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 <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
65 #if defined(__APPLE__) || defined (BSD)
66 # define MDB_USE_POSIX_SEM 1
67 # define MDB_FDATASYNC fsync
68 #elif defined(ANDROID)
69 # define MDB_FDATASYNC fsync
74 #ifdef MDB_USE_POSIX_SEM
75 #include <semaphore.h>
80 #include <valgrind/memcheck.h>
81 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
82 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
83 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
84 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
85 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
87 #define VGMEMP_CREATE(h,r,z)
88 #define VGMEMP_ALLOC(h,a,s)
89 #define VGMEMP_FREE(h,a)
90 #define VGMEMP_DESTROY(h)
91 #define VGMEMP_DEFINED(a,s)
95 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
96 /* Solaris just defines one or the other */
97 # define LITTLE_ENDIAN 1234
98 # define BIG_ENDIAN 4321
99 # ifdef _LITTLE_ENDIAN
100 # define BYTE_ORDER LITTLE_ENDIAN
102 # define BYTE_ORDER BIG_ENDIAN
105 # define BYTE_ORDER __BYTE_ORDER
109 #ifndef LITTLE_ENDIAN
110 #define LITTLE_ENDIAN __LITTLE_ENDIAN
113 #define BIG_ENDIAN __BIG_ENDIAN
116 #if defined(__i386) || defined(__x86_64)
117 #define MISALIGNED_OK 1
123 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
124 # error "Unknown or unsupported endianness (BYTE_ORDER)"
125 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
126 # error "Two's complement, reasonably sized integer types, please"
129 /** @defgroup internal MDB Internals
132 /** @defgroup compat Windows Compatibility Macros
133 * A bunch of macros to minimize the amount of platform-specific ifdefs
134 * needed throughout the rest of the code. When the features this library
135 * needs are similar enough to POSIX to be hidden in a one-or-two line
136 * replacement, this macro approach is used.
140 #define pthread_t DWORD
141 #define pthread_mutex_t HANDLE
142 #define pthread_key_t DWORD
143 #define pthread_self() GetCurrentThreadId()
144 #define pthread_key_create(x,y) \
145 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
146 #define pthread_key_delete(x) TlsFree(x)
147 #define pthread_getspecific(x) TlsGetValue(x)
148 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
149 #define pthread_mutex_unlock(x) ReleaseMutex(x)
150 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
151 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
152 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
153 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
154 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
155 #define getpid() GetCurrentProcessId()
156 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
157 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
158 #define ErrCode() GetLastError()
159 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
160 #define close(fd) CloseHandle(fd)
161 #define munmap(ptr,len) UnmapViewOfFile(ptr)
164 #ifdef MDB_USE_POSIX_SEM
166 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
167 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
168 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
169 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
172 mdb_sem_wait(sem_t *sem)
175 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
180 /** Lock the reader mutex.
182 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
183 /** Unlock the reader mutex.
185 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
187 /** Lock the writer mutex.
188 * Only a single write transaction is allowed at a time. Other writers
189 * will block waiting for this mutex.
191 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
192 /** Unlock the writer mutex.
194 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
195 #endif /* MDB_USE_POSIX_SEM */
197 /** Get the error code for the last failed system function.
199 #define ErrCode() errno
201 /** An abstraction for a file handle.
202 * On POSIX systems file handles are small integers. On Windows
203 * they're opaque pointers.
207 /** A value for an invalid file handle.
208 * Mainly used to initialize file variables and signify that they are
211 #define INVALID_HANDLE_VALUE (-1)
213 /** Get the size of a memory page for the system.
214 * This is the basic size that the platform's memory manager uses, and is
215 * fundamental to the use of memory-mapped files.
217 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
220 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
223 #define MNAME_LEN (sizeof(pthread_mutex_t))
229 /** A flag for opening a file and requesting synchronous data writes.
230 * This is only used when writing a meta page. It's not strictly needed;
231 * we could just do a normal write and then immediately perform a flush.
232 * But if this flag is available it saves us an extra system call.
234 * @note If O_DSYNC is undefined but exists in /usr/include,
235 * preferably set some compiler flag to get the definition.
236 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
239 # define MDB_DSYNC O_DSYNC
243 /** Function for flushing the data of a file. Define this to fsync
244 * if fdatasync() is not supported.
246 #ifndef MDB_FDATASYNC
247 # define MDB_FDATASYNC fdatasync
251 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
262 /** A page number in the database.
263 * Note that 64 bit page numbers are overkill, since pages themselves
264 * already represent 12-13 bits of addressable memory, and the OS will
265 * always limit applications to a maximum of 63 bits of address space.
267 * @note In the #MDB_node structure, we only store 48 bits of this value,
268 * which thus limits us to only 60 bits of addressable data.
270 typedef MDB_ID pgno_t;
272 /** A transaction ID.
273 * See struct MDB_txn.mt_txnid for details.
275 typedef MDB_ID txnid_t;
277 /** @defgroup debug Debug Macros
281 /** Enable debug output.
282 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
283 * read from and written to the database (used for free space management).
288 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
291 # define DPRINTF (void) /* Vararg macros may be unsupported */
293 static int mdb_debug;
294 static txnid_t mdb_debug_start;
296 /** Print a debug message with printf formatting. */
297 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
298 ((void) ((mdb_debug) && \
299 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
301 # define DPRINTF(fmt, ...) ((void) 0)
302 # define MDB_DEBUG_SKIP
304 /** Print a debug string.
305 * The string is printed literally, with no format processing.
307 #define DPUTS(arg) DPRINTF("%s", arg)
310 /** A default memory page size.
311 * The actual size is platform-dependent, but we use this for
312 * boot-strapping. We probably should not be using this any more.
313 * The #GET_PAGESIZE() macro is used to get the actual size.
315 * Note that we don't currently support Huge pages. On Linux,
316 * regular data files cannot use Huge pages, and in general
317 * Huge pages aren't actually pageable. We rely on the OS
318 * demand-pager to read our data and page it out when memory
319 * pressure from other processes is high. So until OSs have
320 * actual paging support for Huge pages, they're not viable.
322 #define MDB_PAGESIZE 4096
324 /** The minimum number of keys required in a database page.
325 * Setting this to a larger value will place a smaller bound on the
326 * maximum size of a data item. Data items larger than this size will
327 * be pushed into overflow pages instead of being stored directly in
328 * the B-tree node. This value used to default to 4. With a page size
329 * of 4096 bytes that meant that any item larger than 1024 bytes would
330 * go into an overflow page. That also meant that on average 2-3KB of
331 * each overflow page was wasted space. The value cannot be lower than
332 * 2 because then there would no longer be a tree structure. With this
333 * value, items larger than 2KB will go into overflow pages, and on
334 * average only 1KB will be wasted.
336 #define MDB_MINKEYS 2
338 /** A stamp that identifies a file as an MDB file.
339 * There's nothing special about this value other than that it is easily
340 * recognizable, and it will reflect any byte order mismatches.
342 #define MDB_MAGIC 0xBEEFC0DE
344 /** The version number for a database's file format. */
345 #define MDB_VERSION 1
347 /** @brief The maximum size of a key in the database.
349 * We require that keys all fit onto a regular page. This limit
350 * could be raised a bit further if needed; to something just
351 * under #MDB_PAGESIZE / #MDB_MINKEYS.
353 * Note that data items in an #MDB_DUPSORT database are actually keys
354 * of a subDB, so they're also limited to this size.
356 #ifndef MDB_MAXKEYSIZE
357 #define MDB_MAXKEYSIZE 511
360 /** @brief The maximum size of a data item.
362 * We only store a 32 bit value for node sizes.
364 #define MAXDATASIZE 0xffffffffUL
369 * This is used for printing a hex dump of a key's contents.
371 #define DKBUF char kbuf[(MDB_MAXKEYSIZE*2+1)]
372 /** Display a key in hex.
374 * Invoke a function to display a key in hex.
376 #define DKEY(x) mdb_dkey(x, kbuf)
378 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
382 /** An invalid page number.
383 * Mainly used to denote an empty tree.
385 #define P_INVALID (~(pgno_t)0)
387 /** Test if the flags \b f are set in a flag word \b w. */
388 #define F_ISSET(w, f) (((w) & (f)) == (f))
390 /** Used for offsets within a single page.
391 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
394 typedef uint16_t indx_t;
396 /** Default size of memory map.
397 * This is certainly too small for any actual applications. Apps should always set
398 * the size explicitly using #mdb_env_set_mapsize().
400 #define DEFAULT_MAPSIZE 1048576
402 /** @defgroup readers Reader Lock Table
403 * Readers don't acquire any locks for their data access. Instead, they
404 * simply record their transaction ID in the reader table. The reader
405 * mutex is needed just to find an empty slot in the reader table. The
406 * slot's address is saved in thread-specific data so that subsequent read
407 * transactions started by the same thread need no further locking to proceed.
409 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
411 * No reader table is used if the database is on a read-only filesystem.
413 * Since the database uses multi-version concurrency control, readers don't
414 * actually need any locking. This table is used to keep track of which
415 * readers are using data from which old transactions, so that we'll know
416 * when a particular old transaction is no longer in use. Old transactions
417 * that have discarded any data pages can then have those pages reclaimed
418 * for use by a later write transaction.
420 * The lock table is constructed such that reader slots are aligned with the
421 * processor's cache line size. Any slot is only ever used by one thread.
422 * This alignment guarantees that there will be no contention or cache
423 * thrashing as threads update their own slot info, and also eliminates
424 * any need for locking when accessing a slot.
426 * A writer thread will scan every slot in the table to determine the oldest
427 * outstanding reader transaction. Any freed pages older than this will be
428 * reclaimed by the writer. The writer doesn't use any locks when scanning
429 * this table. This means that there's no guarantee that the writer will
430 * see the most up-to-date reader info, but that's not required for correct
431 * operation - all we need is to know the upper bound on the oldest reader,
432 * we don't care at all about the newest reader. So the only consequence of
433 * reading stale information here is that old pages might hang around a
434 * while longer before being reclaimed. That's actually good anyway, because
435 * the longer we delay reclaiming old pages, the more likely it is that a
436 * string of contiguous pages can be found after coalescing old pages from
437 * many old transactions together.
440 /** Number of slots in the reader table.
441 * This value was chosen somewhat arbitrarily. 126 readers plus a
442 * couple mutexes fit exactly into 8KB on my development machine.
443 * Applications should set the table size using #mdb_env_set_maxreaders().
445 #define DEFAULT_READERS 126
447 /** The size of a CPU cache line in bytes. We want our lock structures
448 * aligned to this size to avoid false cache line sharing in the
450 * This value works for most CPUs. For Itanium this should be 128.
456 /** The information we store in a single slot of the reader table.
457 * In addition to a transaction ID, we also record the process and
458 * thread ID that owns a slot, so that we can detect stale information,
459 * e.g. threads or processes that went away without cleaning up.
460 * @note We currently don't check for stale records. We simply re-init
461 * the table when we know that we're the only process opening the
464 typedef struct MDB_rxbody {
465 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
466 * Multiple readers that start at the same time will probably have the
467 * same ID here. Again, it's not important to exclude them from
468 * anything; all we need to know is which version of the DB they
469 * started from so we can avoid overwriting any data used in that
470 * particular version.
473 /** The process ID of the process owning this reader txn. */
475 /** The thread ID of the thread owning this txn. */
479 /** The actual reader record, with cacheline padding. */
480 typedef struct MDB_reader {
483 /** shorthand for mrb_txnid */
484 #define mr_txnid mru.mrx.mrb_txnid
485 #define mr_pid mru.mrx.mrb_pid
486 #define mr_tid mru.mrx.mrb_tid
487 /** cache line alignment */
488 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
492 /** The header for the reader table.
493 * The table resides in a memory-mapped file. (This is a different file
494 * than is used for the main database.)
496 * For POSIX the actual mutexes reside in the shared memory of this
497 * mapped file. On Windows, mutexes are named objects allocated by the
498 * kernel; we store the mutex names in this mapped file so that other
499 * processes can grab them. This same approach is also used on
500 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
501 * process-shared POSIX mutexes. For these cases where a named object
502 * is used, the object name is derived from a 64 bit FNV hash of the
503 * environment pathname. As such, naming collisions are extremely
504 * unlikely. If a collision occurs, the results are unpredictable.
506 typedef struct MDB_txbody {
507 /** Stamp identifying this as an MDB file. It must be set
510 /** Version number of this lock file. Must be set to #MDB_VERSION. */
511 uint32_t mtb_version;
512 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
513 char mtb_rmname[MNAME_LEN];
515 /** Mutex protecting access to this table.
516 * This is the reader lock that #LOCK_MUTEX_R acquires.
518 pthread_mutex_t mtb_mutex;
520 /** The ID of the last transaction committed to the database.
521 * This is recorded here only for convenience; the value can always
522 * be determined by reading the main database meta pages.
525 /** The number of slots that have been used in the reader table.
526 * This always records the maximum count, it is not decremented
527 * when readers release their slots.
529 unsigned mtb_numreaders;
532 /** The actual reader table definition. */
533 typedef struct MDB_txninfo {
536 #define mti_magic mt1.mtb.mtb_magic
537 #define mti_version mt1.mtb.mtb_version
538 #define mti_mutex mt1.mtb.mtb_mutex
539 #define mti_rmname mt1.mtb.mtb_rmname
540 #define mti_txnid mt1.mtb.mtb_txnid
541 #define mti_numreaders mt1.mtb.mtb_numreaders
542 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
545 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
546 char mt2_wmname[MNAME_LEN];
547 #define mti_wmname mt2.mt2_wmname
549 pthread_mutex_t mt2_wmutex;
550 #define mti_wmutex mt2.mt2_wmutex
552 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
554 MDB_reader mti_readers[1];
558 /** Common header for all page types.
559 * Overflow records occupy a number of contiguous pages with no
560 * headers on any page after the first.
562 typedef struct MDB_page {
563 #define mp_pgno mp_p.p_pgno
564 #define mp_next mp_p.p_next
566 pgno_t p_pgno; /**< page number */
567 void * p_next; /**< for in-memory list of freed structs */
570 /** @defgroup mdb_page Page Flags
572 * Flags for the page headers.
575 #define P_BRANCH 0x01 /**< branch page */
576 #define P_LEAF 0x02 /**< leaf page */
577 #define P_OVERFLOW 0x04 /**< overflow page */
578 #define P_META 0x08 /**< meta page */
579 #define P_DIRTY 0x10 /**< dirty page */
580 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
581 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
583 uint16_t mp_flags; /**< @ref mdb_page */
584 #define mp_lower mp_pb.pb.pb_lower
585 #define mp_upper mp_pb.pb.pb_upper
586 #define mp_pages mp_pb.pb_pages
589 indx_t pb_lower; /**< lower bound of free space */
590 indx_t pb_upper; /**< upper bound of free space */
592 uint32_t pb_pages; /**< number of overflow pages */
594 indx_t mp_ptrs[1]; /**< dynamic size */
597 /** Size of the page header, excluding dynamic data at the end */
598 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
600 /** Address of first usable data byte in a page, after the header */
601 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
603 /** Number of nodes on a page */
604 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
606 /** The amount of space remaining in the page */
607 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
609 /** The percentage of space used in the page, in tenths of a percent. */
610 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
611 ((env)->me_psize - PAGEHDRSZ))
612 /** The minimum page fill factor, in tenths of a percent.
613 * Pages emptier than this are candidates for merging.
615 #define FILL_THRESHOLD 250
617 /** Test if a page is a leaf page */
618 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
619 /** Test if a page is a LEAF2 page */
620 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
621 /** Test if a page is a branch page */
622 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
623 /** Test if a page is an overflow page */
624 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
625 /** Test if a page is a sub page */
626 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
628 /** The number of overflow pages needed to store the given size. */
629 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
631 /** Header for a single key/data pair within a page.
632 * We guarantee 2-byte alignment for nodes.
634 typedef struct MDB_node {
635 /** lo and hi are used for data size on leaf nodes and for
636 * child pgno on branch nodes. On 64 bit platforms, flags
637 * is also used for pgno. (Branch nodes have no flags).
638 * They are in host byte order in case that lets some
639 * accesses be optimized into a 32-bit word access.
641 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
642 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
643 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
644 /** @defgroup mdb_node Node Flags
646 * Flags for node headers.
649 #define F_BIGDATA 0x01 /**< data put on overflow page */
650 #define F_SUBDATA 0x02 /**< data is a sub-database */
651 #define F_DUPDATA 0x04 /**< data has duplicates */
653 /** valid flags for #mdb_node_add() */
654 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
657 unsigned short mn_flags; /**< @ref mdb_node */
658 unsigned short mn_ksize; /**< key size */
659 char mn_data[1]; /**< key and data are appended here */
662 /** Size of the node header, excluding dynamic data at the end */
663 #define NODESIZE offsetof(MDB_node, mn_data)
665 /** Bit position of top word in page number, for shifting mn_flags */
666 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
668 /** Size of a node in a branch page with a given key.
669 * This is just the node header plus the key, there is no data.
671 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
673 /** Size of a node in a leaf page with a given key and data.
674 * This is node header plus key plus data size.
676 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
678 /** Address of node \b i in page \b p */
679 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
681 /** Address of the key for the node */
682 #define NODEKEY(node) (void *)((node)->mn_data)
684 /** Address of the data for a node */
685 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
687 /** Get the page number pointed to by a branch node */
688 #define NODEPGNO(node) \
689 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
690 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
691 /** Set the page number in a branch node */
692 #define SETPGNO(node,pgno) do { \
693 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
694 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
696 /** Get the size of the data in a leaf node */
697 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
698 /** Set the size of the data for a leaf node */
699 #define SETDSZ(node,size) do { \
700 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
701 /** The size of a key in a node */
702 #define NODEKSZ(node) ((node)->mn_ksize)
704 /** Copy a page number from src to dst */
706 #define COPY_PGNO(dst,src) dst = src
708 #if SIZE_MAX > 4294967295UL
709 #define COPY_PGNO(dst,src) do { \
710 unsigned short *s, *d; \
711 s = (unsigned short *)&(src); \
712 d = (unsigned short *)&(dst); \
719 #define COPY_PGNO(dst,src) do { \
720 unsigned short *s, *d; \
721 s = (unsigned short *)&(src); \
722 d = (unsigned short *)&(dst); \
728 /** The address of a key in a LEAF2 page.
729 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
730 * There are no node headers, keys are stored contiguously.
732 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
734 /** Set the \b node's key into \b key, if requested. */
735 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
736 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
738 /** Information about a single database in the environment. */
739 typedef struct MDB_db {
740 uint32_t md_pad; /**< also ksize for LEAF2 pages */
741 uint16_t md_flags; /**< @ref mdb_dbi_open */
742 uint16_t md_depth; /**< depth of this tree */
743 pgno_t md_branch_pages; /**< number of internal pages */
744 pgno_t md_leaf_pages; /**< number of leaf pages */
745 pgno_t md_overflow_pages; /**< number of overflow pages */
746 size_t md_entries; /**< number of data items */
747 pgno_t md_root; /**< the root page of this tree */
750 /** mdb_dbi_open flags */
751 #define PERSISTENT_FLAGS 0x7fff
752 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
753 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
755 /** Handle for the DB used to track free pages. */
757 /** Handle for the default DB. */
760 /** Meta page content. */
761 typedef struct MDB_meta {
762 /** Stamp identifying this as an MDB file. It must be set
765 /** Version number of this lock file. Must be set to #MDB_VERSION. */
767 void *mm_address; /**< address for fixed mapping */
768 size_t mm_mapsize; /**< size of mmap region */
769 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
770 /** The size of pages used in this DB */
771 #define mm_psize mm_dbs[0].md_pad
772 /** Any persistent environment flags. @ref mdb_env */
773 #define mm_flags mm_dbs[0].md_flags
774 pgno_t mm_last_pg; /**< last used page in file */
775 txnid_t mm_txnid; /**< txnid that committed this page */
778 /** Buffer for a stack-allocated dirty page.
