(srcnode);
6281 key.mv_data = NODEKEY(srcnode);
6283 DPRINTF("update separator for destination page %zu to [%s]",
6284 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6285 mdb_cursor_copy(cdst, &mn);
6288 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6291 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6293 indx_t ix = cdst->mc_ki[cdst->mc_top];
6294 nullkey.mv_size = 0;
6295 cdst->mc_ki[cdst->mc_top] = 0;
6296 rc = mdb_update_key(cdst, &nullkey);
6297 cdst->mc_ki[cdst->mc_top] = ix;
6298 assert(rc == MDB_SUCCESS);
6305 /** Merge one page into another.
6306 * The nodes from the page pointed to by \b csrc will
6307 * be copied to the page pointed to by \b cdst and then
6308 * the \b csrc page will be freed.
6309 * @param[in] csrc Cursor pointing to the source page.
6310 * @param[in] cdst Cursor pointing to the destination page.
6313 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6321 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6322 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6324 assert(csrc->mc_snum > 1); /* can't merge root page */
6325 assert(cdst->mc_snum > 1);
6327 /* Mark dst as dirty. */
6328 if ((rc = mdb_page_touch(cdst)))
6331 /* Move all nodes from src to dst.
6333 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6334 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6335 key.mv_size = csrc->mc_db->md_pad;
6336 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6337 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6338 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6339 if (rc != MDB_SUCCESS)
6341 key.mv_data = (char *)key.mv_data + key.mv_size;
6344 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6345 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6346 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6347 unsigned int snum = csrc->mc_snum;
6349 /* must find the lowest key below src */
6350 mdb_page_search_lowest(csrc);
6351 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6352 key.mv_size = csrc->mc_db->md_pad;
6353 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6355 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6356 key.mv_size = NODEKSZ(s2);
6357 key.mv_data = NODEKEY(s2);
6359 csrc->mc_snum = snum--;
6360 csrc->mc_top = snum;
6362 key.mv_size = srcnode->mn_ksize;
6363 key.mv_data = NODEKEY(srcnode);
6366 data.mv_size = NODEDSZ(srcnode);
6367 data.mv_data = NODEDATA(srcnode);
6368 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6369 if (rc != MDB_SUCCESS)
6374 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6375 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);
6377 /* Unlink the src page from parent and add to free list.
6379 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6380 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6383 rc = mdb_update_key(csrc, &key);
6389 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6390 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6391 csrc->mc_db->md_leaf_pages--;
6393 csrc->mc_db->md_branch_pages--;
6395 /* Adjust other cursors pointing to mp */
6396 MDB_cursor *m2, *m3;
6397 MDB_dbi dbi = csrc->mc_dbi;
6398 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6400 if (csrc->mc_flags & C_SUB)
6403 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6404 if (csrc->mc_flags & C_SUB)
6405 m3 = &m2->mc_xcursor->mx_cursor;
6408 if (m3 == csrc) continue;
6409 if (m3->mc_snum < csrc->mc_snum) continue;
6410 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6411 m3->mc_pg[csrc->mc_top] = mp;
6412 m3->mc_ki[csrc->mc_top] += nkeys;
6416 mdb_cursor_pop(csrc);
6418 return mdb_rebalance(csrc);
6421 /** Copy the contents of a cursor.
6422 * @param[in] csrc The cursor to copy from.
6423 * @param[out] cdst The cursor to copy to.
6426 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6430 cdst->mc_txn = csrc->mc_txn;
6431 cdst->mc_dbi = csrc->mc_dbi;
6432 cdst->mc_db = csrc->mc_db;
6433 cdst->mc_dbx = csrc->mc_dbx;
6434 cdst->mc_snum = csrc->mc_snum;
6435 cdst->mc_top = csrc->mc_top;
6436 cdst->mc_flags = csrc->mc_flags;
6438 for (i=0; i<csrc->mc_snum; i++) {
6439 cdst->mc_pg[i] = csrc->mc_pg[i];
6440 cdst->mc_ki[i] = csrc->mc_ki[i];
6444 /** Rebalance the tree after a delete operation.
6445 * @param[in] mc Cursor pointing to the page where rebalancing
6447 * @return 0 on success, non-zero on failure.
