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.
36 #include <sys/types.h>
38 #include <sys/param.h>
44 #ifdef HAVE_SYS_FILE_H
61 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
62 #include <netinet/in.h>
63 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
66 #if defined(__APPLE__) || defined (BSD)
67 # define MDB_USE_POSIX_SEM 1
68 # define MDB_FDATASYNC fsync
69 #elif defined(ANDROID)
70 # define MDB_FDATASYNC fsync
75 #ifdef MDB_USE_POSIX_SEM
76 #include <semaphore.h>
81 #include <valgrind/memcheck.h>
82 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
83 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
84 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
85 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
86 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
88 #define VGMEMP_CREATE(h,r,z)
89 #define VGMEMP_ALLOC(h,a,s)
90 #define VGMEMP_FREE(h,a)
91 #define VGMEMP_DESTROY(h)
92 #define VGMEMP_DEFINED(a,s)
96 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
97 /* Solaris just defines one or the other */
98 # define LITTLE_ENDIAN 1234
99 # define BIG_ENDIAN 4321
100 # ifdef _LITTLE_ENDIAN
101 # define BYTE_ORDER LITTLE_ENDIAN
103 # define BYTE_ORDER BIG_ENDIAN
106 # define BYTE_ORDER __BYTE_ORDER
110 #ifndef LITTLE_ENDIAN
111 #define LITTLE_ENDIAN __LITTLE_ENDIAN
114 #define BIG_ENDIAN __BIG_ENDIAN
117 #if defined(__i386) || defined(__x86_64)
118 #define MISALIGNED_OK 1
124 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
125 # error "Unknown or unsupported endianness (BYTE_ORDER)"
126 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
127 # error "Two's complement, reasonably sized integer types, please"
130 /** @defgroup internal MDB Internals
133 /** @defgroup compat Windows Compatibility Macros
134 * A bunch of macros to minimize the amount of platform-specific ifdefs
135 * needed throughout the rest of the code. When the features this library
136 * needs are similar enough to POSIX to be hidden in a one-or-two line
137 * replacement, this macro approach is used.
141 #define pthread_t DWORD
142 #define pthread_mutex_t HANDLE
143 #define pthread_key_t DWORD
144 #define pthread_self() GetCurrentThreadId()
145 #define pthread_key_create(x,y) \
146 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
147 #define pthread_key_delete(x) TlsFree(x)
148 #define pthread_getspecific(x) TlsGetValue(x)
149 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
150 #define pthread_mutex_unlock(x) ReleaseMutex(x)
151 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
152 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
153 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
154 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
155 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
156 #define getpid() GetCurrentProcessId()
157 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
158 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
159 #define ErrCode() GetLastError()
160 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
161 #define close(fd) CloseHandle(fd)
162 #define munmap(ptr,len) UnmapViewOfFile(ptr)
165 #ifdef MDB_USE_POSIX_SEM
167 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
168 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
169 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
170 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
173 mdb_sem_wait(sem_t *sem)
176 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
181 /** Lock the reader mutex.
183 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
184 /** Unlock the reader mutex.
186 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
188 /** Lock the writer mutex.
189 * Only a single write transaction is allowed at a time. Other writers
190 * will block waiting for this mutex.
192 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
193 /** Unlock the writer mutex.
195 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
196 #endif /* MDB_USE_POSIX_SEM */
198 /** Get the error code for the last failed system function.
200 #define ErrCode() errno
202 /** An abstraction for a file handle.
203 * On POSIX systems file handles are small integers. On Windows
204 * they're opaque pointers.
208 /** A value for an invalid file handle.
209 * Mainly used to initialize file variables and signify that they are
212 #define INVALID_HANDLE_VALUE (-1)
214 /** Get the size of a memory page for the system.
215 * This is the basic size that the platform's memory manager uses, and is
216 * fundamental to the use of memory-mapped files.
218 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
221 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
224 #define MNAME_LEN (sizeof(pthread_mutex_t))
230 /** A flag for opening a file and requesting synchronous data writes.
231 * This is only used when writing a meta page. It's not strictly needed;
232 * we could just do a normal write and then immediately perform a flush.
233 * But if this flag is available it saves us an extra system call.
235 * @note If O_DSYNC is undefined but exists in /usr/include,
236 * preferably set some compiler flag to get the definition.
237 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
240 # define MDB_DSYNC O_DSYNC
244 /** Function for flushing the data of a file. Define this to fsync
245 * if fdatasync() is not supported.
247 #ifndef MDB_FDATASYNC
248 # define MDB_FDATASYNC fdatasync
252 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
263 /** A page number in the database.
264 * Note that 64 bit page numbers are overkill, since pages themselves
265 * already represent 12-13 bits of addressable memory, and the OS will
266 * always limit applications to a maximum of 63 bits of address space.
268 * @note In the #MDB_node structure, we only store 48 bits of this value,
269 * which thus limits us to only 60 bits of addressable data.
271 typedef MDB_ID pgno_t;
273 /** A transaction ID.
274 * See struct MDB_txn.mt_txnid for details.
276 typedef MDB_ID txnid_t;
278 /** @defgroup debug Debug Macros
282 /** Enable debug output.
283 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
284 * read from and written to the database (used for free space management).
289 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
292 # define DPRINTF (void) /* Vararg macros may be unsupported */
294 static int mdb_debug;
295 static txnid_t mdb_debug_start;
297 /** Print a debug message with printf formatting. */
298 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
299 ((void) ((mdb_debug) && \
300 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
302 # define DPRINTF(fmt, ...) ((void) 0)
303 # define MDB_DEBUG_SKIP
305 /** Print a debug string.
306 * The string is printed literally, with no format processing.
308 #define DPUTS(arg) DPRINTF("%s", arg)
311 /** A default memory page size.
312 * The actual size is platform-dependent, but we use this for
313 * boot-strapping. We probably should not be using this any more.
314 * The #GET_PAGESIZE() macro is used to get the actual size.
316 * Note that we don't currently support Huge pages. On Linux,
317 * regular data files cannot use Huge pages, and in general
318 * Huge pages aren't actually pageable. We rely on the OS
319 * demand-pager to read our data and page it out when memory
320 * pressure from other processes is high. So until OSs have
321 * actual paging support for Huge pages, they're not viable.
323 #define MDB_PAGESIZE 4096
325 /** The minimum number of keys required in a database page.
326 * Setting this to a larger value will place a smaller bound on the
327 * maximum size of a data item. Data items larger than this size will
328 * be pushed into overflow pages instead of being stored directly in
329 * the B-tree node. This value used to default to 4. With a page size
330 * of 4096 bytes that meant that any item larger than 1024 bytes would
331 * go into an overflow page. That also meant that on average 2-3KB of
332 * each overflow page was wasted space. The value cannot be lower than
333 * 2 because then there would no longer be a tree structure. With this
334 * value, items larger than 2KB will go into overflow pages, and on
335 * average only 1KB will be wasted.
337 #define MDB_MINKEYS 2
339 /** A stamp that identifies a file as an MDB file.
340 * There's nothing special about this value other than that it is easily
341 * recognizable, and it will reflect any byte order mismatches.
343 #define MDB_MAGIC 0xBEEFC0DE
345 /** The version number for a database's file format. */
346 #define MDB_VERSION 1
348 /** @brief The maximum size of a key in the database.
350 * We require that keys all fit onto a regular page. This limit
351 * could be raised a bit further if needed; to something just
352 * under #MDB_PAGESIZE / #MDB_MINKEYS.
354 * Note that data items in an #MDB_DUPSORT database are actually keys
355 * of a subDB, so they're also limited to this size.
357 #ifndef MDB_MAXKEYSIZE
358 #define MDB_MAXKEYSIZE 511
361 /** @brief The maximum size of a data item.
363 * We only store a 32 bit value for node sizes.
365 #define MAXDATASIZE 0xffffffffUL
370 * This is used for printing a hex dump of a key's contents.
372 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
373 /** Display a key in hex.
375 * Invoke a function to display a key in hex.
377 #define DKEY(x) mdb_dkey(x, kbuf)
379 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
383 /** An invalid page number.
384 * Mainly used to denote an empty tree.
386 #define P_INVALID (~(pgno_t)0)
388 /** Test if the flags \b f are set in a flag word \b w. */
389 #define F_ISSET(w, f) (((w) & (f)) == (f))
391 /** Used for offsets within a single page.
392 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
395 typedef uint16_t indx_t;
397 /** Default size of memory map.
398 * This is certainly too small for any actual applications. Apps should always set
399 * the size explicitly using #mdb_env_set_mapsize().
401 #define DEFAULT_MAPSIZE 1048576
403 /** @defgroup readers Reader Lock Table
404 * Readers don't acquire any locks for their data access. Instead, they
405 * simply record their transaction ID in the reader table. The reader
406 * mutex is needed just to find an empty slot in the reader table. The
407 * slot's address is saved in thread-specific data so that subsequent read
408 * transactions started by the same thread need no further locking to proceed.
410 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
412 * No reader table is used if the database is on a read-only filesystem.
414 * Since the database uses multi-version concurrency control, readers don't
415 * actually need any locking. This table is used to keep track of which
416 * readers are using data from which old transactions, so that we'll know
417 * when a particular old transaction is no longer in use. Old transactions
418 * that have discarded any data pages can then have those pages reclaimed
419 * for use by a later write transaction.
421 * The lock table is constructed such that reader slots are aligned with the
422 * processor's cache line size. Any slot is only ever used by one thread.
423 * This alignment guarantees that there will be no contention or cache
424 * thrashing as threads update their own slot info, and also eliminates
425 * any need for locking when accessing a slot.
427 * A writer thread will scan every slot in the table to determine the oldest
428 * outstanding reader transaction. Any freed pages older than this will be
429 * reclaimed by the writer. The writer doesn't use any locks when scanning
430 * this table. This means that there's no guarantee that the writer will
431 * see the most up-to-date reader info, but that's not required for correct
432 * operation - all we need is to know the upper bound on the oldest reader,
433 * we don't care at all about the newest reader. So the only consequence of
434 * reading stale information here is that old pages might hang around a
435 * while longer before being reclaimed. That's actually good anyway, because
436 * the longer we delay reclaiming old pages, the more likely it is that a
437 * string of contiguous pages can be found after coalescing old pages from
438 * many old transactions together.
441 /** Number of slots in the reader table.
442 * This value was chosen somewhat arbitrarily. 126 readers plus a
443 * couple mutexes fit exactly into 8KB on my development machine.
444 * Applications should set the table size using #mdb_env_set_maxreaders().
446 #define DEFAULT_READERS 126
448 /** The size of a CPU cache line in bytes. We want our lock structures
449 * aligned to this size to avoid false cache line sharing in the
451 * This value works for most CPUs. For Itanium this should be 128.
457 /** The information we store in a single slot of the reader table.
458 * In addition to a transaction ID, we also record the process and
459 * thread ID that owns a slot, so that we can detect stale information,
460 * e.g. threads or processes that went away without cleaning up.
461 * @note We currently don't check for stale records. We simply re-init
462 * the table when we know that we're the only process opening the
465 typedef struct MDB_rxbody {
466 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
467 * Multiple readers that start at the same time will probably have the
468 * same ID here. Again, it's not important to exclude them from
469 * anything; all we need to know is which version of the DB they
470 * started from so we can avoid overwriting any data used in that
471 * particular version.
474 /** The process ID of the process owning this reader txn. */
476 /** The thread ID of the thread owning this txn. */
480 /** The actual reader record, with cacheline padding. */
481 typedef struct MDB_reader {
484 /** shorthand for mrb_txnid */
485 #define mr_txnid mru.mrx.mrb_txnid
486 #define mr_pid mru.mrx.mrb_pid
487 #define mr_tid mru.mrx.mrb_tid
488 /** cache line alignment */
489 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
493 /** The header for the reader table.
494 * The table resides in a memory-mapped file. (This is a different file
495 * than is used for the main database.)
497 * For POSIX the actual mutexes reside in the shared memory of this
498 * mapped file. On Windows, mutexes are named objects allocated by the
499 * kernel; we store the mutex names in this mapped file so that other
500 * processes can grab them. This same approach is also used on
501 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
502 * process-shared POSIX mutexes. For these cases where a named object
503 * is used, the object name is derived from a 64 bit FNV hash of the
504 * environment pathname. As such, naming collisions are extremely
505 * unlikely. If a collision occurs, the results are unpredictable.
507 typedef struct MDB_txbody {
508 /** Stamp identifying this as an MDB file. It must be set
511 /** Version number of this lock file. Must be set to #MDB_VERSION. */
512 uint32_t mtb_version;
513 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
514 char mtb_rmname[MNAME_LEN];
516 /** Mutex protecting access to this table.
517 * This is the reader lock that #LOCK_MUTEX_R acquires.
519 pthread_mutex_t mtb_mutex;
521 /** The ID of the last transaction committed to the database.
522 * This is recorded here only for convenience; the value can always
523 * be determined by reading the main database meta pages.
526 /** The number of slots that have been used in the reader table.
527 * This always records the maximum count, it is not decremented
528 * when readers release their slots.
530 unsigned mtb_numreaders;
533 /** The actual reader table definition. */
534 typedef struct MDB_txninfo {
537 #define mti_magic mt1.mtb.mtb_magic
538 #define mti_version mt1.mtb.mtb_version
539 #define mti_mutex mt1.mtb.mtb_mutex
540 #define mti_rmname mt1.mtb.mtb_rmname
541 #define mti_txnid mt1.mtb.mtb_txnid
542 #define mti_numreaders mt1.mtb.mtb_numreaders
543 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
546 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
547 char mt2_wmname[MNAME_LEN];
548 #define mti_wmname mt2.mt2_wmname
550 pthread_mutex_t mt2_wmutex;
551 #define mti_wmutex mt2.mt2_wmutex
553 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
555 MDB_reader mti_readers[1];
559 /** Common header for all page types.
560 * Overflow records occupy a number of contiguous pages with no
561 * headers on any page after the first.
563 typedef struct MDB_page {
564 #define mp_pgno mp_p.p_pgno
565 #define mp_next mp_p.p_next
567 pgno_t p_pgno; /**< page number */
568 void * p_next; /**< for in-memory list of freed structs */
571 /** @defgroup mdb_page Page Flags
573 * Flags for the page headers.
576 #define P_BRANCH 0x01 /**< branch page */
577 #define P_LEAF 0x02 /**< leaf page */
578 #define P_OVERFLOW 0x04 /**< overflow page */
579 #define P_META 0x08 /**< meta page */
580 #define P_DIRTY 0x10 /**< dirty page */
581 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
582 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
584 uint16_t mp_flags; /**< @ref mdb_page */
585 #define mp_lower mp_pb.pb.pb_lower
586 #define mp_upper mp_pb.pb.pb_upper
587 #define mp_pages mp_pb.pb_pages
590 indx_t pb_lower; /**< lower bound of free space */
591 indx_t pb_upper; /**< upper bound of free space */
593 uint32_t pb_pages; /**< number of overflow pages */
595 indx_t mp_ptrs[1]; /**< dynamic size */
598 /** Size of the page header, excluding dynamic data at the end */
599 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
601 /** Address of first usable data byte in a page, after the header */
602 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
604 /** Number of nodes on a page */
605 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
607 /** The amount of space remaining in the page */
608 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
610 /** The percentage of space used in the page, in tenths of a percent. */
611 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
612 ((env)->me_psize - PAGEHDRSZ))
613 /** The minimum page fill factor, in tenths of a percent.
614 * Pages emptier than this are candidates for merging.
616 #define FILL_THRESHOLD 250
618 /** Test if a page is a leaf page */
619 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
620 /** Test if a page is a LEAF2 page */
621 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
622 /** Test if a page is a branch page */
623 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
624 /** Test if a page is an overflow page */
625 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
626 /** Test if a page is a sub page */
627 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
629 /** The number of overflow pages needed to store the given size. */
630 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
632 /** Header for a single key/data pair within a page.
633 * We guarantee 2-byte alignment for nodes.
635 typedef struct MDB_node {
636 /** lo and hi are used for data size on leaf nodes and for
637 * child pgno on branch nodes. On 64 bit platforms, flags
638 * is also used for pgno. (Branch nodes have no flags).
639 * They are in host byte order in case that lets some
640 * accesses be optimized into a 32-bit word access.
642 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
643 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
644 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
645 /** @defgroup mdb_node Node Flags
647 * Flags for node headers.
650 #define F_BIGDATA 0x01 /**< data put on overflow page */
651 #define F_SUBDATA 0x02 /**< data is a sub-database */
652 #define F_DUPDATA 0x04 /**< data has duplicates */
654 /** valid flags for #mdb_node_add() */
655 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
658 unsigned short mn_flags; /**< @ref mdb_node */
659 unsigned short mn_ksize; /**< key size */
660 char mn_data[1]; /**< key and data are appended here */
663 /** Size of the node header, excluding dynamic data at the end */
664 #define NODESIZE offsetof(MDB_node, mn_data)
666 /** Bit position of top word in page number, for shifting mn_flags */
667 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
669 /** Size of a node in a branch page with a given key.
670 * This is just the node header plus the key, there is no data.
672 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
674 /** Size of a node in a leaf page with a given key and data.
675 * This is node header plus key plus data size.
677 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
679 /** Address of node \b i in page \b p */
680 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
682 /** Address of the key for the node */
683 #define NODEKEY(node) (void *)((node)->mn_data)
685 /** Address of the data for a node */
686 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
688 /** Get the page number pointed to by a branch node */
689 #define NODEPGNO(node) \
690 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
691 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
692 /** Set the page number in a branch node */
693 #define SETPGNO(node,pgno) do { \
694 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
695 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
697 /** Get the size of the data in a leaf node */
698 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
699 /** Set the size of the data for a leaf node */
700 #define SETDSZ(node,size) do { \
701 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
702 /** The size of a key in a node */
703 #define NODEKSZ(node) ((node)->mn_ksize)
705 /** Copy a page number from src to dst */
707 #define COPY_PGNO(dst,src) dst = src
709 #if SIZE_MAX > 4294967295UL
710 #define COPY_PGNO(dst,src) do { \
711 unsigned short *s, *d; \
712 s = (unsigned short *)&(src); \
713 d = (unsigned short *)&(dst); \
720 #define COPY_PGNO(dst,src) do { \
721 unsigned short *s, *d; \
722 s = (unsigned short *)&(src); \
723 d = (unsigned short *)&(dst); \
729 /** The address of a key in a LEAF2 page.
730 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
731 * There are no node headers, keys are stored contiguously.