779 * The members define size and alignment, and silence type
780 * aliasing warnings. They are not used directly; that could
781 * mean incorrectly using several union members in parallel.
783 typedef union MDB_pagebuf {
784 char mb_raw[MDB_PAGESIZE];
787 char mm_pad[PAGEHDRSZ];
792 /** Auxiliary DB info.
793 * The information here is mostly static/read-only. There is
794 * only a single copy of this record in the environment.
796 typedef struct MDB_dbx {
797 MDB_val md_name; /**< name of the database */
798 MDB_cmp_func *md_cmp; /**< function for comparing keys */
799 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
800 MDB_rel_func *md_rel; /**< user relocate function */
801 void *md_relctx; /**< user-provided context for md_rel */
804 /** A database transaction.
805 * Every operation requires a transaction handle.
808 MDB_txn *mt_parent; /**< parent of a nested txn */
809 MDB_txn *mt_child; /**< nested txn under this txn */
810 pgno_t mt_next_pgno; /**< next unallocated page */
811 /** The ID of this transaction. IDs are integers incrementing from 1.
812 * Only committed write transactions increment the ID. If a transaction
813 * aborts, the ID may be re-used by the next writer.
816 MDB_env *mt_env; /**< the DB environment */
817 /** The list of pages that became unused during this transaction.
821 MDB_ID2L dirty_list; /**< for write txns: modified pages */
822 MDB_reader *reader; /**< this thread's reader table slot or NULL */
824 /** Array of records for each DB known in the environment. */
826 /** Array of MDB_db records for each known DB */
828 /** @defgroup mt_dbflag Transaction DB Flags
832 #define DB_DIRTY 0x01 /**< DB was written in this txn */
833 #define DB_STALE 0x02 /**< DB record is older than txnID */
834 #define DB_NEW 0x04 /**< DB handle opened in this txn */
835 #define DB_VALID 0x08 /**< DB handle is valid */
836 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
838 /** In write txns, array of cursors for each DB */
839 MDB_cursor **mt_cursors;
840 /** Array of flags for each DB */
841 unsigned char *mt_dbflags;
842 /** Number of DB records in use. This number only ever increments;
843 * we don't decrement it when individual DB handles are closed.
847 /** @defgroup mdb_txn Transaction Flags
851 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
852 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
853 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
855 unsigned int mt_flags; /**< @ref mdb_txn */
856 /** dirty_list maxsize - #allocated pages including in parent txns */
857 unsigned int mt_dirty_room;
858 /** Tracks which of the two meta pages was used at the start
859 * of this transaction.
861 unsigned int mt_toggle;
864 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
865 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
866 * raise this on a 64 bit machine.
868 #define CURSOR_STACK 32
872 /** Cursors are used for all DB operations */
874 /** Next cursor on this DB in this txn */
876 /** Original cursor if this is a shadow */
878 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
879 struct MDB_xcursor *mc_xcursor;
880 /** The transaction that owns this cursor */
882 /** The database handle this cursor operates on */
884 /** The database record for this cursor */
886 /** The database auxiliary record for this cursor */
888 /** The @ref mt_dbflag for this database */
889 unsigned char *mc_dbflag;
890 unsigned short mc_snum; /**< number of pushed pages */
891 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
892 /** @defgroup mdb_cursor Cursor Flags
894 * Cursor state flags.
897 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
898 #define C_EOF 0x02 /**< No more data */
899 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
900 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
901 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
902 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
904 unsigned int mc_flags; /**< @ref mdb_cursor */
905 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
906 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
909 /** Context for sorted-dup records.
910 * We could have gone to a fully recursive design, with arbitrarily
911 * deep nesting of sub-databases. But for now we only handle these
912 * levels - main DB, optional sub-DB, sorted-duplicate DB.
914 typedef struct MDB_xcursor {
915 /** A sub-cursor for traversing the Dup DB */
916 MDB_cursor mx_cursor;
917 /** The database record for this Dup DB */
919 /** The auxiliary DB record for this Dup DB */
921 /** The @ref mt_dbflag for this Dup DB */
922 unsigned char mx_dbflag;
925 /** State of FreeDB old pages, stored in the MDB_env */
926 typedef struct MDB_pgstate {
927 txnid_t mf_pglast; /**< ID of last old page record we used */
928 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
929 pgno_t *mf_pgfree; /**< memory to free when dropping me_pghead */
932 /** The database environment. */
934 HANDLE me_fd; /**< The main data file */
935 HANDLE me_lfd; /**< The lock file */
936 HANDLE me_mfd; /**< just for writing the meta pages */
937 /** Failed to update the meta page. Probably an I/O error. */
938 #define MDB_FATAL_ERROR 0x80000000U
939 /** Some fields are initialized. */
940 #define MDB_ENV_ACTIVE 0x20000000U
941 /** me_txkey is set */
942 #define MDB_ENV_TXKEY 0x10000000U
943 uint32_t me_flags; /**< @ref mdb_env */
944 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
945 unsigned int me_maxreaders; /**< size of the reader table */
946 unsigned int me_numreaders; /**< max numreaders set by this env */
947 MDB_dbi me_numdbs; /**< number of DBs opened */
948 MDB_dbi me_maxdbs; /**< size of the DB table */
949 pid_t me_pid; /**< process ID of this env */
950 char *me_path; /**< path to the DB files */
951 char *me_map; /**< the memory map of the data file */
952 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
953 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
954 MDB_txn *me_txn; /**< current write transaction */
955 size_t me_mapsize; /**< size of the data memory map */
956 off_t me_size; /**< current file size */
957 pgno_t me_maxpg; /**< me_mapsize / me_psize */
958 MDB_dbx *me_dbxs; /**< array of static DB info */
959 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
960 pthread_key_t me_txkey; /**< thread-key for readers */
961 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
962 # define me_pglast me_pgstate.mf_pglast
963 # define me_pghead me_pgstate.mf_pghead
964 # define me_pgfree me_pgstate.mf_pgfree
965 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
966 /** IDL of pages that became unused in a write txn */
968 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
969 MDB_ID2L me_dirty_list;
970 /** Max number of freelist items that can fit in a single overflow page */
971 unsigned int me_maxfree_1pg;
972 /** Max size of a node on a page */
973 unsigned int me_nodemax;
975 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
977 #elif defined(MDB_USE_POSIX_SEM)
978 sem_t *me_rmutex; /* Shared mutexes are not supported */
983 /** Nested transaction */
984 typedef struct MDB_ntxn {
985 MDB_txn mnt_txn; /* the transaction */
986 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
989 /** max number of pages to commit in one writev() call */
990 #define MDB_COMMIT_PAGES 64
991 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
992 #undef MDB_COMMIT_PAGES
993 #define MDB_COMMIT_PAGES IOV_MAX
996 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
997 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
998 static int mdb_page_touch(MDB_cursor *mc);
1000 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
1001 static int mdb_page_search_root(MDB_cursor *mc,
1002 MDB_val *key, int modify);
1003 #define MDB_PS_MODIFY 1
1004 #define MDB_PS_ROOTONLY 2
1005 static int mdb_page_search(MDB_cursor *mc,
1006 MDB_val *key, int flags);
1007 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1009 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1010 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1011 pgno_t newpgno, unsigned int nflags);
1013 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1014 static int mdb_env_pick_meta(const MDB_env *env);
1015 static int mdb_env_write_meta(MDB_txn *txn);
1016 static void mdb_env_close0(MDB_env *env, int excl);
1018 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1019 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1020 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1021 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1022 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1023 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1024 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1025 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1026 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1028 static int mdb_rebalance(MDB_cursor *mc);
1029 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1031 static void mdb_cursor_pop(MDB_cursor *mc);
1032 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1034 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1035 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1036 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1037 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1038 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1040 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1041 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1043 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1044 static void mdb_xcursor_init0(MDB_cursor *mc);
1045 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1047 static int mdb_drop0(MDB_cursor *mc, int subs);
1048 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1051 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1055 static SECURITY_DESCRIPTOR mdb_null_sd;
1056 static SECURITY_ATTRIBUTES mdb_all_sa;
1057 static int mdb_sec_inited;
1060 /** Return the library version info. */
1062 mdb_version(int *major, int *minor, int *patch)
1064 if (major) *major = MDB_VERSION_MAJOR;
1065 if (minor) *minor = MDB_VERSION_MINOR;
1066 if (patch) *patch = MDB_VERSION_PATCH;
1067 return MDB_VERSION_STRING;
1070 /** Table of descriptions for MDB @ref errors */
1071 static char *const mdb_errstr[] = {
1072 "MDB_KEYEXIST: Key/data pair already exists",
1073 "MDB_NOTFOUND: No matching key/data pair found",
1074 "MDB_PAGE_NOTFOUND: Requested page not found",
1075 "MDB_CORRUPTED: Located page was wrong type",
1076 "MDB_PANIC: Update of meta page failed",
1077 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1078 "MDB_INVALID: File is not an MDB file",
1079 "MDB_MAP_FULL: Environment mapsize limit reached",
1080 "MDB_DBS_FULL: Environment maxdbs limit reached",
1081 "MDB_READERS_FULL: Environment maxreaders limit reached",
1082 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1083 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1084 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1085 "MDB_PAGE_FULL: Internal error - page has no more space",
1086 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1087 "MDB_INCOMPATIBLE: Database flags changed or would change",
1088 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1092 mdb_strerror(int err)
1096 return ("Successful return: 0");
1098 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1099 i = err - MDB_KEYEXIST;
1100 return mdb_errstr[i];
1103 return strerror(err);
1107 /** Display a key in hexadecimal and return the address of the result.
1108 * @param[in] key the key to display
1109 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1110 * @return The key in hexadecimal form.
1113 mdb_dkey(MDB_val *key, char *buf)
1116 unsigned char *c = key->mv_data;
1122 if (key->mv_size > MDB_MAXKEYSIZE)
1123 return "MDB_MAXKEYSIZE";
1124 /* may want to make this a dynamic check: if the key is mostly
1125 * printable characters, print it as-is instead of converting to hex.
1129 for (i=0; i<key->mv_size; i++)
1130 ptr += sprintf(ptr, "%02x", *c++);
1132 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1137 /** Display all the keys in the page. */
1139 mdb_page_list(MDB_page *mp)
1142 unsigned int i, nkeys, nsize;
1146 nkeys = NUMKEYS(mp);
1147 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1148 for (i=0; i<nkeys; i++) {
1149 node = NODEPTR(mp, i);
1150 key.mv_size = node->mn_ksize;
1151 key.mv_data = node->mn_data;
1152 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1153 if (IS_BRANCH(mp)) {
1154 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1157 if (F_ISSET(node->mn_flags, F_BIGDATA))
1158 nsize += sizeof(pgno_t);
1160 nsize += NODEDSZ(node);
1161 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1167 mdb_cursor_chk(MDB_cursor *mc)
1173 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1174 for (i=0; i<mc->mc_top; i++) {
1176 node = NODEPTR(mp, mc->mc_ki[i]);
1177 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1180 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1186 /** Count all the pages in each DB and in the freelist
1187 * and make sure it matches the actual number of pages
1190 static void mdb_audit(MDB_txn *txn)
1194 MDB_ID freecount, count;
1199 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1200 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1201 freecount += *(MDB_ID *)data.mv_data;
1204 for (i = 0; i<txn->mt_numdbs; i++) {
1205 MDB_xcursor mx, *mxp;
1206 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1207 mdb_cursor_init(&mc, txn, i, mxp);
1208 if (txn->mt_dbs[i].md_root == P_INVALID)
1210 count += txn->mt_dbs[i].md_branch_pages +
1211 txn->mt_dbs[i].md_leaf_pages +
1212 txn->mt_dbs[i].md_overflow_pages;
1213 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1214 mdb_page_search(&mc, NULL, 0);
1218 mp = mc.mc_pg[mc.mc_top];
1219 for (j=0; j<NUMKEYS(mp); j++) {
1220 MDB_node *leaf = NODEPTR(mp, j);
1221 if (leaf->mn_flags & F_SUBDATA) {
1223 memcpy(&db, NODEDATA(leaf), sizeof(db));
1224 count += db.md_branch_pages + db.md_leaf_pages +
1225 db.md_overflow_pages;
1229 while (mdb_cursor_sibling(&mc, 1) == 0);
1232 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1233 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1234 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1240 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1242 return txn->mt_dbxs[dbi].md_cmp(a, b);
1246 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1248 if (txn->mt_dbxs[dbi].md_dcmp)
1249 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1251 return EINVAL; /* too bad you can't distinguish this from a valid result */
1254 /** Allocate a single page.
1255 * Re-use old malloc'd pages first, otherwise just malloc.
1258 mdb_page_malloc(MDB_cursor *mc) {
1260 size_t sz = mc->mc_txn->mt_env->me_psize;
1261 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1262 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1263 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1264 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1265 } else if ((ret = malloc(sz)) != NULL) {
1266 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1272 mdb_page_free(MDB_env *env, MDB_page *mp)
1274 mp->mp_next = env->me_dpages;
1275 VGMEMP_FREE(env, mp);
1276 env->me_dpages = mp;
1279 /** Allocate pages for writing.
1280 * If there are free pages available from older transactions, they
1281 * will be re-used first. Otherwise a new page will be allocated.
1282 * @param[in] mc cursor A cursor handle identifying the transaction and
1283 * database for which we are allocating.
1284 * @param[in] num the number of pages to allocate.
1285 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1286 * will always be satisfied by a single contiguous chunk of memory.
1287 * @return 0 on success, non-zero on failure.
1290 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1292 MDB_txn *txn = mc->mc_txn;
1294 pgno_t pgno = P_INVALID;
1296 txnid_t oldest = 0, last;
1301 /* If our dirty list is already full, we can't do anything */
1302 if (txn->mt_dirty_room == 0)
1303 return MDB_TXN_FULL;
1305 /* The free list won't have any content at all until txn 2 has
1306 * committed. The pages freed by txn 2 will be unreferenced
1307 * after txn 3 commits, and so will be safe to re-use in txn 4.
1309 if (txn->mt_txnid > 3) {
1310 if (!txn->mt_env->me_pghead &&
1311 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1312 /* See if there's anything in the free DB */
1319 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1320 if (!txn->mt_env->me_pglast) {
1321 mdb_page_search(&m2, NULL, 0);
1322 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1323 kptr = (txnid_t *)NODEKEY(leaf);
1328 last = txn->mt_env->me_pglast + 1;
1330 key.mv_data = &last;
1331 key.mv_size = sizeof(last);
1332 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1335 last = *(txnid_t *)key.mv_data;
1341 oldest = txn->mt_txnid - 1;
1342 nr = txn->mt_env->me_txns->mti_numreaders;
1343 r = txn->mt_env->me_txns->mti_readers;
1344 for (i=0; i<nr; i++) {
1345 if (!r[i].mr_pid) continue;
1352 if (oldest > last) {
1353 /* It's usable, grab it.
1357 if (!txn->mt_env->me_pglast) {
1358 mdb_node_read(txn, leaf, &data);
1360 idl = (MDB_ID *) data.mv_data;
1361 /* We might have a zero-length IDL due to freelist growth
1362 * during a prior commit
1365 txn->mt_env->me_pglast = last;
1368 mop = malloc(MDB_IDL_SIZEOF(idl));
1371 txn->mt_env->me_pglast = last;
1372 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1373 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1378 DPRINTF("IDL read txn %zu root %zu num %zu",
1379 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1380 for (i=0; i<idl[0]; i++) {
1381 DPRINTF("IDL %zu", idl[i+1]);
1388 if (txn->mt_env->me_pghead) {
1389 pgno_t *mop = txn->mt_env->me_pghead;
1392 int retry = 1, readit = 0, n2 = num-1;
1393 unsigned int i, j, k;
1395 /* If current list is too short, must fetch more and coalesce */
1396 if (mop[0] < (unsigned)num)
1399 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1401 /* If on freelist, don't try to read more. If what we have
1402 * right now isn't enough just use new pages.
1403 * TODO: get all of this working. Many circular dependencies...
1405 if (mc->mc_dbi == FREE_DBI) {
1413 last = txn->mt_env->me_pglast + 1;
1415 /* We haven't hit the readers list yet? */
1421 oldest = txn->mt_txnid - 1;
1422 nr = txn->mt_env->me_txns->mti_numreaders;
1423 r = txn->mt_env->me_txns->mti_readers;
1424 for (i=0; i<nr; i++) {
1425 if (!r[i].mr_pid) continue;
1432 /* There's nothing we can use on the freelist */
1433 if (oldest - last < 1)
1436 key.mv_data = &last;
1437 key.mv_size = sizeof(last);
1438 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1440 if (rc == MDB_NOTFOUND)
1444 last = *(txnid_t*)key.mv_data;
1447 idl = (MDB_ID *) data.mv_data;
1448 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1451 /* merge in sorted order */
1452 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1454 while (i>0 || j>0) {
1455 if (i && idl[i] < mop[j])
1456 mop2[k--] = idl[i--];
1458 mop2[k--] = mop[j--];
1460 txn->mt_env->me_pglast = last;
1461 free(txn->mt_env->me_pgfree);
1462 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1464 /* Keep trying to read until we have enough */
1465 if (mop[0] < (unsigned)num) {
1470 /* current list has enough pages, but are they contiguous? */
1471 for (i=mop[0]; i>=(unsigned)num; i--) {
1472 if (mop[i-n2] == mop[i] + n2) {
1475 /* move any stragglers down */
1476 for (j=i+num; j<=mop[0]; j++)
1483 /* Stop if we succeeded, or no retries */
1484 if (!retry || pgno != P_INVALID)
1490 /* peel pages off tail, so we only have to truncate the list */
1491 pgno = MDB_IDL_LAST(mop);
1494 if (MDB_IDL_IS_ZERO(mop)) {
1495 free(txn->mt_env->me_pgfree);
1496 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1501 if (pgno == P_INVALID) {
1502 /* DB size is maxed out */
1503 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1504 DPUTS("DB size maxed out");
1505 return MDB_MAP_FULL;
1508 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1509 if (pgno == P_INVALID) {
1510 pgno = txn->mt_next_pgno;
1511 txn->mt_next_pgno += num;
1513 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1516 if (txn->mt_env->me_dpages && num == 1) {
1517 np = txn->mt_env->me_dpages;
1518 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1519 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1520 txn->mt_env->me_dpages = np->mp_next;
1522 size_t sz = txn->mt_env->me_psize * num;
1523 if ((np = malloc(sz)) == NULL)
1525 VGMEMP_ALLOC(txn->mt_env, np, sz);
1527 if (pgno == P_INVALID) {
1528 np->mp_pgno = txn->mt_next_pgno;
1529 txn->mt_next_pgno += num;
1534 mid.mid = np->mp_pgno;
1536 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1537 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1539 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1541 txn->mt_dirty_room--;
1547 /** Copy a page: avoid copying unused portions of the page.
1548 * @param[in] dst page to copy into
1549 * @param[in] src page to copy from
1552 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1554 dst->mp_flags = src->mp_flags | P_DIRTY;
1555 dst->mp_pages = src->mp_pages;
1557 if (IS_LEAF2(src)) {
1558 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1560 unsigned int i, nkeys = NUMKEYS(src);
1561 for (i=0; i<nkeys; i++)
1562 dst->mp_ptrs[i] = src->mp_ptrs[i];
1563 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1564 psize - src->mp_upper);
1568 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1569 * @param[in] mc cursor pointing to the page to be touched
1570 * @return 0 on success, non-zero on failure.
