2 * @brief memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2013 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
38 #include <sys/types.h>
40 #include <sys/param.h>
46 #ifdef HAVE_SYS_FILE_H
63 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
64 #include <netinet/in.h>
65 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
68 #if defined(__APPLE__) || defined (BSD)
69 # define MDB_USE_POSIX_SEM 1
70 # define MDB_FDATASYNC fsync
71 #elif defined(ANDROID)
72 # define MDB_FDATASYNC fsync
77 #ifdef MDB_USE_POSIX_SEM
78 #include <semaphore.h>
83 #include <valgrind/memcheck.h>
84 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
85 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
86 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
87 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
88 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
90 #define VGMEMP_CREATE(h,r,z)
91 #define VGMEMP_ALLOC(h,a,s)
92 #define VGMEMP_FREE(h,a)
93 #define VGMEMP_DESTROY(h)
94 #define VGMEMP_DEFINED(a,s)
98 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
99 /* Solaris just defines one or the other */
100 # define LITTLE_ENDIAN 1234
101 # define BIG_ENDIAN 4321
102 # ifdef _LITTLE_ENDIAN
103 # define BYTE_ORDER LITTLE_ENDIAN
105 # define BYTE_ORDER BIG_ENDIAN
108 # define BYTE_ORDER __BYTE_ORDER
112 #ifndef LITTLE_ENDIAN
113 #define LITTLE_ENDIAN __LITTLE_ENDIAN
116 #define BIG_ENDIAN __BIG_ENDIAN
119 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
120 #define MISALIGNED_OK 1
126 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
127 # error "Unknown or unsupported endianness (BYTE_ORDER)"
128 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
129 # error "Two's complement, reasonably sized integer types, please"
132 /** @defgroup internal MDB Internals
135 /** @defgroup compat Windows Compatibility Macros
136 * A bunch of macros to minimize the amount of platform-specific ifdefs
137 * needed throughout the rest of the code. When the features this library
138 * needs are similar enough to POSIX to be hidden in a one-or-two line
139 * replacement, this macro approach is used.
143 #define pthread_t DWORD
144 #define pthread_mutex_t HANDLE
145 #define pthread_key_t DWORD
146 #define pthread_self() GetCurrentThreadId()
147 #define pthread_key_create(x,y) \
148 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
149 #define pthread_key_delete(x) TlsFree(x)
150 #define pthread_getspecific(x) TlsGetValue(x)
151 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
152 #define pthread_mutex_unlock(x) ReleaseMutex(x)
153 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
154 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
158 #define getpid() GetCurrentProcessId()
159 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
160 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
161 #define ErrCode() GetLastError()
162 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
163 #define close(fd) CloseHandle(fd)
164 #define munmap(ptr,len) UnmapViewOfFile(ptr)
167 #ifdef MDB_USE_POSIX_SEM
169 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
170 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
171 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
172 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
175 mdb_sem_wait(sem_t *sem)
178 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
183 /** Lock the reader mutex.
185 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
186 /** Unlock the reader mutex.
188 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
190 /** Lock the writer mutex.
191 * Only a single write transaction is allowed at a time. Other writers
192 * will block waiting for this mutex.
194 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
195 /** Unlock the writer mutex.
197 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
198 #endif /* MDB_USE_POSIX_SEM */
200 /** Get the error code for the last failed system function.
202 #define ErrCode() errno
204 /** An abstraction for a file handle.
205 * On POSIX systems file handles are small integers. On Windows
206 * they're opaque pointers.
210 /** A value for an invalid file handle.
211 * Mainly used to initialize file variables and signify that they are
214 #define INVALID_HANDLE_VALUE (-1)
216 /** Get the size of a memory page for the system.
217 * This is the basic size that the platform's memory manager uses, and is
218 * fundamental to the use of memory-mapped files.
220 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
223 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
226 #define MNAME_LEN (sizeof(pthread_mutex_t))
232 /** A flag for opening a file and requesting synchronous data writes.
233 * This is only used when writing a meta page. It's not strictly needed;
234 * we could just do a normal write and then immediately perform a flush.
235 * But if this flag is available it saves us an extra system call.
237 * @note If O_DSYNC is undefined but exists in /usr/include,
238 * preferably set some compiler flag to get the definition.
239 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
242 # define MDB_DSYNC O_DSYNC
246 /** Function for flushing the data of a file. Define this to fsync
247 * if fdatasync() is not supported.
249 #ifndef MDB_FDATASYNC
250 # define MDB_FDATASYNC fdatasync
254 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
265 /** A page number in the database.
266 * Note that 64 bit page numbers are overkill, since pages themselves
267 * already represent 12-13 bits of addressable memory, and the OS will
268 * always limit applications to a maximum of 63 bits of address space.
270 * @note In the #MDB_node structure, we only store 48 bits of this value,
271 * which thus limits us to only 60 bits of addressable data.
273 typedef MDB_ID pgno_t;
275 /** A transaction ID.
276 * See struct MDB_txn.mt_txnid for details.
278 typedef MDB_ID txnid_t;
280 /** @defgroup debug Debug Macros
284 /** Enable debug output.
285 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
286 * read from and written to the database (used for free space management).
291 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
294 # define DPRINTF (void) /* Vararg macros may be unsupported */
296 static int mdb_debug;
297 static txnid_t mdb_debug_start;
299 /** Print a debug message with printf formatting. */
300 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
301 ((void) ((mdb_debug) && \
302 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
304 # define DPRINTF(fmt, ...) ((void) 0)
305 # define MDB_DEBUG_SKIP
307 /** Print a debug string.
308 * The string is printed literally, with no format processing.
310 #define DPUTS(arg) DPRINTF("%s", arg)
313 /** A default memory page size.
314 * The actual size is platform-dependent, but we use this for
315 * boot-strapping. We probably should not be using this any more.
316 * The #GET_PAGESIZE() macro is used to get the actual size.
318 * Note that we don't currently support Huge pages. On Linux,
319 * regular data files cannot use Huge pages, and in general
320 * Huge pages aren't actually pageable. We rely on the OS
321 * demand-pager to read our data and page it out when memory
322 * pressure from other processes is high. So until OSs have
323 * actual paging support for Huge pages, they're not viable.
325 #define MDB_PAGESIZE 4096
327 /** The minimum number of keys required in a database page.
328 * Setting this to a larger value will place a smaller bound on the
329 * maximum size of a data item. Data items larger than this size will
330 * be pushed into overflow pages instead of being stored directly in
331 * the B-tree node. This value used to default to 4. With a page size
332 * of 4096 bytes that meant that any item larger than 1024 bytes would
333 * go into an overflow page. That also meant that on average 2-3KB of
334 * each overflow page was wasted space. The value cannot be lower than
335 * 2 because then there would no longer be a tree structure. With this
336 * value, items larger than 2KB will go into overflow pages, and on
337 * average only 1KB will be wasted.
339 #define MDB_MINKEYS 2
341 /** A stamp that identifies a file as an MDB file.
342 * There's nothing special about this value other than that it is easily
343 * recognizable, and it will reflect any byte order mismatches.
345 #define MDB_MAGIC 0xBEEFC0DE
347 /** The version number for a database's file format. */
348 #define MDB_VERSION 1
350 /** @brief The maximum size of a key in the database.
352 * We require that keys all fit onto a regular page. This limit
353 * could be raised a bit further if needed; to something just
354 * under #MDB_PAGESIZE / #MDB_MINKEYS.
356 * Note that data items in an #MDB_DUPSORT database are actually keys
357 * of a subDB, so they're also limited to this size.
359 #ifndef MDB_MAXKEYSIZE
360 #define MDB_MAXKEYSIZE 511
363 /** @brief The maximum size of a data item.
365 * We only store a 32 bit value for node sizes.
367 #define MAXDATASIZE 0xffffffffUL
372 * This is used for printing a hex dump of a key's contents.
374 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
375 /** Display a key in hex.
377 * Invoke a function to display a key in hex.
379 #define DKEY(x) mdb_dkey(x, kbuf)
381 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
385 /** An invalid page number.
386 * Mainly used to denote an empty tree.
388 #define P_INVALID (~(pgno_t)0)
390 /** Test if the flags \b f are set in a flag word \b w. */
391 #define F_ISSET(w, f) (((w) & (f)) == (f))
393 /** Used for offsets within a single page.
394 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
397 typedef uint16_t indx_t;
399 /** Default size of memory map.
400 * This is certainly too small for any actual applications. Apps should always set
401 * the size explicitly using #mdb_env_set_mapsize().
403 #define DEFAULT_MAPSIZE 1048576
405 /** @defgroup readers Reader Lock Table
406 * Readers don't acquire any locks for their data access. Instead, they
407 * simply record their transaction ID in the reader table. The reader
408 * mutex is needed just to find an empty slot in the reader table. The
409 * slot's address is saved in thread-specific data so that subsequent read
410 * transactions started by the same thread need no further locking to proceed.
412 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
414 * No reader table is used if the database is on a read-only filesystem.
416 * Since the database uses multi-version concurrency control, readers don't
417 * actually need any locking. This table is used to keep track of which
418 * readers are using data from which old transactions, so that we'll know
419 * when a particular old transaction is no longer in use. Old transactions
420 * that have discarded any data pages can then have those pages reclaimed
421 * for use by a later write transaction.
423 * The lock table is constructed such that reader slots are aligned with the
424 * processor's cache line size. Any slot is only ever used by one thread.
425 * This alignment guarantees that there will be no contention or cache
426 * thrashing as threads update their own slot info, and also eliminates
427 * any need for locking when accessing a slot.
429 * A writer thread will scan every slot in the table to determine the oldest
430 * outstanding reader transaction. Any freed pages older than this will be
431 * reclaimed by the writer. The writer doesn't use any locks when scanning
432 * this table. This means that there's no guarantee that the writer will
433 * see the most up-to-date reader info, but that's not required for correct
434 * operation - all we need is to know the upper bound on the oldest reader,
435 * we don't care at all about the newest reader. So the only consequence of
436 * reading stale information here is that old pages might hang around a
437 * while longer before being reclaimed. That's actually good anyway, because
438 * the longer we delay reclaiming old pages, the more likely it is that a
439 * string of contiguous pages can be found after coalescing old pages from
440 * many old transactions together.
443 /** Number of slots in the reader table.
444 * This value was chosen somewhat arbitrarily. 126 readers plus a
445 * couple mutexes fit exactly into 8KB on my development machine.
446 * Applications should set the table size using #mdb_env_set_maxreaders().
448 #define DEFAULT_READERS 126
450 /** The size of a CPU cache line in bytes. We want our lock structures
451 * aligned to this size to avoid false cache line sharing in the
453 * This value works for most CPUs. For Itanium this should be 128.
459 /** The information we store in a single slot of the reader table.
460 * In addition to a transaction ID, we also record the process and
461 * thread ID that owns a slot, so that we can detect stale information,
462 * e.g. threads or processes that went away without cleaning up.
463 * @note We currently don't check for stale records. We simply re-init
464 * the table when we know that we're the only process opening the
467 typedef struct MDB_rxbody {
468 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
469 * Multiple readers that start at the same time will probably have the
470 * same ID here. Again, it's not important to exclude them from
471 * anything; all we need to know is which version of the DB they
472 * started from so we can avoid overwriting any data used in that
473 * particular version.
476 /** The process ID of the process owning this reader txn. */
478 /** The thread ID of the thread owning this txn. */
482 /** The actual reader record, with cacheline padding. */
483 typedef struct MDB_reader {
486 /** shorthand for mrb_txnid */
487 #define mr_txnid mru.mrx.mrb_txnid
488 #define mr_pid mru.mrx.mrb_pid
489 #define mr_tid mru.mrx.mrb_tid
490 /** cache line alignment */
491 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
495 /** The header for the reader table.
496 * The table resides in a memory-mapped file. (This is a different file
497 * than is used for the main database.)
499 * For POSIX the actual mutexes reside in the shared memory of this
500 * mapped file. On Windows, mutexes are named objects allocated by the
501 * kernel; we store the mutex names in this mapped file so that other
502 * processes can grab them. This same approach is also used on
503 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
504 * process-shared POSIX mutexes. For these cases where a named object
505 * is used, the object name is derived from a 64 bit FNV hash of the
506 * environment pathname. As such, naming collisions are extremely
507 * unlikely. If a collision occurs, the results are unpredictable.
509 typedef struct MDB_txbody {
510 /** Stamp identifying this as an MDB file. It must be set
513 /** Version number of this lock file. Must be set to #MDB_VERSION. */
514 uint32_t mtb_version;
515 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
516 char mtb_rmname[MNAME_LEN];
518 /** Mutex protecting access to this table.
519 * This is the reader lock that #LOCK_MUTEX_R acquires.
521 pthread_mutex_t mtb_mutex;
523 /** The ID of the last transaction committed to the database.
524 * This is recorded here only for convenience; the value can always
525 * be determined by reading the main database meta pages.
528 /** The number of slots that have been used in the reader table.
529 * This always records the maximum count, it is not decremented
530 * when readers release their slots.
532 unsigned mtb_numreaders;
535 /** The actual reader table definition. */
536 typedef struct MDB_txninfo {
539 #define mti_magic mt1.mtb.mtb_magic
540 #define mti_version mt1.mtb.mtb_version
541 #define mti_mutex mt1.mtb.mtb_mutex
542 #define mti_rmname mt1.mtb.mtb_rmname
543 #define mti_txnid mt1.mtb.mtb_txnid
544 #define mti_numreaders mt1.mtb.mtb_numreaders
545 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
548 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
549 char mt2_wmname[MNAME_LEN];
550 #define mti_wmname mt2.mt2_wmname
552 pthread_mutex_t mt2_wmutex;
553 #define mti_wmutex mt2.mt2_wmutex
555 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
557 MDB_reader mti_readers[1];
561 /** Common header for all page types.
562 * Overflow records occupy a number of contiguous pages with no
563 * headers on any page after the first.
565 typedef struct MDB_page {
566 #define mp_pgno mp_p.p_pgno
567 #define mp_next mp_p.p_next
569 pgno_t p_pgno; /**< page number */
570 void * p_next; /**< for in-memory list of freed structs */
573 /** @defgroup mdb_page Page Flags
575 * Flags for the page headers.
578 #define P_BRANCH 0x01 /**< branch page */
579 #define P_LEAF 0x02 /**< leaf page */
580 #define P_OVERFLOW 0x04 /**< overflow page */
581 #define P_META 0x08 /**< meta page */
582 #define P_DIRTY 0x10 /**< dirty page */
583 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
584 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
586 uint16_t mp_flags; /**< @ref mdb_page */
587 #define mp_lower mp_pb.pb.pb_lower
588 #define mp_upper mp_pb.pb.pb_upper
589 #define mp_pages mp_pb.pb_pages
592 indx_t pb_lower; /**< lower bound of free space */
593 indx_t pb_upper; /**< upper bound of free space */
595 uint32_t pb_pages; /**< number of overflow pages */
597 indx_t mp_ptrs[1]; /**< dynamic size */
600 /** Size of the page header, excluding dynamic data at the end */
601 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
603 /** Address of first usable data byte in a page, after the header */
604 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
606 /** Number of nodes on a page */
607 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
609 /** The amount of space remaining in the page */
610 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
612 /** The percentage of space used in the page, in tenths of a percent. */
613 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
614 ((env)->me_psize - PAGEHDRSZ))
615 /** The minimum page fill factor, in tenths of a percent.
616 * Pages emptier than this are candidates for merging.
618 #define FILL_THRESHOLD 250
620 /** Test if a page is a leaf page */
621 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
622 /** Test if a page is a LEAF2 page */
623 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
624 /** Test if a page is a branch page */
625 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
626 /** Test if a page is an overflow page */
627 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
628 /** Test if a page is a sub page */
629 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
631 /** The number of overflow pages needed to store the given size. */
632 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
634 /** Header for a single key/data pair within a page.
635 * We guarantee 2-byte alignment for nodes.
637 typedef struct MDB_node {
638 /** lo and hi are used for data size on leaf nodes and for
639 * child pgno on branch nodes. On 64 bit platforms, flags
640 * is also used for pgno. (Branch nodes have no flags).
641 * They are in host byte order in case that lets some
642 * accesses be optimized into a 32-bit word access.
644 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
645 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
646 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
647 /** @defgroup mdb_node Node Flags
649 * Flags for node headers.
652 #define F_BIGDATA 0x01 /**< data put on overflow page */
653 #define F_SUBDATA 0x02 /**< data is a sub-database */
654 #define F_DUPDATA 0x04 /**< data has duplicates */
656 /** valid flags for #mdb_node_add() */
657 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
660 unsigned short mn_flags; /**< @ref mdb_node */
661 unsigned short mn_ksize; /**< key size */
662 char mn_data[1]; /**< key and data are appended here */
665 /** Size of the node header, excluding dynamic data at the end */
666 #define NODESIZE offsetof(MDB_node, mn_data)
668 /** Bit position of top word in page number, for shifting mn_flags */
669 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
671 /** Size of a node in a branch page with a given key.
672 * This is just the node header plus the key, there is no data.
674 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
676 /** Size of a node in a leaf page with a given key and data.
677 * This is node header plus key plus data size.
679 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
681 /** Address of node \b i in page \b p */
682 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
684 /** Address of the key for the node */
685 #define NODEKEY(node) (void *)((node)->mn_data)
687 /** Address of the data for a node */
688 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
690 /** Get the page number pointed to by a branch node */
691 #define NODEPGNO(node) \
692 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
693 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
694 /** Set the page number in a branch node */
695 #define SETPGNO(node,pgno) do { \
696 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
697 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
699 /** Get the size of the data in a leaf node */
700 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
701 /** Set the size of the data for a leaf node */
702 #define SETDSZ(node,size) do { \
703 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
704 /** The size of a key in a node */
705 #define NODEKSZ(node) ((node)->mn_ksize)
707 /** Copy a page number from src to dst */
709 #define COPY_PGNO(dst,src) dst = src
711 #if SIZE_MAX > 4294967295UL
712 #define COPY_PGNO(dst,src) do { \
713 unsigned short *s, *d; \
714 s = (unsigned short *)&(src); \
715 d = (unsigned short *)&(dst); \
722 #define COPY_PGNO(dst,src) do { \
723 unsigned short *s, *d; \
724 s = (unsigned short *)&(src); \
725 d = (unsigned short *)&(dst); \
731 /** The address of a key in a LEAF2 page.
732 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
733 * There are no node headers, keys are stored contiguously.
735 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
737 /** Set the \b node's key into \b key, if requested. */
738 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
739 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
741 /** Information about a single database in the environment. */
742 typedef struct MDB_db {
743 uint32_t md_pad; /**< also ksize for LEAF2 pages */
744 uint16_t md_flags; /**< @ref mdb_dbi_open */
745 uint16_t md_depth; /**< depth of this tree */
746 pgno_t md_branch_pages; /**< number of internal pages */
747 pgno_t md_leaf_pages; /**< number of leaf pages */
748 pgno_t md_overflow_pages; /**< number of overflow pages */
749 size_t md_entries; /**< number of data items */
750 pgno_t md_root; /**< the root page of this tree */
753 /** mdb_dbi_open flags */
754 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
755 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
756 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
757 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
759 /** Handle for the DB used to track free pages. */
761 /** Handle for the default DB. */
764 /** Meta page content. */
765 typedef struct MDB_meta {
766 /** Stamp identifying this as an MDB file. It must be set
769 /** Version number of this lock file. Must be set to #MDB_VERSION. */
771 void *mm_address; /**< address for fixed mapping */
772 size_t mm_mapsize; /**< size of mmap region */
773 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
774 /** The size of pages used in this DB */
775 #define mm_psize mm_dbs[0].md_pad
776 /** Any persistent environment flags. @ref mdb_env */
777 #define mm_flags mm_dbs[0].md_flags
778 pgno_t mm_last_pg; /**< last used page in file */
779 txnid_t mm_txnid; /**< txnid that committed this page */
782 /** Buffer for a stack-allocated dirty page.
783 * The members define size and alignment, and silence type
784 * aliasing warnings. They are not used directly; that could
785 * mean incorrectly using several union members in parallel.
787 typedef union MDB_pagebuf {
788 char mb_raw[MDB_PAGESIZE];
791 char mm_pad[PAGEHDRSZ];
796 /** Auxiliary DB info.