733 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
735 /** Set the \b node's key into \b key, if requested. */
736 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
737 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
739 /** Information about a single database in the environment. */
740 typedef struct MDB_db {
741 uint32_t md_pad; /**< also ksize for LEAF2 pages */
742 uint16_t md_flags; /**< @ref mdb_dbi_open */
743 uint16_t md_depth; /**< depth of this tree */
744 pgno_t md_branch_pages; /**< number of internal pages */
745 pgno_t md_leaf_pages; /**< number of leaf pages */
746 pgno_t md_overflow_pages; /**< number of overflow pages */
747 size_t md_entries; /**< number of data items */
748 pgno_t md_root; /**< the root page of this tree */
751 /** mdb_dbi_open flags */
752 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
753 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
754 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
755 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
757 /** Handle for the DB used to track free pages. */
759 /** Handle for the default DB. */
762 /** Meta page content. */
763 typedef struct MDB_meta {
764 /** Stamp identifying this as an MDB file. It must be set
767 /** Version number of this lock file. Must be set to #MDB_VERSION. */
769 void *mm_address; /**< address for fixed mapping */
770 size_t mm_mapsize; /**< size of mmap region */
771 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
772 /** The size of pages used in this DB */
773 #define mm_psize mm_dbs[0].md_pad
774 /** Any persistent environment flags. @ref mdb_env */
775 #define mm_flags mm_dbs[0].md_flags
776 pgno_t mm_last_pg; /**< last used page in file */
777 txnid_t mm_txnid; /**< txnid that committed this page */
780 /** Buffer for a stack-allocated dirty page.
781 * The members define size and alignment, and silence type
782 * aliasing warnings. They are not used directly; that could
783 * mean incorrectly using several union members in parallel.
785 typedef union MDB_pagebuf {
786 char mb_raw[MDB_PAGESIZE];
789 char mm_pad[PAGEHDRSZ];
794 /** Auxiliary DB info.
795 * The information here is mostly static/read-only. There is
796 * only a single copy of this record in the environment.
798 typedef struct MDB_dbx {
799 MDB_val md_name; /**< name of the database */
800 MDB_cmp_func *md_cmp; /**< function for comparing keys */
801 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
802 MDB_rel_func *md_rel; /**< user relocate function */
803 void *md_relctx; /**< user-provided context for md_rel */
806 /** A database transaction.
807 * Every operation requires a transaction handle.
810 MDB_txn *mt_parent; /**< parent of a nested txn */
811 MDB_txn *mt_child; /**< nested txn under this txn */
812 pgno_t mt_next_pgno; /**< next unallocated page */
813 /** The ID of this transaction. IDs are integers incrementing from 1.
814 * Only committed write transactions increment the ID. If a transaction
815 * aborts, the ID may be re-used by the next writer.
818 MDB_env *mt_env; /**< the DB environment */
819 /** The list of pages that became unused during this transaction.
823 MDB_ID2L dirty_list; /**< for write txns: modified pages */
824 MDB_reader *reader; /**< this thread's reader table slot or NULL */
826 /** Array of records for each DB known in the environment. */
828 /** Array of MDB_db records for each known DB */
830 /** @defgroup mt_dbflag Transaction DB Flags
834 #define DB_DIRTY 0x01 /**< DB was written in this txn */
835 #define DB_STALE 0x02 /**< DB record is older than txnID */
836 #define DB_NEW 0x04 /**< DB handle opened in this txn */
837 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
839 /** In write txns, array of cursors for each DB */
840 MDB_cursor **mt_cursors;
841 /** Array of flags for each DB */
842 unsigned char *mt_dbflags;
843 /** Number of DB records in use. This number only ever increments;
844 * we don't decrement it when individual DB handles are closed.
848 /** @defgroup mdb_txn Transaction Flags
852 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
853 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
854 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
856 unsigned int mt_flags; /**< @ref mdb_txn */
857 /** dirty_list maxsize - # of allocated pages allowed, including in parent txns */
858 unsigned int mt_dirty_room;
859 /** Tracks which of the two meta pages was used at the start
860 * of this transaction.
862 unsigned int mt_toggle;
865 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
866 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
867 * raise this on a 64 bit machine.
869 #define CURSOR_STACK 32
873 /** Cursors are used for all DB operations */
875 /** Next cursor on this DB in this txn */
877 /** Original cursor if this is a shadow */
879 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
880 struct MDB_xcursor *mc_xcursor;
881 /** The transaction that owns this cursor */
883 /** The database handle this cursor operates on */
885 /** The database record for this cursor */
887 /** The database auxiliary record for this cursor */
889 /** The @ref mt_dbflag for this database */
890 unsigned char *mc_dbflag;
891 unsigned short mc_snum; /**< number of pushed pages */
892 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
893 /** @defgroup mdb_cursor Cursor Flags
895 * Cursor state flags.
898 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
899 #define C_EOF 0x02 /**< No more data */
900 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
901 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
902 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
903 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
904 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
906 unsigned int mc_flags; /**< @ref mdb_cursor */
907 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
908 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
911 /** Context for sorted-dup records.
912 * We could have gone to a fully recursive design, with arbitrarily
913 * deep nesting of sub-databases. But for now we only handle these
914 * levels - main DB, optional sub-DB, sorted-duplicate DB.
916 typedef struct MDB_xcursor {
917 /** A sub-cursor for traversing the Dup DB */
918 MDB_cursor mx_cursor;
919 /** The database record for this Dup DB */
921 /** The auxiliary DB record for this Dup DB */
923 /** The @ref mt_dbflag for this Dup DB */
924 unsigned char mx_dbflag;
927 /** State of FreeDB old pages, stored in the MDB_env */
928 typedef struct MDB_pgstate {
929 txnid_t mf_pglast; /**< ID of last old page record we used */
930 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
931 pgno_t *mf_pgfree; /**< memory to free when dropping me_pghead */
934 /** The database environment. */
936 HANDLE me_fd; /**< The main data file */
937 HANDLE me_lfd; /**< The lock file */
938 HANDLE me_mfd; /**< just for writing the meta pages */
939 /** Failed to update the meta page. Probably an I/O error. */
940 #define MDB_FATAL_ERROR 0x80000000U
941 /** Some fields are initialized. */
942 #define MDB_ENV_ACTIVE 0x20000000U
943 /** me_txkey is set */
944 #define MDB_ENV_TXKEY 0x10000000U
945 uint32_t me_flags; /**< @ref mdb_env */
946 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
947 unsigned int me_maxreaders; /**< size of the reader table */
948 unsigned int me_numreaders; /**< max numreaders set by this env */
949 MDB_dbi me_numdbs; /**< number of DBs opened */
950 MDB_dbi me_maxdbs; /**< size of the DB table */
951 pid_t me_pid; /**< process ID of this env */
952 char *me_path; /**< path to the DB files */
953 char *me_map; /**< the memory map of the data file */
954 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
955 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
956 MDB_txn *me_txn; /**< current write transaction */
957 size_t me_mapsize; /**< size of the data memory map */
958 off_t me_size; /**< current file size */
959 pgno_t me_maxpg; /**< me_mapsize / me_psize */
960 MDB_dbx *me_dbxs; /**< array of static DB info */
961 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
962 pthread_key_t me_txkey; /**< thread-key for readers */
963 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
964 # define me_pglast me_pgstate.mf_pglast
965 # define me_pghead me_pgstate.mf_pghead
966 # define me_pgfree me_pgstate.mf_pgfree
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 */
973 unsigned int me_maxfree_1pg;
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 /** Allocate pages for writing.
1293 * If there are free pages available from older transactions, they
1294 * will be re-used first. Otherwise a new page will be allocated.
1295 * @param[in] mc cursor A cursor handle identifying the transaction and
1296 * database for which we are allocating.
1297 * @param[in] num the number of pages to allocate.
1298 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1299 * will always be satisfied by a single contiguous chunk of memory.
1300 * @return 0 on success, non-zero on failure.
1303 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1305 MDB_txn *txn = mc->mc_txn;
1307 pgno_t pgno = P_INVALID;
1309 txnid_t oldest = 0, last;
1314 /* If our dirty list is already full, we can't do anything */
1315 if (txn->mt_dirty_room == 0)
1316 return MDB_TXN_FULL;
1318 /* The free list won't have any content at all until txn 2 has
1319 * committed. The pages freed by txn 2 will be unreferenced
1320 * after txn 3 commits, and so will be safe to re-use in txn 4.
1322 if (txn->mt_txnid > 3) {
1323 if (!txn->mt_env->me_pghead &&
1324 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1325 /* See if there's anything in the free DB */
1332 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1333 if (!txn->mt_env->me_pglast) {
1334 mdb_page_search(&m2, NULL, 0);
1335 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1336 kptr = (txnid_t *)NODEKEY(leaf);
1341 last = txn->mt_env->me_pglast + 1;
1343 key.mv_data = &last;
1344 key.mv_size = sizeof(last);
1345 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1348 last = *(txnid_t *)key.mv_data;
1354 oldest = txn->mt_txnid - 1;
1355 nr = txn->mt_env->me_txns->mti_numreaders;
1356 r = txn->mt_env->me_txns->mti_readers;
1357 for (i=0; i<nr; i++) {
1358 if (!r[i].mr_pid) continue;
1365 if (oldest > last) {
1366 /* It's usable, grab it.
1370 if (!txn->mt_env->me_pglast) {
1371 mdb_node_read(txn, leaf, &data);
1373 idl = (MDB_ID *) data.mv_data;
1374 /* We might have a zero-length IDL due to freelist growth
1375 * during a prior commit
1378 txn->mt_env->me_pglast = last;
1381 mop = malloc(MDB_IDL_SIZEOF(idl));
1384 txn->mt_env->me_pglast = last;
1385 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1386 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1391 DPRINTF("IDL read txn %zu root %zu num %zu",
1392 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1393 for (i=0; i<idl[0]; i++) {
1394 DPRINTF("IDL %zu", idl[i+1]);
1401 if (txn->mt_env->me_pghead) {
1402 pgno_t *mop = txn->mt_env->me_pghead;
1405 int retry = 1, readit = 0, n2 = num-1;
1406 unsigned int i, j, k;
1408 /* If current list is too short, must fetch more and coalesce */
1409 if (mop[0] < (unsigned)num)
1412 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1414 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1415 /* If on freelist, don't try to read more. If what we have
1416 * right now isn't enough just use new pages.
1417 * TODO: get all of this working. Many circular dependencies...
1419 if (mc->mc_dbi == FREE_DBI) {
1428 last = txn->mt_env->me_pglast + 1;
1430 /* We haven't hit the readers list yet? */
1436 oldest = txn->mt_txnid - 1;
1437 nr = txn->mt_env->me_txns->mti_numreaders;
1438 r = txn->mt_env->me_txns->mti_readers;
1439 for (i=0; i<nr; i++) {
1440 if (!r[i].mr_pid) continue;
1447 /* There's nothing we can use on the freelist */
1448 if (oldest - last < 1)
1451 key.mv_data = &last;
1452 key.mv_size = sizeof(last);
1453 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1455 if (rc == MDB_NOTFOUND)
1459 last = *(txnid_t*)key.mv_data;
1462 idl = (MDB_ID *) data.mv_data;
1463 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1466 /* merge in sorted order */
1467 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1469 while (i>0 || j>0) {
1470 if (i && idl[i] < mop[j])
1471 mop2[k--] = idl[i--];
1473 mop2[k--] = mop[j--];
1475 txn->mt_env->me_pglast = last;
1476 free(txn->mt_env->me_pgfree);
1477 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1479 /* Keep trying to read until we have enough */
1480 if (mop[0] < (unsigned)num) {
1485 /* current list has enough pages, but are they contiguous? */
1486 for (i=mop[0]; i>=(unsigned)num; i--) {
1487 if (mop[i-n2] == mop[i] + n2) {
1490 /* move any stragglers down */
1491 for (j=i+num; j<=mop[0]; j++)
1498 /* Stop if we succeeded, or no retries */
1499 if (!retry || pgno != P_INVALID)
1505 /* peel pages off tail, so we only have to truncate the list */
1506 pgno = MDB_IDL_LAST(mop);
1509 if (MDB_IDL_IS_ZERO(mop)) {
1510 free(txn->mt_env->me_pgfree);
1511 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1516 if (pgno == P_INVALID) {
1517 /* DB size is maxed out */
1518 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1519 DPUTS("DB size maxed out");
1520 return MDB_MAP_FULL;
1523 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1524 if (pgno == P_INVALID) {
1525 pgno = txn->mt_next_pgno;
1526 txn->mt_next_pgno += num;
1528 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1531 if (!(np = mdb_page_malloc(mc, num)))
1533 if (pgno == P_INVALID) {
1534 np->mp_pgno = txn->mt_next_pgno;
1535 txn->mt_next_pgno += num;
1540 mid.mid = np->mp_pgno;
1542 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1543 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1545 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1547 txn->mt_dirty_room--;
1553 /** Copy a page: avoid copying unused portions of the page.
1554 * @param[in] dst page to copy into
1555 * @param[in] src page to copy from
1556 * @param[in] psize size of a page
1559 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1561 dst->mp_flags = src->mp_flags | P_DIRTY;
1562 dst->mp_pages = src->mp_pages;
1564 if (IS_LEAF2(src)) {
1565 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1567 unsigned int i, nkeys = NUMKEYS(src);
1568 for (i=0; i<nkeys; i++)
1569 dst->mp_ptrs[i] = src->mp_ptrs[i];
1570 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1571 psize - src->mp_upper);
1575 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1576 * @param[in] mc cursor pointing to the page to be touched
1577 * @return 0 on success, non-zero on failure.
1580 mdb_page_touch(MDB_cursor *mc)
1582 MDB_page *mp = mc->mc_pg[mc->mc_top];
1586 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1588 if ((rc = mdb_page_alloc(mc, 1, &np)))
1590 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1591 assert(mp->mp_pgno != np->mp_pgno);
1592 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1594 /* If page isn't full, just copy the used portion */
1595 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1598 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1600 np->mp_flags |= P_DIRTY;
1605 /* Adjust other cursors pointing to mp */
1606 if (mc->mc_flags & C_SUB) {
1607 MDB_cursor *m2, *m3;
1608 MDB_dbi dbi = mc->mc_dbi-1;
1610 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1611 if (m2 == mc) continue;
1612 m3 = &m2->mc_xcursor->mx_cursor;
1613 if (m3->mc_snum < mc->mc_snum) continue;
1614 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1615 m3->mc_pg[mc->mc_top] = mp;
1621 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1622 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1623 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1624 m2->mc_pg[mc->mc_top] = mp;
1628 mc->mc_pg[mc->mc_top] = mp;
1629 /** If this page has a parent, update the parent to point to
1633 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1635 mc->mc_db->md_root = mp->mp_pgno;
1636 } else if (mc->mc_txn->mt_parent) {
1639 /* If txn has a parent, make sure the page is in our
1642 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1643 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1644 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1645 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1646 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1647 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1648 mc->mc_pg[mc->mc_top] = mp;
1653 assert(mc->mc_txn->mt_u.dirty_list[0].mid < MDB_IDL_UM_MAX);
1655 np = mdb_page_malloc(mc, 1);
1658 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1659 mid.mid = np->mp_pgno;
1661 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1669 mdb_env_sync(MDB_env *env, int force)
1672 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1673 if (env->me_flags & MDB_WRITEMAP) {
1674 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1675 ? MS_ASYNC : MS_SYNC;
1676 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1679 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1683 if (MDB_FDATASYNC(env->me_fd))
1690 /** Make shadow copies of all of parent txn's cursors */
1692 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1694 MDB_cursor *mc, *m2;
1695 unsigned int i, j, size;
1697 for (i=0;i<src->mt_numdbs; i++) {
1698 if (src->mt_cursors[i]) {
1699 size = sizeof(MDB_cursor);
1700 if (src->mt_cursors[i]->mc_xcursor)
1701 size += sizeof(MDB_xcursor);
1702 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1709 mc->mc_db = &dst->mt_dbs[i];
1710 mc->mc_dbx = m2->mc_dbx;
1711 mc->mc_dbflag = &dst->mt_dbflags[i];
1712 mc->mc_snum = m2->mc_snum;
1713 mc->mc_top = m2->mc_top;
1714 mc->mc_flags = m2->mc_flags | (C_SHADOW|C_ALLOCD);
1715 for (j=0; j<mc->mc_snum; j++) {
1716 mc->mc_pg[j] = m2->mc_pg[j];
1717 mc->mc_ki[j] = m2->mc_ki[j];
1719 if (m2->mc_xcursor) {
1720 MDB_xcursor *mx, *mx2;
1721 mx = (MDB_xcursor *)(mc+1);
1722 mc->mc_xcursor = mx;
1723 mx2 = m2->mc_xcursor;
1724 mx->mx_db = mx2->mx_db;
1725 mx->mx_dbx = mx2->mx_dbx;
1726 mx->mx_dbflag = mx2->mx_dbflag;
1727 mx->mx_cursor.mc_txn = dst;
1728 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1729 mx->mx_cursor.mc_db = &mx->mx_db;
1730 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1731 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1732 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1733 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1734 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1735 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1736 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1737 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1740 mc->mc_xcursor = NULL;
1742 mc->mc_next = dst->mt_cursors[i];
1743 dst->mt_cursors[i] = mc;
1750 /** Close this write txn's cursors, after optionally merging its shadow
1751 * cursors back into parent's.
1752 * @param[in] txn the transaction handle.
1753 * @param[in] merge 0 to not merge cursors, C_SHADOW to merge.
1754 * @return 0 on success, non-zero on failure.
1757 mdb_cursors_close(MDB_txn *txn, unsigned merge)
1759 MDB_cursor **cursors = txn->mt_cursors, *mc, *next;
1762 for (i = txn->mt_numdbs; --i >= 0; ) {
1763 for (mc = cursors[i]; mc; mc = next) {
1765 if (mc->mc_flags & merge) {
1766 MDB_cursor *m2 = mc->mc_orig;
1767 m2->mc_snum = mc->mc_snum;
1768 m2->mc_top = mc->mc_top;
1769 for (j = mc->mc_snum; --j >= 0; ) {
1770 m2->mc_pg[j] = mc->mc_pg[j];
1771 m2->mc_ki[j] = mc->mc_ki[j];
1774 if (mc->mc_flags & C_ALLOCD)
1782 mdb_txn_reset0(MDB_txn *txn);
1784 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1785 * @param[in] txn the transaction handle to initialize
1786 * @return 0 on success, non-zero on failure.