1573 mdb_page_touch(MDB_cursor *mc)
1575 MDB_page *mp = mc->mc_pg[mc->mc_top];
1579 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1581 if ((rc = mdb_page_alloc(mc, 1, &np)))
1583 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1584 assert(mp->mp_pgno != np->mp_pgno);
1585 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1587 /* If page isn't full, just copy the used portion */
1588 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1591 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1593 np->mp_flags |= P_DIRTY;
1598 /* Adjust other cursors pointing to mp */
1599 if (mc->mc_flags & C_SUB) {
1600 MDB_cursor *m2, *m3;
1601 MDB_dbi dbi = mc->mc_dbi-1;
1603 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1604 if (m2 == mc) continue;
1605 m3 = &m2->mc_xcursor->mx_cursor;
1606 if (m3->mc_snum < mc->mc_snum) continue;
1607 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1608 m3->mc_pg[mc->mc_top] = mp;
1614 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1615 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1616 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1617 m2->mc_pg[mc->mc_top] = mp;
1621 mc->mc_pg[mc->mc_top] = mp;
1622 /** If this page has a parent, update the parent to point to
1626 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1628 mc->mc_db->md_root = mp->mp_pgno;
1629 } else if (mc->mc_txn->mt_parent) {
1632 /* If txn has a parent, make sure the page is in our
1635 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1636 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1637 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1638 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1639 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1640 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1641 mc->mc_pg[mc->mc_top] = mp;
1646 assert(mc->mc_txn->mt_u.dirty_list[0].mid < MDB_IDL_UM_MAX);
1648 np = mdb_page_malloc(mc);
1651 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1652 mid.mid = np->mp_pgno;
1654 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1662 mdb_env_sync(MDB_env *env, int force)
1665 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1666 if (env->me_flags & MDB_WRITEMAP) {
1667 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1668 ? MS_ASYNC : MS_SYNC;
1669 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1672 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1676 if (MDB_FDATASYNC(env->me_fd))
1683 /** Make shadow copies of all of parent txn's cursors */
1685 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1687 MDB_cursor *mc, *m2;
1688 unsigned int i, j, size;
1690 for (i=0;i<src->mt_numdbs; i++) {
1691 if (src->mt_cursors[i]) {
1692 size = sizeof(MDB_cursor);
1693 if (src->mt_cursors[i]->mc_xcursor)
1694 size += sizeof(MDB_xcursor);
1695 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1702 mc->mc_db = &dst->mt_dbs[i];
1703 mc->mc_dbx = m2->mc_dbx;
1704 mc->mc_dbflag = &dst->mt_dbflags[i];
1705 mc->mc_snum = m2->mc_snum;
1706 mc->mc_top = m2->mc_top;
1707 mc->mc_flags = m2->mc_flags | C_SHADOW;
1708 for (j=0; j<mc->mc_snum; j++) {
1709 mc->mc_pg[j] = m2->mc_pg[j];
1710 mc->mc_ki[j] = m2->mc_ki[j];
1712 if (m2->mc_xcursor) {
1713 MDB_xcursor *mx, *mx2;
1714 mx = (MDB_xcursor *)(mc+1);
1715 mc->mc_xcursor = mx;
1716 mx2 = m2->mc_xcursor;
1717 mx->mx_db = mx2->mx_db;
1718 mx->mx_dbx = mx2->mx_dbx;
1719 mx->mx_dbflag = mx2->mx_dbflag;
1720 mx->mx_cursor.mc_txn = dst;
1721 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1722 mx->mx_cursor.mc_db = &mx->mx_db;
1723 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1724 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1725 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1726 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1727 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1728 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1729 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1730 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1733 mc->mc_xcursor = NULL;
1735 mc->mc_next = dst->mt_cursors[i];
1736 dst->mt_cursors[i] = mc;
1743 /** Merge shadow cursors back into parent's */
1745 mdb_cursor_merge(MDB_txn *txn)
1748 for (i=0; i<txn->mt_numdbs; i++) {
1749 if (txn->mt_cursors[i]) {
1751 while ((mc = txn->mt_cursors[i])) {
1752 txn->mt_cursors[i] = mc->mc_next;
1753 if (mc->mc_flags & C_SHADOW) {
1754 MDB_cursor *m2 = mc->mc_orig;
1756 m2->mc_snum = mc->mc_snum;
1757 m2->mc_top = mc->mc_top;
1758 for (j=0; j<mc->mc_snum; j++) {
1759 m2->mc_pg[j] = mc->mc_pg[j];
1760 m2->mc_ki[j] = mc->mc_ki[j];
1763 if (mc->mc_flags & C_ALLOCD)
1771 mdb_txn_reset0(MDB_txn *txn);
1773 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1774 * @param[in] txn the transaction handle to initialize
1775 * @return 0 on success, non-zero on failure. This can only
1776 * fail for read-only transactions, and then only if the
1777 * reader table is full.
1780 mdb_txn_renew0(MDB_txn *txn)
1782 MDB_env *env = txn->mt_env;
1788 txn->mt_numdbs = env->me_numdbs;
1789 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1791 if (txn->mt_flags & MDB_TXN_RDONLY) {
1792 if (!env->me_txns) {
1793 i = mdb_env_pick_meta(env);
1794 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1795 txn->mt_u.reader = NULL;
1797 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
1798 pthread_getspecific(env->me_txkey);
1800 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
1801 return MDB_BAD_RSLOT;
1803 pid_t pid = env->me_pid;
1804 pthread_t tid = pthread_self();
1807 for (i=0; i<env->me_txns->mti_numreaders; i++)
1808 if (env->me_txns->mti_readers[i].mr_pid == 0)
1810 if (i == env->me_maxreaders) {
1811 UNLOCK_MUTEX_R(env);
1812 return MDB_READERS_FULL;
1814 env->me_txns->mti_readers[i].mr_pid = pid;
1815 env->me_txns->mti_readers[i].mr_tid = tid;
1816 if (i >= env->me_txns->mti_numreaders)
1817 env->me_txns->mti_numreaders = i+1;
1818 /* Save numreaders for un-mutexed mdb_env_close() */
1819 env->me_numreaders = env->me_txns->mti_numreaders;
1820 UNLOCK_MUTEX_R(env);
1821 r = &env->me_txns->mti_readers[i];
1822 if (!(env->me_flags & MDB_NOTLS) &&
1823 (rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1824 env->me_txns->mti_readers[i].mr_pid = 0;
1828 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1829 txn->mt_u.reader = r;
1831 txn->mt_toggle = txn->mt_txnid & 1;
1832 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1836 txn->mt_txnid = env->me_txns->mti_txnid;
1837 txn->mt_toggle = txn->mt_txnid & 1;
1838 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1841 if (txn->mt_txnid == mdb_debug_start)
1844 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1845 txn->mt_u.dirty_list = env->me_dirty_list;
1846 txn->mt_u.dirty_list[0].mid = 0;
1847 txn->mt_free_pgs = env->me_free_pgs;
1848 txn->mt_free_pgs[0] = 0;
1852 /* Copy the DB info and flags */
1853 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1854 for (i=2; i<txn->mt_numdbs; i++) {
1855 x = env->me_dbflags[i];
1856 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1857 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1859 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1861 if (env->me_maxpg < txn->mt_next_pgno) {
1862 mdb_txn_reset0(txn);
1863 return MDB_MAP_RESIZED;
1870 mdb_txn_renew(MDB_txn *txn)
1874 if (!txn || txn->mt_numdbs || !(txn->mt_flags & MDB_TXN_RDONLY))
1877 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1878 DPUTS("environment had fatal error, must shutdown!");
1882 rc = mdb_txn_renew0(txn);
1883 if (rc == MDB_SUCCESS) {
1884 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1885 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1886 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1892 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1896 int rc, size, tsize = sizeof(MDB_txn);
1898 if (env->me_flags & MDB_FATAL_ERROR) {
1899 DPUTS("environment had fatal error, must shutdown!");
1902 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1905 /* Nested transactions: Max 1 child, write txns only, no writemap */
1906 if (parent->mt_child ||
1907 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1908 (env->me_flags & MDB_WRITEMAP))
1912 tsize = sizeof(MDB_ntxn);
1914 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1915 if (!(flags & MDB_RDONLY))
1916 size += env->me_maxdbs * sizeof(MDB_cursor *);
1918 if ((txn = calloc(1, size)) == NULL) {
1919 DPRINTF("calloc: %s", strerror(ErrCode()));
1922 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1923 if (flags & MDB_RDONLY) {
1924 txn->mt_flags |= MDB_TXN_RDONLY;
1925 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1927 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1928 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1934 txn->mt_free_pgs = mdb_midl_alloc();
1935 if (!txn->mt_free_pgs) {
1939 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1940 if (!txn->mt_u.dirty_list) {
1941 free(txn->mt_free_pgs);
1945 txn->mt_txnid = parent->mt_txnid;
1946 txn->mt_toggle = parent->mt_toggle;
1947 txn->mt_dirty_room = parent->mt_dirty_room;
1948 txn->mt_u.dirty_list[0].mid = 0;
1949 txn->mt_free_pgs[0] = 0;
1950 txn->mt_next_pgno = parent->mt_next_pgno;
1951 parent->mt_child = txn;
1952 txn->mt_parent = parent;
1953 txn->mt_numdbs = parent->mt_numdbs;
1954 txn->mt_dbxs = parent->mt_dbxs;
1955 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1956 /* Copy parent's mt_dbflags, but clear DB_NEW */
1957 for (i=0; i<txn->mt_numdbs; i++)
1958 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1960 ntxn = (MDB_ntxn *)txn;
1961 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1962 if (env->me_pghead) {
1963 size = MDB_IDL_SIZEOF(env->me_pghead);
1964 env->me_pghead = malloc(size);
1966 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1970 env->me_pgfree = env->me_pghead;
1972 rc = mdb_cursor_shadow(parent, txn);
1974 mdb_txn_reset0(txn);
1976 rc = mdb_txn_renew0(txn);
1982 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1983 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1984 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1990 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1991 * May be called twice for readonly txns: First reset it, then abort.
1992 * @param[in] txn the transaction handle to reset
1995 mdb_txn_reset0(MDB_txn *txn)
1997 MDB_env *env = txn->mt_env;
2000 /* Close any DBI handles opened in this txn */
2001 for (i=2; i<txn->mt_numdbs; i++) {
2002 if (txn->mt_dbflags[i] & DB_NEW) {
2003 char *ptr = env->me_dbxs[i].md_name.mv_data;
2004 env->me_dbxs[i].md_name.mv_data = NULL;
2005 env->me_dbxs[i].md_name.mv_size = 0;
2010 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2011 if (txn->mt_u.reader) {
2012 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2013 if (!(env->me_flags & MDB_NOTLS))
2014 txn->mt_u.reader = NULL; /* txn does not own reader */
2016 txn->mt_numdbs = 0; /* mark txn as reset, do not close DBs again */
2020 /* close(free) all cursors */
2021 for (i=0; i<txn->mt_numdbs; i++) {
2022 if (txn->mt_cursors[i]) {
2024 while ((mc = txn->mt_cursors[i])) {
2025 txn->mt_cursors[i] = mc->mc_next;
2026 if (mc->mc_flags & C_ALLOCD)
2032 if (!(env->me_flags & MDB_WRITEMAP)) {
2033 /* return all dirty pages to dpage list */
2034 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2035 dp = txn->mt_u.dirty_list[i].mptr;
2036 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2037 mdb_page_free(txn->mt_env, dp);
2039 /* large pages just get freed directly */
2040 VGMEMP_FREE(txn->mt_env, dp);
2046 free(env->me_pgfree);
2048 if (txn->mt_parent) {
2049 txn->mt_parent->mt_child = NULL;
2050 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2051 mdb_midl_free(txn->mt_free_pgs);
2052 free(txn->mt_u.dirty_list);
2055 if (mdb_midl_shrink(&txn->mt_free_pgs))
2056 env->me_free_pgs = txn->mt_free_pgs;
2059 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
2060 txn->mt_env->me_pglast = 0;
2063 /* The writer mutex was locked in mdb_txn_begin. */
2064 UNLOCK_MUTEX_W(env);
2069 mdb_txn_reset(MDB_txn *txn)
2074 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2075 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2076 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2078 /* This call is only valid for read-only txns */
2079 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2082 mdb_txn_reset0(txn);
2086 mdb_txn_abort(MDB_txn *txn)
2091 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
2092 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2093 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2096 mdb_txn_abort(txn->mt_child);
2098 mdb_txn_reset0(txn);
2099 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2100 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2101 txn->mt_u.reader->mr_pid = 0;
2107 mdb_txn_commit(MDB_txn *txn)
2115 pgno_t next, freecnt;
2116 txnid_t oldpg_txnid, id;
2119 assert(txn != NULL);
2120 assert(txn->mt_env != NULL);
2122 if (txn->mt_child) {
2123 mdb_txn_commit(txn->mt_child);
2124 txn->mt_child = NULL;
2129 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2130 /* update the DB flags */
2131 for (i = 2; i<txn->mt_numdbs; i++) {
2132 if (txn->mt_dbflags[i] & DB_NEW)
2133 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2135 if (txn->mt_numdbs > env->me_numdbs)
2136 env->me_numdbs = txn->mt_numdbs;
2137 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2142 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2143 DPUTS("error flag is set, can't commit");
2145 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2150 if (txn->mt_parent) {
2151 MDB_txn *parent = txn->mt_parent;
2155 /* Append our free list to parent's */
2156 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2160 mdb_midl_free(txn->mt_free_pgs);
2162 parent->mt_next_pgno = txn->mt_next_pgno;
2163 parent->mt_flags = txn->mt_flags;
2165 /* Merge (and close) our cursors with parent's */
2166 mdb_cursor_merge(txn);
2168 /* Update parent's DB table. */
2169 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2170 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2171 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2172 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2173 for (i=2; i<txn->mt_numdbs; i++) {
2174 /* preserve parent's DB_NEW status */
2175 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2176 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2179 dst = txn->mt_parent->mt_u.dirty_list;
2180 src = txn->mt_u.dirty_list;
2181 /* Find len = length of merging our dirty list with parent's */
2183 dst[0].mid = 0; /* simplify loops */
2184 if (parent->mt_parent) {
2185 len = x + src[0].mid;
2186 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2187 for (i = x; y && i; y--) {
2188 pgno_t yp = src[y].mid;
2189 while (yp < dst[i].mid)
2191 if (yp == dst[i].mid) {
2196 } else { /* Simplify the above for single-ancestor case */
2197 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2199 /* Merge our dirty list with parent's */
2201 for (i = len; y; dst[i--] = src[y--]) {
2202 pgno_t yp = src[y].mid;
2203 while (yp < dst[x].mid)
2204 dst[i--] = dst[x--];
2205 if (yp == dst[x].mid)
2206 free(dst[x--].mptr);
2210 free(txn->mt_u.dirty_list);
2211 parent->mt_dirty_room = txn->mt_dirty_room;
2213 txn->mt_parent->mt_child = NULL;
2214 free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pgfree);
2219 if (txn != env->me_txn) {
2220 DPUTS("attempt to commit unknown transaction");
2225 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2228 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2229 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2231 /* Update DB root pointers */
2232 if (txn->mt_numdbs > 2) {
2235 data.mv_size = sizeof(MDB_db);
2237 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2238 for (i = 2; i < txn->mt_numdbs; i++) {
2239 if (txn->mt_dbflags[i] & DB_DIRTY) {
2240 data.mv_data = &txn->mt_dbs[i];
2241 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2248 /* Save the freelist as of this transaction to the freeDB. This
2249 * can change the freelist, so keep trying until it stabilizes.
2251 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2252 * except the code below can decrease env->me_pglast to split pghead.
2253 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2254 * can only appear in env->me_pghead when env->me_pglast increases.
2255 * Until then, the me_pghead pointer won't move but can become NULL.
2258 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2259 oldpg_txnid = id = 0;
2262 /* should only be one record now */
2263 if (env->me_pghead || env->me_pglast) {
2264 /* make sure first page of freeDB is touched and on freelist */
2265 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2266 if (rc && rc != MDB_NOTFOUND) {
2273 /* Delete IDLs we used from the free list */
2274 if (env->me_pglast) {
2279 rc = mdb_cursor_first(&mc, &key, NULL);
2282 oldpg_txnid = *(txnid_t *)key.mv_data;
2284 assert(oldpg_txnid <= env->me_pglast);
2286 rc = mdb_cursor_del(&mc, 0);
2289 } while (oldpg_txnid < env->me_pglast);
2292 /* Save IDL of pages freed by this txn, to freeDB */
2294 if (freecnt != txn->mt_free_pgs[0]) {
2297 /* make sure last page of freeDB is touched and on freelist */
2298 key.mv_size = MDB_MAXKEYSIZE+1;
2300 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2301 if (rc && rc != MDB_NOTFOUND)
2307 MDB_IDL idl = txn->mt_free_pgs;
2308 mdb_midl_sort(txn->mt_free_pgs);
2309 DPRINTF("IDL write txn %zu root %zu num %zu",
2310 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2311 for (i=1; i<=idl[0]; i++) {
2312 DPRINTF("IDL %zu", idl[i]);
2316 /* write to last page of freeDB */
2317 key.mv_size = sizeof(pgno_t);
2318 key.mv_data = &txn->mt_txnid;
2319 /* The free list can still grow during this call,
2320 * despite the pre-emptive touches above. So retry
2321 * until the reserved space remains big enough.
2324 assert(freecnt < txn->mt_free_pgs[0]);
2325 freecnt = txn->mt_free_pgs[0];
2326 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2327 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2330 } while (freecnt != txn->mt_free_pgs[0]);
2331 mdb_midl_sort(txn->mt_free_pgs);
2332 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2333 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2334 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2337 /* Put back page numbers we took from freeDB but did not use */
2338 if (env->me_pghead) {
2343 mop = env->me_pghead;
2344 id = env->me_pglast;
2345 key.mv_size = sizeof(id);
2347 /* These steps may grow the freelist again
2348 * due to freed overflow pages...
