2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define pthread_t DWORD
144 #define pthread_mutex_t HANDLE
145 #define pthread_key_t DWORD
146 #define pthread_self() GetCurrentThreadId()
147 #define pthread_key_create(x,y) \
148 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
149 #define pthread_key_delete(x) TlsFree(x)
150 #define pthread_getspecific(x) TlsGetValue(x)
151 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
152 #define pthread_mutex_unlock(x) ReleaseMutex(x)
153 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
154 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
158 #define getpid() GetCurrentProcessId()
159 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
160 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
161 #define ErrCode() GetLastError()
162 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
163 #define close(fd) CloseHandle(fd)
164 #define munmap(ptr,len) UnmapViewOfFile(ptr)
167 #ifdef MDB_USE_POSIX_SEM
169 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
170 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
171 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
172 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
175 mdb_sem_wait(sem_t *sem)
178 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
183 /** Lock the reader mutex.
185 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
186 /** Unlock the reader mutex.
188 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
190 /** Lock the writer mutex.
191 * Only a single write transaction is allowed at a time. Other writers
192 * will block waiting for this mutex.
194 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
195 /** Unlock the writer mutex.
197 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
198 #endif /* MDB_USE_POSIX_SEM */
200 /** Get the error code for the last failed system function.
202 #define ErrCode() errno
204 /** An abstraction for a file handle.
205 * On POSIX systems file handles are small integers. On Windows
206 * they're opaque pointers.
210 /** A value for an invalid file handle.
211 * Mainly used to initialize file variables and signify that they are
214 #define INVALID_HANDLE_VALUE (-1)
216 /** Get the size of a memory page for the system.
217 * This is the basic size that the platform's memory manager uses, and is
218 * fundamental to the use of memory-mapped files.
220 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
223 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
226 #define MNAME_LEN (sizeof(pthread_mutex_t))
232 /** A flag for opening a file and requesting synchronous data writes.
233 * This is only used when writing a meta page. It's not strictly needed;
234 * we could just do a normal write and then immediately perform a flush.
235 * But if this flag is available it saves us an extra system call.
237 * @note If O_DSYNC is undefined but exists in /usr/include,
238 * preferably set some compiler flag to get the definition.
239 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
242 # define MDB_DSYNC O_DSYNC
246 /** Function for flushing the data of a file. Define this to fsync
247 * if fdatasync() is not supported.
249 #ifndef MDB_FDATASYNC
250 # define MDB_FDATASYNC fdatasync
254 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
265 /** A page number in the database.
266 * Note that 64 bit page numbers are overkill, since pages themselves
267 * already represent 12-13 bits of addressable memory, and the OS will
268 * always limit applications to a maximum of 63 bits of address space.
270 * @note In the #MDB_node structure, we only store 48 bits of this value,
271 * which thus limits us to only 60 bits of addressable data.
273 typedef MDB_ID pgno_t;
275 /** A transaction ID.
276 * See struct MDB_txn.mt_txnid for details.
278 typedef MDB_ID txnid_t;
280 /** @defgroup debug Debug Macros
284 /** Enable debug output.
285 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
286 * read from and written to the database (used for free space management).
291 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
294 # define DPRINTF (void) /* Vararg macros may be unsupported */
296 static int mdb_debug;
297 static txnid_t mdb_debug_start;
299 /** Print a debug message with printf formatting. */
300 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
301 ((void) ((mdb_debug) && \
302 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
304 # define DPRINTF(fmt, ...) ((void) 0)
305 # define MDB_DEBUG_SKIP
307 /** Print a debug string.
308 * The string is printed literally, with no format processing.
310 #define DPUTS(arg) DPRINTF("%s", arg)
313 /** A default memory page size.
314 * The actual size is platform-dependent, but we use this for
315 * boot-strapping. We probably should not be using this any more.
316 * The #GET_PAGESIZE() macro is used to get the actual size.
318 * Note that we don't currently support Huge pages. On Linux,
319 * regular data files cannot use Huge pages, and in general
320 * Huge pages aren't actually pageable. We rely on the OS
321 * demand-pager to read our data and page it out when memory
322 * pressure from other processes is high. So until OSs have
323 * actual paging support for Huge pages, they're not viable.
325 #define MDB_PAGESIZE 4096
327 /** The minimum number of keys required in a database page.
328 * Setting this to a larger value will place a smaller bound on the
329 * maximum size of a data item. Data items larger than this size will
330 * be pushed into overflow pages instead of being stored directly in
331 * the B-tree node. This value used to default to 4. With a page size
332 * of 4096 bytes that meant that any item larger than 1024 bytes would
333 * go into an overflow page. That also meant that on average 2-3KB of
334 * each overflow page was wasted space. The value cannot be lower than
335 * 2 because then there would no longer be a tree structure. With this
336 * value, items larger than 2KB will go into overflow pages, and on
337 * average only 1KB will be wasted.
339 #define MDB_MINKEYS 2
341 /** A stamp that identifies a file as an MDB file.
342 * There's nothing special about this value other than that it is easily
343 * recognizable, and it will reflect any byte order mismatches.
345 #define MDB_MAGIC 0xBEEFC0DE
347 /** The version number for a database's file format. */
348 #define MDB_VERSION 1
350 /** @brief The maximum size of a key in the database.
352 * We require that keys all fit onto a regular page. This limit
353 * could be raised a bit further if needed; to something just
354 * under #MDB_PAGESIZE / #MDB_MINKEYS.
356 * Note that data items in an #MDB_DUPSORT database are actually keys
357 * of a subDB, so they're also limited to this size.
359 #ifndef MDB_MAXKEYSIZE
360 #define MDB_MAXKEYSIZE 511
363 /** @brief The maximum size of a data item.
365 * We only store a 32 bit value for node sizes.
367 #define MAXDATASIZE 0xffffffffUL
372 * This is used for printing a hex dump of a key's contents.
374 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
375 /** Display a key in hex.
377 * Invoke a function to display a key in hex.
379 #define DKEY(x) mdb_dkey(x, kbuf)
381 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
385 /** An invalid page number.
386 * Mainly used to denote an empty tree.
388 #define P_INVALID (~(pgno_t)0)
390 /** Test if the flags \b f are set in a flag word \b w. */
391 #define F_ISSET(w, f) (((w) & (f)) == (f))
393 /** Used for offsets within a single page.
394 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
397 typedef uint16_t indx_t;
399 /** Default size of memory map.
400 * This is certainly too small for any actual applications. Apps should always set
401 * the size explicitly using #mdb_env_set_mapsize().
403 #define DEFAULT_MAPSIZE 1048576
405 /** @defgroup readers Reader Lock Table
406 * Readers don't acquire any locks for their data access. Instead, they
407 * simply record their transaction ID in the reader table. The reader
408 * mutex is needed just to find an empty slot in the reader table. The
409 * slot's address is saved in thread-specific data so that subsequent read
410 * transactions started by the same thread need no further locking to proceed.
412 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
414 * No reader table is used if the database is on a read-only filesystem.
416 * Since the database uses multi-version concurrency control, readers don't
417 * actually need any locking. This table is used to keep track of which
418 * readers are using data from which old transactions, so that we'll know
419 * when a particular old transaction is no longer in use. Old transactions
420 * that have discarded any data pages can then have those pages reclaimed
421 * for use by a later write transaction.
423 * The lock table is constructed such that reader slots are aligned with the
424 * processor's cache line size. Any slot is only ever used by one thread.
425 * This alignment guarantees that there will be no contention or cache
426 * thrashing as threads update their own slot info, and also eliminates
427 * any need for locking when accessing a slot.
429 * A writer thread will scan every slot in the table to determine the oldest
430 * outstanding reader transaction. Any freed pages older than this will be
431 * reclaimed by the writer. The writer doesn't use any locks when scanning
432 * this table. This means that there's no guarantee that the writer will
433 * see the most up-to-date reader info, but that's not required for correct
434 * operation - all we need is to know the upper bound on the oldest reader,
435 * we don't care at all about the newest reader. So the only consequence of
436 * reading stale information here is that old pages might hang around a
437 * while longer before being reclaimed. That's actually good anyway, because
438 * the longer we delay reclaiming old pages, the more likely it is that a
439 * string of contiguous pages can be found after coalescing old pages from
440 * many old transactions together.
443 /** Number of slots in the reader table.
444 * This value was chosen somewhat arbitrarily. 126 readers plus a
445 * couple mutexes fit exactly into 8KB on my development machine.
446 * Applications should set the table size using #mdb_env_set_maxreaders().
448 #define DEFAULT_READERS 126
450 /** The size of a CPU cache line in bytes. We want our lock structures
451 * aligned to this size to avoid false cache line sharing in the
453 * This value works for most CPUs. For Itanium this should be 128.
459 /** The information we store in a single slot of the reader table.
460 * In addition to a transaction ID, we also record the process and
461 * thread ID that owns a slot, so that we can detect stale information,
462 * e.g. threads or processes that went away without cleaning up.
463 * @note We currently don't check for stale records. We simply re-init
464 * the table when we know that we're the only process opening the
467 typedef struct MDB_rxbody {
468 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
469 * Multiple readers that start at the same time will probably have the
470 * same ID here. Again, it's not important to exclude them from
471 * anything; all we need to know is which version of the DB they
472 * started from so we can avoid overwriting any data used in that
473 * particular version.
476 /** The process ID of the process owning this reader txn. */
478 /** The thread ID of the thread owning this txn. */
482 /** The actual reader record, with cacheline padding. */
483 typedef struct MDB_reader {
486 /** shorthand for mrb_txnid */
487 #define mr_txnid mru.mrx.mrb_txnid
488 #define mr_pid mru.mrx.mrb_pid
489 #define mr_tid mru.mrx.mrb_tid
490 /** cache line alignment */
491 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
495 /** The header for the reader table.
496 * The table resides in a memory-mapped file. (This is a different file
497 * than is used for the main database.)
499 * For POSIX the actual mutexes reside in the shared memory of this
500 * mapped file. On Windows, mutexes are named objects allocated by the
501 * kernel; we store the mutex names in this mapped file so that other
502 * processes can grab them. This same approach is also used on
503 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
504 * process-shared POSIX mutexes. For these cases where a named object
505 * is used, the object name is derived from a 64 bit FNV hash of the
506 * environment pathname. As such, naming collisions are extremely
507 * unlikely. If a collision occurs, the results are unpredictable.
509 typedef struct MDB_txbody {
510 /** Stamp identifying this as an MDB file. It must be set
513 /** Version number of this lock file. Must be set to #MDB_VERSION. */
514 uint32_t mtb_version;
515 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
516 char mtb_rmname[MNAME_LEN];
518 /** Mutex protecting access to this table.
519 * This is the reader lock that #LOCK_MUTEX_R acquires.
521 pthread_mutex_t mtb_mutex;
523 /** The ID of the last transaction committed to the database.
524 * This is recorded here only for convenience; the value can always
525 * be determined by reading the main database meta pages.
528 /** The number of slots that have been used in the reader table.
529 * This always records the maximum count, it is not decremented
530 * when readers release their slots.
532 unsigned mtb_numreaders;
535 /** The actual reader table definition. */
536 typedef struct MDB_txninfo {
539 #define mti_magic mt1.mtb.mtb_magic
540 #define mti_version mt1.mtb.mtb_version
541 #define mti_mutex mt1.mtb.mtb_mutex
542 #define mti_rmname mt1.mtb.mtb_rmname
543 #define mti_txnid mt1.mtb.mtb_txnid
544 #define mti_numreaders mt1.mtb.mtb_numreaders
545 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
548 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
549 char mt2_wmname[MNAME_LEN];
550 #define mti_wmname mt2.mt2_wmname
552 pthread_mutex_t mt2_wmutex;
553 #define mti_wmutex mt2.mt2_wmutex
555 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
557 MDB_reader mti_readers[1];
561 /** Common header for all page types.
562 * Overflow records occupy a number of contiguous pages with no
563 * headers on any page after the first.
565 typedef struct MDB_page {
566 #define mp_pgno mp_p.p_pgno
567 #define mp_next mp_p.p_next
569 pgno_t p_pgno; /**< page number */
570 void * p_next; /**< for in-memory list of freed structs */
573 /** @defgroup mdb_page Page Flags
575 * Flags for the page headers.
578 #define P_BRANCH 0x01 /**< branch page */
579 #define P_LEAF 0x02 /**< leaf page */
580 #define P_OVERFLOW 0x04 /**< overflow page */
581 #define P_META 0x08 /**< meta page */
582 #define P_DIRTY 0x10 /**< dirty page */
583 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
584 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
586 uint16_t mp_flags; /**< @ref mdb_page */
587 #define mp_lower mp_pb.pb.pb_lower
588 #define mp_upper mp_pb.pb.pb_upper
589 #define mp_pages mp_pb.pb_pages
592 indx_t pb_lower; /**< lower bound of free space */
593 indx_t pb_upper; /**< upper bound of free space */
595 uint32_t pb_pages; /**< number of overflow pages */
597 indx_t mp_ptrs[1]; /**< dynamic size */
600 /** Size of the page header, excluding dynamic data at the end */
601 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
603 /** Address of first usable data byte in a page, after the header */
604 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
606 /** Number of nodes on a page */
607 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
609 /** The amount of space remaining in the page */
610 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
612 /** The percentage of space used in the page, in tenths of a percent. */
613 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
614 ((env)->me_psize - PAGEHDRSZ))
615 /** The minimum page fill factor, in tenths of a percent.
616 * Pages emptier than this are candidates for merging.
618 #define FILL_THRESHOLD 250
620 /** Test if a page is a leaf page */
621 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
622 /** Test if a page is a LEAF2 page */
623 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
624 /** Test if a page is a branch page */
625 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
626 /** Test if a page is an overflow page */
627 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
628 /** Test if a page is a sub page */
629 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
631 /** The number of overflow pages needed to store the given size. */
632 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
634 /** Header for a single key/data pair within a page.
635 * We guarantee 2-byte alignment for nodes.
637 typedef struct MDB_node {
638 /** lo and hi are used for data size on leaf nodes and for
639 * child pgno on branch nodes. On 64 bit platforms, flags
640 * is also used for pgno. (Branch nodes have no flags).
641 * They are in host byte order in case that lets some
642 * accesses be optimized into a 32-bit word access.
644 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
645 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
646 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
647 /** @defgroup mdb_node Node Flags
649 * Flags for node headers.
652 #define F_BIGDATA 0x01 /**< data put on overflow page */
653 #define F_SUBDATA 0x02 /**< data is a sub-database */
654 #define F_DUPDATA 0x04 /**< data has duplicates */
656 /** valid flags for #mdb_node_add() */
657 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
660 unsigned short mn_flags; /**< @ref mdb_node */
661 unsigned short mn_ksize; /**< key size */
662 char mn_data[1]; /**< key and data are appended here */
665 /** Size of the node header, excluding dynamic data at the end */
666 #define NODESIZE offsetof(MDB_node, mn_data)
668 /** Bit position of top word in page number, for shifting mn_flags */
669 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
671 /** Size of a node in a branch page with a given key.
672 * This is just the node header plus the key, there is no data.
674 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
676 /** Size of a node in a leaf page with a given key and data.
677 * This is node header plus key plus data size.
679 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
681 /** Address of node \b i in page \b p */
682 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
684 /** Address of the key for the node */
685 #define NODEKEY(node) (void *)((node)->mn_data)
687 /** Address of the data for a node */
688 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
690 /** Get the page number pointed to by a branch node */
691 #define NODEPGNO(node) \
692 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
693 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
694 /** Set the page number in a branch node */
695 #define SETPGNO(node,pgno) do { \
696 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
697 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
699 /** Get the size of the data in a leaf node */
700 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
701 /** Set the size of the data for a leaf node */
702 #define SETDSZ(node,size) do { \
703 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
704 /** The size of a key in a node */
705 #define NODEKSZ(node) ((node)->mn_ksize)
707 /** Copy a page number from src to dst */
709 #define COPY_PGNO(dst,src) dst = src
711 #if SIZE_MAX > 4294967295UL
712 #define COPY_PGNO(dst,src) do { \
713 unsigned short *s, *d; \
714 s = (unsigned short *)&(src); \
715 d = (unsigned short *)&(dst); \
722 #define COPY_PGNO(dst,src) do { \
723 unsigned short *s, *d; \
724 s = (unsigned short *)&(src); \
725 d = (unsigned short *)&(dst); \
731 /** The address of a key in a LEAF2 page.
732 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
733 * There are no node headers, keys are stored contiguously.
735 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
737 /** Set the \b node's key into \b key, if requested. */
738 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
739 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
741 /** Information about a single database in the environment. */
742 typedef struct MDB_db {
743 uint32_t md_pad; /**< also ksize for LEAF2 pages */
744 uint16_t md_flags; /**< @ref mdb_dbi_open */
745 uint16_t md_depth; /**< depth of this tree */
746 pgno_t md_branch_pages; /**< number of internal pages */
747 pgno_t md_leaf_pages; /**< number of leaf pages */
748 pgno_t md_overflow_pages; /**< number of overflow pages */
749 size_t md_entries; /**< number of data items */
750 pgno_t md_root; /**< the root page of this tree */
753 /** mdb_dbi_open flags */
754 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
755 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
756 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
757 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
759 /** Handle for the DB used to track free pages. */
761 /** Handle for the default DB. */
764 /** Meta page content. */
765 typedef struct MDB_meta {
766 /** Stamp identifying this as an MDB file. It must be set
769 /** Version number of this lock file. Must be set to #MDB_VERSION. */
771 void *mm_address; /**< address for fixed mapping */
772 size_t mm_mapsize; /**< size of mmap region */
773 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
774 /** The size of pages used in this DB */
775 #define mm_psize mm_dbs[0].md_pad
776 /** Any persistent environment flags. @ref mdb_env */
777 #define mm_flags mm_dbs[0].md_flags
778 pgno_t mm_last_pg; /**< last used page in file */
779 txnid_t mm_txnid; /**< txnid that committed this page */
782 /** Buffer for a stack-allocated dirty page.
783 * The members define size and alignment, and silence type
784 * aliasing warnings. They are not used directly; that could
785 * mean incorrectly using several union members in parallel.
787 typedef union MDB_pagebuf {
788 char mb_raw[MDB_PAGESIZE];
791 char mm_pad[PAGEHDRSZ];
796 /** Auxiliary DB info.
797 * The information here is mostly static/read-only. There is
798 * only a single copy of this record in the environment.
800 typedef struct MDB_dbx {
801 MDB_val md_name; /**< name of the database */
802 MDB_cmp_func *md_cmp; /**< function for comparing keys */
803 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
804 MDB_rel_func *md_rel; /**< user relocate function */
805 void *md_relctx; /**< user-provided context for md_rel */
808 /** A database transaction.
809 * Every operation requires a transaction handle.
812 MDB_txn *mt_parent; /**< parent of a nested txn */
813 MDB_txn *mt_child; /**< nested txn under this txn */
814 pgno_t mt_next_pgno; /**< next unallocated page */
815 /** The ID of this transaction. IDs are integers incrementing from 1.
816 * Only committed write transactions increment the ID. If a transaction
817 * aborts, the ID may be re-used by the next writer.
820 MDB_env *mt_env; /**< the DB environment */
821 /** The list of pages that became unused during this transaction.
