2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define pthread_t DWORD
144 #define pthread_mutex_t HANDLE
145 #define pthread_key_t DWORD
146 #define pthread_self() GetCurrentThreadId()
147 #define pthread_key_create(x,y) \
148 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
149 #define pthread_key_delete(x) TlsFree(x)
150 #define pthread_getspecific(x) TlsGetValue(x)
151 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
152 #define pthread_mutex_unlock(x) ReleaseMutex(x)
153 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
154 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
158 #define getpid() GetCurrentProcessId()
159 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
160 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
161 #define ErrCode() GetLastError()
162 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
163 #define close(fd) CloseHandle(fd)
164 #define munmap(ptr,len) UnmapViewOfFile(ptr)
167 #ifdef MDB_USE_POSIX_SEM
169 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
170 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
171 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
172 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
175 mdb_sem_wait(sem_t *sem)
178 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
183 /** Lock the reader mutex.
185 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
186 /** Unlock the reader mutex.
188 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
190 /** Lock the writer mutex.
191 * Only a single write transaction is allowed at a time. Other writers
192 * will block waiting for this mutex.
194 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
195 /** Unlock the writer mutex.
197 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
198 #endif /* MDB_USE_POSIX_SEM */
200 /** Get the error code for the last failed system function.
202 #define ErrCode() errno
204 /** An abstraction for a file handle.
205 * On POSIX systems file handles are small integers. On Windows
206 * they're opaque pointers.
210 /** A value for an invalid file handle.
211 * Mainly used to initialize file variables and signify that they are
214 #define INVALID_HANDLE_VALUE (-1)
216 /** Get the size of a memory page for the system.
217 * This is the basic size that the platform's memory manager uses, and is
218 * fundamental to the use of memory-mapped files.
220 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
223 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
226 #define MNAME_LEN (sizeof(pthread_mutex_t))
232 /** A flag for opening a file and requesting synchronous data writes.
233 * This is only used when writing a meta page. It's not strictly needed;
234 * we could just do a normal write and then immediately perform a flush.
235 * But if this flag is available it saves us an extra system call.
237 * @note If O_DSYNC is undefined but exists in /usr/include,
238 * preferably set some compiler flag to get the definition.
239 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
242 # define MDB_DSYNC O_DSYNC
246 /** Function for flushing the data of a file. Define this to fsync
247 * if fdatasync() is not supported.
249 #ifndef MDB_FDATASYNC
250 # define MDB_FDATASYNC fdatasync
254 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
265 /** A page number in the database.
266 * Note that 64 bit page numbers are overkill, since pages themselves
267 * already represent 12-13 bits of addressable memory, and the OS will
268 * always limit applications to a maximum of 63 bits of address space.
270 * @note In the #MDB_node structure, we only store 48 bits of this value,
271 * which thus limits us to only 60 bits of addressable data.
273 typedef MDB_ID pgno_t;
275 /** A transaction ID.
276 * See struct MDB_txn.mt_txnid for details.
278 typedef MDB_ID txnid_t;
280 /** @defgroup debug Debug Macros
284 /** Enable debug output.
285 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
286 * read from and written to the database (used for free space management).
291 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
294 # define DPRINTF (void) /* Vararg macros may be unsupported */
296 static int mdb_debug;
297 static txnid_t mdb_debug_start;
299 /** Print a debug message with printf formatting. */
300 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
301 ((void) ((mdb_debug) && \
302 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
304 # define DPRINTF(fmt, ...) ((void) 0)
305 # define MDB_DEBUG_SKIP
307 /** Print a debug string.
308 * The string is printed literally, with no format processing.
310 #define DPUTS(arg) DPRINTF("%s", arg)
313 /** A default memory page size.
314 * The actual size is platform-dependent, but we use this for
315 * boot-strapping. We probably should not be using this any more.
316 * The #GET_PAGESIZE() macro is used to get the actual size.
318 * Note that we don't currently support Huge pages. On Linux,
319 * regular data files cannot use Huge pages, and in general
320 * Huge pages aren't actually pageable. We rely on the OS
321 * demand-pager to read our data and page it out when memory
322 * pressure from other processes is high. So until OSs have
323 * actual paging support for Huge pages, they're not viable.
325 #define MDB_PAGESIZE 4096
327 /** The minimum number of keys required in a database page.
328 * Setting this to a larger value will place a smaller bound on the
329 * maximum size of a data item. Data items larger than this size will
330 * be pushed into overflow pages instead of being stored directly in
331 * the B-tree node. This value used to default to 4. With a page size
332 * of 4096 bytes that meant that any item larger than 1024 bytes would
333 * go into an overflow page. That also meant that on average 2-3KB of
334 * each overflow page was wasted space. The value cannot be lower than
335 * 2 because then there would no longer be a tree structure. With this
336 * value, items larger than 2KB will go into overflow pages, and on
337 * average only 1KB will be wasted.
339 #define MDB_MINKEYS 2
341 /** A stamp that identifies a file as an MDB file.
342 * There's nothing special about this value other than that it is easily
343 * recognizable, and it will reflect any byte order mismatches.
345 #define MDB_MAGIC 0xBEEFC0DE
347 /** The version number for a database's file format. */
348 #define MDB_VERSION 1
350 /** @brief The maximum size of a key in the database.
352 * We require that keys all fit onto a regular page. This limit
353 * could be raised a bit further if needed; to something just
354 * under #MDB_PAGESIZE / #MDB_MINKEYS.
356 * Note that data items in an #MDB_DUPSORT database are actually keys
357 * of a subDB, so they're also limited to this size.
359 #ifndef MDB_MAXKEYSIZE
360 #define MDB_MAXKEYSIZE 511
363 /** @brief The maximum size of a data item.
365 * We only store a 32 bit value for node sizes.
367 #define MAXDATASIZE 0xffffffffUL
372 * This is used for printing a hex dump of a key's contents.
374 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
375 /** Display a key in hex.
377 * Invoke a function to display a key in hex.
379 #define DKEY(x) mdb_dkey(x, kbuf)
381 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
385 /** An invalid page number.
386 * Mainly used to denote an empty tree.
388 #define P_INVALID (~(pgno_t)0)
390 /** Test if the flags \b f are set in a flag word \b w. */
391 #define F_ISSET(w, f) (((w) & (f)) == (f))
393 /** Used for offsets within a single page.
394 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
397 typedef uint16_t indx_t;
399 /** Default size of memory map.
400 * This is certainly too small for any actual applications. Apps should always set
401 * the size explicitly using #mdb_env_set_mapsize().
403 #define DEFAULT_MAPSIZE 1048576
405 /** @defgroup readers Reader Lock Table
406 * Readers don't acquire any locks for their data access. Instead, they
407 * simply record their transaction ID in the reader table. The reader
408 * mutex is needed just to find an empty slot in the reader table. The
409 * slot's address is saved in thread-specific data so that subsequent read
410 * transactions started by the same thread need no further locking to proceed.
412 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
414 * No reader table is used if the database is on a read-only filesystem.
416 * Since the database uses multi-version concurrency control, readers don't
417 * actually need any locking. This table is used to keep track of which
418 * readers are using data from which old transactions, so that we'll know
419 * when a particular old transaction is no longer in use. Old transactions
420 * that have discarded any data pages can then have those pages reclaimed
421 * for use by a later write transaction.
423 * The lock table is constructed such that reader slots are aligned with the
424 * processor's cache line size. Any slot is only ever used by one thread.
425 * This alignment guarantees that there will be no contention or cache
426 * thrashing as threads update their own slot info, and also eliminates
427 * any need for locking when accessing a slot.
429 * A writer thread will scan every slot in the table to determine the oldest
430 * outstanding reader transaction. Any freed pages older than this will be
431 * reclaimed by the writer. The writer doesn't use any locks when scanning
432 * this table. This means that there's no guarantee that the writer will
433 * see the most up-to-date reader info, but that's not required for correct
434 * operation - all we need is to know the upper bound on the oldest reader,
435 * we don't care at all about the newest reader. So the only consequence of
436 * reading stale information here is that old pages might hang around a
437 * while longer before being reclaimed. That's actually good anyway, because
438 * the longer we delay reclaiming old pages, the more likely it is that a
439 * string of contiguous pages can be found after coalescing old pages from
440 * many old transactions together.
443 /** Number of slots in the reader table.
444 * This value was chosen somewhat arbitrarily. 126 readers plus a
445 * couple mutexes fit exactly into 8KB on my development machine.
446 * Applications should set the table size using #mdb_env_set_maxreaders().
448 #define DEFAULT_READERS 126
450 /** The size of a CPU cache line in bytes. We want our lock structures
451 * aligned to this size to avoid false cache line sharing in the
453 * This value works for most CPUs. For Itanium this should be 128.
459 /** The information we store in a single slot of the reader table.
460 * In addition to a transaction ID, we also record the process and
461 * thread ID that owns a slot, so that we can detect stale information,
462 * e.g. threads or processes that went away without cleaning up.
463 * @note We currently don't check for stale records. We simply re-init
464 * the table when we know that we're the only process opening the
467 typedef struct MDB_rxbody {
468 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
469 * Multiple readers that start at the same time will probably have the
470 * same ID here. Again, it's not important to exclude them from
471 * anything; all we need to know is which version of the DB they
472 * started from so we can avoid overwriting any data used in that
473 * particular version.
476 /** The process ID of the process owning this reader txn. */
478 /** The thread ID of the thread owning this txn. */
482 /** The actual reader record, with cacheline padding. */
483 typedef struct MDB_reader {
486 /** shorthand for mrb_txnid */
487 #define mr_txnid mru.mrx.mrb_txnid
488 #define mr_pid mru.mrx.mrb_pid
489 #define mr_tid mru.mrx.mrb_tid
490 /** cache line alignment */
491 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
495 /** The header for the reader table.
496 * The table resides in a memory-mapped file. (This is a different file
497 * than is used for the main database.)
499 * For POSIX the actual mutexes reside in the shared memory of this
500 * mapped file. On Windows, mutexes are named objects allocated by the
501 * kernel; we store the mutex names in this mapped file so that other
502 * processes can grab them. This same approach is also used on
503 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
504 * process-shared POSIX mutexes. For these cases where a named object
505 * is used, the object name is derived from a 64 bit FNV hash of the
506 * environment pathname. As such, naming collisions are extremely
507 * unlikely. If a collision occurs, the results are unpredictable.
509 typedef struct MDB_txbody {
510 /** Stamp identifying this as an MDB file. It must be set
513 /** Version number of this lock file. Must be set to #MDB_VERSION. */
514 uint32_t mtb_version;
515 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
516 char mtb_rmname[MNAME_LEN];
518 /** Mutex protecting access to this table.
519 * This is the reader lock that #LOCK_MUTEX_R acquires.
521 pthread_mutex_t mtb_mutex;
523 /** The ID of the last transaction committed to the database.
524 * This is recorded here only for convenience; the value can always
525 * be determined by reading the main database meta pages.
528 /** The number of slots that have been used in the reader table.
529 * This always records the maximum count, it is not decremented
530 * when readers release their slots.
532 unsigned mtb_numreaders;
535 /** The actual reader table definition. */
536 typedef struct MDB_txninfo {
539 #define mti_magic mt1.mtb.mtb_magic
540 #define mti_version mt1.mtb.mtb_version
541 #define mti_mutex mt1.mtb.mtb_mutex
542 #define mti_rmname mt1.mtb.mtb_rmname
543 #define mti_txnid mt1.mtb.mtb_txnid
544 #define mti_numreaders mt1.mtb.mtb_numreaders
545 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
548 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
549 char mt2_wmname[MNAME_LEN];
550 #define mti_wmname mt2.mt2_wmname
552 pthread_mutex_t mt2_wmutex;
553 #define mti_wmutex mt2.mt2_wmutex
555 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
557 MDB_reader mti_readers[1];
561 /** Common header for all page types.
562 * Overflow records occupy a number of contiguous pages with no
563 * headers on any page after the first.
565 typedef struct MDB_page {
566 #define mp_pgno mp_p.p_pgno
567 #define mp_next mp_p.p_next
569 pgno_t p_pgno; /**< page number */
570 void * p_next; /**< for in-memory list of freed structs */
573 /** @defgroup mdb_page Page Flags
575 * Flags for the page headers.
578 #define P_BRANCH 0x01 /**< branch page */
579 #define P_LEAF 0x02 /**< leaf page */
580 #define P_OVERFLOW 0x04 /**< overflow page */
581 #define P_META 0x08 /**< meta page */
582 #define P_DIRTY 0x10 /**< dirty page */
583 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
584 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
586 uint16_t mp_flags; /**< @ref mdb_page */
587 #define mp_lower mp_pb.pb.pb_lower
588 #define mp_upper mp_pb.pb.pb_upper
589 #define mp_pages mp_pb.pb_pages
592 indx_t pb_lower; /**< lower bound of free space */
593 indx_t pb_upper; /**< upper bound of free space */
595 uint32_t pb_pages; /**< number of overflow pages */
597 indx_t mp_ptrs[1]; /**< dynamic size */
600 /** Size of the page header, excluding dynamic data at the end */
601 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
603 /** Address of first usable data byte in a page, after the header */
604 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
606 /** Number of nodes on a page */
607 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
609 /** The amount of space remaining in the page */
610 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
612 /** The percentage of space used in the page, in tenths of a percent. */
613 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
614 ((env)->me_psize - PAGEHDRSZ))
615 /** The minimum page fill factor, in tenths of a percent.
616 * Pages emptier than this are candidates for merging.
618 #define FILL_THRESHOLD 250
620 /** Test if a page is a leaf page */
621 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
622 /** Test if a page is a LEAF2 page */
623 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
624 /** Test if a page is a branch page */
625 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
626 /** Test if a page is an overflow page */
627 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
628 /** Test if a page is a sub page */
629 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
631 /** The number of overflow pages needed to store the given size. */
632 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
634 /** Header for a single key/data pair within a page.
635 * We guarantee 2-byte alignment for nodes.
637 typedef struct MDB_node {
638 /** lo and hi are used for data size on leaf nodes and for
639 * child pgno on branch nodes. On 64 bit platforms, flags
640 * is also used for pgno. (Branch nodes have no flags).
641 * They are in host byte order in case that lets some
642 * accesses be optimized into a 32-bit word access.
644 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
645 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
646 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
647 /** @defgroup mdb_node Node Flags
649 * Flags for node headers.
652 #define F_BIGDATA 0x01 /**< data put on overflow page */
653 #define F_SUBDATA 0x02 /**< data is a sub-database */
654 #define F_DUPDATA 0x04 /**< data has duplicates */
656 /** valid flags for #mdb_node_add() */
657 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
660 unsigned short mn_flags; /**< @ref mdb_node */
661 unsigned short mn_ksize; /**< key size */
662 char mn_data[1]; /**< key and data are appended here */
665 /** Size of the node header, excluding dynamic data at the end */
666 #define NODESIZE offsetof(MDB_node, mn_data)
668 /** Bit position of top word in page number, for shifting mn_flags */
669 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
671 /** Size of a node in a branch page with a given key.
672 * This is just the node header plus the key, there is no data.
674 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
676 /** Size of a node in a leaf page with a given key and data.
677 * This is node header plus key plus data size.
679 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
681 /** Address of node \b i in page \b p */
682 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
684 /** Address of the key for the node */
685 #define NODEKEY(node) (void *)((node)->mn_data)
687 /** Address of the data for a node */
688 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
690 /** Get the page number pointed to by a branch node */
691 #define NODEPGNO(node) \
692 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
693 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
694 /** Set the page number in a branch node */
695 #define SETPGNO(node,pgno) do { \
696 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
697 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
699 /** Get the size of the data in a leaf node */
700 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
701 /** Set the size of the data for a leaf node */
702 #define SETDSZ(node,size) do { \
703 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
704 /** The size of a key in a node */
705 #define NODEKSZ(node) ((node)->mn_ksize)
707 /** Copy a page number from src to dst */
709 #define COPY_PGNO(dst,src) dst = src
711 #if SIZE_MAX > 4294967295UL
712 #define COPY_PGNO(dst,src) do { \
713 unsigned short *s, *d; \
714 s = (unsigned short *)&(src); \
715 d = (unsigned short *)&(dst); \
722 #define COPY_PGNO(dst,src) do { \
723 unsigned short *s, *d; \
724 s = (unsigned short *)&(src); \
725 d = (unsigned short *)&(dst); \
731 /** The address of a key in a LEAF2 page.
732 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
733 * There are no node headers, keys are stored contiguously.
735 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
737 /** Set the \b node's key into \b key, if requested. */
738 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
739 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
741 /** Information about a single database in the environment. */
742 typedef struct MDB_db {
743 uint32_t md_pad; /**< also ksize for LEAF2 pages */
744 uint16_t md_flags; /**< @ref mdb_dbi_open */
745 uint16_t md_depth; /**< depth of this tree */
746 pgno_t md_branch_pages; /**< number of internal pages */
747 pgno_t md_leaf_pages; /**< number of leaf pages */
748 pgno_t md_overflow_pages; /**< number of overflow pages */
749 size_t md_entries; /**< number of data items */
750 pgno_t md_root; /**< the root page of this tree */
753 /** mdb_dbi_open flags */
754 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
755 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
756 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
757 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
759 /** Handle for the DB used to track free pages. */
761 /** Handle for the default DB. */
764 /** Meta page content. */
765 typedef struct MDB_meta {
766 /** Stamp identifying this as an MDB file. It must be set
769 /** Version number of this lock file. Must be set to #MDB_VERSION. */
771 void *mm_address; /**< address for fixed mapping */
772 size_t mm_mapsize; /**< size of mmap region */
773 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
774 /** The size of pages used in this DB */
775 #define mm_psize mm_dbs[0].md_pad
776 /** Any persistent environment flags. @ref mdb_env */
777 #define mm_flags mm_dbs[0].md_flags
778 pgno_t mm_last_pg; /**< last used page in file */
779 txnid_t mm_txnid; /**< txnid that committed this page */
782 /** Buffer for a stack-allocated dirty page.
783 * The members define size and alignment, and silence type
784 * aliasing warnings. They are not used directly; that could
785 * mean incorrectly using several union members in parallel.
787 typedef union MDB_pagebuf {
788 char mb_raw[MDB_PAGESIZE];
791 char mm_pad[PAGEHDRSZ];
796 /** Auxiliary DB info.
797 * The information here is mostly static/read-only. There is
798 * only a single copy of this record in the environment.
800 typedef struct MDB_dbx {
801 MDB_val md_name; /**< name of the database */
802 MDB_cmp_func *md_cmp; /**< function for comparing keys */
803 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
804 MDB_rel_func *md_rel; /**< user relocate function */
805 void *md_relctx; /**< user-provided context for md_rel */
808 /** A database transaction.
809 * Every operation requires a transaction handle.
812 MDB_txn *mt_parent; /**< parent of a nested txn */
813 MDB_txn *mt_child; /**< nested txn under this txn */
814 pgno_t mt_next_pgno; /**< next unallocated page */
815 /** The ID of this transaction. IDs are integers incrementing from 1.
816 * Only committed write transactions increment the ID. If a transaction
817 * aborts, the ID may be re-used by the next writer.
820 MDB_env *mt_env; /**< the DB environment */
821 /** The list of pages that became unused during this transaction.
825 MDB_ID2L dirty_list; /**< for write txns: modified pages */
826 MDB_reader *reader; /**< this thread's reader table slot or NULL */
828 /** Array of records for each DB known in the environment. */
830 /** Array of MDB_db records for each known DB */
832 /** @defgroup mt_dbflag Transaction DB Flags
836 #define DB_DIRTY 0x01 /**< DB was written in this txn */
837 #define DB_STALE 0x02 /**< DB record is older than txnID */
838 #define DB_NEW 0x04 /**< DB handle opened in this txn */
839 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
841 /** In write txns, array of cursors for each DB */
842 MDB_cursor **mt_cursors;
843 /** Array of flags for each DB */
844 unsigned char *mt_dbflags;
845 /** Number of DB records in use. This number only ever increments;
846 * we don't decrement it when individual DB handles are closed.