2353 if (orig > env->me_maxfree_1pg && id > 4)
2354 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2355 data.mv_size = (orig + 1) * sizeof(pgno_t);
2356 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2359 assert(!env->me_pghead || env->me_pglast);
2360 /* mop could have been used again here */
2361 if (id != env->me_pglast || env->me_pghead == NULL)
2362 goto again; /* was completely used up */
2363 assert(mop == env->me_pghead);
2364 } while (mop[0] < orig && --i);
2365 memcpy(data.mv_data, mop, data.mv_size);
2368 *(pgno_t *)data.mv_data = orig;
2369 mop[orig] = mop[0] - orig;
2370 env->me_pghead = mop += orig;
2371 /* Save more oldpages at the previous txnid. */
2372 assert(env->me_pglast == id && id == oldpg_txnid);
2373 env->me_pglast = --oldpg_txnid;
2377 /* Check for growth of freelist again */
2378 if (freecnt != txn->mt_free_pgs[0])
2381 free(env->me_pgfree);
2382 env->me_pghead = env->me_pgfree = NULL;
2384 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2385 if (mdb_midl_shrink(&txn->mt_free_pgs))
2386 env->me_free_pgs = txn->mt_free_pgs;
2393 if (env->me_flags & MDB_WRITEMAP) {
2394 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2395 dp = txn->mt_u.dirty_list[i].mptr;
2396 /* clear dirty flag */
2397 dp->mp_flags &= ~P_DIRTY;
2398 txn->mt_u.dirty_list[i].mid = 0;
2400 txn->mt_u.dirty_list[0].mid = 0;
2404 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2410 /* Windows actually supports scatter/gather I/O, but only on
2411 * unbuffered file handles. Since we're relying on the OS page
2412 * cache for all our data, that's self-defeating. So we just
2413 * write pages one at a time. We use the ov structure to set
2414 * the write offset, to at least save the overhead of a Seek
2418 memset(&ov, 0, sizeof(ov));
2419 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2421 dp = txn->mt_u.dirty_list[i].mptr;
2422 DPRINTF("committing page %zu", dp->mp_pgno);
2423 size = dp->mp_pgno * env->me_psize;
2424 ov.Offset = size & 0xffffffff;
2425 ov.OffsetHigh = size >> 16;
2426 ov.OffsetHigh >>= 16;
2427 /* clear dirty flag */
2428 dp->mp_flags &= ~P_DIRTY;
2429 wsize = env->me_psize;
2430 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2431 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2434 DPRINTF("WriteFile: %d", n);
2441 struct iovec iov[MDB_COMMIT_PAGES];
2445 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2446 dp = txn->mt_u.dirty_list[i].mptr;
2447 if (dp->mp_pgno != next) {
2449 rc = writev(env->me_fd, iov, n);
2453 DPUTS("short write, filesystem full?");
2455 DPRINTF("writev: %s", strerror(n));
2462 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2465 DPRINTF("committing page %zu", dp->mp_pgno);
2466 iov[n].iov_len = env->me_psize;
2467 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2468 iov[n].iov_base = (char *)dp;
2469 size += iov[n].iov_len;
2470 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2471 /* clear dirty flag */
2472 dp->mp_flags &= ~P_DIRTY;
2473 if (++n >= MDB_COMMIT_PAGES) {
2483 rc = writev(env->me_fd, iov, n);
2487 DPUTS("short write, filesystem full?");
2489 DPRINTF("writev: %s", strerror(n));
2496 /* Drop the dirty pages.
2498 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2499 dp = txn->mt_u.dirty_list[i].mptr;
2500 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2501 mdb_page_free(txn->mt_env, dp);
2503 VGMEMP_FREE(txn->mt_env, dp);
2506 txn->mt_u.dirty_list[i].mid = 0;
2508 txn->mt_u.dirty_list[0].mid = 0;
2511 if ((n = mdb_env_sync(env, 0)) != 0 ||
2512 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2520 /* update the DB flags */
2521 for (i = 2; i<txn->mt_numdbs; i++) {
2522 if (txn->mt_dbflags[i] & DB_NEW)
2523 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2525 if (txn->mt_numdbs > env->me_numdbs)
2526 env->me_numdbs = txn->mt_numdbs;
2528 UNLOCK_MUTEX_W(env);
2534 /** Read the environment parameters of a DB environment before
2535 * mapping it into memory.
2536 * @param[in] env the environment handle
2537 * @param[out] meta address of where to store the meta information
2538 * @return 0 on success, non-zero on failure.
2541 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2548 /* We don't know the page size yet, so use a minimum value.
2549 * Read both meta pages so we can use the latest one.
2552 for (i=0; i<2; i++) {
2554 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2556 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2561 else if (rc != MDB_PAGESIZE) {
2565 DPRINTF("read: %s", strerror(err));
2569 p = (MDB_page *)&pbuf;
2571 if (!F_ISSET(p->mp_flags, P_META)) {
2572 DPRINTF("page %zu not a meta page", p->mp_pgno);
2577 if (m->mm_magic != MDB_MAGIC) {
2578 DPUTS("meta has invalid magic");
2582 if (m->mm_version != MDB_VERSION) {
2583 DPRINTF("database is version %u, expected version %u",
2584 m->mm_version, MDB_VERSION);
2585 return MDB_VERSION_MISMATCH;
2589 if (m->mm_txnid > meta->mm_txnid)
2590 memcpy(meta, m, sizeof(*m));
2592 memcpy(meta, m, sizeof(*m));
2594 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2596 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2604 /** Write the environment parameters of a freshly created DB environment.
2605 * @param[in] env the environment handle
2606 * @param[out] meta address of where to store the meta information
2607 * @return 0 on success, non-zero on failure.
2610 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2617 DPUTS("writing new meta page");
2619 GET_PAGESIZE(psize);
2621 meta->mm_magic = MDB_MAGIC;
2622 meta->mm_version = MDB_VERSION;
2623 meta->mm_mapsize = env->me_mapsize;
2624 meta->mm_psize = psize;
2625 meta->mm_last_pg = 1;
2626 meta->mm_flags = env->me_flags & 0xffff;
2627 meta->mm_flags |= MDB_INTEGERKEY;
2628 meta->mm_dbs[0].md_root = P_INVALID;
2629 meta->mm_dbs[1].md_root = P_INVALID;
2631 p = calloc(2, psize);
2633 p->mp_flags = P_META;
2636 memcpy(m, meta, sizeof(*meta));
2638 q = (MDB_page *)((char *)p + psize);
2641 q->mp_flags = P_META;
2644 memcpy(m, meta, sizeof(*meta));
2649 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2650 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2651 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2654 lseek(env->me_fd, 0, SEEK_SET);
2655 rc = write(env->me_fd, p, psize * 2);
2656 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2662 /** Update the environment info to commit a transaction.
2663 * @param[in] txn the transaction that's being committed
2664 * @return 0 on success, non-zero on failure.
2667 mdb_env_write_meta(MDB_txn *txn)
2670 MDB_meta meta, metab, *mp;
2672 int rc, len, toggle;
2679 assert(txn != NULL);
2680 assert(txn->mt_env != NULL);
2682 toggle = !txn->mt_toggle;
2683 DPRINTF("writing meta page %d for root page %zu",
2684 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2687 mp = env->me_metas[toggle];
2689 if (env->me_flags & MDB_WRITEMAP) {
2690 /* Persist any increases of mapsize config */
2691 if (env->me_mapsize > mp->mm_mapsize)
2692 mp->mm_mapsize = env->me_mapsize;
2693 mp->mm_dbs[0] = txn->mt_dbs[0];
2694 mp->mm_dbs[1] = txn->mt_dbs[1];
2695 mp->mm_last_pg = txn->mt_next_pgno - 1;
2696 mp->mm_txnid = txn->mt_txnid;
2697 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2698 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2701 ptr += env->me_psize;
2702 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2709 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2710 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2712 ptr = (char *)&meta;
2713 if (env->me_mapsize > mp->mm_mapsize) {
2714 /* Persist any increases of mapsize config */
2715 meta.mm_mapsize = env->me_mapsize;
2716 off = offsetof(MDB_meta, mm_mapsize);
2718 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2720 len = sizeof(MDB_meta) - off;
2723 meta.mm_dbs[0] = txn->mt_dbs[0];
2724 meta.mm_dbs[1] = txn->mt_dbs[1];
2725 meta.mm_last_pg = txn->mt_next_pgno - 1;
2726 meta.mm_txnid = txn->mt_txnid;
2729 off += env->me_psize;
2732 /* Write to the SYNC fd */
2733 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2734 env->me_fd : env->me_mfd;
2737 memset(&ov, 0, sizeof(ov));
2739 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2742 rc = pwrite(mfd, ptr, len, off);
2747 DPUTS("write failed, disk error?");
2748 /* On a failure, the pagecache still contains the new data.
2749 * Write some old data back, to prevent it from being used.
2750 * Use the non-SYNC fd; we know it will fail anyway.
2752 meta.mm_last_pg = metab.mm_last_pg;
2753 meta.mm_txnid = metab.mm_txnid;
2755 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2757 r2 = pwrite(env->me_fd, ptr, len, off);
2760 env->me_flags |= MDB_FATAL_ERROR;
2764 /* Memory ordering issues are irrelevant; since the entire writer
2765 * is wrapped by wmutex, all of these changes will become visible
2766 * after the wmutex is unlocked. Since the DB is multi-version,
2767 * readers will get consistent data regardless of how fresh or
2768 * how stale their view of these values is.
2770 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2775 /** Check both meta pages to see which one is newer.
2776 * @param[in] env the environment handle
2777 * @return meta toggle (0 or 1).
2780 mdb_env_pick_meta(const MDB_env *env)
2782 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2786 mdb_env_create(MDB_env **env)
2790 e = calloc(1, sizeof(MDB_env));
2794 e->me_maxreaders = DEFAULT_READERS;
2795 e->me_maxdbs = e->me_numdbs = 2;
2796 e->me_fd = INVALID_HANDLE_VALUE;
2797 e->me_lfd = INVALID_HANDLE_VALUE;
2798 e->me_mfd = INVALID_HANDLE_VALUE;
2799 #ifdef MDB_USE_POSIX_SEM
2800 e->me_rmutex = SEM_FAILED;
2801 e->me_wmutex = SEM_FAILED;
2803 e->me_pid = getpid();
2804 VGMEMP_CREATE(e,0,0);
2810 mdb_env_set_mapsize(MDB_env *env, size_t size)
2814 env->me_mapsize = size;
2816 env->me_maxpg = env->me_mapsize / env->me_psize;
2821 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2825 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2830 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2832 if (env->me_map || readers < 1)
2834 env->me_maxreaders = readers;
2839 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2841 if (!env || !readers)
2843 *readers = env->me_maxreaders;
2847 /** Further setup required for opening an MDB environment
2850 mdb_env_open2(MDB_env *env)
2852 unsigned int flags = env->me_flags;
2853 int i, newenv = 0, prot;
2857 memset(&meta, 0, sizeof(meta));
2859 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2862 DPUTS("new mdbenv");
2866 /* Was a mapsize configured? */
2867 if (!env->me_mapsize) {
2868 /* If this is a new environment, take the default,
2869 * else use the size recorded in the existing env.
2871 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2872 } else if (env->me_mapsize < meta.mm_mapsize) {
2873 /* If the configured size is smaller, make sure it's
2874 * still big enough. Silently round up to minimum if not.
2876 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2877 if (env->me_mapsize < minsize)
2878 env->me_mapsize = minsize;
2884 LONG sizelo, sizehi;
2885 sizelo = env->me_mapsize & 0xffffffff;
2886 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2888 /* Windows won't create mappings for zero length files.
2889 * Just allocate the maxsize right now.
2892 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2893 if (!SetEndOfFile(env->me_fd))
2895 SetFilePointer(env->me_fd, 0, NULL, 0);
2897 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2898 PAGE_READWRITE : PAGE_READONLY,
2899 sizehi, sizelo, NULL);
2902 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2903 FILE_MAP_WRITE : FILE_MAP_READ,
2904 0, 0, env->me_mapsize, meta.mm_address);
2912 if (flags & MDB_WRITEMAP) {
2914 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2917 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2919 if (env->me_map == MAP_FAILED) {
2923 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2925 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2927 #ifdef POSIX_MADV_RANDOM
2928 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2929 #endif /* POSIX_MADV_RANDOM */
2930 #endif /* MADV_RANDOM */
2934 if (flags & MDB_FIXEDMAP)
2935 meta.mm_address = env->me_map;
2936 i = mdb_env_init_meta(env, &meta);
2937 if (i != MDB_SUCCESS) {
2940 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2941 /* Can happen because the address argument to mmap() is just a
2942 * hint. mmap() can pick another, e.g. if the range is in use.
2943 * The MAP_FIXED flag would prevent that, but then mmap could
2944 * instead unmap existing pages to make room for the new map.
2946 return EBUSY; /* TODO: Make a new MDB_* error code? */
2948 env->me_psize = meta.mm_psize;
2949 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2950 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2952 env->me_maxpg = env->me_mapsize / env->me_psize;
2954 p = (MDB_page *)env->me_map;
2955 env->me_metas[0] = METADATA(p);
2956 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2960 int toggle = mdb_env_pick_meta(env);
2961 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2963 DPRINTF("opened database version %u, pagesize %u",
2964 env->me_metas[0]->mm_version, env->me_psize);
2965 DPRINTF("using meta page %d", toggle);
2966 DPRINTF("depth: %u", db->md_depth);
2967 DPRINTF("entries: %zu", db->md_entries);
2968 DPRINTF("branch pages: %zu", db->md_branch_pages);
2969 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2970 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2971 DPRINTF("root: %zu", db->md_root);
2979 /** Release a reader thread's slot in the reader lock table.
2980 * This function is called automatically when a thread exits.
2981 * @param[in] ptr This points to the slot in the reader lock table.
2984 mdb_env_reader_dest(void *ptr)
2986 MDB_reader *reader = ptr;
2992 /** Junk for arranging thread-specific callbacks on Windows. This is
2993 * necessarily platform and compiler-specific. Windows supports up
2994 * to 1088 keys. Let's assume nobody opens more than 64 environments
2995 * in a single process, for now. They can override this if needed.
2997 #ifndef MAX_TLS_KEYS
2998 #define MAX_TLS_KEYS 64
3000 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3001 static int mdb_tls_nkeys;
3003 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3007 case DLL_PROCESS_ATTACH: break;
3008 case DLL_THREAD_ATTACH: break;
3009 case DLL_THREAD_DETACH:
3010 for (i=0; i<mdb_tls_nkeys; i++) {
3011 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3012 mdb_env_reader_dest(r);
3015 case DLL_PROCESS_DETACH: break;
3020 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3022 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3026 /* Force some symbol references.
3027 * _tls_used forces the linker to create the TLS directory if not already done
3028 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3030 #pragma comment(linker, "/INCLUDE:_tls_used")
3031 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3032 #pragma const_seg(".CRT$XLB")
3033 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3034 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3037 #pragma comment(linker, "/INCLUDE:__tls_used")
3038 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3039 #pragma data_seg(".CRT$XLB")
3040 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3042 #endif /* WIN 32/64 */
3043 #endif /* !__GNUC__ */
3046 /** Downgrade the exclusive lock on the region back to shared */
3048 mdb_env_share_locks(MDB_env *env, int *excl)
3050 int rc = 0, toggle = mdb_env_pick_meta(env);
3052 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3057 /* First acquire a shared lock. The Unlock will
3058 * then release the existing exclusive lock.
3060 memset(&ov, 0, sizeof(ov));
3061 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3064 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3070 struct flock lock_info;
3071 /* The shared lock replaces the existing lock */
3072 memset((void *)&lock_info, 0, sizeof(lock_info));
3073 lock_info.l_type = F_RDLCK;
3074 lock_info.l_whence = SEEK_SET;
3075 lock_info.l_start = 0;
3076 lock_info.l_len = 1;
3077 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3078 (rc = ErrCode()) == EINTR) ;
3079 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3086 /** Try to get exlusive lock, otherwise shared.
3087 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3090 mdb_env_excl_lock(MDB_env *env, int *excl)
3094 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3098 memset(&ov, 0, sizeof(ov));
3099 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3106 struct flock lock_info;
3107 memset((void *)&lock_info, 0, sizeof(lock_info));
3108 lock_info.l_type = F_WRLCK;
3109 lock_info.l_whence = SEEK_SET;
3110 lock_info.l_start = 0;
3111 lock_info.l_len = 1;
3112 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3113 (rc = ErrCode()) == EINTR) ;
3117 # ifdef MDB_USE_POSIX_SEM
3118 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3121 lock_info.l_type = F_RDLCK;
3122 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3123 (rc = ErrCode()) == EINTR) ;
3131 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3133 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3135 * @(#) $Revision: 5.1 $
3136 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3137 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3139 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3143 * Please do not copyright this code. This code is in the public domain.
3145 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3146 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3147 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3148 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3149 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3150 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3151 * PERFORMANCE OF THIS SOFTWARE.
3154 * chongo <Landon Curt Noll> /\oo/\
3155 * http://www.isthe.com/chongo/
3157 * Share and Enjoy! :-)
3160 typedef unsigned long long mdb_hash_t;
3161 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3163 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3164 * @param[in] str string to hash
3165 * @param[in] hval initial value for hash
3166 * @return 64 bit hash
3168 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3169 * hval arg on the first call.
3172 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3174 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3175 unsigned char *end = s + val->mv_size;
3177 * FNV-1a hash each octet of the string
3180 /* xor the bottom with the current octet */
3181 hval ^= (mdb_hash_t)*s++;
3183 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3184 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3185 (hval << 7) + (hval << 8) + (hval << 40);
3187 /* return our new hash value */
3191 /** Hash the string and output the hash in hex.
3192 * @param[in] str string to hash
3193 * @param[out] hexbuf an array of 17 chars to hold the hash
3196 mdb_hash_hex(MDB_val *val, char *hexbuf)
3199 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3200 for (i=0; i<8; i++) {
3201 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3207 /** Open and/or initialize the lock region for the environment.
3208 * @param[in] env The MDB environment.
3209 * @param[in] lpath The pathname of the file used for the lock region.
3210 * @param[in] mode The Unix permissions for the file, if we create it.
3211 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3212 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3213 * @return 0 on success, non-zero on failure.
3216 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3219 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3221 # define MDB_ERRCODE_ROFS EROFS
3222 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3223 # define MDB_CLOEXEC O_CLOEXEC
3226 # define MDB_CLOEXEC 0
3233 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3234 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3235 FILE_ATTRIBUTE_NORMAL, NULL);
3237 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3239 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3241 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3246 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3247 /* Lose record locks when exec*() */
3248 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3249 fcntl(env->me_lfd, F_SETFD, fdflags);
3252 if (!(env->me_flags & MDB_NOTLS)) {
3253 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3256 env->me_flags |= MDB_ENV_TXKEY;
3258 /* Windows TLS callbacks need help finding their TLS info. */
3259 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
3263 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3267 /* Try to get exclusive lock. If we succeed, then
3268 * nobody is using the lock region and we should initialize it.
3270 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3273 size = GetFileSize(env->me_lfd, NULL);
3275 size = lseek(env->me_lfd, 0, SEEK_END);
3277 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3278 if (size < rsize && *excl > 0) {
3280 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3281 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3283 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3287 size = rsize - sizeof(MDB_txninfo);
3288 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3293 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3295 if (!mh) goto fail_errno;
3296 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3298 if (!env->me_txns) goto fail_errno;
3300 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3302 if (m == MAP_FAILED) goto fail_errno;
3308 BY_HANDLE_FILE_INFORMATION stbuf;
3317 if (!mdb_sec_inited) {
3318 InitializeSecurityDescriptor(&mdb_null_sd,
3319 SECURITY_DESCRIPTOR_REVISION);
3320 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3321 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3322 mdb_all_sa.bInheritHandle = FALSE;
3323 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3326 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3327 idbuf.volume = stbuf.dwVolumeSerialNumber;
3328 idbuf.nhigh = stbuf.nFileIndexHigh;
3329 idbuf.nlow = stbuf.nFileIndexLow;
3330 val.mv_data = &idbuf;
3331 val.mv_size = sizeof(idbuf);
3332 mdb_hash_hex(&val, hexbuf);
3333 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3334 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3335 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3336 if (!env->me_rmutex) goto fail_errno;
3337 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3338 if (!env->me_wmutex) goto fail_errno;
3339 #elif defined(MDB_USE_POSIX_SEM)
3348 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3349 idbuf.dev = stbuf.st_dev;
3350 idbuf.ino = stbuf.st_ino;
3351 val.mv_data = &idbuf;
3352 val.mv_size = sizeof(idbuf);
3353 mdb_hash_hex(&val, hexbuf);
3354 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3355 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3356 /* Clean up after a previous run, if needed: Try to
3357 * remove both semaphores before doing anything else.