1789 mdb_txn_renew0(MDB_txn *txn)
1791 MDB_env *env = txn->mt_env;
1794 int rc, new_notls = 0;
1797 txn->mt_numdbs = env->me_numdbs;
1798 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1800 if (txn->mt_flags & MDB_TXN_RDONLY) {
1801 if (!env->me_txns) {
1802 i = mdb_env_pick_meta(env);
1803 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1804 txn->mt_u.reader = NULL;
1806 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1807 pthread_getspecific(env->me_txkey);
1809 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1810 return MDB_BAD_RSLOT;
1812 pid_t pid = env->me_pid;
1813 pthread_t tid = pthread_self();
1816 for (i=0; i<env->me_txns->mti_numreaders; i++)
1817 if (env->me_txns->mti_readers[i].mr_pid == 0)
1819 if (i == env->me_maxreaders) {
1820 UNLOCK_MUTEX_R(env);
1821 return MDB_READERS_FULL;
1823 env->me_txns->mti_readers[i].mr_pid = pid;
1824 env->me_txns->mti_readers[i].mr_tid = tid;
1825 if (i >= env->me_txns->mti_numreaders)
1826 env->me_txns->mti_numreaders = i+1;
1827 /* Save numreaders for un-mutexed mdb_env_close() */
1828 env->me_numreaders = env->me_txns->mti_numreaders;
1829 UNLOCK_MUTEX_R(env);
1830 r = &env->me_txns->mti_readers[i];
1831 new_notls = (env->me_flags & MDB_NOTLS);
1832 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
1837 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1838 txn->mt_u.reader = r;
1840 txn->mt_toggle = txn->mt_txnid & 1;
1841 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1845 txn->mt_txnid = env->me_txns->mti_txnid;
1846 txn->mt_toggle = txn->mt_txnid & 1;
1847 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1850 if (txn->mt_txnid == mdb_debug_start)
1853 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1854 txn->mt_u.dirty_list = env->me_dirty_list;
1855 txn->mt_u.dirty_list[0].mid = 0;
1856 txn->mt_free_pgs = env->me_free_pgs;
1857 txn->mt_free_pgs[0] = 0;
1861 /* Copy the DB info and flags */
1862 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1863 for (i=2; i<txn->mt_numdbs; i++) {
1864 x = env->me_dbflags[i];
1865 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1866 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1868 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1870 if (env->me_maxpg < txn->mt_next_pgno) {
1871 mdb_txn_reset0(txn);
1873 txn->mt_u.reader->mr_pid = 0;
1874 txn->mt_u.reader = NULL;
1876 return MDB_MAP_RESIZED;
1883 mdb_txn_renew(MDB_txn *txn)
1887 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
1890 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1891 DPUTS("environment had fatal error, must shutdown!");
1895 rc = mdb_txn_renew0(txn);
1896 if (rc == MDB_SUCCESS) {
1897 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1898 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1899 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1905 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1909 int rc, size, tsize = sizeof(MDB_txn);
1911 if (env->me_flags & MDB_FATAL_ERROR) {
1912 DPUTS("environment had fatal error, must shutdown!");
1915 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1918 /* Nested transactions: Max 1 child, write txns only, no writemap */
1919 if (parent->mt_child ||
1920 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1921 (env->me_flags & MDB_WRITEMAP))
1925 tsize = sizeof(MDB_ntxn);
1927 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1928 if (!(flags & MDB_RDONLY))
1929 size += env->me_maxdbs * sizeof(MDB_cursor *);
1931 if ((txn = calloc(1, size)) == NULL) {
1932 DPRINTF("calloc: %s", strerror(ErrCode()));
1935 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1936 if (flags & MDB_RDONLY) {
1937 txn->mt_flags |= MDB_TXN_RDONLY;
1938 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1940 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1941 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1947 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1948 if (!txn->mt_u.dirty_list ||
1949 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
1951 free(txn->mt_u.dirty_list);
1955 txn->mt_txnid = parent->mt_txnid;
1956 txn->mt_toggle = parent->mt_toggle;
1957 txn->mt_dirty_room = parent->mt_dirty_room;
1958 txn->mt_u.dirty_list[0].mid = 0;
1959 txn->mt_free_pgs[0] = 0;
1960 txn->mt_next_pgno = parent->mt_next_pgno;
1961 parent->mt_child = txn;
1962 txn->mt_parent = parent;
1963 txn->mt_numdbs = parent->mt_numdbs;
1964 txn->mt_dbxs = parent->mt_dbxs;
1965 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1966 /* Copy parent's mt_dbflags, but clear DB_NEW */
1967 for (i=0; i<txn->mt_numdbs; i++)
1968 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1970 ntxn = (MDB_ntxn *)txn;
1971 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1972 if (env->me_pghead) {
1973 size = MDB_IDL_SIZEOF(env->me_pghead);
1974 env->me_pghead = malloc(size);
1976 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1980 env->me_pgfree = env->me_pghead;
1982 rc = mdb_cursor_shadow(parent, txn);
1984 mdb_txn_reset0(txn);
1986 rc = mdb_txn_renew0(txn);
1992 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1993 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1994 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
2000 /** Export or close DBI handles opened in this txn. */
2002 mdb_dbis_update(MDB_txn *txn, int keep)
2005 MDB_dbi n = txn->mt_numdbs;
2006 MDB_env *env = txn->mt_env;
2007 unsigned char *tdbflags = txn->mt_dbflags;
2009 for (i = n; --i >= 2;) {
2010 if (tdbflags[i] & DB_NEW) {
2012 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2014 char *ptr = env->me_dbxs[i].md_name.mv_data;
2015 env->me_dbxs[i].md_name.mv_data = NULL;
2016 env->me_dbxs[i].md_name.mv_size = 0;
2017 env->me_dbflags[i] = 0;
2022 if (keep && env->me_numdbs < n)
2026 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2027 * May be called twice for readonly txns: First reset it, then abort.
2028 * @param[in] txn the transaction handle to reset
2031 mdb_txn_reset0(MDB_txn *txn)
2033 MDB_env *env = txn->mt_env;
2036 /* Close any DBI handles opened in this txn */
2037 mdb_dbis_update(txn, 0);
2039 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2040 if (txn->mt_u.reader) {
2041 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2042 if (!(env->me_flags & MDB_NOTLS))
2043 txn->mt_u.reader = NULL; /* txn does not own reader */
2045 txn->mt_numdbs = 0; /* close nothing if called again */
2046 txn->mt_dbxs = NULL; /* mark txn as reset */
2050 mdb_cursors_close(txn, 0);
2052 if (!(env->me_flags & MDB_WRITEMAP)) {
2053 /* return all dirty pages to dpage list */
2054 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2055 dp = txn->mt_u.dirty_list[i].mptr;
2056 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2057 mdb_page_free(txn->mt_env, dp);
2059 /* large pages just get freed directly */
2060 VGMEMP_FREE(txn->mt_env, dp);
2066 free(env->me_pgfree);
2068 if (txn->mt_parent) {
2069 txn->mt_parent->mt_child = NULL;
2070 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2071 mdb_midl_free(txn->mt_free_pgs);
2072 free(txn->mt_u.dirty_list);
2075 if (mdb_midl_shrink(&txn->mt_free_pgs))
2076 env->me_free_pgs = txn->mt_free_pgs;
2079 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
2080 txn->mt_env->me_pglast = 0;
2083 /* The writer mutex was locked in mdb_txn_begin. */
2084 UNLOCK_MUTEX_W(env);
2089 mdb_txn_reset(MDB_txn *txn)
2094 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2095 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2096 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2098 /* This call is only valid for read-only txns */
2099 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2102 mdb_txn_reset0(txn);
2106 mdb_txn_abort(MDB_txn *txn)
2111 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2112 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2113 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2116 mdb_txn_abort(txn->mt_child);
2118 mdb_txn_reset0(txn);
2119 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2120 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2121 txn->mt_u.reader->mr_pid = 0;
2127 mdb_txn_commit(MDB_txn *txn)
2135 pgno_t next, freecnt;
2136 txnid_t oldpg_txnid, id;
2139 assert(txn != NULL);
2140 assert(txn->mt_env != NULL);
2142 if (txn->mt_child) {
2143 rc = mdb_txn_commit(txn->mt_child);
2144 txn->mt_child = NULL;
2153 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2154 mdb_dbis_update(txn, 1);
2155 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2160 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2161 DPUTS("error flag is set, can't commit");
2163 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2168 if (txn->mt_parent) {
2169 MDB_txn *parent = txn->mt_parent;
2173 /* Append our free list to parent's */
2174 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2178 mdb_midl_free(txn->mt_free_pgs);
2180 parent->mt_next_pgno = txn->mt_next_pgno;
2181 parent->mt_flags = txn->mt_flags;
2183 /* Merge our cursors into parent's and close them */
2184 mdb_cursors_close(txn, C_SHADOW);
2186 /* Update parent's DB table. */
2187 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2188 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2189 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2190 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2191 for (i=2; i<txn->mt_numdbs; i++) {
2192 /* preserve parent's DB_NEW status */
2193 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2194 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2197 dst = txn->mt_parent->mt_u.dirty_list;
2198 src = txn->mt_u.dirty_list;
2199 /* Find len = length of merging our dirty list with parent's */
2201 dst[0].mid = 0; /* simplify loops */
2202 if (parent->mt_parent) {
2203 len = x + src[0].mid;
2204 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2205 for (i = x; y && i; y--) {
2206 pgno_t yp = src[y].mid;
2207 while (yp < dst[i].mid)
2209 if (yp == dst[i].mid) {
2214 } else { /* Simplify the above for single-ancestor case */
2215 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2217 /* Merge our dirty list with parent's */
2219 for (i = len; y; dst[i--] = src[y--]) {
2220 pgno_t yp = src[y].mid;
2221 while (yp < dst[x].mid)
2222 dst[i--] = dst[x--];
2223 if (yp == dst[x].mid)
2224 free(dst[x--].mptr);
2228 free(txn->mt_u.dirty_list);
2229 parent->mt_dirty_room = txn->mt_dirty_room;
2231 txn->mt_parent->mt_child = NULL;
2232 free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pgfree);
2237 if (txn != env->me_txn) {
2238 DPUTS("attempt to commit unknown transaction");
2243 mdb_cursors_close(txn, 0);
2245 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2248 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2249 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2251 /* Update DB root pointers */
2252 if (txn->mt_numdbs > 2) {
2255 data.mv_size = sizeof(MDB_db);
2257 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2258 for (i = 2; i < txn->mt_numdbs; i++) {
2259 if (txn->mt_dbflags[i] & DB_DIRTY) {
2260 data.mv_data = &txn->mt_dbs[i];
2261 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2268 /* Save the freelist as of this transaction to the freeDB. This
2269 * can change the freelist, so keep trying until it stabilizes.
2271 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2272 * except the code below can decrease env->me_pglast to split pghead.
2273 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2274 * can only appear in env->me_pghead when env->me_pglast increases.
2275 * Until then, the me_pghead pointer won't move but can become NULL.
2278 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2279 oldpg_txnid = id = 0;
2282 /* should only be one record now */
2283 if (env->me_pghead || env->me_pglast) {
2284 /* make sure first page of freeDB is touched and on freelist */
2285 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2286 if (rc && rc != MDB_NOTFOUND) {
2293 /* Delete IDLs we used from the free list */
2294 if (env->me_pglast) {
2299 rc = mdb_cursor_first(&mc, &key, NULL);
2302 oldpg_txnid = *(txnid_t *)key.mv_data;
2304 assert(oldpg_txnid <= env->me_pglast);
2306 rc = mdb_cursor_del(&mc, 0);
2309 } while (oldpg_txnid < env->me_pglast);
2312 /* Save IDL of pages freed by this txn, to freeDB */
2314 if (freecnt != txn->mt_free_pgs[0]) {
2317 /* make sure last page of freeDB is touched and on freelist */
2318 key.mv_size = MDB_MAXKEYSIZE+1;
2320 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2321 if (rc && rc != MDB_NOTFOUND)
2327 MDB_IDL idl = txn->mt_free_pgs;
2328 mdb_midl_sort(txn->mt_free_pgs);
2329 DPRINTF("IDL write txn %zu root %zu num %zu",
2330 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2331 for (i=1; i<=idl[0]; i++) {
2332 DPRINTF("IDL %zu", idl[i]);
2336 /* write to last page of freeDB */
2337 key.mv_size = sizeof(pgno_t);
2338 key.mv_data = &txn->mt_txnid;
2339 /* The free list can still grow during this call,
2340 * despite the pre-emptive touches above. So retry
2341 * until the reserved space remains big enough.
2344 assert(freecnt < txn->mt_free_pgs[0]);
2345 freecnt = txn->mt_free_pgs[0];
2346 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2347 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2350 } while (freecnt != txn->mt_free_pgs[0]);
2351 mdb_midl_sort(txn->mt_free_pgs);
2352 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2353 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2354 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2357 /* Put back page numbers we took from freeDB but did not use */
2358 if (env->me_pghead) {
2363 mop = env->me_pghead;
2364 id = env->me_pglast;
2365 key.mv_size = sizeof(id);
2367 /* These steps may grow the freelist again
2368 * due to freed overflow pages...
2373 if (orig > env->me_maxfree_1pg && id > 4)
2374 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2375 data.mv_size = (orig + 1) * sizeof(pgno_t);
2376 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2379 assert(!env->me_pghead || env->me_pglast);
2380 /* mop could have been used again here */
2381 if (id != env->me_pglast || env->me_pghead == NULL)
2382 goto again; /* was completely used up */
2383 assert(mop == env->me_pghead);
2384 } while (mop[0] < orig && --i);
2385 memcpy(data.mv_data, mop, data.mv_size);
2388 *(pgno_t *)data.mv_data = orig;
2389 mop[orig] = mop[0] - orig;
2390 env->me_pghead = mop += orig;
2391 /* Save more oldpages at the previous txnid. */
2392 assert(env->me_pglast == id && id == oldpg_txnid);
2393 env->me_pglast = --oldpg_txnid;
2397 /* Check for growth of freelist again */
2398 if (freecnt != txn->mt_free_pgs[0])
2401 free(env->me_pgfree);
2402 env->me_pghead = env->me_pgfree = NULL;
2404 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2405 if (mdb_midl_shrink(&txn->mt_free_pgs))
2406 env->me_free_pgs = txn->mt_free_pgs;
2413 if (env->me_flags & MDB_WRITEMAP) {
2414 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2415 dp = txn->mt_u.dirty_list[i].mptr;
2416 /* clear dirty flag */
2417 dp->mp_flags &= ~P_DIRTY;
2418 txn->mt_u.dirty_list[i].mid = 0;
2420 txn->mt_u.dirty_list[0].mid = 0;
2424 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2430 /* Windows actually supports scatter/gather I/O, but only on
2431 * unbuffered file handles. Since we're relying on the OS page
2432 * cache for all our data, that's self-defeating. So we just
2433 * write pages one at a time. We use the ov structure to set
2434 * the write offset, to at least save the overhead of a Seek
2438 memset(&ov, 0, sizeof(ov));
2439 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2441 dp = txn->mt_u.dirty_list[i].mptr;
2442 DPRINTF("committing page %zu", dp->mp_pgno);
2443 size = dp->mp_pgno * env->me_psize;
2444 ov.Offset = size & 0xffffffff;
2445 ov.OffsetHigh = size >> 16;
2446 ov.OffsetHigh >>= 16;
2447 /* clear dirty flag */
2448 dp->mp_flags &= ~P_DIRTY;
2449 wsize = env->me_psize;
2450 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2451 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2454 DPRINTF("WriteFile: %d", n);
2461 struct iovec iov[MDB_COMMIT_PAGES];
2465 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2466 dp = txn->mt_u.dirty_list[i].mptr;
2467 if (dp->mp_pgno != next) {
2469 rc = writev(env->me_fd, iov, n);
2473 DPUTS("short write, filesystem full?");
2475 DPRINTF("writev: %s", strerror(n));
2482 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2485 DPRINTF("committing page %zu", dp->mp_pgno);
2486 iov[n].iov_len = env->me_psize;
2487 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2488 iov[n].iov_base = (char *)dp;
2489 size += iov[n].iov_len;
2490 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2491 /* clear dirty flag */
2492 dp->mp_flags &= ~P_DIRTY;
2493 if (++n >= MDB_COMMIT_PAGES) {
2503 rc = writev(env->me_fd, iov, n);
2507 DPUTS("short write, filesystem full?");
2509 DPRINTF("writev: %s", strerror(n));
2516 /* Drop the dirty pages.
2518 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2519 dp = txn->mt_u.dirty_list[i].mptr;
2520 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2521 mdb_page_free(txn->mt_env, dp);
2523 VGMEMP_FREE(txn->mt_env, dp);
2526 txn->mt_u.dirty_list[i].mid = 0;
2528 txn->mt_u.dirty_list[0].mid = 0;
2531 if ((n = mdb_env_sync(env, 0)) != 0 ||
2532 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2540 mdb_dbis_update(txn, 1);
2542 UNLOCK_MUTEX_W(env);
2548 /** Read the environment parameters of a DB environment before
2549 * mapping it into memory.
2550 * @param[in] env the environment handle
2551 * @param[out] meta address of where to store the meta information
2552 * @return 0 on success, non-zero on failure.
2555 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2562 /* We don't know the page size yet, so use a minimum value.
2563 * Read both meta pages so we can use the latest one.
2566 for (i=0; i<2; i++) {
2568 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2570 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2575 else if (rc != MDB_PAGESIZE) {
2579 DPRINTF("read: %s", strerror(err));
2583 p = (MDB_page *)&pbuf;
2585 if (!F_ISSET(p->mp_flags, P_META)) {
2586 DPRINTF("page %zu not a meta page", p->mp_pgno);
2591 if (m->mm_magic != MDB_MAGIC) {
2592 DPUTS("meta has invalid magic");
2596 if (m->mm_version != MDB_VERSION) {
2597 DPRINTF("database is version %u, expected version %u",
2598 m->mm_version, MDB_VERSION);
2599 return MDB_VERSION_MISMATCH;
2603 if (m->mm_txnid > meta->mm_txnid)
2604 memcpy(meta, m, sizeof(*m));
2606 memcpy(meta, m, sizeof(*m));
2608 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2610 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2618 /** Write the environment parameters of a freshly created DB environment.
2619 * @param[in] env the environment handle
2620 * @param[out] meta address of where to store the meta information
2621 * @return 0 on success, non-zero on failure.
2624 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2631 DPUTS("writing new meta page");
2633 GET_PAGESIZE(psize);
2635 meta->mm_magic = MDB_MAGIC;
2636 meta->mm_version = MDB_VERSION;
2637 meta->mm_mapsize = env->me_mapsize;
2638 meta->mm_psize = psize;
2639 meta->mm_last_pg = 1;
2640 meta->mm_flags = env->me_flags & 0xffff;
2641 meta->mm_flags |= MDB_INTEGERKEY;
2642 meta->mm_dbs[0].md_root = P_INVALID;
2643 meta->mm_dbs[1].md_root = P_INVALID;
2645 p = calloc(2, psize);
2647 p->mp_flags = P_META;
2650 memcpy(m, meta, sizeof(*meta));
2652 q = (MDB_page *)((char *)p + psize);
2655 q->mp_flags = P_META;
2658 memcpy(m, meta, sizeof(*meta));
2663 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2664 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2665 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2668 lseek(env->me_fd, 0, SEEK_SET);
2669 rc = write(env->me_fd, p, psize * 2);
2670 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2676 /** Update the environment info to commit a transaction.
2677 * @param[in] txn the transaction that's being committed
2678 * @return 0 on success, non-zero on failure.