825 MDB_ID2L dirty_list; /**< for write txns: modified pages */
826 MDB_reader *reader; /**< this thread's reader table slot or NULL */
828 /** Array of records for each DB known in the environment. */
830 /** Array of MDB_db records for each known DB */
832 /** @defgroup mt_dbflag Transaction DB Flags
836 #define DB_DIRTY 0x01 /**< DB was written in this txn */
837 #define DB_STALE 0x02 /**< DB record is older than txnID */
838 #define DB_NEW 0x04 /**< DB handle opened in this txn */
839 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
841 /** In write txns, array of cursors for each DB */
842 MDB_cursor **mt_cursors;
843 /** Array of flags for each DB */
844 unsigned char *mt_dbflags;
845 /** Number of DB records in use. This number only ever increments;
846 * we don't decrement it when individual DB handles are closed.
850 /** @defgroup mdb_txn Transaction Flags
854 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
855 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
856 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
858 unsigned int mt_flags; /**< @ref mdb_txn */
859 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
860 unsigned int mt_dirty_room;
861 /** Tracks which of the two meta pages was used at the start
862 * of this transaction.
864 unsigned int mt_toggle;
867 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
868 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
869 * raise this on a 64 bit machine.
871 #define CURSOR_STACK 32
875 /** Cursors are used for all DB operations */
877 /** Next cursor on this DB in this txn */
879 /** Original cursor if this is a shadow */
881 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
882 struct MDB_xcursor *mc_xcursor;
883 /** The transaction that owns this cursor */
885 /** The database handle this cursor operates on */
887 /** The database record for this cursor */
889 /** The database auxiliary record for this cursor */
891 /** The @ref mt_dbflag for this database */
892 unsigned char *mc_dbflag;
893 unsigned short mc_snum; /**< number of pushed pages */
894 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
895 /** @defgroup mdb_cursor Cursor Flags
897 * Cursor state flags.
900 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
901 #define C_EOF 0x02 /**< No more data */
902 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
903 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
904 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
905 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
906 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
908 unsigned int mc_flags; /**< @ref mdb_cursor */
909 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
910 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
913 /** Context for sorted-dup records.
914 * We could have gone to a fully recursive design, with arbitrarily
915 * deep nesting of sub-databases. But for now we only handle these
916 * levels - main DB, optional sub-DB, sorted-duplicate DB.
918 typedef struct MDB_xcursor {
919 /** A sub-cursor for traversing the Dup DB */
920 MDB_cursor mx_cursor;
921 /** The database record for this Dup DB */
923 /** The auxiliary DB record for this Dup DB */
925 /** The @ref mt_dbflag for this Dup DB */
926 unsigned char mx_dbflag;
929 /** State of FreeDB old pages, stored in the MDB_env */
930 typedef struct MDB_pgstate {
931 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
932 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
935 /** The database environment. */
937 HANDLE me_fd; /**< The main data file */
938 HANDLE me_lfd; /**< The lock file */
939 HANDLE me_mfd; /**< just for writing the meta pages */
940 /** Failed to update the meta page. Probably an I/O error. */
941 #define MDB_FATAL_ERROR 0x80000000U
942 /** Some fields are initialized. */
943 #define MDB_ENV_ACTIVE 0x20000000U
944 /** me_txkey is set */
945 #define MDB_ENV_TXKEY 0x10000000U
946 uint32_t me_flags; /**< @ref mdb_env */
947 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
948 unsigned int me_maxreaders; /**< size of the reader table */
949 unsigned int me_numreaders; /**< max numreaders set by this env */
950 MDB_dbi me_numdbs; /**< number of DBs opened */
951 MDB_dbi me_maxdbs; /**< size of the DB table */
952 pid_t me_pid; /**< process ID of this env */
953 char *me_path; /**< path to the DB files */
954 char *me_map; /**< the memory map of the data file */
955 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
956 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
957 MDB_txn *me_txn; /**< current write transaction */
958 size_t me_mapsize; /**< size of the data memory map */
959 off_t me_size; /**< current file size */
960 pgno_t me_maxpg; /**< me_mapsize / me_psize */
961 MDB_dbx *me_dbxs; /**< array of static DB info */
962 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
963 pthread_key_t me_txkey; /**< thread-key for readers */
964 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
965 # define me_pglast me_pgstate.mf_pglast
966 # define me_pghead me_pgstate.mf_pghead
967 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
968 /** IDL of pages that became unused in a write txn */
970 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
971 MDB_ID2L me_dirty_list;
972 /** Max number of freelist items that can fit in a single overflow page */
974 /** Max size of a node on a page */
975 unsigned int me_nodemax;
977 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
979 #elif defined(MDB_USE_POSIX_SEM)
980 sem_t *me_rmutex; /* Shared mutexes are not supported */
985 /** Nested transaction */
986 typedef struct MDB_ntxn {
987 MDB_txn mnt_txn; /* the transaction */
988 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
991 /** max number of pages to commit in one writev() call */
992 #define MDB_COMMIT_PAGES 64
993 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
994 #undef MDB_COMMIT_PAGES
995 #define MDB_COMMIT_PAGES IOV_MAX
998 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
999 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1000 static int mdb_page_touch(MDB_cursor *mc);
1002 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1003 static int mdb_page_search_root(MDB_cursor *mc,
1004 MDB_val *key, int modify);
1005 #define MDB_PS_MODIFY 1
1006 #define MDB_PS_ROOTONLY 2
1007 static int mdb_page_search(MDB_cursor *mc,
1008 MDB_val *key, int flags);
1009 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1011 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1012 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1013 pgno_t newpgno, unsigned int nflags);
1015 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1016 static int mdb_env_pick_meta(const MDB_env *env);
1017 static int mdb_env_write_meta(MDB_txn *txn);
1018 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1019 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1021 static void mdb_env_close0(MDB_env *env, int excl);
1023 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1024 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1025 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1026 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1027 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1028 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1029 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1030 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1031 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1033 static int mdb_rebalance(MDB_cursor *mc);
1034 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1036 static void mdb_cursor_pop(MDB_cursor *mc);
1037 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1039 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1040 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1041 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1042 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1043 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1045 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1046 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1048 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1049 static void mdb_xcursor_init0(MDB_cursor *mc);
1050 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1052 static int mdb_drop0(MDB_cursor *mc, int subs);
1053 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1056 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1060 static SECURITY_DESCRIPTOR mdb_null_sd;
1061 static SECURITY_ATTRIBUTES mdb_all_sa;
1062 static int mdb_sec_inited;
1065 /** Return the library version info. */
1067 mdb_version(int *major, int *minor, int *patch)
1069 if (major) *major = MDB_VERSION_MAJOR;
1070 if (minor) *minor = MDB_VERSION_MINOR;
1071 if (patch) *patch = MDB_VERSION_PATCH;
1072 return MDB_VERSION_STRING;
1075 /** Table of descriptions for MDB @ref errors */
1076 static char *const mdb_errstr[] = {
1077 "MDB_KEYEXIST: Key/data pair already exists",
1078 "MDB_NOTFOUND: No matching key/data pair found",
1079 "MDB_PAGE_NOTFOUND: Requested page not found",
1080 "MDB_CORRUPTED: Located page was wrong type",
1081 "MDB_PANIC: Update of meta page failed",
1082 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1083 "MDB_INVALID: File is not an MDB file",
1084 "MDB_MAP_FULL: Environment mapsize limit reached",
1085 "MDB_DBS_FULL: Environment maxdbs limit reached",
1086 "MDB_READERS_FULL: Environment maxreaders limit reached",
1087 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1088 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1089 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1090 "MDB_PAGE_FULL: Internal error - page has no more space",
1091 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1092 "MDB_INCOMPATIBLE: Database flags changed or would change",
1093 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1097 mdb_strerror(int err)
1101 return ("Successful return: 0");
1103 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1104 i = err - MDB_KEYEXIST;
1105 return mdb_errstr[i];
1108 return strerror(err);
1112 /** Display a key in hexadecimal and return the address of the result.
1113 * @param[in] key the key to display
1114 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1115 * @return The key in hexadecimal form.
1118 mdb_dkey(MDB_val *key, char *buf)
1121 unsigned char *c = key->mv_data;
1127 if (key->mv_size > MDB_MAXKEYSIZE)
1128 return "MDB_MAXKEYSIZE";
1129 /* may want to make this a dynamic check: if the key is mostly
1130 * printable characters, print it as-is instead of converting to hex.
1134 for (i=0; i<key->mv_size; i++)
1135 ptr += sprintf(ptr, "%02x", *c++);
1137 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1142 /** Display all the keys in the page. */
1144 mdb_page_list(MDB_page *mp)
1147 unsigned int i, nkeys, nsize;
1151 nkeys = NUMKEYS(mp);
1152 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1153 for (i=0; i<nkeys; i++) {
1154 node = NODEPTR(mp, i);
1155 key.mv_size = node->mn_ksize;
1156 key.mv_data = node->mn_data;
1157 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1158 if (IS_BRANCH(mp)) {
1159 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1162 if (F_ISSET(node->mn_flags, F_BIGDATA))
1163 nsize += sizeof(pgno_t);
1165 nsize += NODEDSZ(node);
1166 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1172 mdb_cursor_chk(MDB_cursor *mc)
1178 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1179 for (i=0; i<mc->mc_top; i++) {
1181 node = NODEPTR(mp, mc->mc_ki[i]);
1182 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1185 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1191 /** Count all the pages in each DB and in the freelist
1192 * and make sure it matches the actual number of pages
1195 static void mdb_audit(MDB_txn *txn)
1199 MDB_ID freecount, count;
1204 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1205 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1206 freecount += *(MDB_ID *)data.mv_data;
1209 for (i = 0; i<txn->mt_numdbs; i++) {
1211 mdb_cursor_init(&mc, txn, i, &mx);
1212 if (txn->mt_dbs[i].md_root == P_INVALID)
1214 count += txn->mt_dbs[i].md_branch_pages +
1215 txn->mt_dbs[i].md_leaf_pages +
1216 txn->mt_dbs[i].md_overflow_pages;
1217 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1218 mdb_page_search(&mc, NULL, 0);
1222 mp = mc.mc_pg[mc.mc_top];
1223 for (j=0; j<NUMKEYS(mp); j++) {
1224 MDB_node *leaf = NODEPTR(mp, j);
1225 if (leaf->mn_flags & F_SUBDATA) {
1227 memcpy(&db, NODEDATA(leaf), sizeof(db));
1228 count += db.md_branch_pages + db.md_leaf_pages +
1229 db.md_overflow_pages;
1233 while (mdb_cursor_sibling(&mc, 1) == 0);
1236 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1237 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1238 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1244 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1246 return txn->mt_dbxs[dbi].md_cmp(a, b);
1250 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1252 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1255 /** Allocate a page.
1256 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1259 mdb_page_malloc(MDB_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 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1344 /* Get at most <Max_retries> more freeDB records once me_pghead
1345 * has enough pages. If not enough, use new pages from the map.
1346 * If <Paranoid> and mc is updating the freeDB, only get new
1347 * records if me_pghead is empty. Then the freelist cannot play
1348 * catch-up with itself by growing while trying to save it.
1350 enum { Paranoid = 1, Max_retries = 500 };
1352 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1354 int rc, n2 = num-1, retry = Max_retries;
1355 MDB_txn *txn = mc->mc_txn;
1356 MDB_env *env = txn->mt_env;
1357 pgno_t pgno, *mop = env->me_pghead;
1358 unsigned mop_len = mop ? mop[0] : 0;
1361 txnid_t oldest = 0, last;
1367 /* If our dirty list is already full, we can't do anything */
1368 if (txn->mt_dirty_room == 0)
1369 return MDB_TXN_FULL;
1371 for (op = MDB_FIRST;; op = MDB_NEXT) {
1372 unsigned int i, j, k;
1375 pgno_t *idl, old_id, new_id;
1377 /* Seek a big enough contiguous page range. Prefer
1378 * pages at the tail, just truncating the list.
1380 if (mop_len >= (unsigned)num) {
1384 if (mop[i-n2] == pgno+n2) {
1385 mop[0] = mop_len -= num;
1386 /* Move any stragglers down */
1387 for (j = i-n2; j <= mop_len; )
1388 mop[j++] = mop[++i];
1391 } while (--i >= (unsigned)num);
1392 if (Max_retries < INT_MAX && --retry < 0)
1396 if (op == MDB_FIRST) { /* 1st iteration */
1397 /* Prepare to fetch more and coalesce */
1398 oldest = mdb_find_oldest(txn);
1399 last = env->me_pglast;
1400 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1403 key.mv_data = &last; /* will loop up last+1 */
1404 key.mv_size = sizeof(last);
1406 if (Paranoid && mc->mc_dbi == FREE_DBI)
1409 if (Paranoid && retry < 0 && mop_len)
1413 /* Do not fetch more if the record will be too recent */
1416 rc = mdb_cursor_get(&m2, &key, NULL, op);
1418 if (rc == MDB_NOTFOUND)
1422 last = *(txnid_t*)key.mv_data;
1425 np = m2.mc_pg[m2.mc_top];
1426 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1427 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1430 idl = (MDB_ID *) data.mv_data;
1433 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1435 } else if (mop_len+i > mop[-1]) {
1436 if ((rc = mdb_midl_grow(&env->me_pghead, i)) != 0)
1438 mop = env->me_pghead;
1440 env->me_pglast = last;
1442 DPRINTF("IDL read txn %zu root %zu num %u",
1443 last, txn->mt_dbs[FREE_DBI].md_root, i);
1445 DPRINTF("IDL %zu", idl[k]);
1447 /* Merge in descending sorted order */
1450 mop[0] = (pgno_t)-1;
1454 for (; old_id < new_id; old_id = mop[--j])
1461 /* Use new pages from the map when nothing suitable in the freeDB */
1463 if (txn->mt_next_pgno + num >= env->me_maxpg) {
1464 DPUTS("DB size maxed out");
1465 return MDB_MAP_FULL;
1469 if (env->me_flags & MDB_WRITEMAP) {
1470 if (pgno == P_INVALID) {
1471 pgno = txn->mt_next_pgno;
1472 txn->mt_next_pgno += num;
1474 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1476 if (!(np = mdb_page_malloc(mc, num)))
1478 if (pgno == P_INVALID) {
1479 pgno = txn->mt_next_pgno;
1480 txn->mt_next_pgno += num;
1483 mid.mid = np->mp_pgno = pgno;
1485 if (env->me_flags & MDB_WRITEMAP) {
1486 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1488 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1490 txn->mt_dirty_room--;
1496 /** Copy the used portions of a non-overflow page.
1497 * @param[in] dst page to copy into
1498 * @param[in] src page to copy from
1499 * @param[in] psize size of a page
1502 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1504 enum { Align = sizeof(pgno_t) };
1505 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1507 /* If page isn't full, just copy the used portion. Adjust
1508 * alignment so memcpy may copy words instead of bytes.
1510 if ((unused &= -Align) && !IS_LEAF2(src)) {
1512 memcpy(dst, src, (lower + (Align-1)) & -Align);
1513 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1516 memcpy(dst, src, psize - unused);
1520 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1521 * @param[in] mc cursor pointing to the page to be touched
1522 * @return 0 on success, non-zero on failure.
1525 mdb_page_touch(MDB_cursor *mc)
1527 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1528 MDB_cursor *m2, *m3;
1533 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1534 if ((rc = mdb_page_alloc(mc, 1, &np)))
1537 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi,mp->mp_pgno,pgno);
1538 assert(mp->mp_pgno != pgno);
1539 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1540 /* Update the parent page, if any, to point to the new page */
1542 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1543 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1544 SETPGNO(node, pgno);
1546 mc->mc_db->md_root = pgno;
1548 } else if (mc->mc_txn->mt_parent && !(mp->mp_flags & P_SUBP)) {
1549 MDB_ID2 mid, *dl = mc->mc_txn->mt_u.dirty_list;
1551 /* If txn has a parent, make sure the page is in our
1555 unsigned x = mdb_mid2l_search(dl, pgno);
1556 if (x <= dl[0].mid && dl[x].mid == pgno) {
1559 mc->mc_pg[mc->mc_top] = np;
1563 assert(dl[0].mid < MDB_IDL_UM_MAX);
1565 np = mdb_page_malloc(mc, 1);
1570 mdb_mid2l_insert(dl, &mid);
1575 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1577 np->mp_flags |= P_DIRTY;
1579 /* Adjust cursors pointing to mp */
1580 mc->mc_pg[mc->mc_top] = np;
1582 if (mc->mc_flags & C_SUB) {
1584 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1585 m3 = &m2->mc_xcursor->mx_cursor;
1586 if (m3->mc_snum < mc->mc_snum) continue;
1587 if (m3->mc_pg[mc->mc_top] == mp)
1588 m3->mc_pg[mc->mc_top] = np;
1591 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1592 if (m2->mc_snum < mc->mc_snum) continue;
1593 if (m2->mc_pg[mc->mc_top] == mp) {
1594 m2->mc_pg[mc->mc_top] = np;
1595 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1596 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1598 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1599 if (!(leaf->mn_flags & F_SUBDATA))
1600 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1609 mdb_env_sync(MDB_env *env, int force)
1612 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1613 if (env->me_flags & MDB_WRITEMAP) {
1614 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1615 ? MS_ASYNC : MS_SYNC;
1616 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1619 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1623 if (MDB_FDATASYNC(env->me_fd))
1630 /** Make shadow copies of all of parent txn's cursors */
1632 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1634 MDB_cursor *mc, *m2;
1635 unsigned int i, j, size;
1637 for (i=0;i<src->mt_numdbs; i++) {
1638 if (src->mt_cursors[i]) {
1639 size = sizeof(MDB_cursor);
1640 if (src->mt_cursors[i]->mc_xcursor)
1641 size += sizeof(MDB_xcursor);
1642 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1649 mc->mc_db = &dst->mt_dbs[i];
1650 mc->mc_dbx = m2->mc_dbx;
1651 mc->mc_dbflag = &dst->mt_dbflags[i];
1652 mc->mc_snum = m2->mc_snum;
1653 mc->mc_top = m2->mc_top;
1654 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1655 for (j=0; j<mc->mc_snum; j++) {
1656 mc->mc_pg[j] = m2->mc_pg[j];
1657 mc->mc_ki[j] = m2->mc_ki[j];
1659 if (m2->mc_xcursor) {
1660 MDB_xcursor *mx, *mx2;
1661 mx = (MDB_xcursor *)(mc+1);
1662 mc->mc_xcursor = mx;
1663 mx2 = m2->mc_xcursor;
1664 mx->mx_db = mx2->mx_db;
1665 mx->mx_dbx = mx2->mx_dbx;
1666 mx->mx_dbflag = mx2->mx_dbflag;
1667 mx->mx_cursor.mc_txn = dst;
1668 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1669 mx->mx_cursor.mc_db = &mx->mx_db;
1670 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1671 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1672 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1673 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1674 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1675 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1676 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1677 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1680 mc->mc_xcursor = NULL;
1682 mc->mc_next = dst->mt_cursors[i];
1683 dst->mt_cursors[i] = mc;
1690 /** Close this write txn's cursors, after optionally merging its shadow
1691 * cursors back into parent's.
1692 * @param[in] txn the transaction handle.
1693 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1694 * @return 0 on success, non-zero on failure.
1697 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1699 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1702 for (i = txn->mt_numdbs; --i >= 0; ) {
1703 for (mc = cursors[i]; mc; mc = next) {
1705 if (mc->mc_flags & merge) {
1706 MDB_cursor *m2 = mc->mc_orig;
1707 m2->mc_snum = mc->mc_snum;
1708 m2->mc_top = mc->mc_top;
1709 for (j = mc->mc_snum; --j >= 0; ) {
1710 m2->mc_pg[j] = mc->mc_pg[j];
1711 m2->mc_ki[j] = mc->mc_ki[j];
1714 if (mc->mc_flags & C_ALLOCD)
1722 mdb_txn_reset0(MDB_txn *txn);
1724 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1725 * @param[in] txn the transaction handle to initialize
1726 * @return 0 on success, non-zero on failure.