3359 sem_unlink(env->me_txns->mti_rmname);
3360 sem_unlink(env->me_txns->mti_wmname);
3361 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3362 O_CREAT|O_EXCL, mode, 1);
3363 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3364 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3365 O_CREAT|O_EXCL, mode, 1);
3366 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3367 #else /* MDB_USE_POSIX_SEM */
3368 pthread_mutexattr_t mattr;
3370 if ((rc = pthread_mutexattr_init(&mattr))
3371 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3372 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3373 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3375 pthread_mutexattr_destroy(&mattr);
3376 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3378 env->me_txns->mti_version = MDB_VERSION;
3379 env->me_txns->mti_magic = MDB_MAGIC;
3380 env->me_txns->mti_txnid = 0;
3381 env->me_txns->mti_numreaders = 0;
3384 if (env->me_txns->mti_magic != MDB_MAGIC) {
3385 DPUTS("lock region has invalid magic");
3389 if (env->me_txns->mti_version != MDB_VERSION) {
3390 DPRINTF("lock region is version %u, expected version %u",
3391 env->me_txns->mti_version, MDB_VERSION);
3392 rc = MDB_VERSION_MISMATCH;
3396 if (rc != EACCES && rc != EAGAIN) {
3400 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3401 if (!env->me_rmutex) goto fail_errno;
3402 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3403 if (!env->me_wmutex) goto fail_errno;
3404 #elif defined(MDB_USE_POSIX_SEM)
3405 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3406 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3407 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3408 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3419 /** The name of the lock file in the DB environment */
3420 #define LOCKNAME "/lock.mdb"
3421 /** The name of the data file in the DB environment */
3422 #define DATANAME "/data.mdb"
3423 /** The suffix of the lock file when no subdir is used */
3424 #define LOCKSUFF "-lock"
3425 /** Only a subset of the @ref mdb_env flags can be changed
3426 * at runtime. Changing other flags requires closing the
3427 * environment and re-opening it with the new flags.
3429 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3430 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP|MDB_NOTLS)
3433 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3435 int oflags, rc, len, excl = -1;
3436 char *lpath, *dpath;
3438 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3442 if (flags & MDB_NOSUBDIR) {
3443 rc = len + sizeof(LOCKSUFF) + len + 1;
3445 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3450 if (flags & MDB_NOSUBDIR) {
3451 dpath = lpath + len + sizeof(LOCKSUFF);
3452 sprintf(lpath, "%s" LOCKSUFF, path);
3453 strcpy(dpath, path);
3455 dpath = lpath + len + sizeof(LOCKNAME);
3456 sprintf(lpath, "%s" LOCKNAME, path);
3457 sprintf(dpath, "%s" DATANAME, path);
3461 flags |= env->me_flags;
3462 if (flags & MDB_RDONLY) {
3463 /* silently ignore WRITEMAP when we're only getting read access */
3464 flags &= ~MDB_WRITEMAP;
3466 if (!((env->me_free_pgs = mdb_midl_alloc()) &&
3467 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3470 env->me_flags = flags |= MDB_ENV_ACTIVE;
3474 env->me_path = strdup(path);
3475 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3476 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3477 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3482 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3487 if (F_ISSET(flags, MDB_RDONLY)) {
3488 oflags = GENERIC_READ;
3489 len = OPEN_EXISTING;
3491 oflags = GENERIC_READ|GENERIC_WRITE;
3494 mode = FILE_ATTRIBUTE_NORMAL;
3495 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3496 NULL, len, mode, NULL);
3498 if (F_ISSET(flags, MDB_RDONLY))
3501 oflags = O_RDWR | O_CREAT;
3503 env->me_fd = open(dpath, oflags, mode);
3505 if (env->me_fd == INVALID_HANDLE_VALUE) {
3510 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3511 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3512 env->me_mfd = env->me_fd;
3514 /* Synchronous fd for meta writes. Needed even with
3515 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3518 env->me_mfd = CreateFile(dpath, oflags,
3519 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3520 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3522 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3524 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3529 DPRINTF("opened dbenv %p", (void *) env);
3531 rc = mdb_env_share_locks(env, &excl);
3537 mdb_env_close0(env, excl);
3543 /** Destroy resources from mdb_env_open() and clear our readers */
3545 mdb_env_close0(MDB_env *env, int excl)
3549 if (!(env->me_flags & MDB_ENV_ACTIVE))
3552 free(env->me_dbflags);
3555 free(env->me_dirty_list);
3556 if (env->me_free_pgs)
3557 mdb_midl_free(env->me_free_pgs);
3559 if (env->me_flags & MDB_ENV_TXKEY) {
3560 pthread_key_delete(env->me_txkey);
3562 /* Delete our key from the global list */
3563 for (i=0; i<mdb_tls_nkeys; i++)
3564 if (mdb_tls_keys[i] == env->me_txkey) {
3565 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3573 munmap(env->me_map, env->me_mapsize);
3575 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3577 if (env->me_fd != INVALID_HANDLE_VALUE)
3580 pid_t pid = env->me_pid;
3581 /* Clearing readers is done in this function because
3582 * me_txkey with its destructor must be disabled first.
3584 for (i = env->me_numreaders; --i >= 0; )
3585 if (env->me_txns->mti_readers[i].mr_pid == pid)
3586 env->me_txns->mti_readers[i].mr_pid = 0;
3588 if (env->me_rmutex) {
3589 CloseHandle(env->me_rmutex);
3590 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3592 /* Windows automatically destroys the mutexes when
3593 * the last handle closes.
3595 #elif defined(MDB_USE_POSIX_SEM)
3596 if (env->me_rmutex != SEM_FAILED) {
3597 sem_close(env->me_rmutex);
3598 if (env->me_wmutex != SEM_FAILED)
3599 sem_close(env->me_wmutex);
3600 /* If we have the filelock: If we are the
3601 * only remaining user, clean up semaphores.
3604 mdb_env_excl_lock(env, &excl);
3606 sem_unlink(env->me_txns->mti_rmname);
3607 sem_unlink(env->me_txns->mti_wmname);
3611 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3613 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3616 /* Unlock the lockfile. Windows would have unlocked it
3617 * after closing anyway, but not necessarily at once.
3619 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3625 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
3629 mdb_env_copy(MDB_env *env, const char *path)
3631 MDB_txn *txn = NULL;
3635 HANDLE newfd = INVALID_HANDLE_VALUE;
3637 if (env->me_flags & MDB_NOSUBDIR) {
3638 lpath = (char *)path;
3641 len += sizeof(DATANAME);
3642 lpath = malloc(len);
3645 sprintf(lpath, "%s" DATANAME, path);
3648 /* The destination path must exist, but the destination file must not.
3649 * We don't want the OS to cache the writes, since the source data is
3650 * already in the OS cache.
3653 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3654 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3656 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3662 if (!(env->me_flags & MDB_NOSUBDIR))
3664 if (newfd == INVALID_HANDLE_VALUE) {
3669 #ifdef F_NOCACHE /* __APPLE__ */
3670 rc = fcntl(newfd, F_NOCACHE, 1);
3677 /* Do the lock/unlock of the reader mutex before starting the
3678 * write txn. Otherwise other read txns could block writers.
3680 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3685 /* We must start the actual read txn after blocking writers */
3686 mdb_txn_reset0(txn);
3688 /* Temporarily block writers until we snapshot the meta pages */
3691 rc = mdb_txn_renew0(txn);
3693 UNLOCK_MUTEX_W(env);
3698 wsize = env->me_psize * 2;
3702 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3703 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3706 rc = write(newfd, env->me_map, wsize);
3707 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3710 UNLOCK_MUTEX_W(env);
3715 ptr = env->me_map + wsize;
3716 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3717 #define MAX_WRITE 2147483648U
3721 if (wsize > MAX_WRITE)
3725 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3726 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3735 if (wsize > MAX_WRITE)
3739 wres = write(newfd, ptr, w2);
3740 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3749 if (newfd != INVALID_HANDLE_VALUE)
3756 mdb_env_close(MDB_env *env)
3764 for (i = env->me_numdbs; --i > MAIN_DBI; )
3765 free(env->me_dbxs[i].md_name.mv_data);
3767 VGMEMP_DESTROY(env);
3768 while ((dp = env->me_dpages) != NULL) {
3769 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3770 env->me_dpages = dp->mp_next;
3774 mdb_env_close0(env, 0);
3778 /** Compare two items pointing at aligned size_t's */
3780 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3782 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3783 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3786 /** Compare two items pointing at aligned int's */
3788 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3790 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3791 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3794 /** Compare two items pointing at ints of unknown alignment.
3795 * Nodes and keys are guaranteed to be 2-byte aligned.
3798 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3800 #if BYTE_ORDER == LITTLE_ENDIAN
3801 unsigned short *u, *c;
3804 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3805 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3808 } while(!x && u > (unsigned short *)a->mv_data);
3811 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3815 /** Compare two items lexically */
3817 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3824 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3830 diff = memcmp(a->mv_data, b->mv_data, len);
3831 return diff ? diff : len_diff<0 ? -1 : len_diff;
3834 /** Compare two items in reverse byte order */
3836 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3838 const unsigned char *p1, *p2, *p1_lim;
3842 p1_lim = (const unsigned char *)a->mv_data;
3843 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3844 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3846 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3852 while (p1 > p1_lim) {
3853 diff = *--p1 - *--p2;
3857 return len_diff<0 ? -1 : len_diff;
3860 /** Search for key within a page, using binary search.
3861 * Returns the smallest entry larger or equal to the key.
3862 * If exactp is non-null, stores whether the found entry was an exact match
3863 * in *exactp (1 or 0).
3864 * Updates the cursor index with the index of the found entry.
3865 * If no entry larger or equal to the key is found, returns NULL.
3868 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3870 unsigned int i = 0, nkeys;
3873 MDB_page *mp = mc->mc_pg[mc->mc_top];
3874 MDB_node *node = NULL;
3879 nkeys = NUMKEYS(mp);
3884 COPY_PGNO(pgno, mp->mp_pgno);
3885 DPRINTF("searching %u keys in %s %spage %zu",
3886 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3893 low = IS_LEAF(mp) ? 0 : 1;
3895 cmp = mc->mc_dbx->md_cmp;
3897 /* Branch pages have no data, so if using integer keys,
3898 * alignment is guaranteed. Use faster mdb_cmp_int.
3900 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3901 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3908 nodekey.mv_size = mc->mc_db->md_pad;
3909 node = NODEPTR(mp, 0); /* fake */
3910 while (low <= high) {
3911 i = (low + high) >> 1;
3912 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3913 rc = cmp(key, &nodekey);
3914 DPRINTF("found leaf index %u [%s], rc = %i",
3915 i, DKEY(&nodekey), rc);
3924 while (low <= high) {
3925 i = (low + high) >> 1;
3927 node = NODEPTR(mp, i);
3928 nodekey.mv_size = NODEKSZ(node);
3929 nodekey.mv_data = NODEKEY(node);
3931 rc = cmp(key, &nodekey);
3934 DPRINTF("found leaf index %u [%s], rc = %i",
3935 i, DKEY(&nodekey), rc);
3937 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3938 i, DKEY(&nodekey), NODEPGNO(node), rc);
3949 if (rc > 0) { /* Found entry is less than the key. */
3950 i++; /* Skip to get the smallest entry larger than key. */
3952 node = NODEPTR(mp, i);
3955 *exactp = (rc == 0);
3956 /* store the key index */
3957 mc->mc_ki[mc->mc_top] = i;
3959 /* There is no entry larger or equal to the key. */
3962 /* nodeptr is fake for LEAF2 */
3968 mdb_cursor_adjust(MDB_cursor *mc, func)
3972 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3973 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3980 /** Pop a page off the top of the cursor's stack. */
3982 mdb_cursor_pop(MDB_cursor *mc)
3985 #ifndef MDB_DEBUG_SKIP
3986 MDB_page *top = mc->mc_pg[mc->mc_top];
3992 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3993 mc->mc_dbi, (void *) mc);
3997 /** Push a page onto the top of the cursor's stack. */
3999 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4001 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
4002 mc->mc_dbi, (void *) mc);
4004 if (mc->mc_snum >= CURSOR_STACK) {
4005 assert(mc->mc_snum < CURSOR_STACK);
4006 return MDB_CURSOR_FULL;
4009 mc->mc_top = mc->mc_snum++;
4010 mc->mc_pg[mc->mc_top] = mp;
4011 mc->mc_ki[mc->mc_top] = 0;
4016 /** Find the address of the page corresponding to a given page number.
4017 * @param[in] txn the transaction for this access.
4018 * @param[in] pgno the page number for the page to retrieve.
4019 * @param[out] ret address of a pointer where the page's address will be stored.
4020 * @return 0 on success, non-zero on failure.
4023 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
4027 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4028 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4032 MDB_ID2L dl = tx2->mt_u.dirty_list;
4034 unsigned x = mdb_mid2l_search(dl, pgno);
4035 if (x <= dl[0].mid && dl[x].mid == pgno) {
4040 } while ((tx2 = tx2->mt_parent) != NULL);
4043 if (pgno < txn->mt_next_pgno) {
4044 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4046 DPRINTF("page %zu not found", pgno);
4052 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
4055 /** Search for the page a given key should be in.
4056 * Pushes parent pages on the cursor stack. This function continues a
4057 * search on a cursor that has already been initialized. (Usually by
4058 * #mdb_page_search() but also by #mdb_node_move().)
4059 * @param[in,out] mc the cursor for this operation.
4060 * @param[in] key the key to search for. If NULL, search for the lowest
4061 * page. (This is used by #mdb_cursor_first().)
4062 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4063 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4064 * @return 0 on success, non-zero on failure.
4067 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4069 MDB_page *mp = mc->mc_pg[mc->mc_top];
4074 while (IS_BRANCH(mp)) {
4078 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4079 assert(NUMKEYS(mp) > 1);
4080 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4082 if (key == NULL) /* Initialize cursor to first page. */
4084 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4085 /* cursor to last page */
4089 node = mdb_node_search(mc, key, &exact);
4091 i = NUMKEYS(mp) - 1;
4093 i = mc->mc_ki[mc->mc_top];
4102 DPRINTF("following index %u for key [%s]",
4104 assert(i < NUMKEYS(mp));
4105 node = NODEPTR(mp, i);
4107 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
4110 mc->mc_ki[mc->mc_top] = i;
4111 if ((rc = mdb_cursor_push(mc, mp)))
4115 if ((rc = mdb_page_touch(mc)) != 0)
4117 mp = mc->mc_pg[mc->mc_top];
4122 DPRINTF("internal error, index points to a %02X page!?",
4124 return MDB_CORRUPTED;
4127 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4128 key ? DKEY(key) : NULL);
4133 /** Search for the page a given key should be in.
4134 * Pushes parent pages on the cursor stack. This function just sets up
4135 * the search; it finds the root page for \b mc's database and sets this
4136 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4137 * called to complete the search.
4138 * @param[in,out] mc the cursor for this operation.
4139 * @param[in] key the key to search for. If NULL, search for the lowest
4140 * page. (This is used by #mdb_cursor_first().)
4141 * @param[in] modify If true, visited pages are updated with new page numbers.
4142 * @return 0 on success, non-zero on failure.
4145 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4150 /* Make sure the txn is still viable, then find the root from
4151 * the txn's db table.
4153 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4154 DPUTS("transaction has failed, must abort");
4157 /* Make sure we're using an up-to-date root */
4158 if (mc->mc_dbi > MAIN_DBI) {
4159 if ((*mc->mc_dbflag & DB_STALE) ||
4160 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4162 unsigned char dbflag = 0;
4163 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4164 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4167 if (*mc->mc_dbflag & DB_STALE) {
4171 MDB_node *leaf = mdb_node_search(&mc2,
4172 &mc->mc_dbx->md_name, &exact);
4174 return MDB_NOTFOUND;
4175 mdb_node_read(mc->mc_txn, leaf, &data);
4176 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4178 /* The txn may not know this DBI, or another process may
4179 * have dropped and recreated the DB with other flags.
4181 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4182 return MDB_INCOMPATIBLE;
4183 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4185 if (flags & MDB_PS_MODIFY)
4187 *mc->mc_dbflag &= ~DB_STALE;
4188 *mc->mc_dbflag |= dbflag;
4191 root = mc->mc_db->md_root;
4193 if (root == P_INVALID) { /* Tree is empty. */
4194 DPUTS("tree is empty");
4195 return MDB_NOTFOUND;
4200 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4201 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4207 DPRINTF("db %u root page %zu has flags 0x%X",
4208 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4210 if (flags & MDB_PS_MODIFY) {
4211 if ((rc = mdb_page_touch(mc)))
4215 if (flags & MDB_PS_ROOTONLY)
4218 return mdb_page_search_root(mc, key, flags);
4221 /** Return the data associated with a given node.
4222 * @param[in] txn The transaction for this operation.
4223 * @param[in] leaf The node being read.
4224 * @param[out] data Updated to point to the node's data.
4225 * @return 0 on success, non-zero on failure.
4228 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4230 MDB_page *omp; /* overflow page */
4234 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4235 data->mv_size = NODEDSZ(leaf);
4236 data->mv_data = NODEDATA(leaf);
4240 /* Read overflow data.
4242 data->mv_size = NODEDSZ(leaf);
4243 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4244 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4245 DPRINTF("read overflow page %zu failed", pgno);
4248 data->mv_data = METADATA(omp);
4254 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4255 MDB_val *key, MDB_val *data)
4264 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4266 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4269 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4273 mdb_cursor_init(&mc, txn, dbi, &mx);
4274 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4277 /** Find a sibling for a page.
4278 * Replaces the page at the top of the cursor's stack with the
4279 * specified sibling, if one exists.
4280 * @param[in] mc The cursor for this operation.
4281 * @param[in] move_right Non-zero if the right sibling is requested,
4282 * otherwise the left sibling.
4283 * @return 0 on success, non-zero on failure.