2681 mdb_env_write_meta(MDB_txn *txn)
2684 MDB_meta meta, metab, *mp;
2686 int rc, len, toggle;
2693 assert(txn != NULL);
2694 assert(txn->mt_env != NULL);
2696 toggle = !txn->mt_toggle;
2697 DPRINTF("writing meta page %d for root page %zu",
2698 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2701 mp = env->me_metas[toggle];
2703 if (env->me_flags & MDB_WRITEMAP) {
2704 /* Persist any increases of mapsize config */
2705 if (env->me_mapsize > mp->mm_mapsize)
2706 mp->mm_mapsize = env->me_mapsize;
2707 mp->mm_dbs[0] = txn->mt_dbs[0];
2708 mp->mm_dbs[1] = txn->mt_dbs[1];
2709 mp->mm_last_pg = txn->mt_next_pgno - 1;
2710 mp->mm_txnid = txn->mt_txnid;
2711 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2712 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2715 ptr += env->me_psize;
2716 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2723 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2724 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2726 ptr = (char *)&meta;
2727 if (env->me_mapsize > mp->mm_mapsize) {
2728 /* Persist any increases of mapsize config */
2729 meta.mm_mapsize = env->me_mapsize;
2730 off = offsetof(MDB_meta, mm_mapsize);
2732 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2734 len = sizeof(MDB_meta) - off;
2737 meta.mm_dbs[0] = txn->mt_dbs[0];
2738 meta.mm_dbs[1] = txn->mt_dbs[1];
2739 meta.mm_last_pg = txn->mt_next_pgno - 1;
2740 meta.mm_txnid = txn->mt_txnid;
2743 off += env->me_psize;
2746 /* Write to the SYNC fd */
2747 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2748 env->me_fd : env->me_mfd;
2751 memset(&ov, 0, sizeof(ov));
2753 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2756 rc = pwrite(mfd, ptr, len, off);
2761 DPUTS("write failed, disk error?");
2762 /* On a failure, the pagecache still contains the new data.
2763 * Write some old data back, to prevent it from being used.
2764 * Use the non-SYNC fd; we know it will fail anyway.
2766 meta.mm_last_pg = metab.mm_last_pg;
2767 meta.mm_txnid = metab.mm_txnid;
2769 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2771 r2 = pwrite(env->me_fd, ptr, len, off);
2774 env->me_flags |= MDB_FATAL_ERROR;
2778 /* Memory ordering issues are irrelevant; since the entire writer
2779 * is wrapped by wmutex, all of these changes will become visible
2780 * after the wmutex is unlocked. Since the DB is multi-version,
2781 * readers will get consistent data regardless of how fresh or
2782 * how stale their view of these values is.
2784 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2789 /** Check both meta pages to see which one is newer.
2790 * @param[in] env the environment handle
2791 * @return meta toggle (0 or 1).
2794 mdb_env_pick_meta(const MDB_env *env)
2796 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2800 mdb_env_create(MDB_env **env)
2804 e = calloc(1, sizeof(MDB_env));
2808 e->me_maxreaders = DEFAULT_READERS;
2809 e->me_maxdbs = e->me_numdbs = 2;
2810 e->me_fd = INVALID_HANDLE_VALUE;
2811 e->me_lfd = INVALID_HANDLE_VALUE;
2812 e->me_mfd = INVALID_HANDLE_VALUE;
2813 #ifdef MDB_USE_POSIX_SEM
2814 e->me_rmutex = SEM_FAILED;
2815 e->me_wmutex = SEM_FAILED;
2817 e->me_pid = getpid();
2818 VGMEMP_CREATE(e,0,0);
2824 mdb_env_set_mapsize(MDB_env *env, size_t size)
2828 env->me_mapsize = size;
2830 env->me_maxpg = env->me_mapsize / env->me_psize;
2835 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2839 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2844 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2846 if (env->me_map || readers < 1)
2848 env->me_maxreaders = readers;
2853 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2855 if (!env || !readers)
2857 *readers = env->me_maxreaders;
2861 /** Further setup required for opening an MDB environment
2864 mdb_env_open2(MDB_env *env)
2866 unsigned int flags = env->me_flags;
2867 int i, newenv = 0, prot;
2871 memset(&meta, 0, sizeof(meta));
2873 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2876 DPUTS("new mdbenv");
2880 /* Was a mapsize configured? */
2881 if (!env->me_mapsize) {
2882 /* If this is a new environment, take the default,
2883 * else use the size recorded in the existing env.
2885 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2886 } else if (env->me_mapsize < meta.mm_mapsize) {
2887 /* If the configured size is smaller, make sure it's
2888 * still big enough. Silently round up to minimum if not.
2890 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2891 if (env->me_mapsize < minsize)
2892 env->me_mapsize = minsize;
2898 LONG sizelo, sizehi;
2899 sizelo = env->me_mapsize & 0xffffffff;
2900 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2902 /* Windows won't create mappings for zero length files.
2903 * Just allocate the maxsize right now.
2906 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2907 if (!SetEndOfFile(env->me_fd))
2909 SetFilePointer(env->me_fd, 0, NULL, 0);
2911 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2912 PAGE_READWRITE : PAGE_READONLY,
2913 sizehi, sizelo, NULL);
2916 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2917 FILE_MAP_WRITE : FILE_MAP_READ,
2918 0, 0, env->me_mapsize, meta.mm_address);
2926 if (flags & MDB_WRITEMAP) {
2928 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2931 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2933 if (env->me_map == MAP_FAILED) {
2937 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2939 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2941 #ifdef POSIX_MADV_RANDOM
2942 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2943 #endif /* POSIX_MADV_RANDOM */
2944 #endif /* MADV_RANDOM */
2948 if (flags & MDB_FIXEDMAP)
2949 meta.mm_address = env->me_map;
2950 i = mdb_env_init_meta(env, &meta);
2951 if (i != MDB_SUCCESS) {
2954 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2955 /* Can happen because the address argument to mmap() is just a
2956 * hint. mmap() can pick another, e.g. if the range is in use.
2957 * The MAP_FIXED flag would prevent that, but then mmap could
2958 * instead unmap existing pages to make room for the new map.
2960 return EBUSY; /* TODO: Make a new MDB_* error code? */
2962 env->me_psize = meta.mm_psize;
2963 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2964 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2966 env->me_maxpg = env->me_mapsize / env->me_psize;
2968 p = (MDB_page *)env->me_map;
2969 env->me_metas[0] = METADATA(p);
2970 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2974 int toggle = mdb_env_pick_meta(env);
2975 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2977 DPRINTF("opened database version %u, pagesize %u",
2978 env->me_metas[0]->mm_version, env->me_psize);
2979 DPRINTF("using meta page %d", toggle);
2980 DPRINTF("depth: %u", db->md_depth);
2981 DPRINTF("entries: %zu", db->md_entries);
2982 DPRINTF("branch pages: %zu", db->md_branch_pages);
2983 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2984 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2985 DPRINTF("root: %zu", db->md_root);
2993 /** Release a reader thread's slot in the reader lock table.
2994 * This function is called automatically when a thread exits.
2995 * @param[in] ptr This points to the slot in the reader lock table.
2998 mdb_env_reader_dest(void *ptr)
3000 MDB_reader *reader = ptr;
3006 /** Junk for arranging thread-specific callbacks on Windows. This is
3007 * necessarily platform and compiler-specific. Windows supports up
3008 * to 1088 keys. Let's assume nobody opens more than 64 environments
3009 * in a single process, for now. They can override this if needed.
3011 #ifndef MAX_TLS_KEYS
3012 #define MAX_TLS_KEYS 64
3014 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3015 static int mdb_tls_nkeys;
3017 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3021 case DLL_PROCESS_ATTACH: break;
3022 case DLL_THREAD_ATTACH: break;
3023 case DLL_THREAD_DETACH:
3024 for (i=0; i<mdb_tls_nkeys; i++) {
3025 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3026 mdb_env_reader_dest(r);
3029 case DLL_PROCESS_DETACH: break;
3034 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3036 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3040 /* Force some symbol references.
3041 * _tls_used forces the linker to create the TLS directory if not already done
3042 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3044 #pragma comment(linker, "/INCLUDE:_tls_used")
3045 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3046 #pragma const_seg(".CRT$XLB")
3047 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3048 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3051 #pragma comment(linker, "/INCLUDE:__tls_used")
3052 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3053 #pragma data_seg(".CRT$XLB")
3054 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3056 #endif /* WIN 32/64 */
3057 #endif /* !__GNUC__ */
3060 /** Downgrade the exclusive lock on the region back to shared */
3062 mdb_env_share_locks(MDB_env *env, int *excl)
3064 int rc = 0, toggle = mdb_env_pick_meta(env);
3066 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3071 /* First acquire a shared lock. The Unlock will
3072 * then release the existing exclusive lock.
3074 memset(&ov, 0, sizeof(ov));
3075 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3078 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3084 struct flock lock_info;
3085 /* The shared lock replaces the existing lock */
3086 memset((void *)&lock_info, 0, sizeof(lock_info));
3087 lock_info.l_type = F_RDLCK;
3088 lock_info.l_whence = SEEK_SET;
3089 lock_info.l_start = 0;
3090 lock_info.l_len = 1;
3091 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3092 (rc = ErrCode()) == EINTR) ;
3093 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3100 /** Try to get exlusive lock, otherwise shared.
3101 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3104 mdb_env_excl_lock(MDB_env *env, int *excl)
3108 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3112 memset(&ov, 0, sizeof(ov));
3113 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3120 struct flock lock_info;
3121 memset((void *)&lock_info, 0, sizeof(lock_info));
3122 lock_info.l_type = F_WRLCK;
3123 lock_info.l_whence = SEEK_SET;
3124 lock_info.l_start = 0;
3125 lock_info.l_len = 1;
3126 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3127 (rc = ErrCode()) == EINTR) ;
3131 # ifdef MDB_USE_POSIX_SEM
3132 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3135 lock_info.l_type = F_RDLCK;
3136 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3137 (rc = ErrCode()) == EINTR) ;
3145 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3147 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3149 * @(#) $Revision: 5.1 $
3150 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3151 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3153 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3157 * Please do not copyright this code. This code is in the public domain.
3159 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3160 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3161 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3162 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3163 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3164 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3165 * PERFORMANCE OF THIS SOFTWARE.
3168 * chongo <Landon Curt Noll> /\oo/\
3169 * http://www.isthe.com/chongo/
3171 * Share and Enjoy! :-)
3174 typedef unsigned long long mdb_hash_t;
3175 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3177 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3178 * @param[in] str string to hash
3179 * @param[in] hval initial value for hash
3180 * @return 64 bit hash
3182 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3183 * hval arg on the first call.
3186 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3188 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3189 unsigned char *end = s + val->mv_size;
3191 * FNV-1a hash each octet of the string
3194 /* xor the bottom with the current octet */
3195 hval ^= (mdb_hash_t)*s++;
3197 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3198 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3199 (hval << 7) + (hval << 8) + (hval << 40);
3201 /* return our new hash value */
3205 /** Hash the string and output the hash in hex.
3206 * @param[in] str string to hash
3207 * @param[out] hexbuf an array of 17 chars to hold the hash
3210 mdb_hash_hex(MDB_val *val, char *hexbuf)
3213 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3214 for (i=0; i<8; i++) {
3215 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3221 /** Open and/or initialize the lock region for the environment.
3222 * @param[in] env The MDB environment.
3223 * @param[in] lpath The pathname of the file used for the lock region.
3224 * @param[in] mode The Unix permissions for the file, if we create it.
3225 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3226 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3227 * @return 0 on success, non-zero on failure.
3230 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3233 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3235 # define MDB_ERRCODE_ROFS EROFS
3236 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3237 # define MDB_CLOEXEC O_CLOEXEC
3240 # define MDB_CLOEXEC 0
3247 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3248 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3249 FILE_ATTRIBUTE_NORMAL, NULL);
3251 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3253 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3255 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3260 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3261 /* Lose record locks when exec*() */
3262 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3263 fcntl(env->me_lfd, F_SETFD, fdflags);
3266 if (!(env->me_flags & MDB_NOTLS)) {
3267 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3270 env->me_flags |= MDB_ENV_TXKEY;
3272 /* Windows TLS callbacks need help finding their TLS info. */
3273 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3277 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3281 /* Try to get exclusive lock. If we succeed, then
3282 * nobody is using the lock region and we should initialize it.
3284 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3287 size = GetFileSize(env->me_lfd, NULL);
3289 size = lseek(env->me_lfd, 0, SEEK_END);
3291 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3292 if (size < rsize && *excl > 0) {
3294 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3295 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3297 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3301 size = rsize - sizeof(MDB_txninfo);
3302 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3307 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3309 if (!mh) goto fail_errno;
3310 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3312 if (!env->me_txns) goto fail_errno;
3314 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3316 if (m == MAP_FAILED) goto fail_errno;
3322 BY_HANDLE_FILE_INFORMATION stbuf;
3331 if (!mdb_sec_inited) {
3332 InitializeSecurityDescriptor(&mdb_null_sd,
3333 SECURITY_DESCRIPTOR_REVISION);
3334 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3335 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3336 mdb_all_sa.bInheritHandle = FALSE;
3337 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3340 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3341 idbuf.volume = stbuf.dwVolumeSerialNumber;
3342 idbuf.nhigh = stbuf.nFileIndexHigh;
3343 idbuf.nlow = stbuf.nFileIndexLow;
3344 val.mv_data = &idbuf;
3345 val.mv_size = sizeof(idbuf);
3346 mdb_hash_hex(&val, hexbuf);
3347 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3348 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3349 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3350 if (!env->me_rmutex) goto fail_errno;
3351 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3352 if (!env->me_wmutex) goto fail_errno;
3353 #elif defined(MDB_USE_POSIX_SEM)
3362 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3363 idbuf.dev = stbuf.st_dev;
3364 idbuf.ino = stbuf.st_ino;
3365 val.mv_data = &idbuf;
3366 val.mv_size = sizeof(idbuf);
3367 mdb_hash_hex(&val, hexbuf);
3368 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3369 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3370 /* Clean up after a previous run, if needed: Try to
3371 * remove both semaphores before doing anything else.
3373 sem_unlink(env->me_txns->mti_rmname);
3374 sem_unlink(env->me_txns->mti_wmname);
3375 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3376 O_CREAT|O_EXCL, mode, 1);
3377 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3378 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3379 O_CREAT|O_EXCL, mode, 1);
3380 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3381 #else /* MDB_USE_POSIX_SEM */
3382 pthread_mutexattr_t mattr;
3384 if ((rc = pthread_mutexattr_init(&mattr))
3385 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3386 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3387 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3389 pthread_mutexattr_destroy(&mattr);
3390 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3392 env->me_txns->mti_version = MDB_VERSION;
3393 env->me_txns->mti_magic = MDB_MAGIC;
3394 env->me_txns->mti_txnid = 0;
3395 env->me_txns->mti_numreaders = 0;
3398 if (env->me_txns->mti_magic != MDB_MAGIC) {
3399 DPUTS("lock region has invalid magic");
3403 if (env->me_txns->mti_version != MDB_VERSION) {
3404 DPRINTF("lock region is version %u, expected version %u",
3405 env->me_txns->mti_version, MDB_VERSION);
3406 rc = MDB_VERSION_MISMATCH;
3410 if (rc != EACCES && rc != EAGAIN) {
3414 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3415 if (!env->me_rmutex) goto fail_errno;
3416 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3417 if (!env->me_wmutex) goto fail_errno;
3418 #elif defined(MDB_USE_POSIX_SEM)
3419 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3420 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3421 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3422 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3433 /** The name of the lock file in the DB environment */
3434 #define LOCKNAME "/lock.mdb"
3435 /** The name of the data file in the DB environment */
3436 #define DATANAME "/data.mdb"
3437 /** The suffix of the lock file when no subdir is used */
3438 #define LOCKSUFF "-lock"
3439 /** Only a subset of the @ref mdb_env flags can be changed
3440 * at runtime. Changing other flags requires closing the
3441 * environment and re-opening it with the new flags.
3443 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3444 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3447 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3449 int oflags, rc, len, excl = -1;
3450 char *lpath, *dpath;
3452 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3456 if (flags & MDB_NOSUBDIR) {
3457 rc = len + sizeof(LOCKSUFF) + len + 1;
3459 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3464 if (flags & MDB_NOSUBDIR) {
3465 dpath = lpath + len + sizeof(LOCKSUFF);
3466 sprintf(lpath, "%s" LOCKSUFF, path);
3467 strcpy(dpath, path);
3469 dpath = lpath + len + sizeof(LOCKNAME);
3470 sprintf(lpath, "%s" LOCKNAME, path);
3471 sprintf(dpath, "%s" DATANAME, path);
3475 flags |= env->me_flags;
3476 if (flags & MDB_RDONLY) {
3477 /* silently ignore WRITEMAP when we're only getting read access */
3478 flags &= ~MDB_WRITEMAP;
3480 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
3481 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3484 env->me_flags = flags |= MDB_ENV_ACTIVE;
3488 env->me_path = strdup(path);
3489 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3490 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3491 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3496 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3501 if (F_ISSET(flags, MDB_RDONLY)) {
3502 oflags = GENERIC_READ;
3503 len = OPEN_EXISTING;
3505 oflags = GENERIC_READ|GENERIC_WRITE;
3508 mode = FILE_ATTRIBUTE_NORMAL;
3509 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3510 NULL, len, mode, NULL);
3512 if (F_ISSET(flags, MDB_RDONLY))
3515 oflags = O_RDWR | O_CREAT;
3517 env->me_fd = open(dpath, oflags, mode);
3519 if (env->me_fd == INVALID_HANDLE_VALUE) {
3524 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3525 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3526 env->me_mfd = env->me_fd;
3528 /* Synchronous fd for meta writes. Needed even with
3529 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3532 env->me_mfd = CreateFile(dpath, oflags,
3533 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3534 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3536 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3538 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3543 DPRINTF("opened dbenv %p", (void *) env);
3545 rc = mdb_env_share_locks(env, &excl);
3551 mdb_env_close0(env, excl);
3557 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
3559 mdb_env_close0(MDB_env *env, int excl)
3563 if (!(env->me_flags & MDB_ENV_ACTIVE))
3566 /* Doing this here since me_dbxs may not exist during mdb_env_close */
3567 for (i = env->me_maxdbs; --i > MAIN_DBI; )
3568 free(env->me_dbxs[i].md_name.mv_data);
3570 free(env->me_dbflags);
3573 free(env->me_dirty_list);
3574 if (env->me_free_pgs)
3575 mdb_midl_free(env->me_free_pgs);
3577 if (env->me_flags & MDB_ENV_TXKEY) {
3578 pthread_key_delete(env->me_txkey);
3580 /* Delete our key from the global list */
3581 for (i=0; i<mdb_tls_nkeys; i++)
3582 if (mdb_tls_keys[i] == env->me_txkey) {
3583 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3591 munmap(env->me_map, env->me_mapsize);
3593 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3595 if (env->me_fd != INVALID_HANDLE_VALUE)
3598 pid_t pid = env->me_pid;
3599 /* Clearing readers is done in this function because
3600 * me_txkey with its destructor must be disabled first.
3602 for (i = env->me_numreaders; --i >= 0; )
3603 if (env->me_txns->mti_readers[i].mr_pid == pid)
3604 env->me_txns->mti_readers[i].mr_pid = 0;
3606 if (env->me_rmutex) {
3607 CloseHandle(env->me_rmutex);
3608 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3610 /* Windows automatically destroys the mutexes when
3611 * the last handle closes.
3613 #elif defined(MDB_USE_POSIX_SEM)
3614 if (env->me_rmutex != SEM_FAILED) {
3615 sem_close(env->me_rmutex);
3616 if (env->me_wmutex != SEM_FAILED)
3617 sem_close(env->me_wmutex);
3618 /* If we have the filelock: If we are the
3619 * only remaining user, clean up semaphores.
3622 mdb_env_excl_lock(env, &excl);
3624 sem_unlink(env->me_txns->mti_rmname);
3625 sem_unlink(env->me_txns->mti_wmname);
3629 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3631 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3634 /* Unlock the lockfile. Windows would have unlocked it
3635 * after closing anyway, but not necessarily at once.