6450 mdb_rebalance(MDB_cursor *mc)
6454 unsigned int ptop, minkeys;
6457 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6461 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6462 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6463 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6464 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6468 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6469 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6472 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6473 DPRINTF("no need to rebalance page %zu, above fill threshold",
6479 if (mc->mc_snum < 2) {
6480 MDB_page *mp = mc->mc_pg[0];
6482 DPUTS("Can't rebalance a subpage, ignoring");
6485 if (NUMKEYS(mp) == 0) {
6486 DPUTS("tree is completely empty");
6487 mc->mc_db->md_root = P_INVALID;
6488 mc->mc_db->md_depth = 0;
6489 mc->mc_db->md_leaf_pages = 0;
6490 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6491 /* Adjust cursors pointing to mp */
6495 MDB_cursor *m2, *m3;
6496 MDB_dbi dbi = mc->mc_dbi;
6498 if (mc->mc_flags & C_SUB)
6501 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6502 if (mc->mc_flags & C_SUB)
6503 m3 = &m2->mc_xcursor->mx_cursor;
6506 if (m3->mc_snum < mc->mc_snum) continue;
6507 if (m3->mc_pg[0] == mp) {
6513 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6514 DPUTS("collapsing root page!");
6515 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6516 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6517 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6520 mc->mc_db->md_depth--;
6521 mc->mc_db->md_branch_pages--;
6523 /* Adjust other cursors pointing to mp */
6524 MDB_cursor *m2, *m3;
6525 MDB_dbi dbi = mc->mc_dbi;
6527 if (mc->mc_flags & C_SUB)
6530 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6531 if (mc->mc_flags & C_SUB)
6532 m3 = &m2->mc_xcursor->mx_cursor;
6535 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6536 if (m3->mc_pg[0] == mp) {
6537 m3->mc_pg[0] = mc->mc_pg[0];
6544 DPUTS("root page doesn't need rebalancing");
6548 /* The parent (branch page) must have at least 2 pointers,
6549 * otherwise the tree is invalid.
6551 ptop = mc->mc_top-1;
6552 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6554 /* Leaf page fill factor is below the threshold.
6555 * Try to move keys from left or right neighbor, or
6556 * merge with a neighbor page.
6561 mdb_cursor_copy(mc, &mn);
6562 mn.mc_xcursor = NULL;
6564 if (mc->mc_ki[ptop] == 0) {
6565 /* We're the leftmost leaf in our parent.
6567 DPUTS("reading right neighbor");
6569 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6570 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6573 mn.mc_ki[mn.mc_top] = 0;
6574 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6576 /* There is at least one neighbor to the left.
6578 DPUTS("reading left neighbor");
6580 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6581 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6584 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6585 mc->mc_ki[mc->mc_top] = 0;
6588 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6589 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);
6591 /* If the neighbor page is above threshold and has enough keys,
6592 * move one key from it. Otherwise we should try to merge them.
6593 * (A branch page must never have less than 2 keys.)
6595 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6596 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6597 return mdb_node_move(&mn, mc);
6599 if (mc->mc_ki[ptop] == 0)
6600 rc = mdb_page_merge(&mn, mc);
6602 rc = mdb_page_merge(mc, &mn);
6603 mc->mc_flags &= ~C_INITIALIZED;
6608 /** Complete a delete operation started by #mdb_cursor_del(). */
6610 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6614 /* add overflow pages to free list */
6615 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6619 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6620 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
6622 assert(IS_OVERFLOW(omp));
6623 mdb_ovpage_free(mc, omp);
6625 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6626 mc->mc_db->md_entries--;
6627 rc = mdb_rebalance(mc);
6628 if (rc != MDB_SUCCESS)
6629 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6630 /* if mc points past last node in page, invalidate */
6631 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6632 mc->mc_flags &= ~C_INITIALIZED;
6638 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6639 MDB_val *key, MDB_val *data)
6644 MDB_val rdata, *xdata;
6648 assert(key != NULL);
6650 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6652 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6655 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6659 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6663 mdb_cursor_init(&mc, txn, dbi, &mx);
6674 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6676 /* let mdb_page_split know about this cursor if needed:
6677 * delete will trigger a rebalance; if it needs to move
6678 * a node from one page to another, it will have to
6679 * update the parent's separator key(s). If the new sepkey
6680 * is larger than the current one, the parent page may
6681 * run out of space, triggering a split. We need this
6682 * cursor to be consistent until the end of the rebalance.
6684 mc.mc_next = txn->mt_cursors[dbi];
6685 txn->mt_cursors[dbi] = &mc;
6686 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6687 txn->mt_cursors[dbi] = mc.mc_next;
6692 /** Split a page and insert a new node.
6693 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6694 * The cursor will be updated to point to the actual page and index where
6695 * the node got inserted after the split.