4286 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4292 if (mc->mc_snum < 2) {
4293 return MDB_NOTFOUND; /* root has no siblings */
4297 DPRINTF("parent page is page %zu, index %u",
4298 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4300 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4301 : (mc->mc_ki[mc->mc_top] == 0)) {
4302 DPRINTF("no more keys left, moving to %s sibling",
4303 move_right ? "right" : "left");
4304 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4305 /* undo cursor_pop before returning */
4312 mc->mc_ki[mc->mc_top]++;
4314 mc->mc_ki[mc->mc_top]--;
4315 DPRINTF("just moving to %s index key %u",
4316 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4318 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4320 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4321 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4324 mdb_cursor_push(mc, mp);
4326 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4331 /** Move the cursor to the next data item. */
4333 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4339 if (mc->mc_flags & C_EOF) {
4340 return MDB_NOTFOUND;
4343 assert(mc->mc_flags & C_INITIALIZED);
4345 mp = mc->mc_pg[mc->mc_top];
4347 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4348 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4349 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4350 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4351 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4352 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4356 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4357 if (op == MDB_NEXT_DUP)
4358 return MDB_NOTFOUND;
4362 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4364 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4365 DPUTS("=====> move to next sibling page");
4366 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4367 mc->mc_flags |= C_EOF;
4368 mc->mc_flags &= ~C_INITIALIZED;
4369 return MDB_NOTFOUND;
4371 mp = mc->mc_pg[mc->mc_top];
4372 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4374 mc->mc_ki[mc->mc_top]++;
4376 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4377 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4380 key->mv_size = mc->mc_db->md_pad;
4381 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4385 assert(IS_LEAF(mp));
4386 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4388 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4389 mdb_xcursor_init1(mc, leaf);
4392 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4395 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4396 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4397 if (rc != MDB_SUCCESS)
4402 MDB_GET_KEY(leaf, key);
4406 /** Move the cursor to the previous data item. */
4408 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4414 assert(mc->mc_flags & C_INITIALIZED);
4416 mp = mc->mc_pg[mc->mc_top];
4418 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4419 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4420 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4421 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4422 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4423 if (op != MDB_PREV || rc == MDB_SUCCESS)
4426 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4427 if (op == MDB_PREV_DUP)
4428 return MDB_NOTFOUND;
4433 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4435 if (mc->mc_ki[mc->mc_top] == 0) {
4436 DPUTS("=====> move to prev sibling page");
4437 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4438 mc->mc_flags &= ~C_INITIALIZED;
4439 return MDB_NOTFOUND;
4441 mp = mc->mc_pg[mc->mc_top];
4442 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4443 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4445 mc->mc_ki[mc->mc_top]--;
4447 mc->mc_flags &= ~C_EOF;
4449 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4450 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4453 key->mv_size = mc->mc_db->md_pad;
4454 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4458 assert(IS_LEAF(mp));
4459 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4461 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4462 mdb_xcursor_init1(mc, leaf);
4465 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4468 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4469 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4470 if (rc != MDB_SUCCESS)
4475 MDB_GET_KEY(leaf, key);
4479 /** Set the cursor on a specific data item. */
4481 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4482 MDB_cursor_op op, int *exactp)
4486 MDB_node *leaf = NULL;
4491 assert(key->mv_size > 0);
4493 /* See if we're already on the right page */
4494 if (mc->mc_flags & C_INITIALIZED) {
4497 mp = mc->mc_pg[mc->mc_top];
4499 mc->mc_ki[mc->mc_top] = 0;
4500 return MDB_NOTFOUND;
4502 if (mp->mp_flags & P_LEAF2) {
4503 nodekey.mv_size = mc->mc_db->md_pad;
4504 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4506 leaf = NODEPTR(mp, 0);
4507 MDB_GET_KEY(leaf, &nodekey);
4509 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4511 /* Probably happens rarely, but first node on the page
4512 * was the one we wanted.
4514 mc->mc_ki[mc->mc_top] = 0;
4521 unsigned int nkeys = NUMKEYS(mp);
4523 if (mp->mp_flags & P_LEAF2) {
4524 nodekey.mv_data = LEAF2KEY(mp,
4525 nkeys-1, nodekey.mv_size);
4527 leaf = NODEPTR(mp, nkeys-1);
4528 MDB_GET_KEY(leaf, &nodekey);
4530 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4532 /* last node was the one we wanted */
4533 mc->mc_ki[mc->mc_top] = nkeys-1;
4539 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4540 /* This is definitely the right page, skip search_page */
4541 if (mp->mp_flags & P_LEAF2) {
4542 nodekey.mv_data = LEAF2KEY(mp,
4543 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4545 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4546 MDB_GET_KEY(leaf, &nodekey);
4548 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4550 /* current node was the one we wanted */
4560 /* If any parents have right-sibs, search.
4561 * Otherwise, there's nothing further.
4563 for (i=0; i<mc->mc_top; i++)
4565 NUMKEYS(mc->mc_pg[i])-1)
4567 if (i == mc->mc_top) {
4568 /* There are no other pages */
4569 mc->mc_ki[mc->mc_top] = nkeys;
4570 return MDB_NOTFOUND;
4574 /* There are no other pages */
4575 mc->mc_ki[mc->mc_top] = 0;
4576 return MDB_NOTFOUND;
4580 rc = mdb_page_search(mc, key, 0);
4581 if (rc != MDB_SUCCESS)
4584 mp = mc->mc_pg[mc->mc_top];
4585 assert(IS_LEAF(mp));
4588 leaf = mdb_node_search(mc, key, exactp);
4589 if (exactp != NULL && !*exactp) {
4590 /* MDB_SET specified and not an exact match. */
4591 return MDB_NOTFOUND;
4595 DPUTS("===> inexact leaf not found, goto sibling");
4596 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4597 return rc; /* no entries matched */
4598 mp = mc->mc_pg[mc->mc_top];
4599 assert(IS_LEAF(mp));
4600 leaf = NODEPTR(mp, 0);
4604 mc->mc_flags |= C_INITIALIZED;
4605 mc->mc_flags &= ~C_EOF;
4608 key->mv_size = mc->mc_db->md_pad;
4609 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4613 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4614 mdb_xcursor_init1(mc, leaf);
4617 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4618 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4619 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4622 if (op == MDB_GET_BOTH) {
4628 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4629 if (rc != MDB_SUCCESS)
4632 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4634 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4636 rc = mc->mc_dbx->md_dcmp(data, &d2);
4638 if (op == MDB_GET_BOTH || rc > 0)
4639 return MDB_NOTFOUND;
4644 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4645 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4650 /* The key already matches in all other cases */
4651 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4652 MDB_GET_KEY(leaf, key);
4653 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4658 /** Move the cursor to the first item in the database. */
4660 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4665 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4666 rc = mdb_page_search(mc, NULL, 0);
4667 if (rc != MDB_SUCCESS)
4670 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4672 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4673 mc->mc_flags |= C_INITIALIZED;
4674 mc->mc_flags &= ~C_EOF;
4676 mc->mc_ki[mc->mc_top] = 0;
4678 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4679 key->mv_size = mc->mc_db->md_pad;
4680 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4685 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4686 mdb_xcursor_init1(mc, leaf);
4687 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4692 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4693 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4697 MDB_GET_KEY(leaf, key);
4701 /** Move the cursor to the last item in the database. */
4703 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4708 if (!(mc->mc_flags & C_EOF)) {
4710 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4713 lkey.mv_size = MDB_MAXKEYSIZE+1;
4714 lkey.mv_data = NULL;
4715 rc = mdb_page_search(mc, &lkey, 0);
4716 if (rc != MDB_SUCCESS)
4719 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4721 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4723 mc->mc_flags |= C_INITIALIZED|C_EOF;
4724 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4726 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4727 key->mv_size = mc->mc_db->md_pad;
4728 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4733 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4734 mdb_xcursor_init1(mc, leaf);
4735 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4740 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4741 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4746 MDB_GET_KEY(leaf, key);
4751 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4760 case MDB_GET_CURRENT:
4761 if (!(mc->mc_flags & C_INITIALIZED)) {
4764 MDB_page *mp = mc->mc_pg[mc->mc_top];
4766 mc->mc_ki[mc->mc_top] = 0;
4772 key->mv_size = mc->mc_db->md_pad;
4773 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4775 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4776 MDB_GET_KEY(leaf, key);
4778 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4779 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4781 rc = mdb_node_read(mc->mc_txn, leaf, data);
4788 case MDB_GET_BOTH_RANGE:
4789 if (data == NULL || mc->mc_xcursor == NULL) {
4797 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4799 } else if (op == MDB_SET_RANGE)
4800 rc = mdb_cursor_set(mc, key, data, op, NULL);
4802 rc = mdb_cursor_set(mc, key, data, op, &exact);
4804 case MDB_GET_MULTIPLE:
4806 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4807 !(mc->mc_flags & C_INITIALIZED)) {
4812 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4813 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4816 case MDB_NEXT_MULTIPLE:
4818 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4822 if (!(mc->mc_flags & C_INITIALIZED))
4823 rc = mdb_cursor_first(mc, key, data);
4825 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4826 if (rc == MDB_SUCCESS) {
4827 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4830 mx = &mc->mc_xcursor->mx_cursor;
4831 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4833 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4834 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4842 case MDB_NEXT_NODUP:
4843 if (!(mc->mc_flags & C_INITIALIZED))
4844 rc = mdb_cursor_first(mc, key, data);
4846 rc = mdb_cursor_next(mc, key, data, op);
4850 case MDB_PREV_NODUP:
4851 if (!(mc->mc_flags & C_INITIALIZED)) {
4852 rc = mdb_cursor_last(mc, key, data);
4853 mc->mc_flags |= C_INITIALIZED;
4854 mc->mc_ki[mc->mc_top]++;
4856 rc = mdb_cursor_prev(mc, key, data, op);
4859 rc = mdb_cursor_first(mc, key, data);
4863 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4864 !(mc->mc_flags & C_INITIALIZED) ||
4865 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4869 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4872 rc = mdb_cursor_last(mc, key, data);
4876 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4877 !(mc->mc_flags & C_INITIALIZED) ||
4878 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4882 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4885 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4893 /** Touch all the pages in the cursor stack.
4894 * Makes sure all the pages are writable, before attempting a write operation.
4895 * @param[in] mc The cursor to operate on.
4898 mdb_cursor_touch(MDB_cursor *mc)
4902 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4905 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4906 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4907 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4910 *mc->mc_dbflag |= DB_DIRTY;
4912 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4913 rc = mdb_page_touch(mc);
4917 mc->mc_top = mc->mc_snum-1;
4922 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4925 MDB_node *leaf = NULL;
4926 MDB_val xdata, *rdata, dkey;
4929 int do_sub = 0, insert = 0;
4930 unsigned int mcount = 0;
4934 char dbuf[MDB_MAXKEYSIZE+1];
4935 unsigned int nflags;
4938 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4941 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4944 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4947 #if SIZE_MAX > MAXDATASIZE
4948 if (data->mv_size > MAXDATASIZE)
4952 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4953 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4957 if (flags == MDB_CURRENT) {
4958 if (!(mc->mc_flags & C_INITIALIZED))
4961 } else if (mc->mc_db->md_root == P_INVALID) {
4963 /* new database, write a root leaf page */
4964 DPUTS("allocating new root leaf page");
4965 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4969 mdb_cursor_push(mc, np);
4970 mc->mc_db->md_root = np->mp_pgno;
4971 mc->mc_db->md_depth++;
4972 *mc->mc_dbflag |= DB_DIRTY;
4973 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4975 np->mp_flags |= P_LEAF2;
4976 mc->mc_flags |= C_INITIALIZED;
4982 if (flags & MDB_APPEND) {
4984 rc = mdb_cursor_last(mc, &k2, &d2);
4986 rc = mc->mc_dbx->md_cmp(key, &k2);
4989 mc->mc_ki[mc->mc_top]++;
4991 /* new key is <= last key */
4996 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4998 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4999 DPRINTF("duplicate key [%s]", DKEY(key));
5001 return MDB_KEYEXIST;
5003 if (rc && rc != MDB_NOTFOUND)
5007 /* Cursor is positioned, now make sure all pages are writable */
5008 rc2 = mdb_cursor_touch(mc);
5013 /* The key already exists */
5014 if (rc == MDB_SUCCESS) {
5015 /* there's only a key anyway, so this is a no-op */
5016 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5017 unsigned int ksize = mc->mc_db->md_pad;
5018 if (key->mv_size != ksize)
5020 if (flags == MDB_CURRENT) {
5021 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5022 memcpy(ptr, key->mv_data, ksize);
5027 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5030 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5031 /* Was a single item before, must convert now */
5033 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5034 /* Just overwrite the current item */
5035 if (flags == MDB_CURRENT)
5038 dkey.mv_size = NODEDSZ(leaf);
5039 dkey.mv_data = NODEDATA(leaf);
5040 #if UINT_MAX < SIZE_MAX
5041 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5042 #ifdef MISALIGNED_OK
5043 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5045 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5048 /* if data matches, ignore it */
5049 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5050 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5052 /* create a fake page for the dup items */
5053 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5054 dkey.mv_data = dbuf;
5055 fp = (MDB_page *)&pbuf;
5056 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5057 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5058 fp->mp_lower = PAGEHDRSZ;
5059 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5060 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5061 fp->mp_flags |= P_LEAF2;
5062 fp->mp_pad = data->mv_size;
5063 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5065 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5066 (dkey.mv_size & 1) + (data->mv_size & 1);
5068 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5071 xdata.mv_size = fp->mp_upper;
5076 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5077 /* See if we need to convert from fake page to subDB */
5079 unsigned int offset;
5082 fp = NODEDATA(leaf);
5083 if (flags == MDB_CURRENT) {
5085 fp->mp_flags |= P_DIRTY;
5086 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5087 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5091 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5092 offset = fp->mp_pad;
5093 if (SIZELEFT(fp) >= offset)
5095 offset *= 4; /* space for 4 more */
5097 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5099 offset += offset & 1;
5100 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5101 offset >= mc->mc_txn->mt_env->me_nodemax) {
5102 /* yes, convert it */
5104 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5105 dummy.md_pad = fp->mp_pad;
5106 dummy.md_flags = MDB_DUPFIXED;
5107 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5108 dummy.md_flags |= MDB_INTEGERKEY;
5111 dummy.md_branch_pages = 0;
5112 dummy.md_leaf_pages = 1;
5113 dummy.md_overflow_pages = 0;
5114 dummy.md_entries = NUMKEYS(fp);
5116 xdata.mv_size = sizeof(MDB_db);
5117 xdata.mv_data = &dummy;
5118 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5120 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5121 flags |= F_DUPDATA|F_SUBDATA;
5122 dummy.md_root = mp->mp_pgno;
5124 /* no, just grow it */
5126 xdata.mv_size = NODEDSZ(leaf) + offset;
5127 xdata.mv_data = &pbuf;
5128 mp = (MDB_page *)&pbuf;
5129 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5132 mp->mp_flags = fp->mp_flags | P_DIRTY;
5133 mp->mp_pad = fp->mp_pad;
5134 mp->mp_lower = fp->mp_lower;
5135 mp->mp_upper = fp->mp_upper + offset;
5137 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5139 nsize = NODEDSZ(leaf) - fp->mp_upper;
5140 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5141 for (i=0; i<NUMKEYS(fp); i++)
5142 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5144 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5148 /* data is on sub-DB, just store it */
5149 flags |= F_DUPDATA|F_SUBDATA;
5153 /* overflow page overwrites need special handling */
5154 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5157 int ovpages, dpages;
5159 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5160 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5161 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5162 mdb_page_get(mc->mc_txn, pg, &omp);
5163 /* Is the ov page writable and large enough? */
5164 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5165 /* yes, overwrite it. Note in this case we don't
5166 * bother to try shrinking the node if the new data
5167 * is smaller than the overflow threshold.
5169 if (F_ISSET(flags, MDB_RESERVE))
5170 data->mv_data = METADATA(omp);
5172 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5175 /* no, free ovpages */
5177 mc->mc_db->md_overflow_pages -= ovpages;
5178 for (i=0; i<ovpages; i++) {
5179 DPRINTF("freed ov page %zu", pg);
5180 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5184 } else if (NODEDSZ(leaf) == data->mv_size) {
5185 /* same size, just replace it. Note that we could
5186 * also reuse this node if the new data is smaller,
5187 * but instead we opt to shrink the node in that case.
5189 if (F_ISSET(flags, MDB_RESERVE))
5190 data->mv_data = NODEDATA(leaf);
5191 else if (data->mv_size)
5192 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5194 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5197 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5198 mc->mc_db->md_entries--;
5200 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5207 nflags = flags & NODE_ADD_FLAGS;
5208 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5209 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5210 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5211 nflags &= ~MDB_APPEND;
5213 nflags |= MDB_SPLIT_REPLACE;
5214 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5216 /* There is room already in this leaf page. */
5217 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5218 if (rc == 0 && !do_sub && insert) {
5219 /* Adjust other cursors pointing to mp */
5220 MDB_cursor *m2, *m3;
5221 MDB_dbi dbi = mc->mc_dbi;
5222 unsigned i = mc->mc_top;
5223 MDB_page *mp = mc->mc_pg[i];
5225 if (mc->mc_flags & C_SUB)
5228 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5229 if (mc->mc_flags & C_SUB)
5230 m3 = &m2->mc_xcursor->mx_cursor;
5233 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5234 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5241 if (rc != MDB_SUCCESS)
5242 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5244 /* Now store the actual data in the child DB. Note that we're
5245 * storing the user data in the keys field, so there are strict
5246 * size limits on dupdata. The actual data fields of the child
5247 * DB are all zero size.
5254 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5255 if (flags & MDB_CURRENT) {
5256 xflags = MDB_CURRENT;
5258 mdb_xcursor_init1(mc, leaf);
5259 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5261 /* converted, write the original data first */
5263 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5267 /* Adjust other cursors pointing to mp */
5269 unsigned i = mc->mc_top;
5270 MDB_page *mp = mc->mc_pg[i];
5272 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5273 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5274 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5275 mdb_xcursor_init1(m2, leaf);
5279 /* we've done our job */
5282 if (flags & MDB_APPENDDUP)
5283 xflags |= MDB_APPEND;
5284 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5285 if (flags & F_SUBDATA) {
5286 void *db = NODEDATA(leaf);
5287 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5290 /* sub-writes might have failed so check rc again.
5291 * Don't increment count if we just replaced an existing item.
5293 if (!rc && !(flags & MDB_CURRENT))
5294 mc->mc_db->md_entries++;
5295 if (flags & MDB_MULTIPLE) {
5297 if (mcount < data[1].mv_size) {
5298 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5299 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5305 /* If we succeeded and the key didn't exist before, make sure
5306 * the cursor is marked valid.
5309 mc->mc_flags |= C_INITIALIZED;
5314 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5319 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5322 if (!(mc->mc_flags & C_INITIALIZED))
5325 rc = mdb_cursor_touch(mc);
5329 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5331 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5332 if (flags != MDB_NODUPDATA) {
5333 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5334 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5336 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5337 /* If sub-DB still has entries, we're done */
5338 if (mc->mc_xcursor->mx_db.md_entries) {
5339 if (leaf->mn_flags & F_SUBDATA) {
5340 /* update subDB info */
5341 void *db = NODEDATA(leaf);
5342 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5344 /* shrink fake page */
5345 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5347 mc->mc_db->md_entries--;
5350 /* otherwise fall thru and delete the sub-DB */
5353 if (leaf->mn_flags & F_SUBDATA) {
5354 /* add all the child DB's pages to the free list */
5355 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5356 if (rc == MDB_SUCCESS) {
5357 mc->mc_db->md_entries -=
5358 mc->mc_xcursor->mx_db.md_entries;
5363 return mdb_cursor_del0(mc, leaf);
5366 /** Allocate and initialize new pages for a database.
5367 * @param[in] mc a cursor on the database being added to.
5368 * @param[in] flags flags defining what type of page is being allocated.
5369 * @param[in] num the number of pages to allocate. This is usually 1,
5370 * unless allocating overflow pages for a large record.
5371 * @param[out] mp Address of a page, or NULL on failure.
5372 * @return 0 on success, non-zero on failure.
5375 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5380 if ((rc = mdb_page_alloc(mc, num, &np)))
5382 DPRINTF("allocated new mpage %zu, page size %u",
5383 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5384 np->mp_flags = flags | P_DIRTY;
5385 np->mp_lower = PAGEHDRSZ;
5386 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5389 mc->mc_db->md_branch_pages++;
5390 else if (IS_LEAF(np))
5391 mc->mc_db->md_leaf_pages++;
5392 else if (IS_OVERFLOW(np)) {
5393 mc->mc_db->md_overflow_pages += num;
5401 /** Calculate the size of a leaf node.
5402 * The size depends on the environment's page size; if a data item
5403 * is too large it will be put onto an overflow page and the node
5404 * size will only include the key and not the data. Sizes are always
5405 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5406 * of the #MDB_node headers.
5407 * @param[in] env The environment handle.
5408 * @param[in] key The key for the node.
5409 * @param[in] data The data for the node.
5410 * @return The number of bytes needed to store the node.