1729 mdb_txn_renew0(MDB_txn *txn)
1731 MDB_env *env = txn->mt_env;
1734 int rc, new_notls = 0;
1737 txn->mt_numdbs = env->me_numdbs;
1738 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1740 if (txn->mt_flags & MDB_TXN_RDONLY) {
1741 if (!env->me_txns) {
1742 i = mdb_env_pick_meta(env);
1743 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1744 txn->mt_u.reader = NULL;
1746 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1747 pthread_getspecific(env->me_txkey);
1749 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1750 return MDB_BAD_RSLOT;
1752 pid_t pid = env->me_pid;
1753 pthread_t tid = pthread_self();
1756 for (i=0; i<env->me_txns->mti_numreaders; i++)
1757 if (env->me_txns->mti_readers[i].mr_pid == 0)
1759 if (i == env->me_maxreaders) {
1760 UNLOCK_MUTEX_R(env);
1761 return MDB_READERS_FULL;
1763 env->me_txns->mti_readers[i].mr_pid = pid;
1764 env->me_txns->mti_readers[i].mr_tid = tid;
1765 if (i >= env->me_txns->mti_numreaders)
1766 env->me_txns->mti_numreaders = i+1;
1767 /* Save numreaders for un-mutexed mdb_env_close() */
1768 env->me_numreaders = env->me_txns->mti_numreaders;
1769 UNLOCK_MUTEX_R(env);
1770 r = &env->me_txns->mti_readers[i];
1771 new_notls = (env->me_flags & MDB_NOTLS);
1772 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1777 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1778 txn->mt_u.reader = r;
1780 txn->mt_toggle = txn->mt_txnid & 1;
1781 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1785 txn->mt_txnid = env->me_txns->mti_txnid;
1786 txn->mt_toggle = txn->mt_txnid & 1;
1787 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1790 if (txn->mt_txnid == mdb_debug_start)
1793 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1794 txn->mt_u.dirty_list = env->me_dirty_list;
1795 txn->mt_u.dirty_list[0].mid = 0;
1796 txn->mt_free_pgs = env->me_free_pgs;
1797 txn->mt_free_pgs[0] = 0;
1801 /* Copy the DB info and flags */
1802 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1803 for (i=2; i<txn->mt_numdbs; i++) {
1804 x = env->me_dbflags[i];
1805 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1806 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1808 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1810 if (env->me_maxpg < txn->mt_next_pgno) {
1811 mdb_txn_reset0(txn);
1813 txn->mt_u.reader->mr_pid = 0;
1814 txn->mt_u.reader = NULL;
1816 return MDB_MAP_RESIZED;
1823 mdb_txn_renew(MDB_txn *txn)
1827 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1830 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1831 DPUTS("environment had fatal error, must shutdown!");
1835 rc = mdb_txn_renew0(txn);
1836 if (rc == MDB_SUCCESS) {
1837 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1838 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1839 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1845 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1849 int rc, size, tsize = sizeof(MDB_txn);
1851 if (env->me_flags & MDB_FATAL_ERROR) {
1852 DPUTS("environment had fatal error, must shutdown!");
1855 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1858 /* Nested transactions: Max 1 child, write txns only, no writemap */
1859 if (parent->mt_child ||
1860 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1861 (env->me_flags & MDB_WRITEMAP))
1865 tsize = sizeof(MDB_ntxn);
1867 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1868 if (!(flags & MDB_RDONLY))
1869 size += env->me_maxdbs * sizeof(MDB_cursor *);
1871 if ((txn = calloc(1, size)) == NULL) {
1872 DPRINTF("calloc: %s", strerror(ErrCode()));
1875 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1876 if (flags & MDB_RDONLY) {
1877 txn->mt_flags |= MDB_TXN_RDONLY;
1878 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1880 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1881 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1887 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1888 if (!txn->mt_u.dirty_list ||
1889 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1891 free(txn->mt_u.dirty_list);
1895 txn->mt_txnid = parent->mt_txnid;
1896 txn->mt_toggle = parent->mt_toggle;
1897 txn->mt_dirty_room = parent->mt_dirty_room;
1898 txn->mt_u.dirty_list[0].mid = 0;
1899 txn->mt_next_pgno = parent->mt_next_pgno;
1900 parent->mt_child = txn;
1901 txn->mt_parent = parent;
1902 txn->mt_numdbs = parent->mt_numdbs;
1903 txn->mt_flags = parent->mt_flags;
1904 txn->mt_dbxs = parent->mt_dbxs;
1905 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1906 /* Copy parent's mt_dbflags, but clear DB_NEW */
1907 for (i=0; i<txn->mt_numdbs; i++)
1908 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1910 ntxn = (MDB_ntxn *)txn;
1911 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1912 if (env->me_pghead) {
1913 size = MDB_IDL_SIZEOF(env->me_pghead);
1914 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1916 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1921 rc = mdb_cursor_shadow(parent, txn);
1923 mdb_txn_reset0(txn);
1925 rc = mdb_txn_renew0(txn);
1931 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1932 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1933 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1939 /** Export or close DBI handles opened in this txn. */
1941 mdb_dbis_update(MDB_txn *txn, int keep)
1944 MDB_dbi n = txn->mt_numdbs;
1945 MDB_env *env = txn->mt_env;
1946 unsigned char *tdbflags = txn->mt_dbflags;
1948 for (i = n; --i >= 2;) {
1949 if (tdbflags[i] & DB_NEW) {
1951 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
1953 char *ptr = env->me_dbxs[i].md_name.mv_data;
1954 env->me_dbxs[i].md_name.mv_data = NULL;
1955 env->me_dbxs[i].md_name.mv_size = 0;
1956 env->me_dbflags[i] = 0;
1961 if (keep && env->me_numdbs < n)
1965 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1966 * May be called twice for readonly txns: First reset it, then abort.
1967 * @param[in] txn the transaction handle to reset
1970 mdb_txn_reset0(MDB_txn *txn)
1972 MDB_env *env = txn->mt_env;
1974 /* Close any DBI handles opened in this txn */
1975 mdb_dbis_update(txn, 0);
1977 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1978 if (txn->mt_u.reader) {
1979 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1980 if (!(env->me_flags & MDB_NOTLS))
1981 txn->mt_u.reader = NULL; /* txn does not own reader */
1983 txn->mt_numdbs = 0; /* close nothing if called again */
1984 txn->mt_dbxs = NULL; /* mark txn as reset */
1986 mdb_cursors_close(txn, 0);
1988 if (!(env->me_flags & MDB_WRITEMAP)) {
1989 mdb_dlist_free(txn);
1991 mdb_midl_free(env->me_pghead);
1993 if (txn->mt_parent) {
1994 txn->mt_parent->mt_child = NULL;
1995 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
1996 mdb_midl_free(txn->mt_free_pgs);
1997 free(txn->mt_u.dirty_list);
2000 if (mdb_midl_shrink(&txn->mt_free_pgs))
2001 env->me_free_pgs = txn->mt_free_pgs;
2004 txn->mt_env->me_pghead = NULL;
2005 txn->mt_env->me_pglast = 0;
2008 /* The writer mutex was locked in mdb_txn_begin. */
2009 UNLOCK_MUTEX_W(env);
2014 mdb_txn_reset(MDB_txn *txn)
2019 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2020 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2021 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2023 /* This call is only valid for read-only txns */
2024 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2027 mdb_txn_reset0(txn);
2031 mdb_txn_abort(MDB_txn *txn)
2036 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2037 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2038 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2041 mdb_txn_abort(txn->mt_child);
2043 mdb_txn_reset0(txn);
2044 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2045 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2046 txn->mt_u.reader->mr_pid = 0;
2051 /** Save the freelist as of this transaction to the freeDB.
2052 * This changes the freelist. Keep trying until it stabilizes.
2055 mdb_freelist_save(MDB_txn *txn)
2057 /* env->me_pghead[] can grow and shrink during this call.
2058 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2059 * Page numbers cannot disappear from txn->mt_free_pgs[].
2062 MDB_env *env = txn->mt_env;
2063 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2064 txnid_t pglast = 0, head_id = 0;
2065 pgno_t freecnt = 0, *free_pgs, *mop;
2066 ssize_t head_room = 0, total_room = 0, mop_len;
2068 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2070 if (env->me_pghead) {
2071 /* Make sure first page of freeDB is touched and on freelist */
2072 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2073 if (rc && rc != MDB_NOTFOUND)
2078 /* Come back here after each Put() in case freelist changed */
2081 /* If using records from freeDB which we have not yet
2082 * deleted, delete them and any we reserved for me_pghead.
2084 while (pglast < env->me_pglast) {
2085 rc = mdb_cursor_first(&mc, &key, NULL);
2088 pglast = head_id = *(txnid_t *)key.mv_data;
2089 total_room = head_room = 0;
2090 assert(pglast <= env->me_pglast);
2091 rc = mdb_cursor_del(&mc, 0);
2096 /* Save the IDL of pages freed by this txn, to a single record */
2097 if (freecnt < txn->mt_free_pgs[0]) {
2099 /* Make sure last page of freeDB is touched and on freelist */
2100 key.mv_size = MDB_MAXKEYSIZE+1;
2102 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2103 if (rc && rc != MDB_NOTFOUND)
2106 free_pgs = txn->mt_free_pgs;
2107 /* Write to last page of freeDB */
2108 key.mv_size = sizeof(txn->mt_txnid);
2109 key.mv_data = &txn->mt_txnid;
2111 freecnt = free_pgs[0];
2112 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2113 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2116 /* Retry if mt_free_pgs[] grew during the Put() */
2117 free_pgs = txn->mt_free_pgs;
2118 } while (freecnt < free_pgs[0]);
2119 mdb_midl_sort(free_pgs);
2120 memcpy(data.mv_data, free_pgs, data.mv_size);
2123 unsigned int i = free_pgs[0];
2124 DPRINTF("IDL write txn %zu root %zu num %u",
2125 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2127 DPRINTF("IDL %zu", free_pgs[i]);
2133 mop = env->me_pghead;
2134 mop_len = mop ? mop[0] : 0;
2136 /* Reserve records for me_pghead[]. Split it if multi-page,
2137 * to avoid searching freeDB for a page range. Use keys in
2138 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2140 if (total_room >= mop_len) {
2141 if (total_room == mop_len || --more < 0)
2143 } else if (head_room >= maxfree_1pg && head_id > 1) {
2144 /* Keep current record (overflow page), add a new one */
2148 /* (Re)write {key = head_id, IDL length = head_room} */
2149 total_room -= head_room;
2150 head_room = mop_len - total_room;
2151 if (head_room > maxfree_1pg && head_id > 1) {
2152 /* Overflow multi-page for part of me_pghead */
2153 head_room /= head_id; /* amortize page sizes */
2154 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2155 } else if (head_room < 0) {
2156 /* Rare case, not bothering to delete this record */
2159 key.mv_size = sizeof(head_id);
2160 key.mv_data = &head_id;
2161 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2162 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2165 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2166 total_room += head_room;
2169 /* Fill in the reserved, touched me_pghead records. Avoid write ops
2170 * so they cannot rearrange anything, just read the destinations.
2177 rc = mdb_cursor_first(&mc, &key, &data);
2178 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2179 MDB_IDL dest = data.mv_data;
2180 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2182 assert(len >= 0 && *(txnid_t*)key.mv_data <= env->me_pglast);
2186 memcpy(dest, mop -= len, len * sizeof(MDB_ID));
2187 if (! (mop_len -= len))
2195 mdb_txn_commit(MDB_txn *txn)
2205 assert(txn != NULL);
2206 assert(txn->mt_env != NULL);
2208 if (txn->mt_child) {
2209 rc = mdb_txn_commit(txn->mt_child);
2210 txn->mt_child = NULL;
2219 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2220 mdb_dbis_update(txn, 1);
2221 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2226 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2227 DPUTS("error flag is set, can't commit");
2229 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2234 if (txn->mt_parent) {
2235 MDB_txn *parent = txn->mt_parent;
2239 /* Append our free list to parent's */
2240 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2244 mdb_midl_free(txn->mt_free_pgs);
2246 parent->mt_next_pgno = txn->mt_next_pgno;
2247 parent->mt_flags = txn->mt_flags;
2249 /* Merge our cursors into parent's and close them */
2250 mdb_cursors_close(txn, C_SHADOW);
2252 /* Update parent's DB table. */
2253 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2254 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2255 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2256 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2257 for (i=2; i<txn->mt_numdbs; i++) {
2258 /* preserve parent's DB_NEW status */
2259 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2260 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2263 dst = txn->mt_parent->mt_u.dirty_list;
2264 src = txn->mt_u.dirty_list;
2265 /* Find len = length of merging our dirty list with parent's */
2267 dst[0].mid = 0; /* simplify loops */
2268 if (parent->mt_parent) {
2269 len = x + src[0].mid;
2270 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2271 for (i = x; y && i; y--) {
2272 pgno_t yp = src[y].mid;
2273 while (yp < dst[i].mid)
2275 if (yp == dst[i].mid) {
2280 } else { /* Simplify the above for single-ancestor case */
2281 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2283 /* Merge our dirty list with parent's */
2285 for (i = len; y; dst[i--] = src[y--]) {
2286 pgno_t yp = src[y].mid;
2287 while (yp < dst[x].mid)
2288 dst[i--] = dst[x--];
2289 if (yp == dst[x].mid)
2290 free(dst[x--].mptr);
2294 free(txn->mt_u.dirty_list);
2295 parent->mt_dirty_room = txn->mt_dirty_room;
2297 txn->mt_parent->mt_child = NULL;
2298 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2303 if (txn != env->me_txn) {
2304 DPUTS("attempt to commit unknown transaction");
2309 mdb_cursors_close(txn, 0);
2311 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2314 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2315 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2317 /* Update DB root pointers */
2318 if (txn->mt_numdbs > 2) {
2322 data.mv_size = sizeof(MDB_db);
2324 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2325 for (i = 2; i < txn->mt_numdbs; i++) {
2326 if (txn->mt_dbflags[i] & DB_DIRTY) {
2327 data.mv_data = &txn->mt_dbs[i];
2328 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2335 rc = mdb_freelist_save(txn);
2339 mdb_midl_free(env->me_pghead);
2340 env->me_pghead = NULL;
2341 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2342 if (mdb_midl_shrink(&txn->mt_free_pgs))
2343 env->me_free_pgs = txn->mt_free_pgs;
2350 if (env->me_flags & MDB_WRITEMAP) {
2351 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2352 dp = txn->mt_u.dirty_list[i].mptr;
2353 /* clear dirty flag */
2354 dp->mp_flags &= ~P_DIRTY;
2356 txn->mt_u.dirty_list[0].mid = 0;
2360 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2366 /* Windows actually supports scatter/gather I/O, but only on
2367 * unbuffered file handles. Since we're relying on the OS page
2368 * cache for all our data, that's self-defeating. So we just
2369 * write pages one at a time. We use the ov structure to set
2370 * the write offset, to at least save the overhead of a Seek
2374 memset(&ov, 0, sizeof(ov));
2375 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2377 dp = txn->mt_u.dirty_list[i].mptr;
2378 DPRINTF("committing page %zu", dp->mp_pgno);
2379 size = dp->mp_pgno * env->me_psize;
2380 ov.Offset = size & 0xffffffff;
2381 ov.OffsetHigh = size >> 16;
2382 ov.OffsetHigh >>= 16;
2383 /* clear dirty flag */
2384 dp->mp_flags &= ~P_DIRTY;
2385 wsize = env->me_psize;
2386 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2387 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2390 DPRINTF("WriteFile: %d", n);
2397 struct iovec iov[MDB_COMMIT_PAGES];
2401 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2402 dp = txn->mt_u.dirty_list[i].mptr;
2403 if (dp->mp_pgno != next) {
2405 rc = writev(env->me_fd, iov, n);
2409 DPUTS("short write, filesystem full?");
2411 DPRINTF("writev: %s", strerror(n));
2418 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2421 DPRINTF("committing page %zu", dp->mp_pgno);
2422 iov[n].iov_len = env->me_psize;
2423 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2424 iov[n].iov_base = (char *)dp;
2425 size += iov[n].iov_len;
2426 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2427 /* clear dirty flag */
2428 dp->mp_flags &= ~P_DIRTY;
2429 if (++n >= MDB_COMMIT_PAGES) {
2439 rc = writev(env->me_fd, iov, n);
2443 DPUTS("short write, filesystem full?");
2445 DPRINTF("writev: %s", strerror(n));
2452 mdb_dlist_free(txn);
2455 if ((n = mdb_env_sync(env, 0)) != 0 ||
2456 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2464 mdb_dbis_update(txn, 1);
2466 UNLOCK_MUTEX_W(env);
2476 /** Read the environment parameters of a DB environment before
2477 * mapping it into memory.
2478 * @param[in] env the environment handle
2479 * @param[out] meta address of where to store the meta information
2480 * @return 0 on success, non-zero on failure.
2483 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2490 /* We don't know the page size yet, so use a minimum value.
2491 * Read both meta pages so we can use the latest one.
2494 for (i=0; i<2; i++) {
2496 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2498 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2503 else if (rc != MDB_PAGESIZE) {
2507 DPRINTF("read: %s", strerror(err));
2511 p = (MDB_page *)&pbuf;
2513 if (!F_ISSET(p->mp_flags, P_META)) {
2514 DPRINTF("page %zu not a meta page", p->mp_pgno);
2519 if (m->mm_magic != MDB_MAGIC) {
2520 DPUTS("meta has invalid magic");
2524 if (m->mm_version != MDB_VERSION) {
2525 DPRINTF("database is version %u, expected version %u",
2526 m->mm_version, MDB_VERSION);
2527 return MDB_VERSION_MISMATCH;
2531 if (m->mm_txnid > meta->mm_txnid)
2532 memcpy(meta, m, sizeof(*m));
2534 memcpy(meta, m, sizeof(*m));
2536 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2538 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2546 /** Write the environment parameters of a freshly created DB environment.
2547 * @param[in] env the environment handle
2548 * @param[out] meta address of where to store the meta information
2549 * @return 0 on success, non-zero on failure.
2552 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2559 DPUTS("writing new meta page");
2561 GET_PAGESIZE(psize);
2563 meta->mm_magic = MDB_MAGIC;
2564 meta->mm_version = MDB_VERSION;
2565 meta->mm_mapsize = env->me_mapsize;
2566 meta->mm_psize = psize;
2567 meta->mm_last_pg = 1;
2568 meta->mm_flags = env->me_flags & 0xffff;
2569 meta->mm_flags |= MDB_INTEGERKEY;
2570 meta->mm_dbs[0].md_root = P_INVALID;
2571 meta->mm_dbs[1].md_root = P_INVALID;
2573 p = calloc(2, psize);
2575 p->mp_flags = P_META;
2578 memcpy(m, meta, sizeof(*meta));
2580 q = (MDB_page *)((char *)p + psize);
2583 q->mp_flags = P_META;
2586 memcpy(m, meta, sizeof(*meta));
2591 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2592 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2593 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2596 lseek(env->me_fd, 0, SEEK_SET);
2597 rc = write(env->me_fd, p, psize * 2);
2598 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2604 /** Update the environment info to commit a transaction.
2605 * @param[in] txn the transaction that's being committed
2606 * @return 0 on success, non-zero on failure.