6696 * @param[in] newkey The key for the newly inserted node.
6697 * @param[in] newdata The data for the newly inserted node.
6698 * @param[in] newpgno The page number, if the new node is a branch node.
6699 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6700 * @return 0 on success, non-zero on failure.
6703 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6704 unsigned int nflags)
6707 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6710 unsigned int i, j, split_indx, nkeys, pmax;
6712 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6714 MDB_page *mp, *rp, *pp;
6719 mp = mc->mc_pg[mc->mc_top];
6720 newindx = mc->mc_ki[mc->mc_top];
6722 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6723 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6724 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6726 /* Create a right sibling. */
6727 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6729 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6731 if (mc->mc_snum < 2) {
6732 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6734 /* shift current top to make room for new parent */
6735 mc->mc_pg[1] = mc->mc_pg[0];
6736 mc->mc_ki[1] = mc->mc_ki[0];
6739 mc->mc_db->md_root = pp->mp_pgno;
6740 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6741 mc->mc_db->md_depth++;
6744 /* Add left (implicit) pointer. */
6745 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6746 /* undo the pre-push */
6747 mc->mc_pg[0] = mc->mc_pg[1];
6748 mc->mc_ki[0] = mc->mc_ki[1];
6749 mc->mc_db->md_root = mp->mp_pgno;
6750 mc->mc_db->md_depth--;
6757 ptop = mc->mc_top-1;
6758 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6761 mc->mc_flags |= C_SPLITTING;
6762 mdb_cursor_copy(mc, &mn);
6763 mn.mc_pg[mn.mc_top] = rp;
6764 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6766 if (nflags & MDB_APPEND) {
6767 mn.mc_ki[mn.mc_top] = 0;
6769 split_indx = newindx;
6774 nkeys = NUMKEYS(mp);
6775 split_indx = nkeys / 2;
6776 if (newindx < split_indx)
6782 unsigned int lsize, rsize, ksize;
6783 /* Move half of the keys to the right sibling */
6785 x = mc->mc_ki[mc->mc_top] - split_indx;
6786 ksize = mc->mc_db->md_pad;
6787 split = LEAF2KEY(mp, split_indx, ksize);
6788 rsize = (nkeys - split_indx) * ksize;
6789 lsize = (nkeys - split_indx) * sizeof(indx_t);
6790 mp->mp_lower -= lsize;
6791 rp->mp_lower += lsize;
6792 mp->mp_upper += rsize - lsize;
6793 rp->mp_upper -= rsize - lsize;
6794 sepkey.mv_size = ksize;
6795 if (newindx == split_indx) {
6796 sepkey.mv_data = newkey->mv_data;
6798 sepkey.mv_data = split;
6801 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6802 memcpy(rp->mp_ptrs, split, rsize);
6803 sepkey.mv_data = rp->mp_ptrs;
6804 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6805 memcpy(ins, newkey->mv_data, ksize);
6806 mp->mp_lower += sizeof(indx_t);
6807 mp->mp_upper -= ksize - sizeof(indx_t);
6810 memcpy(rp->mp_ptrs, split, x * ksize);
6811 ins = LEAF2KEY(rp, x, ksize);
6812 memcpy(ins, newkey->mv_data, ksize);
6813 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6814 rp->mp_lower += sizeof(indx_t);
6815 rp->mp_upper -= ksize - sizeof(indx_t);
6816 mc->mc_ki[mc->mc_top] = x;
6817 mc->mc_pg[mc->mc_top] = rp;
6822 /* For leaf pages, check the split point based on what
6823 * fits where, since otherwise mdb_node_add can fail.
6825 * This check is only needed when the data items are
6826 * relatively large, such that being off by one will
6827 * make the difference between success or failure.
6829 * It's also relevant if a page happens to be laid out
6830 * such that one half of its nodes are all "small" and
6831 * the other half of its nodes are "large." If the new
6832 * item is also "large" and falls on the half with
6833 * "large" nodes, it also may not fit.
6836 unsigned int psize, nsize;
6837 /* Maximum free space in an empty page */
6838 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6839 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6840 if ((nkeys < 20) || (nsize > pmax/16)) {
6841 if (newindx <= split_indx) {
6844 for (i=0; i<split_indx; i++) {
6845 node = NODEPTR(mp, i);
6846 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6847 if (F_ISSET(node->mn_flags, F_BIGDATA))
6848 psize += sizeof(pgno_t);
6850 psize += NODEDSZ(node);
6854 split_indx = newindx;
6865 for (i=nkeys-1; i>=split_indx; i--) {
6866 node = NODEPTR(mp, i);
6867 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6868 if (F_ISSET(node->mn_flags, F_BIGDATA))
6869 psize += sizeof(pgno_t);
6871 psize += NODEDSZ(node);
6875 split_indx = newindx;
6886 /* First find the separating key between the split pages.