5413 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5417 sz = LEAFSIZE(key, data);
5418 if (sz >= env->me_nodemax) {
5419 /* put on overflow page */
5420 sz -= data->mv_size - sizeof(pgno_t);
5424 return sz + sizeof(indx_t);
5427 /** Calculate the size of a branch node.
5428 * The size should depend on the environment's page size but since
5429 * we currently don't support spilling large keys onto overflow
5430 * pages, it's simply the size of the #MDB_node header plus the
5431 * size of the key. Sizes are always rounded up to an even number
5432 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5433 * @param[in] env The environment handle.
5434 * @param[in] key The key for the node.
5435 * @return The number of bytes needed to store the node.
5438 mdb_branch_size(MDB_env *env, MDB_val *key)
5443 if (sz >= env->me_nodemax) {
5444 /* put on overflow page */
5445 /* not implemented */
5446 /* sz -= key->size - sizeof(pgno_t); */
5449 return sz + sizeof(indx_t);
5452 /** Add a node to the page pointed to by the cursor.
5453 * @param[in] mc The cursor for this operation.
5454 * @param[in] indx The index on the page where the new node should be added.
5455 * @param[in] key The key for the new node.
5456 * @param[in] data The data for the new node, if any.
5457 * @param[in] pgno The page number, if adding a branch node.
5458 * @param[in] flags Flags for the node.
5459 * @return 0 on success, non-zero on failure. Possible errors are:
5461 * <li>ENOMEM - failed to allocate overflow pages for the node.
5462 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5463 * should never happen since all callers already calculate the
5464 * page's free space before calling this function.
5468 mdb_node_add(MDB_cursor *mc, indx_t indx,
5469 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5472 size_t node_size = NODESIZE;
5475 MDB_page *mp = mc->mc_pg[mc->mc_top];
5476 MDB_page *ofp = NULL; /* overflow page */
5479 assert(mp->mp_upper >= mp->mp_lower);
5481 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5482 IS_LEAF(mp) ? "leaf" : "branch",
5483 IS_SUBP(mp) ? "sub-" : "",
5484 mp->mp_pgno, indx, data ? data->mv_size : 0,
5485 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5488 /* Move higher keys up one slot. */
5489 int ksize = mc->mc_db->md_pad, dif;
5490 char *ptr = LEAF2KEY(mp, indx, ksize);
5491 dif = NUMKEYS(mp) - indx;
5493 memmove(ptr+ksize, ptr, dif*ksize);
5494 /* insert new key */
5495 memcpy(ptr, key->mv_data, ksize);
5497 /* Just using these for counting */
5498 mp->mp_lower += sizeof(indx_t);
5499 mp->mp_upper -= ksize - sizeof(indx_t);
5504 node_size += key->mv_size;
5508 if (F_ISSET(flags, F_BIGDATA)) {
5509 /* Data already on overflow page. */
5510 node_size += sizeof(pgno_t);
5511 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5512 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5514 /* Put data on overflow page. */
5515 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5516 data->mv_size, node_size+data->mv_size);
5517 node_size += sizeof(pgno_t);
5518 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5520 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5523 node_size += data->mv_size;
5526 node_size += node_size & 1;
5528 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5529 DPRINTF("not enough room in page %zu, got %u ptrs",
5530 mp->mp_pgno, NUMKEYS(mp));
5531 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5532 mp->mp_upper - mp->mp_lower);
5533 DPRINTF("node size = %zu", node_size);
5534 return MDB_PAGE_FULL;
5537 /* Move higher pointers up one slot. */
5538 for (i = NUMKEYS(mp); i > indx; i--)
5539 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5541 /* Adjust free space offsets. */
5542 ofs = mp->mp_upper - node_size;
5543 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5544 mp->mp_ptrs[indx] = ofs;
5546 mp->mp_lower += sizeof(indx_t);
5548 /* Write the node data. */
5549 node = NODEPTR(mp, indx);
5550 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5551 node->mn_flags = flags;
5553 SETDSZ(node,data->mv_size);
5558 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5563 if (F_ISSET(flags, F_BIGDATA))
5564 memcpy(node->mn_data + key->mv_size, data->mv_data,
5566 else if (F_ISSET(flags, MDB_RESERVE))
5567 data->mv_data = node->mn_data + key->mv_size;
5569 memcpy(node->mn_data + key->mv_size, data->mv_data,
5572 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5574 if (F_ISSET(flags, MDB_RESERVE))
5575 data->mv_data = METADATA(ofp);
5577 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5584 /** Delete the specified node from a page.
5585 * @param[in] mp The page to operate on.
5586 * @param[in] indx The index of the node to delete.
5587 * @param[in] ksize The size of a node. Only used if the page is
5588 * part of a #MDB_DUPFIXED database.
5591 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5594 indx_t i, j, numkeys, ptr;
5601 COPY_PGNO(pgno, mp->mp_pgno);
5602 DPRINTF("delete node %u on %s page %zu", indx,
5603 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5606 assert(indx < NUMKEYS(mp));
5609 int x = NUMKEYS(mp) - 1 - indx;
5610 base = LEAF2KEY(mp, indx, ksize);
5612 memmove(base, base + ksize, x * ksize);
5613 mp->mp_lower -= sizeof(indx_t);
5614 mp->mp_upper += ksize - sizeof(indx_t);
5618 node = NODEPTR(mp, indx);
5619 sz = NODESIZE + node->mn_ksize;
5621 if (F_ISSET(node->mn_flags, F_BIGDATA))
5622 sz += sizeof(pgno_t);
5624 sz += NODEDSZ(node);
5628 ptr = mp->mp_ptrs[indx];
5629 numkeys = NUMKEYS(mp);
5630 for (i = j = 0; i < numkeys; i++) {
5632 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5633 if (mp->mp_ptrs[i] < ptr)
5634 mp->mp_ptrs[j] += sz;
5639 base = (char *)mp + mp->mp_upper;
5640 memmove(base + sz, base, ptr - mp->mp_upper);
5642 mp->mp_lower -= sizeof(indx_t);
5646 /** Compact the main page after deleting a node on a subpage.
5647 * @param[in] mp The main page to operate on.
5648 * @param[in] indx The index of the subpage on the main page.
5651 mdb_node_shrink(MDB_page *mp, indx_t indx)
5658 indx_t i, numkeys, ptr;
5660 node = NODEPTR(mp, indx);
5661 sp = (MDB_page *)NODEDATA(node);
5662 osize = NODEDSZ(node);
5664 delta = sp->mp_upper - sp->mp_lower;
5665 SETDSZ(node, osize - delta);
5666 xp = (MDB_page *)((char *)sp + delta);
5668 /* shift subpage upward */
5670 nsize = NUMKEYS(sp) * sp->mp_pad;
5671 memmove(METADATA(xp), METADATA(sp), nsize);
5674 nsize = osize - sp->mp_upper;
5675 numkeys = NUMKEYS(sp);
5676 for (i=numkeys-1; i>=0; i--)
5677 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5679 xp->mp_upper = sp->mp_lower;
5680 xp->mp_lower = sp->mp_lower;
5681 xp->mp_flags = sp->mp_flags;
5682 xp->mp_pad = sp->mp_pad;
5683 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5685 /* shift lower nodes upward */
5686 ptr = mp->mp_ptrs[indx];
5687 numkeys = NUMKEYS(mp);
5688 for (i = 0; i < numkeys; i++) {
5689 if (mp->mp_ptrs[i] <= ptr)
5690 mp->mp_ptrs[i] += delta;
5693 base = (char *)mp + mp->mp_upper;
5694 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5695 mp->mp_upper += delta;
5698 /** Initial setup of a sorted-dups cursor.
5699 * Sorted duplicates are implemented as a sub-database for the given key.
5700 * The duplicate data items are actually keys of the sub-database.
5701 * Operations on the duplicate data items are performed using a sub-cursor
5702 * initialized when the sub-database is first accessed. This function does
5703 * the preliminary setup of the sub-cursor, filling in the fields that
5704 * depend only on the parent DB.
5705 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5708 mdb_xcursor_init0(MDB_cursor *mc)
5710 MDB_xcursor *mx = mc->mc_xcursor;
5712 mx->mx_cursor.mc_xcursor = NULL;
5713 mx->mx_cursor.mc_txn = mc->mc_txn;
5714 mx->mx_cursor.mc_db = &mx->mx_db;
5715 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5716 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5717 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5718 mx->mx_cursor.mc_snum = 0;
5719 mx->mx_cursor.mc_top = 0;
5720 mx->mx_cursor.mc_flags = C_SUB;
5721 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5722 mx->mx_dbx.md_dcmp = NULL;
5723 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5726 /** Final setup of a sorted-dups cursor.
5727 * Sets up the fields that depend on the data from the main cursor.
5728 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5729 * @param[in] node The data containing the #MDB_db record for the
5730 * sorted-dup database.
5733 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5735 MDB_xcursor *mx = mc->mc_xcursor;
5737 if (node->mn_flags & F_SUBDATA) {
5738 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5739 mx->mx_cursor.mc_pg[0] = 0;
5740 mx->mx_cursor.mc_snum = 0;
5741 mx->mx_cursor.mc_flags = C_SUB;
5743 MDB_page *fp = NODEDATA(node);
5744 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5745 mx->mx_db.md_flags = 0;
5746 mx->mx_db.md_depth = 1;
5747 mx->mx_db.md_branch_pages = 0;
5748 mx->mx_db.md_leaf_pages = 1;
5749 mx->mx_db.md_overflow_pages = 0;
5750 mx->mx_db.md_entries = NUMKEYS(fp);
5751 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5752 mx->mx_cursor.mc_snum = 1;
5753 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5754 mx->mx_cursor.mc_top = 0;
5755 mx->mx_cursor.mc_pg[0] = fp;
5756 mx->mx_cursor.mc_ki[0] = 0;
5757 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5758 mx->mx_db.md_flags = MDB_DUPFIXED;
5759 mx->mx_db.md_pad = fp->mp_pad;
5760 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5761 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5764 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5766 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5768 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5769 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5770 #if UINT_MAX < SIZE_MAX
5771 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5772 #ifdef MISALIGNED_OK
5773 mx->mx_dbx.md_cmp = mdb_cmp_long;
5775 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5780 /** Initialize a cursor for a given transaction and database. */
5782 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5787 mc->mc_db = &txn->mt_dbs[dbi];
5788 mc->mc_dbx = &txn->mt_dbxs[dbi];
5789 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5794 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5796 mc->mc_xcursor = mx;
5797 mdb_xcursor_init0(mc);
5799 mc->mc_xcursor = NULL;
5801 if (*mc->mc_dbflag & DB_STALE) {
5802 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5807 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5810 MDB_xcursor *mx = NULL;
5811 size_t size = sizeof(MDB_cursor);
5813 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5816 /* Allow read access to the freelist */
5817 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5820 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5821 size += sizeof(MDB_xcursor);
5823 if ((mc = malloc(size)) != NULL) {
5824 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5825 mx = (MDB_xcursor *)(mc + 1);
5827 mdb_cursor_init(mc, txn, dbi, mx);
5828 if (txn->mt_cursors) {
5829 mc->mc_next = txn->mt_cursors[dbi];
5830 txn->mt_cursors[dbi] = mc;
5832 mc->mc_flags |= C_ALLOCD;
5843 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5845 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5848 if (txn->mt_cursors)
5851 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5855 /* Return the count of duplicate data items for the current key */
5857 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5861 if (mc == NULL || countp == NULL)
5864 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5867 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5868 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5871 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5874 *countp = mc->mc_xcursor->mx_db.md_entries;
5880 mdb_cursor_close(MDB_cursor *mc)
5883 /* remove from txn, if tracked */
5884 if (mc->mc_txn->mt_cursors) {
5885 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5886 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5888 *prev = mc->mc_next;
5890 if (mc->mc_flags & C_ALLOCD)
5896 mdb_cursor_txn(MDB_cursor *mc)
5898 if (!mc) return NULL;
5903 mdb_cursor_dbi(MDB_cursor *mc)
5909 /** Replace the key for a node with a new key.
5910 * @param[in] mp The page containing the node to operate on.
5911 * @param[in] indx The index of the node to operate on.
5912 * @param[in] key The new key to use.
5913 * @return 0 on success, non-zero on failure.
5916 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5923 indx_t ptr, i, numkeys, indx;
5926 indx = mc->mc_ki[mc->mc_top];
5927 mp = mc->mc_pg[mc->mc_top];
5928 node = NODEPTR(mp, indx);
5929 ptr = mp->mp_ptrs[indx];
5933 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5934 k2.mv_data = NODEKEY(node);
5935 k2.mv_size = node->mn_ksize;
5936 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5938 mdb_dkey(&k2, kbuf2),
5944 delta0 = delta = key->mv_size - node->mn_ksize;
5946 /* Must be 2-byte aligned. If new key is
5947 * shorter by 1, the shift will be skipped.
5949 delta += (delta & 1);
5951 if (delta > 0 && SIZELEFT(mp) < delta) {
5953 /* not enough space left, do a delete and split */
5954 DPRINTF("Not enough room, delta = %d, splitting...", delta);
5955 pgno = NODEPGNO(node);
5956 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5957 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
5960 numkeys = NUMKEYS(mp);
5961 for (i = 0; i < numkeys; i++) {
5962 if (mp->mp_ptrs[i] <= ptr)
5963 mp->mp_ptrs[i] -= delta;
5966 base = (char *)mp + mp->mp_upper;
5967 len = ptr - mp->mp_upper + NODESIZE;
5968 memmove(base - delta, base, len);
5969 mp->mp_upper -= delta;
5971 node = NODEPTR(mp, indx);
5974 /* But even if no shift was needed, update ksize */
5976 node->mn_ksize = key->mv_size;
5979 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5985 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
5987 /** Move a node from csrc to cdst.
5990 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5997 unsigned short flags;
6001 /* Mark src and dst as dirty. */
6002 if ((rc = mdb_page_touch(csrc)) ||
6003 (rc = mdb_page_touch(cdst)))
6006 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6007 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
6008 key.mv_size = csrc->mc_db->md_pad;
6009 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6011 data.mv_data = NULL;
6015 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
6016 assert(!((long)srcnode&1));
6017 srcpg = NODEPGNO(srcnode);
6018 flags = srcnode->mn_flags;
6019 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6020 unsigned int snum = csrc->mc_snum;
6022 /* must find the lowest key below src */
6023 mdb_page_search_root(csrc, NULL, 0);
6024 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6025 key.mv_size = csrc->mc_db->md_pad;
6026 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6028 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6029 key.mv_size = NODEKSZ(s2);
6030 key.mv_data = NODEKEY(s2);
6032 csrc->mc_snum = snum--;
6033 csrc->mc_top = snum;
6035 key.mv_size = NODEKSZ(srcnode);
6036 key.mv_data = NODEKEY(srcnode);
6038 data.mv_size = NODEDSZ(srcnode);
6039 data.mv_data = NODEDATA(srcnode);
6041 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6042 unsigned int snum = cdst->mc_snum;
6045 /* must find the lowest key below dst */
6046 mdb_page_search_root(cdst, NULL, 0);
6047 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6048 bkey.mv_size = cdst->mc_db->md_pad;
6049 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6051 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6052 bkey.mv_size = NODEKSZ(s2);
6053 bkey.mv_data = NODEKEY(s2);
6055 cdst->mc_snum = snum--;
6056 cdst->mc_top = snum;
6057 mdb_cursor_copy(cdst, &mn);
6059 rc = mdb_update_key(&mn, &bkey);
6064 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6065 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6066 csrc->mc_ki[csrc->mc_top],
6068 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6069 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6071 /* Add the node to the destination page.
6073 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6074 if (rc != MDB_SUCCESS)
6077 /* Delete the node from the source page.
6079 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6082 /* Adjust other cursors pointing to mp */
6083 MDB_cursor *m2, *m3;
6084 MDB_dbi dbi = csrc->mc_dbi;
6085 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6087 if (csrc->mc_flags & C_SUB)
6090 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6091 if (m2 == csrc) continue;
6092 if (csrc->mc_flags & C_SUB)
6093 m3 = &m2->mc_xcursor->mx_cursor;
6096 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6097 csrc->mc_ki[csrc->mc_top]) {
6098 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6099 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6104 /* Update the parent separators.
6106 if (csrc->mc_ki[csrc->mc_top] == 0) {
6107 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6108 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6109 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6111 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6112 key.mv_size = NODEKSZ(srcnode);
6113 key.mv_data = NODEKEY(srcnode);
6115 DPRINTF("update separator for source page %zu to [%s]",
6116 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6117 mdb_cursor_copy(csrc, &mn);
6120 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6123 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6125 indx_t ix = csrc->mc_ki[csrc->mc_top];
6126 nullkey.mv_size = 0;
6127 csrc->mc_ki[csrc->mc_top] = 0;
6128 rc = mdb_update_key(csrc, &nullkey);
6129 csrc->mc_ki[csrc->mc_top] = ix;
6130 assert(rc == MDB_SUCCESS);
6134 if (cdst->mc_ki[cdst->mc_top] == 0) {
6135 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6136 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6137 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6139 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6140 key.mv_size = NODEKSZ(srcnode);
6141 key.mv_data = NODEKEY(srcnode);
6143 DPRINTF("update separator for destination page %zu to [%s]",
6144 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6145 mdb_cursor_copy(cdst, &mn);
6148 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6151 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6153 indx_t ix = cdst->mc_ki[cdst->mc_top];
6154 nullkey.mv_size = 0;
6155 cdst->mc_ki[cdst->mc_top] = 0;
6156 rc = mdb_update_key(cdst, &nullkey);
6157 cdst->mc_ki[cdst->mc_top] = ix;
6158 assert(rc == MDB_SUCCESS);
6165 /** Merge one page into another.
6166 * The nodes from the page pointed to by \b csrc will
6167 * be copied to the page pointed to by \b cdst and then
6168 * the \b csrc page will be freed.
6169 * @param[in] csrc Cursor pointing to the source page.
6170 * @param[in] cdst Cursor pointing to the destination page.
6173 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6181 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6182 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6184 assert(csrc->mc_snum > 1); /* can't merge root page */
6185 assert(cdst->mc_snum > 1);
6187 /* Mark dst as dirty. */
6188 if ((rc = mdb_page_touch(cdst)))
6191 /* Move all nodes from src to dst.
6193 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6194 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6195 key.mv_size = csrc->mc_db->md_pad;
6196 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6197 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6198 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6199 if (rc != MDB_SUCCESS)
6201 key.mv_data = (char *)key.mv_data + key.mv_size;
6204 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6205 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6206 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6207 unsigned int snum = csrc->mc_snum;
6209 /* must find the lowest key below src */
6210 mdb_page_search_root(csrc, NULL, 0);
6211 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6212 key.mv_size = csrc->mc_db->md_pad;
6213 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6215 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6216 key.mv_size = NODEKSZ(s2);
6217 key.mv_data = NODEKEY(s2);
6219 csrc->mc_snum = snum--;
6220 csrc->mc_top = snum;
6222 key.mv_size = srcnode->mn_ksize;
6223 key.mv_data = NODEKEY(srcnode);
6226 data.mv_size = NODEDSZ(srcnode);
6227 data.mv_data = NODEDATA(srcnode);
6228 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6229 if (rc != MDB_SUCCESS)
6234 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6235 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);
6237 /* Unlink the src page from parent and add to free list.