797 * The information here is mostly static/read-only. There is
798 * only a single copy of this record in the environment.
800 typedef struct MDB_dbx {
801 MDB_val md_name; /**< name of the database */
802 MDB_cmp_func *md_cmp; /**< function for comparing keys */
803 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
804 MDB_rel_func *md_rel; /**< user relocate function */
805 void *md_relctx; /**< user-provided context for md_rel */
808 /** A database transaction.
809 * Every operation requires a transaction handle.
812 MDB_txn *mt_parent; /**< parent of a nested txn */
813 MDB_txn *mt_child; /**< nested txn under this txn */
814 pgno_t mt_next_pgno; /**< next unallocated page */
815 /** The ID of this transaction. IDs are integers incrementing from 1.
816 * Only committed write transactions increment the ID. If a transaction
817 * aborts, the ID may be re-used by the next writer.
820 MDB_env *mt_env; /**< the DB environment */
821 /** The list of pages that became unused during this transaction.
825 MDB_ID2L dirty_list; /**< for write txns: modified pages */
826 MDB_reader *reader; /**< this thread's reader table slot or NULL */
828 /** Array of records for each DB known in the environment. */
830 /** Array of MDB_db records for each known DB */
832 /** @defgroup mt_dbflag Transaction DB Flags
836 #define DB_DIRTY 0x01 /**< DB was written in this txn */
837 #define DB_STALE 0x02 /**< DB record is older than txnID */
838 #define DB_NEW 0x04 /**< DB handle opened in this txn */
839 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
841 /** In write txns, array of cursors for each DB */
842 MDB_cursor **mt_cursors;
843 /** Array of flags for each DB */
844 unsigned char *mt_dbflags;
845 /** Number of DB records in use. This number only ever increments;
846 * we don't decrement it when individual DB handles are closed.
850 /** @defgroup mdb_txn Transaction Flags
854 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
855 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
856 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
858 unsigned int mt_flags; /**< @ref mdb_txn */
859 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
860 unsigned int mt_dirty_room;
861 /** Tracks which of the two meta pages was used at the start
862 * of this transaction.
864 unsigned int mt_toggle;
867 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
868 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
869 * raise this on a 64 bit machine.
871 #define CURSOR_STACK 32
875 /** Cursors are used for all DB operations */
877 /** Next cursor on this DB in this txn */
879 /** Original cursor if this is a shadow */
881 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
882 struct MDB_xcursor *mc_xcursor;
883 /** The transaction that owns this cursor */
885 /** The database handle this cursor operates on */
887 /** The database record for this cursor */
889 /** The database auxiliary record for this cursor */
891 /** The @ref mt_dbflag for this database */
892 unsigned char *mc_dbflag;
893 unsigned short mc_snum; /**< number of pushed pages */
894 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
895 /** @defgroup mdb_cursor Cursor Flags
897 * Cursor state flags.
900 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
901 #define C_EOF 0x02 /**< No more data */
902 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
903 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
904 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
905 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
906 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
908 unsigned int mc_flags; /**< @ref mdb_cursor */
909 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
910 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
913 /** Context for sorted-dup records.
914 * We could have gone to a fully recursive design, with arbitrarily
915 * deep nesting of sub-databases. But for now we only handle these
916 * levels - main DB, optional sub-DB, sorted-duplicate DB.
918 typedef struct MDB_xcursor {
919 /** A sub-cursor for traversing the Dup DB */
920 MDB_cursor mx_cursor;
921 /** The database record for this Dup DB */
923 /** The auxiliary DB record for this Dup DB */
925 /** The @ref mt_dbflag for this Dup DB */
926 unsigned char mx_dbflag;
929 /** State of FreeDB old pages, stored in the MDB_env */
930 typedef struct MDB_pgstate {
931 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
932 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
935 /** The database environment. */
937 HANDLE me_fd; /**< The main data file */
938 HANDLE me_lfd; /**< The lock file */
939 HANDLE me_mfd; /**< just for writing the meta pages */
940 /** Failed to update the meta page. Probably an I/O error. */
941 #define MDB_FATAL_ERROR 0x80000000U
942 /** Some fields are initialized. */
943 #define MDB_ENV_ACTIVE 0x20000000U
944 /** me_txkey is set */
945 #define MDB_ENV_TXKEY 0x10000000U
946 uint32_t me_flags; /**< @ref mdb_env */
947 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
948 unsigned int me_maxreaders; /**< size of the reader table */
949 unsigned int me_numreaders; /**< max numreaders set by this env */
950 MDB_dbi me_numdbs; /**< number of DBs opened */
951 MDB_dbi me_maxdbs; /**< size of the DB table */
952 pid_t me_pid; /**< process ID of this env */
953 char *me_path; /**< path to the DB files */
954 char *me_map; /**< the memory map of the data file */
955 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
956 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
957 MDB_txn *me_txn; /**< current write transaction */
958 size_t me_mapsize; /**< size of the data memory map */
959 off_t me_size; /**< current file size */
960 pgno_t me_maxpg; /**< me_mapsize / me_psize */
961 MDB_dbx *me_dbxs; /**< array of static DB info */
962 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
963 pthread_key_t me_txkey; /**< thread-key for readers */
964 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
965 # define me_pglast me_pgstate.mf_pglast
966 # define me_pghead me_pgstate.mf_pghead
967 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
968 /** IDL of pages that became unused in a write txn */
970 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
971 MDB_ID2L me_dirty_list;
972 /** Max number of freelist items that can fit in a single overflow page */
974 /** Max size of a node on a page */
975 unsigned int me_nodemax;
977 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
979 #elif defined(MDB_USE_POSIX_SEM)
980 sem_t *me_rmutex; /* Shared mutexes are not supported */
985 /** Nested transaction */
986 typedef struct MDB_ntxn {
987 MDB_txn mnt_txn; /* the transaction */
988 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
991 /** max number of pages to commit in one writev() call */
992 #define MDB_COMMIT_PAGES 64
993 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
994 #undef MDB_COMMIT_PAGES
995 #define MDB_COMMIT_PAGES IOV_MAX
998 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
999 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1000 static int mdb_page_touch(MDB_cursor *mc);
1002 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1003 static int mdb_page_search_root(MDB_cursor *mc,
1004 MDB_val *key, int modify);
1005 #define MDB_PS_MODIFY 1
1006 #define MDB_PS_ROOTONLY 2
1007 static int mdb_page_search(MDB_cursor *mc,
1008 MDB_val *key, int flags);
1009 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1011 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1012 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1013 pgno_t newpgno, unsigned int nflags);
1015 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1016 static int mdb_env_pick_meta(const MDB_env *env);
1017 static int mdb_env_write_meta(MDB_txn *txn);
1018 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1019 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1021 static void mdb_env_close0(MDB_env *env, int excl);
1023 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1024 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1025 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1026 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1027 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1028 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1029 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1030 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1031 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1033 static int mdb_rebalance(MDB_cursor *mc);
1034 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1036 static void mdb_cursor_pop(MDB_cursor *mc);
1037 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1039 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1040 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1041 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1042 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1043 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1045 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1046 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1048 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1049 static void mdb_xcursor_init0(MDB_cursor *mc);
1050 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1052 static int mdb_drop0(MDB_cursor *mc, int subs);
1053 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1056 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1060 static SECURITY_DESCRIPTOR mdb_null_sd;
1061 static SECURITY_ATTRIBUTES mdb_all_sa;
1062 static int mdb_sec_inited;
1065 /** Return the library version info. */
1067 mdb_version(int *major, int *minor, int *patch)
1069 if (major) *major = MDB_VERSION_MAJOR;
1070 if (minor) *minor = MDB_VERSION_MINOR;
1071 if (patch) *patch = MDB_VERSION_PATCH;
1072 return MDB_VERSION_STRING;
1075 /** Table of descriptions for MDB @ref errors */
1076 static char *const mdb_errstr[] = {
1077 "MDB_KEYEXIST: Key/data pair already exists",
1078 "MDB_NOTFOUND: No matching key/data pair found",
1079 "MDB_PAGE_NOTFOUND: Requested page not found",
1080 "MDB_CORRUPTED: Located page was wrong type",
1081 "MDB_PANIC: Update of meta page failed",
1082 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1083 "MDB_INVALID: File is not an MDB file",
1084 "MDB_MAP_FULL: Environment mapsize limit reached",
1085 "MDB_DBS_FULL: Environment maxdbs limit reached",
1086 "MDB_READERS_FULL: Environment maxreaders limit reached",
1087 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1088 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1089 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1090 "MDB_PAGE_FULL: Internal error - page has no more space",
1091 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1092 "MDB_INCOMPATIBLE: Database flags changed or would change",
1093 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1097 mdb_strerror(int err)
1101 return ("Successful return: 0");
1103 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1104 i = err - MDB_KEYEXIST;
1105 return mdb_errstr[i];
1108 return strerror(err);
1112 /** Display a key in hexadecimal and return the address of the result.
1113 * @param[in] key the key to display
1114 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1115 * @return The key in hexadecimal form.
1118 mdb_dkey(MDB_val *key, char *buf)
1121 unsigned char *c = key->mv_data;
1127 if (key->mv_size > MDB_MAXKEYSIZE)
1128 return "MDB_MAXKEYSIZE";
1129 /* may want to make this a dynamic check: if the key is mostly
1130 * printable characters, print it as-is instead of converting to hex.
1134 for (i=0; i<key->mv_size; i++)
1135 ptr += sprintf(ptr, "%02x", *c++);
1137 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1142 /** Display all the keys in the page. */
1144 mdb_page_list(MDB_page *mp)
1147 unsigned int i, nkeys, nsize;
1151 nkeys = NUMKEYS(mp);
1152 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1153 for (i=0; i<nkeys; i++) {
1154 node = NODEPTR(mp, i);
1155 key.mv_size = node->mn_ksize;
1156 key.mv_data = node->mn_data;
1157 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1158 if (IS_BRANCH(mp)) {
1159 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1162 if (F_ISSET(node->mn_flags, F_BIGDATA))
1163 nsize += sizeof(pgno_t);
1165 nsize += NODEDSZ(node);
1166 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1172 mdb_cursor_chk(MDB_cursor *mc)
1178 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1179 for (i=0; i<mc->mc_top; i++) {
1181 node = NODEPTR(mp, mc->mc_ki[i]);
1182 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1185 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1191 /** Count all the pages in each DB and in the freelist
1192 * and make sure it matches the actual number of pages
1195 static void mdb_audit(MDB_txn *txn)
1199 MDB_ID freecount, count;
1204 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1205 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1206 freecount += *(MDB_ID *)data.mv_data;
1209 for (i = 0; i<txn->mt_numdbs; i++) {
1211 mdb_cursor_init(&mc, txn, i, &mx);
1212 if (txn->mt_dbs[i].md_root == P_INVALID)
1214 count += txn->mt_dbs[i].md_branch_pages +
1215 txn->mt_dbs[i].md_leaf_pages +
1216 txn->mt_dbs[i].md_overflow_pages;
1217 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1218 mdb_page_search(&mc, NULL, 0);
1222 mp = mc.mc_pg[mc.mc_top];
1223 for (j=0; j<NUMKEYS(mp); j++) {
1224 MDB_node *leaf = NODEPTR(mp, j);
1225 if (leaf->mn_flags & F_SUBDATA) {
1227 memcpy(&db, NODEDATA(leaf), sizeof(db));
1228 count += db.md_branch_pages + db.md_leaf_pages +
1229 db.md_overflow_pages;
1233 while (mdb_cursor_sibling(&mc, 1) == 0);
1236 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1237 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1238 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1244 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1246 return txn->mt_dbxs[dbi].md_cmp(a, b);
1250 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1252 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1255 /** Allocate a page.
1256 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1259 mdb_page_malloc(MDB_cursor *mc, unsigned num)
1261 MDB_env *env = mc->mc_txn->mt_env;
1262 MDB_page *ret = env->me_dpages;
1263 size_t sz = env->me_psize;
1266 VGMEMP_ALLOC(env, ret, sz);
1267 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1268 env->me_dpages = ret->mp_next;
1274 if ((ret = malloc(sz)) != NULL) {
1275 VGMEMP_ALLOC(env, ret, sz);
1280 /** Free a single page.
1281 * Saves single pages to a list, for future reuse.
1282 * (This is not used for multi-page overflow pages.)
1285 mdb_page_free(MDB_env *env, MDB_page *mp)
1287 mp->mp_next = env->me_dpages;
1288 VGMEMP_FREE(env, mp);
1289 env->me_dpages = mp;
1292 /* Return all dirty pages to dpage list */
1294 mdb_dlist_free(MDB_txn *txn)
1296 MDB_env *env = txn->mt_env;
1297 MDB_ID2L dl = txn->mt_u.dirty_list;
1298 unsigned i, n = dl[0].mid;
1300 for (i = 1; i <= n; i++) {
1301 MDB_page *dp = dl[i].mptr;
1302 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1303 mdb_page_free(env, dp);
1305 /* large pages just get freed directly */
1306 VGMEMP_FREE(env, dp);
1313 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1315 mdb_find_oldest(MDB_txn *txn)
1318 txnid_t mr, oldest = txn->mt_txnid - 1;
1319 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1320 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1330 /** Allocate pages for writing.
1331 * If there are free pages available from older transactions, they
1332 * will be re-used first. Otherwise a new page will be allocated.
1333 * @param[in] mc cursor A cursor handle identifying the transaction and
1334 * database for which we are allocating.
1335 * @param[in] num the number of pages to allocate.
1336 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1337 * will always be satisfied by a single contiguous chunk of memory.
1338 * @return 0 on success, non-zero on failure.
1341 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1343 MDB_txn *txn = mc->mc_txn;
1344 MDB_env *env = txn->mt_env;
1345 pgno_t pgno = P_INVALID, *mop = env->me_pghead;
1346 unsigned mop_len = mop ? mop[0] : 0;
1349 txnid_t oldest = 0, last;
1354 /* If our dirty list is already full, we can't do anything */
1355 if (txn->mt_dirty_room == 0)
1356 return MDB_TXN_FULL;
1358 /* Pages freed by txn#1 (after allocating but discarding them)
1359 * are used when txn#1 is unreferenced, i.e. txn#3.
1361 if (txn->mt_txnid >= 3) {
1362 if (!mop_len && txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1363 /* See if there's anything in the free DB */
1369 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1370 if (!txn->mt_env->me_pglast) {
1371 mdb_page_search(&m2, NULL, 0);
1372 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1373 kptr = (txnid_t *)NODEKEY(leaf);
1378 last = txn->mt_env->me_pglast + 1;
1380 key.mv_data = &last;
1381 key.mv_size = sizeof(last);
1382 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1385 last = *(txnid_t *)key.mv_data;
1389 oldest = mdb_find_oldest(txn);
1391 if (oldest > last) {
1392 /* It's usable, grab it.
1396 if (!txn->mt_env->me_pglast) {
1397 mdb_node_read(txn, leaf, &data);
1399 idl = (MDB_ID *) data.mv_data;
1402 if (!(env->me_pghead = mop = mdb_midl_alloc(mop_len)))
1404 } else if (mop_len > mop[-1]) {
1405 if ((rc = mdb_midl_grow(&env->me_pghead, mop_len)) != 0)
1407 mop = env->me_pghead;
1409 txn->mt_env->me_pglast = last;
1410 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1415 DPRINTF("IDL read txn %zu root %zu num %zu",
1416 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1417 for (i = idl[0]; i; i--)
1418 DPRINTF("IDL %zu", idl[i]);
1421 /* We might have a zero-length IDL due to freelist growth
1422 * during a prior commit
1432 int retry = 1, readit = 0, n2 = num-1;
1433 unsigned int i, j, k;
1435 /* If current list is too short, must fetch more and coalesce */
1436 if (mop[0] < (unsigned)num)
1439 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1441 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1442 /* If on freelist, don't try to read more. If what we have
1443 * right now isn't enough just use new pages.
1444 * TODO: get all of this working. Many circular dependencies...
1446 if (mc->mc_dbi == FREE_DBI) {
1453 pgno_t *idl, old_id, new_id;
1455 last = txn->mt_env->me_pglast + 1;
1457 /* We haven't hit the readers list yet? */
1459 oldest = mdb_find_oldest(txn);
1462 /* There's nothing we can use on the freelist */
1463 if (oldest - last < 1)
1466 key.mv_data = &last;
1467 key.mv_size = sizeof(last);
1468 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1470 if (rc == MDB_NOTFOUND)
1474 last = *(txnid_t*)key.mv_data;
1477 idl = (MDB_ID *) data.mv_data;
1479 if (mop_len+i > mop[-1]) {
1480 if ((rc = mdb_midl_grow(&env->me_pghead, i)) != 0)
1482 mop = env->me_pghead;
1485 DPRINTF("IDL read txn %zu root %zu num %u",
1486 last, txn->mt_dbs[FREE_DBI].md_root, i);
1488 DPRINTF("IDL %zu", idl[k]);
1490 /* merge in sorted order */
1497 for (; old_id < new_id; old_id = mop[--j])
1502 txn->mt_env->me_pglast = last;
1503 /* Keep trying to read until we have enough */
1504 if (mop[0] < (unsigned)num) {
1509 /* current list has enough pages, but are they contiguous? */
1510 for (i=mop[0]; i>=(unsigned)num; i--) {
1511 if (mop[i-n2] == mop[i] + n2) {
1514 /* move any stragglers down */
1515 for (j=i+num; j<=mop[0]; j++)
1522 /* Stop if we succeeded, or no retries */
1523 if (!retry || pgno != P_INVALID)
1529 /* peel pages off tail, so we only have to truncate the list */
1530 pgno = MDB_IDL_LAST(mop);
1536 if (pgno == P_INVALID) {
1537 /* DB size is maxed out */
1538 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1539 DPUTS("DB size maxed out");
1540 return MDB_MAP_FULL;
1543 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1544 if (pgno == P_INVALID) {
1545 pgno = txn->mt_next_pgno;
1546 txn->mt_next_pgno += num;
1548 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1551 if (!(np = mdb_page_malloc(mc, num)))
1553 if (pgno == P_INVALID) {
1554 np->mp_pgno = txn->mt_next_pgno;
1555 txn->mt_next_pgno += num;
1560 mid.mid = np->mp_pgno;
1562 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1563 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1565 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1567 txn->mt_dirty_room--;
1573 /** Copy the used portions of a non-overflow page.
1574 * @param[in] dst page to copy into
1575 * @param[in] src page to copy from
1576 * @param[in] psize size of a page
1579 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1581 enum { Align = sizeof(pgno_t) };
1582 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
1584 /* If page isn't full, just copy the used portion. Adjust
1585 * alignment so memcpy may copy words instead of bytes.
1587 if ((unused &= -Align) && !IS_LEAF2(src)) {
1589 memcpy(dst, src, (lower + (Align-1)) & -Align);
1590 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
1593 memcpy(dst, src, psize - unused);
1597 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1598 * @param[in] mc cursor pointing to the page to be touched
1599 * @return 0 on success, non-zero on failure.