6887 * The case where newindx == split_indx is ambiguous; the
6888 * new item could go to the new page or stay on the original
6889 * page. If newpos == 1 it goes to the new page.
6891 if (newindx == split_indx && newpos) {
6892 sepkey.mv_size = newkey->mv_size;
6893 sepkey.mv_data = newkey->mv_data;
6895 node = NODEPTR(mp, split_indx);
6896 sepkey.mv_size = node->mn_ksize;
6897 sepkey.mv_data = NODEKEY(node);
6901 DPRINTF("separator is [%s]", DKEY(&sepkey));
6903 /* Copy separator key to the parent.
6905 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6909 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6912 if (mn.mc_snum == mc->mc_snum) {
6913 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6914 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6915 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6916 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6921 /* Right page might now have changed parent.
6922 * Check if left page also changed parent.
6924 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6925 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6926 for (i=0; i<ptop; i++) {
6927 mc->mc_pg[i] = mn.mc_pg[i];
6928 mc->mc_ki[i] = mn.mc_ki[i];
6930 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6931 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6935 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6938 mc->mc_flags ^= C_SPLITTING;
6939 if (rc != MDB_SUCCESS) {
6942 if (nflags & MDB_APPEND) {
6943 mc->mc_pg[mc->mc_top] = rp;
6944 mc->mc_ki[mc->mc_top] = 0;
6945 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6948 for (i=0; i<mc->mc_top; i++)
6949 mc->mc_ki[i] = mn.mc_ki[i];
6956 /* Move half of the keys to the right sibling. */
6958 /* grab a page to hold a temporary copy */
6959 copy = mdb_page_malloc(mc, 1);
6963 copy->mp_pgno = mp->mp_pgno;
6964 copy->mp_flags = mp->mp_flags;
6965 copy->mp_lower = PAGEHDRSZ;
6966 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6967 mc->mc_pg[mc->mc_top] = copy;
6968 for (i = j = 0; i <= nkeys; j++) {
6969 if (i == split_indx) {
6970 /* Insert in right sibling. */
6971 /* Reset insert index for right sibling. */
6972 if (i != newindx || (newpos ^ ins_new)) {
6974 mc->mc_pg[mc->mc_top] = rp;
6978 if (i == newindx && !ins_new) {
6979 /* Insert the original entry that caused the split. */
6980 rkey.mv_data = newkey->mv_data;
6981 rkey.mv_size = newkey->mv_size;
6990 /* Update index for the new key. */
6991 mc->mc_ki[mc->mc_top] = j;
6992 } else if (i == nkeys) {
6995 node = NODEPTR(mp, i);
6996 rkey.mv_data = NODEKEY(node);
6997 rkey.mv_size = node->mn_ksize;
6999 xdata.mv_data = NODEDATA(node);
7000 xdata.mv_size = NODEDSZ(node);
7003 pgno = NODEPGNO(node);
7004 flags = node->mn_flags;
7009 if (!IS_LEAF(mp) && j == 0) {
7010 /* First branch index doesn't need key data. */
7014 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7018 nkeys = NUMKEYS(copy);
7019 for (i=0; i<nkeys; i++)
7020 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7021 mp->mp_lower = copy->mp_lower;
7022 mp->mp_upper = copy->mp_upper;
7023 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7024 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7026 /* reset back to original page */
7027 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7028 mc->mc_pg[mc->mc_top] = mp;
7029 if (nflags & MDB_RESERVE) {
7030 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7031 if (!(node->mn_flags & F_BIGDATA))
7032 newdata->mv_data = NODEDATA(node);
7036 /* Make sure mc_ki is still valid.