850 /** @defgroup mdb_txn Transaction Flags
854 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
855 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
856 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
858 unsigned int mt_flags; /**< @ref mdb_txn */
859 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
860 unsigned int mt_dirty_room;
861 /** Tracks which of the two meta pages was used at the start
862 * of this transaction.
864 unsigned int mt_toggle;
867 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
868 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
869 * raise this on a 64 bit machine.
871 #define CURSOR_STACK 32
875 /** Cursors are used for all DB operations */
877 /** Next cursor on this DB in this txn */
879 /** Original cursor if this is a shadow */
881 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
882 struct MDB_xcursor *mc_xcursor;
883 /** The transaction that owns this cursor */
885 /** The database handle this cursor operates on */
887 /** The database record for this cursor */
889 /** The database auxiliary record for this cursor */
891 /** The @ref mt_dbflag for this database */
892 unsigned char *mc_dbflag;
893 unsigned short mc_snum; /**< number of pushed pages */
894 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
895 /** @defgroup mdb_cursor Cursor Flags
897 * Cursor state flags.
900 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
901 #define C_EOF 0x02 /**< No more data */
902 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
903 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
904 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
905 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
906 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
908 unsigned int mc_flags; /**< @ref mdb_cursor */
909 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
910 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
913 /** Context for sorted-dup records.
914 * We could have gone to a fully recursive design, with arbitrarily
915 * deep nesting of sub-databases. But for now we only handle these
916 * levels - main DB, optional sub-DB, sorted-duplicate DB.
918 typedef struct MDB_xcursor {
919 /** A sub-cursor for traversing the Dup DB */
920 MDB_cursor mx_cursor;
921 /** The database record for this Dup DB */
923 /** The auxiliary DB record for this Dup DB */
925 /** The @ref mt_dbflag for this Dup DB */
926 unsigned char mx_dbflag;
929 /** State of FreeDB old pages, stored in the MDB_env */
930 typedef struct MDB_pgstate {
931 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
932 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
935 /** The database environment. */
937 HANDLE me_fd; /**< The main data file */
938 HANDLE me_lfd; /**< The lock file */
939 HANDLE me_mfd; /**< just for writing the meta pages */
940 /** Failed to update the meta page. Probably an I/O error. */
941 #define MDB_FATAL_ERROR 0x80000000U
942 /** Some fields are initialized. */
943 #define MDB_ENV_ACTIVE 0x20000000U
944 /** me_txkey is set */
945 #define MDB_ENV_TXKEY 0x10000000U
946 uint32_t me_flags; /**< @ref mdb_env */
947 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
948 unsigned int me_maxreaders; /**< size of the reader table */
949 unsigned int me_numreaders; /**< max numreaders set by this env */
950 MDB_dbi me_numdbs; /**< number of DBs opened */
951 MDB_dbi me_maxdbs; /**< size of the DB table */
952 pid_t me_pid; /**< process ID of this env */
953 char *me_path; /**< path to the DB files */
954 char *me_map; /**< the memory map of the data file */
955 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
956 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
957 MDB_txn *me_txn; /**< current write transaction */
958 size_t me_mapsize; /**< size of the data memory map */
959 off_t me_size; /**< current file size */
960 pgno_t me_maxpg; /**< me_mapsize / me_psize */
961 MDB_dbx *me_dbxs; /**< array of static DB info */
962 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
963 pthread_key_t me_txkey; /**< thread-key for readers */
964 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
965 # define me_pglast me_pgstate.mf_pglast
966 # define me_pghead me_pgstate.mf_pghead
967 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
968 /** IDL of pages that became unused in a write txn */
970 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
971 MDB_ID2L me_dirty_list;
972 /** Max number of freelist items that can fit in a single overflow page */
974 /** Max size of a node on a page */
975 unsigned int me_nodemax;
977 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
979 #elif defined(MDB_USE_POSIX_SEM)
980 sem_t *me_rmutex; /* Shared mutexes are not supported */
985 /** Nested transaction */
986 typedef struct MDB_ntxn {
987 MDB_txn mnt_txn; /* the transaction */
988 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
991 /** max number of pages to commit in one writev() call */
992 #define MDB_COMMIT_PAGES 64
993 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
994 #undef MDB_COMMIT_PAGES
995 #define MDB_COMMIT_PAGES IOV_MAX
998 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
999 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1000 static int mdb_page_touch(MDB_cursor *mc);
1002 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1003 static int mdb_page_search_root(MDB_cursor *mc,
1004 MDB_val *key, int modify);
1005 #define MDB_PS_MODIFY 1
1006 #define MDB_PS_ROOTONLY 2
1007 static int mdb_page_search(MDB_cursor *mc,
1008 MDB_val *key, int flags);
1009 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1011 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1012 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1013 pgno_t newpgno, unsigned int nflags);
1015 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1016 static int mdb_env_pick_meta(const MDB_env *env);
1017 static int mdb_env_write_meta(MDB_txn *txn);
1018 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1019 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1021 static void mdb_env_close0(MDB_env *env, int excl);
1023 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1024 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1025 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1026 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1027 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1028 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1029 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1030 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1031 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1033 static int mdb_rebalance(MDB_cursor *mc);
1034 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1036 static void mdb_cursor_pop(MDB_cursor *mc);
1037 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1039 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1040 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1041 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1042 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1043 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1045 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1046 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1048 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1049 static void mdb_xcursor_init0(MDB_cursor *mc);
1050 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1052 static int mdb_drop0(MDB_cursor *mc, int subs);
1053 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1056 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1060 static SECURITY_DESCRIPTOR mdb_null_sd;
1061 static SECURITY_ATTRIBUTES mdb_all_sa;
1062 static int mdb_sec_inited;
1065 /** Return the library version info. */
1067 mdb_version(int *major, int *minor, int *patch)
1069 if (major) *major = MDB_VERSION_MAJOR;
1070 if (minor) *minor = MDB_VERSION_MINOR;
1071 if (patch) *patch = MDB_VERSION_PATCH;
1072 return MDB_VERSION_STRING;
1075 /** Table of descriptions for MDB @ref errors */
1076 static char *const mdb_errstr[] = {
1077 "MDB_KEYEXIST: Key/data pair already exists",
1078 "MDB_NOTFOUND: No matching key/data pair found",
1079 "MDB_PAGE_NOTFOUND: Requested page not found",
1080 "MDB_CORRUPTED: Located page was wrong type",
1081 "MDB_PANIC: Update of meta page failed",
1082 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1083 "MDB_INVALID: File is not an MDB file",
1084 "MDB_MAP_FULL: Environment mapsize limit reached",
1085 "MDB_DBS_FULL: Environment maxdbs limit reached",
1086 "MDB_READERS_FULL: Environment maxreaders limit reached",
1087 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1088 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1089 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1090 "MDB_PAGE_FULL: Internal error - page has no more space",
1091 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1092 "MDB_INCOMPATIBLE: Database flags changed or would change",
1093 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1097 mdb_strerror(int err)
1101 return ("Successful return: 0");
1103 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1104 i = err - MDB_KEYEXIST;
1105 return mdb_errstr[i];
1108 return strerror(err);
1112 /** Display a key in hexadecimal and return the address of the result.
1113 * @param[in] key the key to display
1114 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1115 * @return The key in hexadecimal form.
1118 mdb_dkey(MDB_val *key, char *buf)
1121 unsigned char *c = key->mv_data;
1127 if (key->mv_size > MDB_MAXKEYSIZE)
1128 return "MDB_MAXKEYSIZE";
1129 /* may want to make this a dynamic check: if the key is mostly
1130 * printable characters, print it as-is instead of converting to hex.
1134 for (i=0; i<key->mv_size; i++)
1135 ptr += sprintf(ptr, "%02x", *c++);
1137 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1142 /** Display all the keys in the page. */
1144 mdb_page_list(MDB_page *mp)
1147 unsigned int i, nkeys, nsize;
1151 nkeys = NUMKEYS(mp);
1152 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1153 for (i=0; i<nkeys; i++) {
1154 node = NODEPTR(mp, i);
1155 key.mv_size = node->mn_ksize;
1156 key.mv_data = node->mn_data;
1157 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1158 if (IS_BRANCH(mp)) {
1159 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1162 if (F_ISSET(node->mn_flags, F_BIGDATA))
1163 nsize += sizeof(pgno_t);
1165 nsize += NODEDSZ(node);
1166 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1172 mdb_cursor_chk(MDB_cursor *mc)
1178 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1179 for (i=0; i<mc->mc_top; i++) {
1181 node = NODEPTR(mp, mc->mc_ki[i]);
1182 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1185 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1191 /** Count all the pages in each DB and in the freelist
1192 * and make sure it matches the actual number of pages
1195 static void mdb_audit(MDB_txn *txn)
1199 MDB_ID freecount, count;
1204 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1205 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1206 freecount += *(MDB_ID *)data.mv_data;
1209 for (i = 0; i<txn->mt_numdbs; i++) {
1211 mdb_cursor_init(&mc, txn, i, &mx);
1212 if (txn->mt_dbs[i].md_root == P_INVALID)
1214 count += txn->mt_dbs[i].md_branch_pages +
1215 txn->mt_dbs[i].md_leaf_pages +
1216 txn->mt_dbs[i].md_overflow_pages;
1217 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1218 mdb_page_search(&mc, NULL, 0);
1222 mp = mc.mc_pg[mc.mc_top];
1223 for (j=0; j<NUMKEYS(mp); j++) {
1224 MDB_node *leaf = NODEPTR(mp, j);
1225 if (leaf->mn_flags & F_SUBDATA) {
1227 memcpy(&db, NODEDATA(leaf), sizeof(db));
1228 count += db.md_branch_pages + db.md_leaf_pages +
1229 db.md_overflow_pages;
1233 while (mdb_cursor_sibling(&mc, 1) == 0);
1236 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1237 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1238 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1244 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1246 return txn->mt_dbxs[dbi].md_cmp(a, b);
1250 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1252 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1255 /** Allocate a page.
1256 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1259 mdb_page_malloc(MDB_txn *txn, unsigned num)
1261 MDB_env *env = txn->mt_env;
1262 MDB_page *ret = env->me_dpages;
1263 size_t sz = env->me_psize;
1266 VGMEMP_ALLOC(env, ret, sz);
1267 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1268 env->me_dpages = ret->mp_next;
1274 if ((ret = malloc(sz)) != NULL) {
1275 VGMEMP_ALLOC(env, ret, sz);
1280 /** Free a single page.
1281 * Saves single pages to a list, for future reuse.
1282 * (This is not used for multi-page overflow pages.)
1285 mdb_page_free(MDB_env *env, MDB_page *mp)
1287 mp->mp_next = env->me_dpages;
1288 VGMEMP_FREE(env, mp);
1289 env->me_dpages = mp;
1292 /* Free a dirty page */
1294 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1296 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1297 mdb_page_free(env, dp);
1299 /* large pages just get freed directly */
1300 VGMEMP_FREE(env, dp);
1305 /* Return all dirty pages to dpage list */
1307 mdb_dlist_free(MDB_txn *txn)
1309 MDB_env *env = txn->mt_env;
1310 MDB_ID2L dl = txn->mt_u.dirty_list;
1311 unsigned i, n = dl[0].mid;
1313 for (i = 1; i <= n; i++) {
1314 mdb_dpage_free(env, dl[i].mptr);
1319 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1321 mdb_find_oldest(MDB_txn *txn)
1324 txnid_t mr, oldest = txn->mt_txnid - 1;
1325 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1326 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1336 /** Allocate pages for writing.
1337 * If there are free pages available from older transactions, they
1338 * will be re-used first. Otherwise a new page will be allocated.
1339 * @param[in] mc cursor A cursor handle identifying the transaction and
1340 * database for which we are allocating.
1341 * @param[in] num the number of pages to allocate.
1342 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1343 * will always be satisfied by a single contiguous chunk of memory.
1344 * @return 0 on success, non-zero on failure.
1347 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1349 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1350 /* Get at most <Max_retries> more freeDB records once me_pghead
1351 * has enough pages. If not enough, use new pages from the map.
1352 * If <Paranoid> and mc is updating the freeDB, only get new
1353 * records if me_pghead is empty. Then the freelist cannot play
1354 * catch-up with itself by growing while trying to save it.
1356 enum { Paranoid = 1, Max_retries = 500 };
1358 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1360 int rc, n2 = num-1, retry = Max_retries;
1361 MDB_txn *txn = mc->mc_txn;
1362 MDB_env *env = txn->mt_env;
1363 pgno_t pgno, *mop = env->me_pghead;
1364 unsigned mop_len = mop ? mop[0] : 0;
1367 txnid_t oldest = 0, last;
1373 /* If our dirty list is already full, we can't do anything */
1374 if (txn->mt_dirty_room == 0)
1375 return MDB_TXN_FULL;
1377 for (op = MDB_FIRST;; op = MDB_NEXT) {
1378 unsigned int i, j, k;
1381 pgno_t *idl, old_id, new_id;
1383 /* Seek a big enough contiguous page range. Prefer
1384 * pages at the tail, just truncating the list.
1386 if (mop_len >= (unsigned)num) {
1390 if (mop[i-n2] == pgno+n2) {
1391 mop[0] = mop_len -= num;
1392 /* Move any stragglers down */
1393 for (j = i-n2; j <= mop_len; )
1394 mop[j++] = mop[++i];
1397 } while (--i >= (unsigned)num);
1398 if (Max_retries < INT_MAX && --retry < 0)
1402 if (op == MDB_FIRST) { /* 1st iteration */
1403 /* Prepare to fetch more and coalesce */
1404 oldest = mdb_find_oldest(txn);
1405 last = env->me_pglast;
1406 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1409 key.mv_data = &last; /* will loop up last+1 */
1410 key.mv_size = sizeof(last);
1412 if (Paranoid && mc->mc_dbi == FREE_DBI)
1415 if (Paranoid && retry < 0 && mop_len)
1419 /* Do not fetch more if the record will be too recent */
1422 rc = mdb_cursor_get(&m2, &key, NULL, op);
1424 if (rc == MDB_NOTFOUND)
1428 last = *(txnid_t*)key.mv_data;
1431 np = m2.mc_pg[m2.mc_top];
1432 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1433 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1436 idl = (MDB_ID *) data.mv_data;
1439 if (!(env->me_pghead = mop = mdb_midl_alloc(i)))
1442 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1444 mop = env->me_pghead;
1446 env->me_pglast = last;
1448 DPRINTF("IDL read txn %zu root %zu num %u",
1449 last, txn->mt_dbs[FREE_DBI].md_root, i);
1451 DPRINTF("IDL %zu", idl[k]);
1453 /* Merge in descending sorted order */
1456 mop[0] = (pgno_t)-1;
1460 for (; old_id < new_id; old_id = mop[--j])
1467 /* Use new pages from the map when nothing suitable in the freeDB */
1469 if (txn->mt_next_pgno + num >= env->me_maxpg) {
1470 DPUTS("DB size maxed out");
1471 return MDB_MAP_FULL;
1475 if (env->me_flags & MDB_WRITEMAP) {
1476 if (pgno == P_INVALID) {
1477 pgno = txn->mt_next_pgno;
1478 txn->mt_next_pgno += num;
1480 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
1482 if (!(np = mdb_page_malloc(txn, num)))
1484 if (pgno == P_INVALID) {
1485 pgno = txn->mt_next_pgno;
1486 txn->mt_next_pgno += num;
1489 mid.mid = np->mp_pgno = pgno;
1491 if (env->me_flags & MDB_WRITEMAP) {
1492 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1494 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1496 txn->mt_dirty_room--;
1502 /** Copy the used portions of a non-overflow page.
1503 * @param[in] dst page to copy into
1504 * @param[in] src page to copy from
1505 * @param[in] psize size of a page
1508 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1510 enum { Align = sizeof(pgno_t) };
1511 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1513 /* If page isn't full, just copy the used portion. Adjust
1514 * alignment so memcpy may copy words instead of bytes.
1516 if ((unused &= -Align) && !IS_LEAF2(src)) {
1518 memcpy(dst, src, (lower + (Align-1)) & -Align);
1519 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1522 memcpy(dst, src, psize - unused);
1526 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1527 * @param[in] mc cursor pointing to the page to be touched
1528 * @return 0 on success, non-zero on failure.
1531 mdb_page_touch(MDB_cursor *mc)
1533 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1534 MDB_txn *txn = mc->mc_txn;
1535 MDB_cursor *m2, *m3;
1540 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1541 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
1542 (rc = mdb_page_alloc(mc, 1, &np)))
1545 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi,mp->mp_pgno,pgno);
1546 assert(mp->mp_pgno != pgno);
1547 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
1548 /* Update the parent page, if any, to point to the new page */
1550 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1551 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1552 SETPGNO(node, pgno);
1554 mc->mc_db->md_root = pgno;
1556 } else if (txn->mt_parent && !IS_SUBP(mp)) {
1557 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
1559 /* If txn has a parent, make sure the page is in our
1563 unsigned x = mdb_mid2l_search(dl, pgno);
1564 if (x <= dl[0].mid && dl[x].mid == pgno) {
1567 mc->mc_pg[mc->mc_top] = np;
1571 assert(dl[0].mid < MDB_IDL_UM_MAX);
1573 np = mdb_page_malloc(txn, 1);
1578 mdb_mid2l_insert(dl, &mid);
1583 mdb_page_copy(np, mp, txn->mt_env->me_psize);
1585 np->mp_flags |= P_DIRTY;
1587 /* Adjust cursors pointing to mp */
1588 mc->mc_pg[mc->mc_top] = np;
1590 if (mc->mc_flags & C_SUB) {
1592 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1593 m3 = &m2->mc_xcursor->mx_cursor;
1594 if (m3->mc_snum < mc->mc_snum) continue;
1595 if (m3->mc_pg[mc->mc_top] == mp)
1596 m3->mc_pg[mc->mc_top] = np;
1599 for (m2 = txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1600 if (m2->mc_snum < mc->mc_snum) continue;
1601 if (m2->mc_pg[mc->mc_top] == mp) {
1602 m2->mc_pg[mc->mc_top] = np;
1603 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1604 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1606 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1607 if (!(leaf->mn_flags & F_SUBDATA))
1608 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1617 mdb_env_sync(MDB_env *env, int force)
1620 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1621 if (env->me_flags & MDB_WRITEMAP) {
1622 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1623 ? MS_ASYNC : MS_SYNC;
1624 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1627 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1631 if (MDB_FDATASYNC(env->me_fd))
1638 /** Make shadow copies of all of parent txn's cursors */
1640 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1642 MDB_cursor *mc, *m2;
1643 unsigned int i, j, size;
1645 for (i=0;i<src->mt_numdbs; i++) {
1646 if (src->mt_cursors[i]) {
1647 size = sizeof(MDB_cursor);
1648 if (src->mt_cursors[i]->mc_xcursor)
1649 size += sizeof(MDB_xcursor);
1650 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1657 mc->mc_db = &dst->mt_dbs[i];
1658 mc->mc_dbx = m2->mc_dbx;
1659 mc->mc_dbflag = &dst->mt_dbflags[i];
1660 mc->mc_snum = m2->mc_snum;
1661 mc->mc_top = m2->mc_top;
1662 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1663 for (j=0; j<mc->mc_snum; j++) {
1664 mc->mc_pg[j] = m2->mc_pg[j];
1665 mc->mc_ki[j] = m2->mc_ki[j];
1667 if (m2->mc_xcursor) {
1668 MDB_xcursor *mx, *mx2;
1669 mx = (MDB_xcursor *)(mc+1);
1670 mc->mc_xcursor = mx;
1671 mx2 = m2->mc_xcursor;
1672 mx->mx_db = mx2->mx_db;
1673 mx->mx_dbx = mx2->mx_dbx;
1674 mx->mx_dbflag = mx2->mx_dbflag;
1675 mx->mx_cursor.mc_txn = dst;
1676 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1677 mx->mx_cursor.mc_db = &mx->mx_db;
1678 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1679 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1680 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1681 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1682 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1683 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1684 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1685 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1688 mc->mc_xcursor = NULL;
1690 mc->mc_next = dst->mt_cursors[i];
1691 dst->mt_cursors[i] = mc;
1698 /** Close this write txn's cursors, after optionally merging its shadow
1699 * cursors back into parent's.