2609 mdb_env_write_meta(MDB_txn *txn)
2612 MDB_meta meta, metab, *mp;
2614 int rc, len, toggle;
2621 assert(txn != NULL);
2622 assert(txn->mt_env != NULL);
2624 toggle = !txn->mt_toggle;
2625 DPRINTF("writing meta page %d for root page %zu",
2626 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2629 mp = env->me_metas[toggle];
2631 if (env->me_flags & MDB_WRITEMAP) {
2632 /* Persist any increases of mapsize config */
2633 if (env->me_mapsize > mp->mm_mapsize)
2634 mp->mm_mapsize = env->me_mapsize;
2635 mp->mm_dbs[0] = txn->mt_dbs[0];
2636 mp->mm_dbs[1] = txn->mt_dbs[1];
2637 mp->mm_last_pg = txn->mt_next_pgno - 1;
2638 mp->mm_txnid = txn->mt_txnid;
2639 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2640 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2643 ptr += env->me_psize;
2644 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2651 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2652 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2654 ptr = (char *)&meta;
2655 if (env->me_mapsize > mp->mm_mapsize) {
2656 /* Persist any increases of mapsize config */
2657 meta.mm_mapsize = env->me_mapsize;
2658 off = offsetof(MDB_meta, mm_mapsize);
2660 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2662 len = sizeof(MDB_meta) - off;
2665 meta.mm_dbs[0] = txn->mt_dbs[0];
2666 meta.mm_dbs[1] = txn->mt_dbs[1];
2667 meta.mm_last_pg = txn->mt_next_pgno - 1;
2668 meta.mm_txnid = txn->mt_txnid;
2671 off += env->me_psize;
2674 /* Write to the SYNC fd */
2675 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2676 env->me_fd : env->me_mfd;
2679 memset(&ov, 0, sizeof(ov));
2681 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2684 rc = pwrite(mfd, ptr, len, off);
2689 DPUTS("write failed, disk error?");
2690 /* On a failure, the pagecache still contains the new data.
2691 * Write some old data back, to prevent it from being used.
2692 * Use the non-SYNC fd; we know it will fail anyway.
2694 meta.mm_last_pg = metab.mm_last_pg;
2695 meta.mm_txnid = metab.mm_txnid;
2697 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2699 r2 = pwrite(env->me_fd, ptr, len, off);
2702 env->me_flags |= MDB_FATAL_ERROR;
2706 /* Memory ordering issues are irrelevant; since the entire writer
2707 * is wrapped by wmutex, all of these changes will become visible
2708 * after the wmutex is unlocked. Since the DB is multi-version,
2709 * readers will get consistent data regardless of how fresh or
2710 * how stale their view of these values is.
2712 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2717 /** Check both meta pages to see which one is newer.
2718 * @param[in] env the environment handle
2719 * @return meta toggle (0 or 1).
2722 mdb_env_pick_meta(const MDB_env *env)
2724 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2728 mdb_env_create(MDB_env **env)
2732 e = calloc(1, sizeof(MDB_env));
2736 e->me_maxreaders = DEFAULT_READERS;
2737 e->me_maxdbs = e->me_numdbs = 2;
2738 e->me_fd = INVALID_HANDLE_VALUE;
2739 e->me_lfd = INVALID_HANDLE_VALUE;
2740 e->me_mfd = INVALID_HANDLE_VALUE;
2741 #ifdef MDB_USE_POSIX_SEM
2742 e->me_rmutex = SEM_FAILED;
2743 e->me_wmutex = SEM_FAILED;
2745 e->me_pid = getpid();
2746 VGMEMP_CREATE(e,0,0);
2752 mdb_env_set_mapsize(MDB_env *env, size_t size)
2756 env->me_mapsize = size;
2758 env->me_maxpg = env->me_mapsize / env->me_psize;
2763 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2767 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2772 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2774 if (env->me_map || readers < 1)
2776 env->me_maxreaders = readers;
2781 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2783 if (!env || !readers)
2785 *readers = env->me_maxreaders;
2789 /** Further setup required for opening an MDB environment
2792 mdb_env_open2(MDB_env *env)
2794 unsigned int flags = env->me_flags;
2795 int i, newenv = 0, prot;
2799 memset(&meta, 0, sizeof(meta));
2801 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2804 DPUTS("new mdbenv");
2808 /* Was a mapsize configured? */
2809 if (!env->me_mapsize) {
2810 /* If this is a new environment, take the default,
2811 * else use the size recorded in the existing env.
2813 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2814 } else if (env->me_mapsize < meta.mm_mapsize) {
2815 /* If the configured size is smaller, make sure it's
2816 * still big enough. Silently round up to minimum if not.
2818 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2819 if (env->me_mapsize < minsize)
2820 env->me_mapsize = minsize;
2826 LONG sizelo, sizehi;
2827 sizelo = env->me_mapsize & 0xffffffff;
2828 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2830 /* Windows won't create mappings for zero length files.
2831 * Just allocate the maxsize right now.
2834 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2835 if (!SetEndOfFile(env->me_fd))
2837 SetFilePointer(env->me_fd, 0, NULL, 0);
2839 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2840 PAGE_READWRITE : PAGE_READONLY,
2841 sizehi, sizelo, NULL);
2844 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2845 FILE_MAP_WRITE : FILE_MAP_READ,
2846 0, 0, env->me_mapsize, meta.mm_address);
2854 if (flags & MDB_WRITEMAP) {
2856 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2859 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2861 if (env->me_map == MAP_FAILED) {
2865 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2867 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2869 #ifdef POSIX_MADV_RANDOM
2870 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2871 #endif /* POSIX_MADV_RANDOM */
2872 #endif /* MADV_RANDOM */
2876 if (flags & MDB_FIXEDMAP)
2877 meta.mm_address = env->me_map;
2878 i = mdb_env_init_meta(env, &meta);
2879 if (i != MDB_SUCCESS) {
2882 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2883 /* Can happen because the address argument to mmap() is just a
2884 * hint. mmap() can pick another, e.g. if the range is in use.
2885 * The MAP_FIXED flag would prevent that, but then mmap could
2886 * instead unmap existing pages to make room for the new map.
2888 return EBUSY; /* TODO: Make a new MDB_* error code? */
2890 env->me_psize = meta.mm_psize;
2891 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2892 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2894 env->me_maxpg = env->me_mapsize / env->me_psize;
2896 p = (MDB_page *)env->me_map;
2897 env->me_metas[0] = METADATA(p);
2898 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2902 int toggle = mdb_env_pick_meta(env);
2903 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2905 DPRINTF("opened database version %u, pagesize %u",
2906 env->me_metas[0]->mm_version, env->me_psize);
2907 DPRINTF("using meta page %d", toggle);
2908 DPRINTF("depth: %u", db->md_depth);
2909 DPRINTF("entries: %zu", db->md_entries);
2910 DPRINTF("branch pages: %zu", db->md_branch_pages);
2911 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2912 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2913 DPRINTF("root: %zu", db->md_root);
2921 /** Release a reader thread's slot in the reader lock table.
2922 * This function is called automatically when a thread exits.
2923 * @param[in] ptr This points to the slot in the reader lock table.
2926 mdb_env_reader_dest(void *ptr)
2928 MDB_reader *reader = ptr;
2934 /** Junk for arranging thread-specific callbacks on Windows. This is
2935 * necessarily platform and compiler-specific. Windows supports up
2936 * to 1088 keys. Let's assume nobody opens more than 64 environments
2937 * in a single process, for now. They can override this if needed.
2939 #ifndef MAX_TLS_KEYS
2940 #define MAX_TLS_KEYS 64
2942 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2943 static int mdb_tls_nkeys;
2945 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2949 case DLL_PROCESS_ATTACH: break;
2950 case DLL_THREAD_ATTACH: break;
2951 case DLL_THREAD_DETACH:
2952 for (i=0; i<mdb_tls_nkeys; i++) {
2953 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2954 mdb_env_reader_dest(r);
2957 case DLL_PROCESS_DETACH: break;
2962 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2964 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2968 /* Force some symbol references.
2969 * _tls_used forces the linker to create the TLS directory if not already done
2970 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2972 #pragma comment(linker, "/INCLUDE:_tls_used")
2973 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2974 #pragma const_seg(".CRT$XLB")
2975 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2976 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2979 #pragma comment(linker, "/INCLUDE:__tls_used")
2980 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2981 #pragma data_seg(".CRT$XLB")
2982 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2984 #endif /* WIN 32/64 */
2985 #endif /* !__GNUC__ */
2988 /** Downgrade the exclusive lock on the region back to shared */
2990 mdb_env_share_locks(MDB_env *env, int *excl)
2992 int rc = 0, toggle = mdb_env_pick_meta(env);
2994 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2999 /* First acquire a shared lock. The Unlock will
3000 * then release the existing exclusive lock.
3002 memset(&ov, 0, sizeof(ov));
3003 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3006 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3012 struct flock lock_info;
3013 /* The shared lock replaces the existing lock */
3014 memset((void *)&lock_info, 0, sizeof(lock_info));
3015 lock_info.l_type = F_RDLCK;
3016 lock_info.l_whence = SEEK_SET;
3017 lock_info.l_start = 0;
3018 lock_info.l_len = 1;
3019 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3020 (rc = ErrCode()) == EINTR) ;
3021 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3028 /** Try to get exlusive lock, otherwise shared.
3029 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3032 mdb_env_excl_lock(MDB_env *env, int *excl)
3036 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3040 memset(&ov, 0, sizeof(ov));
3041 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3048 struct flock lock_info;
3049 memset((void *)&lock_info, 0, sizeof(lock_info));
3050 lock_info.l_type = F_WRLCK;
3051 lock_info.l_whence = SEEK_SET;
3052 lock_info.l_start = 0;
3053 lock_info.l_len = 1;
3054 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3055 (rc = ErrCode()) == EINTR) ;
3059 # ifdef MDB_USE_POSIX_SEM
3060 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3063 lock_info.l_type = F_RDLCK;
3064 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3065 (rc = ErrCode()) == EINTR) ;
3073 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3075 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3077 * @(#) $Revision: 5.1 $
3078 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3079 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3081 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3085 * Please do not copyright this code. This code is in the public domain.
3087 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3088 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3089 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3090 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3091 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3092 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3093 * PERFORMANCE OF THIS SOFTWARE.
3096 * chongo <Landon Curt Noll> /\oo/\
3097 * http://www.isthe.com/chongo/
3099 * Share and Enjoy! :-)
3102 typedef unsigned long long mdb_hash_t;
3103 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3105 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3106 * @param[in] str string to hash
3107 * @param[in] hval initial value for hash
3108 * @return 64 bit hash
3110 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3111 * hval arg on the first call.
3114 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3116 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3117 unsigned char *end = s + val->mv_size;
3119 * FNV-1a hash each octet of the string
3122 /* xor the bottom with the current octet */
3123 hval ^= (mdb_hash_t)*s++;
3125 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3126 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3127 (hval << 7) + (hval << 8) + (hval << 40);
3129 /* return our new hash value */
3133 /** Hash the string and output the hash in hex.
3134 * @param[in] str string to hash
3135 * @param[out] hexbuf an array of 17 chars to hold the hash
3138 mdb_hash_hex(MDB_val *val, char *hexbuf)
3141 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3142 for (i=0; i<8; i++) {
3143 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3149 /** Open and/or initialize the lock region for the environment.
3150 * @param[in] env The MDB environment.
3151 * @param[in] lpath The pathname of the file used for the lock region.
3152 * @param[in] mode The Unix permissions for the file, if we create it.
3153 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3154 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3155 * @return 0 on success, non-zero on failure.
3158 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3161 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3163 # define MDB_ERRCODE_ROFS EROFS
3164 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3165 # define MDB_CLOEXEC O_CLOEXEC
3168 # define MDB_CLOEXEC 0
3175 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3176 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3177 FILE_ATTRIBUTE_NORMAL, NULL);
3179 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3181 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3183 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3188 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3189 /* Lose record locks when exec*() */
3190 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3191 fcntl(env->me_lfd, F_SETFD, fdflags);
3194 if (!(env->me_flags & MDB_NOTLS)) {
3195 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3198 env->me_flags |= MDB_ENV_TXKEY;
3200 /* Windows TLS callbacks need help finding their TLS info. */
3201 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3205 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3209 /* Try to get exclusive lock. If we succeed, then
3210 * nobody is using the lock region and we should initialize it.
3212 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3215 size = GetFileSize(env->me_lfd, NULL);
3217 size = lseek(env->me_lfd, 0, SEEK_END);
3219 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3220 if (size < rsize && *excl > 0) {
3222 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3223 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3225 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3229 size = rsize - sizeof(MDB_txninfo);
3230 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3235 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3237 if (!mh) goto fail_errno;
3238 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3240 if (!env->me_txns) goto fail_errno;
3242 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3244 if (m == MAP_FAILED) goto fail_errno;
3250 BY_HANDLE_FILE_INFORMATION stbuf;
3259 if (!mdb_sec_inited) {
3260 InitializeSecurityDescriptor(&mdb_null_sd,
3261 SECURITY_DESCRIPTOR_REVISION);
3262 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3263 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3264 mdb_all_sa.bInheritHandle = FALSE;
3265 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3268 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3269 idbuf.volume = stbuf.dwVolumeSerialNumber;
3270 idbuf.nhigh = stbuf.nFileIndexHigh;
3271 idbuf.nlow = stbuf.nFileIndexLow;
3272 val.mv_data = &idbuf;
3273 val.mv_size = sizeof(idbuf);
3274 mdb_hash_hex(&val, hexbuf);
3275 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3276 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3277 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3278 if (!env->me_rmutex) goto fail_errno;
3279 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3280 if (!env->me_wmutex) goto fail_errno;
3281 #elif defined(MDB_USE_POSIX_SEM)
3290 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3291 idbuf.dev = stbuf.st_dev;
3292 idbuf.ino = stbuf.st_ino;
3293 val.mv_data = &idbuf;
3294 val.mv_size = sizeof(idbuf);
3295 mdb_hash_hex(&val, hexbuf);
3296 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3297 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3298 /* Clean up after a previous run, if needed: Try to
3299 * remove both semaphores before doing anything else.
3301 sem_unlink(env->me_txns->mti_rmname);
3302 sem_unlink(env->me_txns->mti_wmname);
3303 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3304 O_CREAT|O_EXCL, mode, 1);
3305 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3306 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3307 O_CREAT|O_EXCL, mode, 1);
3308 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3309 #else /* MDB_USE_POSIX_SEM */
3310 pthread_mutexattr_t mattr;
3312 if ((rc = pthread_mutexattr_init(&mattr))
3313 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3314 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3315 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3317 pthread_mutexattr_destroy(&mattr);
3318 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3320 env->me_txns->mti_version = MDB_VERSION;
3321 env->me_txns->mti_magic = MDB_MAGIC;
3322 env->me_txns->mti_txnid = 0;
3323 env->me_txns->mti_numreaders = 0;
3326 if (env->me_txns->mti_magic != MDB_MAGIC) {
3327 DPUTS("lock region has invalid magic");
3331 if (env->me_txns->mti_version != MDB_VERSION) {
3332 DPRINTF("lock region is version %u, expected version %u",
3333 env->me_txns->mti_version, MDB_VERSION);
3334 rc = MDB_VERSION_MISMATCH;
3338 if (rc != EACCES && rc != EAGAIN) {
3342 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3343 if (!env->me_rmutex) goto fail_errno;
3344 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3345 if (!env->me_wmutex) goto fail_errno;
3346 #elif defined(MDB_USE_POSIX_SEM)
3347 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3348 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3349 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3350 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3361 /** The name of the lock file in the DB environment */
3362 #define LOCKNAME "/lock.mdb"
3363 /** The name of the data file in the DB environment */
3364 #define DATANAME "/data.mdb"
3365 /** The suffix of the lock file when no subdir is used */
3366 #define LOCKSUFF "-lock"
3367 /** Only a subset of the @ref mdb_env flags can be changed
3368 * at runtime. Changing other flags requires closing the
3369 * environment and re-opening it with the new flags.
3371 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3372 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3375 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3377 int oflags, rc, len, excl = -1;
3378 char *lpath, *dpath;
3380 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3384 if (flags & MDB_NOSUBDIR) {
3385 rc = len + sizeof(LOCKSUFF) + len + 1;
3387 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3392 if (flags & MDB_NOSUBDIR) {
3393 dpath = lpath + len + sizeof(LOCKSUFF);
3394 sprintf(lpath, "%s" LOCKSUFF, path);
3395 strcpy(dpath, path);
3397 dpath = lpath + len + sizeof(LOCKNAME);
3398 sprintf(lpath, "%s" LOCKNAME, path);
3399 sprintf(dpath, "%s" DATANAME, path);
3403 flags |= env->me_flags;
3404 if (flags & MDB_RDONLY) {
3405 /* silently ignore WRITEMAP when we're only getting read access */
3406 flags &= ~MDB_WRITEMAP;
3408 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3409 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3412 env->me_flags = flags |= MDB_ENV_ACTIVE;
3416 env->me_path = strdup(path);
3417 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3418 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3419 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3424 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3429 if (F_ISSET(flags, MDB_RDONLY)) {
3430 oflags = GENERIC_READ;
3431 len = OPEN_EXISTING;
3433 oflags = GENERIC_READ|GENERIC_WRITE;
3436 mode = FILE_ATTRIBUTE_NORMAL;
3437 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3438 NULL, len, mode, NULL);
3440 if (F_ISSET(flags, MDB_RDONLY))
3443 oflags = O_RDWR | O_CREAT;
3445 env->me_fd = open(dpath, oflags, mode);
3447 if (env->me_fd == INVALID_HANDLE_VALUE) {
3452 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3453 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3454 env->me_mfd = env->me_fd;
3456 /* Synchronous fd for meta writes. Needed even with
3457 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3460 env->me_mfd = CreateFile(dpath, oflags,
3461 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3462 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3464 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3466 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3471 DPRINTF("opened dbenv %p", (void *) env);
3473 rc = mdb_env_share_locks(env, &excl);
3479 mdb_env_close0(env, excl);
3485 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3487 mdb_env_close0(MDB_env *env, int excl)
3491 if (!(env->me_flags & MDB_ENV_ACTIVE))
3494 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3495 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3496 free(env->me_dbxs[i].md_name.mv_data);
3498 free(env->me_dbflags);
3501 free(env->me_dirty_list);
3502 if (env->me_free_pgs)
3503 mdb_midl_free(env->me_free_pgs);
3505 if (env->me_flags & MDB_ENV_TXKEY) {
3506 pthread_key_delete(env->me_txkey);
3508 /* Delete our key from the global list */
3509 for (i=0; i<mdb_tls_nkeys; i++)
3510 if (mdb_tls_keys[i] == env->me_txkey) {
3511 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3519 munmap(env->me_map, env->me_mapsize);
3521 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3523 if (env->me_fd != INVALID_HANDLE_VALUE)
3526 pid_t pid = env->me_pid;
3527 /* Clearing readers is done in this function because
3528 * me_txkey with its destructor must be disabled first.
3530 for (i = env->me_numreaders; --i >= 0; )
3531 if (env->me_txns->mti_readers[i].mr_pid == pid)
3532 env->me_txns->mti_readers[i].mr_pid = 0;
3534 if (env->me_rmutex) {
3535 CloseHandle(env->me_rmutex);
3536 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3538 /* Windows automatically destroys the mutexes when
3539 * the last handle closes.
3541 #elif defined(MDB_USE_POSIX_SEM)
3542 if (env->me_rmutex != SEM_FAILED) {
3543 sem_close(env->me_rmutex);
3544 if (env->me_wmutex != SEM_FAILED)
3545 sem_close(env->me_wmutex);
3546 /* If we have the filelock: If we are the
3547 * only remaining user, clean up semaphores.
3550 mdb_env_excl_lock(env, &excl);
3552 sem_unlink(env->me_txns->mti_rmname);
3553 sem_unlink(env->me_txns->mti_wmname);
3557 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3559 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3562 /* Unlock the lockfile. Windows would have unlocked it
3563 * after closing anyway, but not necessarily at once.