3637 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3643 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3647 mdb_env_copy(MDB_env *env, const char *path)
3649 MDB_txn *txn = NULL;
3653 HANDLE newfd = INVALID_HANDLE_VALUE;
3655 if (env->me_flags & MDB_NOSUBDIR) {
3656 lpath = (char *)path;
3659 len += sizeof(DATANAME);
3660 lpath = malloc(len);
3663 sprintf(lpath, "%s" DATANAME, path);
3666 /* The destination path must exist, but the destination file must not.
3667 * We don't want the OS to cache the writes, since the source data is
3668 * already in the OS cache.
3671 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3672 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3674 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3680 if (!(env->me_flags & MDB_NOSUBDIR))
3682 if (newfd == INVALID_HANDLE_VALUE) {
3687 #ifdef F_NOCACHE /* __APPLE__ */
3688 rc = fcntl(newfd, F_NOCACHE, 1);
3695 /* Do the lock/unlock of the reader mutex before starting the
3696 * write txn. Otherwise other read txns could block writers.
3698 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3703 /* We must start the actual read txn after blocking writers */
3704 mdb_txn_reset0(txn);
3706 /* Temporarily block writers until we snapshot the meta pages */
3709 rc = mdb_txn_renew0(txn);
3711 UNLOCK_MUTEX_W(env);
3716 wsize = env->me_psize * 2;
3720 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3721 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3724 rc = write(newfd, env->me_map, wsize);
3725 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3728 UNLOCK_MUTEX_W(env);
3733 ptr = env->me_map + wsize;
3734 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3735 #define MAX_WRITE 2147483648U
3739 if (wsize > MAX_WRITE)
3743 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3744 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3753 if (wsize > MAX_WRITE)
3757 wres = write(newfd, ptr, w2);
3758 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3767 if (newfd != INVALID_HANDLE_VALUE)
3774 mdb_env_close(MDB_env *env)
3781 VGMEMP_DESTROY(env);
3782 while ((dp = env->me_dpages) != NULL) {
3783 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3784 env->me_dpages = dp->mp_next;
3788 mdb_env_close0(env, 0);
3792 /** Compare two items pointing at aligned size_t's */
3794 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3796 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3797 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3800 /** Compare two items pointing at aligned int's */
3802 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3804 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3805 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3808 /** Compare two items pointing at ints of unknown alignment.
3809 * Nodes and keys are guaranteed to be 2-byte aligned.
3812 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3814 #if BYTE_ORDER == LITTLE_ENDIAN
3815 unsigned short *u, *c;
3818 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3819 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3822 } while(!x && u > (unsigned short *)a->mv_data);
3825 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3829 /** Compare two items lexically */
3831 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3838 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3844 diff = memcmp(a->mv_data, b->mv_data, len);
3845 return diff ? diff : len_diff<0 ? -1 : len_diff;
3848 /** Compare two items in reverse byte order */
3850 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3852 const unsigned char *p1, *p2, *p1_lim;
3856 p1_lim = (const unsigned char *)a->mv_data;
3857 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3858 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3860 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3866 while (p1 > p1_lim) {
3867 diff = *--p1 - *--p2;
3871 return len_diff<0 ? -1 : len_diff;
3874 /** Search for key within a page, using binary search.
3875 * Returns the smallest entry larger or equal to the key.
3876 * If exactp is non-null, stores whether the found entry was an exact match
3877 * in *exactp (1 or 0).
3878 * Updates the cursor index with the index of the found entry.
3879 * If no entry larger or equal to the key is found, returns NULL.
3882 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3884 unsigned int i = 0, nkeys;
3887 MDB_page *mp = mc->mc_pg[mc->mc_top];
3888 MDB_node *node = NULL;
3893 nkeys = NUMKEYS(mp);
3898 COPY_PGNO(pgno, mp->mp_pgno);
3899 DPRINTF("searching %u keys in %s %spage %zu",
3900 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3907 low = IS_LEAF(mp) ? 0 : 1;
3909 cmp = mc->mc_dbx->md_cmp;
3911 /* Branch pages have no data, so if using integer keys,
3912 * alignment is guaranteed. Use faster mdb_cmp_int.
3914 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3915 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3922 nodekey.mv_size = mc->mc_db->md_pad;
3923 node = NODEPTR(mp, 0); /* fake */
3924 while (low <= high) {
3925 i = (low + high) >> 1;
3926 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3927 rc = cmp(key, &nodekey);
3928 DPRINTF("found leaf index %u [%s], rc = %i",
3929 i, DKEY(&nodekey), rc);
3938 while (low <= high) {
3939 i = (low + high) >> 1;
3941 node = NODEPTR(mp, i);
3942 nodekey.mv_size = NODEKSZ(node);
3943 nodekey.mv_data = NODEKEY(node);
3945 rc = cmp(key, &nodekey);
3948 DPRINTF("found leaf index %u [%s], rc = %i",
3949 i, DKEY(&nodekey), rc);
3951 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3952 i, DKEY(&nodekey), NODEPGNO(node), rc);
3963 if (rc > 0) { /* Found entry is less than the key. */
3964 i++; /* Skip to get the smallest entry larger than key. */
3966 node = NODEPTR(mp, i);
3969 *exactp = (rc == 0);
3970 /* store the key index */
3971 mc->mc_ki[mc->mc_top] = i;
3973 /* There is no entry larger or equal to the key. */
3976 /* nodeptr is fake for LEAF2 */
3982 mdb_cursor_adjust(MDB_cursor *mc, func)
3986 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3987 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3994 /** Pop a page off the top of the cursor's stack. */
3996 mdb_cursor_pop(MDB_cursor *mc)
3999 #ifndef MDB_DEBUG_SKIP
4000 MDB_page *top = mc->mc_pg[mc->mc_top];
4006 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
4007 mc->mc_dbi, (void *) mc);
4011 /** Push a page onto the top of the cursor's stack. */
4013 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4015 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4016 mc->mc_dbi, (void *) mc);
4018 if (mc->mc_snum >= CURSOR_STACK) {
4019 assert(mc->mc_snum < CURSOR_STACK);
4020 return MDB_CURSOR_FULL;
4023 mc->mc_top = mc->mc_snum++;
4024 mc->mc_pg[mc->mc_top] = mp;
4025 mc->mc_ki[mc->mc_top] = 0;
4030 /** Find the address of the page corresponding to a given page number.
4031 * @param[in] txn the transaction for this access.
4032 * @param[in] pgno the page number for the page to retrieve.
4033 * @param[out] ret address of a pointer where the page's address will be stored.
4034 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4035 * @return 0 on success, non-zero on failure.
4038 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4043 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4044 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4049 MDB_ID2L dl = tx2->mt_u.dirty_list;
4051 unsigned x = mdb_mid2l_search(dl, pgno);
4052 if (x <= dl[0].mid && dl[x].mid == pgno) {
4058 } while ((tx2 = tx2->mt_parent) != NULL);
4061 if (pgno < txn->mt_next_pgno) {
4063 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4065 DPRINTF("page %zu not found", pgno);
4067 return MDB_PAGE_NOTFOUND;
4077 /** Search for the page a given key should be in.
4078 * Pushes parent pages on the cursor stack. This function continues a
4079 * search on a cursor that has already been initialized. (Usually by
4080 * #mdb_page_search() but also by #mdb_node_move().)
4081 * @param[in,out] mc the cursor for this operation.
4082 * @param[in] key the key to search for. If NULL, search for the lowest
4083 * page. (This is used by #mdb_cursor_first().)
4084 * @param[in] modify If true, visited pages are updated with new page numbers.
4085 * @return 0 on success, non-zero on failure.
4088 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4090 MDB_page *mp = mc->mc_pg[mc->mc_top];
4095 while (IS_BRANCH(mp)) {
4099 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4100 assert(NUMKEYS(mp) > 1);
4101 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4103 if (key == NULL) /* Initialize cursor to first page. */
4105 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4106 /* cursor to last page */
4110 node = mdb_node_search(mc, key, &exact);
4112 i = NUMKEYS(mp) - 1;
4114 i = mc->mc_ki[mc->mc_top];
4123 DPRINTF("following index %u for key [%s]",
4125 assert(i < NUMKEYS(mp));
4126 node = NODEPTR(mp, i);
4128 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4131 mc->mc_ki[mc->mc_top] = i;
4132 if ((rc = mdb_cursor_push(mc, mp)))
4136 if ((rc = mdb_page_touch(mc)) != 0)
4138 mp = mc->mc_pg[mc->mc_top];
4143 DPRINTF("internal error, index points to a %02X page!?",
4145 return MDB_CORRUPTED;
4148 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4149 key ? DKEY(key) : NULL);
4154 /** Search for the lowest key under the current branch page.
4155 * This just bypasses a NUMKEYS check in the current page
4156 * before calling mdb_page_search_root(), because the callers
4157 * are all in situations where the current page is known to
4161 mdb_page_search_lowest(MDB_cursor *mc)
4163 MDB_page *mp = mc->mc_pg[mc->mc_top];
4164 MDB_node *node = NODEPTR(mp, 0);
4167 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4170 mc->mc_ki[mc->mc_top] = 0;
4171 if ((rc = mdb_cursor_push(mc, mp)))
4173 return mdb_page_search_root(mc, NULL, 0);
4176 /** Search for the page a given key should be in.
4177 * Pushes parent pages on the cursor stack. This function just sets up
4178 * the search; it finds the root page for \b mc's database and sets this
4179 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4180 * called to complete the search.
4181 * @param[in,out] mc the cursor for this operation.
4182 * @param[in] key the key to search for. If NULL, search for the lowest
4183 * page. (This is used by #mdb_cursor_first().)
4184 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4185 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4186 * @return 0 on success, non-zero on failure.
4189 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4194 /* Make sure the txn is still viable, then find the root from
4195 * the txn's db table.
4197 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4198 DPUTS("transaction has failed, must abort");
4201 /* Make sure we're using an up-to-date root */
4202 if (mc->mc_dbi > MAIN_DBI) {
4203 if ((*mc->mc_dbflag & DB_STALE) ||
4204 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4206 unsigned char dbflag = 0;
4207 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4208 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4211 if (*mc->mc_dbflag & DB_STALE) {
4215 MDB_node *leaf = mdb_node_search(&mc2,
4216 &mc->mc_dbx->md_name, &exact);
4218 return MDB_NOTFOUND;
4219 mdb_node_read(mc->mc_txn, leaf, &data);
4220 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4222 /* The txn may not know this DBI, or another process may
4223 * have dropped and recreated the DB with other flags.
4225 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4226 return MDB_INCOMPATIBLE;
4227 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4229 if (flags & MDB_PS_MODIFY)
4231 *mc->mc_dbflag &= ~DB_STALE;
4232 *mc->mc_dbflag |= dbflag;
4235 root = mc->mc_db->md_root;
4237 if (root == P_INVALID) { /* Tree is empty. */
4238 DPUTS("tree is empty");
4239 return MDB_NOTFOUND;
4244 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4245 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4251 DPRINTF("db %u root page %zu has flags 0x%X",
4252 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4254 if (flags & MDB_PS_MODIFY) {
4255 if ((rc = mdb_page_touch(mc)))
4259 if (flags & MDB_PS_ROOTONLY)
4262 return mdb_page_search_root(mc, key, flags);
4265 /** Return the data associated with a given node.
4266 * @param[in] txn The transaction for this operation.
4267 * @param[in] leaf The node being read.
4268 * @param[out] data Updated to point to the node's data.
4269 * @return 0 on success, non-zero on failure.
4272 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4274 MDB_page *omp; /* overflow page */
4278 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4279 data->mv_size = NODEDSZ(leaf);
4280 data->mv_data = NODEDATA(leaf);
4284 /* Read overflow data.
4286 data->mv_size = NODEDSZ(leaf);
4287 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4288 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
4289 DPRINTF("read overflow page %zu failed", pgno);
4292 data->mv_data = METADATA(omp);
4298 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4299 MDB_val *key, MDB_val *data)
4308 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4310 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4313 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4317 mdb_cursor_init(&mc, txn, dbi, &mx);
4318 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4321 /** Find a sibling for a page.
4322 * Replaces the page at the top of the cursor's stack with the
4323 * specified sibling, if one exists.
4324 * @param[in] mc The cursor for this operation.
4325 * @param[in] move_right Non-zero if the right sibling is requested,
4326 * otherwise the left sibling.
4327 * @return 0 on success, non-zero on failure.
4330 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4336 if (mc->mc_snum < 2) {
4337 return MDB_NOTFOUND; /* root has no siblings */
4341 DPRINTF("parent page is page %zu, index %u",
4342 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4344 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4345 : (mc->mc_ki[mc->mc_top] == 0)) {
4346 DPRINTF("no more keys left, moving to %s sibling",
4347 move_right ? "right" : "left");
4348 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4349 /* undo cursor_pop before returning */
4356 mc->mc_ki[mc->mc_top]++;
4358 mc->mc_ki[mc->mc_top]--;
4359 DPRINTF("just moving to %s index key %u",
4360 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4362 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4364 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4365 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL) != 0))
4368 mdb_cursor_push(mc, mp);
4370 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4375 /** Move the cursor to the next data item. */
4377 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4383 if (mc->mc_flags & C_EOF) {
4384 return MDB_NOTFOUND;
4387 assert(mc->mc_flags & C_INITIALIZED);
4389 mp = mc->mc_pg[mc->mc_top];
4391 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4392 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4393 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4394 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4395 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4396 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4400 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4401 if (op == MDB_NEXT_DUP)
4402 return MDB_NOTFOUND;
4406 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4408 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4409 DPUTS("=====> move to next sibling page");
4410 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4411 mc->mc_flags |= C_EOF;
4412 mc->mc_flags &= ~C_INITIALIZED;
4413 return MDB_NOTFOUND;
4415 mp = mc->mc_pg[mc->mc_top];
4416 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4418 mc->mc_ki[mc->mc_top]++;
4420 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4421 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4424 key->mv_size = mc->mc_db->md_pad;
4425 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4429 assert(IS_LEAF(mp));
4430 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4432 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4433 mdb_xcursor_init1(mc, leaf);
4436 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4439 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4440 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4441 if (rc != MDB_SUCCESS)
4446 MDB_GET_KEY(leaf, key);
4450 /** Move the cursor to the previous data item. */
4452 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4458 assert(mc->mc_flags & C_INITIALIZED);
4460 mp = mc->mc_pg[mc->mc_top];
4462 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4463 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4464 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4465 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4466 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4467 if (op != MDB_PREV || rc == MDB_SUCCESS)
4470 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4471 if (op == MDB_PREV_DUP)
4472 return MDB_NOTFOUND;
4477 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4479 if (mc->mc_ki[mc->mc_top] == 0) {
4480 DPUTS("=====> move to prev sibling page");
4481 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4482 mc->mc_flags &= ~C_INITIALIZED;
4483 return MDB_NOTFOUND;
4485 mp = mc->mc_pg[mc->mc_top];
4486 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4487 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4489 mc->mc_ki[mc->mc_top]--;
4491 mc->mc_flags &= ~C_EOF;
4493 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4494 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4497 key->mv_size = mc->mc_db->md_pad;
4498 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4502 assert(IS_LEAF(mp));
4503 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4505 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4506 mdb_xcursor_init1(mc, leaf);
4509 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4512 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4513 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4514 if (rc != MDB_SUCCESS)
4519 MDB_GET_KEY(leaf, key);
4523 /** Set the cursor on a specific data item. */
4525 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4526 MDB_cursor_op op, int *exactp)
4530 MDB_node *leaf = NULL;
4535 assert(key->mv_size > 0);
4537 /* See if we're already on the right page */
4538 if (mc->mc_flags & C_INITIALIZED) {
4541 mp = mc->mc_pg[mc->mc_top];
4543 mc->mc_ki[mc->mc_top] = 0;
4544 return MDB_NOTFOUND;
4546 if (mp->mp_flags & P_LEAF2) {
4547 nodekey.mv_size = mc->mc_db->md_pad;
4548 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4550 leaf = NODEPTR(mp, 0);
4551 MDB_GET_KEY(leaf, &nodekey);
4553 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4555 /* Probably happens rarely, but first node on the page
4556 * was the one we wanted.
4558 mc->mc_ki[mc->mc_top] = 0;
4565 unsigned int nkeys = NUMKEYS(mp);
4567 if (mp->mp_flags & P_LEAF2) {
4568 nodekey.mv_data = LEAF2KEY(mp,
4569 nkeys-1, nodekey.mv_size);
4571 leaf = NODEPTR(mp, nkeys-1);
4572 MDB_GET_KEY(leaf, &nodekey);
4574 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4576 /* last node was the one we wanted */
4577 mc->mc_ki[mc->mc_top] = nkeys-1;
4583 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4584 /* This is definitely the right page, skip search_page */
4585 if (mp->mp_flags & P_LEAF2) {
4586 nodekey.mv_data = LEAF2KEY(mp,
4587 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4589 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4590 MDB_GET_KEY(leaf, &nodekey);
4592 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4594 /* current node was the one we wanted */
4604 /* If any parents have right-sibs, search.
4605 * Otherwise, there's nothing further.