1700 * @param[in] txn the transaction handle.
1701 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1702 * @return 0 on success, non-zero on failure.
1705 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1707 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1710 for (i = txn->mt_numdbs; --i >= 0; ) {
1711 for (mc = cursors[i]; mc; mc = next) {
1713 if (mc->mc_flags & merge) {
1714 MDB_cursor *m2 = mc->mc_orig;
1715 m2->mc_snum = mc->mc_snum;
1716 m2->mc_top = mc->mc_top;
1717 for (j = mc->mc_snum; --j >= 0; ) {
1718 m2->mc_pg[j] = mc->mc_pg[j];
1719 m2->mc_ki[j] = mc->mc_ki[j];
1722 if (mc->mc_flags & C_ALLOCD)
1730 mdb_txn_reset0(MDB_txn *txn);
1732 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1733 * @param[in] txn the transaction handle to initialize
1734 * @return 0 on success, non-zero on failure.
1737 mdb_txn_renew0(MDB_txn *txn)
1739 MDB_env *env = txn->mt_env;
1742 int rc, new_notls = 0;
1745 txn->mt_numdbs = env->me_numdbs;
1746 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1748 if (txn->mt_flags & MDB_TXN_RDONLY) {
1749 if (!env->me_txns) {
1750 i = mdb_env_pick_meta(env);
1751 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1752 txn->mt_u.reader = NULL;
1754 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1755 pthread_getspecific(env->me_txkey);
1757 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1758 return MDB_BAD_RSLOT;
1760 pid_t pid = env->me_pid;
1761 pthread_t tid = pthread_self();
1764 for (i=0; i<env->me_txns->mti_numreaders; i++)
1765 if (env->me_txns->mti_readers[i].mr_pid == 0)
1767 if (i == env->me_maxreaders) {
1768 UNLOCK_MUTEX_R(env);
1769 return MDB_READERS_FULL;
1771 env->me_txns->mti_readers[i].mr_pid = pid;
1772 env->me_txns->mti_readers[i].mr_tid = tid;
1773 if (i >= env->me_txns->mti_numreaders)
1774 env->me_txns->mti_numreaders = i+1;
1775 /* Save numreaders for un-mutexed mdb_env_close() */
1776 env->me_numreaders = env->me_txns->mti_numreaders;
1777 UNLOCK_MUTEX_R(env);
1778 r = &env->me_txns->mti_readers[i];
1779 new_notls = (env->me_flags & MDB_NOTLS);
1780 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1785 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1786 txn->mt_u.reader = r;
1788 txn->mt_toggle = txn->mt_txnid & 1;
1789 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1793 txn->mt_txnid = env->me_txns->mti_txnid;
1794 txn->mt_toggle = txn->mt_txnid & 1;
1795 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1798 if (txn->mt_txnid == mdb_debug_start)
1801 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1802 txn->mt_u.dirty_list = env->me_dirty_list;
1803 txn->mt_u.dirty_list[0].mid = 0;
1804 txn->mt_free_pgs = env->me_free_pgs;
1805 txn->mt_free_pgs[0] = 0;
1809 /* Copy the DB info and flags */
1810 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1811 for (i=2; i<txn->mt_numdbs; i++) {
1812 x = env->me_dbflags[i];
1813 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1814 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1816 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1818 if (env->me_maxpg < txn->mt_next_pgno) {
1819 mdb_txn_reset0(txn);
1821 txn->mt_u.reader->mr_pid = 0;
1822 txn->mt_u.reader = NULL;
1824 return MDB_MAP_RESIZED;
1831 mdb_txn_renew(MDB_txn *txn)
1835 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1838 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1839 DPUTS("environment had fatal error, must shutdown!");
1843 rc = mdb_txn_renew0(txn);
1844 if (rc == MDB_SUCCESS) {
1845 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1846 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1847 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1853 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1857 int rc, size, tsize = sizeof(MDB_txn);
1859 if (env->me_flags & MDB_FATAL_ERROR) {
1860 DPUTS("environment had fatal error, must shutdown!");
1863 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1866 /* Nested transactions: Max 1 child, write txns only, no writemap */
1867 if (parent->mt_child ||
1868 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1869 (env->me_flags & MDB_WRITEMAP))
1873 tsize = sizeof(MDB_ntxn);
1875 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1876 if (!(flags & MDB_RDONLY))
1877 size += env->me_maxdbs * sizeof(MDB_cursor *);
1879 if ((txn = calloc(1, size)) == NULL) {
1880 DPRINTF("calloc: %s", strerror(ErrCode()));
1883 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1884 if (flags & MDB_RDONLY) {
1885 txn->mt_flags |= MDB_TXN_RDONLY;
1886 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1888 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1889 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1895 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1896 if (!txn->mt_u.dirty_list ||
1897 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1899 free(txn->mt_u.dirty_list);
1903 txn->mt_txnid = parent->mt_txnid;
1904 txn->mt_toggle = parent->mt_toggle;
1905 txn->mt_dirty_room = parent->mt_dirty_room;
1906 txn->mt_u.dirty_list[0].mid = 0;
1907 txn->mt_next_pgno = parent->mt_next_pgno;
1908 parent->mt_child = txn;
1909 txn->mt_parent = parent;
1910 txn->mt_numdbs = parent->mt_numdbs;
1911 txn->mt_flags = parent->mt_flags;
1912 txn->mt_dbxs = parent->mt_dbxs;
1913 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1914 /* Copy parent's mt_dbflags, but clear DB_NEW */
1915 for (i=0; i<txn->mt_numdbs; i++)
1916 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1918 ntxn = (MDB_ntxn *)txn;
1919 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1920 if (env->me_pghead) {
1921 size = MDB_IDL_SIZEOF(env->me_pghead);
1922 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1924 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1929 rc = mdb_cursor_shadow(parent, txn);
1931 mdb_txn_reset0(txn);
1933 rc = mdb_txn_renew0(txn);
1939 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1940 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1941 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1947 /** Export or close DBI handles opened in this txn. */
1949 mdb_dbis_update(MDB_txn *txn, int keep)
1952 MDB_dbi n = txn->mt_numdbs;
1953 MDB_env *env = txn->mt_env;
1954 unsigned char *tdbflags = txn->mt_dbflags;
1956 for (i = n; --i >= 2;) {
1957 if (tdbflags[i] & DB_NEW) {
1959 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
1961 char *ptr = env->me_dbxs[i].md_name.mv_data;
1962 env->me_dbxs[i].md_name.mv_data = NULL;
1963 env->me_dbxs[i].md_name.mv_size = 0;
1964 env->me_dbflags[i] = 0;
1969 if (keep && env->me_numdbs < n)
1973 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1974 * May be called twice for readonly txns: First reset it, then abort.
1975 * @param[in] txn the transaction handle to reset
1978 mdb_txn_reset0(MDB_txn *txn)
1980 MDB_env *env = txn->mt_env;
1982 /* Close any DBI handles opened in this txn */
1983 mdb_dbis_update(txn, 0);
1985 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1986 if (txn->mt_u.reader) {
1987 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1988 if (!(env->me_flags & MDB_NOTLS))
1989 txn->mt_u.reader = NULL; /* txn does not own reader */
1991 txn->mt_numdbs = 0; /* close nothing if called again */
1992 txn->mt_dbxs = NULL; /* mark txn as reset */
1994 mdb_cursors_close(txn, 0);
1996 if (!(env->me_flags & MDB_WRITEMAP)) {
1997 mdb_dlist_free(txn);
1999 mdb_midl_free(env->me_pghead);
2001 if (txn->mt_parent) {
2002 txn->mt_parent->mt_child = NULL;
2003 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2004 mdb_midl_free(txn->mt_free_pgs);
2005 free(txn->mt_u.dirty_list);
2009 if (mdb_midl_shrink(&txn->mt_free_pgs))
2010 env->me_free_pgs = txn->mt_free_pgs;
2011 env->me_pghead = NULL;
2015 /* The writer mutex was locked in mdb_txn_begin. */
2016 UNLOCK_MUTEX_W(env);
2021 mdb_txn_reset(MDB_txn *txn)
2026 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2027 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2028 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2030 /* This call is only valid for read-only txns */
2031 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2034 mdb_txn_reset0(txn);
2038 mdb_txn_abort(MDB_txn *txn)
2043 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2044 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2045 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2048 mdb_txn_abort(txn->mt_child);
2050 mdb_txn_reset0(txn);
2051 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2052 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2053 txn->mt_u.reader->mr_pid = 0;
2058 /** Save the freelist as of this transaction to the freeDB.
2059 * This changes the freelist. Keep trying until it stabilizes.
2062 mdb_freelist_save(MDB_txn *txn)
2064 /* env->me_pghead[] can grow and shrink during this call.
2065 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2066 * Page numbers cannot disappear from txn->mt_free_pgs[].
2069 MDB_env *env = txn->mt_env;
2070 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2071 txnid_t pglast = 0, head_id = 0;
2072 pgno_t freecnt = 0, *free_pgs, *mop;
2073 ssize_t head_room = 0, total_room = 0, mop_len;
2075 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2077 if (env->me_pghead) {
2078 /* Make sure first page of freeDB is touched and on freelist */
2079 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2080 if (rc && rc != MDB_NOTFOUND)
2085 /* Come back here after each Put() in case freelist changed */
2088 /* If using records from freeDB which we have not yet
2089 * deleted, delete them and any we reserved for me_pghead.
2091 while (pglast < env->me_pglast) {
2092 rc = mdb_cursor_first(&mc, &key, NULL);
2095 pglast = head_id = *(txnid_t *)key.mv_data;
2096 total_room = head_room = 0;
2097 assert(pglast <= env->me_pglast);
2098 rc = mdb_cursor_del(&mc, 0);
2103 /* Save the IDL of pages freed by this txn, to a single record */
2104 if (freecnt < txn->mt_free_pgs[0]) {
2106 /* Make sure last page of freeDB is touched and on freelist */
2107 key.mv_size = MDB_MAXKEYSIZE+1;
2109 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2110 if (rc && rc != MDB_NOTFOUND)
2113 free_pgs = txn->mt_free_pgs;
2114 /* Write to last page of freeDB */
2115 key.mv_size = sizeof(txn->mt_txnid);
2116 key.mv_data = &txn->mt_txnid;
2118 freecnt = free_pgs[0];
2119 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2120 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2123 /* Retry if mt_free_pgs[] grew during the Put() */
2124 free_pgs = txn->mt_free_pgs;
2125 } while (freecnt < free_pgs[0]);
2126 mdb_midl_sort(free_pgs);
2127 memcpy(data.mv_data, free_pgs, data.mv_size);
2130 unsigned int i = free_pgs[0];
2131 DPRINTF("IDL write txn %zu root %zu num %u",
2132 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2134 DPRINTF("IDL %zu", free_pgs[i]);
2140 mop = env->me_pghead;
2141 mop_len = mop ? mop[0] : 0;
2143 /* Reserve records for me_pghead[]. Split it if multi-page,
2144 * to avoid searching freeDB for a page range. Use keys in
2145 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2147 if (total_room >= mop_len) {
2148 if (total_room == mop_len || --more < 0)
2150 } else if (head_room >= maxfree_1pg && head_id > 1) {
2151 /* Keep current record (overflow page), add a new one */
2155 /* (Re)write {key = head_id, IDL length = head_room} */
2156 total_room -= head_room;
2157 head_room = mop_len - total_room;
2158 if (head_room > maxfree_1pg && head_id > 1) {
2159 /* Overflow multi-page for part of me_pghead */
2160 head_room /= head_id; /* amortize page sizes */
2161 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2162 } else if (head_room < 0) {
2163 /* Rare case, not bothering to delete this record */
2166 key.mv_size = sizeof(head_id);
2167 key.mv_data = &head_id;
2168 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2169 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2172 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2173 total_room += head_room;
2176 /* Fill in the reserved, touched me_pghead records. Avoid write ops
2177 * so they cannot rearrange anything, just read the destinations.
2184 rc = mdb_cursor_first(&mc, &key, &data);
2185 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2186 MDB_IDL dest = data.mv_data;
2187 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2189 assert(len >= 0 && *(txnid_t*)key.mv_data <= env->me_pglast);
2193 memcpy(dest, mop -= len, len * sizeof(MDB_ID));
2194 if (! (mop_len -= len))
2201 /** Flush dirty pages to the map, after clearing their dirty flag.
2204 mdb_page_flush(MDB_txn *txn)
2206 MDB_env *env = txn->mt_env;
2207 MDB_ID2L dl = txn->mt_u.dirty_list;
2208 unsigned psize = env->me_psize;
2209 int i, pagecount = dl[0].mid, rc;
2210 size_t size, pos = 0;
2215 memset(&ov, 0, sizeof(ov));
2217 struct iovec iov[MDB_COMMIT_PAGES];
2218 ssize_t wpos, wsize, wres;
2219 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2223 if (env->me_flags & MDB_WRITEMAP) {
2224 /* Clear dirty flags */
2225 for (i = pagecount; i; i--) {
2227 dp->mp_flags &= ~P_DIRTY;
2233 /* Write the pages */
2235 if (i <= pagecount) {
2238 /* clear dirty flag */
2239 dp->mp_flags &= ~P_DIRTY;
2242 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2247 /* Windows actually supports scatter/gather I/O, but only on
2248 * unbuffered file handles. Since we're relying on the OS page
2249 * cache for all our data, that's self-defeating. So we just
2250 * write pages one at a time. We use the ov structure to set
2251 * the write offset, to at least save the overhead of a Seek
2254 DPRINTF("committing page %zu", pgno);
2255 ov.Offset = pos & 0xffffffff;
2256 ov.OffsetHigh = pos >> 16;
2257 ov.OffsetHigh >>= 16;
2258 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2260 DPRINTF("WriteFile: %d", rc);
2264 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
2265 if (pos != next_pos || n == MDB_COMMIT_PAGES) {
2267 /* Write previous page(s) */
2268 lseek(env->me_fd, wpos, SEEK_SET);
2269 wres = writev(env->me_fd, iov, n);
2270 if (wres != wsize) {
2273 DPRINTF("writev: %s", strerror(rc));
2275 DPUTS("short write, filesystem full?");
2286 DPRINTF("committing page %zu", pgno);
2287 next_pos = pos + size;
2288 iov[n].iov_len = size;
2289 iov[n].iov_base = (char *)dp;
2295 mdb_dlist_free(txn);
2301 mdb_txn_commit(MDB_txn *txn)
2307 assert(txn != NULL);
2308 assert(txn->mt_env != NULL);
2310 if (txn->mt_child) {
2311 rc = mdb_txn_commit(txn->mt_child);
2312 txn->mt_child = NULL;
2319 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2320 mdb_dbis_update(txn, 1);
2321 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2326 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2327 DPUTS("error flag is set, can't commit");
2329 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2334 if (txn->mt_parent) {
2335 MDB_txn *parent = txn->mt_parent;
2339 /* Append our free list to parent's */
2340 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
2343 mdb_midl_free(txn->mt_free_pgs);
2345 parent->mt_next_pgno = txn->mt_next_pgno;
2346 parent->mt_flags = txn->mt_flags;
2348 /* Merge our cursors into parent's and close them */
2349 mdb_cursors_close(txn, C_SHADOW);
2351 /* Update parent's DB table. */
2352 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2353 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2354 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2355 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2356 for (i=2; i<txn->mt_numdbs; i++) {
2357 /* preserve parent's DB_NEW status */
2358 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2359 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2362 dst = txn->mt_parent->mt_u.dirty_list;
2363 src = txn->mt_u.dirty_list;
2364 /* Find len = length of merging our dirty list with parent's */
2366 dst[0].mid = 0; /* simplify loops */
2367 if (parent->mt_parent) {
2368 len = x + src[0].mid;
2369 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2370 for (i = x; y && i; y--) {
2371 pgno_t yp = src[y].mid;
2372 while (yp < dst[i].mid)
2374 if (yp == dst[i].mid) {
2379 } else { /* Simplify the above for single-ancestor case */
2380 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2382 /* Merge our dirty list with parent's */
2384 for (i = len; y; dst[i--] = src[y--]) {
2385 pgno_t yp = src[y].mid;
2386 while (yp < dst[x].mid)
2387 dst[i--] = dst[x--];
2388 if (yp == dst[x].mid)
2389 free(dst[x--].mptr);
2393 free(txn->mt_u.dirty_list);
2394 parent->mt_dirty_room = txn->mt_dirty_room;
2396 txn->mt_parent->mt_child = NULL;
2397 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2402 if (txn != env->me_txn) {
2403 DPUTS("attempt to commit unknown transaction");
2408 mdb_cursors_close(txn, 0);
2410 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2413 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2414 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2416 /* Update DB root pointers */
2417 if (txn->mt_numdbs > 2) {
2421 data.mv_size = sizeof(MDB_db);
2423 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2424 for (i = 2; i < txn->mt_numdbs; i++) {
2425 if (txn->mt_dbflags[i] & DB_DIRTY) {
2426 data.mv_data = &txn->mt_dbs[i];
2427 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2434 rc = mdb_freelist_save(txn);
2438 mdb_midl_free(env->me_pghead);
2439 env->me_pghead = NULL;
2440 if (mdb_midl_shrink(&txn->mt_free_pgs))
2441 env->me_free_pgs = txn->mt_free_pgs;
2447 if ((rc = mdb_page_flush(txn)) ||
2448 (rc = mdb_env_sync(env, 0)) ||
2449 (rc = mdb_env_write_meta(txn)))
2455 mdb_dbis_update(txn, 1);
2457 UNLOCK_MUTEX_W(env);
2467 /** Read the environment parameters of a DB environment before
2468 * mapping it into memory.
2469 * @param[in] env the environment handle
2470 * @param[out] meta address of where to store the meta information
2471 * @return 0 on success, non-zero on failure.
2474 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2481 /* We don't know the page size yet, so use a minimum value.
2482 * Read both meta pages so we can use the latest one.
2485 for (i=0; i<2; i++) {
2487 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2489 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2494 else if (rc != MDB_PAGESIZE) {
2498 DPRINTF("read: %s", strerror(err));
2502 p = (MDB_page *)&pbuf;
2504 if (!F_ISSET(p->mp_flags, P_META)) {
2505 DPRINTF("page %zu not a meta page", p->mp_pgno);
2510 if (m->mm_magic != MDB_MAGIC) {
2511 DPUTS("meta has invalid magic");
2515 if (m->mm_version != MDB_VERSION) {
2516 DPRINTF("database is version %u, expected version %u",
2517 m->mm_version, MDB_VERSION);
2518 return MDB_VERSION_MISMATCH;
2522 if (m->mm_txnid > meta->mm_txnid)
2523 memcpy(meta, m, sizeof(*m));
2525 memcpy(meta, m, sizeof(*m));
2527 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2529 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2537 /** Write the environment parameters of a freshly created DB environment.
2538 * @param[in] env the environment handle
2539 * @param[out] meta address of where to store the meta information
2540 * @return 0 on success, non-zero on failure.