7038 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7039 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7040 for (i=0; i<ptop; i++) {
7041 mc->mc_pg[i] = mn.mc_pg[i];
7042 mc->mc_ki[i] = mn.mc_ki[i];
7044 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7045 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7049 /* return tmp page to freelist */
7050 mdb_page_free(mc->mc_txn->mt_env, copy);
7053 /* Adjust other cursors pointing to mp */
7054 MDB_cursor *m2, *m3;
7055 MDB_dbi dbi = mc->mc_dbi;
7056 int fixup = NUMKEYS(mp);
7058 if (mc->mc_flags & C_SUB)
7061 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7062 if (mc->mc_flags & C_SUB)
7063 m3 = &m2->mc_xcursor->mx_cursor;
7068 if (!(m3->mc_flags & C_INITIALIZED))
7070 if (m3->mc_flags & C_SPLITTING)
7075 for (k=m3->mc_top; k>=0; k--) {
7076 m3->mc_ki[k+1] = m3->mc_ki[k];
7077 m3->mc_pg[k+1] = m3->mc_pg[k];
7079 if (m3->mc_ki[0] >= split_indx) {
7084 m3->mc_pg[0] = mc->mc_pg[0];
7088 if (m3->mc_pg[mc->mc_top] == mp) {
7089 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7090 m3->mc_ki[mc->mc_top]++;
7091 if (m3->mc_ki[mc->mc_top] >= fixup) {
7092 m3->mc_pg[mc->mc_top] = rp;
7093 m3->mc_ki[mc->mc_top] -= fixup;
7094 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7096 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7097 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7106 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7107 MDB_val *key, MDB_val *data, unsigned int flags)
7112 assert(key != NULL);
7113 assert(data != NULL);
7115 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7118 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7122 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7126 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7129 mdb_cursor_init(&mc, txn, dbi, &mx);
7130 return mdb_cursor_put(&mc, key, data, flags);
7134 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7136 if ((flag & CHANGEABLE) != flag)
7139 env->me_flags |= flag;
7141 env->me_flags &= ~flag;
7146 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7151 *arg = env->me_flags;
7156 mdb_env_get_path(MDB_env *env, const char **arg)
7161 *arg = env->me_path;
7165 /** Common code for #mdb_stat() and #mdb_env_stat().
7166 * @param[in] env the environment to operate in.
7167 * @param[in] db the #MDB_db record containing the stats to return.
7168 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7169 * @return 0, this function always succeeds.
7172 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7174 arg->ms_psize = env->me_psize;
7175 arg->ms_depth = db->md_depth;
7176 arg->ms_branch_pages = db->md_branch_pages;
7177 arg->ms_leaf_pages = db->md_leaf_pages;
7178 arg->ms_overflow_pages = db->md_overflow_pages;
7179 arg->ms_entries = db->md_entries;
7184 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7188 if (env == NULL || arg == NULL)
7191 toggle = mdb_env_pick_meta(env);
7193 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7197 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7201 if (env == NULL || arg == NULL)
7204 toggle = mdb_env_pick_meta(env);
7205 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7206 arg->me_mapsize = env->me_mapsize;
7207 arg->me_maxreaders = env->me_maxreaders;
7208 arg->me_numreaders = env->me_numreaders;
7209 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7210 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7214 /** Set the default comparison functions for a database.
7215 * Called immediately after a database is opened to set the defaults.
7216 * The user can then override them with #mdb_set_compare() or
7217 * #mdb_set_dupsort().
7218 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7219 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7222 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7224 uint16_t f = txn->mt_dbs[dbi].md_flags;
7226 txn->mt_dbxs[dbi].md_cmp =
7227 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7228 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7230 txn->mt_dbxs[dbi].md_dcmp =
7231 !(f & MDB_DUPSORT) ? 0 :
7232 ((f & MDB_INTEGERDUP)
7233 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7234 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7237 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7242 int rc, dbflag, exact;
7243 unsigned int unused = 0;
7246 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7247 mdb_default_cmp(txn, FREE_DBI);
7250 if ((flags & VALID_FLAGS) != flags)
7256 if (flags & PERSISTENT_FLAGS) {
7257 uint16_t f2 = flags & PERSISTENT_FLAGS;
7258 /* make sure flag changes get committed */
7259 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7260 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7261 txn->mt_flags |= MDB_TXN_DIRTY;
7264 mdb_default_cmp(txn, MAIN_DBI);
7268 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7269 mdb_default_cmp(txn, MAIN_DBI);
7272 /* Is the DB already open? */
7274 for (i=2; i<txn->mt_numdbs; i++) {
7275 if (!txn->mt_dbxs[i].md_name.mv_size) {
7276 /* Remember this free slot */
7277 if (!unused) unused = i;
7280 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7281 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7287 /* If no free slot and max hit, fail */