1602 mdb_page_touch(MDB_cursor *mc)
1604 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
1605 MDB_cursor *m2, *m3;
1610 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1611 if ((rc = mdb_page_alloc(mc, 1, &np)))
1614 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi,mp->mp_pgno,pgno);
1615 assert(mp->mp_pgno != pgno);
1616 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1617 /* Update the parent page, if any, to point to the new page */
1619 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
1620 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
1621 SETPGNO(node, pgno);
1623 mc->mc_db->md_root = pgno;
1625 } else if (mc->mc_txn->mt_parent && !(mp->mp_flags & P_SUBP)) {
1626 MDB_ID2 mid, *dl = mc->mc_txn->mt_u.dirty_list;
1628 /* If txn has a parent, make sure the page is in our
1632 unsigned x = mdb_mid2l_search(dl, pgno);
1633 if (x <= dl[0].mid && dl[x].mid == pgno) {
1636 mc->mc_pg[mc->mc_top] = np;
1640 assert(dl[0].mid < MDB_IDL_UM_MAX);
1642 np = mdb_page_malloc(mc, 1);
1647 mdb_mid2l_insert(dl, &mid);
1652 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1654 np->mp_flags |= P_DIRTY;
1656 /* Adjust cursors pointing to mp */
1657 mc->mc_pg[mc->mc_top] = np;
1659 if (mc->mc_flags & C_SUB) {
1661 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1662 m3 = &m2->mc_xcursor->mx_cursor;
1663 if (m3->mc_snum < mc->mc_snum) continue;
1664 if (m3->mc_pg[mc->mc_top] == mp)
1665 m3->mc_pg[mc->mc_top] = np;
1668 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1669 if (m2->mc_snum < mc->mc_snum) continue;
1670 if (m2->mc_pg[mc->mc_top] == mp) {
1671 m2->mc_pg[mc->mc_top] = np;
1672 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
1673 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
1675 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
1676 if (!(leaf->mn_flags & F_SUBDATA))
1677 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
1686 mdb_env_sync(MDB_env *env, int force)
1689 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1690 if (env->me_flags & MDB_WRITEMAP) {
1691 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1692 ? MS_ASYNC : MS_SYNC;
1693 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1696 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1700 if (MDB_FDATASYNC(env->me_fd))
1707 /** Make shadow copies of all of parent txn's cursors */
1709 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1711 MDB_cursor *mc, *m2;
1712 unsigned int i, j, size;
1714 for (i=0;i<src->mt_numdbs; i++) {
1715 if (src->mt_cursors[i]) {
1716 size = sizeof(MDB_cursor);
1717 if (src->mt_cursors[i]->mc_xcursor)
1718 size += sizeof(MDB_xcursor);
1719 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1726 mc->mc_db = &dst->mt_dbs[i];
1727 mc->mc_dbx = m2->mc_dbx;
1728 mc->mc_dbflag = &dst->mt_dbflags[i];
1729 mc->mc_snum = m2->mc_snum;
1730 mc->mc_top = m2->mc_top;
1731 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1732 for (j=0; j<mc->mc_snum; j++) {
1733 mc->mc_pg[j] = m2->mc_pg[j];
1734 mc->mc_ki[j] = m2->mc_ki[j];
1736 if (m2->mc_xcursor) {
1737 MDB_xcursor *mx, *mx2;
1738 mx = (MDB_xcursor *)(mc+1);
1739 mc->mc_xcursor = mx;
1740 mx2 = m2->mc_xcursor;
1741 mx->mx_db = mx2->mx_db;
1742 mx->mx_dbx = mx2->mx_dbx;
1743 mx->mx_dbflag = mx2->mx_dbflag;
1744 mx->mx_cursor.mc_txn = dst;
1745 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1746 mx->mx_cursor.mc_db = &mx->mx_db;
1747 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1748 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1749 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1750 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1751 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1752 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1753 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1754 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1757 mc->mc_xcursor = NULL;
1759 mc->mc_next = dst->mt_cursors[i];
1760 dst->mt_cursors[i] = mc;
1767 /** Close this write txn's cursors, after optionally merging its shadow
1768 * cursors back into parent's.
1769 * @param[in] txn the transaction handle.
1770 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1771 * @return 0 on success, non-zero on failure.
1774 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1776 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1779 for (i = txn->mt_numdbs; --i >= 0; ) {
1780 for (mc = cursors[i]; mc; mc = next) {
1782 if (mc->mc_flags & merge) {
1783 MDB_cursor *m2 = mc->mc_orig;
1784 m2->mc_snum = mc->mc_snum;
1785 m2->mc_top = mc->mc_top;
1786 for (j = mc->mc_snum; --j >= 0; ) {
1787 m2->mc_pg[j] = mc->mc_pg[j];
1788 m2->mc_ki[j] = mc->mc_ki[j];
1791 if (mc->mc_flags & C_ALLOCD)
1799 mdb_txn_reset0(MDB_txn *txn);
1801 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1802 * @param[in] txn the transaction handle to initialize
1803 * @return 0 on success, non-zero on failure.
1806 mdb_txn_renew0(MDB_txn *txn)
1808 MDB_env *env = txn->mt_env;
1811 int rc, new_notls = 0;
1814 txn->mt_numdbs = env->me_numdbs;
1815 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1817 if (txn->mt_flags & MDB_TXN_RDONLY) {
1818 if (!env->me_txns) {
1819 i = mdb_env_pick_meta(env);
1820 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1821 txn->mt_u.reader = NULL;
1823 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1824 pthread_getspecific(env->me_txkey);
1826 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1827 return MDB_BAD_RSLOT;
1829 pid_t pid = env->me_pid;
1830 pthread_t tid = pthread_self();
1833 for (i=0; i<env->me_txns->mti_numreaders; i++)
1834 if (env->me_txns->mti_readers[i].mr_pid == 0)
1836 if (i == env->me_maxreaders) {
1837 UNLOCK_MUTEX_R(env);
1838 return MDB_READERS_FULL;
1840 env->me_txns->mti_readers[i].mr_pid = pid;
1841 env->me_txns->mti_readers[i].mr_tid = tid;
1842 if (i >= env->me_txns->mti_numreaders)
1843 env->me_txns->mti_numreaders = i+1;
1844 /* Save numreaders for un-mutexed mdb_env_close() */
1845 env->me_numreaders = env->me_txns->mti_numreaders;
1846 UNLOCK_MUTEX_R(env);
1847 r = &env->me_txns->mti_readers[i];
1848 new_notls = (env->me_flags & MDB_NOTLS);
1849 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1854 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1855 txn->mt_u.reader = r;
1857 txn->mt_toggle = txn->mt_txnid & 1;
1858 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1862 txn->mt_txnid = env->me_txns->mti_txnid;
1863 txn->mt_toggle = txn->mt_txnid & 1;
1864 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1867 if (txn->mt_txnid == mdb_debug_start)
1870 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1871 txn->mt_u.dirty_list = env->me_dirty_list;
1872 txn->mt_u.dirty_list[0].mid = 0;
1873 txn->mt_free_pgs = env->me_free_pgs;
1874 txn->mt_free_pgs[0] = 0;
1878 /* Copy the DB info and flags */
1879 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1880 for (i=2; i<txn->mt_numdbs; i++) {
1881 x = env->me_dbflags[i];
1882 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1883 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1885 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1887 if (env->me_maxpg < txn->mt_next_pgno) {
1888 mdb_txn_reset0(txn);
1890 txn->mt_u.reader->mr_pid = 0;
1891 txn->mt_u.reader = NULL;
1893 return MDB_MAP_RESIZED;
1900 mdb_txn_renew(MDB_txn *txn)
1904 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1907 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1908 DPUTS("environment had fatal error, must shutdown!");
1912 rc = mdb_txn_renew0(txn);
1913 if (rc == MDB_SUCCESS) {
1914 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1915 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1916 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1922 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1926 int rc, size, tsize = sizeof(MDB_txn);
1928 if (env->me_flags & MDB_FATAL_ERROR) {
1929 DPUTS("environment had fatal error, must shutdown!");
1932 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1935 /* Nested transactions: Max 1 child, write txns only, no writemap */
1936 if (parent->mt_child ||
1937 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1938 (env->me_flags & MDB_WRITEMAP))
1942 tsize = sizeof(MDB_ntxn);
1944 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1945 if (!(flags & MDB_RDONLY))
1946 size += env->me_maxdbs * sizeof(MDB_cursor *);
1948 if ((txn = calloc(1, size)) == NULL) {
1949 DPRINTF("calloc: %s", strerror(ErrCode()));
1952 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1953 if (flags & MDB_RDONLY) {
1954 txn->mt_flags |= MDB_TXN_RDONLY;
1955 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1957 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1958 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1964 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1965 if (!txn->mt_u.dirty_list ||
1966 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1968 free(txn->mt_u.dirty_list);
1972 txn->mt_txnid = parent->mt_txnid;
1973 txn->mt_toggle = parent->mt_toggle;
1974 txn->mt_dirty_room = parent->mt_dirty_room;
1975 txn->mt_u.dirty_list[0].mid = 0;
1976 txn->mt_next_pgno = parent->mt_next_pgno;
1977 parent->mt_child = txn;
1978 txn->mt_parent = parent;
1979 txn->mt_numdbs = parent->mt_numdbs;
1980 txn->mt_flags = parent->mt_flags;
1981 txn->mt_dbxs = parent->mt_dbxs;
1982 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1983 /* Copy parent's mt_dbflags, but clear DB_NEW */
1984 for (i=0; i<txn->mt_numdbs; i++)
1985 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1987 ntxn = (MDB_ntxn *)txn;
1988 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1989 if (env->me_pghead) {
1990 size = MDB_IDL_SIZEOF(env->me_pghead);
1991 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
1993 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1998 rc = mdb_cursor_shadow(parent, txn);
2000 mdb_txn_reset0(txn);
2002 rc = mdb_txn_renew0(txn);
2008 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
2009 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2010 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2016 /** Export or close DBI handles opened in this txn. */
2018 mdb_dbis_update(MDB_txn *txn, int keep)
2021 MDB_dbi n = txn->mt_numdbs;
2022 MDB_env *env = txn->mt_env;
2023 unsigned char *tdbflags = txn->mt_dbflags;
2025 for (i = n; --i >= 2;) {
2026 if (tdbflags[i] & DB_NEW) {
2028 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2030 char *ptr = env->me_dbxs[i].md_name.mv_data;
2031 env->me_dbxs[i].md_name.mv_data = NULL;
2032 env->me_dbxs[i].md_name.mv_size = 0;
2033 env->me_dbflags[i] = 0;
2038 if (keep && env->me_numdbs < n)
2042 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2043 * May be called twice for readonly txns: First reset it, then abort.
2044 * @param[in] txn the transaction handle to reset
2047 mdb_txn_reset0(MDB_txn *txn)
2049 MDB_env *env = txn->mt_env;
2051 /* Close any DBI handles opened in this txn */
2052 mdb_dbis_update(txn, 0);
2054 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2055 if (txn->mt_u.reader) {
2056 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2057 if (!(env->me_flags & MDB_NOTLS))
2058 txn->mt_u.reader = NULL; /* txn does not own reader */
2060 txn->mt_numdbs = 0; /* close nothing if called again */
2061 txn->mt_dbxs = NULL; /* mark txn as reset */
2063 mdb_cursors_close(txn, 0);
2065 if (!(env->me_flags & MDB_WRITEMAP)) {
2066 mdb_dlist_free(txn);
2068 mdb_midl_free(env->me_pghead);
2070 if (txn->mt_parent) {
2071 txn->mt_parent->mt_child = NULL;
2072 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2073 mdb_midl_free(txn->mt_free_pgs);
2074 free(txn->mt_u.dirty_list);
2077 if (mdb_midl_shrink(&txn->mt_free_pgs))
2078 env->me_free_pgs = txn->mt_free_pgs;
2081 txn->mt_env->me_pghead = NULL;
2082 txn->mt_env->me_pglast = 0;
2085 /* The writer mutex was locked in mdb_txn_begin. */
2086 UNLOCK_MUTEX_W(env);
2091 mdb_txn_reset(MDB_txn *txn)
2096 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2097 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2098 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2100 /* This call is only valid for read-only txns */
2101 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2104 mdb_txn_reset0(txn);
2108 mdb_txn_abort(MDB_txn *txn)
2113 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2114 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2115 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2118 mdb_txn_abort(txn->mt_child);
2120 mdb_txn_reset0(txn);
2121 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2122 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2123 txn->mt_u.reader->mr_pid = 0;
2128 /** Save the freelist as of this transaction to the freeDB.
2129 * This changes the freelist. Keep trying until it stabilizes.
2132 mdb_freelist_save(MDB_txn *txn)
2134 /* env->me_pghead[] can grow and shrink during this call.
2135 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2136 * Page numbers cannot disappear from txn->mt_free_pgs[].
2139 MDB_env *env = txn->mt_env;
2140 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2141 txnid_t pglast = 0, head_id = 0;
2142 pgno_t freecnt = 0, *free_pgs, *mop;
2143 ssize_t head_room = 0, total_room = 0, mop_len;
2145 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2147 if (env->me_pghead) {
2148 /* Make sure first page of freeDB is touched and on freelist */
2149 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2150 if (rc && rc != MDB_NOTFOUND)
2155 /* Come back here after each Put() in case freelist changed */
2158 /* If using records from freeDB which we have not yet
2159 * deleted, delete them and any we reserved for me_pghead.
2161 while (pglast < env->me_pglast) {
2162 rc = mdb_cursor_first(&mc, &key, NULL);
2165 pglast = head_id = *(txnid_t *)key.mv_data;
2166 total_room = head_room = 0;
2167 assert(pglast <= env->me_pglast);
2168 rc = mdb_cursor_del(&mc, 0);
2173 /* Save the IDL of pages freed by this txn, to a single record */
2174 if (freecnt < txn->mt_free_pgs[0]) {
2176 /* Make sure last page of freeDB is touched and on freelist */
2177 key.mv_size = MDB_MAXKEYSIZE+1;
2179 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2180 if (rc && rc != MDB_NOTFOUND)
2183 free_pgs = txn->mt_free_pgs;
2184 /* Write to last page of freeDB */
2185 key.mv_size = sizeof(txn->mt_txnid);
2186 key.mv_data = &txn->mt_txnid;
2188 freecnt = free_pgs[0];
2189 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2190 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2193 /* Retry if mt_free_pgs[] grew during the Put() */
2194 free_pgs = txn->mt_free_pgs;
2195 } while (freecnt < free_pgs[0]);
2196 mdb_midl_sort(free_pgs);
2197 memcpy(data.mv_data, free_pgs, data.mv_size);
2200 unsigned int i = free_pgs[0];
2201 DPRINTF("IDL write txn %zu root %zu num %u",
2202 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i);
2204 DPRINTF("IDL %zu", free_pgs[i]);
2210 mop = env->me_pghead;
2211 mop_len = mop ? mop[0] : 0;
2213 /* Reserve records for me_pghead[]. Split it if multi-page,
2214 * to avoid searching freeDB for a page range. Use keys in
2215 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2217 if (total_room >= mop_len) {
2218 if (total_room == mop_len || --more < 0)
2220 } else if (head_room >= maxfree_1pg && head_id > 1) {
2221 /* Keep current record (overflow page), add a new one */
2225 /* (Re)write {key = head_id, IDL length = head_room} */
2226 total_room -= head_room;
2227 head_room = mop_len - total_room;
2228 if (head_room > maxfree_1pg && head_id > 1) {
2229 /* Overflow multi-page for part of me_pghead */
2230 head_room /= head_id; /* amortize page sizes */
2231 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2232 } else if (head_room < 0) {
2233 /* Rare case, not bothering to delete this record */
2236 key.mv_size = sizeof(head_id);
2237 key.mv_data = &head_id;
2238 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2239 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2242 *(MDB_ID *)data.mv_data = 0; /* IDL is initially empty */
2243 total_room += head_room;
2246 /* Fill in the reserved, touched me_pghead records. Avoid write ops
2247 * so they cannot rearrange anything, just read the destinations.
2254 rc = mdb_cursor_first(&mc, &key, &data);
2255 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2256 MDB_IDL dest = data.mv_data;
2257 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2259 assert(len >= 0 && *(txnid_t*)key.mv_data <= env->me_pglast);
2263 memcpy(dest, mop -= len, len * sizeof(MDB_ID));
2264 if (! (mop_len -= len))
2272 mdb_txn_commit(MDB_txn *txn)
2282 assert(txn != NULL);
2283 assert(txn->mt_env != NULL);
2285 if (txn->mt_child) {
2286 rc = mdb_txn_commit(txn->mt_child);
2287 txn->mt_child = NULL;
2296 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2297 mdb_dbis_update(txn, 1);
2298 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2303 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2304 DPUTS("error flag is set, can't commit");
2306 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2311 if (txn->mt_parent) {
2312 MDB_txn *parent = txn->mt_parent;
2316 /* Append our free list to parent's */
2317 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2321 mdb_midl_free(txn->mt_free_pgs);
2323 parent->mt_next_pgno = txn->mt_next_pgno;
2324 parent->mt_flags = txn->mt_flags;
2326 /* Merge our cursors into parent's and close them */
2327 mdb_cursors_close(txn, C_SHADOW);
2329 /* Update parent's DB table. */
2330 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2331 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2332 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2333 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2334 for (i=2; i<txn->mt_numdbs; i++) {
2335 /* preserve parent's DB_NEW status */
2336 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2337 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2340 dst = txn->mt_parent->mt_u.dirty_list;
2341 src = txn->mt_u.dirty_list;
2342 /* Find len = length of merging our dirty list with parent's */
2344 dst[0].mid = 0; /* simplify loops */
2345 if (parent->mt_parent) {
2346 len = x + src[0].mid;
2347 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2348 for (i = x; y && i; y--) {
2349 pgno_t yp = src[y].mid;
2350 while (yp < dst[i].mid)
2352 if (yp == dst[i].mid) {
2357 } else { /* Simplify the above for single-ancestor case */
2358 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2360 /* Merge our dirty list with parent's */
2362 for (i = len; y; dst[i--] = src[y--]) {
2363 pgno_t yp = src[y].mid;
2364 while (yp < dst[x].mid)
2365 dst[i--] = dst[x--];
2366 if (yp == dst[x].mid)
2367 free(dst[x--].mptr);
2371 free(txn->mt_u.dirty_list);
2372 parent->mt_dirty_room = txn->mt_dirty_room;
2374 txn->mt_parent->mt_child = NULL;
2375 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
2380 if (txn != env->me_txn) {
2381 DPUTS("attempt to commit unknown transaction");
2386 mdb_cursors_close(txn, 0);
2388 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2391 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2392 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2394 /* Update DB root pointers */
2395 if (txn->mt_numdbs > 2) {
2399 data.mv_size = sizeof(MDB_db);
2401 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2402 for (i = 2; i < txn->mt_numdbs; i++) {
2403 if (txn->mt_dbflags[i] & DB_DIRTY) {
2404 data.mv_data = &txn->mt_dbs[i];
2405 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2412 rc = mdb_freelist_save(txn);
2416 mdb_midl_free(env->me_pghead);
2417 env->me_pghead = NULL;
2418 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2419 if (mdb_midl_shrink(&txn->mt_free_pgs))
2420 env->me_free_pgs = txn->mt_free_pgs;
2427 if (env->me_flags & MDB_WRITEMAP) {
2428 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2429 dp = txn->mt_u.dirty_list[i].mptr;
2430 /* clear dirty flag */
2431 dp->mp_flags &= ~P_DIRTY;
2433 txn->mt_u.dirty_list[0].mid = 0;
2437 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2443 /* Windows actually supports scatter/gather I/O, but only on
2444 * unbuffered file handles. Since we're relying on the OS page
2445 * cache for all our data, that's self-defeating. So we just
2446 * write pages one at a time. We use the ov structure to set
2447 * the write offset, to at least save the overhead of a Seek
2451 memset(&ov, 0, sizeof(ov));
2452 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2454 dp = txn->mt_u.dirty_list[i].mptr;
2455 DPRINTF("committing page %zu", dp->mp_pgno);
2456 size = dp->mp_pgno * env->me_psize;
2457 ov.Offset = size & 0xffffffff;
2458 ov.OffsetHigh = size >> 16;
2459 ov.OffsetHigh >>= 16;
2460 /* clear dirty flag */
2461 dp->mp_flags &= ~P_DIRTY;
2462 wsize = env->me_psize;
2463 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2464 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2467 DPRINTF("WriteFile: %d", n);
2474 struct iovec iov[MDB_COMMIT_PAGES];
2478 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2479 dp = txn->mt_u.dirty_list[i].mptr;
2480 if (dp->mp_pgno != next) {
2482 rc = writev(env->me_fd, iov, n);
2486 DPUTS("short write, filesystem full?");
2488 DPRINTF("writev: %s", strerror(n));
2495 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2498 DPRINTF("committing page %zu", dp->mp_pgno);
2499 iov[n].iov_len = env->me_psize;
2500 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2501 iov[n].iov_base = (char *)dp;
2502 size += iov[n].iov_len;
2503 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2504 /* clear dirty flag */
2505 dp->mp_flags &= ~P_DIRTY;
2506 if (++n >= MDB_COMMIT_PAGES) {
2516 rc = writev(env->me_fd, iov, n);
2520 DPUTS("short write, filesystem full?");
2522 DPRINTF("writev: %s", strerror(n));
2529 mdb_dlist_free(txn);
2532 if ((n = mdb_env_sync(env, 0)) != 0 ||
2533 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2541 mdb_dbis_update(txn, 1);
2543 UNLOCK_MUTEX_W(env);
2553 /** Read the environment parameters of a DB environment before
2554 * mapping it into memory.