2543 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2550 DPUTS("writing new meta page");
2552 GET_PAGESIZE(psize);
2554 meta->mm_magic = MDB_MAGIC;
2555 meta->mm_version = MDB_VERSION;
2556 meta->mm_mapsize = env->me_mapsize;
2557 meta->mm_psize = psize;
2558 meta->mm_last_pg = 1;
2559 meta->mm_flags = env->me_flags & 0xffff;
2560 meta->mm_flags |= MDB_INTEGERKEY;
2561 meta->mm_dbs[0].md_root = P_INVALID;
2562 meta->mm_dbs[1].md_root = P_INVALID;
2564 p = calloc(2, psize);
2566 p->mp_flags = P_META;
2569 memcpy(m, meta, sizeof(*meta));
2571 q = (MDB_page *)((char *)p + psize);
2574 q->mp_flags = P_META;
2577 memcpy(m, meta, sizeof(*meta));
2582 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2583 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2584 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2587 lseek(env->me_fd, 0, SEEK_SET);
2588 rc = write(env->me_fd, p, psize * 2);
2589 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2595 /** Update the environment info to commit a transaction.
2596 * @param[in] txn the transaction that's being committed
2597 * @return 0 on success, non-zero on failure.
2600 mdb_env_write_meta(MDB_txn *txn)
2603 MDB_meta meta, metab, *mp;
2605 int rc, len, toggle;
2612 assert(txn != NULL);
2613 assert(txn->mt_env != NULL);
2615 toggle = !txn->mt_toggle;
2616 DPRINTF("writing meta page %d for root page %zu",
2617 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2620 mp = env->me_metas[toggle];
2622 if (env->me_flags & MDB_WRITEMAP) {
2623 /* Persist any increases of mapsize config */
2624 if (env->me_mapsize > mp->mm_mapsize)
2625 mp->mm_mapsize = env->me_mapsize;
2626 mp->mm_dbs[0] = txn->mt_dbs[0];
2627 mp->mm_dbs[1] = txn->mt_dbs[1];
2628 mp->mm_last_pg = txn->mt_next_pgno - 1;
2629 mp->mm_txnid = txn->mt_txnid;
2630 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2631 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2634 ptr += env->me_psize;
2635 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2642 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2643 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2645 ptr = (char *)&meta;
2646 if (env->me_mapsize > mp->mm_mapsize) {
2647 /* Persist any increases of mapsize config */
2648 meta.mm_mapsize = env->me_mapsize;
2649 off = offsetof(MDB_meta, mm_mapsize);
2651 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2653 len = sizeof(MDB_meta) - off;
2656 meta.mm_dbs[0] = txn->mt_dbs[0];
2657 meta.mm_dbs[1] = txn->mt_dbs[1];
2658 meta.mm_last_pg = txn->mt_next_pgno - 1;
2659 meta.mm_txnid = txn->mt_txnid;
2662 off += env->me_psize;
2665 /* Write to the SYNC fd */
2666 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2667 env->me_fd : env->me_mfd;
2670 memset(&ov, 0, sizeof(ov));
2672 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2675 rc = pwrite(mfd, ptr, len, off);
2680 DPUTS("write failed, disk error?");
2681 /* On a failure, the pagecache still contains the new data.
2682 * Write some old data back, to prevent it from being used.
2683 * Use the non-SYNC fd; we know it will fail anyway.
2685 meta.mm_last_pg = metab.mm_last_pg;
2686 meta.mm_txnid = metab.mm_txnid;
2688 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2690 r2 = pwrite(env->me_fd, ptr, len, off);
2693 env->me_flags |= MDB_FATAL_ERROR;
2697 /* Memory ordering issues are irrelevant; since the entire writer
2698 * is wrapped by wmutex, all of these changes will become visible
2699 * after the wmutex is unlocked. Since the DB is multi-version,
2700 * readers will get consistent data regardless of how fresh or
2701 * how stale their view of these values is.
2703 env->me_txns->mti_txnid = txn->mt_txnid;
2708 /** Check both meta pages to see which one is newer.
2709 * @param[in] env the environment handle
2710 * @return meta toggle (0 or 1).
2713 mdb_env_pick_meta(const MDB_env *env)
2715 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2719 mdb_env_create(MDB_env **env)
2723 e = calloc(1, sizeof(MDB_env));
2727 e->me_maxreaders = DEFAULT_READERS;
2728 e->me_maxdbs = e->me_numdbs = 2;
2729 e->me_fd = INVALID_HANDLE_VALUE;
2730 e->me_lfd = INVALID_HANDLE_VALUE;
2731 e->me_mfd = INVALID_HANDLE_VALUE;
2732 #ifdef MDB_USE_POSIX_SEM
2733 e->me_rmutex = SEM_FAILED;
2734 e->me_wmutex = SEM_FAILED;
2736 e->me_pid = getpid();
2737 VGMEMP_CREATE(e,0,0);
2743 mdb_env_set_mapsize(MDB_env *env, size_t size)
2747 env->me_mapsize = size;
2749 env->me_maxpg = env->me_mapsize / env->me_psize;
2754 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2758 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2763 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2765 if (env->me_map || readers < 1)
2767 env->me_maxreaders = readers;
2772 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2774 if (!env || !readers)
2776 *readers = env->me_maxreaders;
2780 /** Further setup required for opening an MDB environment
2783 mdb_env_open2(MDB_env *env)
2785 unsigned int flags = env->me_flags;
2786 int i, newenv = 0, prot;
2790 memset(&meta, 0, sizeof(meta));
2792 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2795 DPUTS("new mdbenv");
2799 /* Was a mapsize configured? */
2800 if (!env->me_mapsize) {
2801 /* If this is a new environment, take the default,
2802 * else use the size recorded in the existing env.
2804 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2805 } else if (env->me_mapsize < meta.mm_mapsize) {
2806 /* If the configured size is smaller, make sure it's
2807 * still big enough. Silently round up to minimum if not.
2809 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2810 if (env->me_mapsize < minsize)
2811 env->me_mapsize = minsize;
2817 LONG sizelo, sizehi;
2818 sizelo = env->me_mapsize & 0xffffffff;
2819 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2821 /* Windows won't create mappings for zero length files.
2822 * Just allocate the maxsize right now.
2825 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2826 if (!SetEndOfFile(env->me_fd))
2828 SetFilePointer(env->me_fd, 0, NULL, 0);
2830 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2831 PAGE_READWRITE : PAGE_READONLY,
2832 sizehi, sizelo, NULL);
2835 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2836 FILE_MAP_WRITE : FILE_MAP_READ,
2837 0, 0, env->me_mapsize, meta.mm_address);
2845 if (flags & MDB_WRITEMAP) {
2847 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2850 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2852 if (env->me_map == MAP_FAILED) {
2856 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2858 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2860 #ifdef POSIX_MADV_RANDOM
2861 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2862 #endif /* POSIX_MADV_RANDOM */
2863 #endif /* MADV_RANDOM */
2867 if (flags & MDB_FIXEDMAP)
2868 meta.mm_address = env->me_map;
2869 i = mdb_env_init_meta(env, &meta);
2870 if (i != MDB_SUCCESS) {
2873 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2874 /* Can happen because the address argument to mmap() is just a
2875 * hint. mmap() can pick another, e.g. if the range is in use.
2876 * The MAP_FIXED flag would prevent that, but then mmap could
2877 * instead unmap existing pages to make room for the new map.
2879 return EBUSY; /* TODO: Make a new MDB_* error code? */
2881 env->me_psize = meta.mm_psize;
2882 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2883 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2885 env->me_maxpg = env->me_mapsize / env->me_psize;
2887 p = (MDB_page *)env->me_map;
2888 env->me_metas[0] = METADATA(p);
2889 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2893 int toggle = mdb_env_pick_meta(env);
2894 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2896 DPRINTF("opened database version %u, pagesize %u",
2897 env->me_metas[0]->mm_version, env->me_psize);
2898 DPRINTF("using meta page %d", toggle);
2899 DPRINTF("depth: %u", db->md_depth);
2900 DPRINTF("entries: %zu", db->md_entries);
2901 DPRINTF("branch pages: %zu", db->md_branch_pages);
2902 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2903 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2904 DPRINTF("root: %zu", db->md_root);
2912 /** Release a reader thread's slot in the reader lock table.
2913 * This function is called automatically when a thread exits.
2914 * @param[in] ptr This points to the slot in the reader lock table.
2917 mdb_env_reader_dest(void *ptr)
2919 MDB_reader *reader = ptr;
2925 /** Junk for arranging thread-specific callbacks on Windows. This is
2926 * necessarily platform and compiler-specific. Windows supports up
2927 * to 1088 keys. Let's assume nobody opens more than 64 environments
2928 * in a single process, for now. They can override this if needed.
2930 #ifndef MAX_TLS_KEYS
2931 #define MAX_TLS_KEYS 64
2933 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2934 static int mdb_tls_nkeys;
2936 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2940 case DLL_PROCESS_ATTACH: break;
2941 case DLL_THREAD_ATTACH: break;
2942 case DLL_THREAD_DETACH:
2943 for (i=0; i<mdb_tls_nkeys; i++) {
2944 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2945 mdb_env_reader_dest(r);
2948 case DLL_PROCESS_DETACH: break;
2953 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2955 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2959 /* Force some symbol references.
2960 * _tls_used forces the linker to create the TLS directory if not already done
2961 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2963 #pragma comment(linker, "/INCLUDE:_tls_used")
2964 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2965 #pragma const_seg(".CRT$XLB")
2966 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2967 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2970 #pragma comment(linker, "/INCLUDE:__tls_used")
2971 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2972 #pragma data_seg(".CRT$XLB")
2973 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2975 #endif /* WIN 32/64 */
2976 #endif /* !__GNUC__ */
2979 /** Downgrade the exclusive lock on the region back to shared */
2981 mdb_env_share_locks(MDB_env *env, int *excl)
2983 int rc = 0, toggle = mdb_env_pick_meta(env);
2985 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2990 /* First acquire a shared lock. The Unlock will
2991 * then release the existing exclusive lock.
2993 memset(&ov, 0, sizeof(ov));
2994 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2997 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3003 struct flock lock_info;
3004 /* The shared lock replaces the existing lock */
3005 memset((void *)&lock_info, 0, sizeof(lock_info));
3006 lock_info.l_type = F_RDLCK;
3007 lock_info.l_whence = SEEK_SET;
3008 lock_info.l_start = 0;
3009 lock_info.l_len = 1;
3010 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3011 (rc = ErrCode()) == EINTR) ;
3012 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3019 /** Try to get exlusive lock, otherwise shared.
3020 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3023 mdb_env_excl_lock(MDB_env *env, int *excl)
3027 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3031 memset(&ov, 0, sizeof(ov));
3032 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3039 struct flock lock_info;
3040 memset((void *)&lock_info, 0, sizeof(lock_info));
3041 lock_info.l_type = F_WRLCK;
3042 lock_info.l_whence = SEEK_SET;
3043 lock_info.l_start = 0;
3044 lock_info.l_len = 1;
3045 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3046 (rc = ErrCode()) == EINTR) ;
3050 # ifdef MDB_USE_POSIX_SEM
3051 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3054 lock_info.l_type = F_RDLCK;
3055 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3056 (rc = ErrCode()) == EINTR) ;
3064 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3066 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3068 * @(#) $Revision: 5.1 $
3069 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3070 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3072 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3076 * Please do not copyright this code. This code is in the public domain.
3078 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3079 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3080 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3081 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3082 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3083 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3084 * PERFORMANCE OF THIS SOFTWARE.
3087 * chongo <Landon Curt Noll> /\oo/\
3088 * http://www.isthe.com/chongo/
3090 * Share and Enjoy! :-)
3093 typedef unsigned long long mdb_hash_t;
3094 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3096 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3097 * @param[in] str string to hash
3098 * @param[in] hval initial value for hash
3099 * @return 64 bit hash
3101 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3102 * hval arg on the first call.
3105 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3107 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3108 unsigned char *end = s + val->mv_size;
3110 * FNV-1a hash each octet of the string
3113 /* xor the bottom with the current octet */
3114 hval ^= (mdb_hash_t)*s++;
3116 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3117 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3118 (hval << 7) + (hval << 8) + (hval << 40);
3120 /* return our new hash value */
3124 /** Hash the string and output the hash in hex.
3125 * @param[in] str string to hash
3126 * @param[out] hexbuf an array of 17 chars to hold the hash
3129 mdb_hash_hex(MDB_val *val, char *hexbuf)
3132 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3133 for (i=0; i<8; i++) {
3134 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3140 /** Open and/or initialize the lock region for the environment.
3141 * @param[in] env The MDB environment.
3142 * @param[in] lpath The pathname of the file used for the lock region.
3143 * @param[in] mode The Unix permissions for the file, if we create it.
3144 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3145 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3146 * @return 0 on success, non-zero on failure.
3149 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3152 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3154 # define MDB_ERRCODE_ROFS EROFS
3155 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3156 # define MDB_CLOEXEC O_CLOEXEC
3159 # define MDB_CLOEXEC 0
3166 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3167 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3168 FILE_ATTRIBUTE_NORMAL, NULL);
3170 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3172 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3174 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3179 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3180 /* Lose record locks when exec*() */
3181 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3182 fcntl(env->me_lfd, F_SETFD, fdflags);
3185 if (!(env->me_flags & MDB_NOTLS)) {
3186 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3189 env->me_flags |= MDB_ENV_TXKEY;
3191 /* Windows TLS callbacks need help finding their TLS info. */
3192 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3196 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3200 /* Try to get exclusive lock. If we succeed, then
3201 * nobody is using the lock region and we should initialize it.
3203 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3206 size = GetFileSize(env->me_lfd, NULL);
3208 size = lseek(env->me_lfd, 0, SEEK_END);
3210 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3211 if (size < rsize && *excl > 0) {
3213 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3214 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3216 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3220 size = rsize - sizeof(MDB_txninfo);
3221 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3226 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3228 if (!mh) goto fail_errno;
3229 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3231 if (!env->me_txns) goto fail_errno;
3233 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3235 if (m == MAP_FAILED) goto fail_errno;
3241 BY_HANDLE_FILE_INFORMATION stbuf;
3250 if (!mdb_sec_inited) {
3251 InitializeSecurityDescriptor(&mdb_null_sd,
3252 SECURITY_DESCRIPTOR_REVISION);
3253 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3254 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3255 mdb_all_sa.bInheritHandle = FALSE;
3256 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3259 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3260 idbuf.volume = stbuf.dwVolumeSerialNumber;
3261 idbuf.nhigh = stbuf.nFileIndexHigh;
3262 idbuf.nlow = stbuf.nFileIndexLow;
3263 val.mv_data = &idbuf;
3264 val.mv_size = sizeof(idbuf);
3265 mdb_hash_hex(&val, hexbuf);
3266 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3267 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3268 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3269 if (!env->me_rmutex) goto fail_errno;
3270 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3271 if (!env->me_wmutex) goto fail_errno;
3272 #elif defined(MDB_USE_POSIX_SEM)
3281 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3282 idbuf.dev = stbuf.st_dev;
3283 idbuf.ino = stbuf.st_ino;
3284 val.mv_data = &idbuf;
3285 val.mv_size = sizeof(idbuf);
3286 mdb_hash_hex(&val, hexbuf);
3287 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3288 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3289 /* Clean up after a previous run, if needed: Try to
3290 * remove both semaphores before doing anything else.
3292 sem_unlink(env->me_txns->mti_rmname);
3293 sem_unlink(env->me_txns->mti_wmname);
3294 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3295 O_CREAT|O_EXCL, mode, 1);
3296 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3297 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3298 O_CREAT|O_EXCL, mode, 1);
3299 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3300 #else /* MDB_USE_POSIX_SEM */
3301 pthread_mutexattr_t mattr;
3303 if ((rc = pthread_mutexattr_init(&mattr))
3304 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3305 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3306 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3308 pthread_mutexattr_destroy(&mattr);
3309 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3311 env->me_txns->mti_version = MDB_VERSION;
3312 env->me_txns->mti_magic = MDB_MAGIC;
3313 env->me_txns->mti_txnid = 0;
3314 env->me_txns->mti_numreaders = 0;
3317 if (env->me_txns->mti_magic != MDB_MAGIC) {
3318 DPUTS("lock region has invalid magic");
3322 if (env->me_txns->mti_version != MDB_VERSION) {
3323 DPRINTF("lock region is version %u, expected version %u",
3324 env->me_txns->mti_version, MDB_VERSION);
3325 rc = MDB_VERSION_MISMATCH;
3329 if (rc != EACCES && rc != EAGAIN) {
3333 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3334 if (!env->me_rmutex) goto fail_errno;
3335 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3336 if (!env->me_wmutex) goto fail_errno;
3337 #elif defined(MDB_USE_POSIX_SEM)
3338 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3339 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3340 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3341 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3352 /** The name of the lock file in the DB environment */
3353 #define LOCKNAME "/lock.mdb"
3354 /** The name of the data file in the DB environment */
3355 #define DATANAME "/data.mdb"
3356 /** The suffix of the lock file when no subdir is used */
3357 #define LOCKSUFF "-lock"
3358 /** Only a subset of the @ref mdb_env flags can be changed
3359 * at runtime. Changing other flags requires closing the
3360 * environment and re-opening it with the new flags.
3362 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3363 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3366 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3368 int oflags, rc, len, excl = -1;
3369 char *lpath, *dpath;
3371 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3375 if (flags & MDB_NOSUBDIR) {
3376 rc = len + sizeof(LOCKSUFF) + len + 1;
3378 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3383 if (flags & MDB_NOSUBDIR) {
3384 dpath = lpath + len + sizeof(LOCKSUFF);
3385 sprintf(lpath, "%s" LOCKSUFF, path);
3386 strcpy(dpath, path);
3388 dpath = lpath + len + sizeof(LOCKNAME);
3389 sprintf(lpath, "%s" LOCKNAME, path);
3390 sprintf(dpath, "%s" DATANAME, path);
3394 flags |= env->me_flags;
3395 if (flags & MDB_RDONLY) {
3396 /* silently ignore WRITEMAP when we're only getting read access */
3397 flags &= ~MDB_WRITEMAP;
3399 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3400 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3403 env->me_flags = flags |= MDB_ENV_ACTIVE;
3407 env->me_path = strdup(path);
3408 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3409 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3410 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3415 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3420 if (F_ISSET(flags, MDB_RDONLY)) {
3421 oflags = GENERIC_READ;
3422 len = OPEN_EXISTING;
3424 oflags = GENERIC_READ|GENERIC_WRITE;
3427 mode = FILE_ATTRIBUTE_NORMAL;
3428 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3429 NULL, len, mode, NULL);
3431 if (F_ISSET(flags, MDB_RDONLY))
3434 oflags = O_RDWR | O_CREAT;
3436 env->me_fd = open(dpath, oflags, mode);
3438 if (env->me_fd == INVALID_HANDLE_VALUE) {
3443 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3444 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3445 env->me_mfd = env->me_fd;
3447 /* Synchronous fd for meta writes. Needed even with
3448 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3451 env->me_mfd = CreateFile(dpath, oflags,
3452 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3453 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3455 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3457 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3462 DPRINTF("opened dbenv %p", (void *) env);
3464 rc = mdb_env_share_locks(env, &excl);
3470 mdb_env_close0(env, excl);
3476 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3478 mdb_env_close0(MDB_env *env, int excl)
3482 if (!(env->me_flags & MDB_ENV_ACTIVE))
3485 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3486 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3487 free(env->me_dbxs[i].md_name.mv_data);
3489 free(env->me_dbflags);
3492 free(env->me_dirty_list);
3493 mdb_midl_free(env->me_free_pgs);
3495 if (env->me_flags & MDB_ENV_TXKEY) {
3496 pthread_key_delete(env->me_txkey);
3498 /* Delete our key from the global list */
3499 for (i=0; i<mdb_tls_nkeys; i++)
3500 if (mdb_tls_keys[i] == env->me_txkey) {
3501 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3509 munmap(env->me_map, env->me_mapsize);
3511 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3513 if (env->me_fd != INVALID_HANDLE_VALUE)
3516 pid_t pid = env->me_pid;
3517 /* Clearing readers is done in this function because
3518 * me_txkey with its destructor must be disabled first.
3520 for (i = env->me_numreaders; --i >= 0; )
3521 if (env->me_txns->mti_readers[i].mr_pid == pid)
3522 env->me_txns->mti_readers[i].mr_pid = 0;
3524 if (env->me_rmutex) {
3525 CloseHandle(env->me_rmutex);
3526 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3528 /* Windows automatically destroys the mutexes when
3529 * the last handle closes.