3565 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3571 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3575 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3577 MDB_txn *txn = NULL;
3582 /* Do the lock/unlock of the reader mutex before starting the
3583 * write txn. Otherwise other read txns could block writers.
3585 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3590 /* We must start the actual read txn after blocking writers */
3591 mdb_txn_reset0(txn);
3593 /* Temporarily block writers until we snapshot the meta pages */
3596 rc = mdb_txn_renew0(txn);
3598 UNLOCK_MUTEX_W(env);
3603 wsize = env->me_psize * 2;
3607 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3608 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3611 rc = write(fd, env->me_map, wsize);
3612 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3615 UNLOCK_MUTEX_W(env);
3620 ptr = env->me_map + wsize;
3621 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3622 #define MAX_WRITE 2147483648U
3626 if (wsize > MAX_WRITE)
3630 rc = WriteFile(fd, ptr, w2, &len, NULL);
3631 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3640 if (wsize > MAX_WRITE)
3644 wres = write(fd, ptr, w2);
3645 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3658 mdb_env_copy(MDB_env *env, const char *path)
3662 HANDLE newfd = INVALID_HANDLE_VALUE;
3664 if (env->me_flags & MDB_NOSUBDIR) {
3665 lpath = (char *)path;
3668 len += sizeof(DATANAME);
3669 lpath = malloc(len);
3672 sprintf(lpath, "%s" DATANAME, path);
3675 /* The destination path must exist, but the destination file must not.
3676 * We don't want the OS to cache the writes, since the source data is
3677 * already in the OS cache.
3680 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3681 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3683 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3689 if (!(env->me_flags & MDB_NOSUBDIR))
3691 if (newfd == INVALID_HANDLE_VALUE) {
3696 #ifdef F_NOCACHE /* __APPLE__ */
3697 rc = fcntl(newfd, F_NOCACHE, 1);
3704 rc = mdb_env_copyfd(env, newfd);
3707 if (newfd != INVALID_HANDLE_VALUE)
3714 mdb_env_close(MDB_env *env)
3721 VGMEMP_DESTROY(env);
3722 while ((dp = env->me_dpages) != NULL) {
3723 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3724 env->me_dpages = dp->mp_next;
3728 mdb_env_close0(env, 0);
3732 /** Compare two items pointing at aligned size_t's */
3734 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3736 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3737 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3740 /** Compare two items pointing at aligned int's */
3742 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3744 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3745 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3748 /** Compare two items pointing at ints of unknown alignment.
3749 * Nodes and keys are guaranteed to be 2-byte aligned.
3752 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3754 #if BYTE_ORDER == LITTLE_ENDIAN
3755 unsigned short *u, *c;
3758 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3759 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3762 } while(!x && u > (unsigned short *)a->mv_data);
3765 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3769 /** Compare two items lexically */
3771 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3778 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3784 diff = memcmp(a->mv_data, b->mv_data, len);
3785 return diff ? diff : len_diff<0 ? -1 : len_diff;
3788 /** Compare two items in reverse byte order */
3790 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3792 const unsigned char *p1, *p2, *p1_lim;
3796 p1_lim = (const unsigned char *)a->mv_data;
3797 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3798 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3800 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3806 while (p1 > p1_lim) {
3807 diff = *--p1 - *--p2;
3811 return len_diff<0 ? -1 : len_diff;
3814 /** Search for key within a page, using binary search.
3815 * Returns the smallest entry larger or equal to the key.
3816 * If exactp is non-null, stores whether the found entry was an exact match
3817 * in *exactp (1 or 0).
3818 * Updates the cursor index with the index of the found entry.
3819 * If no entry larger or equal to the key is found, returns NULL.
3822 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3824 unsigned int i = 0, nkeys;
3827 MDB_page *mp = mc->mc_pg[mc->mc_top];
3828 MDB_node *node = NULL;
3833 nkeys = NUMKEYS(mp);
3838 COPY_PGNO(pgno, mp->mp_pgno);
3839 DPRINTF("searching %u keys in %s %spage %zu",
3840 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3847 low = IS_LEAF(mp) ? 0 : 1;
3849 cmp = mc->mc_dbx->md_cmp;
3851 /* Branch pages have no data, so if using integer keys,
3852 * alignment is guaranteed. Use faster mdb_cmp_int.
3854 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3855 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3862 nodekey.mv_size = mc->mc_db->md_pad;
3863 node = NODEPTR(mp, 0); /* fake */
3864 while (low <= high) {
3865 i = (low + high) >> 1;
3866 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3867 rc = cmp(key, &nodekey);
3868 DPRINTF("found leaf index %u [%s], rc = %i",
3869 i, DKEY(&nodekey), rc);
3878 while (low <= high) {
3879 i = (low + high) >> 1;
3881 node = NODEPTR(mp, i);
3882 nodekey.mv_size = NODEKSZ(node);
3883 nodekey.mv_data = NODEKEY(node);
3885 rc = cmp(key, &nodekey);
3888 DPRINTF("found leaf index %u [%s], rc = %i",
3889 i, DKEY(&nodekey), rc);
3891 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3892 i, DKEY(&nodekey), NODEPGNO(node), rc);
3903 if (rc > 0) { /* Found entry is less than the key. */
3904 i++; /* Skip to get the smallest entry larger than key. */
3906 node = NODEPTR(mp, i);
3909 *exactp = (rc == 0);
3910 /* store the key index */
3911 mc->mc_ki[mc->mc_top] = i;
3913 /* There is no entry larger or equal to the key. */
3916 /* nodeptr is fake for LEAF2 */
3922 mdb_cursor_adjust(MDB_cursor *mc, func)
3926 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3927 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3934 /** Pop a page off the top of the cursor's stack. */
3936 mdb_cursor_pop(MDB_cursor *mc)
3939 #ifndef MDB_DEBUG_SKIP
3940 MDB_page *top = mc->mc_pg[mc->mc_top];
3946 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3947 mc->mc_dbi, (void *) mc);
3951 /** Push a page onto the top of the cursor's stack. */
3953 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3955 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3956 mc->mc_dbi, (void *) mc);
3958 if (mc->mc_snum >= CURSOR_STACK) {
3959 assert(mc->mc_snum < CURSOR_STACK);
3960 return MDB_CURSOR_FULL;
3963 mc->mc_top = mc->mc_snum++;
3964 mc->mc_pg[mc->mc_top] = mp;
3965 mc->mc_ki[mc->mc_top] = 0;
3970 /** Find the address of the page corresponding to a given page number.
3971 * @param[in] txn the transaction for this access.
3972 * @param[in] pgno the page number for the page to retrieve.
3973 * @param[out] ret address of a pointer where the page's address will be stored.
3974 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
3975 * @return 0 on success, non-zero on failure.
3978 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
3983 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3984 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3989 MDB_ID2L dl = tx2->mt_u.dirty_list;
3991 unsigned x = mdb_mid2l_search(dl, pgno);
3992 if (x <= dl[0].mid && dl[x].mid == pgno) {
3998 } while ((tx2 = tx2->mt_parent) != NULL);
4001 if (pgno < txn->mt_next_pgno) {
4003 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4005 DPRINTF("page %zu not found", pgno);
4007 return MDB_PAGE_NOTFOUND;
4017 /** Search for the page a given key should be in.
4018 * Pushes parent pages on the cursor stack. This function continues a
4019 * search on a cursor that has already been initialized. (Usually by
4020 * #mdb_page_search() but also by #mdb_node_move().)
4021 * @param[in,out] mc the cursor for this operation.
4022 * @param[in] key the key to search for. If NULL, search for the lowest
4023 * page. (This is used by #mdb_cursor_first().)
4024 * @param[in] modify If true, visited pages are updated with new page numbers.
4025 * @return 0 on success, non-zero on failure.
4028 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4030 MDB_page *mp = mc->mc_pg[mc->mc_top];
4035 while (IS_BRANCH(mp)) {
4039 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4040 assert(NUMKEYS(mp) > 1);
4041 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4043 if (key == NULL) /* Initialize cursor to first page. */
4045 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4046 /* cursor to last page */
4050 node = mdb_node_search(mc, key, &exact);
4052 i = NUMKEYS(mp) - 1;
4054 i = mc->mc_ki[mc->mc_top];
4063 DPRINTF("following index %u for key [%s]",
4065 assert(i < NUMKEYS(mp));
4066 node = NODEPTR(mp, i);
4068 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4071 mc->mc_ki[mc->mc_top] = i;
4072 if ((rc = mdb_cursor_push(mc, mp)))
4076 if ((rc = mdb_page_touch(mc)) != 0)
4078 mp = mc->mc_pg[mc->mc_top];
4083 DPRINTF("internal error, index points to a %02X page!?",
4085 return MDB_CORRUPTED;
4088 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4089 key ? DKEY(key) : NULL);
4094 /** Search for the lowest key under the current branch page.
4095 * This just bypasses a NUMKEYS check in the current page
4096 * before calling mdb_page_search_root(), because the callers
4097 * are all in situations where the current page is known to
4101 mdb_page_search_lowest(MDB_cursor *mc)
4103 MDB_page *mp = mc->mc_pg[mc->mc_top];
4104 MDB_node *node = NODEPTR(mp, 0);
4107 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4110 mc->mc_ki[mc->mc_top] = 0;
4111 if ((rc = mdb_cursor_push(mc, mp)))
4113 return mdb_page_search_root(mc, NULL, 0);
4116 /** Search for the page a given key should be in.
4117 * Pushes parent pages on the cursor stack. This function just sets up
4118 * the search; it finds the root page for \b mc's database and sets this
4119 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4120 * called to complete the search.
4121 * @param[in,out] mc the cursor for this operation.
4122 * @param[in] key the key to search for. If NULL, search for the lowest
4123 * page. (This is used by #mdb_cursor_first().)
4124 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4125 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4126 * @return 0 on success, non-zero on failure.
4129 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4134 /* Make sure the txn is still viable, then find the root from
4135 * the txn's db table.
4137 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4138 DPUTS("transaction has failed, must abort");
4141 /* Make sure we're using an up-to-date root */
4142 if (mc->mc_dbi > MAIN_DBI) {
4143 if ((*mc->mc_dbflag & DB_STALE) ||
4144 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4146 unsigned char dbflag = 0;
4147 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4148 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4151 if (*mc->mc_dbflag & DB_STALE) {
4155 MDB_node *leaf = mdb_node_search(&mc2,
4156 &mc->mc_dbx->md_name, &exact);
4158 return MDB_NOTFOUND;
4159 mdb_node_read(mc->mc_txn, leaf, &data);
4160 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4162 /* The txn may not know this DBI, or another process may
4163 * have dropped and recreated the DB with other flags.
4165 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4166 return MDB_INCOMPATIBLE;
4167 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4169 if (flags & MDB_PS_MODIFY)
4171 *mc->mc_dbflag &= ~DB_STALE;
4172 *mc->mc_dbflag |= dbflag;
4175 root = mc->mc_db->md_root;
4177 if (root == P_INVALID) { /* Tree is empty. */
4178 DPUTS("tree is empty");
4179 return MDB_NOTFOUND;
4184 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4185 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4191 DPRINTF("db %u root page %zu has flags 0x%X",
4192 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4194 if (flags & MDB_PS_MODIFY) {
4195 if ((rc = mdb_page_touch(mc)))
4199 if (flags & MDB_PS_ROOTONLY)
4202 return mdb_page_search_root(mc, key, flags);
4206 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4208 MDB_txn *txn = mc->mc_txn;
4209 pgno_t pg = mp->mp_pgno;
4210 unsigned i, ovpages = mp->mp_pages;
4213 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4214 /* If the page is dirty we just acquired it, so we should
4215 * give it back to our current free list, if any.
4216 * Not currently supported in nested txns.
4217 * Otherwise put it onto the list of pages we freed in this txn.
4219 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && txn->mt_env->me_pghead) {
4221 pgno_t *mop = txn->mt_env->me_pghead;
4222 MDB_ID2 *dl, ix, iy;
4223 /* Prepare to insert pg */
4224 j = mop[0] + ovpages;
4226 rc = mdb_midl_grow(&mop, ovpages);
4229 txn->mt_env->me_pghead = mop;
4231 /* Remove from dirty list */
4232 dl = txn->mt_u.dirty_list;
4234 for (ix = dl[x]; ix.mid != pg; ix = iy) {
4241 return MDB_CORRUPTED;
4244 /* Insert in me_pghead */
4245 for (i = mop[0]; i && mop[i] < pg; i--)
4251 for (i=0; i<ovpages; i++) {
4252 mdb_midl_append(&txn->mt_free_pgs, pg);
4256 mc->mc_db->md_overflow_pages -= ovpages;
4260 /** Return the data associated with a given node.
4261 * @param[in] txn The transaction for this operation.
4262 * @param[in] leaf The node being read.
4263 * @param[out] data Updated to point to the node's data.
4264 * @return 0 on success, non-zero on failure.
4267 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4269 MDB_page *omp; /* overflow page */
4273 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4274 data->mv_size = NODEDSZ(leaf);
4275 data->mv_data = NODEDATA(leaf);
4279 /* Read overflow data.
4281 data->mv_size = NODEDSZ(leaf);
4282 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4283 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4284 DPRINTF("read overflow page %zu failed", pgno);
4287 data->mv_data = METADATA(omp);
4293 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4294 MDB_val *key, MDB_val *data)
4303 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4305 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4308 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4312 mdb_cursor_init(&mc, txn, dbi, &mx);
4313 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4316 /** Find a sibling for a page.
4317 * Replaces the page at the top of the cursor's stack with the
4318 * specified sibling, if one exists.
4319 * @param[in] mc The cursor for this operation.
4320 * @param[in] move_right Non-zero if the right sibling is requested,
4321 * otherwise the left sibling.
4322 * @return 0 on success, non-zero on failure.
4325 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4331 if (mc->mc_snum < 2) {
4332 return MDB_NOTFOUND; /* root has no siblings */
4336 DPRINTF("parent page is page %zu, index %u",
4337 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4339 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4340 : (mc->mc_ki[mc->mc_top] == 0)) {
4341 DPRINTF("no more keys left, moving to %s sibling",
4342 move_right ? "right" : "left");
4343 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4344 /* undo cursor_pop before returning */
4351 mc->mc_ki[mc->mc_top]++;
4353 mc->mc_ki[mc->mc_top]--;
4354 DPRINTF("just moving to %s index key %u",
4355 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4357 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4359 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4360 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4363 mdb_cursor_push(mc, mp);
4365 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4370 /** Move the cursor to the next data item. */
4372 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4378 if (mc->mc_flags & C_EOF) {
4379 return MDB_NOTFOUND;
4382 assert(mc->mc_flags & C_INITIALIZED);
4384 mp = mc->mc_pg[mc->mc_top];
4386 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4387 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4388 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4389 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4390 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4391 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4395 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4396 if (op == MDB_NEXT_DUP)
4397 return MDB_NOTFOUND;
4401 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4403 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4404 DPUTS("=====> move to next sibling page");
4405 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4406 mc->mc_flags |= C_EOF;
4407 mc->mc_flags &= ~C_INITIALIZED;
4408 return MDB_NOTFOUND;
4410 mp = mc->mc_pg[mc->mc_top];
4411 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4413 mc->mc_ki[mc->mc_top]++;
4415 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4416 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4419 key->mv_size = mc->mc_db->md_pad;
4420 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4424 assert(IS_LEAF(mp));
4425 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4427 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4428 mdb_xcursor_init1(mc, leaf);
4431 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4434 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4435 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4436 if (rc != MDB_SUCCESS)
4441 MDB_GET_KEY(leaf, key);
4445 /** Move the cursor to the previous data item. */
4447 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4453 assert(mc->mc_flags & C_INITIALIZED);
4455 mp = mc->mc_pg[mc->mc_top];
4457 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4458 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4459 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4460 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4461 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4462 if (op != MDB_PREV || rc == MDB_SUCCESS)
4465 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4466 if (op == MDB_PREV_DUP)
4467 return MDB_NOTFOUND;
4472 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4474 if (mc->mc_ki[mc->mc_top] == 0) {
4475 DPUTS("=====> move to prev sibling page");
4476 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4477 mc->mc_flags &= ~C_INITIALIZED;
4478 return MDB_NOTFOUND;
4480 mp = mc->mc_pg[mc->mc_top];
4481 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4482 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4484 mc->mc_ki[mc->mc_top]--;
4486 mc->mc_flags &= ~C_EOF;
4488 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4489 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4492 key->mv_size = mc->mc_db->md_pad;
4493 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4497 assert(IS_LEAF(mp));
4498 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4500 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4501 mdb_xcursor_init1(mc, leaf);
4504 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4507 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4508 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4509 if (rc != MDB_SUCCESS)
4514 MDB_GET_KEY(leaf, key);
4518 /** Set the cursor on a specific data item. */
4520 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4521 MDB_cursor_op op, int *exactp)
4525 MDB_node *leaf = NULL;
4530 assert(key->mv_size > 0);
4532 /* See if we're already on the right page */
4533 if (mc->mc_flags & C_INITIALIZED) {
4536 mp = mc->mc_pg[mc->mc_top];
4538 mc->mc_ki[mc->mc_top] = 0;
4539 return MDB_NOTFOUND;
4541 if (mp->mp_flags & P_LEAF2) {
4542 nodekey.mv_size = mc->mc_db->md_pad;
4543 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4545 leaf = NODEPTR(mp, 0);
4546 MDB_GET_KEY(leaf, &nodekey);
4548 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4550 /* Probably happens rarely, but first node on the page
4551 * was the one we wanted.
4553 mc->mc_ki[mc->mc_top] = 0;
4560 unsigned int nkeys = NUMKEYS(mp);
4562 if (mp->mp_flags & P_LEAF2) {
4563 nodekey.mv_data = LEAF2KEY(mp,
4564 nkeys-1, nodekey.mv_size);
4566 leaf = NODEPTR(mp, nkeys-1);
4567 MDB_GET_KEY(leaf, &nodekey);
4569 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4571 /* last node was the one we wanted */
4572 mc->mc_ki[mc->mc_top] = nkeys-1;
4578 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4579 /* This is definitely the right page, skip search_page */
4580 if (mp->mp_flags & P_LEAF2) {
4581 nodekey.mv_data = LEAF2KEY(mp,
4582 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4584 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4585 MDB_GET_KEY(leaf, &nodekey);
4587 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4589 /* current node was the one we wanted */
4599 /* If any parents have right-sibs, search.
4600 * Otherwise, there's nothing further.