4607 for (i=0; i<mc->mc_top; i++)
4609 NUMKEYS(mc->mc_pg[i])-1)
4611 if (i == mc->mc_top) {
4612 /* There are no other pages */
4613 mc->mc_ki[mc->mc_top] = nkeys;
4614 return MDB_NOTFOUND;
4618 /* There are no other pages */
4619 mc->mc_ki[mc->mc_top] = 0;
4620 return MDB_NOTFOUND;
4624 rc = mdb_page_search(mc, key, 0);
4625 if (rc != MDB_SUCCESS)
4628 mp = mc->mc_pg[mc->mc_top];
4629 assert(IS_LEAF(mp));
4632 leaf = mdb_node_search(mc, key, exactp);
4633 if (exactp != NULL && !*exactp) {
4634 /* MDB_SET specified and not an exact match. */
4635 return MDB_NOTFOUND;
4639 DPUTS("===> inexact leaf not found, goto sibling");
4640 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4641 return rc; /* no entries matched */
4642 mp = mc->mc_pg[mc->mc_top];
4643 assert(IS_LEAF(mp));
4644 leaf = NODEPTR(mp, 0);
4648 mc->mc_flags |= C_INITIALIZED;
4649 mc->mc_flags &= ~C_EOF;
4652 key->mv_size = mc->mc_db->md_pad;
4653 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4657 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4658 mdb_xcursor_init1(mc, leaf);
4661 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4662 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4663 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4666 if (op == MDB_GET_BOTH) {
4672 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4673 if (rc != MDB_SUCCESS)
4676 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4678 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4680 rc = mc->mc_dbx->md_dcmp(data, &d2);
4682 if (op == MDB_GET_BOTH || rc > 0)
4683 return MDB_NOTFOUND;
4688 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4689 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4694 /* The key already matches in all other cases */
4695 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4696 MDB_GET_KEY(leaf, key);
4697 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4702 /** Move the cursor to the first item in the database. */
4704 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4709 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4710 rc = mdb_page_search(mc, NULL, 0);
4711 if (rc != MDB_SUCCESS)
4714 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4716 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4717 mc->mc_flags |= C_INITIALIZED;
4718 mc->mc_flags &= ~C_EOF;
4720 mc->mc_ki[mc->mc_top] = 0;
4722 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4723 key->mv_size = mc->mc_db->md_pad;
4724 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4729 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4730 mdb_xcursor_init1(mc, leaf);
4731 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4736 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4737 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4741 MDB_GET_KEY(leaf, key);
4745 /** Move the cursor to the last item in the database. */
4747 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4752 if (!(mc->mc_flags & C_EOF)) {
4754 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4757 lkey.mv_size = MDB_MAXKEYSIZE+1;
4758 lkey.mv_data = NULL;
4759 rc = mdb_page_search(mc, &lkey, 0);
4760 if (rc != MDB_SUCCESS)
4763 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4766 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4767 mc->mc_flags |= C_INITIALIZED|C_EOF;
4768 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4770 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4771 key->mv_size = mc->mc_db->md_pad;
4772 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4777 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4778 mdb_xcursor_init1(mc, leaf);
4779 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4784 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4785 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4790 MDB_GET_KEY(leaf, key);
4795 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4804 case MDB_GET_CURRENT:
4805 if (!(mc->mc_flags & C_INITIALIZED)) {
4808 MDB_page *mp = mc->mc_pg[mc->mc_top];
4810 mc->mc_ki[mc->mc_top] = 0;
4816 key->mv_size = mc->mc_db->md_pad;
4817 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4819 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4820 MDB_GET_KEY(leaf, key);
4822 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4823 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4825 rc = mdb_node_read(mc->mc_txn, leaf, data);
4832 case MDB_GET_BOTH_RANGE:
4833 if (data == NULL || mc->mc_xcursor == NULL) {
4841 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4843 } else if (op == MDB_SET_RANGE)
4844 rc = mdb_cursor_set(mc, key, data, op, NULL);
4846 rc = mdb_cursor_set(mc, key, data, op, &exact);
4848 case MDB_GET_MULTIPLE:
4850 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4851 !(mc->mc_flags & C_INITIALIZED)) {
4856 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4857 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4860 case MDB_NEXT_MULTIPLE:
4862 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4866 if (!(mc->mc_flags & C_INITIALIZED))
4867 rc = mdb_cursor_first(mc, key, data);
4869 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4870 if (rc == MDB_SUCCESS) {
4871 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4874 mx = &mc->mc_xcursor->mx_cursor;
4875 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4877 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4878 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4886 case MDB_NEXT_NODUP:
4887 if (!(mc->mc_flags & C_INITIALIZED))
4888 rc = mdb_cursor_first(mc, key, data);
4890 rc = mdb_cursor_next(mc, key, data, op);
4894 case MDB_PREV_NODUP:
4895 if (!(mc->mc_flags & C_INITIALIZED)) {
4896 rc = mdb_cursor_last(mc, key, data);
4897 mc->mc_flags |= C_INITIALIZED;
4898 mc->mc_ki[mc->mc_top]++;
4900 rc = mdb_cursor_prev(mc, key, data, op);
4903 rc = mdb_cursor_first(mc, key, data);
4907 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4908 !(mc->mc_flags & C_INITIALIZED) ||
4909 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4913 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4916 rc = mdb_cursor_last(mc, key, data);
4920 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4921 !(mc->mc_flags & C_INITIALIZED) ||
4922 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4926 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4929 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4937 /** Touch all the pages in the cursor stack.
4938 * Makes sure all the pages are writable, before attempting a write operation.
4939 * @param[in] mc The cursor to operate on.
4942 mdb_cursor_touch(MDB_cursor *mc)
4946 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4949 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
4950 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4953 *mc->mc_dbflag |= DB_DIRTY;
4955 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4956 rc = mdb_page_touch(mc);
4960 mc->mc_top = mc->mc_snum-1;
4965 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4968 MDB_node *leaf = NULL;
4969 MDB_val xdata, *rdata, dkey;
4972 int do_sub = 0, insert = 0;
4973 unsigned int mcount = 0;
4977 char dbuf[MDB_MAXKEYSIZE+1];
4978 unsigned int nflags;
4981 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4984 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4987 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4990 #if SIZE_MAX > MAXDATASIZE
4991 if (data->mv_size > MAXDATASIZE)
4995 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4996 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
5000 if (flags == MDB_CURRENT) {
5001 if (!(mc->mc_flags & C_INITIALIZED))
5004 } else if (mc->mc_db->md_root == P_INVALID) {
5006 /* new database, write a root leaf page */
5007 DPUTS("allocating new root leaf page");
5008 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
5012 mdb_cursor_push(mc, np);
5013 mc->mc_db->md_root = np->mp_pgno;
5014 mc->mc_db->md_depth++;
5015 *mc->mc_dbflag |= DB_DIRTY;
5016 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5018 np->mp_flags |= P_LEAF2;
5019 mc->mc_flags |= C_INITIALIZED;
5025 if (flags & MDB_APPEND) {
5027 rc = mdb_cursor_last(mc, &k2, &d2);
5029 rc = mc->mc_dbx->md_cmp(key, &k2);
5032 mc->mc_ki[mc->mc_top]++;
5034 /* new key is <= last key */
5039 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5041 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5042 DPRINTF("duplicate key [%s]", DKEY(key));
5044 return MDB_KEYEXIST;
5046 if (rc && rc != MDB_NOTFOUND)
5050 /* Cursor is positioned, now make sure all pages are writable */
5051 rc2 = mdb_cursor_touch(mc);
5056 /* The key already exists */
5057 if (rc == MDB_SUCCESS) {
5058 /* there's only a key anyway, so this is a no-op */
5059 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5060 unsigned int ksize = mc->mc_db->md_pad;
5061 if (key->mv_size != ksize)
5063 if (flags == MDB_CURRENT) {
5064 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5065 memcpy(ptr, key->mv_data, ksize);
5070 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5073 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5074 /* Was a single item before, must convert now */
5076 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5077 /* Just overwrite the current item */
5078 if (flags == MDB_CURRENT)
5081 dkey.mv_size = NODEDSZ(leaf);
5082 dkey.mv_data = NODEDATA(leaf);
5083 #if UINT_MAX < SIZE_MAX
5084 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5085 #ifdef MISALIGNED_OK
5086 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5088 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5091 /* if data matches, ignore it */
5092 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5093 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5095 /* create a fake page for the dup items */
5096 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5097 dkey.mv_data = dbuf;
5098 fp = (MDB_page *)&pbuf;
5099 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5100 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5101 fp->mp_lower = PAGEHDRSZ;
5102 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5103 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5104 fp->mp_flags |= P_LEAF2;
5105 fp->mp_pad = data->mv_size;
5106 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5108 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5109 (dkey.mv_size & 1) + (data->mv_size & 1);
5111 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5114 xdata.mv_size = fp->mp_upper;
5119 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5120 /* See if we need to convert from fake page to subDB */
5122 unsigned int offset;
5125 fp = NODEDATA(leaf);
5126 if (flags == MDB_CURRENT) {
5128 fp->mp_flags |= P_DIRTY;
5129 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5130 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5134 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5135 offset = fp->mp_pad;
5136 if (SIZELEFT(fp) >= offset)
5138 offset *= 4; /* space for 4 more */
5140 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5142 offset += offset & 1;
5143 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5144 offset >= mc->mc_txn->mt_env->me_nodemax) {
5145 /* yes, convert it */
5147 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5148 dummy.md_pad = fp->mp_pad;
5149 dummy.md_flags = MDB_DUPFIXED;
5150 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5151 dummy.md_flags |= MDB_INTEGERKEY;
5154 dummy.md_branch_pages = 0;
5155 dummy.md_leaf_pages = 1;
5156 dummy.md_overflow_pages = 0;
5157 dummy.md_entries = NUMKEYS(fp);
5159 xdata.mv_size = sizeof(MDB_db);
5160 xdata.mv_data = &dummy;
5161 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5163 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5164 flags |= F_DUPDATA|F_SUBDATA;
5165 dummy.md_root = mp->mp_pgno;
5167 /* no, just grow it */
5169 xdata.mv_size = NODEDSZ(leaf) + offset;
5170 xdata.mv_data = &pbuf;
5171 mp = (MDB_page *)&pbuf;
5172 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5175 mp->mp_flags = fp->mp_flags | P_DIRTY;
5176 mp->mp_pad = fp->mp_pad;
5177 mp->mp_lower = fp->mp_lower;
5178 mp->mp_upper = fp->mp_upper + offset;
5180 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5182 nsize = NODEDSZ(leaf) - fp->mp_upper;
5183 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5184 for (i=0; i<NUMKEYS(fp); i++)
5185 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5187 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5191 /* data is on sub-DB, just store it */
5192 flags |= F_DUPDATA|F_SUBDATA;
5196 /* overflow page overwrites need special handling */
5197 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5200 int ovpages, dpages;
5202 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5203 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5204 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
5206 ovpages = omp->mp_pages;
5208 /* Is the ov page writable and large enough? */
5209 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5210 /* yes, overwrite it. Note in this case we don't
5211 * bother to try shrinking the page if the new data
5212 * is smaller than the overflow threshold.
5214 SETDSZ(leaf, data->mv_size);
5215 if (F_ISSET(flags, MDB_RESERVE))
5216 data->mv_data = METADATA(omp);
5218 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5221 /* no, free ovpages */
5223 mc->mc_db->md_overflow_pages -= ovpages;
5224 for (i=0; i<ovpages; i++) {
5225 DPRINTF("freed ov page %zu", pg);
5226 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5230 } else if (NODEDSZ(leaf) == data->mv_size) {
5231 /* same size, just replace it. Note that we could
5232 * also reuse this node if the new data is smaller,
5233 * but instead we opt to shrink the node in that case.
5235 if (F_ISSET(flags, MDB_RESERVE))
5236 data->mv_data = NODEDATA(leaf);
5237 else if (data->mv_size)
5238 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5240 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5243 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5244 mc->mc_db->md_entries--;
5246 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5253 nflags = flags & NODE_ADD_FLAGS;
5254 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5255 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5256 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5257 nflags &= ~MDB_APPEND;
5259 nflags |= MDB_SPLIT_REPLACE;
5260 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5262 /* There is room already in this leaf page. */
5263 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5264 if (rc == 0 && !do_sub && insert) {
5265 /* Adjust other cursors pointing to mp */
5266 MDB_cursor *m2, *m3;
5267 MDB_dbi dbi = mc->mc_dbi;
5268 unsigned i = mc->mc_top;
5269 MDB_page *mp = mc->mc_pg[i];
5271 if (mc->mc_flags & C_SUB)
5274 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5275 if (mc->mc_flags & C_SUB)
5276 m3 = &m2->mc_xcursor->mx_cursor;
5279 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5280 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5287 if (rc != MDB_SUCCESS)
5288 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5290 /* Now store the actual data in the child DB. Note that we're
5291 * storing the user data in the keys field, so there are strict
5292 * size limits on dupdata. The actual data fields of the child
5293 * DB are all zero size.
5300 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5301 if (flags & MDB_CURRENT) {
5302 xflags = MDB_CURRENT;
5304 mdb_xcursor_init1(mc, leaf);
5305 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5307 /* converted, write the original data first */
5309 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5313 /* Adjust other cursors pointing to mp */
5315 unsigned i = mc->mc_top;
5316 MDB_page *mp = mc->mc_pg[i];
5318 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5319 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5320 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5321 mdb_xcursor_init1(m2, leaf);
5325 /* we've done our job */
5328 if (flags & MDB_APPENDDUP)
5329 xflags |= MDB_APPEND;
5330 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5331 if (flags & F_SUBDATA) {
5332 void *db = NODEDATA(leaf);
5333 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5336 /* sub-writes might have failed so check rc again.
5337 * Don't increment count if we just replaced an existing item.
5339 if (!rc && !(flags & MDB_CURRENT))
5340 mc->mc_db->md_entries++;
5341 if (flags & MDB_MULTIPLE) {
5343 if (mcount < data[1].mv_size) {
5344 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5345 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5351 /* If we succeeded and the key didn't exist before, make sure
5352 * the cursor is marked valid.
5355 mc->mc_flags |= C_INITIALIZED;
5360 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5365 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5368 if (!(mc->mc_flags & C_INITIALIZED))
5371 rc = mdb_cursor_touch(mc);
5375 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5377 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5378 if (flags != MDB_NODUPDATA) {
5379 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5380 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5382 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5383 /* If sub-DB still has entries, we're done */
5384 if (mc->mc_xcursor->mx_db.md_entries) {
5385 if (leaf->mn_flags & F_SUBDATA) {
5386 /* update subDB info */
5387 void *db = NODEDATA(leaf);
5388 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5390 /* shrink fake page */
5391 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5393 mc->mc_db->md_entries--;
5396 /* otherwise fall thru and delete the sub-DB */
5399 if (leaf->mn_flags & F_SUBDATA) {
5400 /* add all the child DB's pages to the free list */
5401 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5402 if (rc == MDB_SUCCESS) {
5403 mc->mc_db->md_entries -=
5404 mc->mc_xcursor->mx_db.md_entries;
5409 return mdb_cursor_del0(mc, leaf);
5412 /** Allocate and initialize new pages for a database.
5413 * @param[in] mc a cursor on the database being added to.
5414 * @param[in] flags flags defining what type of page is being allocated.
5415 * @param[in] num the number of pages to allocate. This is usually 1,
5416 * unless allocating overflow pages for a large record.
5417 * @param[out] mp Address of a page, or NULL on failure.
5418 * @return 0 on success, non-zero on failure.
5421 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5426 if ((rc = mdb_page_alloc(mc, num, &np)))
5428 DPRINTF("allocated new mpage %zu, page size %u",
5429 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5430 np->mp_flags = flags | P_DIRTY;
5431 np->mp_lower = PAGEHDRSZ;
5432 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5435 mc->mc_db->md_branch_pages++;
5436 else if (IS_LEAF(np))
5437 mc->mc_db->md_leaf_pages++;
5438 else if (IS_OVERFLOW(np)) {
5439 mc->mc_db->md_overflow_pages += num;
5447 /** Calculate the size of a leaf node.
5448 * The size depends on the environment's page size; if a data item
5449 * is too large it will be put onto an overflow page and the node
5450 * size will only include the key and not the data. Sizes are always
5451 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5452 * of the #MDB_node headers.
5453 * @param[in] env The environment handle.
5454 * @param[in] key The key for the node.
5455 * @param[in] data The data for the node.
5456 * @return The number of bytes needed to store the node.
5459 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5463 sz = LEAFSIZE(key, data);
5464 if (sz >= env->me_nodemax) {
5465 /* put on overflow page */
5466 sz -= data->mv_size - sizeof(pgno_t);
5470 return sz + sizeof(indx_t);
5473 /** Calculate the size of a branch node.
5474 * The size should depend on the environment's page size but since
5475 * we currently don't support spilling large keys onto overflow
5476 * pages, it's simply the size of the #MDB_node header plus the
5477 * size of the key. Sizes are always rounded up to an even number
5478 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5479 * @param[in] env The environment handle.
5480 * @param[in] key The key for the node.
5481 * @return The number of bytes needed to store the node.
5484 mdb_branch_size(MDB_env *env, MDB_val *key)
5489 if (sz >= env->me_nodemax) {
5490 /* put on overflow page */
5491 /* not implemented */
5492 /* sz -= key->size - sizeof(pgno_t); */
5495 return sz + sizeof(indx_t);
5498 /** Add a node to the page pointed to by the cursor.
5499 * @param[in] mc The cursor for this operation.
5500 * @param[in] indx The index on the page where the new node should be added.
5501 * @param[in] key The key for the new node.
5502 * @param[in] data The data for the new node, if any.
5503 * @param[in] pgno The page number, if adding a branch node.
5504 * @param[in] flags Flags for the node.
5505 * @return 0 on success, non-zero on failure. Possible errors are:
5507 * <li>ENOMEM - failed to allocate overflow pages for the node.
5508 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5509 * should never happen since all callers already calculate the
5510 * page's free space before calling this function.
5514 mdb_node_add(MDB_cursor *mc, indx_t indx,
5515 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5518 size_t node_size = NODESIZE;
5521 MDB_page *mp = mc->mc_pg[mc->mc_top];
5522 MDB_page *ofp = NULL; /* overflow page */
5525 assert(mp->mp_upper >= mp->mp_lower);
5527 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5528 IS_LEAF(mp) ? "leaf" : "branch",
5529 IS_SUBP(mp) ? "sub-" : "",
5530 mp->mp_pgno, indx, data ? data->mv_size : 0,
5531 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5534 /* Move higher keys up one slot. */
5535 int ksize = mc->mc_db->md_pad, dif;
5536 char *ptr = LEAF2KEY(mp, indx, ksize);
5537 dif = NUMKEYS(mp) - indx;
5539 memmove(ptr+ksize, ptr, dif*ksize);
5540 /* insert new key */
5541 memcpy(ptr, key->mv_data, ksize);
5543 /* Just using these for counting */
5544 mp->mp_lower += sizeof(indx_t);
5545 mp->mp_upper -= ksize - sizeof(indx_t);
5550 node_size += key->mv_size;
5554 if (F_ISSET(flags, F_BIGDATA)) {
5555 /* Data already on overflow page. */
5556 node_size += sizeof(pgno_t);
5557 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5558 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5560 /* Put data on overflow page. */
5561 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5562 data->mv_size, node_size+data->mv_size);
5563 node_size += sizeof(pgno_t);
5564 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5566 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5569 node_size += data->mv_size;
5572 node_size += node_size & 1;
5574 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5575 DPRINTF("not enough room in page %zu, got %u ptrs",
5576 mp->mp_pgno, NUMKEYS(mp));
5577 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5578 mp->mp_upper - mp->mp_lower);
5579 DPRINTF("node size = %zu", node_size);
5580 return MDB_PAGE_FULL;
5583 /* Move higher pointers up one slot. */
5584 for (i = NUMKEYS(mp); i > indx; i--)
5585 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5587 /* Adjust free space offsets. */
5588 ofs = mp->mp_upper - node_size;
5589 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5590 mp->mp_ptrs[indx] = ofs;
5592 mp->mp_lower += sizeof(indx_t);
5594 /* Write the node data. */
5595 node = NODEPTR(mp, indx);
5596 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5597 node->mn_flags = flags;
5599 SETDSZ(node,data->mv_size);
5604 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5609 if (F_ISSET(flags, F_BIGDATA))
5610 memcpy(node->mn_data + key->mv_size, data->mv_data,
5612 else if (F_ISSET(flags, MDB_RESERVE))
5613 data->mv_data = node->mn_data + key->mv_size;
5615 memcpy(node->mn_data + key->mv_size, data->mv_data,
5618 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5620 if (F_ISSET(flags, MDB_RESERVE))
5621 data->mv_data = METADATA(ofp);
5623 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5630 /** Delete the specified node from a page.
5631 * @param[in] mp The page to operate on.
5632 * @param[in] indx The index of the node to delete.
5633 * @param[in] ksize The size of a node. Only used if the page is
5634 * part of a #MDB_DUPFIXED database.