3531 #elif defined(MDB_USE_POSIX_SEM)
3532 if (env->me_rmutex != SEM_FAILED) {
3533 sem_close(env->me_rmutex);
3534 if (env->me_wmutex != SEM_FAILED)
3535 sem_close(env->me_wmutex);
3536 /* If we have the filelock: If we are the
3537 * only remaining user, clean up semaphores.
3540 mdb_env_excl_lock(env, &excl);
3542 sem_unlink(env->me_txns->mti_rmname);
3543 sem_unlink(env->me_txns->mti_wmname);
3547 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3549 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3552 /* Unlock the lockfile. Windows would have unlocked it
3553 * after closing anyway, but not necessarily at once.
3555 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3561 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3565 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3567 MDB_txn *txn = NULL;
3572 /* Do the lock/unlock of the reader mutex before starting the
3573 * write txn. Otherwise other read txns could block writers.
3575 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3580 /* We must start the actual read txn after blocking writers */
3581 mdb_txn_reset0(txn);
3583 /* Temporarily block writers until we snapshot the meta pages */
3586 rc = mdb_txn_renew0(txn);
3588 UNLOCK_MUTEX_W(env);
3593 wsize = env->me_psize * 2;
3597 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3598 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3601 rc = write(fd, env->me_map, wsize);
3602 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3605 UNLOCK_MUTEX_W(env);
3610 ptr = env->me_map + wsize;
3611 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3612 #define MAX_WRITE 2147483648U
3616 if (wsize > MAX_WRITE)
3620 rc = WriteFile(fd, ptr, w2, &len, NULL);
3621 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3630 if (wsize > MAX_WRITE)
3634 wres = write(fd, ptr, w2);
3635 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3648 mdb_env_copy(MDB_env *env, const char *path)
3652 HANDLE newfd = INVALID_HANDLE_VALUE;
3654 if (env->me_flags & MDB_NOSUBDIR) {
3655 lpath = (char *)path;
3658 len += sizeof(DATANAME);
3659 lpath = malloc(len);
3662 sprintf(lpath, "%s" DATANAME, path);
3665 /* The destination path must exist, but the destination file must not.
3666 * We don't want the OS to cache the writes, since the source data is
3667 * already in the OS cache.
3670 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3671 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3673 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3679 if (!(env->me_flags & MDB_NOSUBDIR))
3681 if (newfd == INVALID_HANDLE_VALUE) {
3686 #ifdef F_NOCACHE /* __APPLE__ */
3687 rc = fcntl(newfd, F_NOCACHE, 1);
3694 rc = mdb_env_copyfd(env, newfd);
3697 if (newfd != INVALID_HANDLE_VALUE)
3704 mdb_env_close(MDB_env *env)
3711 VGMEMP_DESTROY(env);
3712 while ((dp = env->me_dpages) != NULL) {
3713 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3714 env->me_dpages = dp->mp_next;
3718 mdb_env_close0(env, 0);
3722 /** Compare two items pointing at aligned size_t's */
3724 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3726 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3727 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3730 /** Compare two items pointing at aligned int's */
3732 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3734 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3735 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3738 /** Compare two items pointing at ints of unknown alignment.
3739 * Nodes and keys are guaranteed to be 2-byte aligned.
3742 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3744 #if BYTE_ORDER == LITTLE_ENDIAN
3745 unsigned short *u, *c;
3748 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3749 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3752 } while(!x && u > (unsigned short *)a->mv_data);
3755 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3759 /** Compare two items lexically */
3761 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3768 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3774 diff = memcmp(a->mv_data, b->mv_data, len);
3775 return diff ? diff : len_diff<0 ? -1 : len_diff;
3778 /** Compare two items in reverse byte order */
3780 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3782 const unsigned char *p1, *p2, *p1_lim;
3786 p1_lim = (const unsigned char *)a->mv_data;
3787 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3788 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3790 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3796 while (p1 > p1_lim) {
3797 diff = *--p1 - *--p2;
3801 return len_diff<0 ? -1 : len_diff;
3804 /** Search for key within a page, using binary search.
3805 * Returns the smallest entry larger or equal to the key.
3806 * If exactp is non-null, stores whether the found entry was an exact match
3807 * in *exactp (1 or 0).
3808 * Updates the cursor index with the index of the found entry.
3809 * If no entry larger or equal to the key is found, returns NULL.
3812 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3814 unsigned int i = 0, nkeys;
3817 MDB_page *mp = mc->mc_pg[mc->mc_top];
3818 MDB_node *node = NULL;
3823 nkeys = NUMKEYS(mp);
3828 COPY_PGNO(pgno, mp->mp_pgno);
3829 DPRINTF("searching %u keys in %s %spage %zu",
3830 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3837 low = IS_LEAF(mp) ? 0 : 1;
3839 cmp = mc->mc_dbx->md_cmp;
3841 /* Branch pages have no data, so if using integer keys,
3842 * alignment is guaranteed. Use faster mdb_cmp_int.
3844 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3845 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3852 nodekey.mv_size = mc->mc_db->md_pad;
3853 node = NODEPTR(mp, 0); /* fake */
3854 while (low <= high) {
3855 i = (low + high) >> 1;
3856 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3857 rc = cmp(key, &nodekey);
3858 DPRINTF("found leaf index %u [%s], rc = %i",
3859 i, DKEY(&nodekey), rc);
3868 while (low <= high) {
3869 i = (low + high) >> 1;
3871 node = NODEPTR(mp, i);
3872 nodekey.mv_size = NODEKSZ(node);
3873 nodekey.mv_data = NODEKEY(node);
3875 rc = cmp(key, &nodekey);
3878 DPRINTF("found leaf index %u [%s], rc = %i",
3879 i, DKEY(&nodekey), rc);
3881 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3882 i, DKEY(&nodekey), NODEPGNO(node), rc);
3893 if (rc > 0) { /* Found entry is less than the key. */
3894 i++; /* Skip to get the smallest entry larger than key. */
3896 node = NODEPTR(mp, i);
3899 *exactp = (rc == 0);
3900 /* store the key index */
3901 mc->mc_ki[mc->mc_top] = i;
3903 /* There is no entry larger or equal to the key. */
3906 /* nodeptr is fake for LEAF2 */
3912 mdb_cursor_adjust(MDB_cursor *mc, func)
3916 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3917 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3924 /** Pop a page off the top of the cursor's stack. */
3926 mdb_cursor_pop(MDB_cursor *mc)
3929 #ifndef MDB_DEBUG_SKIP
3930 MDB_page *top = mc->mc_pg[mc->mc_top];
3936 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3937 mc->mc_dbi, (void *) mc);
3941 /** Push a page onto the top of the cursor's stack. */
3943 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3945 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3946 mc->mc_dbi, (void *) mc);
3948 if (mc->mc_snum >= CURSOR_STACK) {
3949 assert(mc->mc_snum < CURSOR_STACK);
3950 return MDB_CURSOR_FULL;
3953 mc->mc_top = mc->mc_snum++;
3954 mc->mc_pg[mc->mc_top] = mp;
3955 mc->mc_ki[mc->mc_top] = 0;
3960 /** Find the address of the page corresponding to a given page number.
3961 * @param[in] txn the transaction for this access.
3962 * @param[in] pgno the page number for the page to retrieve.
3963 * @param[out] ret address of a pointer where the page's address will be stored.
3964 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
3965 * @return 0 on success, non-zero on failure.
3968 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
3973 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
3974 (txn->mt_env->me_flags & MDB_WRITEMAP)))
3979 MDB_ID2L dl = tx2->mt_u.dirty_list;
3981 unsigned x = mdb_mid2l_search(dl, pgno);
3982 if (x <= dl[0].mid && dl[x].mid == pgno) {
3988 } while ((tx2 = tx2->mt_parent) != NULL);
3991 if (pgno < txn->mt_next_pgno) {
3993 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3995 DPRINTF("page %zu not found", pgno);
3997 return MDB_PAGE_NOTFOUND;
4007 /** Search for the page a given key should be in.
4008 * Pushes parent pages on the cursor stack. This function continues a
4009 * search on a cursor that has already been initialized. (Usually by
4010 * #mdb_page_search() but also by #mdb_node_move().)
4011 * @param[in,out] mc the cursor for this operation.
4012 * @param[in] key the key to search for. If NULL, search for the lowest
4013 * page. (This is used by #mdb_cursor_first().)
4014 * @param[in] modify If true, visited pages are updated with new page numbers.
4015 * @return 0 on success, non-zero on failure.
4018 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4020 MDB_page *mp = mc->mc_pg[mc->mc_top];
4025 while (IS_BRANCH(mp)) {
4029 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4030 assert(NUMKEYS(mp) > 1);
4031 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4033 if (key == NULL) /* Initialize cursor to first page. */
4035 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4036 /* cursor to last page */
4040 node = mdb_node_search(mc, key, &exact);
4042 i = NUMKEYS(mp) - 1;
4044 i = mc->mc_ki[mc->mc_top];
4053 DPRINTF("following index %u for key [%s]",
4055 assert(i < NUMKEYS(mp));
4056 node = NODEPTR(mp, i);
4058 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4061 mc->mc_ki[mc->mc_top] = i;
4062 if ((rc = mdb_cursor_push(mc, mp)))
4066 if ((rc = mdb_page_touch(mc)) != 0)
4068 mp = mc->mc_pg[mc->mc_top];
4073 DPRINTF("internal error, index points to a %02X page!?",
4075 return MDB_CORRUPTED;
4078 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4079 key ? DKEY(key) : NULL);
4084 /** Search for the lowest key under the current branch page.
4085 * This just bypasses a NUMKEYS check in the current page
4086 * before calling mdb_page_search_root(), because the callers
4087 * are all in situations where the current page is known to
4091 mdb_page_search_lowest(MDB_cursor *mc)
4093 MDB_page *mp = mc->mc_pg[mc->mc_top];
4094 MDB_node *node = NODEPTR(mp, 0);
4097 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4100 mc->mc_ki[mc->mc_top] = 0;
4101 if ((rc = mdb_cursor_push(mc, mp)))
4103 return mdb_page_search_root(mc, NULL, 0);
4106 /** Search for the page a given key should be in.
4107 * Pushes parent pages on the cursor stack. This function just sets up
4108 * the search; it finds the root page for \b mc's database and sets this
4109 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4110 * called to complete the search.
4111 * @param[in,out] mc the cursor for this operation.
4112 * @param[in] key the key to search for. If NULL, search for the lowest
4113 * page. (This is used by #mdb_cursor_first().)
4114 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4115 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4116 * @return 0 on success, non-zero on failure.
4119 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4124 /* Make sure the txn is still viable, then find the root from
4125 * the txn's db table.
4127 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4128 DPUTS("transaction has failed, must abort");
4131 /* Make sure we're using an up-to-date root */
4132 if (mc->mc_dbi > MAIN_DBI) {
4133 if ((*mc->mc_dbflag & DB_STALE) ||
4134 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4136 unsigned char dbflag = 0;
4137 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4138 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4141 if (*mc->mc_dbflag & DB_STALE) {
4145 MDB_node *leaf = mdb_node_search(&mc2,
4146 &mc->mc_dbx->md_name, &exact);
4148 return MDB_NOTFOUND;
4149 mdb_node_read(mc->mc_txn, leaf, &data);
4150 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4152 /* The txn may not know this DBI, or another process may
4153 * have dropped and recreated the DB with other flags.
4155 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4156 return MDB_INCOMPATIBLE;
4157 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4159 if (flags & MDB_PS_MODIFY)
4161 *mc->mc_dbflag &= ~DB_STALE;
4162 *mc->mc_dbflag |= dbflag;
4165 root = mc->mc_db->md_root;
4167 if (root == P_INVALID) { /* Tree is empty. */
4168 DPUTS("tree is empty");
4169 return MDB_NOTFOUND;
4174 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4175 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4181 DPRINTF("db %u root page %zu has flags 0x%X",
4182 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4184 if (flags & MDB_PS_MODIFY) {
4185 if ((rc = mdb_page_touch(mc)))
4189 if (flags & MDB_PS_ROOTONLY)
4192 return mdb_page_search_root(mc, key, flags);
4196 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4198 MDB_txn *txn = mc->mc_txn;
4199 pgno_t pg = mp->mp_pgno;
4200 unsigned i, ovpages = mp->mp_pages;
4203 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4204 /* If the page is dirty we just acquired it, so we should
4205 * give it back to our current free list, if any.
4206 * Not currently supported in nested txns.
4207 * Otherwise put it onto the list of pages we freed in this txn.
4209 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && txn->mt_env->me_pghead) {
4212 MDB_ID2 *dl, ix, iy;
4213 rc = mdb_midl_need(&txn->mt_env->me_pghead, ovpages);
4216 /* Remove from dirty list */
4217 dl = txn->mt_u.dirty_list;
4219 for (ix = dl[x]; ix.mid != pg; ix = iy) {
4226 return MDB_CORRUPTED;
4229 /* Insert in me_pghead */
4230 mop = txn->mt_env->me_pghead;
4231 j = mop[0] + ovpages;
4232 for (i = mop[0]; i && mop[i] < pg; i--)
4238 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
4242 mc->mc_db->md_overflow_pages -= ovpages;
4246 /** Return the data associated with a given node.
4247 * @param[in] txn The transaction for this operation.
4248 * @param[in] leaf The node being read.
4249 * @param[out] data Updated to point to the node's data.
4250 * @return 0 on success, non-zero on failure.
4253 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4255 MDB_page *omp; /* overflow page */
4259 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4260 data->mv_size = NODEDSZ(leaf);
4261 data->mv_data = NODEDATA(leaf);
4265 /* Read overflow data.
4267 data->mv_size = NODEDSZ(leaf);
4268 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4269 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4270 DPRINTF("read overflow page %zu failed", pgno);
4273 data->mv_data = METADATA(omp);
4279 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4280 MDB_val *key, MDB_val *data)
4289 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4291 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4294 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4298 mdb_cursor_init(&mc, txn, dbi, &mx);
4299 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4302 /** Find a sibling for a page.
4303 * Replaces the page at the top of the cursor's stack with the
4304 * specified sibling, if one exists.
4305 * @param[in] mc The cursor for this operation.
4306 * @param[in] move_right Non-zero if the right sibling is requested,
4307 * otherwise the left sibling.
4308 * @return 0 on success, non-zero on failure.
4311 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4317 if (mc->mc_snum < 2) {
4318 return MDB_NOTFOUND; /* root has no siblings */
4322 DPRINTF("parent page is page %zu, index %u",
4323 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4325 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4326 : (mc->mc_ki[mc->mc_top] == 0)) {
4327 DPRINTF("no more keys left, moving to %s sibling",
4328 move_right ? "right" : "left");
4329 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4330 /* undo cursor_pop before returning */
4337 mc->mc_ki[mc->mc_top]++;
4339 mc->mc_ki[mc->mc_top]--;
4340 DPRINTF("just moving to %s index key %u",
4341 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4343 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4345 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4346 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4349 mdb_cursor_push(mc, mp);
4351 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4356 /** Move the cursor to the next data item. */
4358 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4364 if (mc->mc_flags & C_EOF) {
4365 return MDB_NOTFOUND;
4368 assert(mc->mc_flags & C_INITIALIZED);
4370 mp = mc->mc_pg[mc->mc_top];
4372 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4373 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4374 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4375 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4376 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4377 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4381 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4382 if (op == MDB_NEXT_DUP)
4383 return MDB_NOTFOUND;
4387 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4389 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4390 DPUTS("=====> move to next sibling page");
4391 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4392 mc->mc_flags |= C_EOF;
4393 mc->mc_flags &= ~C_INITIALIZED;
4394 return MDB_NOTFOUND;
4396 mp = mc->mc_pg[mc->mc_top];
4397 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4399 mc->mc_ki[mc->mc_top]++;
4401 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4402 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4405 key->mv_size = mc->mc_db->md_pad;
4406 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4410 assert(IS_LEAF(mp));
4411 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4413 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4414 mdb_xcursor_init1(mc, leaf);
4417 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4420 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4421 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4422 if (rc != MDB_SUCCESS)
4427 MDB_GET_KEY(leaf, key);
4431 /** Move the cursor to the previous data item. */
4433 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4439 assert(mc->mc_flags & C_INITIALIZED);
4441 mp = mc->mc_pg[mc->mc_top];
4443 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4444 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4445 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4446 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4447 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4448 if (op != MDB_PREV || rc == MDB_SUCCESS)
4451 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4452 if (op == MDB_PREV_DUP)
4453 return MDB_NOTFOUND;
4458 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4460 if (mc->mc_ki[mc->mc_top] == 0) {
4461 DPUTS("=====> move to prev sibling page");
4462 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4463 mc->mc_flags &= ~C_INITIALIZED;
4464 return MDB_NOTFOUND;
4466 mp = mc->mc_pg[mc->mc_top];
4467 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4468 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4470 mc->mc_ki[mc->mc_top]--;
4472 mc->mc_flags &= ~C_EOF;
4474 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4475 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4478 key->mv_size = mc->mc_db->md_pad;
4479 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4483 assert(IS_LEAF(mp));
4484 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4486 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4487 mdb_xcursor_init1(mc, leaf);
4490 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4493 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4494 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4495 if (rc != MDB_SUCCESS)
4500 MDB_GET_KEY(leaf, key);
4504 /** Set the cursor on a specific data item. */
4506 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4507 MDB_cursor_op op, int *exactp)
4511 MDB_node *leaf = NULL;
4516 assert(key->mv_size > 0);
4518 /* See if we're already on the right page */
4519 if (mc->mc_flags & C_INITIALIZED) {
4522 mp = mc->mc_pg[mc->mc_top];
4524 mc->mc_ki[mc->mc_top] = 0;
4525 return MDB_NOTFOUND;
4527 if (mp->mp_flags & P_LEAF2) {
4528 nodekey.mv_size = mc->mc_db->md_pad;
4529 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4531 leaf = NODEPTR(mp, 0);
4532 MDB_GET_KEY(leaf, &nodekey);
4534 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4536 /* Probably happens rarely, but first node on the page
4537 * was the one we wanted.
4539 mc->mc_ki[mc->mc_top] = 0;
4546 unsigned int nkeys = NUMKEYS(mp);
4548 if (mp->mp_flags & P_LEAF2) {
4549 nodekey.mv_data = LEAF2KEY(mp,
4550 nkeys-1, nodekey.mv_size);
4552 leaf = NODEPTR(mp, nkeys-1);
4553 MDB_GET_KEY(leaf, &nodekey);
4555 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4557 /* last node was the one we wanted */
4558 mc->mc_ki[mc->mc_top] = nkeys-1;
4564 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4565 /* This is definitely the right page, skip search_page */
4566 if (mp->mp_flags & P_LEAF2) {
4567 nodekey.mv_data = LEAF2KEY(mp,
4568 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4570 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4571 MDB_GET_KEY(leaf, &nodekey);
4573 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4575 /* current node was the one we wanted */
4585 /* If any parents have right-sibs, search.
4586 * Otherwise, there's nothing further.