2555 * @param[in] env the environment handle
2556 * @param[out] meta address of where to store the meta information
2557 * @return 0 on success, non-zero on failure.
2560 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2567 /* We don't know the page size yet, so use a minimum value.
2568 * Read both meta pages so we can use the latest one.
2571 for (i=0; i<2; i++) {
2573 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2575 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2580 else if (rc != MDB_PAGESIZE) {
2584 DPRINTF("read: %s", strerror(err));
2588 p = (MDB_page *)&pbuf;
2590 if (!F_ISSET(p->mp_flags, P_META)) {
2591 DPRINTF("page %zu not a meta page", p->mp_pgno);
2596 if (m->mm_magic != MDB_MAGIC) {
2597 DPUTS("meta has invalid magic");
2601 if (m->mm_version != MDB_VERSION) {
2602 DPRINTF("database is version %u, expected version %u",
2603 m->mm_version, MDB_VERSION);
2604 return MDB_VERSION_MISMATCH;
2608 if (m->mm_txnid > meta->mm_txnid)
2609 memcpy(meta, m, sizeof(*m));
2611 memcpy(meta, m, sizeof(*m));
2613 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2615 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2623 /** Write the environment parameters of a freshly created DB environment.
2624 * @param[in] env the environment handle
2625 * @param[out] meta address of where to store the meta information
2626 * @return 0 on success, non-zero on failure.
2629 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2636 DPUTS("writing new meta page");
2638 GET_PAGESIZE(psize);
2640 meta->mm_magic = MDB_MAGIC;
2641 meta->mm_version = MDB_VERSION;
2642 meta->mm_mapsize = env->me_mapsize;
2643 meta->mm_psize = psize;
2644 meta->mm_last_pg = 1;
2645 meta->mm_flags = env->me_flags & 0xffff;
2646 meta->mm_flags |= MDB_INTEGERKEY;
2647 meta->mm_dbs[0].md_root = P_INVALID;
2648 meta->mm_dbs[1].md_root = P_INVALID;
2650 p = calloc(2, psize);
2652 p->mp_flags = P_META;
2655 memcpy(m, meta, sizeof(*meta));
2657 q = (MDB_page *)((char *)p + psize);
2660 q->mp_flags = P_META;
2663 memcpy(m, meta, sizeof(*meta));
2668 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2669 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2670 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2673 lseek(env->me_fd, 0, SEEK_SET);
2674 rc = write(env->me_fd, p, psize * 2);
2675 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2681 /** Update the environment info to commit a transaction.
2682 * @param[in] txn the transaction that's being committed
2683 * @return 0 on success, non-zero on failure.
2686 mdb_env_write_meta(MDB_txn *txn)
2689 MDB_meta meta, metab, *mp;
2691 int rc, len, toggle;
2698 assert(txn != NULL);
2699 assert(txn->mt_env != NULL);
2701 toggle = !txn->mt_toggle;
2702 DPRINTF("writing meta page %d for root page %zu",
2703 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2706 mp = env->me_metas[toggle];
2708 if (env->me_flags & MDB_WRITEMAP) {
2709 /* Persist any increases of mapsize config */
2710 if (env->me_mapsize > mp->mm_mapsize)
2711 mp->mm_mapsize = env->me_mapsize;
2712 mp->mm_dbs[0] = txn->mt_dbs[0];
2713 mp->mm_dbs[1] = txn->mt_dbs[1];
2714 mp->mm_last_pg = txn->mt_next_pgno - 1;
2715 mp->mm_txnid = txn->mt_txnid;
2716 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2717 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2720 ptr += env->me_psize;
2721 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2728 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2729 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2731 ptr = (char *)&meta;
2732 if (env->me_mapsize > mp->mm_mapsize) {
2733 /* Persist any increases of mapsize config */
2734 meta.mm_mapsize = env->me_mapsize;
2735 off = offsetof(MDB_meta, mm_mapsize);
2737 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2739 len = sizeof(MDB_meta) - off;
2742 meta.mm_dbs[0] = txn->mt_dbs[0];
2743 meta.mm_dbs[1] = txn->mt_dbs[1];
2744 meta.mm_last_pg = txn->mt_next_pgno - 1;
2745 meta.mm_txnid = txn->mt_txnid;
2748 off += env->me_psize;
2751 /* Write to the SYNC fd */
2752 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2753 env->me_fd : env->me_mfd;
2756 memset(&ov, 0, sizeof(ov));
2758 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2761 rc = pwrite(mfd, ptr, len, off);
2766 DPUTS("write failed, disk error?");
2767 /* On a failure, the pagecache still contains the new data.
2768 * Write some old data back, to prevent it from being used.
2769 * Use the non-SYNC fd; we know it will fail anyway.
2771 meta.mm_last_pg = metab.mm_last_pg;
2772 meta.mm_txnid = metab.mm_txnid;
2774 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2776 r2 = pwrite(env->me_fd, ptr, len, off);
2779 env->me_flags |= MDB_FATAL_ERROR;
2783 /* Memory ordering issues are irrelevant; since the entire writer
2784 * is wrapped by wmutex, all of these changes will become visible
2785 * after the wmutex is unlocked. Since the DB is multi-version,
2786 * readers will get consistent data regardless of how fresh or
2787 * how stale their view of these values is.
2789 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2794 /** Check both meta pages to see which one is newer.
2795 * @param[in] env the environment handle
2796 * @return meta toggle (0 or 1).
2799 mdb_env_pick_meta(const MDB_env *env)
2801 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2805 mdb_env_create(MDB_env **env)
2809 e = calloc(1, sizeof(MDB_env));
2813 e->me_maxreaders = DEFAULT_READERS;
2814 e->me_maxdbs = e->me_numdbs = 2;
2815 e->me_fd = INVALID_HANDLE_VALUE;
2816 e->me_lfd = INVALID_HANDLE_VALUE;
2817 e->me_mfd = INVALID_HANDLE_VALUE;
2818 #ifdef MDB_USE_POSIX_SEM
2819 e->me_rmutex = SEM_FAILED;
2820 e->me_wmutex = SEM_FAILED;
2822 e->me_pid = getpid();
2823 VGMEMP_CREATE(e,0,0);
2829 mdb_env_set_mapsize(MDB_env *env, size_t size)
2833 env->me_mapsize = size;
2835 env->me_maxpg = env->me_mapsize / env->me_psize;
2840 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2844 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2849 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2851 if (env->me_map || readers < 1)
2853 env->me_maxreaders = readers;
2858 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2860 if (!env || !readers)
2862 *readers = env->me_maxreaders;
2866 /** Further setup required for opening an MDB environment
2869 mdb_env_open2(MDB_env *env)
2871 unsigned int flags = env->me_flags;
2872 int i, newenv = 0, prot;
2876 memset(&meta, 0, sizeof(meta));
2878 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2881 DPUTS("new mdbenv");
2885 /* Was a mapsize configured? */
2886 if (!env->me_mapsize) {
2887 /* If this is a new environment, take the default,
2888 * else use the size recorded in the existing env.
2890 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2891 } else if (env->me_mapsize < meta.mm_mapsize) {
2892 /* If the configured size is smaller, make sure it's
2893 * still big enough. Silently round up to minimum if not.
2895 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2896 if (env->me_mapsize < minsize)
2897 env->me_mapsize = minsize;
2903 LONG sizelo, sizehi;
2904 sizelo = env->me_mapsize & 0xffffffff;
2905 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2907 /* Windows won't create mappings for zero length files.
2908 * Just allocate the maxsize right now.
2911 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2912 if (!SetEndOfFile(env->me_fd))
2914 SetFilePointer(env->me_fd, 0, NULL, 0);
2916 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2917 PAGE_READWRITE : PAGE_READONLY,
2918 sizehi, sizelo, NULL);
2921 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2922 FILE_MAP_WRITE : FILE_MAP_READ,
2923 0, 0, env->me_mapsize, meta.mm_address);
2931 if (flags & MDB_WRITEMAP) {
2933 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2936 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2938 if (env->me_map == MAP_FAILED) {
2942 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2944 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2946 #ifdef POSIX_MADV_RANDOM
2947 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2948 #endif /* POSIX_MADV_RANDOM */
2949 #endif /* MADV_RANDOM */
2953 if (flags & MDB_FIXEDMAP)
2954 meta.mm_address = env->me_map;
2955 i = mdb_env_init_meta(env, &meta);
2956 if (i != MDB_SUCCESS) {
2959 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2960 /* Can happen because the address argument to mmap() is just a
2961 * hint. mmap() can pick another, e.g. if the range is in use.
2962 * The MAP_FIXED flag would prevent that, but then mmap could
2963 * instead unmap existing pages to make room for the new map.
2965 return EBUSY; /* TODO: Make a new MDB_* error code? */
2967 env->me_psize = meta.mm_psize;
2968 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2969 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2971 env->me_maxpg = env->me_mapsize / env->me_psize;
2973 p = (MDB_page *)env->me_map;
2974 env->me_metas[0] = METADATA(p);
2975 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2979 int toggle = mdb_env_pick_meta(env);
2980 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2982 DPRINTF("opened database version %u, pagesize %u",
2983 env->me_metas[0]->mm_version, env->me_psize);
2984 DPRINTF("using meta page %d", toggle);
2985 DPRINTF("depth: %u", db->md_depth);
2986 DPRINTF("entries: %zu", db->md_entries);
2987 DPRINTF("branch pages: %zu", db->md_branch_pages);
2988 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2989 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2990 DPRINTF("root: %zu", db->md_root);
2998 /** Release a reader thread's slot in the reader lock table.
2999 * This function is called automatically when a thread exits.
3000 * @param[in] ptr This points to the slot in the reader lock table.
3003 mdb_env_reader_dest(void *ptr)
3005 MDB_reader *reader = ptr;
3011 /** Junk for arranging thread-specific callbacks on Windows. This is
3012 * necessarily platform and compiler-specific. Windows supports up
3013 * to 1088 keys. Let's assume nobody opens more than 64 environments
3014 * in a single process, for now. They can override this if needed.
3016 #ifndef MAX_TLS_KEYS
3017 #define MAX_TLS_KEYS 64
3019 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3020 static int mdb_tls_nkeys;
3022 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3026 case DLL_PROCESS_ATTACH: break;
3027 case DLL_THREAD_ATTACH: break;
3028 case DLL_THREAD_DETACH:
3029 for (i=0; i<mdb_tls_nkeys; i++) {
3030 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3031 mdb_env_reader_dest(r);
3034 case DLL_PROCESS_DETACH: break;
3039 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3041 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3045 /* Force some symbol references.
3046 * _tls_used forces the linker to create the TLS directory if not already done
3047 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3049 #pragma comment(linker, "/INCLUDE:_tls_used")
3050 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3051 #pragma const_seg(".CRT$XLB")
3052 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3053 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3056 #pragma comment(linker, "/INCLUDE:__tls_used")
3057 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3058 #pragma data_seg(".CRT$XLB")
3059 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3061 #endif /* WIN 32/64 */
3062 #endif /* !__GNUC__ */
3065 /** Downgrade the exclusive lock on the region back to shared */
3067 mdb_env_share_locks(MDB_env *env, int *excl)
3069 int rc = 0, toggle = mdb_env_pick_meta(env);
3071 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3076 /* First acquire a shared lock. The Unlock will
3077 * then release the existing exclusive lock.
3079 memset(&ov, 0, sizeof(ov));
3080 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3083 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3089 struct flock lock_info;
3090 /* The shared lock replaces the existing lock */
3091 memset((void *)&lock_info, 0, sizeof(lock_info));
3092 lock_info.l_type = F_RDLCK;
3093 lock_info.l_whence = SEEK_SET;
3094 lock_info.l_start = 0;
3095 lock_info.l_len = 1;
3096 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3097 (rc = ErrCode()) == EINTR) ;
3098 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3105 /** Try to get exlusive lock, otherwise shared.
3106 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3109 mdb_env_excl_lock(MDB_env *env, int *excl)
3113 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3117 memset(&ov, 0, sizeof(ov));
3118 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3125 struct flock lock_info;
3126 memset((void *)&lock_info, 0, sizeof(lock_info));
3127 lock_info.l_type = F_WRLCK;
3128 lock_info.l_whence = SEEK_SET;
3129 lock_info.l_start = 0;
3130 lock_info.l_len = 1;
3131 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3132 (rc = ErrCode()) == EINTR) ;
3136 # ifdef MDB_USE_POSIX_SEM
3137 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3140 lock_info.l_type = F_RDLCK;
3141 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3142 (rc = ErrCode()) == EINTR) ;
3150 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3152 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3154 * @(#) $Revision: 5.1 $
3155 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3156 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3158 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3162 * Please do not copyright this code. This code is in the public domain.
3164 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3165 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3166 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3167 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3168 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3169 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3170 * PERFORMANCE OF THIS SOFTWARE.
3173 * chongo <Landon Curt Noll> /\oo/\
3174 * http://www.isthe.com/chongo/
3176 * Share and Enjoy! :-)
3179 typedef unsigned long long mdb_hash_t;
3180 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3182 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3183 * @param[in] str string to hash
3184 * @param[in] hval initial value for hash
3185 * @return 64 bit hash
3187 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3188 * hval arg on the first call.
3191 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3193 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3194 unsigned char *end = s + val->mv_size;
3196 * FNV-1a hash each octet of the string
3199 /* xor the bottom with the current octet */
3200 hval ^= (mdb_hash_t)*s++;
3202 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3203 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3204 (hval << 7) + (hval << 8) + (hval << 40);
3206 /* return our new hash value */
3210 /** Hash the string and output the hash in hex.
3211 * @param[in] str string to hash
3212 * @param[out] hexbuf an array of 17 chars to hold the hash
3215 mdb_hash_hex(MDB_val *val, char *hexbuf)
3218 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3219 for (i=0; i<8; i++) {
3220 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3226 /** Open and/or initialize the lock region for the environment.
3227 * @param[in] env The MDB environment.
3228 * @param[in] lpath The pathname of the file used for the lock region.
3229 * @param[in] mode The Unix permissions for the file, if we create it.
3230 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3231 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3232 * @return 0 on success, non-zero on failure.
3235 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3238 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3240 # define MDB_ERRCODE_ROFS EROFS
3241 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3242 # define MDB_CLOEXEC O_CLOEXEC
3245 # define MDB_CLOEXEC 0
3252 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3253 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3254 FILE_ATTRIBUTE_NORMAL, NULL);
3256 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3258 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3260 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3265 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3266 /* Lose record locks when exec*() */
3267 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3268 fcntl(env->me_lfd, F_SETFD, fdflags);
3271 if (!(env->me_flags & MDB_NOTLS)) {
3272 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3275 env->me_flags |= MDB_ENV_TXKEY;
3277 /* Windows TLS callbacks need help finding their TLS info. */
3278 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3282 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3286 /* Try to get exclusive lock. If we succeed, then
3287 * nobody is using the lock region and we should initialize it.
3289 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3292 size = GetFileSize(env->me_lfd, NULL);
3294 size = lseek(env->me_lfd, 0, SEEK_END);
3296 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3297 if (size < rsize && *excl > 0) {
3299 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3300 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3302 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3306 size = rsize - sizeof(MDB_txninfo);
3307 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3312 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3314 if (!mh) goto fail_errno;
3315 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3317 if (!env->me_txns) goto fail_errno;
3319 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3321 if (m == MAP_FAILED) goto fail_errno;
3327 BY_HANDLE_FILE_INFORMATION stbuf;
3336 if (!mdb_sec_inited) {
3337 InitializeSecurityDescriptor(&mdb_null_sd,
3338 SECURITY_DESCRIPTOR_REVISION);
3339 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3340 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3341 mdb_all_sa.bInheritHandle = FALSE;
3342 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3345 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3346 idbuf.volume = stbuf.dwVolumeSerialNumber;
3347 idbuf.nhigh = stbuf.nFileIndexHigh;
3348 idbuf.nlow = stbuf.nFileIndexLow;
3349 val.mv_data = &idbuf;
3350 val.mv_size = sizeof(idbuf);
3351 mdb_hash_hex(&val, hexbuf);
3352 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3353 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3354 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3355 if (!env->me_rmutex) goto fail_errno;
3356 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3357 if (!env->me_wmutex) goto fail_errno;
3358 #elif defined(MDB_USE_POSIX_SEM)
3367 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3368 idbuf.dev = stbuf.st_dev;
3369 idbuf.ino = stbuf.st_ino;
3370 val.mv_data = &idbuf;
3371 val.mv_size = sizeof(idbuf);
3372 mdb_hash_hex(&val, hexbuf);
3373 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3374 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3375 /* Clean up after a previous run, if needed: Try to
3376 * remove both semaphores before doing anything else.
3378 sem_unlink(env->me_txns->mti_rmname);
3379 sem_unlink(env->me_txns->mti_wmname);
3380 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3381 O_CREAT|O_EXCL, mode, 1);
3382 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3383 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3384 O_CREAT|O_EXCL, mode, 1);
3385 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3386 #else /* MDB_USE_POSIX_SEM */
3387 pthread_mutexattr_t mattr;
3389 if ((rc = pthread_mutexattr_init(&mattr))
3390 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3391 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3392 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3394 pthread_mutexattr_destroy(&mattr);
3395 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3397 env->me_txns->mti_version = MDB_VERSION;
3398 env->me_txns->mti_magic = MDB_MAGIC;
3399 env->me_txns->mti_txnid = 0;
3400 env->me_txns->mti_numreaders = 0;
3403 if (env->me_txns->mti_magic != MDB_MAGIC) {
3404 DPUTS("lock region has invalid magic");
3408 if (env->me_txns->mti_version != MDB_VERSION) {
3409 DPRINTF("lock region is version %u, expected version %u",
3410 env->me_txns->mti_version, MDB_VERSION);
3411 rc = MDB_VERSION_MISMATCH;
3415 if (rc != EACCES && rc != EAGAIN) {
3419 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3420 if (!env->me_rmutex) goto fail_errno;
3421 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3422 if (!env->me_wmutex) goto fail_errno;
3423 #elif defined(MDB_USE_POSIX_SEM)
3424 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3425 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3426 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3427 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3438 /** The name of the lock file in the DB environment */
3439 #define LOCKNAME "/lock.mdb"
3440 /** The name of the data file in the DB environment */
3441 #define DATANAME "/data.mdb"
3442 /** The suffix of the lock file when no subdir is used */
3443 #define LOCKSUFF "-lock"
3444 /** Only a subset of the @ref mdb_env flags can be changed
3445 * at runtime. Changing other flags requires closing the
3446 * environment and re-opening it with the new flags.
3448 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3449 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3452 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3454 int oflags, rc, len, excl = -1;
3455 char *lpath, *dpath;
3457 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3461 if (flags & MDB_NOSUBDIR) {
3462 rc = len + sizeof(LOCKSUFF) + len + 1;
3464 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3469 if (flags & MDB_NOSUBDIR) {
3470 dpath = lpath + len + sizeof(LOCKSUFF);
3471 sprintf(lpath, "%s" LOCKSUFF, path);
3472 strcpy(dpath, path);
3474 dpath = lpath + len + sizeof(LOCKNAME);
3475 sprintf(lpath, "%s" LOCKNAME, path);
3476 sprintf(dpath, "%s" DATANAME, path);
3480 flags |= env->me_flags;
3481 if (flags & MDB_RDONLY) {
3482 /* silently ignore WRITEMAP when we're only getting read access */
3483 flags &= ~MDB_WRITEMAP;
3485 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3486 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3489 env->me_flags = flags |= MDB_ENV_ACTIVE;
3493 env->me_path = strdup(path);
3494 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3495 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3496 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3501 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3506 if (F_ISSET(flags, MDB_RDONLY)) {
3507 oflags = GENERIC_READ;
3508 len = OPEN_EXISTING;
3510 oflags = GENERIC_READ|GENERIC_WRITE;
3513 mode = FILE_ATTRIBUTE_NORMAL;
3514 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3515 NULL, len, mode, NULL);
3517 if (F_ISSET(flags, MDB_RDONLY))
3520 oflags = O_RDWR | O_CREAT;
3522 env->me_fd = open(dpath, oflags, mode);
3524 if (env->me_fd == INVALID_HANDLE_VALUE) {
3529 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3530 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3531 env->me_mfd = env->me_fd;
3533 /* Synchronous fd for meta writes. Needed even with
3534 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3537 env->me_mfd = CreateFile(dpath, oflags,
3538 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3539 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3541 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3543 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3548 DPRINTF("opened dbenv %p", (void *) env);
3550 rc = mdb_env_share_locks(env, &excl);
3556 mdb_env_close0(env, excl);
3562 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3564 mdb_env_close0(MDB_env *env, int excl)
3568 if (!(env->me_flags & MDB_ENV_ACTIVE))
3571 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3572 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3573 free(env->me_dbxs[i].md_name.mv_data);
3575 free(env->me_dbflags);
3578 free(env->me_dirty_list);
3579 if (env->me_free_pgs)
3580 mdb_midl_free(env->me_free_pgs);
3582 if (env->me_flags & MDB_ENV_TXKEY) {
3583 pthread_key_delete(env->me_txkey);
3585 /* Delete our key from the global list */
3586 for (i=0; i<mdb_tls_nkeys; i++)
3587 if (mdb_tls_keys[i] == env->me_txkey) {
3588 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3596 munmap(env->me_map, env->me_mapsize);
3598 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3600 if (env->me_fd != INVALID_HANDLE_VALUE)
3603 pid_t pid = env->me_pid;
3604 /* Clearing readers is done in this function because
3605 * me_txkey with its destructor must be disabled first.