4602 for (i=0; i<mc->mc_top; i++)
4604 NUMKEYS(mc->mc_pg[i])-1)
4606 if (i == mc->mc_top) {
4607 /* There are no other pages */
4608 mc->mc_ki[mc->mc_top] = nkeys;
4609 return MDB_NOTFOUND;
4613 /* There are no other pages */
4614 mc->mc_ki[mc->mc_top] = 0;
4615 return MDB_NOTFOUND;
4619 rc = mdb_page_search(mc, key, 0);
4620 if (rc != MDB_SUCCESS)
4623 mp = mc->mc_pg[mc->mc_top];
4624 assert(IS_LEAF(mp));
4627 leaf = mdb_node_search(mc, key, exactp);
4628 if (exactp != NULL && !*exactp) {
4629 /* MDB_SET specified and not an exact match. */
4630 return MDB_NOTFOUND;
4634 DPUTS("===> inexact leaf not found, goto sibling");
4635 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4636 return rc; /* no entries matched */
4637 mp = mc->mc_pg[mc->mc_top];
4638 assert(IS_LEAF(mp));
4639 leaf = NODEPTR(mp, 0);
4643 mc->mc_flags |= C_INITIALIZED;
4644 mc->mc_flags &= ~C_EOF;
4647 key->mv_size = mc->mc_db->md_pad;
4648 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4652 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4653 mdb_xcursor_init1(mc, leaf);
4656 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4657 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4658 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4661 if (op == MDB_GET_BOTH) {
4667 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4668 if (rc != MDB_SUCCESS)
4671 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4673 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4675 rc = mc->mc_dbx->md_dcmp(data, &d2);
4677 if (op == MDB_GET_BOTH || rc > 0)
4678 return MDB_NOTFOUND;
4683 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4684 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4689 /* The key already matches in all other cases */
4690 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4691 MDB_GET_KEY(leaf, key);
4692 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4697 /** Move the cursor to the first item in the database. */
4699 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4704 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4705 rc = mdb_page_search(mc, NULL, 0);
4706 if (rc != MDB_SUCCESS)
4709 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4711 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4712 mc->mc_flags |= C_INITIALIZED;
4713 mc->mc_flags &= ~C_EOF;
4715 mc->mc_ki[mc->mc_top] = 0;
4717 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4718 key->mv_size = mc->mc_db->md_pad;
4719 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4724 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4725 mdb_xcursor_init1(mc, leaf);
4726 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4731 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4732 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4736 MDB_GET_KEY(leaf, key);
4740 /** Move the cursor to the last item in the database. */
4742 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4747 if (!(mc->mc_flags & C_EOF)) {
4749 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4752 lkey.mv_size = MDB_MAXKEYSIZE+1;
4753 lkey.mv_data = NULL;
4754 rc = mdb_page_search(mc, &lkey, 0);
4755 if (rc != MDB_SUCCESS)
4758 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4761 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4762 mc->mc_flags |= C_INITIALIZED|C_EOF;
4763 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4765 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4766 key->mv_size = mc->mc_db->md_pad;
4767 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4772 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4773 mdb_xcursor_init1(mc, leaf);
4774 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4779 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4780 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4785 MDB_GET_KEY(leaf, key);
4790 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4799 case MDB_GET_CURRENT:
4800 if (!(mc->mc_flags & C_INITIALIZED)) {
4803 MDB_page *mp = mc->mc_pg[mc->mc_top];
4805 mc->mc_ki[mc->mc_top] = 0;
4811 key->mv_size = mc->mc_db->md_pad;
4812 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4814 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4815 MDB_GET_KEY(leaf, key);
4817 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4818 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4820 rc = mdb_node_read(mc->mc_txn, leaf, data);
4827 case MDB_GET_BOTH_RANGE:
4828 if (data == NULL || mc->mc_xcursor == NULL) {
4836 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4838 } else if (op == MDB_SET_RANGE)
4839 rc = mdb_cursor_set(mc, key, data, op, NULL);
4841 rc = mdb_cursor_set(mc, key, data, op, &exact);
4843 case MDB_GET_MULTIPLE:
4845 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4846 !(mc->mc_flags & C_INITIALIZED)) {
4851 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4852 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4855 case MDB_NEXT_MULTIPLE:
4857 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4861 if (!(mc->mc_flags & C_INITIALIZED))
4862 rc = mdb_cursor_first(mc, key, data);
4864 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4865 if (rc == MDB_SUCCESS) {
4866 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4869 mx = &mc->mc_xcursor->mx_cursor;
4870 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4872 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4873 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4881 case MDB_NEXT_NODUP:
4882 if (!(mc->mc_flags & C_INITIALIZED))
4883 rc = mdb_cursor_first(mc, key, data);
4885 rc = mdb_cursor_next(mc, key, data, op);
4889 case MDB_PREV_NODUP:
4890 if (!(mc->mc_flags & C_INITIALIZED)) {
4891 rc = mdb_cursor_last(mc, key, data);
4894 mc->mc_flags |= C_INITIALIZED;
4895 mc->mc_ki[mc->mc_top]++;
4897 rc = mdb_cursor_prev(mc, key, data, op);
4900 rc = mdb_cursor_first(mc, key, data);
4904 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4905 !(mc->mc_flags & C_INITIALIZED) ||
4906 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4910 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4913 rc = mdb_cursor_last(mc, key, data);
4917 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4918 !(mc->mc_flags & C_INITIALIZED) ||
4919 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4923 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4926 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4934 /** Touch all the pages in the cursor stack.
4935 * Makes sure all the pages are writable, before attempting a write operation.
4936 * @param[in] mc The cursor to operate on.
4939 mdb_cursor_touch(MDB_cursor *mc)
4943 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4946 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
4947 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4950 *mc->mc_dbflag |= DB_DIRTY;
4952 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4953 rc = mdb_page_touch(mc);
4957 mc->mc_top = mc->mc_snum-1;
4962 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4965 MDB_node *leaf = NULL;
4966 MDB_val xdata, *rdata, dkey;
4969 int do_sub = 0, insert = 0;
4970 unsigned int mcount = 0;
4974 char dbuf[MDB_MAXKEYSIZE+1];
4975 unsigned int nflags;
4978 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4981 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4984 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4987 #if SIZE_MAX > MAXDATASIZE
4988 if (data->mv_size > MAXDATASIZE)
4992 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4993 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4997 if (flags == MDB_CURRENT) {
4998 if (!(mc->mc_flags & C_INITIALIZED))
5001 } else if (mc->mc_db->md_root == P_INVALID) {
5003 /* new database, write a root leaf page */
5004 DPUTS("allocating new root leaf page");
5005 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5009 mdb_cursor_push(mc, np);
5010 mc->mc_db->md_root = np->mp_pgno;
5011 mc->mc_db->md_depth++;
5012 *mc->mc_dbflag |= DB_DIRTY;
5013 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5015 np->mp_flags |= P_LEAF2;
5016 mc->mc_flags |= C_INITIALIZED;
5022 if (flags & MDB_APPEND) {
5024 rc = mdb_cursor_last(mc, &k2, &d2);
5026 rc = mc->mc_dbx->md_cmp(key, &k2);
5029 mc->mc_ki[mc->mc_top]++;
5031 /* new key is <= last key */
5036 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5038 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5039 DPRINTF("duplicate key [%s]", DKEY(key));
5041 return MDB_KEYEXIST;
5043 if (rc && rc != MDB_NOTFOUND)
5047 /* Cursor is positioned, now make sure all pages are writable */
5048 rc2 = mdb_cursor_touch(mc);
5053 /* The key already exists */
5054 if (rc == MDB_SUCCESS) {
5055 /* there's only a key anyway, so this is a no-op */
5056 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5057 unsigned int ksize = mc->mc_db->md_pad;
5058 if (key->mv_size != ksize)
5060 if (flags == MDB_CURRENT) {
5061 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5062 memcpy(ptr, key->mv_data, ksize);
5067 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5070 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5071 /* Was a single item before, must convert now */
5073 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5074 /* Just overwrite the current item */
5075 if (flags == MDB_CURRENT)
5078 dkey.mv_size = NODEDSZ(leaf);
5079 dkey.mv_data = NODEDATA(leaf);
5080 #if UINT_MAX < SIZE_MAX
5081 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5082 #ifdef MISALIGNED_OK
5083 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5085 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5088 /* if data matches, ignore it */
5089 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5090 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5092 /* create a fake page for the dup items */
5093 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5094 dkey.mv_data = dbuf;
5095 fp = (MDB_page *)&pbuf;
5096 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5097 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5098 fp->mp_lower = PAGEHDRSZ;
5099 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5100 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5101 fp->mp_flags |= P_LEAF2;
5102 fp->mp_pad = data->mv_size;
5103 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5105 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5106 (dkey.mv_size & 1) + (data->mv_size & 1);
5108 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5111 xdata.mv_size = fp->mp_upper;
5116 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5117 /* See if we need to convert from fake page to subDB */
5119 unsigned int offset;
5123 fp = NODEDATA(leaf);
5124 if (flags == MDB_CURRENT) {
5126 fp->mp_flags |= P_DIRTY;
5127 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5128 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5132 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5133 offset = fp->mp_pad;
5134 if (SIZELEFT(fp) >= offset)
5136 offset *= 4; /* space for 4 more */
5138 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5140 offset += offset & 1;
5141 fp_flags = fp->mp_flags;
5142 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5143 offset >= mc->mc_txn->mt_env->me_nodemax) {
5144 /* yes, convert it */
5146 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5147 dummy.md_pad = fp->mp_pad;
5148 dummy.md_flags = MDB_DUPFIXED;
5149 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5150 dummy.md_flags |= MDB_INTEGERKEY;
5153 dummy.md_branch_pages = 0;
5154 dummy.md_leaf_pages = 1;
5155 dummy.md_overflow_pages = 0;
5156 dummy.md_entries = NUMKEYS(fp);
5158 xdata.mv_size = sizeof(MDB_db);
5159 xdata.mv_data = &dummy;
5160 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5162 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5163 flags |= F_DUPDATA|F_SUBDATA;
5164 dummy.md_root = mp->mp_pgno;
5165 fp_flags &= ~P_SUBP;
5167 /* no, just grow it */
5169 xdata.mv_size = NODEDSZ(leaf) + offset;
5170 xdata.mv_data = &pbuf;
5171 mp = (MDB_page *)&pbuf;
5172 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5175 mp->mp_flags = fp_flags | P_DIRTY;
5176 mp->mp_pad = fp->mp_pad;
5177 mp->mp_lower = fp->mp_lower;
5178 mp->mp_upper = fp->mp_upper + offset;
5180 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5182 nsize = NODEDSZ(leaf) - fp->mp_upper;
5183 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5184 for (i=0; i<NUMKEYS(fp); i++)
5185 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5187 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5191 /* data is on sub-DB, just store it */
5192 flags |= F_DUPDATA|F_SUBDATA;
5196 /* overflow page overwrites need special handling */
5197 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5200 unsigned psize = mc->mc_txn->mt_env->me_psize;
5201 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5203 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5204 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5206 ovpages = omp->mp_pages;
5208 /* Is the ov page writable and large enough? */
5209 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5210 /* yes, overwrite it. Note in this case we don't
5211 * bother to try shrinking the page if the new data
5212 * is smaller than the overflow threshold.
5215 /* It is writable only in a parent txn */
5216 size_t sz = (size_t) psize * ovpages, off;
5217 MDB_page *np = mdb_page_malloc(mc, ovpages);
5223 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5224 if (!(flags & MDB_RESERVE)) {
5225 /* Copy end of page, adjusting alignment so
5226 * compiler may copy words instead of bytes.
5228 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5229 memcpy((size_t *)((char *)np + off),
5230 (size_t *)((char *)omp + off), sz - off);
5233 memcpy(np, omp, sz); /* Copy beginning of page */
5236 SETDSZ(leaf, data->mv_size);
5237 if (F_ISSET(flags, MDB_RESERVE))
5238 data->mv_data = METADATA(omp);
5240 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5243 mdb_ovpage_free(mc, omp);
5245 } else if (NODEDSZ(leaf) == data->mv_size) {
5246 /* same size, just replace it. Note that we could
5247 * also reuse this node if the new data is smaller,
5248 * but instead we opt to shrink the node in that case.
5250 if (F_ISSET(flags, MDB_RESERVE))
5251 data->mv_data = NODEDATA(leaf);
5252 else if (data->mv_size)
5253 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5255 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5258 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5259 mc->mc_db->md_entries--;
5261 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5268 nflags = flags & NODE_ADD_FLAGS;
5269 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5270 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5271 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5272 nflags &= ~MDB_APPEND;
5274 nflags |= MDB_SPLIT_REPLACE;
5275 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5277 /* There is room already in this leaf page. */
5278 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5279 if (rc == 0 && !do_sub && insert) {
5280 /* Adjust other cursors pointing to mp */
5281 MDB_cursor *m2, *m3;
5282 MDB_dbi dbi = mc->mc_dbi;
5283 unsigned i = mc->mc_top;
5284 MDB_page *mp = mc->mc_pg[i];
5286 if (mc->mc_flags & C_SUB)
5289 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5290 if (mc->mc_flags & C_SUB)
5291 m3 = &m2->mc_xcursor->mx_cursor;
5294 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5295 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5302 if (rc != MDB_SUCCESS)
5303 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5305 /* Now store the actual data in the child DB. Note that we're
5306 * storing the user data in the keys field, so there are strict
5307 * size limits on dupdata. The actual data fields of the child
5308 * DB are all zero size.
5315 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5316 if (flags & MDB_CURRENT) {
5317 xflags = MDB_CURRENT;
5319 mdb_xcursor_init1(mc, leaf);
5320 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5322 /* converted, write the original data first */
5324 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5328 /* Adjust other cursors pointing to mp */
5330 unsigned i = mc->mc_top;
5331 MDB_page *mp = mc->mc_pg[i];
5333 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5334 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5335 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5336 mdb_xcursor_init1(m2, leaf);
5340 /* we've done our job */
5343 if (flags & MDB_APPENDDUP)
5344 xflags |= MDB_APPEND;
5345 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5346 if (flags & F_SUBDATA) {
5347 void *db = NODEDATA(leaf);
5348 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5351 /* sub-writes might have failed so check rc again.
5352 * Don't increment count if we just replaced an existing item.
5354 if (!rc && !(flags & MDB_CURRENT))
5355 mc->mc_db->md_entries++;
5356 if (flags & MDB_MULTIPLE) {
5358 if (mcount < data[1].mv_size) {
5359 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5360 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5366 /* If we succeeded and the key didn't exist before, make sure
5367 * the cursor is marked valid.
5370 mc->mc_flags |= C_INITIALIZED;
5375 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5380 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5383 if (!(mc->mc_flags & C_INITIALIZED))
5386 rc = mdb_cursor_touch(mc);
5390 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5392 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5393 if (flags != MDB_NODUPDATA) {
5394 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5395 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5397 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5398 /* If sub-DB still has entries, we're done */
5399 if (mc->mc_xcursor->mx_db.md_entries) {
5400 if (leaf->mn_flags & F_SUBDATA) {
5401 /* update subDB info */
5402 void *db = NODEDATA(leaf);
5403 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5406 /* shrink fake page */
5407 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5408 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5409 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5410 /* fix other sub-DB cursors pointed at this fake page */
5411 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5412 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5413 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5414 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5415 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5418 mc->mc_db->md_entries--;
5421 /* otherwise fall thru and delete the sub-DB */
5424 if (leaf->mn_flags & F_SUBDATA) {
5425 /* add all the child DB's pages to the free list */
5426 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5427 if (rc == MDB_SUCCESS) {
5428 mc->mc_db->md_entries -=
5429 mc->mc_xcursor->mx_db.md_entries;
5434 return mdb_cursor_del0(mc, leaf);
5437 /** Allocate and initialize new pages for a database.
5438 * @param[in] mc a cursor on the database being added to.
5439 * @param[in] flags flags defining what type of page is being allocated.
5440 * @param[in] num the number of pages to allocate. This is usually 1,
5441 * unless allocating overflow pages for a large record.
5442 * @param[out] mp Address of a page, or NULL on failure.
5443 * @return 0 on success, non-zero on failure.
5446 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5451 if ((rc = mdb_page_alloc(mc, num, &np)))
5453 DPRINTF("allocated new mpage %zu, page size %u",
5454 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5455 np->mp_flags = flags | P_DIRTY;
5456 np->mp_lower = PAGEHDRSZ;
5457 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5460 mc->mc_db->md_branch_pages++;
5461 else if (IS_LEAF(np))
5462 mc->mc_db->md_leaf_pages++;
5463 else if (IS_OVERFLOW(np)) {
5464 mc->mc_db->md_overflow_pages += num;
5472 /** Calculate the size of a leaf node.
5473 * The size depends on the environment's page size; if a data item
5474 * is too large it will be put onto an overflow page and the node
5475 * size will only include the key and not the data. Sizes are always
5476 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5477 * of the #MDB_node headers.
5478 * @param[in] env The environment handle.
5479 * @param[in] key The key for the node.
5480 * @param[in] data The data for the node.
5481 * @return The number of bytes needed to store the node.
5484 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5488 sz = LEAFSIZE(key, data);
5489 if (sz >= env->me_nodemax) {
5490 /* put on overflow page */
5491 sz -= data->mv_size - sizeof(pgno_t);
5495 return sz + sizeof(indx_t);
5498 /** Calculate the size of a branch node.
5499 * The size should depend on the environment's page size but since
5500 * we currently don't support spilling large keys onto overflow
5501 * pages, it's simply the size of the #MDB_node header plus the
5502 * size of the key. Sizes are always rounded up to an even number
5503 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5504 * @param[in] env The environment handle.
5505 * @param[in] key The key for the node.
5506 * @return The number of bytes needed to store the node.
5509 mdb_branch_size(MDB_env *env, MDB_val *key)
5514 if (sz >= env->me_nodemax) {
5515 /* put on overflow page */
5516 /* not implemented */
5517 /* sz -= key->size - sizeof(pgno_t); */
5520 return sz + sizeof(indx_t);
5523 /** Add a node to the page pointed to by the cursor.
5524 * @param[in] mc The cursor for this operation.
5525 * @param[in] indx The index on the page where the new node should be added.
5526 * @param[in] key The key for the new node.
5527 * @param[in] data The data for the new node, if any.
5528 * @param[in] pgno The page number, if adding a branch node.
5529 * @param[in] flags Flags for the node.
5530 * @return 0 on success, non-zero on failure. Possible errors are:
5532 * <li>ENOMEM - failed to allocate overflow pages for the node.
5533 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5534 * should never happen since all callers already calculate the
5535 * page's free space before calling this function.
5539 mdb_node_add(MDB_cursor *mc, indx_t indx,
5540 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5543 size_t node_size = NODESIZE;
5546 MDB_page *mp = mc->mc_pg[mc->mc_top];
5547 MDB_page *ofp = NULL; /* overflow page */
5550 assert(mp->mp_upper >= mp->mp_lower);
5552 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5553 IS_LEAF(mp) ? "leaf" : "branch",
5554 IS_SUBP(mp) ? "sub-" : "",
5555 mp->mp_pgno, indx, data ? data->mv_size : 0,
5556 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5559 /* Move higher keys up one slot. */
5560 int ksize = mc->mc_db->md_pad, dif;
5561 char *ptr = LEAF2KEY(mp, indx, ksize);
5562 dif = NUMKEYS(mp) - indx;
5564 memmove(ptr+ksize, ptr, dif*ksize);
5565 /* insert new key */
5566 memcpy(ptr, key->mv_data, ksize);
5568 /* Just using these for counting */
5569 mp->mp_lower += sizeof(indx_t);
5570 mp->mp_upper -= ksize - sizeof(indx_t);
5575 node_size += key->mv_size;
5579 if (F_ISSET(flags, F_BIGDATA)) {
5580 /* Data already on overflow page. */
5581 node_size += sizeof(pgno_t);
5582 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5583 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5585 /* Put data on overflow page. */
5586 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5587 data->mv_size, node_size+data->mv_size);
5588 node_size += sizeof(pgno_t);
5589 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5591 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5594 node_size += data->mv_size;
5597 node_size += node_size & 1;
5599 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5600 DPRINTF("not enough room in page %zu, got %u ptrs",
5601 mp->mp_pgno, NUMKEYS(mp));
5602 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5603 mp->mp_upper - mp->mp_lower);
5604 DPRINTF("node size = %zu", node_size);
5605 return MDB_PAGE_FULL;
5608 /* Move higher pointers up one slot. */
5609 for (i = NUMKEYS(mp); i > indx; i--)
5610 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5612 /* Adjust free space offsets. */
5613 ofs = mp->mp_upper - node_size;
5614 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5615 mp->mp_ptrs[indx] = ofs;
5617 mp->mp_lower += sizeof(indx_t);
5619 /* Write the node data. */
5620 node = NODEPTR(mp, indx);
5621 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5622 node->mn_flags = flags;
5624 SETDSZ(node,data->mv_size);
5629 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5634 if (F_ISSET(flags, F_BIGDATA))
5635 memcpy(node->mn_data + key->mv_size, data->mv_data,
5637 else if (F_ISSET(flags, MDB_RESERVE))
5638 data->mv_data = node->mn_data + key->mv_size;
5640 memcpy(node->mn_data + key->mv_size, data->mv_data,
5643 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5645 if (F_ISSET(flags, MDB_RESERVE))
5646 data->mv_data = METADATA(ofp);
5648 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5655 /** Delete the specified node from a page.