4588 for (i=0; i<mc->mc_top; i++)
4590 NUMKEYS(mc->mc_pg[i])-1)
4592 if (i == mc->mc_top) {
4593 /* There are no other pages */
4594 mc->mc_ki[mc->mc_top] = nkeys;
4595 return MDB_NOTFOUND;
4599 /* There are no other pages */
4600 mc->mc_ki[mc->mc_top] = 0;
4601 return MDB_NOTFOUND;
4605 rc = mdb_page_search(mc, key, 0);
4606 if (rc != MDB_SUCCESS)
4609 mp = mc->mc_pg[mc->mc_top];
4610 assert(IS_LEAF(mp));
4613 leaf = mdb_node_search(mc, key, exactp);
4614 if (exactp != NULL && !*exactp) {
4615 /* MDB_SET specified and not an exact match. */
4616 return MDB_NOTFOUND;
4620 DPUTS("===> inexact leaf not found, goto sibling");
4621 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4622 return rc; /* no entries matched */
4623 mp = mc->mc_pg[mc->mc_top];
4624 assert(IS_LEAF(mp));
4625 leaf = NODEPTR(mp, 0);
4629 mc->mc_flags |= C_INITIALIZED;
4630 mc->mc_flags &= ~C_EOF;
4633 key->mv_size = mc->mc_db->md_pad;
4634 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4638 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4639 mdb_xcursor_init1(mc, leaf);
4642 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4643 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4644 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4647 if (op == MDB_GET_BOTH) {
4653 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4654 if (rc != MDB_SUCCESS)
4657 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4659 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4661 rc = mc->mc_dbx->md_dcmp(data, &d2);
4663 if (op == MDB_GET_BOTH || rc > 0)
4664 return MDB_NOTFOUND;
4669 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4670 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4675 /* The key already matches in all other cases */
4676 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4677 MDB_GET_KEY(leaf, key);
4678 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4683 /** Move the cursor to the first item in the database. */
4685 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4690 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4691 rc = mdb_page_search(mc, NULL, 0);
4692 if (rc != MDB_SUCCESS)
4695 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4697 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4698 mc->mc_flags |= C_INITIALIZED;
4699 mc->mc_flags &= ~C_EOF;
4701 mc->mc_ki[mc->mc_top] = 0;
4703 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4704 key->mv_size = mc->mc_db->md_pad;
4705 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4710 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4711 mdb_xcursor_init1(mc, leaf);
4712 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4717 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4718 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4722 MDB_GET_KEY(leaf, key);
4726 /** Move the cursor to the last item in the database. */
4728 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4733 if (!(mc->mc_flags & C_EOF)) {
4735 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4738 lkey.mv_size = MDB_MAXKEYSIZE+1;
4739 lkey.mv_data = NULL;
4740 rc = mdb_page_search(mc, &lkey, 0);
4741 if (rc != MDB_SUCCESS)
4744 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4747 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4748 mc->mc_flags |= C_INITIALIZED|C_EOF;
4749 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4751 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4752 key->mv_size = mc->mc_db->md_pad;
4753 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4758 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4759 mdb_xcursor_init1(mc, leaf);
4760 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4765 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4766 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4771 MDB_GET_KEY(leaf, key);
4776 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4785 case MDB_GET_CURRENT:
4786 if (!(mc->mc_flags & C_INITIALIZED)) {
4789 MDB_page *mp = mc->mc_pg[mc->mc_top];
4791 mc->mc_ki[mc->mc_top] = 0;
4797 key->mv_size = mc->mc_db->md_pad;
4798 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4800 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4801 MDB_GET_KEY(leaf, key);
4803 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4804 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4806 rc = mdb_node_read(mc->mc_txn, leaf, data);
4813 case MDB_GET_BOTH_RANGE:
4814 if (data == NULL || mc->mc_xcursor == NULL) {
4822 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4824 } else if (op == MDB_SET_RANGE)
4825 rc = mdb_cursor_set(mc, key, data, op, NULL);
4827 rc = mdb_cursor_set(mc, key, data, op, &exact);
4829 case MDB_GET_MULTIPLE:
4831 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4832 !(mc->mc_flags & C_INITIALIZED)) {
4837 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4838 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4841 case MDB_NEXT_MULTIPLE:
4843 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4847 if (!(mc->mc_flags & C_INITIALIZED))
4848 rc = mdb_cursor_first(mc, key, data);
4850 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4851 if (rc == MDB_SUCCESS) {
4852 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4855 mx = &mc->mc_xcursor->mx_cursor;
4856 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4858 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4859 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4867 case MDB_NEXT_NODUP:
4868 if (!(mc->mc_flags & C_INITIALIZED))
4869 rc = mdb_cursor_first(mc, key, data);
4871 rc = mdb_cursor_next(mc, key, data, op);
4875 case MDB_PREV_NODUP:
4876 if (!(mc->mc_flags & C_INITIALIZED)) {
4877 rc = mdb_cursor_last(mc, key, data);
4880 mc->mc_flags |= C_INITIALIZED;
4881 mc->mc_ki[mc->mc_top]++;
4883 rc = mdb_cursor_prev(mc, key, data, op);
4886 rc = mdb_cursor_first(mc, key, data);
4890 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4891 !(mc->mc_flags & C_INITIALIZED) ||
4892 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4896 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4899 rc = mdb_cursor_last(mc, key, data);
4903 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4904 !(mc->mc_flags & C_INITIALIZED) ||
4905 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4909 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4912 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4920 /** Touch all the pages in the cursor stack.
4921 * Makes sure all the pages are writable, before attempting a write operation.
4922 * @param[in] mc The cursor to operate on.
4925 mdb_cursor_touch(MDB_cursor *mc)
4929 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4932 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
4933 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4936 *mc->mc_dbflag |= DB_DIRTY;
4938 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4939 rc = mdb_page_touch(mc);
4943 mc->mc_top = mc->mc_snum-1;
4948 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4951 MDB_node *leaf = NULL;
4952 MDB_val xdata, *rdata, dkey;
4955 int do_sub = 0, insert = 0;
4956 unsigned int mcount = 0;
4960 char dbuf[MDB_MAXKEYSIZE+1];
4961 unsigned int nflags;
4964 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4967 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4970 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4973 #if SIZE_MAX > MAXDATASIZE
4974 if (data->mv_size > MAXDATASIZE)
4978 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4979 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4983 if (flags == MDB_CURRENT) {
4984 if (!(mc->mc_flags & C_INITIALIZED))
4987 } else if (mc->mc_db->md_root == P_INVALID) {
4989 /* new database, write a root leaf page */
4990 DPUTS("allocating new root leaf page");
4991 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4995 mdb_cursor_push(mc, np);
4996 mc->mc_db->md_root = np->mp_pgno;
4997 mc->mc_db->md_depth++;
4998 *mc->mc_dbflag |= DB_DIRTY;
4999 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5001 np->mp_flags |= P_LEAF2;
5002 mc->mc_flags |= C_INITIALIZED;
5008 if (flags & MDB_APPEND) {
5010 rc = mdb_cursor_last(mc, &k2, &d2);
5012 rc = mc->mc_dbx->md_cmp(key, &k2);
5015 mc->mc_ki[mc->mc_top]++;
5017 /* new key is <= last key */
5022 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5024 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5025 DPRINTF("duplicate key [%s]", DKEY(key));
5027 return MDB_KEYEXIST;
5029 if (rc && rc != MDB_NOTFOUND)
5033 /* Cursor is positioned, now make sure all pages are writable */
5034 rc2 = mdb_cursor_touch(mc);
5039 /* The key already exists */
5040 if (rc == MDB_SUCCESS) {
5041 /* there's only a key anyway, so this is a no-op */
5042 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5043 unsigned int ksize = mc->mc_db->md_pad;
5044 if (key->mv_size != ksize)
5046 if (flags == MDB_CURRENT) {
5047 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5048 memcpy(ptr, key->mv_data, ksize);
5053 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5056 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5057 /* Was a single item before, must convert now */
5059 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5060 /* Just overwrite the current item */
5061 if (flags == MDB_CURRENT)
5064 dkey.mv_size = NODEDSZ(leaf);
5065 dkey.mv_data = NODEDATA(leaf);
5066 #if UINT_MAX < SIZE_MAX
5067 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5068 #ifdef MISALIGNED_OK
5069 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5071 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5074 /* if data matches, ignore it */
5075 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5076 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5078 /* create a fake page for the dup items */
5079 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5080 dkey.mv_data = dbuf;
5081 fp = (MDB_page *)&pbuf;
5082 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5083 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5084 fp->mp_lower = PAGEHDRSZ;
5085 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5086 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5087 fp->mp_flags |= P_LEAF2;
5088 fp->mp_pad = data->mv_size;
5089 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5091 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5092 (dkey.mv_size & 1) + (data->mv_size & 1);
5094 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5097 xdata.mv_size = fp->mp_upper;
5102 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5103 /* See if we need to convert from fake page to subDB */
5105 unsigned int offset;
5109 fp = NODEDATA(leaf);
5110 if (flags == MDB_CURRENT) {
5112 fp->mp_flags |= P_DIRTY;
5113 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5114 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5118 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5119 offset = fp->mp_pad;
5120 if (SIZELEFT(fp) >= offset)
5122 offset *= 4; /* space for 4 more */
5124 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5126 offset += offset & 1;
5127 fp_flags = fp->mp_flags;
5128 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5129 offset >= mc->mc_txn->mt_env->me_nodemax) {
5130 /* yes, convert it */
5132 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5133 dummy.md_pad = fp->mp_pad;
5134 dummy.md_flags = MDB_DUPFIXED;
5135 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5136 dummy.md_flags |= MDB_INTEGERKEY;
5139 dummy.md_branch_pages = 0;
5140 dummy.md_leaf_pages = 1;
5141 dummy.md_overflow_pages = 0;
5142 dummy.md_entries = NUMKEYS(fp);
5144 xdata.mv_size = sizeof(MDB_db);
5145 xdata.mv_data = &dummy;
5146 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5148 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5149 flags |= F_DUPDATA|F_SUBDATA;
5150 dummy.md_root = mp->mp_pgno;
5151 fp_flags &= ~P_SUBP;
5153 /* no, just grow it */
5155 xdata.mv_size = NODEDSZ(leaf) + offset;
5156 xdata.mv_data = &pbuf;
5157 mp = (MDB_page *)&pbuf;
5158 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5161 mp->mp_flags = fp_flags | P_DIRTY;
5162 mp->mp_pad = fp->mp_pad;
5163 mp->mp_lower = fp->mp_lower;
5164 mp->mp_upper = fp->mp_upper + offset;
5166 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5168 nsize = NODEDSZ(leaf) - fp->mp_upper;
5169 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5170 for (i=0; i<NUMKEYS(fp); i++)
5171 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5173 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5177 /* data is on sub-DB, just store it */
5178 flags |= F_DUPDATA|F_SUBDATA;
5182 /* overflow page overwrites need special handling */
5183 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5186 unsigned psize = mc->mc_txn->mt_env->me_psize;
5187 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5189 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5190 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5192 ovpages = omp->mp_pages;
5194 /* Is the ov page writable and large enough? */
5195 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5196 /* yes, overwrite it. Note in this case we don't
5197 * bother to try shrinking the page if the new data
5198 * is smaller than the overflow threshold.
5201 /* It is writable only in a parent txn */
5202 size_t sz = (size_t) psize * ovpages, off;
5203 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
5209 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5210 if (!(flags & MDB_RESERVE)) {
5211 /* Copy end of page, adjusting alignment so
5212 * compiler may copy words instead of bytes.
5214 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5215 memcpy((size_t *)((char *)np + off),
5216 (size_t *)((char *)omp + off), sz - off);
5219 memcpy(np, omp, sz); /* Copy beginning of page */
5222 SETDSZ(leaf, data->mv_size);
5223 if (F_ISSET(flags, MDB_RESERVE))
5224 data->mv_data = METADATA(omp);
5226 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5229 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
5232 } else if (NODEDSZ(leaf) == data->mv_size) {
5233 /* same size, just replace it. Note that we could
5234 * also reuse this node if the new data is smaller,
5235 * but instead we opt to shrink the node in that case.
5237 if (F_ISSET(flags, MDB_RESERVE))
5238 data->mv_data = NODEDATA(leaf);
5239 else if (data->mv_size)
5240 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5242 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5245 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5246 mc->mc_db->md_entries--;
5248 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5255 nflags = flags & NODE_ADD_FLAGS;
5256 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5257 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5258 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5259 nflags &= ~MDB_APPEND;
5261 nflags |= MDB_SPLIT_REPLACE;
5262 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5264 /* There is room already in this leaf page. */
5265 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5266 if (rc == 0 && !do_sub && insert) {
5267 /* Adjust other cursors pointing to mp */
5268 MDB_cursor *m2, *m3;
5269 MDB_dbi dbi = mc->mc_dbi;
5270 unsigned i = mc->mc_top;
5271 MDB_page *mp = mc->mc_pg[i];
5273 if (mc->mc_flags & C_SUB)
5276 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5277 if (mc->mc_flags & C_SUB)
5278 m3 = &m2->mc_xcursor->mx_cursor;
5281 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5282 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5289 if (rc != MDB_SUCCESS)
5290 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5292 /* Now store the actual data in the child DB. Note that we're
5293 * storing the user data in the keys field, so there are strict
5294 * size limits on dupdata. The actual data fields of the child
5295 * DB are all zero size.
5302 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5303 if (flags & MDB_CURRENT) {
5304 xflags = MDB_CURRENT;
5306 mdb_xcursor_init1(mc, leaf);
5307 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5309 /* converted, write the original data first */
5311 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5315 /* Adjust other cursors pointing to mp */
5317 unsigned i = mc->mc_top;
5318 MDB_page *mp = mc->mc_pg[i];
5320 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5321 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5322 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5323 mdb_xcursor_init1(m2, leaf);
5327 /* we've done our job */
5330 if (flags & MDB_APPENDDUP)
5331 xflags |= MDB_APPEND;
5332 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5333 if (flags & F_SUBDATA) {
5334 void *db = NODEDATA(leaf);
5335 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5338 /* sub-writes might have failed so check rc again.
5339 * Don't increment count if we just replaced an existing item.
5341 if (!rc && !(flags & MDB_CURRENT))
5342 mc->mc_db->md_entries++;
5343 if (flags & MDB_MULTIPLE) {
5345 if (mcount < data[1].mv_size) {
5346 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5347 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5353 /* If we succeeded and the key didn't exist before, make sure
5354 * the cursor is marked valid.
5357 mc->mc_flags |= C_INITIALIZED;
5362 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5367 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5370 if (!(mc->mc_flags & C_INITIALIZED))
5373 rc = mdb_cursor_touch(mc);
5377 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5379 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5380 if (flags != MDB_NODUPDATA) {
5381 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5382 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5384 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5385 /* If sub-DB still has entries, we're done */
5386 if (mc->mc_xcursor->mx_db.md_entries) {
5387 if (leaf->mn_flags & F_SUBDATA) {
5388 /* update subDB info */
5389 void *db = NODEDATA(leaf);
5390 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5393 /* shrink fake page */
5394 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5395 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5396 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5397 /* fix other sub-DB cursors pointed at this fake page */
5398 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5399 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5400 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5401 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5402 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5405 mc->mc_db->md_entries--;
5408 /* otherwise fall thru and delete the sub-DB */
5411 if (leaf->mn_flags & F_SUBDATA) {
5412 /* add all the child DB's pages to the free list */
5413 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5414 if (rc == MDB_SUCCESS) {
5415 mc->mc_db->md_entries -=
5416 mc->mc_xcursor->mx_db.md_entries;
5421 return mdb_cursor_del0(mc, leaf);
5424 /** Allocate and initialize new pages for a database.
5425 * @param[in] mc a cursor on the database being added to.
5426 * @param[in] flags flags defining what type of page is being allocated.
5427 * @param[in] num the number of pages to allocate. This is usually 1,
5428 * unless allocating overflow pages for a large record.
5429 * @param[out] mp Address of a page, or NULL on failure.
5430 * @return 0 on success, non-zero on failure.
5433 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5438 if ((rc = mdb_page_alloc(mc, num, &np)))
5440 DPRINTF("allocated new mpage %zu, page size %u",
5441 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5442 np->mp_flags = flags | P_DIRTY;
5443 np->mp_lower = PAGEHDRSZ;
5444 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5447 mc->mc_db->md_branch_pages++;
5448 else if (IS_LEAF(np))
5449 mc->mc_db->md_leaf_pages++;
5450 else if (IS_OVERFLOW(np)) {
5451 mc->mc_db->md_overflow_pages += num;
5459 /** Calculate the size of a leaf node.
5460 * The size depends on the environment's page size; if a data item
5461 * is too large it will be put onto an overflow page and the node
5462 * size will only include the key and not the data. Sizes are always
5463 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5464 * of the #MDB_node headers.
5465 * @param[in] env The environment handle.
5466 * @param[in] key The key for the node.
5467 * @param[in] data The data for the node.
5468 * @return The number of bytes needed to store the node.
5471 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5475 sz = LEAFSIZE(key, data);
5476 if (sz >= env->me_nodemax) {
5477 /* put on overflow page */
5478 sz -= data->mv_size - sizeof(pgno_t);
5482 return sz + sizeof(indx_t);
5485 /** Calculate the size of a branch node.
5486 * The size should depend on the environment's page size but since
5487 * we currently don't support spilling large keys onto overflow
5488 * pages, it's simply the size of the #MDB_node header plus the
5489 * size of the key. Sizes are always rounded up to an even number
5490 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5491 * @param[in] env The environment handle.
5492 * @param[in] key The key for the node.
5493 * @return The number of bytes needed to store the node.
5496 mdb_branch_size(MDB_env *env, MDB_val *key)
5501 if (sz >= env->me_nodemax) {
5502 /* put on overflow page */
5503 /* not implemented */
5504 /* sz -= key->size - sizeof(pgno_t); */
5507 return sz + sizeof(indx_t);
5510 /** Add a node to the page pointed to by the cursor.
5511 * @param[in] mc The cursor for this operation.
5512 * @param[in] indx The index on the page where the new node should be added.
5513 * @param[in] key The key for the new node.
5514 * @param[in] data The data for the new node, if any.
5515 * @param[in] pgno The page number, if adding a branch node.
5516 * @param[in] flags Flags for the node.
5517 * @return 0 on success, non-zero on failure. Possible errors are:
5519 * <li>ENOMEM - failed to allocate overflow pages for the node.
5520 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5521 * should never happen since all callers already calculate the
5522 * page's free space before calling this function.
5526 mdb_node_add(MDB_cursor *mc, indx_t indx,
5527 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5530 size_t node_size = NODESIZE;
5533 MDB_page *mp = mc->mc_pg[mc->mc_top];
5534 MDB_page *ofp = NULL; /* overflow page */
5537 assert(mp->mp_upper >= mp->mp_lower);
5539 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5540 IS_LEAF(mp) ? "leaf" : "branch",
5541 IS_SUBP(mp) ? "sub-" : "",
5542 mp->mp_pgno, indx, data ? data->mv_size : 0,
5543 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5546 /* Move higher keys up one slot. */
5547 int ksize = mc->mc_db->md_pad, dif;
5548 char *ptr = LEAF2KEY(mp, indx, ksize);
5549 dif = NUMKEYS(mp) - indx;
5551 memmove(ptr+ksize, ptr, dif*ksize);
5552 /* insert new key */
5553 memcpy(ptr, key->mv_data, ksize);
5555 /* Just using these for counting */
5556 mp->mp_lower += sizeof(indx_t);
5557 mp->mp_upper -= ksize - sizeof(indx_t);
5562 node_size += key->mv_size;
5566 if (F_ISSET(flags, F_BIGDATA)) {
5567 /* Data already on overflow page. */
5568 node_size += sizeof(pgno_t);
5569 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5570 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5572 /* Put data on overflow page. */
5573 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5574 data->mv_size, node_size+data->mv_size);
5575 node_size += sizeof(pgno_t);
5576 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5578 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5581 node_size += data->mv_size;
5584 node_size += node_size & 1;
5586 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5587 DPRINTF("not enough room in page %zu, got %u ptrs",
5588 mp->mp_pgno, NUMKEYS(mp));
5589 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5590 mp->mp_upper - mp->mp_lower);
5591 DPRINTF("node size = %zu", node_size);
5592 return MDB_PAGE_FULL;
5595 /* Move higher pointers up one slot. */
5596 for (i = NUMKEYS(mp); i > indx; i--)
5597 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5599 /* Adjust free space offsets. */
5600 ofs = mp->mp_upper - node_size;
5601 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5602 mp->mp_ptrs[indx] = ofs;
5604 mp->mp_lower += sizeof(indx_t);
5606 /* Write the node data. */
5607 node = NODEPTR(mp, indx);
5608 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5609 node->mn_flags = flags;
5611 SETDSZ(node,data->mv_size);
5616 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5621 if (F_ISSET(flags, F_BIGDATA))
5622 memcpy(node->mn_data + key->mv_size, data->mv_data,
5624 else if (F_ISSET(flags, MDB_RESERVE))
5625 data->mv_data = node->mn_data + key->mv_size;
5627 memcpy(node->mn_data + key->mv_size, data->mv_data,
5630 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5632 if (F_ISSET(flags, MDB_RESERVE))
5633 data->mv_data = METADATA(ofp);
5635 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5642 /** Delete the specified node from a page.