7288 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7289 return MDB_DBS_FULL;
7291 /* Cannot mix named databases with some mainDB flags */
7292 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7293 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7295 /* Find the DB info */
7296 dbflag = DB_NEW|DB_VALID;
7299 key.mv_data = (void *)name;
7300 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7301 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7302 if (rc == MDB_SUCCESS) {
7303 /* make sure this is actually a DB */
7304 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7305 if (!(node->mn_flags & F_SUBDATA))
7307 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7308 /* Create if requested */
7310 data.mv_size = sizeof(MDB_db);
7311 data.mv_data = &dummy;
7312 memset(&dummy, 0, sizeof(dummy));
7313 dummy.md_root = P_INVALID;
7314 dummy.md_flags = flags & PERSISTENT_FLAGS;
7315 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7319 /* OK, got info, add to table */
7320 if (rc == MDB_SUCCESS) {
7321 unsigned int slot = unused ? unused : txn->mt_numdbs;
7322 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7323 txn->mt_dbxs[slot].md_name.mv_size = len;
7324 txn->mt_dbxs[slot].md_rel = NULL;
7325 txn->mt_dbflags[slot] = dbflag;
7326 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7328 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7329 mdb_default_cmp(txn, slot);
7338 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7340 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7343 if (txn->mt_dbflags[dbi] & DB_STALE) {
7346 /* Stale, must read the DB's root. cursor_init does it for us. */
7347 mdb_cursor_init(&mc, txn, dbi, &mx);
7349 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7352 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7355 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7357 ptr = env->me_dbxs[dbi].md_name.mv_data;
7358 env->me_dbxs[dbi].md_name.mv_data = NULL;
7359 env->me_dbxs[dbi].md_name.mv_size = 0;
7360 env->me_dbflags[dbi] = 0;
7364 /** Add all the DB's pages to the free list.
7365 * @param[in] mc Cursor on the DB to free.
7366 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7367 * @return 0 on success, non-zero on failure.
7370 mdb_drop0(MDB_cursor *mc, int subs)
7374 rc = mdb_page_search(mc, NULL, 0);
7375 if (rc == MDB_SUCCESS) {
7380 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7381 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7384 mdb_cursor_copy(mc, &mx);
7385 while (mc->mc_snum > 0) {
7386 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7387 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7388 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7389 if (ni->mn_flags & F_BIGDATA) {
7393 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7394 rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL);
7397 assert(IS_OVERFLOW(omp));
7398 ovpages = omp->mp_pages;
7399 for (j=0; j<ovpages; j++) {
7400 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7403 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7404 mdb_xcursor_init1(mc, ni);
7405 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7411 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7413 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7416 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7421 mc->mc_ki[mc->mc_top] = i;
7422 rc = mdb_cursor_sibling(mc, 1);
7424 /* no more siblings, go back to beginning
7425 * of previous level.
7429 for (i=1; i<mc->mc_snum; i++) {
7431 mc->mc_pg[i] = mx.mc_pg[i];
7436 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7437 mc->mc_db->md_root);
7442 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7444 MDB_cursor *mc, *m2;
7447 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7450 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7453 rc = mdb_cursor_open(txn, dbi, &mc);
7457 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7458 /* Invalidate the dropped DB's cursors */
7459 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7460 m2->mc_flags &= ~C_INITIALIZED;
7464 /* Can't delete the main DB */
7465 if (del && dbi > MAIN_DBI) {
7466 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7468 txn->mt_dbflags[dbi] = DB_STALE;
7469 mdb_dbi_close(txn->mt_env, dbi);
7472 /* reset the DB record, mark it dirty */
7473 txn->mt_dbflags[dbi] |= DB_DIRTY;
7474 txn->mt_dbs[dbi].md_depth = 0;
7475 txn->mt_dbs[dbi].md_branch_pages = 0;
7476 txn->mt_dbs[dbi].md_leaf_pages = 0;
7477 txn->mt_dbs[dbi].md_overflow_pages = 0;
7478 txn->mt_dbs[dbi].md_entries = 0;
7479 txn->mt_dbs[dbi].md_root = P_INVALID;
7481 txn->mt_flags |= MDB_TXN_DIRTY;
7484 mdb_cursor_close(mc);
7488 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7490 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7493 txn->mt_dbxs[dbi].md_cmp = cmp;
7497 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7499 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7502 txn->mt_dbxs[dbi].md_dcmp = cmp;
7506 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7508 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7511 txn->mt_dbxs[dbi].md_rel = rel;
7515 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7517 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7520 txn->mt_dbxs[dbi].md_relctx = ctx;