5656 * @param[in] mp The page to operate on.
5657 * @param[in] indx The index of the node to delete.
5658 * @param[in] ksize The size of a node. Only used if the page is
5659 * part of a #MDB_DUPFIXED database.
5662 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5665 indx_t i, j, numkeys, ptr;
5672 COPY_PGNO(pgno, mp->mp_pgno);
5673 DPRINTF("delete node %u on %s page %zu", indx,
5674 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5677 assert(indx < NUMKEYS(mp));
5680 int x = NUMKEYS(mp) - 1 - indx;
5681 base = LEAF2KEY(mp, indx, ksize);
5683 memmove(base, base + ksize, x * ksize);
5684 mp->mp_lower -= sizeof(indx_t);
5685 mp->mp_upper += ksize - sizeof(indx_t);
5689 node = NODEPTR(mp, indx);
5690 sz = NODESIZE + node->mn_ksize;
5692 if (F_ISSET(node->mn_flags, F_BIGDATA))
5693 sz += sizeof(pgno_t);
5695 sz += NODEDSZ(node);
5699 ptr = mp->mp_ptrs[indx];
5700 numkeys = NUMKEYS(mp);
5701 for (i = j = 0; i < numkeys; i++) {
5703 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5704 if (mp->mp_ptrs[i] < ptr)
5705 mp->mp_ptrs[j] += sz;
5710 base = (char *)mp + mp->mp_upper;
5711 memmove(base + sz, base, ptr - mp->mp_upper);
5713 mp->mp_lower -= sizeof(indx_t);
5717 /** Compact the main page after deleting a node on a subpage.
5718 * @param[in] mp The main page to operate on.
5719 * @param[in] indx The index of the subpage on the main page.
5722 mdb_node_shrink(MDB_page *mp, indx_t indx)
5729 indx_t i, numkeys, ptr;
5731 node = NODEPTR(mp, indx);
5732 sp = (MDB_page *)NODEDATA(node);
5733 osize = NODEDSZ(node);
5735 delta = sp->mp_upper - sp->mp_lower;
5736 SETDSZ(node, osize - delta);
5737 xp = (MDB_page *)((char *)sp + delta);
5739 /* shift subpage upward */
5741 nsize = NUMKEYS(sp) * sp->mp_pad;
5742 memmove(METADATA(xp), METADATA(sp), nsize);
5745 nsize = osize - sp->mp_upper;
5746 numkeys = NUMKEYS(sp);
5747 for (i=numkeys-1; i>=0; i--)
5748 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5750 xp->mp_upper = sp->mp_lower;
5751 xp->mp_lower = sp->mp_lower;
5752 xp->mp_flags = sp->mp_flags;
5753 xp->mp_pad = sp->mp_pad;
5754 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5756 /* shift lower nodes upward */
5757 ptr = mp->mp_ptrs[indx];
5758 numkeys = NUMKEYS(mp);
5759 for (i = 0; i < numkeys; i++) {
5760 if (mp->mp_ptrs[i] <= ptr)
5761 mp->mp_ptrs[i] += delta;
5764 base = (char *)mp + mp->mp_upper;
5765 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5766 mp->mp_upper += delta;
5769 /** Initial setup of a sorted-dups cursor.
5770 * Sorted duplicates are implemented as a sub-database for the given key.
5771 * The duplicate data items are actually keys of the sub-database.
5772 * Operations on the duplicate data items are performed using a sub-cursor
5773 * initialized when the sub-database is first accessed. This function does
5774 * the preliminary setup of the sub-cursor, filling in the fields that
5775 * depend only on the parent DB.
5776 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5779 mdb_xcursor_init0(MDB_cursor *mc)
5781 MDB_xcursor *mx = mc->mc_xcursor;
5783 mx->mx_cursor.mc_xcursor = NULL;
5784 mx->mx_cursor.mc_txn = mc->mc_txn;
5785 mx->mx_cursor.mc_db = &mx->mx_db;
5786 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5787 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5788 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5789 mx->mx_cursor.mc_snum = 0;
5790 mx->mx_cursor.mc_top = 0;
5791 mx->mx_cursor.mc_flags = C_SUB;
5792 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5793 mx->mx_dbx.md_dcmp = NULL;
5794 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5797 /** Final setup of a sorted-dups cursor.
5798 * Sets up the fields that depend on the data from the main cursor.
5799 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5800 * @param[in] node The data containing the #MDB_db record for the
5801 * sorted-dup database.
5804 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5806 MDB_xcursor *mx = mc->mc_xcursor;
5808 if (node->mn_flags & F_SUBDATA) {
5809 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5810 mx->mx_cursor.mc_pg[0] = 0;
5811 mx->mx_cursor.mc_snum = 0;
5812 mx->mx_cursor.mc_flags = C_SUB;
5814 MDB_page *fp = NODEDATA(node);
5815 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5816 mx->mx_db.md_flags = 0;
5817 mx->mx_db.md_depth = 1;
5818 mx->mx_db.md_branch_pages = 0;
5819 mx->mx_db.md_leaf_pages = 1;
5820 mx->mx_db.md_overflow_pages = 0;
5821 mx->mx_db.md_entries = NUMKEYS(fp);
5822 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5823 mx->mx_cursor.mc_snum = 1;
5824 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5825 mx->mx_cursor.mc_top = 0;
5826 mx->mx_cursor.mc_pg[0] = fp;
5827 mx->mx_cursor.mc_ki[0] = 0;
5828 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5829 mx->mx_db.md_flags = MDB_DUPFIXED;
5830 mx->mx_db.md_pad = fp->mp_pad;
5831 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5832 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5835 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5837 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5839 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5840 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5841 #if UINT_MAX < SIZE_MAX
5842 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5843 #ifdef MISALIGNED_OK
5844 mx->mx_dbx.md_cmp = mdb_cmp_long;
5846 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5851 /** Initialize a cursor for a given transaction and database. */
5853 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5858 mc->mc_db = &txn->mt_dbs[dbi];
5859 mc->mc_dbx = &txn->mt_dbxs[dbi];
5860 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5865 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5867 mc->mc_xcursor = mx;
5868 mdb_xcursor_init0(mc);
5870 mc->mc_xcursor = NULL;
5872 if (*mc->mc_dbflag & DB_STALE) {
5873 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5878 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5881 size_t size = sizeof(MDB_cursor);
5883 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5886 /* Allow read access to the freelist */
5887 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5890 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5891 size += sizeof(MDB_xcursor);
5893 if ((mc = malloc(size)) != NULL) {
5894 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5895 if (txn->mt_cursors) {
5896 mc->mc_next = txn->mt_cursors[dbi];
5897 txn->mt_cursors[dbi] = mc;
5898 mc->mc_flags |= C_UNTRACK;
5900 mc->mc_flags |= C_ALLOCD;
5911 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5915 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5918 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5921 flags = mc->mc_flags;
5923 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5925 mc->mc_flags |= (flags & C_ALLOCD);
5929 /* Return the count of duplicate data items for the current key */
5931 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5935 if (mc == NULL || countp == NULL)
5938 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5941 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5942 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5945 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5948 *countp = mc->mc_xcursor->mx_db.md_entries;
5954 mdb_cursor_close(MDB_cursor *mc)
5957 /* remove from txn, if tracked */
5958 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
5959 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5960 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5962 *prev = mc->mc_next;
5964 if (mc->mc_flags & C_ALLOCD)
5970 mdb_cursor_txn(MDB_cursor *mc)
5972 if (!mc) return NULL;
5977 mdb_cursor_dbi(MDB_cursor *mc)
5983 /** Replace the key for a node with a new key.
5984 * @param[in] mc Cursor pointing to the node to operate on.
5985 * @param[in] key The new key to use.
5986 * @return 0 on success, non-zero on failure.
5989 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5996 indx_t ptr, i, numkeys, indx;
5999 indx = mc->mc_ki[mc->mc_top];
6000 mp = mc->mc_pg[mc->mc_top];
6001 node = NODEPTR(mp, indx);
6002 ptr = mp->mp_ptrs[indx];
6006 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6007 k2.mv_data = NODEKEY(node);
6008 k2.mv_size = node->mn_ksize;
6009 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6011 mdb_dkey(&k2, kbuf2),
6017 delta0 = delta = key->mv_size - node->mn_ksize;
6019 /* Must be 2-byte aligned. If new key is
6020 * shorter by 1, the shift will be skipped.
6022 delta += (delta & 1);
6024 if (delta > 0 && SIZELEFT(mp) < delta) {
6026 /* not enough space left, do a delete and split */
6027 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6028 pgno = NODEPGNO(node);
6029 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6030 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6033 numkeys = NUMKEYS(mp);
6034 for (i = 0; i < numkeys; i++) {
6035 if (mp->mp_ptrs[i] <= ptr)
6036 mp->mp_ptrs[i] -= delta;
6039 base = (char *)mp + mp->mp_upper;
6040 len = ptr - mp->mp_upper + NODESIZE;
6041 memmove(base - delta, base, len);
6042 mp->mp_upper -= delta;
6044 node = NODEPTR(mp, indx);
6047 /* But even if no shift was needed, update ksize */
6049 node->mn_ksize = key->mv_size;
6052 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6058 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6060 /** Move a node from csrc to cdst.
6063 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6070 unsigned short flags;
6074 /* Mark src and dst as dirty. */
6075 if ((rc = mdb_page_touch(csrc)) ||
6076 (rc = mdb_page_touch(cdst)))
6079 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6080 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6081 key.mv_size = csrc->mc_db->md_pad;
6082 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6084 data.mv_data = NULL;
6088 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6089 assert(!((long)srcnode&1));
6090 srcpg = NODEPGNO(srcnode);
6091 flags = srcnode->mn_flags;
6092 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6093 unsigned int snum = csrc->mc_snum;
6095 /* must find the lowest key below src */
6096 mdb_page_search_lowest(csrc);
6097 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6098 key.mv_size = csrc->mc_db->md_pad;
6099 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6101 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6102 key.mv_size = NODEKSZ(s2);
6103 key.mv_data = NODEKEY(s2);
6105 csrc->mc_snum = snum--;
6106 csrc->mc_top = snum;
6108 key.mv_size = NODEKSZ(srcnode);
6109 key.mv_data = NODEKEY(srcnode);
6111 data.mv_size = NODEDSZ(srcnode);
6112 data.mv_data = NODEDATA(srcnode);
6114 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6115 unsigned int snum = cdst->mc_snum;
6118 /* must find the lowest key below dst */
6119 mdb_page_search_lowest(cdst);
6120 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6121 bkey.mv_size = cdst->mc_db->md_pad;
6122 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6124 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6125 bkey.mv_size = NODEKSZ(s2);
6126 bkey.mv_data = NODEKEY(s2);
6128 cdst->mc_snum = snum--;
6129 cdst->mc_top = snum;
6130 mdb_cursor_copy(cdst, &mn);
6132 rc = mdb_update_key(&mn, &bkey);
6137 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6138 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6139 csrc->mc_ki[csrc->mc_top],
6141 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6142 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6144 /* Add the node to the destination page.
6146 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6147 if (rc != MDB_SUCCESS)
6150 /* Delete the node from the source page.
6152 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6155 /* Adjust other cursors pointing to mp */
6156 MDB_cursor *m2, *m3;
6157 MDB_dbi dbi = csrc->mc_dbi;
6158 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6160 if (csrc->mc_flags & C_SUB)
6163 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6164 if (csrc->mc_flags & C_SUB)
6165 m3 = &m2->mc_xcursor->mx_cursor;
6168 if (m3 == csrc) continue;
6169 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6170 csrc->mc_ki[csrc->mc_top]) {
6171 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6172 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6177 /* Update the parent separators.
6179 if (csrc->mc_ki[csrc->mc_top] == 0) {
6180 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6181 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6182 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6184 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6185 key.mv_size = NODEKSZ(srcnode);
6186 key.mv_data = NODEKEY(srcnode);
6188 DPRINTF("update separator for source page %zu to [%s]",
6189 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6190 mdb_cursor_copy(csrc, &mn);
6193 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6196 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6198 indx_t ix = csrc->mc_ki[csrc->mc_top];
6199 nullkey.mv_size = 0;
6200 csrc->mc_ki[csrc->mc_top] = 0;
6201 rc = mdb_update_key(csrc, &nullkey);
6202 csrc->mc_ki[csrc->mc_top] = ix;
6203 assert(rc == MDB_SUCCESS);
6207 if (cdst->mc_ki[cdst->mc_top] == 0) {
6208 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6209 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6210 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6212 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6213 key.mv_size = NODEKSZ(srcnode);
6214 key.mv_data = NODEKEY(srcnode);
6216 DPRINTF("update separator for destination page %zu to [%s]",
6217 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6218 mdb_cursor_copy(cdst, &mn);
6221 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6224 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6226 indx_t ix = cdst->mc_ki[cdst->mc_top];
6227 nullkey.mv_size = 0;
6228 cdst->mc_ki[cdst->mc_top] = 0;
6229 rc = mdb_update_key(cdst, &nullkey);
6230 cdst->mc_ki[cdst->mc_top] = ix;
6231 assert(rc == MDB_SUCCESS);
6238 /** Merge one page into another.
6239 * The nodes from the page pointed to by \b csrc will
6240 * be copied to the page pointed to by \b cdst and then
6241 * the \b csrc page will be freed.
6242 * @param[in] csrc Cursor pointing to the source page.
6243 * @param[in] cdst Cursor pointing to the destination page.
6246 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6254 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6255 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6257 assert(csrc->mc_snum > 1); /* can't merge root page */
6258 assert(cdst->mc_snum > 1);
6260 /* Mark dst as dirty. */
6261 if ((rc = mdb_page_touch(cdst)))
6264 /* Move all nodes from src to dst.
6266 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6267 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6268 key.mv_size = csrc->mc_db->md_pad;
6269 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6270 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6271 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6272 if (rc != MDB_SUCCESS)
6274 key.mv_data = (char *)key.mv_data + key.mv_size;
6277 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6278 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6279 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6280 unsigned int snum = csrc->mc_snum;
6282 /* must find the lowest key below src */
6283 mdb_page_search_lowest(csrc);
6284 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6285 key.mv_size = csrc->mc_db->md_pad;
6286 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6288 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6289 key.mv_size = NODEKSZ(s2);
6290 key.mv_data = NODEKEY(s2);
6292 csrc->mc_snum = snum--;
6293 csrc->mc_top = snum;
6295 key.mv_size = srcnode->mn_ksize;
6296 key.mv_data = NODEKEY(srcnode);
6299 data.mv_size = NODEDSZ(srcnode);
6300 data.mv_data = NODEDATA(srcnode);
6301 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6302 if (rc != MDB_SUCCESS)
6307 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6308 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);
6310 /* Unlink the src page from parent and add to free list.
6312 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6313 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6316 rc = mdb_update_key(csrc, &key);
6322 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6323 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6324 csrc->mc_db->md_leaf_pages--;
6326 csrc->mc_db->md_branch_pages--;
6328 /* Adjust other cursors pointing to mp */
6329 MDB_cursor *m2, *m3;
6330 MDB_dbi dbi = csrc->mc_dbi;
6331 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6333 if (csrc->mc_flags & C_SUB)
6336 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6337 if (csrc->mc_flags & C_SUB)
6338 m3 = &m2->mc_xcursor->mx_cursor;
6341 if (m3 == csrc) continue;
6342 if (m3->mc_snum < csrc->mc_snum) continue;
6343 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6344 m3->mc_pg[csrc->mc_top] = mp;
6345 m3->mc_ki[csrc->mc_top] += nkeys;
6349 mdb_cursor_pop(csrc);
6351 return mdb_rebalance(csrc);
6354 /** Copy the contents of a cursor.
6355 * @param[in] csrc The cursor to copy from.
6356 * @param[out] cdst The cursor to copy to.
6359 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6363 cdst->mc_txn = csrc->mc_txn;
6364 cdst->mc_dbi = csrc->mc_dbi;
6365 cdst->mc_db = csrc->mc_db;
6366 cdst->mc_dbx = csrc->mc_dbx;
6367 cdst->mc_snum = csrc->mc_snum;
6368 cdst->mc_top = csrc->mc_top;
6369 cdst->mc_flags = csrc->mc_flags;
6371 for (i=0; i<csrc->mc_snum; i++) {
6372 cdst->mc_pg[i] = csrc->mc_pg[i];
6373 cdst->mc_ki[i] = csrc->mc_ki[i];
6377 /** Rebalance the tree after a delete operation.
6378 * @param[in] mc Cursor pointing to the page where rebalancing
6380 * @return 0 on success, non-zero on failure.
6383 mdb_rebalance(MDB_cursor *mc)
6387 unsigned int ptop, minkeys;
6390 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6394 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6395 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6396 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6397 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6401 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6402 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6405 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6406 DPRINTF("no need to rebalance page %zu, above fill threshold",
6412 if (mc->mc_snum < 2) {
6413 MDB_page *mp = mc->mc_pg[0];
6415 DPUTS("Can't rebalance a subpage, ignoring");
6418 if (NUMKEYS(mp) == 0) {
6419 DPUTS("tree is completely empty");
6420 mc->mc_db->md_root = P_INVALID;
6421 mc->mc_db->md_depth = 0;
6422 mc->mc_db->md_leaf_pages = 0;
6423 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6424 /* Adjust cursors pointing to mp */
6428 MDB_cursor *m2, *m3;
6429 MDB_dbi dbi = mc->mc_dbi;
6431 if (mc->mc_flags & C_SUB)
6434 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6435 if (mc->mc_flags & C_SUB)
6436 m3 = &m2->mc_xcursor->mx_cursor;
6439 if (m3->mc_snum < mc->mc_snum) continue;
6440 if (m3->mc_pg[0] == mp) {
6446 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6447 DPUTS("collapsing root page!");
6448 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6449 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6450 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6453 mc->mc_db->md_depth--;
6454 mc->mc_db->md_branch_pages--;
6456 /* Adjust other cursors pointing to mp */
6457 MDB_cursor *m2, *m3;
6458 MDB_dbi dbi = mc->mc_dbi;
6460 if (mc->mc_flags & C_SUB)
6463 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6464 if (mc->mc_flags & C_SUB)
6465 m3 = &m2->mc_xcursor->mx_cursor;
6468 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6469 if (m3->mc_pg[0] == mp) {
6470 m3->mc_pg[0] = mc->mc_pg[0];
6477 DPUTS("root page doesn't need rebalancing");
6481 /* The parent (branch page) must have at least 2 pointers,
6482 * otherwise the tree is invalid.
6484 ptop = mc->mc_top-1;
6485 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6487 /* Leaf page fill factor is below the threshold.
6488 * Try to move keys from left or right neighbor, or
6489 * merge with a neighbor page.
6494 mdb_cursor_copy(mc, &mn);
6495 mn.mc_xcursor = NULL;
6497 if (mc->mc_ki[ptop] == 0) {
6498 /* We're the leftmost leaf in our parent.