5643 * @param[in] mp The page to operate on.
5644 * @param[in] indx The index of the node to delete.
5645 * @param[in] ksize The size of a node. Only used if the page is
5646 * part of a #MDB_DUPFIXED database.
5649 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5652 indx_t i, j, numkeys, ptr;
5659 COPY_PGNO(pgno, mp->mp_pgno);
5660 DPRINTF("delete node %u on %s page %zu", indx,
5661 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5664 assert(indx < NUMKEYS(mp));
5667 int x = NUMKEYS(mp) - 1 - indx;
5668 base = LEAF2KEY(mp, indx, ksize);
5670 memmove(base, base + ksize, x * ksize);
5671 mp->mp_lower -= sizeof(indx_t);
5672 mp->mp_upper += ksize - sizeof(indx_t);
5676 node = NODEPTR(mp, indx);
5677 sz = NODESIZE + node->mn_ksize;
5679 if (F_ISSET(node->mn_flags, F_BIGDATA))
5680 sz += sizeof(pgno_t);
5682 sz += NODEDSZ(node);
5686 ptr = mp->mp_ptrs[indx];
5687 numkeys = NUMKEYS(mp);
5688 for (i = j = 0; i < numkeys; i++) {
5690 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5691 if (mp->mp_ptrs[i] < ptr)
5692 mp->mp_ptrs[j] += sz;
5697 base = (char *)mp + mp->mp_upper;
5698 memmove(base + sz, base, ptr - mp->mp_upper);
5700 mp->mp_lower -= sizeof(indx_t);
5704 /** Compact the main page after deleting a node on a subpage.
5705 * @param[in] mp The main page to operate on.
5706 * @param[in] indx The index of the subpage on the main page.
5709 mdb_node_shrink(MDB_page *mp, indx_t indx)
5716 indx_t i, numkeys, ptr;
5718 node = NODEPTR(mp, indx);
5719 sp = (MDB_page *)NODEDATA(node);
5720 osize = NODEDSZ(node);
5722 delta = sp->mp_upper - sp->mp_lower;
5723 SETDSZ(node, osize - delta);
5724 xp = (MDB_page *)((char *)sp + delta);
5726 /* shift subpage upward */
5728 nsize = NUMKEYS(sp) * sp->mp_pad;
5729 memmove(METADATA(xp), METADATA(sp), nsize);
5732 nsize = osize - sp->mp_upper;
5733 numkeys = NUMKEYS(sp);
5734 for (i=numkeys-1; i>=0; i--)
5735 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5737 xp->mp_upper = sp->mp_lower;
5738 xp->mp_lower = sp->mp_lower;
5739 xp->mp_flags = sp->mp_flags;
5740 xp->mp_pad = sp->mp_pad;
5741 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5743 /* shift lower nodes upward */
5744 ptr = mp->mp_ptrs[indx];
5745 numkeys = NUMKEYS(mp);
5746 for (i = 0; i < numkeys; i++) {
5747 if (mp->mp_ptrs[i] <= ptr)
5748 mp->mp_ptrs[i] += delta;
5751 base = (char *)mp + mp->mp_upper;
5752 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5753 mp->mp_upper += delta;
5756 /** Initial setup of a sorted-dups cursor.
5757 * Sorted duplicates are implemented as a sub-database for the given key.
5758 * The duplicate data items are actually keys of the sub-database.
5759 * Operations on the duplicate data items are performed using a sub-cursor
5760 * initialized when the sub-database is first accessed. This function does
5761 * the preliminary setup of the sub-cursor, filling in the fields that
5762 * depend only on the parent DB.
5763 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5766 mdb_xcursor_init0(MDB_cursor *mc)
5768 MDB_xcursor *mx = mc->mc_xcursor;
5770 mx->mx_cursor.mc_xcursor = NULL;
5771 mx->mx_cursor.mc_txn = mc->mc_txn;
5772 mx->mx_cursor.mc_db = &mx->mx_db;
5773 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5774 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5775 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5776 mx->mx_cursor.mc_snum = 0;
5777 mx->mx_cursor.mc_top = 0;
5778 mx->mx_cursor.mc_flags = C_SUB;
5779 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5780 mx->mx_dbx.md_dcmp = NULL;
5781 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5784 /** Final setup of a sorted-dups cursor.
5785 * Sets up the fields that depend on the data from the main cursor.
5786 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5787 * @param[in] node The data containing the #MDB_db record for the
5788 * sorted-dup database.
5791 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5793 MDB_xcursor *mx = mc->mc_xcursor;
5795 if (node->mn_flags & F_SUBDATA) {
5796 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5797 mx->mx_cursor.mc_pg[0] = 0;
5798 mx->mx_cursor.mc_snum = 0;
5799 mx->mx_cursor.mc_flags = C_SUB;
5801 MDB_page *fp = NODEDATA(node);
5802 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5803 mx->mx_db.md_flags = 0;
5804 mx->mx_db.md_depth = 1;
5805 mx->mx_db.md_branch_pages = 0;
5806 mx->mx_db.md_leaf_pages = 1;
5807 mx->mx_db.md_overflow_pages = 0;
5808 mx->mx_db.md_entries = NUMKEYS(fp);
5809 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5810 mx->mx_cursor.mc_snum = 1;
5811 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5812 mx->mx_cursor.mc_top = 0;
5813 mx->mx_cursor.mc_pg[0] = fp;
5814 mx->mx_cursor.mc_ki[0] = 0;
5815 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5816 mx->mx_db.md_flags = MDB_DUPFIXED;
5817 mx->mx_db.md_pad = fp->mp_pad;
5818 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5819 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5822 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5824 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5826 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5827 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5828 #if UINT_MAX < SIZE_MAX
5829 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5830 #ifdef MISALIGNED_OK
5831 mx->mx_dbx.md_cmp = mdb_cmp_long;
5833 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5838 /** Initialize a cursor for a given transaction and database. */
5840 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5845 mc->mc_db = &txn->mt_dbs[dbi];
5846 mc->mc_dbx = &txn->mt_dbxs[dbi];
5847 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5852 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5854 mc->mc_xcursor = mx;
5855 mdb_xcursor_init0(mc);
5857 mc->mc_xcursor = NULL;
5859 if (*mc->mc_dbflag & DB_STALE) {
5860 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5865 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5868 size_t size = sizeof(MDB_cursor);
5870 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5873 /* Allow read access to the freelist */
5874 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5877 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5878 size += sizeof(MDB_xcursor);
5880 if ((mc = malloc(size)) != NULL) {
5881 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5882 if (txn->mt_cursors) {
5883 mc->mc_next = txn->mt_cursors[dbi];
5884 txn->mt_cursors[dbi] = mc;
5885 mc->mc_flags |= C_UNTRACK;
5887 mc->mc_flags |= C_ALLOCD;
5898 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5902 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5905 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5908 flags = mc->mc_flags;
5910 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5912 mc->mc_flags |= (flags & C_ALLOCD);
5916 /* Return the count of duplicate data items for the current key */
5918 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5922 if (mc == NULL || countp == NULL)
5925 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5928 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5929 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5932 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5935 *countp = mc->mc_xcursor->mx_db.md_entries;
5941 mdb_cursor_close(MDB_cursor *mc)
5944 /* remove from txn, if tracked */
5945 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
5946 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5947 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5949 *prev = mc->mc_next;
5951 if (mc->mc_flags & C_ALLOCD)
5957 mdb_cursor_txn(MDB_cursor *mc)
5959 if (!mc) return NULL;
5964 mdb_cursor_dbi(MDB_cursor *mc)
5970 /** Replace the key for a node with a new key.
5971 * @param[in] mc Cursor pointing to the node to operate on.
5972 * @param[in] key The new key to use.
5973 * @return 0 on success, non-zero on failure.
5976 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5983 indx_t ptr, i, numkeys, indx;
5986 indx = mc->mc_ki[mc->mc_top];
5987 mp = mc->mc_pg[mc->mc_top];
5988 node = NODEPTR(mp, indx);
5989 ptr = mp->mp_ptrs[indx];
5993 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5994 k2.mv_data = NODEKEY(node);
5995 k2.mv_size = node->mn_ksize;
5996 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5998 mdb_dkey(&k2, kbuf2),
6004 delta0 = delta = key->mv_size - node->mn_ksize;
6006 /* Must be 2-byte aligned. If new key is
6007 * shorter by 1, the shift will be skipped.
6009 delta += (delta & 1);
6011 if (delta > 0 && SIZELEFT(mp) < delta) {
6013 /* not enough space left, do a delete and split */
6014 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6015 pgno = NODEPGNO(node);
6016 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6017 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6020 numkeys = NUMKEYS(mp);
6021 for (i = 0; i < numkeys; i++) {
6022 if (mp->mp_ptrs[i] <= ptr)
6023 mp->mp_ptrs[i] -= delta;
6026 base = (char *)mp + mp->mp_upper;
6027 len = ptr - mp->mp_upper + NODESIZE;
6028 memmove(base - delta, base, len);
6029 mp->mp_upper -= delta;
6031 node = NODEPTR(mp, indx);
6034 /* But even if no shift was needed, update ksize */
6036 node->mn_ksize = key->mv_size;
6039 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6045 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6047 /** Move a node from csrc to cdst.
6050 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6057 unsigned short flags;
6061 /* Mark src and dst as dirty. */
6062 if ((rc = mdb_page_touch(csrc)) ||
6063 (rc = mdb_page_touch(cdst)))
6066 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6067 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6068 key.mv_size = csrc->mc_db->md_pad;
6069 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6071 data.mv_data = NULL;
6075 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6076 assert(!((long)srcnode&1));
6077 srcpg = NODEPGNO(srcnode);
6078 flags = srcnode->mn_flags;
6079 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6080 unsigned int snum = csrc->mc_snum;
6082 /* must find the lowest key below src */
6083 mdb_page_search_lowest(csrc);
6084 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6085 key.mv_size = csrc->mc_db->md_pad;
6086 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6088 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6089 key.mv_size = NODEKSZ(s2);
6090 key.mv_data = NODEKEY(s2);
6092 csrc->mc_snum = snum--;
6093 csrc->mc_top = snum;
6095 key.mv_size = NODEKSZ(srcnode);
6096 key.mv_data = NODEKEY(srcnode);
6098 data.mv_size = NODEDSZ(srcnode);
6099 data.mv_data = NODEDATA(srcnode);
6101 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6102 unsigned int snum = cdst->mc_snum;
6105 /* must find the lowest key below dst */
6106 mdb_page_search_lowest(cdst);
6107 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6108 bkey.mv_size = cdst->mc_db->md_pad;
6109 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6111 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6112 bkey.mv_size = NODEKSZ(s2);
6113 bkey.mv_data = NODEKEY(s2);
6115 cdst->mc_snum = snum--;
6116 cdst->mc_top = snum;
6117 mdb_cursor_copy(cdst, &mn);
6119 rc = mdb_update_key(&mn, &bkey);
6124 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6125 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6126 csrc->mc_ki[csrc->mc_top],
6128 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6129 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6131 /* Add the node to the destination page.
6133 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6134 if (rc != MDB_SUCCESS)
6137 /* Delete the node from the source page.
6139 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6142 /* Adjust other cursors pointing to mp */
6143 MDB_cursor *m2, *m3;
6144 MDB_dbi dbi = csrc->mc_dbi;
6145 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6147 if (csrc->mc_flags & C_SUB)
6150 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6151 if (csrc->mc_flags & C_SUB)
6152 m3 = &m2->mc_xcursor->mx_cursor;
6155 if (m3 == csrc) continue;
6156 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6157 csrc->mc_ki[csrc->mc_top]) {
6158 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6159 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6164 /* Update the parent separators.
6166 if (csrc->mc_ki[csrc->mc_top] == 0) {
6167 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6168 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6169 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6171 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6172 key.mv_size = NODEKSZ(srcnode);
6173 key.mv_data = NODEKEY(srcnode);
6175 DPRINTF("update separator for source page %zu to [%s]",
6176 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6177 mdb_cursor_copy(csrc, &mn);
6180 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6183 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6185 indx_t ix = csrc->mc_ki[csrc->mc_top];
6186 nullkey.mv_size = 0;
6187 csrc->mc_ki[csrc->mc_top] = 0;
6188 rc = mdb_update_key(csrc, &nullkey);
6189 csrc->mc_ki[csrc->mc_top] = ix;
6190 assert(rc == MDB_SUCCESS);
6194 if (cdst->mc_ki[cdst->mc_top] == 0) {
6195 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6196 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6197 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6199 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6200 key.mv_size = NODEKSZ(srcnode);
6201 key.mv_data = NODEKEY(srcnode);
6203 DPRINTF("update separator for destination page %zu to [%s]",
6204 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6205 mdb_cursor_copy(cdst, &mn);
6208 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6211 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6213 indx_t ix = cdst->mc_ki[cdst->mc_top];
6214 nullkey.mv_size = 0;
6215 cdst->mc_ki[cdst->mc_top] = 0;
6216 rc = mdb_update_key(cdst, &nullkey);
6217 cdst->mc_ki[cdst->mc_top] = ix;
6218 assert(rc == MDB_SUCCESS);
6225 /** Merge one page into another.
6226 * The nodes from the page pointed to by \b csrc will
6227 * be copied to the page pointed to by \b cdst and then
6228 * the \b csrc page will be freed.
6229 * @param[in] csrc Cursor pointing to the source page.
6230 * @param[in] cdst Cursor pointing to the destination page.
6233 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6241 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6242 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6244 assert(csrc->mc_snum > 1); /* can't merge root page */
6245 assert(cdst->mc_snum > 1);
6247 /* Mark dst as dirty. */
6248 if ((rc = mdb_page_touch(cdst)))
6251 /* Move all nodes from src to dst.
6253 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6254 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6255 key.mv_size = csrc->mc_db->md_pad;
6256 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6257 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6258 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6259 if (rc != MDB_SUCCESS)
6261 key.mv_data = (char *)key.mv_data + key.mv_size;
6264 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6265 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6266 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6267 unsigned int snum = csrc->mc_snum;
6269 /* must find the lowest key below src */
6270 mdb_page_search_lowest(csrc);
6271 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6272 key.mv_size = csrc->mc_db->md_pad;
6273 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6275 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6276 key.mv_size = NODEKSZ(s2);
6277 key.mv_data = NODEKEY(s2);
6279 csrc->mc_snum = snum--;
6280 csrc->mc_top = snum;
6282 key.mv_size = srcnode->mn_ksize;
6283 key.mv_data = NODEKEY(srcnode);
6286 data.mv_size = NODEDSZ(srcnode);
6287 data.mv_data = NODEDATA(srcnode);
6288 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6289 if (rc != MDB_SUCCESS)
6294 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6295 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);
6297 /* Unlink the src page from parent and add to free list.
6299 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6300 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6303 rc = mdb_update_key(csrc, &key);
6309 rc = mdb_midl_append(&csrc->mc_txn->mt_free_pgs,
6310 csrc->mc_pg[csrc->mc_top]->mp_pgno);
6313 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6314 csrc->mc_db->md_leaf_pages--;
6316 csrc->mc_db->md_branch_pages--;
6318 /* Adjust other cursors pointing to mp */
6319 MDB_cursor *m2, *m3;
6320 MDB_dbi dbi = csrc->mc_dbi;
6321 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6323 if (csrc->mc_flags & C_SUB)
6326 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6327 if (csrc->mc_flags & C_SUB)
6328 m3 = &m2->mc_xcursor->mx_cursor;
6331 if (m3 == csrc) continue;
6332 if (m3->mc_snum < csrc->mc_snum) continue;
6333 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6334 m3->mc_pg[csrc->mc_top] = mp;
6335 m3->mc_ki[csrc->mc_top] += nkeys;
6339 mdb_cursor_pop(csrc);
6341 return mdb_rebalance(csrc);
6344 /** Copy the contents of a cursor.
6345 * @param[in] csrc The cursor to copy from.
6346 * @param[out] cdst The cursor to copy to.
6349 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6353 cdst->mc_txn = csrc->mc_txn;
6354 cdst->mc_dbi = csrc->mc_dbi;
6355 cdst->mc_db = csrc->mc_db;
6356 cdst->mc_dbx = csrc->mc_dbx;
6357 cdst->mc_snum = csrc->mc_snum;
6358 cdst->mc_top = csrc->mc_top;
6359 cdst->mc_flags = csrc->mc_flags;
6361 for (i=0; i<csrc->mc_snum; i++) {
6362 cdst->mc_pg[i] = csrc->mc_pg[i];
6363 cdst->mc_ki[i] = csrc->mc_ki[i];
6367 /** Rebalance the tree after a delete operation.
6368 * @param[in] mc Cursor pointing to the page where rebalancing
6370 * @return 0 on success, non-zero on failure.
6373 mdb_rebalance(MDB_cursor *mc)
6377 unsigned int ptop, minkeys;
6380 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6384 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6385 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6386 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6387 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6391 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6392 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6395 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6396 DPRINTF("no need to rebalance page %zu, above fill threshold",
6402 if (mc->mc_snum < 2) {
6403 MDB_page *mp = mc->mc_pg[0];
6405 DPUTS("Can't rebalance a subpage, ignoring");
6408 if (NUMKEYS(mp) == 0) {
6409 DPUTS("tree is completely empty");
6410 mc->mc_db->md_root = P_INVALID;
6411 mc->mc_db->md_depth = 0;
6412 mc->mc_db->md_leaf_pages = 0;
6413 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6416 /* Adjust cursors pointing to mp */
6420 MDB_cursor *m2, *m3;
6421 MDB_dbi dbi = mc->mc_dbi;
6423 if (mc->mc_flags & C_SUB)
6426 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6427 if (mc->mc_flags & C_SUB)
6428 m3 = &m2->mc_xcursor->mx_cursor;
6431 if (m3->mc_snum < mc->mc_snum) continue;
6432 if (m3->mc_pg[0] == mp) {
6438 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6439 DPUTS("collapsing root page!");
6440 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6443 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6444 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6447 mc->mc_db->md_depth--;
6448 mc->mc_db->md_branch_pages--;
6450 /* Adjust other cursors pointing to mp */
6451 MDB_cursor *m2, *m3;
6452 MDB_dbi dbi = mc->mc_dbi;
6454 if (mc->mc_flags & C_SUB)
6457 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6458 if (mc->mc_flags & C_SUB)
6459 m3 = &m2->mc_xcursor->mx_cursor;
6462 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6463 if (m3->mc_pg[0] == mp) {
6464 m3->mc_pg[0] = mc->mc_pg[0];
6471 DPUTS("root page doesn't need rebalancing");
6475 /* The parent (branch page) must have at least 2 pointers,
6476 * otherwise the tree is invalid.
6478 ptop = mc->mc_top-1;
6479 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6481 /* Leaf page fill factor is below the threshold.
6482 * Try to move keys from left or right neighbor, or
6483 * merge with a neighbor page.
6488 mdb_cursor_copy(mc, &mn);
6489 mn.mc_xcursor = NULL;
6491 if (mc->mc_ki[ptop] == 0) {
6492 /* We're the leftmost leaf in our parent.
6494 DPUTS("reading right neighbor");
6496 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6497 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6500 mn.mc_ki[mn.mc_top] = 0;
6501 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6503 /* There is at least one neighbor to the left.