3607 for (i = env->me_numreaders; --i >= 0; )
3608 if (env->me_txns->mti_readers[i].mr_pid == pid)
3609 env->me_txns->mti_readers[i].mr_pid = 0;
3611 if (env->me_rmutex) {
3612 CloseHandle(env->me_rmutex);
3613 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3615 /* Windows automatically destroys the mutexes when
3616 * the last handle closes.
3618 #elif defined(MDB_USE_POSIX_SEM)
3619 if (env->me_rmutex != SEM_FAILED) {
3620 sem_close(env->me_rmutex);
3621 if (env->me_wmutex != SEM_FAILED)
3622 sem_close(env->me_wmutex);
3623 /* If we have the filelock: If we are the
3624 * only remaining user, clean up semaphores.
3627 mdb_env_excl_lock(env, &excl);
3629 sem_unlink(env->me_txns->mti_rmname);
3630 sem_unlink(env->me_txns->mti_wmname);
3634 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3636 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3639 /* Unlock the lockfile. Windows would have unlocked it
3640 * after closing anyway, but not necessarily at once.
3642 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3648 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3652 mdb_env_copyfd(MDB_env *env, HANDLE fd)
3654 MDB_txn *txn = NULL;
3659 /* Do the lock/unlock of the reader mutex before starting the
3660 * write txn. Otherwise other read txns could block writers.
3662 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3667 /* We must start the actual read txn after blocking writers */
3668 mdb_txn_reset0(txn);
3670 /* Temporarily block writers until we snapshot the meta pages */
3673 rc = mdb_txn_renew0(txn);
3675 UNLOCK_MUTEX_W(env);
3680 wsize = env->me_psize * 2;
3684 rc = WriteFile(fd, env->me_map, wsize, &len, NULL);
3685 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3688 rc = write(fd, env->me_map, wsize);
3689 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3692 UNLOCK_MUTEX_W(env);
3697 ptr = env->me_map + wsize;
3698 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3699 #define MAX_WRITE 2147483648U
3703 if (wsize > MAX_WRITE)
3707 rc = WriteFile(fd, ptr, w2, &len, NULL);
3708 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3717 if (wsize > MAX_WRITE)
3721 wres = write(fd, ptr, w2);
3722 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3735 mdb_env_copy(MDB_env *env, const char *path)
3739 HANDLE newfd = INVALID_HANDLE_VALUE;
3741 if (env->me_flags & MDB_NOSUBDIR) {
3742 lpath = (char *)path;
3745 len += sizeof(DATANAME);
3746 lpath = malloc(len);
3749 sprintf(lpath, "%s" DATANAME, path);
3752 /* The destination path must exist, but the destination file must not.
3753 * We don't want the OS to cache the writes, since the source data is
3754 * already in the OS cache.
3757 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3758 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3760 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3766 if (!(env->me_flags & MDB_NOSUBDIR))
3768 if (newfd == INVALID_HANDLE_VALUE) {
3773 #ifdef F_NOCACHE /* __APPLE__ */
3774 rc = fcntl(newfd, F_NOCACHE, 1);
3781 rc = mdb_env_copyfd(env, newfd);
3784 if (newfd != INVALID_HANDLE_VALUE)
3791 mdb_env_close(MDB_env *env)
3798 VGMEMP_DESTROY(env);
3799 while ((dp = env->me_dpages) != NULL) {
3800 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3801 env->me_dpages = dp->mp_next;
3805 mdb_env_close0(env, 0);
3809 /** Compare two items pointing at aligned size_t's */
3811 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3813 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3814 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3817 /** Compare two items pointing at aligned int's */
3819 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3821 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3822 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3825 /** Compare two items pointing at ints of unknown alignment.
3826 * Nodes and keys are guaranteed to be 2-byte aligned.
3829 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3831 #if BYTE_ORDER == LITTLE_ENDIAN
3832 unsigned short *u, *c;
3835 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3836 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3839 } while(!x && u > (unsigned short *)a->mv_data);
3842 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3846 /** Compare two items lexically */
3848 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3855 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3861 diff = memcmp(a->mv_data, b->mv_data, len);
3862 return diff ? diff : len_diff<0 ? -1 : len_diff;
3865 /** Compare two items in reverse byte order */
3867 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3869 const unsigned char *p1, *p2, *p1_lim;
3873 p1_lim = (const unsigned char *)a->mv_data;
3874 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3875 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3877 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3883 while (p1 > p1_lim) {
3884 diff = *--p1 - *--p2;
3888 return len_diff<0 ? -1 : len_diff;
3891 /** Search for key within a page, using binary search.
3892 * Returns the smallest entry larger or equal to the key.
3893 * If exactp is non-null, stores whether the found entry was an exact match
3894 * in *exactp (1 or 0).
3895 * Updates the cursor index with the index of the found entry.
3896 * If no entry larger or equal to the key is found, returns NULL.
3899 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3901 unsigned int i = 0, nkeys;
3904 MDB_page *mp = mc->mc_pg[mc->mc_top];
3905 MDB_node *node = NULL;
3910 nkeys = NUMKEYS(mp);
3915 COPY_PGNO(pgno, mp->mp_pgno);
3916 DPRINTF("searching %u keys in %s %spage %zu",
3917 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3924 low = IS_LEAF(mp) ? 0 : 1;
3926 cmp = mc->mc_dbx->md_cmp;
3928 /* Branch pages have no data, so if using integer keys,
3929 * alignment is guaranteed. Use faster mdb_cmp_int.
3931 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3932 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3939 nodekey.mv_size = mc->mc_db->md_pad;
3940 node = NODEPTR(mp, 0); /* fake */
3941 while (low <= high) {
3942 i = (low + high) >> 1;
3943 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3944 rc = cmp(key, &nodekey);
3945 DPRINTF("found leaf index %u [%s], rc = %i",
3946 i, DKEY(&nodekey), rc);
3955 while (low <= high) {
3956 i = (low + high) >> 1;
3958 node = NODEPTR(mp, i);
3959 nodekey.mv_size = NODEKSZ(node);
3960 nodekey.mv_data = NODEKEY(node);
3962 rc = cmp(key, &nodekey);
3965 DPRINTF("found leaf index %u [%s], rc = %i",
3966 i, DKEY(&nodekey), rc);
3968 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3969 i, DKEY(&nodekey), NODEPGNO(node), rc);
3980 if (rc > 0) { /* Found entry is less than the key. */
3981 i++; /* Skip to get the smallest entry larger than key. */
3983 node = NODEPTR(mp, i);
3986 *exactp = (rc == 0);
3987 /* store the key index */
3988 mc->mc_ki[mc->mc_top] = i;
3990 /* There is no entry larger or equal to the key. */
3993 /* nodeptr is fake for LEAF2 */
3999 mdb_cursor_adjust(MDB_cursor *mc, func)
4003 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4004 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4011 /** Pop a page off the top of the cursor's stack. */
4013 mdb_cursor_pop(MDB_cursor *mc)
4016 #ifndef MDB_DEBUG_SKIP
4017 MDB_page *top = mc->mc_pg[mc->mc_top];
4023 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
4024 mc->mc_dbi, (void *) mc);
4028 /** Push a page onto the top of the cursor's stack. */
4030 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4032 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4033 mc->mc_dbi, (void *) mc);
4035 if (mc->mc_snum >= CURSOR_STACK) {
4036 assert(mc->mc_snum < CURSOR_STACK);
4037 return MDB_CURSOR_FULL;
4040 mc->mc_top = mc->mc_snum++;
4041 mc->mc_pg[mc->mc_top] = mp;
4042 mc->mc_ki[mc->mc_top] = 0;
4047 /** Find the address of the page corresponding to a given page number.
4048 * @param[in] txn the transaction for this access.
4049 * @param[in] pgno the page number for the page to retrieve.
4050 * @param[out] ret address of a pointer where the page's address will be stored.
4051 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4052 * @return 0 on success, non-zero on failure.
4055 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4060 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4061 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4066 MDB_ID2L dl = tx2->mt_u.dirty_list;
4068 unsigned x = mdb_mid2l_search(dl, pgno);
4069 if (x <= dl[0].mid && dl[x].mid == pgno) {
4075 } while ((tx2 = tx2->mt_parent) != NULL);
4078 if (pgno < txn->mt_next_pgno) {
4080 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4082 DPRINTF("page %zu not found", pgno);
4084 return MDB_PAGE_NOTFOUND;
4094 /** Search for the page a given key should be in.
4095 * Pushes parent pages on the cursor stack. This function continues a
4096 * search on a cursor that has already been initialized. (Usually by
4097 * #mdb_page_search() but also by #mdb_node_move().)
4098 * @param[in,out] mc the cursor for this operation.
4099 * @param[in] key the key to search for. If NULL, search for the lowest
4100 * page. (This is used by #mdb_cursor_first().)
4101 * @param[in] modify If true, visited pages are updated with new page numbers.
4102 * @return 0 on success, non-zero on failure.
4105 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4107 MDB_page *mp = mc->mc_pg[mc->mc_top];
4112 while (IS_BRANCH(mp)) {
4116 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4117 assert(NUMKEYS(mp) > 1);
4118 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4120 if (key == NULL) /* Initialize cursor to first page. */
4122 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4123 /* cursor to last page */
4127 node = mdb_node_search(mc, key, &exact);
4129 i = NUMKEYS(mp) - 1;
4131 i = mc->mc_ki[mc->mc_top];
4140 DPRINTF("following index %u for key [%s]",
4142 assert(i < NUMKEYS(mp));
4143 node = NODEPTR(mp, i);
4145 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4148 mc->mc_ki[mc->mc_top] = i;
4149 if ((rc = mdb_cursor_push(mc, mp)))
4153 if ((rc = mdb_page_touch(mc)) != 0)
4155 mp = mc->mc_pg[mc->mc_top];
4160 DPRINTF("internal error, index points to a %02X page!?",
4162 return MDB_CORRUPTED;
4165 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4166 key ? DKEY(key) : NULL);
4171 /** Search for the lowest key under the current branch page.
4172 * This just bypasses a NUMKEYS check in the current page
4173 * before calling mdb_page_search_root(), because the callers
4174 * are all in situations where the current page is known to
4178 mdb_page_search_lowest(MDB_cursor *mc)
4180 MDB_page *mp = mc->mc_pg[mc->mc_top];
4181 MDB_node *node = NODEPTR(mp, 0);
4184 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4187 mc->mc_ki[mc->mc_top] = 0;
4188 if ((rc = mdb_cursor_push(mc, mp)))
4190 return mdb_page_search_root(mc, NULL, 0);
4193 /** Search for the page a given key should be in.
4194 * Pushes parent pages on the cursor stack. This function just sets up
4195 * the search; it finds the root page for \b mc's database and sets this
4196 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4197 * called to complete the search.
4198 * @param[in,out] mc the cursor for this operation.
4199 * @param[in] key the key to search for. If NULL, search for the lowest
4200 * page. (This is used by #mdb_cursor_first().)
4201 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4202 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4203 * @return 0 on success, non-zero on failure.
4206 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4211 /* Make sure the txn is still viable, then find the root from
4212 * the txn's db table.
4214 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4215 DPUTS("transaction has failed, must abort");
4218 /* Make sure we're using an up-to-date root */
4219 if (mc->mc_dbi > MAIN_DBI) {
4220 if ((*mc->mc_dbflag & DB_STALE) ||
4221 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4223 unsigned char dbflag = 0;
4224 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4225 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4228 if (*mc->mc_dbflag & DB_STALE) {
4232 MDB_node *leaf = mdb_node_search(&mc2,
4233 &mc->mc_dbx->md_name, &exact);
4235 return MDB_NOTFOUND;
4236 mdb_node_read(mc->mc_txn, leaf, &data);
4237 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4239 /* The txn may not know this DBI, or another process may
4240 * have dropped and recreated the DB with other flags.
4242 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4243 return MDB_INCOMPATIBLE;
4244 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4246 if (flags & MDB_PS_MODIFY)
4248 *mc->mc_dbflag &= ~DB_STALE;
4249 *mc->mc_dbflag |= dbflag;
4252 root = mc->mc_db->md_root;
4254 if (root == P_INVALID) { /* Tree is empty. */
4255 DPUTS("tree is empty");
4256 return MDB_NOTFOUND;
4261 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4262 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4268 DPRINTF("db %u root page %zu has flags 0x%X",
4269 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4271 if (flags & MDB_PS_MODIFY) {
4272 if ((rc = mdb_page_touch(mc)))
4276 if (flags & MDB_PS_ROOTONLY)
4279 return mdb_page_search_root(mc, key, flags);
4283 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4285 MDB_txn *txn = mc->mc_txn;
4286 pgno_t pg = mp->mp_pgno;
4287 unsigned i, ovpages = mp->mp_pages;
4290 DPRINTF("free ov page %zu (%d)", pg, ovpages);
4291 /* If the page is dirty we just acquired it, so we should
4292 * give it back to our current free list, if any.
4293 * Not currently supported in nested txns.
4294 * Otherwise put it onto the list of pages we freed in this txn.
4296 if ((mp->mp_flags & P_DIRTY) && !txn->mt_parent && txn->mt_env->me_pghead) {
4298 pgno_t *mop = txn->mt_env->me_pghead;
4299 MDB_ID2 *dl, ix, iy;
4300 /* Prepare to insert pg */
4301 j = mop[0] + ovpages;
4303 rc = mdb_midl_grow(&mop, ovpages);
4306 txn->mt_env->me_pghead = mop;
4308 /* Remove from dirty list */
4309 dl = txn->mt_u.dirty_list;
4311 for (ix = dl[x]; ix.mid != pg; ix = iy) {
4318 return MDB_CORRUPTED;
4321 /* Insert in me_pghead */
4322 for (i = mop[0]; i && mop[i] < pg; i--)
4328 for (i=0; i<ovpages; i++) {
4329 mdb_midl_append(&txn->mt_free_pgs, pg);
4333 mc->mc_db->md_overflow_pages -= ovpages;
4337 /** Return the data associated with a given node.
4338 * @param[in] txn The transaction for this operation.
4339 * @param[in] leaf The node being read.
4340 * @param[out] data Updated to point to the node's data.
4341 * @return 0 on success, non-zero on failure.
4344 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4346 MDB_page *omp; /* overflow page */
4350 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4351 data->mv_size = NODEDSZ(leaf);
4352 data->mv_data = NODEDATA(leaf);
4356 /* Read overflow data.
4358 data->mv_size = NODEDSZ(leaf);
4359 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4360 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4361 DPRINTF("read overflow page %zu failed", pgno);
4364 data->mv_data = METADATA(omp);
4370 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4371 MDB_val *key, MDB_val *data)
4380 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4382 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4385 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4389 mdb_cursor_init(&mc, txn, dbi, &mx);
4390 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4393 /** Find a sibling for a page.
4394 * Replaces the page at the top of the cursor's stack with the
4395 * specified sibling, if one exists.
4396 * @param[in] mc The cursor for this operation.
4397 * @param[in] move_right Non-zero if the right sibling is requested,
4398 * otherwise the left sibling.
4399 * @return 0 on success, non-zero on failure.
4402 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4408 if (mc->mc_snum < 2) {
4409 return MDB_NOTFOUND; /* root has no siblings */
4413 DPRINTF("parent page is page %zu, index %u",
4414 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4416 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4417 : (mc->mc_ki[mc->mc_top] == 0)) {
4418 DPRINTF("no more keys left, moving to %s sibling",
4419 move_right ? "right" : "left");
4420 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4421 /* undo cursor_pop before returning */
4428 mc->mc_ki[mc->mc_top]++;
4430 mc->mc_ki[mc->mc_top]--;
4431 DPRINTF("just moving to %s index key %u",
4432 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4434 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4436 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4437 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4440 mdb_cursor_push(mc, mp);
4442 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4447 /** Move the cursor to the next data item. */
4449 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4455 if (mc->mc_flags & C_EOF) {
4456 return MDB_NOTFOUND;
4459 assert(mc->mc_flags & C_INITIALIZED);
4461 mp = mc->mc_pg[mc->mc_top];
4463 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4464 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4465 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4466 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4467 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4468 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4472 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4473 if (op == MDB_NEXT_DUP)
4474 return MDB_NOTFOUND;
4478 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4480 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4481 DPUTS("=====> move to next sibling page");
4482 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4483 mc->mc_flags |= C_EOF;
4484 mc->mc_flags &= ~C_INITIALIZED;
4485 return MDB_NOTFOUND;
4487 mp = mc->mc_pg[mc->mc_top];
4488 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4490 mc->mc_ki[mc->mc_top]++;
4492 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4493 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4496 key->mv_size = mc->mc_db->md_pad;
4497 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4501 assert(IS_LEAF(mp));
4502 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4504 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4505 mdb_xcursor_init1(mc, leaf);
4508 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4511 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4512 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4513 if (rc != MDB_SUCCESS)
4518 MDB_GET_KEY(leaf, key);
4522 /** Move the cursor to the previous data item. */
4524 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4530 assert(mc->mc_flags & C_INITIALIZED);
4532 mp = mc->mc_pg[mc->mc_top];
4534 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4535 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4536 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4537 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4538 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4539 if (op != MDB_PREV || rc == MDB_SUCCESS)
4542 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4543 if (op == MDB_PREV_DUP)
4544 return MDB_NOTFOUND;
4549 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4551 if (mc->mc_ki[mc->mc_top] == 0) {
4552 DPUTS("=====> move to prev sibling page");
4553 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4554 mc->mc_flags &= ~C_INITIALIZED;
4555 return MDB_NOTFOUND;
4557 mp = mc->mc_pg[mc->mc_top];
4558 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4559 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4561 mc->mc_ki[mc->mc_top]--;
4563 mc->mc_flags &= ~C_EOF;
4565 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4566 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4569 key->mv_size = mc->mc_db->md_pad;
4570 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4574 assert(IS_LEAF(mp));
4575 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4577 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4578 mdb_xcursor_init1(mc, leaf);
4581 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4584 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4585 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4586 if (rc != MDB_SUCCESS)
4591 MDB_GET_KEY(leaf, key);
4595 /** Set the cursor on a specific data item. */
4597 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4598 MDB_cursor_op op, int *exactp)
4602 MDB_node *leaf = NULL;
4607 assert(key->mv_size > 0);
4609 /* See if we're already on the right page */
4610 if (mc->mc_flags & C_INITIALIZED) {
4613 mp = mc->mc_pg[mc->mc_top];
4615 mc->mc_ki[mc->mc_top] = 0;
4616 return MDB_NOTFOUND;
4618 if (mp->mp_flags & P_LEAF2) {
4619 nodekey.mv_size = mc->mc_db->md_pad;
4620 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4622 leaf = NODEPTR(mp, 0);
4623 MDB_GET_KEY(leaf, &nodekey);
4625 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4627 /* Probably happens rarely, but first node on the page
4628 * was the one we wanted.