5637 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5640 indx_t i, j, numkeys, ptr;
5647 COPY_PGNO(pgno, mp->mp_pgno);
5648 DPRINTF("delete node %u on %s page %zu", indx,
5649 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5652 assert(indx < NUMKEYS(mp));
5655 int x = NUMKEYS(mp) - 1 - indx;
5656 base = LEAF2KEY(mp, indx, ksize);
5658 memmove(base, base + ksize, x * ksize);
5659 mp->mp_lower -= sizeof(indx_t);
5660 mp->mp_upper += ksize - sizeof(indx_t);
5664 node = NODEPTR(mp, indx);
5665 sz = NODESIZE + node->mn_ksize;
5667 if (F_ISSET(node->mn_flags, F_BIGDATA))
5668 sz += sizeof(pgno_t);
5670 sz += NODEDSZ(node);
5674 ptr = mp->mp_ptrs[indx];
5675 numkeys = NUMKEYS(mp);
5676 for (i = j = 0; i < numkeys; i++) {
5678 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5679 if (mp->mp_ptrs[i] < ptr)
5680 mp->mp_ptrs[j] += sz;
5685 base = (char *)mp + mp->mp_upper;
5686 memmove(base + sz, base, ptr - mp->mp_upper);
5688 mp->mp_lower -= sizeof(indx_t);
5692 /** Compact the main page after deleting a node on a subpage.
5693 * @param[in] mp The main page to operate on.
5694 * @param[in] indx The index of the subpage on the main page.
5697 mdb_node_shrink(MDB_page *mp, indx_t indx)
5704 indx_t i, numkeys, ptr;
5706 node = NODEPTR(mp, indx);
5707 sp = (MDB_page *)NODEDATA(node);
5708 osize = NODEDSZ(node);
5710 delta = sp->mp_upper - sp->mp_lower;
5711 SETDSZ(node, osize - delta);
5712 xp = (MDB_page *)((char *)sp + delta);
5714 /* shift subpage upward */
5716 nsize = NUMKEYS(sp) * sp->mp_pad;
5717 memmove(METADATA(xp), METADATA(sp), nsize);
5720 nsize = osize - sp->mp_upper;
5721 numkeys = NUMKEYS(sp);
5722 for (i=numkeys-1; i>=0; i--)
5723 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5725 xp->mp_upper = sp->mp_lower;
5726 xp->mp_lower = sp->mp_lower;
5727 xp->mp_flags = sp->mp_flags;
5728 xp->mp_pad = sp->mp_pad;
5729 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5731 /* shift lower nodes upward */
5732 ptr = mp->mp_ptrs[indx];
5733 numkeys = NUMKEYS(mp);
5734 for (i = 0; i < numkeys; i++) {
5735 if (mp->mp_ptrs[i] <= ptr)
5736 mp->mp_ptrs[i] += delta;
5739 base = (char *)mp + mp->mp_upper;
5740 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5741 mp->mp_upper += delta;
5744 /** Initial setup of a sorted-dups cursor.
5745 * Sorted duplicates are implemented as a sub-database for the given key.
5746 * The duplicate data items are actually keys of the sub-database.
5747 * Operations on the duplicate data items are performed using a sub-cursor
5748 * initialized when the sub-database is first accessed. This function does
5749 * the preliminary setup of the sub-cursor, filling in the fields that
5750 * depend only on the parent DB.
5751 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5754 mdb_xcursor_init0(MDB_cursor *mc)
5756 MDB_xcursor *mx = mc->mc_xcursor;
5758 mx->mx_cursor.mc_xcursor = NULL;
5759 mx->mx_cursor.mc_txn = mc->mc_txn;
5760 mx->mx_cursor.mc_db = &mx->mx_db;
5761 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5762 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5763 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5764 mx->mx_cursor.mc_snum = 0;
5765 mx->mx_cursor.mc_top = 0;
5766 mx->mx_cursor.mc_flags = C_SUB;
5767 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5768 mx->mx_dbx.md_dcmp = NULL;
5769 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5772 /** Final setup of a sorted-dups cursor.
5773 * Sets up the fields that depend on the data from the main cursor.
5774 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5775 * @param[in] node The data containing the #MDB_db record for the
5776 * sorted-dup database.
5779 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5781 MDB_xcursor *mx = mc->mc_xcursor;
5783 if (node->mn_flags & F_SUBDATA) {
5784 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5785 mx->mx_cursor.mc_pg[0] = 0;
5786 mx->mx_cursor.mc_snum = 0;
5787 mx->mx_cursor.mc_flags = C_SUB;
5789 MDB_page *fp = NODEDATA(node);
5790 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5791 mx->mx_db.md_flags = 0;
5792 mx->mx_db.md_depth = 1;
5793 mx->mx_db.md_branch_pages = 0;
5794 mx->mx_db.md_leaf_pages = 1;
5795 mx->mx_db.md_overflow_pages = 0;
5796 mx->mx_db.md_entries = NUMKEYS(fp);
5797 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5798 mx->mx_cursor.mc_snum = 1;
5799 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5800 mx->mx_cursor.mc_top = 0;
5801 mx->mx_cursor.mc_pg[0] = fp;
5802 mx->mx_cursor.mc_ki[0] = 0;
5803 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5804 mx->mx_db.md_flags = MDB_DUPFIXED;
5805 mx->mx_db.md_pad = fp->mp_pad;
5806 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5807 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5810 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5812 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5814 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5815 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5816 #if UINT_MAX < SIZE_MAX
5817 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5818 #ifdef MISALIGNED_OK
5819 mx->mx_dbx.md_cmp = mdb_cmp_long;
5821 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5826 /** Initialize a cursor for a given transaction and database. */
5828 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5833 mc->mc_db = &txn->mt_dbs[dbi];
5834 mc->mc_dbx = &txn->mt_dbxs[dbi];
5835 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5840 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5842 mc->mc_xcursor = mx;
5843 mdb_xcursor_init0(mc);
5845 mc->mc_xcursor = NULL;
5847 if (*mc->mc_dbflag & DB_STALE) {
5848 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5853 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5856 size_t size = sizeof(MDB_cursor);
5858 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5861 /* Allow read access to the freelist */
5862 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5865 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5866 size += sizeof(MDB_xcursor);
5868 if ((mc = malloc(size)) != NULL) {
5869 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
5870 if (txn->mt_cursors) {
5871 mc->mc_next = txn->mt_cursors[dbi];
5872 txn->mt_cursors[dbi] = mc;
5873 mc->mc_flags |= C_UNTRACK;
5875 mc->mc_flags |= C_ALLOCD;
5886 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5890 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5893 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
5896 flags = mc->mc_flags;
5898 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5900 mc->mc_flags |= (flags & C_ALLOCD);
5904 /* Return the count of duplicate data items for the current key */
5906 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5910 if (mc == NULL || countp == NULL)
5913 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5916 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5917 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5920 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5923 *countp = mc->mc_xcursor->mx_db.md_entries;
5929 mdb_cursor_close(MDB_cursor *mc)
5932 /* remove from txn, if tracked */
5933 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
5934 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5935 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5937 *prev = mc->mc_next;
5939 if (mc->mc_flags & C_ALLOCD)
5945 mdb_cursor_txn(MDB_cursor *mc)
5947 if (!mc) return NULL;
5952 mdb_cursor_dbi(MDB_cursor *mc)
5958 /** Replace the key for a node with a new key.
5959 * @param[in] mc Cursor pointing to the node to operate on.
5960 * @param[in] key The new key to use.
5961 * @return 0 on success, non-zero on failure.
5964 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5971 indx_t ptr, i, numkeys, indx;
5974 indx = mc->mc_ki[mc->mc_top];
5975 mp = mc->mc_pg[mc->mc_top];
5976 node = NODEPTR(mp, indx);
5977 ptr = mp->mp_ptrs[indx];
5981 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5982 k2.mv_data = NODEKEY(node);
5983 k2.mv_size = node->mn_ksize;
5984 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5986 mdb_dkey(&k2, kbuf2),
5992 delta0 = delta = key->mv_size - node->mn_ksize;
5994 /* Must be 2-byte aligned. If new key is
5995 * shorter by 1, the shift will be skipped.
5997 delta += (delta & 1);
5999 if (delta > 0 && SIZELEFT(mp) < delta) {
6001 /* not enough space left, do a delete and split */
6002 DPRINTF("Not enough room, delta = %d, splitting...", delta);
6003 pgno = NODEPGNO(node);
6004 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
6005 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6008 numkeys = NUMKEYS(mp);
6009 for (i = 0; i < numkeys; i++) {
6010 if (mp->mp_ptrs[i] <= ptr)
6011 mp->mp_ptrs[i] -= delta;
6014 base = (char *)mp + mp->mp_upper;
6015 len = ptr - mp->mp_upper + NODESIZE;
6016 memmove(base - delta, base, len);
6017 mp->mp_upper -= delta;
6019 node = NODEPTR(mp, indx);
6022 /* But even if no shift was needed, update ksize */
6024 node->mn_ksize = key->mv_size;
6027 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6033 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
6035 /** Move a node from csrc to cdst.
6038 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
6045 unsigned short flags;
6049 /* Mark src and dst as dirty. */
6050 if ((rc = mdb_page_touch(csrc)) ||
6051 (rc = mdb_page_touch(cdst)))
6054 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6055 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6056 key.mv_size = csrc->mc_db->md_pad;
6057 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6059 data.mv_data = NULL;
6063 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6064 assert(!((long)srcnode&1));
6065 srcpg = NODEPGNO(srcnode);
6066 flags = srcnode->mn_flags;
6067 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6068 unsigned int snum = csrc->mc_snum;
6070 /* must find the lowest key below src */
6071 mdb_page_search_lowest(csrc);
6072 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6073 key.mv_size = csrc->mc_db->md_pad;
6074 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6076 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6077 key.mv_size = NODEKSZ(s2);
6078 key.mv_data = NODEKEY(s2);
6080 csrc->mc_snum = snum--;
6081 csrc->mc_top = snum;
6083 key.mv_size = NODEKSZ(srcnode);
6084 key.mv_data = NODEKEY(srcnode);
6086 data.mv_size = NODEDSZ(srcnode);
6087 data.mv_data = NODEDATA(srcnode);
6089 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6090 unsigned int snum = cdst->mc_snum;
6093 /* must find the lowest key below dst */
6094 mdb_page_search_lowest(cdst);
6095 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6096 bkey.mv_size = cdst->mc_db->md_pad;
6097 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6099 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6100 bkey.mv_size = NODEKSZ(s2);
6101 bkey.mv_data = NODEKEY(s2);
6103 cdst->mc_snum = snum--;
6104 cdst->mc_top = snum;
6105 mdb_cursor_copy(cdst, &mn);
6107 rc = mdb_update_key(&mn, &bkey);
6112 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6113 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6114 csrc->mc_ki[csrc->mc_top],
6116 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6117 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6119 /* Add the node to the destination page.
6121 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6122 if (rc != MDB_SUCCESS)
6125 /* Delete the node from the source page.
6127 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6130 /* Adjust other cursors pointing to mp */
6131 MDB_cursor *m2, *m3;
6132 MDB_dbi dbi = csrc->mc_dbi;
6133 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6135 if (csrc->mc_flags & C_SUB)
6138 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6139 if (m2 == csrc) continue;
6140 if (csrc->mc_flags & C_SUB)
6141 m3 = &m2->mc_xcursor->mx_cursor;
6144 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6145 csrc->mc_ki[csrc->mc_top]) {
6146 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6147 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6152 /* Update the parent separators.
6154 if (csrc->mc_ki[csrc->mc_top] == 0) {
6155 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6156 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6157 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6159 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6160 key.mv_size = NODEKSZ(srcnode);
6161 key.mv_data = NODEKEY(srcnode);
6163 DPRINTF("update separator for source page %zu to [%s]",
6164 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6165 mdb_cursor_copy(csrc, &mn);
6168 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6171 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6173 indx_t ix = csrc->mc_ki[csrc->mc_top];
6174 nullkey.mv_size = 0;
6175 csrc->mc_ki[csrc->mc_top] = 0;
6176 rc = mdb_update_key(csrc, &nullkey);
6177 csrc->mc_ki[csrc->mc_top] = ix;
6178 assert(rc == MDB_SUCCESS);
6182 if (cdst->mc_ki[cdst->mc_top] == 0) {
6183 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6184 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6185 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6187 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6188 key.mv_size = NODEKSZ(srcnode);
6189 key.mv_data = NODEKEY(srcnode);
6191 DPRINTF("update separator for destination page %zu to [%s]",
6192 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6193 mdb_cursor_copy(cdst, &mn);
6196 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6199 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6201 indx_t ix = cdst->mc_ki[cdst->mc_top];
6202 nullkey.mv_size = 0;
6203 cdst->mc_ki[cdst->mc_top] = 0;
6204 rc = mdb_update_key(cdst, &nullkey);
6205 cdst->mc_ki[cdst->mc_top] = ix;
6206 assert(rc == MDB_SUCCESS);
6213 /** Merge one page into another.
6214 * The nodes from the page pointed to by \b csrc will
6215 * be copied to the page pointed to by \b cdst and then
6216 * the \b csrc page will be freed.
6217 * @param[in] csrc Cursor pointing to the source page.
6218 * @param[in] cdst Cursor pointing to the destination page.
6221 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6229 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6230 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6232 assert(csrc->mc_snum > 1); /* can't merge root page */
6233 assert(cdst->mc_snum > 1);
6235 /* Mark dst as dirty. */
6236 if ((rc = mdb_page_touch(cdst)))
6239 /* Move all nodes from src to dst.
6241 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6242 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6243 key.mv_size = csrc->mc_db->md_pad;
6244 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6245 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6246 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6247 if (rc != MDB_SUCCESS)
6249 key.mv_data = (char *)key.mv_data + key.mv_size;
6252 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6253 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6254 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6255 unsigned int snum = csrc->mc_snum;
6257 /* must find the lowest key below src */
6258 mdb_page_search_lowest(csrc);
6259 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6260 key.mv_size = csrc->mc_db->md_pad;
6261 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6263 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6264 key.mv_size = NODEKSZ(s2);
6265 key.mv_data = NODEKEY(s2);
6267 csrc->mc_snum = snum--;
6268 csrc->mc_top = snum;
6270 key.mv_size = srcnode->mn_ksize;
6271 key.mv_data = NODEKEY(srcnode);
6274 data.mv_size = NODEDSZ(srcnode);
6275 data.mv_data = NODEDATA(srcnode);
6276 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6277 if (rc != MDB_SUCCESS)
6282 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6283 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);
6285 /* Unlink the src page from parent and add to free list.
6287 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6288 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6291 rc = mdb_update_key(csrc, &key);
6297 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6298 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6299 csrc->mc_db->md_leaf_pages--;
6301 csrc->mc_db->md_branch_pages--;
6303 /* Adjust other cursors pointing to mp */
6304 MDB_cursor *m2, *m3;
6305 MDB_dbi dbi = csrc->mc_dbi;
6306 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6308 if (csrc->mc_flags & C_SUB)
6311 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6312 if (csrc->mc_flags & C_SUB)
6313 m3 = &m2->mc_xcursor->mx_cursor;
6316 if (m3 == csrc) continue;
6317 if (m3->mc_snum < csrc->mc_snum) continue;
6318 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6319 m3->mc_pg[csrc->mc_top] = mp;
6320 m3->mc_ki[csrc->mc_top] += nkeys;
6324 mdb_cursor_pop(csrc);
6326 return mdb_rebalance(csrc);
6329 /** Copy the contents of a cursor.
6330 * @param[in] csrc The cursor to copy from.
6331 * @param[out] cdst The cursor to copy to.
6334 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6338 cdst->mc_txn = csrc->mc_txn;
6339 cdst->mc_dbi = csrc->mc_dbi;
6340 cdst->mc_db = csrc->mc_db;
6341 cdst->mc_dbx = csrc->mc_dbx;
6342 cdst->mc_snum = csrc->mc_snum;
6343 cdst->mc_top = csrc->mc_top;
6344 cdst->mc_flags = csrc->mc_flags;
6346 for (i=0; i<csrc->mc_snum; i++) {
6347 cdst->mc_pg[i] = csrc->mc_pg[i];
6348 cdst->mc_ki[i] = csrc->mc_ki[i];
6352 /** Rebalance the tree after a delete operation.
6353 * @param[in] mc Cursor pointing to the page where rebalancing
6355 * @return 0 on success, non-zero on failure.
6358 mdb_rebalance(MDB_cursor *mc)
6362 unsigned int ptop, minkeys;
6365 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6369 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6370 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6371 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6372 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6376 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6377 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6380 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6381 DPRINTF("no need to rebalance page %zu, above fill threshold",
6387 if (mc->mc_snum < 2) {
6388 MDB_page *mp = mc->mc_pg[0];
6390 DPUTS("Can't rebalance a subpage, ignoring");
6393 if (NUMKEYS(mp) == 0) {
6394 DPUTS("tree is completely empty");
6395 mc->mc_db->md_root = P_INVALID;
6396 mc->mc_db->md_depth = 0;
6397 mc->mc_db->md_leaf_pages = 0;
6398 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6402 /* Adjust other cursors pointing to mp */
6403 MDB_cursor *m2, *m3;
6404 MDB_dbi dbi = mc->mc_dbi;
6406 if (mc->mc_flags & C_SUB)
6409 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6410 if (m2 == mc) continue;
6411 if (mc->mc_flags & C_SUB)
6412 m3 = &m2->mc_xcursor->mx_cursor;
6415 if (m3->mc_snum < mc->mc_snum) continue;
6416 if (m3->mc_pg[0] == mp) {
6422 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6423 DPUTS("collapsing root page!");
6424 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6425 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6426 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
6429 mc->mc_db->md_depth--;
6430 mc->mc_db->md_branch_pages--;
6432 /* Adjust other cursors pointing to mp */
6433 MDB_cursor *m2, *m3;
6434 MDB_dbi dbi = mc->mc_dbi;
6436 if (mc->mc_flags & C_SUB)
6439 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6440 if (m2 == mc) continue;
6441 if (mc->mc_flags & C_SUB)
6442 m3 = &m2->mc_xcursor->mx_cursor;
6445 if (m3->mc_snum < mc->mc_snum) continue;
6446 if (m3->mc_pg[0] == mp) {
6447 m3->mc_pg[0] = mc->mc_pg[0];
6454 DPUTS("root page doesn't need rebalancing");
6458 /* The parent (branch page) must have at least 2 pointers,
6459 * otherwise the tree is invalid.
6461 ptop = mc->mc_top-1;
6462 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6464 /* Leaf page fill factor is below the threshold.
6465 * Try to move keys from left or right neighbor, or
6466 * merge with a neighbor page.
6471 mdb_cursor_copy(mc, &mn);
6472 mn.mc_xcursor = NULL;
6474 if (mc->mc_ki[ptop] == 0) {
6475 /* We're the leftmost leaf in our parent.
6477 DPUTS("reading right neighbor");
6479 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6480 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6483 mn.mc_ki[mn.mc_top] = 0;
6484 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6486 /* There is at least one neighbor to the left.
6488 DPUTS("reading left neighbor");
6490 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6491 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
6494 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6495 mc->mc_ki[mc->mc_top] = 0;
6498 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6499 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);
6501 /* If the neighbor page is above threshold and has enough keys,
6502 * move one key from it. Otherwise we should try to merge them.