6505 DPUTS("reading left neighbor");
6507 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6508 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6511 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6512 mc->mc_ki[mc->mc_top] = 0;
6515 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6516 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);
6518 /* If the neighbor page is above threshold and has enough keys,
6519 * move one key from it. Otherwise we should try to merge them.
6520 * (A branch page must never have less than 2 keys.)
6522 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6523 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6524 return mdb_node_move(&mn, mc);
6526 if (mc->mc_ki[ptop] == 0)
6527 rc = mdb_page_merge(&mn, mc);
6529 rc = mdb_page_merge(mc, &mn);
6530 mc->mc_flags &= ~C_INITIALIZED;
6535 /** Complete a delete operation started by #mdb_cursor_del(). */
6537 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6541 /* add overflow pages to free list */
6542 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6546 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6547 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6548 (rc = mdb_ovpage_free(mc, omp)))
6551 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6552 mc->mc_db->md_entries--;
6553 rc = mdb_rebalance(mc);
6554 if (rc != MDB_SUCCESS)
6555 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6556 /* if mc points past last node in page, invalidate */
6557 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6558 mc->mc_flags &= ~C_INITIALIZED;
6564 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6565 MDB_val *key, MDB_val *data)
6570 MDB_val rdata, *xdata;
6574 assert(key != NULL);
6576 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6578 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6581 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6585 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6589 mdb_cursor_init(&mc, txn, dbi, &mx);
6600 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6602 /* let mdb_page_split know about this cursor if needed:
6603 * delete will trigger a rebalance; if it needs to move
6604 * a node from one page to another, it will have to
6605 * update the parent's separator key(s). If the new sepkey
6606 * is larger than the current one, the parent page may
6607 * run out of space, triggering a split. We need this
6608 * cursor to be consistent until the end of the rebalance.
6610 mc.mc_next = txn->mt_cursors[dbi];
6611 txn->mt_cursors[dbi] = &mc;
6612 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6613 txn->mt_cursors[dbi] = mc.mc_next;
6618 /** Split a page and insert a new node.
6619 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6620 * The cursor will be updated to point to the actual page and index where
6621 * the node got inserted after the split.
6622 * @param[in] newkey The key for the newly inserted node.
6623 * @param[in] newdata The data for the newly inserted node.
6624 * @param[in] newpgno The page number, if the new node is a branch node.
6625 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6626 * @return 0 on success, non-zero on failure.
6629 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6630 unsigned int nflags)
6633 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6636 unsigned int i, j, split_indx, nkeys, pmax;
6638 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6640 MDB_page *mp, *rp, *pp;
6645 mp = mc->mc_pg[mc->mc_top];
6646 newindx = mc->mc_ki[mc->mc_top];
6648 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6649 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6650 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6652 /* Create a right sibling. */
6653 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6655 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6657 if (mc->mc_snum < 2) {
6658 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6660 /* shift current top to make room for new parent */
6661 mc->mc_pg[1] = mc->mc_pg[0];
6662 mc->mc_ki[1] = mc->mc_ki[0];
6665 mc->mc_db->md_root = pp->mp_pgno;
6666 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6667 mc->mc_db->md_depth++;
6670 /* Add left (implicit) pointer. */
6671 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6672 /* undo the pre-push */
6673 mc->mc_pg[0] = mc->mc_pg[1];
6674 mc->mc_ki[0] = mc->mc_ki[1];
6675 mc->mc_db->md_root = mp->mp_pgno;
6676 mc->mc_db->md_depth--;
6683 ptop = mc->mc_top-1;
6684 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6687 mc->mc_flags |= C_SPLITTING;
6688 mdb_cursor_copy(mc, &mn);
6689 mn.mc_pg[mn.mc_top] = rp;
6690 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6692 if (nflags & MDB_APPEND) {
6693 mn.mc_ki[mn.mc_top] = 0;
6695 split_indx = newindx;
6700 nkeys = NUMKEYS(mp);
6701 split_indx = nkeys / 2;
6702 if (newindx < split_indx)
6708 unsigned int lsize, rsize, ksize;
6709 /* Move half of the keys to the right sibling */
6711 x = mc->mc_ki[mc->mc_top] - split_indx;
6712 ksize = mc->mc_db->md_pad;
6713 split = LEAF2KEY(mp, split_indx, ksize);
6714 rsize = (nkeys - split_indx) * ksize;
6715 lsize = (nkeys - split_indx) * sizeof(indx_t);
6716 mp->mp_lower -= lsize;
6717 rp->mp_lower += lsize;
6718 mp->mp_upper += rsize - lsize;
6719 rp->mp_upper -= rsize - lsize;
6720 sepkey.mv_size = ksize;
6721 if (newindx == split_indx) {
6722 sepkey.mv_data = newkey->mv_data;
6724 sepkey.mv_data = split;
6727 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6728 memcpy(rp->mp_ptrs, split, rsize);
6729 sepkey.mv_data = rp->mp_ptrs;
6730 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6731 memcpy(ins, newkey->mv_data, ksize);
6732 mp->mp_lower += sizeof(indx_t);
6733 mp->mp_upper -= ksize - sizeof(indx_t);
6736 memcpy(rp->mp_ptrs, split, x * ksize);
6737 ins = LEAF2KEY(rp, x, ksize);
6738 memcpy(ins, newkey->mv_data, ksize);
6739 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6740 rp->mp_lower += sizeof(indx_t);
6741 rp->mp_upper -= ksize - sizeof(indx_t);
6742 mc->mc_ki[mc->mc_top] = x;
6743 mc->mc_pg[mc->mc_top] = rp;
6748 /* For leaf pages, check the split point based on what
6749 * fits where, since otherwise mdb_node_add can fail.
6751 * This check is only needed when the data items are
6752 * relatively large, such that being off by one will
6753 * make the difference between success or failure.
6755 * It's also relevant if a page happens to be laid out
6756 * such that one half of its nodes are all "small" and
6757 * the other half of its nodes are "large." If the new
6758 * item is also "large" and falls on the half with
6759 * "large" nodes, it also may not fit.
6762 unsigned int psize, nsize;
6763 /* Maximum free space in an empty page */
6764 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6765 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6766 if ((nkeys < 20) || (nsize > pmax/16)) {
6767 if (newindx <= split_indx) {
6770 for (i=0; i<split_indx; i++) {
6771 node = NODEPTR(mp, i);
6772 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6773 if (F_ISSET(node->mn_flags, F_BIGDATA))
6774 psize += sizeof(pgno_t);
6776 psize += NODEDSZ(node);
6780 split_indx = newindx;
6791 for (i=nkeys-1; i>=split_indx; i--) {
6792 node = NODEPTR(mp, i);
6793 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6794 if (F_ISSET(node->mn_flags, F_BIGDATA))
6795 psize += sizeof(pgno_t);
6797 psize += NODEDSZ(node);
6801 split_indx = newindx;
6812 /* First find the separating key between the split pages.
6813 * The case where newindx == split_indx is ambiguous; the
6814 * new item could go to the new page or stay on the original
6815 * page. If newpos == 1 it goes to the new page.
6817 if (newindx == split_indx && newpos) {
6818 sepkey.mv_size = newkey->mv_size;
6819 sepkey.mv_data = newkey->mv_data;
6821 node = NODEPTR(mp, split_indx);
6822 sepkey.mv_size = node->mn_ksize;
6823 sepkey.mv_data = NODEKEY(node);
6827 DPRINTF("separator is [%s]", DKEY(&sepkey));
6829 /* Copy separator key to the parent.
6831 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6835 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6838 if (mn.mc_snum == mc->mc_snum) {
6839 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6840 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6841 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6842 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6847 /* Right page might now have changed parent.
6848 * Check if left page also changed parent.
6850 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6851 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6852 for (i=0; i<ptop; i++) {
6853 mc->mc_pg[i] = mn.mc_pg[i];
6854 mc->mc_ki[i] = mn.mc_ki[i];
6856 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6857 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6861 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6864 mc->mc_flags ^= C_SPLITTING;
6865 if (rc != MDB_SUCCESS) {
6868 if (nflags & MDB_APPEND) {
6869 mc->mc_pg[mc->mc_top] = rp;
6870 mc->mc_ki[mc->mc_top] = 0;
6871 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6874 for (i=0; i<mc->mc_top; i++)
6875 mc->mc_ki[i] = mn.mc_ki[i];
6882 /* Move half of the keys to the right sibling. */
6884 /* grab a page to hold a temporary copy */
6885 copy = mdb_page_malloc(mc->mc_txn, 1);
6889 copy->mp_pgno = mp->mp_pgno;
6890 copy->mp_flags = mp->mp_flags;
6891 copy->mp_lower = PAGEHDRSZ;
6892 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6893 mc->mc_pg[mc->mc_top] = copy;
6894 for (i = j = 0; i <= nkeys; j++) {
6895 if (i == split_indx) {
6896 /* Insert in right sibling. */
6897 /* Reset insert index for right sibling. */
6898 if (i != newindx || (newpos ^ ins_new)) {
6900 mc->mc_pg[mc->mc_top] = rp;
6904 if (i == newindx && !ins_new) {
6905 /* Insert the original entry that caused the split. */
6906 rkey.mv_data = newkey->mv_data;
6907 rkey.mv_size = newkey->mv_size;
6916 /* Update index for the new key. */
6917 mc->mc_ki[mc->mc_top] = j;
6918 } else if (i == nkeys) {
6921 node = NODEPTR(mp, i);
6922 rkey.mv_data = NODEKEY(node);
6923 rkey.mv_size = node->mn_ksize;
6925 xdata.mv_data = NODEDATA(node);
6926 xdata.mv_size = NODEDSZ(node);
6929 pgno = NODEPGNO(node);
6930 flags = node->mn_flags;
6935 if (!IS_LEAF(mp) && j == 0) {
6936 /* First branch index doesn't need key data. */
6940 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6944 nkeys = NUMKEYS(copy);
6945 for (i=0; i<nkeys; i++)
6946 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6947 mp->mp_lower = copy->mp_lower;
6948 mp->mp_upper = copy->mp_upper;
6949 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6950 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6952 /* reset back to original page */
6953 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6954 mc->mc_pg[mc->mc_top] = mp;
6955 if (nflags & MDB_RESERVE) {
6956 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6957 if (!(node->mn_flags & F_BIGDATA))
6958 newdata->mv_data = NODEDATA(node);
6962 /* Make sure mc_ki is still valid.
6964 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6965 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6966 for (i=0; i<ptop; i++) {
6967 mc->mc_pg[i] = mn.mc_pg[i];
6968 mc->mc_ki[i] = mn.mc_ki[i];
6970 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6971 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6975 /* return tmp page to freelist */
6976 mdb_page_free(mc->mc_txn->mt_env, copy);
6979 /* Adjust other cursors pointing to mp */
6980 MDB_cursor *m2, *m3;
6981 MDB_dbi dbi = mc->mc_dbi;
6982 int fixup = NUMKEYS(mp);
6984 if (mc->mc_flags & C_SUB)
6987 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6988 if (mc->mc_flags & C_SUB)
6989 m3 = &m2->mc_xcursor->mx_cursor;
6994 if (!(m3->mc_flags & C_INITIALIZED))
6996 if (m3->mc_flags & C_SPLITTING)
7001 for (k=m3->mc_top; k>=0; k--) {
7002 m3->mc_ki[k+1] = m3->mc_ki[k];
7003 m3->mc_pg[k+1] = m3->mc_pg[k];
7005 if (m3->mc_ki[0] >= split_indx) {
7010 m3->mc_pg[0] = mc->mc_pg[0];
7014 if (m3->mc_pg[mc->mc_top] == mp) {
7015 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7016 m3->mc_ki[mc->mc_top]++;
7017 if (m3->mc_ki[mc->mc_top] >= fixup) {
7018 m3->mc_pg[mc->mc_top] = rp;
7019 m3->mc_ki[mc->mc_top] -= fixup;
7020 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7022 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7023 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7032 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7033 MDB_val *key, MDB_val *data, unsigned int flags)
7038 assert(key != NULL);
7039 assert(data != NULL);
7041 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7044 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7048 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7052 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7055 mdb_cursor_init(&mc, txn, dbi, &mx);
7056 return mdb_cursor_put(&mc, key, data, flags);
7060 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7062 if ((flag & CHANGEABLE) != flag)
7065 env->me_flags |= flag;
7067 env->me_flags &= ~flag;
7072 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7077 *arg = env->me_flags;
7082 mdb_env_get_path(MDB_env *env, const char **arg)
7087 *arg = env->me_path;
7091 /** Common code for #mdb_stat() and #mdb_env_stat().
7092 * @param[in] env the environment to operate in.
7093 * @param[in] db the #MDB_db record containing the stats to return.
7094 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7095 * @return 0, this function always succeeds.
7098 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7100 arg->ms_psize = env->me_psize;
7101 arg->ms_depth = db->md_depth;
7102 arg->ms_branch_pages = db->md_branch_pages;
7103 arg->ms_leaf_pages = db->md_leaf_pages;
7104 arg->ms_overflow_pages = db->md_overflow_pages;
7105 arg->ms_entries = db->md_entries;
7110 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7114 if (env == NULL || arg == NULL)
7117 toggle = mdb_env_pick_meta(env);
7119 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7123 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7127 if (env == NULL || arg == NULL)
7130 toggle = mdb_env_pick_meta(env);
7131 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7132 arg->me_mapsize = env->me_mapsize;
7133 arg->me_maxreaders = env->me_maxreaders;
7134 arg->me_numreaders = env->me_numreaders;
7135 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7136 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7140 /** Set the default comparison functions for a database.
7141 * Called immediately after a database is opened to set the defaults.
7142 * The user can then override them with #mdb_set_compare() or
7143 * #mdb_set_dupsort().
7144 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7145 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7148 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7150 uint16_t f = txn->mt_dbs[dbi].md_flags;
7152 txn->mt_dbxs[dbi].md_cmp =
7153 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7154 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7156 txn->mt_dbxs[dbi].md_dcmp =
7157 !(f & MDB_DUPSORT) ? 0 :
7158 ((f & MDB_INTEGERDUP)
7159 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7160 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7163 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7168 int rc, dbflag, exact;
7169 unsigned int unused = 0;
7172 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7173 mdb_default_cmp(txn, FREE_DBI);
7176 if ((flags & VALID_FLAGS) != flags)
7182 if (flags & PERSISTENT_FLAGS) {
7183 uint16_t f2 = flags & PERSISTENT_FLAGS;
7184 /* make sure flag changes get committed */
7185 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7186 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7187 txn->mt_flags |= MDB_TXN_DIRTY;
7190 mdb_default_cmp(txn, MAIN_DBI);
7194 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7195 mdb_default_cmp(txn, MAIN_DBI);
7198 /* Is the DB already open? */
7200 for (i=2; i<txn->mt_numdbs; i++) {
7201 if (!txn->mt_dbxs[i].md_name.mv_size) {
7202 /* Remember this free slot */
7203 if (!unused) unused = i;
7206 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7207 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7213 /* If no free slot and max hit, fail */
7214 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7215 return MDB_DBS_FULL;
7217 /* Cannot mix named databases with some mainDB flags */
7218 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7219 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7221 /* Find the DB info */
7222 dbflag = DB_NEW|DB_VALID;
7225 key.mv_data = (void *)name;
7226 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7227 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7228 if (rc == MDB_SUCCESS) {
7229 /* make sure this is actually a DB */
7230 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7231 if (!(node->mn_flags & F_SUBDATA))
7233 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7234 /* Create if requested */
7236 data.mv_size = sizeof(MDB_db);
7237 data.mv_data = &dummy;
7238 memset(&dummy, 0, sizeof(dummy));
7239 dummy.md_root = P_INVALID;
7240 dummy.md_flags = flags & PERSISTENT_FLAGS;
7241 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7245 /* OK, got info, add to table */
7246 if (rc == MDB_SUCCESS) {
7247 unsigned int slot = unused ? unused : txn->mt_numdbs;
7248 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7249 txn->mt_dbxs[slot].md_name.mv_size = len;
7250 txn->mt_dbxs[slot].md_rel = NULL;
7251 txn->mt_dbflags[slot] = dbflag;
7252 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7254 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7255 mdb_default_cmp(txn, slot);
7264 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7266 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7269 if (txn->mt_dbflags[dbi] & DB_STALE) {
7272 /* Stale, must read the DB's root. cursor_init does it for us. */
7273 mdb_cursor_init(&mc, txn, dbi, &mx);
7275 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7278 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7281 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7283 ptr = env->me_dbxs[dbi].md_name.mv_data;
7284 env->me_dbxs[dbi].md_name.mv_data = NULL;
7285 env->me_dbxs[dbi].md_name.mv_size = 0;
7286 env->me_dbflags[dbi] = 0;
7290 /** Add all the DB's pages to the free list.
7291 * @param[in] mc Cursor on the DB to free.
7292 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7293 * @return 0 on success, non-zero on failure.
7296 mdb_drop0(MDB_cursor *mc, int subs)
7300 rc = mdb_page_search(mc, NULL, 0);
7301 if (rc == MDB_SUCCESS) {
7302 MDB_txn *txn = mc->mc_txn;
7307 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7308 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7311 mdb_cursor_copy(mc, &mx);
7312 while (mc->mc_snum > 0) {
7313 MDB_page *mp = mc->mc_pg[mc->mc_top];
7314 unsigned n = NUMKEYS(mp);
7316 for (i=0; i<n; i++) {
7317 ni = NODEPTR(mp, i);
7318 if (ni->mn_flags & F_BIGDATA) {
7321 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7322 rc = mdb_page_get(txn, pg, &omp, NULL);
7325 assert(IS_OVERFLOW(omp));
7326 rc = mdb_midl_append_range(&txn->mt_free_pgs,
7330 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7331 mdb_xcursor_init1(mc, ni);
7332 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7338 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
7340 for (i=0; i<n; i++) {
7342 ni = NODEPTR(mp, i);
7345 mdb_midl_xappend(txn->mt_free_pgs, pg);
7350 mc->mc_ki[mc->mc_top] = i;
7351 rc = mdb_cursor_sibling(mc, 1);
7353 /* no more siblings, go back to beginning
7354 * of previous level.
7358 for (i=1; i<mc->mc_snum; i++) {
7360 mc->mc_pg[i] = mx.mc_pg[i];
7365 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
7366 } else if (rc == MDB_NOTFOUND) {
7372 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7374 MDB_cursor *mc, *m2;
7377 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7380 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7383 rc = mdb_cursor_open(txn, dbi, &mc);
7387 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7388 /* Invalidate the dropped DB's cursors */
7389 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7390 m2->mc_flags &= ~C_INITIALIZED;
7394 /* Can't delete the main DB */
7395 if (del && dbi > MAIN_DBI) {
7396 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7398 txn->mt_dbflags[dbi] = DB_STALE;
7399 mdb_dbi_close(txn->mt_env, dbi);
7402 /* reset the DB record, mark it dirty */
7403 txn->mt_dbflags[dbi] |= DB_DIRTY;
7404 txn->mt_dbs[dbi].md_depth = 0;
7405 txn->mt_dbs[dbi].md_branch_pages = 0;
7406 txn->mt_dbs[dbi].md_leaf_pages = 0;
7407 txn->mt_dbs[dbi].md_overflow_pages = 0;
7408 txn->mt_dbs[dbi].md_entries = 0;
7409 txn->mt_dbs[dbi].md_root = P_INVALID;
7411 txn->mt_flags |= MDB_TXN_DIRTY;
7414 mdb_cursor_close(mc);
7418 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7420 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7423 txn->mt_dbxs[dbi].md_cmp = cmp;
7427 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7429 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7432 txn->mt_dbxs[dbi].md_dcmp = cmp;
7436 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7438 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7441 txn->mt_dbxs[dbi].md_rel = rel;
7445 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7447 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7450 txn->mt_dbxs[dbi].md_relctx = ctx;
projects / openldap / blob