4630 mc->mc_ki[mc->mc_top] = 0;
4637 unsigned int nkeys = NUMKEYS(mp);
4639 if (mp->mp_flags & P_LEAF2) {
4640 nodekey.mv_data = LEAF2KEY(mp,
4641 nkeys-1, nodekey.mv_size);
4643 leaf = NODEPTR(mp, nkeys-1);
4644 MDB_GET_KEY(leaf, &nodekey);
4646 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4648 /* last node was the one we wanted */
4649 mc->mc_ki[mc->mc_top] = nkeys-1;
4655 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4656 /* This is definitely the right page, skip search_page */
4657 if (mp->mp_flags & P_LEAF2) {
4658 nodekey.mv_data = LEAF2KEY(mp,
4659 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4661 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4662 MDB_GET_KEY(leaf, &nodekey);
4664 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4666 /* current node was the one we wanted */
4676 /* If any parents have right-sibs, search.
4677 * Otherwise, there's nothing further.
4679 for (i=0; i<mc->mc_top; i++)
4681 NUMKEYS(mc->mc_pg[i])-1)
4683 if (i == mc->mc_top) {
4684 /* There are no other pages */
4685 mc->mc_ki[mc->mc_top] = nkeys;
4686 return MDB_NOTFOUND;
4690 /* There are no other pages */
4691 mc->mc_ki[mc->mc_top] = 0;
4692 return MDB_NOTFOUND;
4696 rc = mdb_page_search(mc, key, 0);
4697 if (rc != MDB_SUCCESS)
4700 mp = mc->mc_pg[mc->mc_top];
4701 assert(IS_LEAF(mp));
4704 leaf = mdb_node_search(mc, key, exactp);
4705 if (exactp != NULL && !*exactp) {
4706 /* MDB_SET specified and not an exact match. */
4707 return MDB_NOTFOUND;
4711 DPUTS("===> inexact leaf not found, goto sibling");
4712 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4713 return rc; /* no entries matched */
4714 mp = mc->mc_pg[mc->mc_top];
4715 assert(IS_LEAF(mp));
4716 leaf = NODEPTR(mp, 0);
4720 mc->mc_flags |= C_INITIALIZED;
4721 mc->mc_flags &= ~C_EOF;
4724 key->mv_size = mc->mc_db->md_pad;
4725 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4729 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4730 mdb_xcursor_init1(mc, leaf);
4733 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4734 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4735 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4738 if (op == MDB_GET_BOTH) {
4744 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4745 if (rc != MDB_SUCCESS)
4748 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4750 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4752 rc = mc->mc_dbx->md_dcmp(data, &d2);
4754 if (op == MDB_GET_BOTH || rc > 0)
4755 return MDB_NOTFOUND;
4760 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4761 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4766 /* The key already matches in all other cases */
4767 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4768 MDB_GET_KEY(leaf, key);
4769 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4774 /** Move the cursor to the first item in the database. */
4776 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4781 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4782 rc = mdb_page_search(mc, NULL, 0);
4783 if (rc != MDB_SUCCESS)
4786 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4788 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4789 mc->mc_flags |= C_INITIALIZED;
4790 mc->mc_flags &= ~C_EOF;
4792 mc->mc_ki[mc->mc_top] = 0;
4794 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4795 key->mv_size = mc->mc_db->md_pad;
4796 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4801 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4802 mdb_xcursor_init1(mc, leaf);
4803 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4808 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4809 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4813 MDB_GET_KEY(leaf, key);
4817 /** Move the cursor to the last item in the database. */
4819 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4824 if (!(mc->mc_flags & C_EOF)) {
4826 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4829 lkey.mv_size = MDB_MAXKEYSIZE+1;
4830 lkey.mv_data = NULL;
4831 rc = mdb_page_search(mc, &lkey, 0);
4832 if (rc != MDB_SUCCESS)
4835 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4838 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4839 mc->mc_flags |= C_INITIALIZED|C_EOF;
4840 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4842 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4843 key->mv_size = mc->mc_db->md_pad;
4844 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4849 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4850 mdb_xcursor_init1(mc, leaf);
4851 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4856 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4857 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4862 MDB_GET_KEY(leaf, key);
4867 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4876 case MDB_GET_CURRENT:
4877 if (!(mc->mc_flags & C_INITIALIZED)) {
4880 MDB_page *mp = mc->mc_pg[mc->mc_top];
4882 mc->mc_ki[mc->mc_top] = 0;
4888 key->mv_size = mc->mc_db->md_pad;
4889 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4891 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4892 MDB_GET_KEY(leaf, key);
4894 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4895 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4897 rc = mdb_node_read(mc->mc_txn, leaf, data);
4904 case MDB_GET_BOTH_RANGE:
4905 if (data == NULL || mc->mc_xcursor == NULL) {
4913 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4915 } else if (op == MDB_SET_RANGE)
4916 rc = mdb_cursor_set(mc, key, data, op, NULL);
4918 rc = mdb_cursor_set(mc, key, data, op, &exact);
4920 case MDB_GET_MULTIPLE:
4922 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4923 !(mc->mc_flags & C_INITIALIZED)) {
4928 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4929 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4932 case MDB_NEXT_MULTIPLE:
4934 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4938 if (!(mc->mc_flags & C_INITIALIZED))
4939 rc = mdb_cursor_first(mc, key, data);
4941 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4942 if (rc == MDB_SUCCESS) {
4943 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4946 mx = &mc->mc_xcursor->mx_cursor;
4947 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4949 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4950 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4958 case MDB_NEXT_NODUP:
4959 if (!(mc->mc_flags & C_INITIALIZED))
4960 rc = mdb_cursor_first(mc, key, data);
4962 rc = mdb_cursor_next(mc, key, data, op);
4966 case MDB_PREV_NODUP:
4967 if (!(mc->mc_flags & C_INITIALIZED)) {
4968 rc = mdb_cursor_last(mc, key, data);
4971 mc->mc_flags |= C_INITIALIZED;
4972 mc->mc_ki[mc->mc_top]++;
4974 rc = mdb_cursor_prev(mc, key, data, op);
4977 rc = mdb_cursor_first(mc, key, data);
4981 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4982 !(mc->mc_flags & C_INITIALIZED) ||
4983 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4987 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4990 rc = mdb_cursor_last(mc, key, data);
4994 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4995 !(mc->mc_flags & C_INITIALIZED) ||
4996 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5000 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5003 DPRINTF("unhandled/unimplemented cursor operation %u", op);
5011 /** Touch all the pages in the cursor stack.
5012 * Makes sure all the pages are writable, before attempting a write operation.
5013 * @param[in] mc The cursor to operate on.
5016 mdb_cursor_touch(MDB_cursor *mc)
5020 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5023 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5024 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5027 *mc->mc_dbflag |= DB_DIRTY;
5029 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
5030 rc = mdb_page_touch(mc);
5034 mc->mc_top = mc->mc_snum-1;
5039 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5042 MDB_node *leaf = NULL;
5043 MDB_val xdata, *rdata, dkey;
5046 int do_sub = 0, insert = 0;
5047 unsigned int mcount = 0;
5051 char dbuf[MDB_MAXKEYSIZE+1];
5052 unsigned int nflags;
5055 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5058 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
5061 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
5064 #if SIZE_MAX > MAXDATASIZE
5065 if (data->mv_size > MAXDATASIZE)
5069 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
5070 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5074 if (flags == MDB_CURRENT) {
5075 if (!(mc->mc_flags & C_INITIALIZED))
5078 } else if (mc->mc_db->md_root == P_INVALID) {
5080 /* new database, write a root leaf page */
5081 DPUTS("allocating new root leaf page");
5082 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5086 mdb_cursor_push(mc, np);
5087 mc->mc_db->md_root = np->mp_pgno;
5088 mc->mc_db->md_depth++;
5089 *mc->mc_dbflag |= DB_DIRTY;
5090 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5092 np->mp_flags |= P_LEAF2;
5093 mc->mc_flags |= C_INITIALIZED;
5099 if (flags & MDB_APPEND) {
5101 rc = mdb_cursor_last(mc, &k2, &d2);
5103 rc = mc->mc_dbx->md_cmp(key, &k2);
5106 mc->mc_ki[mc->mc_top]++;
5108 /* new key is <= last key */
5113 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5115 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5116 DPRINTF("duplicate key [%s]", DKEY(key));
5118 return MDB_KEYEXIST;
5120 if (rc && rc != MDB_NOTFOUND)
5124 /* Cursor is positioned, now make sure all pages are writable */
5125 rc2 = mdb_cursor_touch(mc);
5130 /* The key already exists */
5131 if (rc == MDB_SUCCESS) {
5132 /* there's only a key anyway, so this is a no-op */
5133 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5134 unsigned int ksize = mc->mc_db->md_pad;
5135 if (key->mv_size != ksize)
5137 if (flags == MDB_CURRENT) {
5138 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5139 memcpy(ptr, key->mv_data, ksize);
5144 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5147 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5148 /* Was a single item before, must convert now */
5150 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5151 /* Just overwrite the current item */
5152 if (flags == MDB_CURRENT)
5155 dkey.mv_size = NODEDSZ(leaf);
5156 dkey.mv_data = NODEDATA(leaf);
5157 #if UINT_MAX < SIZE_MAX
5158 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5159 #ifdef MISALIGNED_OK
5160 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5162 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5165 /* if data matches, ignore it */
5166 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5167 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5169 /* create a fake page for the dup items */
5170 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5171 dkey.mv_data = dbuf;
5172 fp = (MDB_page *)&pbuf;
5173 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5174 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5175 fp->mp_lower = PAGEHDRSZ;
5176 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5177 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5178 fp->mp_flags |= P_LEAF2;
5179 fp->mp_pad = data->mv_size;
5180 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5182 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5183 (dkey.mv_size & 1) + (data->mv_size & 1);
5185 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5188 xdata.mv_size = fp->mp_upper;
5193 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5194 /* See if we need to convert from fake page to subDB */
5196 unsigned int offset;
5200 fp = NODEDATA(leaf);
5201 if (flags == MDB_CURRENT) {
5203 fp->mp_flags |= P_DIRTY;
5204 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5205 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5209 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5210 offset = fp->mp_pad;
5211 if (SIZELEFT(fp) >= offset)
5213 offset *= 4; /* space for 4 more */
5215 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5217 offset += offset & 1;
5218 fp_flags = fp->mp_flags;
5219 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5220 offset >= mc->mc_txn->mt_env->me_nodemax) {
5221 /* yes, convert it */
5223 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5224 dummy.md_pad = fp->mp_pad;
5225 dummy.md_flags = MDB_DUPFIXED;
5226 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5227 dummy.md_flags |= MDB_INTEGERKEY;
5230 dummy.md_branch_pages = 0;
5231 dummy.md_leaf_pages = 1;
5232 dummy.md_overflow_pages = 0;
5233 dummy.md_entries = NUMKEYS(fp);
5235 xdata.mv_size = sizeof(MDB_db);
5236 xdata.mv_data = &dummy;
5237 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5239 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5240 flags |= F_DUPDATA|F_SUBDATA;
5241 dummy.md_root = mp->mp_pgno;
5242 fp_flags &= ~P_SUBP;
5244 /* no, just grow it */
5246 xdata.mv_size = NODEDSZ(leaf) + offset;
5247 xdata.mv_data = &pbuf;
5248 mp = (MDB_page *)&pbuf;
5249 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5252 mp->mp_flags = fp_flags | P_DIRTY;
5253 mp->mp_pad = fp->mp_pad;
5254 mp->mp_lower = fp->mp_lower;
5255 mp->mp_upper = fp->mp_upper + offset;
5257 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5259 nsize = NODEDSZ(leaf) - fp->mp_upper;
5260 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5261 for (i=0; i<NUMKEYS(fp); i++)
5262 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5264 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5268 /* data is on sub-DB, just store it */
5269 flags |= F_DUPDATA|F_SUBDATA;
5273 /* overflow page overwrites need special handling */
5274 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5277 unsigned psize = mc->mc_txn->mt_env->me_psize;
5278 int level, ovpages, dpages = OVPAGES(data->mv_size, psize);
5280 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5281 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
5283 ovpages = omp->mp_pages;
5285 /* Is the ov page writable and large enough? */
5286 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5287 /* yes, overwrite it. Note in this case we don't
5288 * bother to try shrinking the page if the new data
5289 * is smaller than the overflow threshold.
5292 /* It is writable only in a parent txn */
5293 size_t sz = (size_t) psize * ovpages, off;
5294 MDB_page *np = mdb_page_malloc(mc, ovpages);
5300 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
5301 if (!(flags & MDB_RESERVE)) {
5302 /* Copy end of page, adjusting alignment so
5303 * compiler may copy words instead of bytes.
5305 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
5306 memcpy((size_t *)((char *)np + off),
5307 (size_t *)((char *)omp + off), sz - off);
5310 memcpy(np, omp, sz); /* Copy beginning of page */
5313 SETDSZ(leaf, data->mv_size);
5314 if (F_ISSET(flags, MDB_RESERVE))
5315 data->mv_data = METADATA(omp);
5317 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5320 mdb_ovpage_free(mc, omp);
5322 } else if (NODEDSZ(leaf) == data->mv_size) {
5323 /* same size, just replace it. Note that we could
5324 * also reuse this node if the new data is smaller,
5325 * but instead we opt to shrink the node in that case.
5327 if (F_ISSET(flags, MDB_RESERVE))
5328 data->mv_data = NODEDATA(leaf);
5329 else if (data->mv_size)
5330 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5332 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5335 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5336 mc->mc_db->md_entries--;
5338 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5345 nflags = flags & NODE_ADD_FLAGS;
5346 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5347 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5348 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5349 nflags &= ~MDB_APPEND;
5351 nflags |= MDB_SPLIT_REPLACE;
5352 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5354 /* There is room already in this leaf page. */
5355 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5356 if (rc == 0 && !do_sub && insert) {
5357 /* Adjust other cursors pointing to mp */
5358 MDB_cursor *m2, *m3;
5359 MDB_dbi dbi = mc->mc_dbi;
5360 unsigned i = mc->mc_top;
5361 MDB_page *mp = mc->mc_pg[i];
5363 if (mc->mc_flags & C_SUB)
5366 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5367 if (mc->mc_flags & C_SUB)
5368 m3 = &m2->mc_xcursor->mx_cursor;
5371 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5372 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5379 if (rc != MDB_SUCCESS)
5380 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5382 /* Now store the actual data in the child DB. Note that we're
5383 * storing the user data in the keys field, so there are strict
5384 * size limits on dupdata. The actual data fields of the child
5385 * DB are all zero size.
5392 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5393 if (flags & MDB_CURRENT) {
5394 xflags = MDB_CURRENT;
5396 mdb_xcursor_init1(mc, leaf);
5397 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5399 /* converted, write the original data first */
5401 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5405 /* Adjust other cursors pointing to mp */
5407 unsigned i = mc->mc_top;
5408 MDB_page *mp = mc->mc_pg[i];
5410 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5411 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5412 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5413 mdb_xcursor_init1(m2, leaf);
5417 /* we've done our job */
5420 if (flags & MDB_APPENDDUP)
5421 xflags |= MDB_APPEND;
5422 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5423 if (flags & F_SUBDATA) {
5424 void *db = NODEDATA(leaf);
5425 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5428 /* sub-writes might have failed so check rc again.
5429 * Don't increment count if we just replaced an existing item.
5431 if (!rc && !(flags & MDB_CURRENT))
5432 mc->mc_db->md_entries++;
5433 if (flags & MDB_MULTIPLE) {
5435 if (mcount < data[1].mv_size) {
5436 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5437 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5443 /* If we succeeded and the key didn't exist before, make sure
5444 * the cursor is marked valid.
5447 mc->mc_flags |= C_INITIALIZED;
5452 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5457 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5460 if (!(mc->mc_flags & C_INITIALIZED))
5463 rc = mdb_cursor_touch(mc);
5467 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5469 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5470 if (flags != MDB_NODUPDATA) {
5471 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5472 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5474 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5475 /* If sub-DB still has entries, we're done */
5476 if (mc->mc_xcursor->mx_db.md_entries) {
5477 if (leaf->mn_flags & F_SUBDATA) {
5478 /* update subDB info */
5479 void *db = NODEDATA(leaf);
5480 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5483 /* shrink fake page */
5484 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5485 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5486 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5487 /* fix other sub-DB cursors pointed at this fake page */
5488 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5489 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5490 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top] &&
5491 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
5492 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5495 mc->mc_db->md_entries--;
5498 /* otherwise fall thru and delete the sub-DB */
5501 if (leaf->mn_flags & F_SUBDATA) {
5502 /* add all the child DB's pages to the free list */
5503 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5504 if (rc == MDB_SUCCESS) {
5505 mc->mc_db->md_entries -=
5506 mc->mc_xcursor->mx_db.md_entries;
5511 return mdb_cursor_del0(mc, leaf);
5514 /** Allocate and initialize new pages for a database.
5515 * @param[in] mc a cursor on the database being added to.
5516 * @param[in] flags flags defining what type of page is being allocated.
5517 * @param[in] num the number of pages to allocate. This is usually 1,
5518 * unless allocating overflow pages for a large record.
5519 * @param[out] mp Address of a page, or NULL on failure.
5520 * @return 0 on success, non-zero on failure.
5523 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5528 if ((rc = mdb_page_alloc(mc, num, &np)))
5530 DPRINTF("allocated new mpage %zu, page size %u",
5531 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5532 np->mp_flags = flags | P_DIRTY;
5533 np->mp_lower = PAGEHDRSZ;
5534 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5537 mc->mc_db->md_branch_pages++;
5538 else if (IS_LEAF(np))
5539 mc->mc_db->md_leaf_pages++;
5540 else if (IS_OVERFLOW(np)) {
5541 mc->mc_db->md_overflow_pages += num;
5549 /** Calculate the size of a leaf node.
5550 * The size depends on the environment's page size; if a data item
5551 * is too large it will be put onto an overflow page and the node
5552 * size will only include the key and not the data. Sizes are always
5553 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5554 * of the #MDB_node headers.
5555 * @param[in] env The environment handle.
5556 * @param[in] key The key for the node.
5557 * @param[in] data The data for the node.
5558 * @return The number of bytes needed to store the node.
5561 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5565 sz = LEAFSIZE(key, data);
5566 if (sz >= env->me_nodemax) {
5567 /* put on overflow page */
5568 sz -= data->mv_size - sizeof(pgno_t);
5572 return sz + sizeof(indx_t);
5575 /** Calculate the size of a branch node.
5576 * The size should depend on the environment's page size but since
5577 * we currently don't support spilling large keys onto overflow
5578 * pages, it's simply the size of the #MDB_node header plus the
5579 * size of the key. Sizes are always rounded up to an even number
5580 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5581 * @param[in] env The environment handle.
5582 * @param[in] key The key for the node.
5583 * @return The number of bytes needed to store the node.
5586 mdb_branch_size(MDB_env *env, MDB_val *key)
5591 if (sz >= env->me_nodemax) {
5592 /* put on overflow page */
5593 /* not implemented */
5594 /* sz -= key->size - sizeof(pgno_t); */
5597 return sz + sizeof(indx_t);
5600 /** Add a node to the page pointed to by the cursor.
5601 * @param[in] mc The cursor for this operation.
5602 * @param[in] indx The index on the page where the new node should be added.
5603 * @param[in] key The key for the new node.
5604 * @param[in] data The data for the new node, if any.
5605 * @param[in] pgno The page number, if adding a branch node.
5606 * @param[in] flags Flags for the node.
5607 * @return 0 on success, non-zero on failure. Possible errors are:
5609 * <li>ENOMEM - failed to allocate overflow pages for the node.
5610 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5611 * should never happen since all callers already calculate the
5612 * page's free space before calling this function.