6503 * (A branch page must never have less than 2 keys.)
6505 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6506 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6507 return mdb_node_move(&mn, mc);
6509 if (mc->mc_ki[ptop] == 0)
6510 rc = mdb_page_merge(&mn, mc);
6512 rc = mdb_page_merge(mc, &mn);
6513 mc->mc_flags &= ~C_INITIALIZED;
6518 /** Complete a delete operation started by #mdb_cursor_del(). */
6520 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6524 /* add overflow pages to free list */
6525 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6530 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6531 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) != 0)
6533 assert(IS_OVERFLOW(omp));
6534 ovpages = omp->mp_pages;
6535 mc->mc_db->md_overflow_pages -= ovpages;
6536 for (i=0; i<ovpages; i++) {
6537 DPRINTF("freed ov page %zu", pg);
6538 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6542 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6543 mc->mc_db->md_entries--;
6544 rc = mdb_rebalance(mc);
6545 if (rc != MDB_SUCCESS)
6546 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6547 /* if mc points past last node in page, invalidate */
6548 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6549 mc->mc_flags &= ~C_INITIALIZED;
6555 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6556 MDB_val *key, MDB_val *data)
6561 MDB_val rdata, *xdata;
6565 assert(key != NULL);
6567 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6569 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6572 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6576 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6580 mdb_cursor_init(&mc, txn, dbi, &mx);
6591 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6593 /* let mdb_page_split know about this cursor if needed:
6594 * delete will trigger a rebalance; if it needs to move
6595 * a node from one page to another, it will have to
6596 * update the parent's separator key(s). If the new sepkey
6597 * is larger than the current one, the parent page may
6598 * run out of space, triggering a split. We need this
6599 * cursor to be consistent until the end of the rebalance.
6601 mc.mc_next = txn->mt_cursors[dbi];
6602 txn->mt_cursors[dbi] = &mc;
6603 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6604 txn->mt_cursors[dbi] = mc.mc_next;
6609 /** Split a page and insert a new node.
6610 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6611 * The cursor will be updated to point to the actual page and index where
6612 * the node got inserted after the split.
6613 * @param[in] newkey The key for the newly inserted node.
6614 * @param[in] newdata The data for the newly inserted node.
6615 * @param[in] newpgno The page number, if the new node is a branch node.
6616 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6617 * @return 0 on success, non-zero on failure.
6620 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6621 unsigned int nflags)
6624 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6627 unsigned int i, j, split_indx, nkeys, pmax;
6629 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6631 MDB_page *mp, *rp, *pp;
6636 mp = mc->mc_pg[mc->mc_top];
6637 newindx = mc->mc_ki[mc->mc_top];
6639 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6640 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6641 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6643 /* Create a right sibling. */
6644 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6646 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6648 if (mc->mc_snum < 2) {
6649 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6651 /* shift current top to make room for new parent */
6652 mc->mc_pg[1] = mc->mc_pg[0];
6653 mc->mc_ki[1] = mc->mc_ki[0];
6656 mc->mc_db->md_root = pp->mp_pgno;
6657 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6658 mc->mc_db->md_depth++;
6661 /* Add left (implicit) pointer. */
6662 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6663 /* undo the pre-push */
6664 mc->mc_pg[0] = mc->mc_pg[1];
6665 mc->mc_ki[0] = mc->mc_ki[1];
6666 mc->mc_db->md_root = mp->mp_pgno;
6667 mc->mc_db->md_depth--;
6674 ptop = mc->mc_top-1;
6675 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6678 mc->mc_flags |= C_SPLITTING;
6679 mdb_cursor_copy(mc, &mn);
6680 mn.mc_pg[mn.mc_top] = rp;
6681 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6683 if (nflags & MDB_APPEND) {
6684 mn.mc_ki[mn.mc_top] = 0;
6686 split_indx = newindx;
6691 nkeys = NUMKEYS(mp);
6692 split_indx = nkeys / 2;
6693 if (newindx < split_indx)
6699 unsigned int lsize, rsize, ksize;
6700 /* Move half of the keys to the right sibling */
6702 x = mc->mc_ki[mc->mc_top] - split_indx;
6703 ksize = mc->mc_db->md_pad;
6704 split = LEAF2KEY(mp, split_indx, ksize);
6705 rsize = (nkeys - split_indx) * ksize;
6706 lsize = (nkeys - split_indx) * sizeof(indx_t);
6707 mp->mp_lower -= lsize;
6708 rp->mp_lower += lsize;
6709 mp->mp_upper += rsize - lsize;
6710 rp->mp_upper -= rsize - lsize;
6711 sepkey.mv_size = ksize;
6712 if (newindx == split_indx) {
6713 sepkey.mv_data = newkey->mv_data;
6715 sepkey.mv_data = split;
6718 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6719 memcpy(rp->mp_ptrs, split, rsize);
6720 sepkey.mv_data = rp->mp_ptrs;
6721 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6722 memcpy(ins, newkey->mv_data, ksize);
6723 mp->mp_lower += sizeof(indx_t);
6724 mp->mp_upper -= ksize - sizeof(indx_t);
6727 memcpy(rp->mp_ptrs, split, x * ksize);
6728 ins = LEAF2KEY(rp, x, ksize);
6729 memcpy(ins, newkey->mv_data, ksize);
6730 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6731 rp->mp_lower += sizeof(indx_t);
6732 rp->mp_upper -= ksize - sizeof(indx_t);
6733 mc->mc_ki[mc->mc_top] = x;
6734 mc->mc_pg[mc->mc_top] = rp;
6739 /* For leaf pages, check the split point based on what
6740 * fits where, since otherwise mdb_node_add can fail.
6742 * This check is only needed when the data items are
6743 * relatively large, such that being off by one will
6744 * make the difference between success or failure.
6746 * It's also relevant if a page happens to be laid out
6747 * such that one half of its nodes are all "small" and
6748 * the other half of its nodes are "large." If the new
6749 * item is also "large" and falls on the half with
6750 * "large" nodes, it also may not fit.
6753 unsigned int psize, nsize;
6754 /* Maximum free space in an empty page */
6755 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6756 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6757 if ((nkeys < 20) || (nsize > pmax/16)) {
6758 if (newindx <= split_indx) {
6761 for (i=0; i<split_indx; i++) {
6762 node = NODEPTR(mp, i);
6763 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6764 if (F_ISSET(node->mn_flags, F_BIGDATA))
6765 psize += sizeof(pgno_t);
6767 psize += NODEDSZ(node);
6771 split_indx = newindx;
6782 for (i=nkeys-1; i>=split_indx; i--) {
6783 node = NODEPTR(mp, i);
6784 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6785 if (F_ISSET(node->mn_flags, F_BIGDATA))
6786 psize += sizeof(pgno_t);
6788 psize += NODEDSZ(node);
6792 split_indx = newindx;
6803 /* First find the separating key between the split pages.
6804 * The case where newindx == split_indx is ambiguous; the
6805 * new item could go to the new page or stay on the original
6806 * page. If newpos == 1 it goes to the new page.
6808 if (newindx == split_indx && newpos) {
6809 sepkey.mv_size = newkey->mv_size;
6810 sepkey.mv_data = newkey->mv_data;
6812 node = NODEPTR(mp, split_indx);
6813 sepkey.mv_size = node->mn_ksize;
6814 sepkey.mv_data = NODEKEY(node);
6818 DPRINTF("separator is [%s]", DKEY(&sepkey));
6820 /* Copy separator key to the parent.
6822 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6826 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6829 if (mn.mc_snum == mc->mc_snum) {
6830 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6831 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6832 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6833 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6838 /* Right page might now have changed parent.
6839 * Check if left page also changed parent.
6841 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6842 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6843 for (i=0; i<ptop; i++) {
6844 mc->mc_pg[i] = mn.mc_pg[i];
6845 mc->mc_ki[i] = mn.mc_ki[i];
6847 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6848 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6852 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6855 mc->mc_flags ^= C_SPLITTING;
6856 if (rc != MDB_SUCCESS) {
6859 if (nflags & MDB_APPEND) {
6860 mc->mc_pg[mc->mc_top] = rp;
6861 mc->mc_ki[mc->mc_top] = 0;
6862 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6865 for (i=0; i<mc->mc_top; i++)
6866 mc->mc_ki[i] = mn.mc_ki[i];
6873 /* Move half of the keys to the right sibling. */
6875 /* grab a page to hold a temporary copy */
6876 copy = mdb_page_malloc(mc, 1);
6880 copy->mp_pgno = mp->mp_pgno;
6881 copy->mp_flags = mp->mp_flags;
6882 copy->mp_lower = PAGEHDRSZ;
6883 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6884 mc->mc_pg[mc->mc_top] = copy;
6885 for (i = j = 0; i <= nkeys; j++) {
6886 if (i == split_indx) {
6887 /* Insert in right sibling. */
6888 /* Reset insert index for right sibling. */
6889 if (i != newindx || (newpos ^ ins_new)) {
6891 mc->mc_pg[mc->mc_top] = rp;
6895 if (i == newindx && !ins_new) {
6896 /* Insert the original entry that caused the split. */
6897 rkey.mv_data = newkey->mv_data;
6898 rkey.mv_size = newkey->mv_size;
6907 /* Update index for the new key. */
6908 mc->mc_ki[mc->mc_top] = j;
6909 } else if (i == nkeys) {
6912 node = NODEPTR(mp, i);
6913 rkey.mv_data = NODEKEY(node);
6914 rkey.mv_size = node->mn_ksize;
6916 xdata.mv_data = NODEDATA(node);
6917 xdata.mv_size = NODEDSZ(node);
6920 pgno = NODEPGNO(node);
6921 flags = node->mn_flags;
6926 if (!IS_LEAF(mp) && j == 0) {
6927 /* First branch index doesn't need key data. */
6931 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6935 nkeys = NUMKEYS(copy);
6936 for (i=0; i<nkeys; i++)
6937 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6938 mp->mp_lower = copy->mp_lower;
6939 mp->mp_upper = copy->mp_upper;
6940 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6941 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6943 /* reset back to original page */
6944 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6945 mc->mc_pg[mc->mc_top] = mp;
6946 if (nflags & MDB_RESERVE) {
6947 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6948 if (!(node->mn_flags & F_BIGDATA))
6949 newdata->mv_data = NODEDATA(node);
6953 /* Make sure mc_ki is still valid.
6955 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6956 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6957 for (i=0; i<ptop; i++) {
6958 mc->mc_pg[i] = mn.mc_pg[i];
6959 mc->mc_ki[i] = mn.mc_ki[i];
6961 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6962 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6966 /* return tmp page to freelist */
6967 mdb_page_free(mc->mc_txn->mt_env, copy);
6970 /* Adjust other cursors pointing to mp */
6971 MDB_cursor *m2, *m3;
6972 MDB_dbi dbi = mc->mc_dbi;
6973 int fixup = NUMKEYS(mp);
6975 if (mc->mc_flags & C_SUB)
6978 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6979 if (m2 == mc) continue;
6980 if (mc->mc_flags & C_SUB)
6981 m3 = &m2->mc_xcursor->mx_cursor;
6984 if (!(m3->mc_flags & C_INITIALIZED))
6986 if (m3->mc_flags & C_SPLITTING)
6991 for (k=m3->mc_top; k>=0; k--) {
6992 m3->mc_ki[k+1] = m3->mc_ki[k];
6993 m3->mc_pg[k+1] = m3->mc_pg[k];
6995 if (m3->mc_ki[0] >= split_indx) {
7000 m3->mc_pg[0] = mc->mc_pg[0];
7004 if (m3->mc_pg[mc->mc_top] == mp) {
7005 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
7006 m3->mc_ki[mc->mc_top]++;
7007 if (m3->mc_ki[mc->mc_top] >= fixup) {
7008 m3->mc_pg[mc->mc_top] = rp;
7009 m3->mc_ki[mc->mc_top] -= fixup;
7010 m3->mc_ki[ptop] = mn.mc_ki[ptop];
7012 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
7013 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
7022 mdb_put(MDB_txn *txn, MDB_dbi dbi,
7023 MDB_val *key, MDB_val *data, unsigned int flags)
7028 assert(key != NULL);
7029 assert(data != NULL);
7031 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7034 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
7038 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
7042 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
7045 mdb_cursor_init(&mc, txn, dbi, &mx);
7046 return mdb_cursor_put(&mc, key, data, flags);
7050 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
7052 if ((flag & CHANGEABLE) != flag)
7055 env->me_flags |= flag;
7057 env->me_flags &= ~flag;
7062 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7067 *arg = env->me_flags;
7072 mdb_env_get_path(MDB_env *env, const char **arg)
7077 *arg = env->me_path;
7081 /** Common code for #mdb_stat() and #mdb_env_stat().
7082 * @param[in] env the environment to operate in.
7083 * @param[in] db the #MDB_db record containing the stats to return.
7084 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7085 * @return 0, this function always succeeds.
7088 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7090 arg->ms_psize = env->me_psize;
7091 arg->ms_depth = db->md_depth;
7092 arg->ms_branch_pages = db->md_branch_pages;
7093 arg->ms_leaf_pages = db->md_leaf_pages;
7094 arg->ms_overflow_pages = db->md_overflow_pages;
7095 arg->ms_entries = db->md_entries;
7100 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7104 if (env == NULL || arg == NULL)
7107 toggle = mdb_env_pick_meta(env);
7109 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7113 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7117 if (env == NULL || arg == NULL)
7120 toggle = mdb_env_pick_meta(env);
7121 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7122 arg->me_mapsize = env->me_mapsize;
7123 arg->me_maxreaders = env->me_maxreaders;
7124 arg->me_numreaders = env->me_numreaders;
7125 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7126 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7130 /** Set the default comparison functions for a database.
7131 * Called immediately after a database is opened to set the defaults.
7132 * The user can then override them with #mdb_set_compare() or
7133 * #mdb_set_dupsort().
7134 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7135 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7138 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7140 uint16_t f = txn->mt_dbs[dbi].md_flags;
7142 txn->mt_dbxs[dbi].md_cmp =
7143 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7144 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7146 txn->mt_dbxs[dbi].md_dcmp =
7147 !(f & MDB_DUPSORT) ? 0 :
7148 ((f & MDB_INTEGERDUP)
7149 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7150 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7153 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7158 int rc, dbflag, exact;
7159 unsigned int unused = 0;
7162 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7163 mdb_default_cmp(txn, FREE_DBI);
7166 if ((flags & VALID_FLAGS) != flags)
7172 if (flags & PERSISTENT_FLAGS) {
7173 uint16_t f2 = flags & PERSISTENT_FLAGS;
7174 /* make sure flag changes get committed */
7175 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7176 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7177 txn->mt_flags |= MDB_TXN_DIRTY;
7180 mdb_default_cmp(txn, MAIN_DBI);
7184 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7185 mdb_default_cmp(txn, MAIN_DBI);
7188 /* Is the DB already open? */
7190 for (i=2; i<txn->mt_numdbs; i++) {
7191 if (!txn->mt_dbxs[i].md_name.mv_size) {
7192 /* Remember this free slot */
7193 if (!unused) unused = i;
7196 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7197 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7203 /* If no free slot and max hit, fail */
7204 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7205 return MDB_DBS_FULL;
7207 /* Find the DB info */
7208 dbflag = DB_NEW|DB_VALID;
7211 key.mv_data = (void *)name;
7212 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7213 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7214 if (rc == MDB_SUCCESS) {
7215 /* make sure this is actually a DB */
7216 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7217 if (!(node->mn_flags & F_SUBDATA))
7219 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7220 /* Create if requested */
7222 data.mv_size = sizeof(MDB_db);
7223 data.mv_data = &dummy;
7224 memset(&dummy, 0, sizeof(dummy));
7225 dummy.md_root = P_INVALID;
7226 dummy.md_flags = flags & PERSISTENT_FLAGS;
7227 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7231 /* OK, got info, add to table */
7232 if (rc == MDB_SUCCESS) {
7233 unsigned int slot = unused ? unused : txn->mt_numdbs;
7234 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7235 txn->mt_dbxs[slot].md_name.mv_size = len;
7236 txn->mt_dbxs[slot].md_rel = NULL;
7237 txn->mt_dbflags[slot] = dbflag;
7238 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7240 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7241 mdb_default_cmp(txn, slot);
7250 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7252 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7255 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7258 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7261 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7263 ptr = env->me_dbxs[dbi].md_name.mv_data;
7264 env->me_dbxs[dbi].md_name.mv_data = NULL;
7265 env->me_dbxs[dbi].md_name.mv_size = 0;
7266 env->me_dbflags[dbi] = 0;
7270 /** Add all the DB's pages to the free list.
7271 * @param[in] mc Cursor on the DB to free.
7272 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7273 * @return 0 on success, non-zero on failure.
7276 mdb_drop0(MDB_cursor *mc, int subs)
7280 rc = mdb_page_search(mc, NULL, 0);
7281 if (rc == MDB_SUCCESS) {
7286 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7287 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7290 mdb_cursor_copy(mc, &mx);
7291 while (mc->mc_snum > 0) {
7292 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7293 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7294 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7295 if (ni->mn_flags & F_BIGDATA) {
7299 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7300 rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL);
7303 assert(IS_OVERFLOW(omp));
7304 ovpages = omp->mp_pages;
7305 for (j=0; j<ovpages; j++) {
7306 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7309 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7310 mdb_xcursor_init1(mc, ni);
7311 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7317 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7319 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7322 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7327 mc->mc_ki[mc->mc_top] = i;
7328 rc = mdb_cursor_sibling(mc, 1);
7330 /* no more siblings, go back to beginning
7331 * of previous level.
7335 for (i=1; i<mc->mc_snum; i++) {
7337 mc->mc_pg[i] = mx.mc_pg[i];
7342 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7343 mc->mc_db->md_root);
7348 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7353 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7356 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7359 rc = mdb_cursor_open(txn, dbi, &mc);
7363 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7367 /* Can't delete the main DB */
7368 if (del && dbi > MAIN_DBI) {
7369 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7371 txn->mt_dbflags[dbi] = DB_STALE;
7372 mdb_dbi_close(txn->mt_env, dbi);
7375 /* reset the DB record, mark it dirty */
7376 txn->mt_dbflags[dbi] |= DB_DIRTY;
7377 txn->mt_dbs[dbi].md_depth = 0;
7378 txn->mt_dbs[dbi].md_branch_pages = 0;
7379 txn->mt_dbs[dbi].md_leaf_pages = 0;
7380 txn->mt_dbs[dbi].md_overflow_pages = 0;
7381 txn->mt_dbs[dbi].md_entries = 0;
7382 txn->mt_dbs[dbi].md_root = P_INVALID;
7384 txn->mt_flags |= MDB_TXN_DIRTY;
7387 mdb_cursor_close(mc);
7391 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7393 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7396 txn->mt_dbxs[dbi].md_cmp = cmp;
7400 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7402 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7405 txn->mt_dbxs[dbi].md_dcmp = cmp;
7409 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7411 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7414 txn->mt_dbxs[dbi].md_rel = rel;
7418 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7420 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7423 txn->mt_dbxs[dbi].md_relctx = ctx;