6500 DPUTS("reading right neighbor");
6502 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6503 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6506 mn.mc_ki[mn.mc_top] = 0;
6507 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6509 /* There is at least one neighbor to the left.
6511 DPUTS("reading left neighbor");
6513 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6514 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6517 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6518 mc->mc_ki[mc->mc_top] = 0;
6521 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6522 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);
6524 /* If the neighbor page is above threshold and has enough keys,
6525 * move one key from it. Otherwise we should try to merge them.
6526 * (A branch page must never have less than 2 keys.)
6528 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6529 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6530 return mdb_node_move(&mn, mc);
6532 if (mc->mc_ki[ptop] == 0)
6533 rc = mdb_page_merge(&mn, mc);
6535 rc = mdb_page_merge(mc, &mn);
6536 mc->mc_flags &= ~C_INITIALIZED;
6541 /** Complete a delete operation started by #mdb_cursor_del(). */
6543 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6547 /* add overflow pages to free list */
6548 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6552 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6553 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
6555 assert(IS_OVERFLOW(omp));
6556 mdb_ovpage_free(mc, omp);
6558 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6559 mc->mc_db->md_entries--;
6560 rc = mdb_rebalance(mc);
6561 if (rc != MDB_SUCCESS)
6562 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6563 /* if mc points past last node in page, invalidate */
6564 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6565 mc->mc_flags &= ~C_INITIALIZED;
6571 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6572 MDB_val *key, MDB_val *data)
6577 MDB_val rdata, *xdata;
6581 assert(key != NULL);
6583 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6585 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6588 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6592 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6596 mdb_cursor_init(&mc, txn, dbi, &mx);
6607 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6609 /* let mdb_page_split know about this cursor if needed:
6610 * delete will trigger a rebalance; if it needs to move
6611 * a node from one page to another, it will have to
6612 * update the parent's separator key(s). If the new sepkey
6613 * is larger than the current one, the parent page may
6614 * run out of space, triggering a split. We need this
6615 * cursor to be consistent until the end of the rebalance.
6617 mc.mc_next = txn->mt_cursors[dbi];
6618 txn->mt_cursors[dbi] = &mc;
6619 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6620 txn->mt_cursors[dbi] = mc.mc_next;
6625 /** Split a page and insert a new node.
6626 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6627 * The cursor will be updated to point to the actual page and index where
6628 * the node got inserted after the split.
6629 * @param[in] newkey The key for the newly inserted node.
6630 * @param[in] newdata The data for the newly inserted node.
6631 * @param[in] newpgno The page number, if the new node is a branch node.
6632 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6633 * @return 0 on success, non-zero on failure.
6636 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6637 unsigned int nflags)
6640 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6643 unsigned int i, j, split_indx, nkeys, pmax;
6645 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6647 MDB_page *mp, *rp, *pp;
6652 mp = mc->mc_pg[mc->mc_top];
6653 newindx = mc->mc_ki[mc->mc_top];
6655 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6656 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6657 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6659 /* Create a right sibling. */
6660 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6662 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6664 if (mc->mc_snum < 2) {
6665 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6667 /* shift current top to make room for new parent */
6668 mc->mc_pg[1] = mc->mc_pg[0];
6669 mc->mc_ki[1] = mc->mc_ki[0];
6672 mc->mc_db->md_root = pp->mp_pgno;
6673 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6674 mc->mc_db->md_depth++;
6677 /* Add left (implicit) pointer. */
6678 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6679 /* undo the pre-push */
6680 mc->mc_pg[0] = mc->mc_pg[1];
6681 mc->mc_ki[0] = mc->mc_ki[1];
6682 mc->mc_db->md_root = mp->mp_pgno;
6683 mc->mc_db->md_depth--;
6690 ptop = mc->mc_top-1;
6691 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6694 mc->mc_flags |= C_SPLITTING;
6695 mdb_cursor_copy(mc, &mn);
6696 mn.mc_pg[mn.mc_top] = rp;
6697 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6699 if (nflags & MDB_APPEND) {
6700 mn.mc_ki[mn.mc_top] = 0;
6702 split_indx = newindx;
6707 nkeys = NUMKEYS(mp);
6708 split_indx = nkeys / 2;
6709 if (newindx < split_indx)
6715 unsigned int lsize, rsize, ksize;
6716 /* Move half of the keys to the right sibling */
6718 x = mc->mc_ki[mc->mc_top] - split_indx;
6719 ksize = mc->mc_db->md_pad;
6720 split = LEAF2KEY(mp, split_indx, ksize);
6721 rsize = (nkeys - split_indx) * ksize;
6722 lsize = (nkeys - split_indx) * sizeof(indx_t);
6723 mp->mp_lower -= lsize;
6724 rp->mp_lower += lsize;
6725 mp->mp_upper += rsize - lsize;
6726 rp->mp_upper -= rsize - lsize;
6727 sepkey.mv_size = ksize;
6728 if (newindx == split_indx) {
6729 sepkey.mv_data = newkey->mv_data;
6731 sepkey.mv_data = split;
6734 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6735 memcpy(rp->mp_ptrs, split, rsize);
6736 sepkey.mv_data = rp->mp_ptrs;
6737 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6738 memcpy(ins, newkey->mv_data, ksize);
6739 mp->mp_lower += sizeof(indx_t);
6740 mp->mp_upper -= ksize - sizeof(indx_t);
6743 memcpy(rp->mp_ptrs, split, x * ksize);
6744 ins = LEAF2KEY(rp, x, ksize);
6745 memcpy(ins, newkey->mv_data, ksize);
6746 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6747 rp->mp_lower += sizeof(indx_t);
6748 rp->mp_upper -= ksize - sizeof(indx_t);
6749 mc->mc_ki[mc->mc_top] = x;
6750 mc->mc_pg[mc->mc_top] = rp;
6755 /* For leaf pages, check the split point based on what
6756 * fits where, since otherwise mdb_node_add can fail.
6758 * This check is only needed when the data items are
6759 * relatively large, such that being off by one will
6760 * make the difference between success or failure.
6762 * It's also relevant if a page happens to be laid out
6763 * such that one half of its nodes are all "small" and
6764 * the other half of its nodes are "large." If the new
6765 * item is also "large" and falls on the half with
6766 * "large" nodes, it also may not fit.
6769 unsigned int psize, nsize;
6770 /* Maximum free space in an empty page */
6771 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6772 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6773 if ((nkeys < 20) || (nsize > pmax/16)) {
6774 if (newindx <= split_indx) {
6777 for (i=0; i<split_indx; i++) {
6778 node = NODEPTR(mp, i);
6779 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6780 if (F_ISSET(node->mn_flags, F_BIGDATA))
6781 psize += sizeof(pgno_t);
6783 psize += NODEDSZ(node);
6787 split_indx = newindx;
6798 for (i=nkeys-1; i>=split_indx; i--) {
6799 node = NODEPTR(mp, i);
6800 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6801 if (F_ISSET(node->mn_flags, F_BIGDATA))
6802 psize += sizeof(pgno_t);
6804 psize += NODEDSZ(node);
6808 split_indx = newindx;
6819 /* First find the separating key between the split pages.
6820 * The case where newindx == split_indx is ambiguous; the
6821 * new item could go to the new page or stay on the original
6822 * page. If newpos == 1 it goes to the new page.
6824 if (newindx == split_indx && newpos) {
6825 sepkey.mv_size = newkey->mv_size;
6826 sepkey.mv_data = newkey->mv_data;
6828 node = NODEPTR(mp, split_indx);
6829 sepkey.mv_size = node->mn_ksize;
6830 sepkey.mv_data = NODEKEY(node);
6834 DPRINTF("separator is [%s]", DKEY(&sepkey));
6836 /* Copy separator key to the parent.
6838 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6842 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6845 if (mn.mc_snum == mc->mc_snum) {
6846 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6847 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6848 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6849 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6854 /* Right page might now have changed parent.
6855 * Check if left page also changed parent.
6857 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6858 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6859 for (i=0; i<ptop; i++) {
6860 mc->mc_pg[i] = mn.mc_pg[i];
6861 mc->mc_ki[i] = mn.mc_ki[i];
6863 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6864 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6868 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6871 mc->mc_flags ^= C_SPLITTING;
6872 if (rc != MDB_SUCCESS) {
6875 if (nflags & MDB_APPEND) {
6876 mc->mc_pg[mc->mc_top] = rp;
6877 mc->mc_ki[mc->mc_top] = 0;
6878 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6881 for (i=0; i<mc->mc_top; i++)
6882 mc->mc_ki[i] = mn.mc_ki[i];
6889 /* Move half of the keys to the right sibling. */
6891 /* grab a page to hold a temporary copy */
6892 copy = mdb_page_malloc(mc, 1);
6896 copy->mp_pgno = mp->mp_pgno;
6897 copy->mp_flags = mp->mp_flags;
6898 copy->mp_lower = PAGEHDRSZ;
6899 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6900 mc->mc_pg[mc->mc_top] = copy;
6901 for (i = j = 0; i <= nkeys; j++) {
6902 if (i == split_indx) {
6903 /* Insert in right sibling. */
6904 /* Reset insert index for right sibling. */
6905 if (i != newindx || (newpos ^ ins_new)) {
6907 mc->mc_pg[mc->mc_top] = rp;
6911 if (i == newindx && !ins_new) {
6912 /* Insert the original entry that caused the split. */
6913 rkey.mv_data = newkey->mv_data;
6914 rkey.mv_size = newkey->mv_size;
6923 /* Update index for the new key. */
6924 mc->mc_ki[mc->mc_top] = j;
6925 } else if (i == nkeys) {
6928 node = NODEPTR(mp, i);
6929 rkey.mv_data = NODEKEY(node);
6930 rkey.mv_size = node->mn_ksize;
6932 xdata.mv_data = NODEDATA(node);
6933 xdata.mv_size = NODEDSZ(node);
6936 pgno = NODEPGNO(node);
6937 flags = node->mn_flags;
6942 if (!IS_LEAF(mp) && j == 0) {
6943 /* First branch index doesn't need key data. */
6947 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6951 nkeys = NUMKEYS(copy);
6952 for (i=0; i<nkeys; i++)
6953 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6954 mp->mp_lower = copy->mp_lower;
6955 mp->mp_upper = copy->mp_upper;
6956 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6957 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6959 /* reset back to original page */
6960 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6961 mc->mc_pg[mc->mc_top] = mp;
6962 if (nflags & MDB_RESERVE) {
6963 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6964 if (!(node->mn_flags & F_BIGDATA))
6965 newdata->mv_data = NODEDATA(node);
6969 /* Make sure mc_ki is still valid.
6971 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6972 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6973 for (i=0; i<ptop; i++) {
6974 mc->mc_pg[i] = mn.mc_pg[i];
6975 mc->mc_ki[i] = mn.mc_ki[i];
6977 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6978 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6982 /* return tmp page to freelist */
6983 mdb_page_free(mc->mc_txn->mt_env, copy);
6986 /* Adjust other cursors pointing to mp */
6987 MDB_cursor *m2, *m3;
6988 MDB_dbi dbi = mc->mc_dbi;
6989 int fixup = NUMKEYS(mp);
6991 if (mc->mc_flags & C_SUB)
6994 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6995 if (mc->mc_flags & C_SUB)
6996 m3 = &m2->mc_xcursor->mx_cursor;
7001 if (!(m3->mc_flags & C_INITIALIZED))
7003 if (m3->mc_flags & C_SPLITTING)
7008 for (k=m3->mc_top; k>=0; k--) {
7009 m3->mc_ki[k+1] = m3->mc_ki[k];
7010 m3->mc_pg[k+1] = m3->mc_pg[k];
7012 if (m3->mc_ki[0] >= split_indx) {
7017 m3->mc_pg[0] = mc->mc_pg[0];
7021 if (m3->mc_pg[mc->mc_top] == mp) {
7022 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7023 m3->mc_ki[mc->mc_top]++;
7024 if (m3->mc_ki[mc->mc_top] >= fixup) {
7025 m3->mc_pg[mc->mc_top] = rp;
7026 m3->mc_ki[mc->mc_top] -= fixup;
7027 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7029 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7030 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7039 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7040 MDB_val *key, MDB_val *data, unsigned int flags)
7045 assert(key != NULL);
7046 assert(data != NULL);
7048 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7051 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7055 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7059 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7062 mdb_cursor_init(&mc, txn, dbi, &mx);
7063 return mdb_cursor_put(&mc, key, data, flags);
7067 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7069 if ((flag & CHANGEABLE) != flag)
7072 env->me_flags |= flag;
7074 env->me_flags &= ~flag;
7079 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7084 *arg = env->me_flags;
7089 mdb_env_get_path(MDB_env *env, const char **arg)
7094 *arg = env->me_path;
7098 /** Common code for #mdb_stat() and #mdb_env_stat().
7099 * @param[in] env the environment to operate in.
7100 * @param[in] db the #MDB_db record containing the stats to return.
7101 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7102 * @return 0, this function always succeeds.
7105 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7107 arg->ms_psize = env->me_psize;
7108 arg->ms_depth = db->md_depth;
7109 arg->ms_branch_pages = db->md_branch_pages;
7110 arg->ms_leaf_pages = db->md_leaf_pages;
7111 arg->ms_overflow_pages = db->md_overflow_pages;
7112 arg->ms_entries = db->md_entries;
7117 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7121 if (env == NULL || arg == NULL)
7124 toggle = mdb_env_pick_meta(env);
7126 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7130 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7134 if (env == NULL || arg == NULL)
7137 toggle = mdb_env_pick_meta(env);
7138 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7139 arg->me_mapsize = env->me_mapsize;
7140 arg->me_maxreaders = env->me_maxreaders;
7141 arg->me_numreaders = env->me_numreaders;
7142 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7143 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7147 /** Set the default comparison functions for a database.
7148 * Called immediately after a database is opened to set the defaults.
7149 * The user can then override them with #mdb_set_compare() or
7150 * #mdb_set_dupsort().
7151 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7152 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7155 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7157 uint16_t f = txn->mt_dbs[dbi].md_flags;
7159 txn->mt_dbxs[dbi].md_cmp =
7160 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7161 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7163 txn->mt_dbxs[dbi].md_dcmp =
7164 !(f & MDB_DUPSORT) ? 0 :
7165 ((f & MDB_INTEGERDUP)
7166 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7167 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7170 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7175 int rc, dbflag, exact;
7176 unsigned int unused = 0;
7179 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7180 mdb_default_cmp(txn, FREE_DBI);
7183 if ((flags & VALID_FLAGS) != flags)
7189 if (flags & PERSISTENT_FLAGS) {
7190 uint16_t f2 = flags & PERSISTENT_FLAGS;
7191 /* make sure flag changes get committed */
7192 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7193 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7194 txn->mt_flags |= MDB_TXN_DIRTY;
7197 mdb_default_cmp(txn, MAIN_DBI);
7201 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7202 mdb_default_cmp(txn, MAIN_DBI);
7205 /* Is the DB already open? */
7207 for (i=2; i<txn->mt_numdbs; i++) {
7208 if (!txn->mt_dbxs[i].md_name.mv_size) {
7209 /* Remember this free slot */
7210 if (!unused) unused = i;
7213 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7214 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7220 /* If no free slot and max hit, fail */
7221 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7222 return MDB_DBS_FULL;
7224 /* Cannot mix named databases with some mainDB flags */
7225 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7226 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7228 /* Find the DB info */
7229 dbflag = DB_NEW|DB_VALID;
7232 key.mv_data = (void *)name;
7233 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7234 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7235 if (rc == MDB_SUCCESS) {
7236 /* make sure this is actually a DB */
7237 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7238 if (!(node->mn_flags & F_SUBDATA))
7240 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7241 /* Create if requested */
7243 data.mv_size = sizeof(MDB_db);
7244 data.mv_data = &dummy;
7245 memset(&dummy, 0, sizeof(dummy));
7246 dummy.md_root = P_INVALID;
7247 dummy.md_flags = flags & PERSISTENT_FLAGS;
7248 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7252 /* OK, got info, add to table */
7253 if (rc == MDB_SUCCESS) {
7254 unsigned int slot = unused ? unused : txn->mt_numdbs;
7255 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7256 txn->mt_dbxs[slot].md_name.mv_size = len;
7257 txn->mt_dbxs[slot].md_rel = NULL;
7258 txn->mt_dbflags[slot] = dbflag;
7259 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7261 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7262 mdb_default_cmp(txn, slot);
7271 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7273 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7276 if (txn->mt_dbflags[dbi] & DB_STALE) {
7279 /* Stale, must read the DB's root. cursor_init does it for us. */
7280 mdb_cursor_init(&mc, txn, dbi, &mx);
7282 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7285 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7288 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7290 ptr = env->me_dbxs[dbi].md_name.mv_data;
7291 env->me_dbxs[dbi].md_name.mv_data = NULL;
7292 env->me_dbxs[dbi].md_name.mv_size = 0;
7293 env->me_dbflags[dbi] = 0;
7297 /** Add all the DB's pages to the free list.
7298 * @param[in] mc Cursor on the DB to free.
7299 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7300 * @return 0 on success, non-zero on failure.
7303 mdb_drop0(MDB_cursor *mc, int subs)
7307 rc = mdb_page_search(mc, NULL, 0);
7308 if (rc == MDB_SUCCESS) {
7313 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7314 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7317 mdb_cursor_copy(mc, &mx);
7318 while (mc->mc_snum > 0) {
7319 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7320 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7321 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7322 if (ni->mn_flags & F_BIGDATA) {
7326 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7327 rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL);
7330 assert(IS_OVERFLOW(omp));
7331 ovpages = omp->mp_pages;
7332 for (j=0; j<ovpages; j++) {
7333 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7336 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7337 mdb_xcursor_init1(mc, ni);
7338 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7344 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7346 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7349 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7354 mc->mc_ki[mc->mc_top] = i;
7355 rc = mdb_cursor_sibling(mc, 1);
7357 /* no more siblings, go back to beginning
7358 * of previous level.
7362 for (i=1; i<mc->mc_snum; i++) {
7364 mc->mc_pg[i] = mx.mc_pg[i];
7369 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7370 mc->mc_db->md_root);
7375 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7377 MDB_cursor *mc, *m2;
7380 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7383 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7386 rc = mdb_cursor_open(txn, dbi, &mc);
7390 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7391 /* Invalidate the dropped DB's cursors */
7392 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7393 m2->mc_flags &= ~C_INITIALIZED;
7397 /* Can't delete the main DB */
7398 if (del && dbi > MAIN_DBI) {
7399 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7401 txn->mt_dbflags[dbi] = DB_STALE;
7402 mdb_dbi_close(txn->mt_env, dbi);
7405 /* reset the DB record, mark it dirty */
7406 txn->mt_dbflags[dbi] |= DB_DIRTY;
7407 txn->mt_dbs[dbi].md_depth = 0;
7408 txn->mt_dbs[dbi].md_branch_pages = 0;
7409 txn->mt_dbs[dbi].md_leaf_pages = 0;
7410 txn->mt_dbs[dbi].md_overflow_pages = 0;
7411 txn->mt_dbs[dbi].md_entries = 0;
7412 txn->mt_dbs[dbi].md_root = P_INVALID;
7414 txn->mt_flags |= MDB_TXN_DIRTY;
7417 mdb_cursor_close(mc);
7421 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7423 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7426 txn->mt_dbxs[dbi].md_cmp = cmp;
7430 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7432 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7435 txn->mt_dbxs[dbi].md_dcmp = cmp;
7439 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7441 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7444 txn->mt_dbxs[dbi].md_rel = rel;
7448 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7450 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7453 txn->mt_dbxs[dbi].md_relctx = ctx;