5616 mdb_node_add(MDB_cursor *mc, indx_t indx,
5617 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5620 size_t node_size = NODESIZE;
5623 MDB_page *mp = mc->mc_pg[mc->mc_top];
5624 MDB_page *ofp = NULL; /* overflow page */
5627 assert(mp->mp_upper >= mp->mp_lower);
5629 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5630 IS_LEAF(mp) ? "leaf" : "branch",
5631 IS_SUBP(mp) ? "sub-" : "",
5632 mp->mp_pgno, indx, data ? data->mv_size : 0,
5633 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5636 /* Move higher keys up one slot. */
5637 int ksize = mc->mc_db->md_pad, dif;
5638 char *ptr = LEAF2KEY(mp, indx, ksize);
5639 dif = NUMKEYS(mp) - indx;
5641 memmove(ptr+ksize, ptr, dif*ksize);
5642 /* insert new key */
5643 memcpy(ptr, key->mv_data, ksize);
5645 /* Just using these for counting */
5646 mp->mp_lower += sizeof(indx_t);
5647 mp->mp_upper -= ksize - sizeof(indx_t);
5652 node_size += key->mv_size;
5656 if (F_ISSET(flags, F_BIGDATA)) {
5657 /* Data already on overflow page. */
5658 node_size += sizeof(pgno_t);
5659 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5660 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5662 /* Put data on overflow page. */
5663 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5664 data->mv_size, node_size+data->mv_size);
5665 node_size += sizeof(pgno_t);
5666 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5668 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5671 node_size += data->mv_size;
5674 node_size += node_size & 1;
5676 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5677 DPRINTF("not enough room in page %zu, got %u ptrs",
5678 mp->mp_pgno, NUMKEYS(mp));
5679 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5680 mp->mp_upper - mp->mp_lower);
5681 DPRINTF("node size = %zu", node_size);
5682 return MDB_PAGE_FULL;
5685 /* Move higher pointers up one slot. */
5686 for (i = NUMKEYS(mp); i > indx; i--)
5687 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5689 /* Adjust free space offsets. */
5690 ofs = mp->mp_upper - node_size;
5691 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5692 mp->mp_ptrs[indx] = ofs;
5694 mp->mp_lower += sizeof(indx_t);
5696 /* Write the node data. */
5697 node = NODEPTR(mp, indx);
5698 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5699 node->mn_flags = flags;
5701 SETDSZ(node,data->mv_size);
5706 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5711 if (F_ISSET(flags, F_BIGDATA))
5712 memcpy(node->mn_data + key->mv_size, data->mv_data,
5714 else if (F_ISSET(flags, MDB_RESERVE))
5715 data->mv_data = node->mn_data + key->mv_size;
5717 memcpy(node->mn_data + key->mv_size, data->mv_data,
5720 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5722 if (F_ISSET(flags, MDB_RESERVE))
5723 data->mv_data = METADATA(ofp);
5725 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5732 /** Delete the specified node from a page.
5733 * @param[in] mp The page to operate on.
5734 * @param[in] indx The index of the node to delete.
5735 * @param[in] ksize The size of a node. Only used if the page is
5736 * part of a #MDB_DUPFIXED database.
5739 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5742 indx_t i, j, numkeys, ptr;
5749 COPY_PGNO(pgno, mp->mp_pgno);
5750 DPRINTF("delete node %u on %s page %zu", indx,
5751 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5754 assert(indx < NUMKEYS(mp));
5757 int x = NUMKEYS(mp) - 1 - indx;
5758 base = LEAF2KEY(mp, indx, ksize);
5760 memmove(base, base + ksize, x * ksize);
5761 mp->mp_lower -= sizeof(indx_t);
5762 mp->mp_upper += ksize - sizeof(indx_t);
5766 node = NODEPTR(mp, indx);
5767 sz = NODESIZE + node->mn_ksize;
5769 if (F_ISSET(node->mn_flags, F_BIGDATA))
5770 sz += sizeof(pgno_t);
5772 sz += NODEDSZ(node);
5776 ptr = mp->mp_ptrs[indx];
5777 numkeys = NUMKEYS(mp);
5778 for (i = j = 0; i < numkeys; i++) {
5780 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5781 if (mp->mp_ptrs[i] < ptr)
5782 mp->mp_ptrs[j] += sz;
5787 base = (char *)mp + mp->mp_upper;
5788 memmove(base + sz, base, ptr - mp->mp_upper);
5790 mp->mp_lower -= sizeof(indx_t);
5794 /** Compact the main page after deleting a node on a subpage.
5795 * @param[in] mp The main page to operate on.
5796 * @param[in] indx The index of the subpage on the main page.
5799 mdb_node_shrink(MDB_page *mp, indx_t indx)
5806 indx_t i, numkeys, ptr;
5808 node = NODEPTR(mp, indx);
5809 sp = (MDB_page *)NODEDATA(node);
5810 osize = NODEDSZ(node);
5812 delta = sp->mp_upper - sp->mp_lower;
5813 SETDSZ(node, osize - delta);
5814 xp = (MDB_page *)((char *)sp + delta);
5816 /* shift subpage upward */
5818 nsize = NUMKEYS(sp) * sp->mp_pad;
5819 memmove(METADATA(xp), METADATA(sp), nsize);
5822 nsize = osize - sp->mp_upper;
5823 numkeys = NUMKEYS(sp);
5824 for (i=numkeys-1; i>=0; i--)
5825 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5827 xp->mp_upper = sp->mp_lower;
5828 xp->mp_lower = sp->mp_lower;
5829 xp->mp_flags = sp->mp_flags;
5830 xp->mp_pad = sp->mp_pad;
5831 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5833 /* shift lower nodes upward */
5834 ptr = mp->mp_ptrs[indx];
5835 numkeys = NUMKEYS(mp);
5836 for (i = 0; i < numkeys; i++) {
5837 if (mp->mp_ptrs[i] <= ptr)
5838 mp->mp_ptrs[i] += delta;
5841 base = (char *)mp + mp->mp_upper;
5842 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5843 mp->mp_upper += delta;
5846 /** Initial setup of a sorted-dups cursor.
5847 * Sorted duplicates are implemented as a sub-database for the given key.
5848 * The duplicate data items are actually keys of the sub-database.
5849 * Operations on the duplicate data items are performed using a sub-cursor
5850 * initialized when the sub-database is first accessed. This function does
5851 * the preliminary setup of the sub-cursor, filling in the fields that
5852 * depend only on the parent DB.
5853 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5856 mdb_xcursor_init0(MDB_cursor *mc)
5858 MDB_xcursor *mx = mc->mc_xcursor;
5860 mx->mx_cursor.mc_xcursor = NULL;
5861 mx->mx_cursor.mc_txn = mc->mc_txn;
5862 mx->mx_cursor.mc_db = &mx->mx_db;
5863 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5864 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5865 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5866 mx->mx_cursor.mc_snum = 0;
5867 mx->mx_cursor.mc_top = 0;
5868 mx->mx_cursor.mc_flags = C_SUB;
5869 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5870 mx->mx_dbx.md_dcmp = NULL;
5871 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5874 /** Final setup of a sorted-dups cursor.
5875 * Sets up the fields that depend on the data from the main cursor.
5876 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5877 * @param[in] node The data containing the #MDB_db record for the
5878 * sorted-dup database.
5881 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5883 MDB_xcursor *mx = mc->mc_xcursor;
5885 if (node->mn_flags & F_SUBDATA) {
5886 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5887 mx->mx_cursor.mc_pg[0] = 0;
5888 mx->mx_cursor.mc_snum = 0;
5889 mx->mx_cursor.mc_flags = C_SUB;
5891 MDB_page *fp = NODEDATA(node);
5892 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5893 mx->mx_db.md_flags = 0;
5894 mx->mx_db.md_depth = 1;
5895 mx->mx_db.md_branch_pages = 0;
5896 mx->mx_db.md_leaf_pages = 1;
5897 mx->mx_db.md_overflow_pages = 0;
5898 mx->mx_db.md_entries = NUMKEYS(fp);
5899 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5900 mx->mx_cursor.mc_snum = 1;
5901 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5902 mx->mx_cursor.mc_top = 0;
5903 mx->mx_cursor.mc_pg[0] = fp;
5904 mx->mx_cursor.mc_ki[0] = 0;
5905 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5906 mx->mx_db.md_flags = MDB_DUPFIXED;
5907 mx->mx_db.md_pad = fp->mp_pad;
5908 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5909 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5912 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5914 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5916 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5917 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5918 #if UINT_MAX < SIZE_MAX
5919 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5920 #ifdef MISALIGNED_OK
5921 mx->mx_dbx.md_cmp = mdb_cmp_long;
5923 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5928 /** Initialize a cursor for a given transaction and database. */
5930 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5935 mc->mc_db = &txn->mt_dbs[dbi];
5936 mc->mc_dbx = &txn->mt_dbxs[dbi];
5937 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5942 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5944 mc->mc_xcursor = mx;
5945 mdb_xcursor_init0(mc);
5947 mc->mc_xcursor = NULL;
5949 if (*mc->mc_dbflag & DB_STALE) {
5950 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5955 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5958 size_t size = sizeof(MDB_cursor);
5960 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5963 /* Allow read access to the freelist */
5964 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5967 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5968 size += sizeof(MDB_xcursor);
5970 if ((mc = malloc(size)) != NULL) {
5971 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5972 if (txn->mt_cursors) {
5973 mc->mc_next = txn->mt_cursors[dbi];
5974 txn->mt_cursors[dbi] = mc;
5975 mc->mc_flags |= C_UNTRACK;
5977 mc->mc_flags |= C_ALLOCD;
5988 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5992 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5995 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5998 flags = mc->mc_flags;
6000 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6002 mc->mc_flags |= (flags & C_ALLOCD);
6006 /* Return the count of duplicate data items for the current key */
6008 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6012 if (mc == NULL || countp == NULL)
6015 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
6018 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6019 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6022 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6025 *countp = mc->mc_xcursor->mx_db.md_entries;
6031 mdb_cursor_close(MDB_cursor *mc)
6034 /* remove from txn, if tracked */
6035 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6036 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6037 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6039 *prev = mc->mc_next;
6041 if (mc->mc_flags & C_ALLOCD)
6047 mdb_cursor_txn(MDB_cursor *mc)
6049 if (!mc) return NULL;
6054 mdb_cursor_dbi(MDB_cursor *mc)
6060 /** Replace the key for a node with a new key.
6061 * @param[in] mc Cursor pointing to the node to operate on.
6062 * @param[in] key The new key to use.
6063 * @return 0 on success, non-zero on failure.
6066 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6073 indx_t ptr, i, numkeys, indx;
6076 indx = mc->mc_ki[mc->mc_top];
6077 mp = mc->mc_pg[mc->mc_top];
6078 node = NODEPTR(mp, indx);
6079 ptr = mp->mp_ptrs[indx];
6083 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
6084 k2.mv_data = NODEKEY(node);
6085 k2.mv_size = node->mn_ksize;
6086 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
6088 mdb_dkey(&k2, kbuf2),
6094 delta0 = delta = key->mv_size - node->mn_ksize;
6096 /* Must be 2-byte aligned. If new key is
6097 * shorter by 1, the shift will be skipped.
6099 delta += (delta & 1);
6101 if (delta > 0 && SIZELEFT(mp) < delta) {
6103 /* not enough space left, do a delete and split */
6104 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6105 pgno = NODEPGNO(node);
6106 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6107 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6110 numkeys = NUMKEYS(mp);
6111 for (i = 0; i < numkeys; i++) {
6112 if (mp->mp_ptrs[i] <= ptr)
6113 mp->mp_ptrs[i] -= delta;
6116 base = (char *)mp + mp->mp_upper;
6117 len = ptr - mp->mp_upper + NODESIZE;
6118 memmove(base - delta, base, len);
6119 mp->mp_upper -= delta;
6121 node = NODEPTR(mp, indx);
6124 /* But even if no shift was needed, update ksize */
6126 node->mn_ksize = key->mv_size;
6129 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6135 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6137 /** Move a node from csrc to cdst.
6140 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6147 unsigned short flags;
6151 /* Mark src and dst as dirty. */
6152 if ((rc = mdb_page_touch(csrc)) ||
6153 (rc = mdb_page_touch(cdst)))
6156 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6157 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6158 key.mv_size = csrc->mc_db->md_pad;
6159 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6161 data.mv_data = NULL;
6165 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6166 assert(!((long)srcnode&1));
6167 srcpg = NODEPGNO(srcnode);
6168 flags = srcnode->mn_flags;
6169 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6170 unsigned int snum = csrc->mc_snum;
6172 /* must find the lowest key below src */
6173 mdb_page_search_lowest(csrc);
6174 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6175 key.mv_size = csrc->mc_db->md_pad;
6176 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6178 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6179 key.mv_size = NODEKSZ(s2);
6180 key.mv_data = NODEKEY(s2);
6182 csrc->mc_snum = snum--;
6183 csrc->mc_top = snum;
6185 key.mv_size = NODEKSZ(srcnode);
6186 key.mv_data = NODEKEY(srcnode);
6188 data.mv_size = NODEDSZ(srcnode);
6189 data.mv_data = NODEDATA(srcnode);
6191 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6192 unsigned int snum = cdst->mc_snum;
6195 /* must find the lowest key below dst */
6196 mdb_page_search_lowest(cdst);
6197 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6198 bkey.mv_size = cdst->mc_db->md_pad;
6199 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6201 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6202 bkey.mv_size = NODEKSZ(s2);
6203 bkey.mv_data = NODEKEY(s2);
6205 cdst->mc_snum = snum--;
6206 cdst->mc_top = snum;
6207 mdb_cursor_copy(cdst, &mn);
6209 rc = mdb_update_key(&mn, &bkey);
6214 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6215 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6216 csrc->mc_ki[csrc->mc_top],
6218 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6219 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6221 /* Add the node to the destination page.
6223 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6224 if (rc != MDB_SUCCESS)
6227 /* Delete the node from the source page.
6229 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6232 /* Adjust other cursors pointing to mp */
6233 MDB_cursor *m2, *m3;
6234 MDB_dbi dbi = csrc->mc_dbi;
6235 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6237 if (csrc->mc_flags & C_SUB)
6240 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6241 if (csrc->mc_flags & C_SUB)
6242 m3 = &m2->mc_xcursor->mx_cursor;
6245 if (m3 == csrc) continue;
6246 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6247 csrc->mc_ki[csrc->mc_top]) {
6248 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6249 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6254 /* Update the parent separators.
6256 if (csrc->mc_ki[csrc->mc_top] == 0) {
6257 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6258 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6259 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6261 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6262 key.mv_size = NODEKSZ(srcnode);
6263 key.mv_data = NODEKEY(srcnode);
6265 DPRINTF("update separator for source page %zu to [%s]",
6266 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6267 mdb_cursor_copy(csrc, &mn);
6270 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6273 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6275 indx_t ix = csrc->mc_ki[csrc->mc_top];
6276 nullkey.mv_size = 0;
6277 csrc->mc_ki[csrc->mc_top] = 0;
6278 rc = mdb_update_key(csrc, &nullkey);
6279 csrc->mc_ki[csrc->mc_top] = ix;
6280 assert(rc == MDB_SUCCESS);
6284 if (cdst->mc_ki[cdst->mc_top] == 0) {
6285 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6286 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6287 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6289 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6290 key.mv_size = NODEKSZ(srcnode);
6291 key.mv_data = NODEKEY(srcnode);
6293 DPRINTF("update separator for destination page %zu to [%s]",
6294 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6295 mdb_cursor_copy(cdst, &mn);
6298 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6301 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6303 indx_t ix = cdst->mc_ki[cdst->mc_top];
6304 nullkey.mv_size = 0;
6305 cdst->mc_ki[cdst->mc_top] = 0;
6306 rc = mdb_update_key(cdst, &nullkey);
6307 cdst->mc_ki[cdst->mc_top] = ix;
6308 assert(rc == MDB_SUCCESS);
6315 /** Merge one page into another.
6316 * The nodes from the page pointed to by \b csrc will
6317 * be copied to the page pointed to by \b cdst and then
6318 * the \b csrc page will be freed.
6319 * @param[in] csrc Cursor pointing to the source page.
6320 * @param[in] cdst Cursor pointing to the destination page.
6323 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6331 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6332 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6334 assert(csrc->mc_snum > 1); /* can't merge root page */
6335 assert(cdst->mc_snum > 1);
6337 /* Mark dst as dirty. */
6338 if ((rc = mdb_page_touch(cdst)))
6341 /* Move all nodes from src to dst.
6343 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6344 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6345 key.mv_size = csrc->mc_db->md_pad;
6346 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6347 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6348 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6349 if (rc != MDB_SUCCESS)
6351 key.mv_data = (char *)key.mv_data + key.mv_size;
6354 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6355 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6356 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6357 unsigned int snum = csrc->mc_snum;
6359 /* must find the lowest key below src */
6360 mdb_page_search_lowest(csrc);
6361 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6362 key.mv_size = csrc->mc_db->md_pad;
6363 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6365 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6366 key.mv_size = NODEKSZ(s2);
6367 key.mv_data = NODEKEY(s2);
6369 csrc->mc_snum = snum--;
6370 csrc->mc_top = snum;
6372 key.mv_size = srcnode->mn_ksize;
6373 key.mv_data = NODEKEY(srcnode);
6376 data.mv_size = NODEDSZ(srcnode);
6377 data.mv_data = NODEDATA(srcnode);
6378 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6379 if (rc != MDB_SUCCESS)
6384 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6385 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);
6387 /* Unlink the src page from parent and add to free list.
6389 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6390 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6393 rc = mdb_update_key(csrc, &key);
6399 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6400 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6401 csrc->mc_db->md_leaf_pages--;
6403 csrc->mc_db->md_branch_pages--;
6405 /* Adjust other cursors pointing to mp */
6406 MDB_cursor *m2, *m3;
6407 MDB_dbi dbi = csrc->mc_dbi;
6408 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6410 if (csrc->mc_flags & C_SUB)
6413 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6414 if (csrc->mc_flags & C_SUB)
6415 m3 = &m2->mc_xcursor->mx_cursor;
6418 if (m3 == csrc) continue;
6419 if (m3->mc_snum < csrc->mc_snum) continue;
6420 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6421 m3->mc_pg[csrc->mc_top] = mp;
6422 m3->mc_ki[csrc->mc_top] += nkeys;
6426 mdb_cursor_pop(csrc);
6428 return mdb_rebalance(csrc);
6431 /** Copy the contents of a cursor.
6432 * @param[in] csrc The cursor to copy from.
6433 * @param[out] cdst The cursor to copy to.
6436 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6440 cdst->mc_txn = csrc->mc_txn;
6441 cdst->mc_dbi = csrc->mc_dbi;
6442 cdst->mc_db = csrc->mc_db;
6443 cdst->mc_dbx = csrc->mc_dbx;
6444 cdst->mc_snum = csrc->mc_snum;
6445 cdst->mc_top = csrc->mc_top;
6446 cdst->mc_flags = csrc->mc_flags;
6448 for (i=0; i<csrc->mc_snum; i++) {
6449 cdst->mc_pg[i] = csrc->mc_pg[i];
6450 cdst->mc_ki[i] = csrc->mc_ki[i];
6454 /** Rebalance the tree after a delete operation.
6455 * @param[in] mc Cursor pointing to the page where rebalancing
6457 * @return 0 on success, non-zero on failure.
6460 mdb_rebalance(MDB_cursor *mc)
6464 unsigned int ptop, minkeys;
6467 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6471 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6472 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6473 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6474 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6478 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6479 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6482 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6483 DPRINTF("no need to rebalance page %zu, above fill threshold",
6489 if (mc->mc_snum < 2) {
6490 MDB_page *mp = mc->mc_pg[0];
6492 DPUTS("Can't rebalance a subpage, ignoring");
6495 if (NUMKEYS(mp) == 0) {
6496 DPUTS("tree is completely empty");
6497 mc->mc_db->md_root = P_INVALID;
6498 mc->mc_db->md_depth = 0;
6499 mc->mc_db->md_leaf_pages = 0;
6500 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6501 /* Adjust cursors pointing to mp */
6505 MDB_cursor *m2, *m3;
6506 MDB_dbi dbi = mc->mc_dbi;
6508 if (mc->mc_flags & C_SUB)
6511 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6512 if (mc->mc_flags & C_SUB)
6513 m3 = &m2->mc_xcursor->mx_cursor;
6516 if (m3->mc_snum < mc->mc_snum) continue;
6517 if (m3->mc_pg[0] == mp) {
6523 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6524 DPUTS("collapsing root page!");
6525 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6526 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6527 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6530 mc->mc_db->md_depth--;
6531 mc->mc_db->md_branch_pages--;
6533 /* Adjust other cursors pointing to mp */
6534 MDB_cursor *m2, *m3;
6535 MDB_dbi dbi = mc->mc_dbi;
6537 if (mc->mc_flags & C_SUB)
6540 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6541 if (mc->mc_flags & C_SUB)
6542 m3 = &m2->mc_xcursor->mx_cursor;
6545 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6546 if (m3->mc_pg[0] == mp) {
6547 m3->mc_pg[0] = mc->mc_pg[0];
6554 DPUTS("root page doesn't need rebalancing");
6558 /* The parent (branch page) must have at least 2 pointers,
6559 * otherwise the tree is invalid.
6561 ptop = mc->mc_top-1;
6562 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6564 /* Leaf page fill factor is below the threshold.
6565 * Try to move keys from left or right neighbor, or
6566 * merge with a neighbor page.
6571 mdb_cursor_copy(mc, &mn);
6572 mn.mc_xcursor = NULL;
6574 if (mc->mc_ki[ptop] == 0) {
6575 /* We're the leftmost leaf in our parent.
6577 DPUTS("reading right neighbor");
6579 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6580 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6583 mn.mc_ki[mn.mc_top] = 0;
6584 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6586 /* There is at least one neighbor to the left.
6588 DPUTS("reading left neighbor");
6590 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6591 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6594 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6595 mc->mc_ki[mc->mc_top] = 0;
6598 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6599 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);
6601 /* If the neighbor page is above threshold and has enough keys,
6602 * move one key from it. Otherwise we should try to merge them.
6603 * (A branch page must never have less than 2 keys.)
6605 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6606 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6607 return mdb_node_move(&mn, mc);
6609 if (mc->mc_ki[ptop] == 0)
6610 rc = mdb_page_merge(&mn, mc);
6612 rc = mdb_page_merge(mc, &mn);
6613 mc->mc_flags &= ~C_INITIALIZED;
6618 /** Complete a delete operation started by #mdb_cursor_del(). */
6620 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6624 /* add overflow pages to free list */
6625 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6629 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6630 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
6632 assert(IS_OVERFLOW(omp));
6633 mdb_ovpage_free(mc, omp);
6635 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6636 mc->mc_db->md_entries--;
6637 rc = mdb_rebalance(mc);
6638 if (rc != MDB_SUCCESS)
6639 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6640 /* if mc points past last node in page, invalidate */
6641 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6642 mc->mc_flags &= ~C_INITIALIZED;
6648 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6649 MDB_val *key, MDB_val *data)
6654 MDB_val rdata, *xdata;
6658 assert(key != NULL);
6660 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6662 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6665 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6669 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6673 mdb_cursor_init(&mc, txn, dbi, &mx);
6684 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6686 /* let mdb_page_split know about this cursor if needed:
6687 * delete will trigger a rebalance; if it needs to move
6688 * a node from one page to another, it will have to
6689 * update the parent's separator key(s). If the new sepkey
6690 * is larger than the current one, the parent page may
6691 * run out of space, triggering a split. We need this
6692 * cursor to be consistent until the end of the rebalance.
6694 mc.mc_next = txn->mt_cursors[dbi];
6695 txn->mt_cursors[dbi] = &mc;
6696 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6697 txn->mt_cursors[dbi] = mc.mc_next;
6702 /** Split a page and insert a new node.
6703 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6704 * The cursor will be updated to point to the actual page and index where
6705 * the node got inserted after the split.
6706 * @param[in] newkey The key for the newly inserted node.
6707 * @param[in] newdata The data for the newly inserted node.
6708 * @param[in] newpgno The page number, if the new node is a branch node.
6709 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6710 * @return 0 on success, non-zero on failure.
6713 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6714 unsigned int nflags)
6717 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6720 unsigned int i, j, split_indx, nkeys, pmax;
6722 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6724 MDB_page *mp, *rp, *pp;
6729 mp = mc->mc_pg[mc->mc_top];
6730 newindx = mc->mc_ki[mc->mc_top];
6732 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6733 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6734 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6736 /* Create a right sibling. */
6737 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6739 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6741 if (mc->mc_snum < 2) {
6742 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6744 /* shift current top to make room for new parent */
6745 mc->mc_pg[1] = mc->mc_pg[0];
6746 mc->mc_ki[1] = mc->mc_ki[0];
6749 mc->mc_db->md_root = pp->mp_pgno;
6750 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6751 mc->mc_db->md_depth++;
6754 /* Add left (implicit) pointer. */
6755 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6756 /* undo the pre-push */
6757 mc->mc_pg[0] = mc->mc_pg[1];
6758 mc->mc_ki[0] = mc->mc_ki[1];
6759 mc->mc_db->md_root = mp->mp_pgno;
6760 mc->mc_db->md_depth--;
6767 ptop = mc->mc_top-1;
6768 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6771 mc->mc_flags |= C_SPLITTING;
6772 mdb_cursor_copy(mc, &mn);
6773 mn.mc_pg[mn.mc_top] = rp;
6774 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6776 if (nflags & MDB_APPEND) {
6777 mn.mc_ki[mn.mc_top] = 0;
6779 split_indx = newindx;
6784 nkeys = NUMKEYS(mp);
6785 split_indx = nkeys / 2;
6786 if (newindx < split_indx)
6792 unsigned int lsize, rsize, ksize;
6793 /* Move half of the keys to the right sibling */
6795 x = mc->mc_ki[mc->mc_top] - split_indx;
6796 ksize = mc->mc_db->md_pad;
6797 split = LEAF2KEY(mp, split_indx, ksize);
6798 rsize = (nkeys - split_indx) * ksize;
6799 lsize = (nkeys - split_indx) * sizeof(indx_t);
6800 mp->mp_lower -= lsize;
6801 rp->mp_lower += lsize;
6802 mp->mp_upper += rsize - lsize;
6803 rp->mp_upper -= rsize - lsize;
6804 sepkey.mv_size = ksize;
6805 if (newindx == split_indx) {
6806 sepkey.mv_data = newkey->mv_data;
6808 sepkey.mv_data = split;
6811 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6812 memcpy(rp->mp_ptrs, split, rsize);
6813 sepkey.mv_data = rp->mp_ptrs;
6814 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6815 memcpy(ins, newkey->mv_data, ksize);
6816 mp->mp_lower += sizeof(indx_t);
6817 mp->mp_upper -= ksize - sizeof(indx_t);
6820 memcpy(rp->mp_ptrs, split, x * ksize);
6821 ins = LEAF2KEY(rp, x, ksize);
6822 memcpy(ins, newkey->mv_data, ksize);
6823 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6824 rp->mp_lower += sizeof(indx_t);
6825 rp->mp_upper -= ksize - sizeof(indx_t);
6826 mc->mc_ki[mc->mc_top] = x;
6827 mc->mc_pg[mc->mc_top] = rp;
6832 /* For leaf pages, check the split point based on what
6833 * fits where, since otherwise mdb_node_add can fail.
6835 * This check is only needed when the data items are
6836 * relatively large, such that being off by one will
6837 * make the difference between success or failure.
6839 * It's also relevant if a page happens to be laid out
6840 * such that one half of its nodes are all "small" and
6841 * the other half of its nodes are "large." If the new
6842 * item is also "large" and falls on the half with
6843 * "large" nodes, it also may not fit.
6846 unsigned int psize, nsize;
6847 /* Maximum free space in an empty page */
6848 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6849 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6850 if ((nkeys < 20) || (nsize > pmax/16)) {
6851 if (newindx <= split_indx) {
6854 for (i=0; i<split_indx; i++) {
6855 node = NODEPTR(mp, i);
6856 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6857 if (F_ISSET(node->mn_flags, F_BIGDATA))
6858 psize += sizeof(pgno_t);
6860 psize += NODEDSZ(node);
6864 split_indx = newindx;
6875 for (i=nkeys-1; i>=split_indx; i--) {
6876 node = NODEPTR(mp, i);
6877 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6878 if (F_ISSET(node->mn_flags, F_BIGDATA))
6879 psize += sizeof(pgno_t);
6881 psize += NODEDSZ(node);
6885 split_indx = newindx;
6896 /* First find the separating key between the split pages.
6897 * The case where newindx == split_indx is ambiguous; the
6898 * new item could go to the new page or stay on the original
6899 * page. If newpos == 1 it goes to the new page.
6901 if (newindx == split_indx && newpos) {
6902 sepkey.mv_size = newkey->mv_size;
6903 sepkey.mv_data = newkey->mv_data;
6905 node = NODEPTR(mp, split_indx);
6906 sepkey.mv_size = node->mn_ksize;
6907 sepkey.mv_data = NODEKEY(node);
6911 DPRINTF("separator is [%s]", DKEY(&sepkey));
6913 /* Copy separator key to the parent.
6915 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6919 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6922 if (mn.mc_snum == mc->mc_snum) {
6923 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6924 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6925 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6926 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6931 /* Right page might now have changed parent.
6932 * Check if left page also changed parent.
6934 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6935 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6936 for (i=0; i<ptop; i++) {
6937 mc->mc_pg[i] = mn.mc_pg[i];
6938 mc->mc_ki[i] = mn.mc_ki[i];
6940 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6941 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6945 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6948 mc->mc_flags ^= C_SPLITTING;
6949 if (rc != MDB_SUCCESS) {
6952 if (nflags & MDB_APPEND) {
6953 mc->mc_pg[mc->mc_top] = rp;
6954 mc->mc_ki[mc->mc_top] = 0;
6955 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6958 for (i=0; i<mc->mc_top; i++)
6959 mc->mc_ki[i] = mn.mc_ki[i];
6966 /* Move half of the keys to the right sibling. */
6968 /* grab a page to hold a temporary copy */
6969 copy = mdb_page_malloc(mc, 1);
6973 copy->mp_pgno = mp->mp_pgno;
6974 copy->mp_flags = mp->mp_flags;
6975 copy->mp_lower = PAGEHDRSZ;
6976 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6977 mc->mc_pg[mc->mc_top] = copy;
6978 for (i = j = 0; i <= nkeys; j++) {
6979 if (i == split_indx) {
6980 /* Insert in right sibling. */
6981 /* Reset insert index for right sibling. */
6982 if (i != newindx || (newpos ^ ins_new)) {
6984 mc->mc_pg[mc->mc_top] = rp;
6988 if (i == newindx && !ins_new) {
6989 /* Insert the original entry that caused the split. */
6990 rkey.mv_data = newkey->mv_data;
6991 rkey.mv_size = newkey->mv_size;
7000 /* Update index for the new key. */
7001 mc->mc_ki[mc->mc_top] = j;
7002 } else if (i == nkeys) {
7005 node = NODEPTR(mp, i);
7006 rkey.mv_data = NODEKEY(node);
7007 rkey.mv_size = node->mn_ksize;
7009 xdata.mv_data = NODEDATA(node);
7010 xdata.mv_size = NODEDSZ(node);
7013 pgno = NODEPGNO(node);
7014 flags = node->mn_flags;
7019 if (!IS_LEAF(mp) && j == 0) {
7020 /* First branch index doesn't need key data. */
7024 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7028 nkeys = NUMKEYS(copy);
7029 for (i=0; i<nkeys; i++)
7030 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7031 mp->mp_lower = copy->mp_lower;
7032 mp->mp_upper = copy->mp_upper;
7033 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7034 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
7036 /* reset back to original page */
7037 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
7038 mc->mc_pg[mc->mc_top] = mp;
7039 if (nflags & MDB_RESERVE) {
7040 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7041 if (!(node->mn_flags & F_BIGDATA))
7042 newdata->mv_data = NODEDATA(node);
7046 /* Make sure mc_ki is still valid.
7048 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7049 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7050 for (i=0; i<ptop; i++) {
7051 mc->mc_pg[i] = mn.mc_pg[i];
7052 mc->mc_ki[i] = mn.mc_ki[i];
7054 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7055 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7059 /* return tmp page to freelist */
7060 mdb_page_free(mc->mc_txn->mt_env, copy);
7063 /* Adjust other cursors pointing to mp */
7064 MDB_cursor *m2, *m3;
7065 MDB_dbi dbi = mc->mc_dbi;
7066 int fixup = NUMKEYS(mp);
7068 if (mc->mc_flags & C_SUB)
7071 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7072 if (mc->mc_flags & C_SUB)
7073 m3 = &m2->mc_xcursor->mx_cursor;
7078 if (!(m3->mc_flags & C_INITIALIZED))
7080 if (m3->mc_flags & C_SPLITTING)
7085 for (k=m3->mc_top; k>=0; k--) {
7086 m3->mc_ki[k+1] = m3->mc_ki[k];
7087 m3->mc_pg[k+1] = m3->mc_pg[k];
7089 if (m3->mc_ki[0] >= split_indx) {
7094 m3->mc_pg[0] = mc->mc_pg[0];
7098 if (m3->mc_pg[mc->mc_top] == mp) {
7099 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7100 m3->mc_ki[mc->mc_top]++;
7101 if (m3->mc_ki[mc->mc_top] >= fixup) {
7102 m3->mc_pg[mc->mc_top] = rp;
7103 m3->mc_ki[mc->mc_top] -= fixup;
7104 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7106 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7107 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7116 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7117 MDB_val *key, MDB_val *data, unsigned int flags)
7122 assert(key != NULL);
7123 assert(data != NULL);
7125 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7128 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7132 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7136 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7139 mdb_cursor_init(&mc, txn, dbi, &mx);
7140 return mdb_cursor_put(&mc, key, data, flags);
7144 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7146 if ((flag & CHANGEABLE) != flag)
7149 env->me_flags |= flag;
7151 env->me_flags &= ~flag;
7156 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7161 *arg = env->me_flags;
7166 mdb_env_get_path(MDB_env *env, const char **arg)
7171 *arg = env->me_path;
7175 /** Common code for #mdb_stat() and #mdb_env_stat().
7176 * @param[in] env the environment to operate in.
7177 * @param[in] db the #MDB_db record containing the stats to return.
7178 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7179 * @return 0, this function always succeeds.
7182 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7184 arg->ms_psize = env->me_psize;
7185 arg->ms_depth = db->md_depth;
7186 arg->ms_branch_pages = db->md_branch_pages;
7187 arg->ms_leaf_pages = db->md_leaf_pages;
7188 arg->ms_overflow_pages = db->md_overflow_pages;
7189 arg->ms_entries = db->md_entries;
7194 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7198 if (env == NULL || arg == NULL)
7201 toggle = mdb_env_pick_meta(env);
7203 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7207 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7211 if (env == NULL || arg == NULL)
7214 toggle = mdb_env_pick_meta(env);
7215 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7216 arg->me_mapsize = env->me_mapsize;
7217 arg->me_maxreaders = env->me_maxreaders;
7218 arg->me_numreaders = env->me_numreaders;
7219 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7220 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7224 /** Set the default comparison functions for a database.
7225 * Called immediately after a database is opened to set the defaults.
7226 * The user can then override them with #mdb_set_compare() or
7227 * #mdb_set_dupsort().
7228 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7229 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7232 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7234 uint16_t f = txn->mt_dbs[dbi].md_flags;
7236 txn->mt_dbxs[dbi].md_cmp =
7237 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7238 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7240 txn->mt_dbxs[dbi].md_dcmp =
7241 !(f & MDB_DUPSORT) ? 0 :
7242 ((f & MDB_INTEGERDUP)
7243 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7244 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7247 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7252 int rc, dbflag, exact;
7253 unsigned int unused = 0;
7256 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7257 mdb_default_cmp(txn, FREE_DBI);
7260 if ((flags & VALID_FLAGS) != flags)
7266 if (flags & PERSISTENT_FLAGS) {
7267 uint16_t f2 = flags & PERSISTENT_FLAGS;
7268 /* make sure flag changes get committed */
7269 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7270 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7271 txn->mt_flags |= MDB_TXN_DIRTY;
7274 mdb_default_cmp(txn, MAIN_DBI);
7278 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7279 mdb_default_cmp(txn, MAIN_DBI);
7282 /* Is the DB already open? */
7284 for (i=2; i<txn->mt_numdbs; i++) {
7285 if (!txn->mt_dbxs[i].md_name.mv_size) {
7286 /* Remember this free slot */
7287 if (!unused) unused = i;
7290 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7291 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7297 /* If no free slot and max hit, fail */
7298 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7299 return MDB_DBS_FULL;
7301 /* Cannot mix named databases with some mainDB flags */
7302 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
7303 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
7305 /* Find the DB info */
7306 dbflag = DB_NEW|DB_VALID;
7309 key.mv_data = (void *)name;
7310 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7311 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7312 if (rc == MDB_SUCCESS) {
7313 /* make sure this is actually a DB */
7314 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7315 if (!(node->mn_flags & F_SUBDATA))
7317 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7318 /* Create if requested */
7320 data.mv_size = sizeof(MDB_db);
7321 data.mv_data = &dummy;
7322 memset(&dummy, 0, sizeof(dummy));
7323 dummy.md_root = P_INVALID;
7324 dummy.md_flags = flags & PERSISTENT_FLAGS;
7325 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7329 /* OK, got info, add to table */
7330 if (rc == MDB_SUCCESS) {
7331 unsigned int slot = unused ? unused : txn->mt_numdbs;
7332 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7333 txn->mt_dbxs[slot].md_name.mv_size = len;
7334 txn->mt_dbxs[slot].md_rel = NULL;
7335 txn->mt_dbflags[slot] = dbflag;
7336 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7338 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7339 mdb_default_cmp(txn, slot);
7348 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7350 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7353 if (txn->mt_dbflags[dbi] & DB_STALE) {
7356 /* Stale, must read the DB's root. cursor_init does it for us. */
7357 mdb_cursor_init(&mc, txn, dbi, &mx);
7359 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7362 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7365 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7367 ptr = env->me_dbxs[dbi].md_name.mv_data;
7368 env->me_dbxs[dbi].md_name.mv_data = NULL;
7369 env->me_dbxs[dbi].md_name.mv_size = 0;
7370 env->me_dbflags[dbi] = 0;
7374 /** Add all the DB's pages to the free list.
7375 * @param[in] mc Cursor on the DB to free.
7376 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7377 * @return 0 on success, non-zero on failure.
7380 mdb_drop0(MDB_cursor *mc, int subs)
7384 rc = mdb_page_search(mc, NULL, 0);
7385 if (rc == MDB_SUCCESS) {
7390 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7391 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7394 mdb_cursor_copy(mc, &mx);
7395 while (mc->mc_snum > 0) {
7396 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7397 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7398 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7399 if (ni->mn_flags & F_BIGDATA) {
7403 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7404 rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL);
7407 assert(IS_OVERFLOW(omp));
7408 ovpages = omp->mp_pages;
7409 for (j=0; j<ovpages; j++) {
7410 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7413 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7414 mdb_xcursor_init1(mc, ni);
7415 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7421 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7423 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7426 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7431 mc->mc_ki[mc->mc_top] = i;
7432 rc = mdb_cursor_sibling(mc, 1);
7434 /* no more siblings, go back to beginning
7435 * of previous level.
7439 for (i=1; i<mc->mc_snum; i++) {
7441 mc->mc_pg[i] = mx.mc_pg[i];
7446 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7447 mc->mc_db->md_root);
7452 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7454 MDB_cursor *mc, *m2;
7457 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7460 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7463 rc = mdb_cursor_open(txn, dbi, &mc);
7467 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7468 /* Invalidate the dropped DB's cursors */
7469 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
7470 m2->mc_flags &= ~C_INITIALIZED;
7474 /* Can't delete the main DB */
7475 if (del && dbi > MAIN_DBI) {
7476 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7478 txn->mt_dbflags[dbi] = DB_STALE;
7479 mdb_dbi_close(txn->mt_env, dbi);
7482 /* reset the DB record, mark it dirty */
7483 txn->mt_dbflags[dbi] |= DB_DIRTY;
7484 txn->mt_dbs[dbi].md_depth = 0;
7485 txn->mt_dbs[dbi].md_branch_pages = 0;
7486 txn->mt_dbs[dbi].md_leaf_pages = 0;
7487 txn->mt_dbs[dbi].md_overflow_pages = 0;
7488 txn->mt_dbs[dbi].md_entries = 0;
7489 txn->mt_dbs[dbi].md_root = P_INVALID;
7491 txn->mt_flags |= MDB_TXN_DIRTY;
7494 mdb_cursor_close(mc);
7498 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7500 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7503 txn->mt_dbxs[dbi].md_cmp = cmp;
7507 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7509 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7512 txn->mt_dbxs[dbi].md_dcmp = cmp;
7516 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7518 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7521 txn->mt_dbxs[dbi].md_rel = rel;
7525 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7527 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7530 txn->mt_dbxs[dbi].md_relctx = ctx;