6239 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6240 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6243 rc = mdb_update_key(csrc, &key);
6249 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6250 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6251 csrc->mc_db->md_leaf_pages--;
6253 csrc->mc_db->md_branch_pages--;
6255 /* Adjust other cursors pointing to mp */
6256 MDB_cursor *m2, *m3;
6257 MDB_dbi dbi = csrc->mc_dbi;
6258 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6260 if (csrc->mc_flags & C_SUB)
6263 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6264 if (csrc->mc_flags & C_SUB)
6265 m3 = &m2->mc_xcursor->mx_cursor;
6268 if (m3 == csrc) continue;
6269 if (m3->mc_snum < csrc->mc_snum) continue;
6270 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6271 m3->mc_pg[csrc->mc_top] = mp;
6272 m3->mc_ki[csrc->mc_top] += nkeys;
6276 mdb_cursor_pop(csrc);
6278 return mdb_rebalance(csrc);
6281 /** Copy the contents of a cursor.
6282 * @param[in] csrc The cursor to copy from.
6283 * @param[out] cdst The cursor to copy to.
6286 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6290 cdst->mc_txn = csrc->mc_txn;
6291 cdst->mc_dbi = csrc->mc_dbi;
6292 cdst->mc_db = csrc->mc_db;
6293 cdst->mc_dbx = csrc->mc_dbx;
6294 cdst->mc_snum = csrc->mc_snum;
6295 cdst->mc_top = csrc->mc_top;
6296 cdst->mc_flags = csrc->mc_flags;
6298 for (i=0; i<csrc->mc_snum; i++) {
6299 cdst->mc_pg[i] = csrc->mc_pg[i];
6300 cdst->mc_ki[i] = csrc->mc_ki[i];
6304 /** Rebalance the tree after a delete operation.
6305 * @param[in] mc Cursor pointing to the page where rebalancing
6307 * @return 0 on success, non-zero on failure.
6310 mdb_rebalance(MDB_cursor *mc)
6314 unsigned int ptop, minkeys;
6317 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6321 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6322 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6323 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6324 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6328 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6329 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6332 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6333 DPRINTF("no need to rebalance page %zu, above fill threshold",
6339 if (mc->mc_snum < 2) {
6340 MDB_page *mp = mc->mc_pg[0];
6342 DPUTS("Can't rebalance a subpage, ignoring");
6345 if (NUMKEYS(mp) == 0) {
6346 DPUTS("tree is completely empty");
6347 mc->mc_db->md_root = P_INVALID;
6348 mc->mc_db->md_depth = 0;
6349 mc->mc_db->md_leaf_pages = 0;
6350 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6354 /* Adjust other cursors pointing to mp */
6355 MDB_cursor *m2, *m3;
6356 MDB_dbi dbi = mc->mc_dbi;
6358 if (mc->mc_flags & C_SUB)
6361 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6362 if (m2 == mc) continue;
6363 if (mc->mc_flags & C_SUB)
6364 m3 = &m2->mc_xcursor->mx_cursor;
6367 if (m3->mc_snum < mc->mc_snum) continue;
6368 if (m3->mc_pg[0] == mp) {
6374 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6375 DPUTS("collapsing root page!");
6376 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6377 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6378 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6381 mc->mc_db->md_depth--;
6382 mc->mc_db->md_branch_pages--;
6384 /* Adjust other cursors pointing to mp */
6385 MDB_cursor *m2, *m3;
6386 MDB_dbi dbi = mc->mc_dbi;
6388 if (mc->mc_flags & C_SUB)
6391 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6392 if (m2 == mc) continue;
6393 if (mc->mc_flags & C_SUB)
6394 m3 = &m2->mc_xcursor->mx_cursor;
6397 if (m3->mc_snum < mc->mc_snum) continue;
6398 if (m3->mc_pg[0] == mp) {
6399 m3->mc_pg[0] = mc->mc_pg[0];
6406 DPUTS("root page doesn't need rebalancing");
6410 /* The parent (branch page) must have at least 2 pointers,
6411 * otherwise the tree is invalid.
6413 ptop = mc->mc_top-1;
6414 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6416 /* Leaf page fill factor is below the threshold.
6417 * Try to move keys from left or right neighbor, or
6418 * merge with a neighbor page.
6423 mdb_cursor_copy(mc, &mn);
6424 mn.mc_xcursor = NULL;
6426 if (mc->mc_ki[ptop] == 0) {
6427 /* We're the leftmost leaf in our parent.
6429 DPUTS("reading right neighbor");
6431 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6432 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6434 mn.mc_ki[mn.mc_top] = 0;
6435 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6437 /* There is at least one neighbor to the left.
6439 DPUTS("reading left neighbor");
6441 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6442 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6444 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6445 mc->mc_ki[mc->mc_top] = 0;
6448 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6449 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);
6451 /* If the neighbor page is above threshold and has enough keys,
6452 * move one key from it. Otherwise we should try to merge them.
6453 * (A branch page must never have less than 2 keys.)
6455 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6456 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6457 return mdb_node_move(&mn, mc);
6459 if (mc->mc_ki[ptop] == 0)
6460 rc = mdb_page_merge(&mn, mc);
6462 rc = mdb_page_merge(mc, &mn);
6463 mc->mc_flags &= ~C_INITIALIZED;
6468 /** Complete a delete operation started by #mdb_cursor_del(). */
6470 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6474 /* add overflow pages to free list */
6475 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6479 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6480 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6481 mc->mc_db->md_overflow_pages -= ovpages;
6482 for (i=0; i<ovpages; i++) {
6483 DPRINTF("freed ov page %zu", pg);
6484 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6488 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6489 mc->mc_db->md_entries--;
6490 rc = mdb_rebalance(mc);
6491 if (rc != MDB_SUCCESS)
6492 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6493 /* if mc points past last node in page, invalidate */
6494 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6495 mc->mc_flags &= ~C_INITIALIZED;
6501 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6502 MDB_val *key, MDB_val *data)
6507 MDB_val rdata, *xdata;
6511 assert(key != NULL);
6513 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6515 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6518 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6522 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6526 mdb_cursor_init(&mc, txn, dbi, &mx);
6537 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6539 /* let mdb_page_split know about this cursor if needed:
6540 * delete will trigger a rebalance; if it needs to move
6541 * a node from one page to another, it will have to
6542 * update the parent's separator key(s). If the new sepkey
6543 * is larger than the current one, the parent page may
6544 * run out of space, triggering a split. We need this
6545 * cursor to be consistent until the end of the rebalance.
6547 mc.mc_next = txn->mt_cursors[dbi];
6548 txn->mt_cursors[dbi] = &mc;
6549 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6550 txn->mt_cursors[dbi] = mc.mc_next;
6555 /** Split a page and insert a new node.
6556 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6557 * The cursor will be updated to point to the actual page and index where
6558 * the node got inserted after the split.
6559 * @param[in] newkey The key for the newly inserted node.
6560 * @param[in] newdata The data for the newly inserted node.
6561 * @param[in] newpgno The page number, if the new node is a branch node.
6562 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6563 * @return 0 on success, non-zero on failure.
6566 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6567 unsigned int nflags)
6570 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6573 unsigned int i, j, split_indx, nkeys, pmax;
6575 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6577 MDB_page *mp, *rp, *pp;
6582 mp = mc->mc_pg[mc->mc_top];
6583 newindx = mc->mc_ki[mc->mc_top];
6585 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6586 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6587 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6589 /* Create a right sibling. */
6590 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6592 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6594 if (mc->mc_snum < 2) {
6595 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6597 /* shift current top to make room for new parent */
6598 mc->mc_pg[1] = mc->mc_pg[0];
6599 mc->mc_ki[1] = mc->mc_ki[0];
6602 mc->mc_db->md_root = pp->mp_pgno;
6603 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6604 mc->mc_db->md_depth++;
6607 /* Add left (implicit) pointer. */
6608 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6609 /* undo the pre-push */
6610 mc->mc_pg[0] = mc->mc_pg[1];
6611 mc->mc_ki[0] = mc->mc_ki[1];
6612 mc->mc_db->md_root = mp->mp_pgno;
6613 mc->mc_db->md_depth--;
6620 ptop = mc->mc_top-1;
6621 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6624 mc->mc_flags |= C_SPLITTING;
6625 mdb_cursor_copy(mc, &mn);
6626 mn.mc_pg[mn.mc_top] = rp;
6627 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6629 if (nflags & MDB_APPEND) {
6630 mn.mc_ki[mn.mc_top] = 0;
6632 split_indx = newindx;
6637 nkeys = NUMKEYS(mp);
6638 split_indx = nkeys / 2;
6639 if (newindx < split_indx)
6645 unsigned int lsize, rsize, ksize;
6646 /* Move half of the keys to the right sibling */
6648 x = mc->mc_ki[mc->mc_top] - split_indx;
6649 ksize = mc->mc_db->md_pad;
6650 split = LEAF2KEY(mp, split_indx, ksize);
6651 rsize = (nkeys - split_indx) * ksize;
6652 lsize = (nkeys - split_indx) * sizeof(indx_t);
6653 mp->mp_lower -= lsize;
6654 rp->mp_lower += lsize;
6655 mp->mp_upper += rsize - lsize;
6656 rp->mp_upper -= rsize - lsize;
6657 sepkey.mv_size = ksize;
6658 if (newindx == split_indx) {
6659 sepkey.mv_data = newkey->mv_data;
6661 sepkey.mv_data = split;
6664 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6665 memcpy(rp->mp_ptrs, split, rsize);
6666 sepkey.mv_data = rp->mp_ptrs;
6667 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6668 memcpy(ins, newkey->mv_data, ksize);
6669 mp->mp_lower += sizeof(indx_t);
6670 mp->mp_upper -= ksize - sizeof(indx_t);
6673 memcpy(rp->mp_ptrs, split, x * ksize);
6674 ins = LEAF2KEY(rp, x, ksize);
6675 memcpy(ins, newkey->mv_data, ksize);
6676 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6677 rp->mp_lower += sizeof(indx_t);
6678 rp->mp_upper -= ksize - sizeof(indx_t);
6679 mc->mc_ki[mc->mc_top] = x;
6680 mc->mc_pg[mc->mc_top] = rp;
6685 /* For leaf pages, check the split point based on what
6686 * fits where, since otherwise mdb_node_add can fail.
6688 * This check is only needed when the data items are
6689 * relatively large, such that being off by one will
6690 * make the difference between success or failure.
6692 * It's also relevant if a page happens to be laid out
6693 * such that one half of its nodes are all "small" and
6694 * the other half of its nodes are "large." If the new
6695 * item is also "large" and falls on the half with
6696 * "large" nodes, it also may not fit.
6699 unsigned int psize, nsize;
6700 /* Maximum free space in an empty page */
6701 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6702 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6703 if ((nkeys < 20) || (nsize > pmax/16)) {
6704 if (newindx <= split_indx) {
6707 for (i=0; i<split_indx; i++) {
6708 node = NODEPTR(mp, i);
6709 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6710 if (F_ISSET(node->mn_flags, F_BIGDATA))
6711 psize += sizeof(pgno_t);
6713 psize += NODEDSZ(node);
6717 split_indx = newindx;
6728 for (i=nkeys-1; i>=split_indx; i--) {
6729 node = NODEPTR(mp, i);
6730 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6731 if (F_ISSET(node->mn_flags, F_BIGDATA))
6732 psize += sizeof(pgno_t);
6734 psize += NODEDSZ(node);
6738 split_indx = newindx;
6749 /* First find the separating key between the split pages.
6750 * The case where newindx == split_indx is ambiguous; the
6751 * new item could go to the new page or stay on the original
6752 * page. If newpos == 1 it goes to the new page.
6754 if (newindx == split_indx && newpos) {
6755 sepkey.mv_size = newkey->mv_size;
6756 sepkey.mv_data = newkey->mv_data;
6758 node = NODEPTR(mp, split_indx);
6759 sepkey.mv_size = node->mn_ksize;
6760 sepkey.mv_data = NODEKEY(node);
6764 DPRINTF("separator is [%s]", DKEY(&sepkey));
6766 /* Copy separator key to the parent.
6768 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6772 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6775 if (mn.mc_snum == mc->mc_snum) {
6776 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6777 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6778 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6779 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6784 /* Right page might now have changed parent.
6785 * Check if left page also changed parent.
6787 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6788 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6789 for (i=0; i<ptop; i++) {
6790 mc->mc_pg[i] = mn.mc_pg[i];
6791 mc->mc_ki[i] = mn.mc_ki[i];
6793 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6794 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6798 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6801 mc->mc_flags ^= C_SPLITTING;
6802 if (rc != MDB_SUCCESS) {
6805 if (nflags & MDB_APPEND) {
6806 mc->mc_pg[mc->mc_top] = rp;
6807 mc->mc_ki[mc->mc_top] = 0;
6808 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6811 for (i=0; i<mc->mc_top; i++)
6812 mc->mc_ki[i] = mn.mc_ki[i];
6819 /* Move half of the keys to the right sibling. */
6821 /* grab a page to hold a temporary copy */
6822 copy = mdb_page_malloc(mc);
6826 copy->mp_pgno = mp->mp_pgno;
6827 copy->mp_flags = mp->mp_flags;
6828 copy->mp_lower = PAGEHDRSZ;
6829 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6830 mc->mc_pg[mc->mc_top] = copy;
6831 for (i = j = 0; i <= nkeys; j++) {
6832 if (i == split_indx) {
6833 /* Insert in right sibling. */
6834 /* Reset insert index for right sibling. */
6835 if (i != newindx || (newpos ^ ins_new)) {
6837 mc->mc_pg[mc->mc_top] = rp;
6841 if (i == newindx && !ins_new) {
6842 /* Insert the original entry that caused the split. */
6843 rkey.mv_data = newkey->mv_data;
6844 rkey.mv_size = newkey->mv_size;
6853 /* Update index for the new key. */
6854 mc->mc_ki[mc->mc_top] = j;
6855 } else if (i == nkeys) {
6858 node = NODEPTR(mp, i);
6859 rkey.mv_data = NODEKEY(node);
6860 rkey.mv_size = node->mn_ksize;
6862 xdata.mv_data = NODEDATA(node);
6863 xdata.mv_size = NODEDSZ(node);
6866 pgno = NODEPGNO(node);
6867 flags = node->mn_flags;
6872 if (!IS_LEAF(mp) && j == 0) {
6873 /* First branch index doesn't need key data. */
6877 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6881 nkeys = NUMKEYS(copy);
6882 for (i=0; i<nkeys; i++)
6883 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6884 mp->mp_lower = copy->mp_lower;
6885 mp->mp_upper = copy->mp_upper;
6886 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6887 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6889 /* reset back to original page */
6890 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6891 mc->mc_pg[mc->mc_top] = mp;
6892 if (nflags & MDB_RESERVE) {
6893 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6894 if (!(node->mn_flags & F_BIGDATA))
6895 newdata->mv_data = NODEDATA(node);
6901 /* return tmp page to freelist */
6902 mdb_page_free(mc->mc_txn->mt_env, copy);
6905 /* Adjust other cursors pointing to mp */
6906 MDB_cursor *m2, *m3;
6907 MDB_dbi dbi = mc->mc_dbi;
6908 int fixup = NUMKEYS(mp);
6910 if (mc->mc_flags & C_SUB)
6913 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6914 if (m2 == mc) continue;
6915 if (mc->mc_flags & C_SUB)
6916 m3 = &m2->mc_xcursor->mx_cursor;
6919 if (!(m3->mc_flags & C_INITIALIZED))
6921 if (m3->mc_flags & C_SPLITTING)
6926 for (k=m3->mc_top; k>=0; k--) {
6927 m3->mc_ki[k+1] = m3->mc_ki[k];
6928 m3->mc_pg[k+1] = m3->mc_pg[k];
6930 if (m3->mc_ki[0] >= split_indx) {
6935 m3->mc_pg[0] = mc->mc_pg[0];
6939 if (m3->mc_pg[mc->mc_top] == mp) {
6940 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6941 m3->mc_ki[mc->mc_top]++;
6942 if (m3->mc_ki[mc->mc_top] >= fixup) {
6943 m3->mc_pg[mc->mc_top] = rp;
6944 m3->mc_ki[mc->mc_top] -= fixup;
6945 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6947 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6948 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6957 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6958 MDB_val *key, MDB_val *data, unsigned int flags)
6963 assert(key != NULL);
6964 assert(data != NULL);
6966 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6969 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6973 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6977 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
6980 mdb_cursor_init(&mc, txn, dbi, &mx);
6981 return mdb_cursor_put(&mc, key, data, flags);
6985 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6987 if ((flag & CHANGEABLE) != flag)
6990 env->me_flags |= flag;
6992 env->me_flags &= ~flag;
6997 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
7002 *arg = env->me_flags;
7007 mdb_env_get_path(MDB_env *env, const char **arg)
7012 *arg = env->me_path;
7016 /** Common code for #mdb_stat() and #mdb_env_stat().
7017 * @param[in] env the environment to operate in.
7018 * @param[in] db the #MDB_db record containing the stats to return.
7019 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
7020 * @return 0, this function always succeeds.
7023 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7025 arg->ms_psize = env->me_psize;
7026 arg->ms_depth = db->md_depth;
7027 arg->ms_branch_pages = db->md_branch_pages;
7028 arg->ms_leaf_pages = db->md_leaf_pages;
7029 arg->ms_overflow_pages = db->md_overflow_pages;
7030 arg->ms_entries = db->md_entries;
7035 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7039 if (env == NULL || arg == NULL)
7042 toggle = mdb_env_pick_meta(env);
7044 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7048 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7052 if (env == NULL || arg == NULL)
7055 toggle = mdb_env_pick_meta(env);
7056 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7057 arg->me_mapsize = env->me_mapsize;
7058 arg->me_maxreaders = env->me_maxreaders;
7059 arg->me_numreaders = env->me_numreaders;
7060 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7061 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7065 /** Set the default comparison functions for a database.
7066 * Called immediately after a database is opened to set the defaults.
7067 * The user can then override them with #mdb_set_compare() or
7068 * #mdb_set_dupsort().
7069 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7070 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7073 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7075 uint16_t f = txn->mt_dbs[dbi].md_flags;
7077 txn->mt_dbxs[dbi].md_cmp =
7078 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7079 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7081 txn->mt_dbxs[dbi].md_dcmp =
7082 !(f & MDB_DUPSORT) ? 0 :
7083 ((f & MDB_INTEGERDUP)
7084 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7085 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7088 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7093 int rc, dbflag, exact;
7094 unsigned int unused = 0;
7097 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7098 mdb_default_cmp(txn, FREE_DBI);
7101 if ((flags & VALID_FLAGS) != flags)
7107 if (flags & PERSISTENT_FLAGS) {
7108 uint16_t f2 = flags & PERSISTENT_FLAGS;
7109 /* make sure flag changes get committed */
7110 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7111 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7112 txn->mt_flags |= MDB_TXN_DIRTY;
7115 mdb_default_cmp(txn, MAIN_DBI);
7119 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7120 mdb_default_cmp(txn, MAIN_DBI);
7123 /* Is the DB already open? */
7125 for (i=2; i<txn->mt_numdbs; i++) {
7126 if (!txn->mt_dbxs[i].md_name.mv_size) {
7127 /* Remember this free slot */
7128 if (!unused) unused = i;
7131 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7132 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7138 /* If no free slot and max hit, fail */
7139 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7140 return MDB_DBS_FULL;
7142 /* Find the DB info */
7143 dbflag = DB_NEW|DB_VALID;
7146 key.mv_data = (void *)name;
7147 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7148 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7149 if (rc == MDB_SUCCESS) {
7150 /* make sure this is actually a DB */
7151 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7152 if (!(node->mn_flags & F_SUBDATA))
7154 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7155 /* Create if requested */
7157 data.mv_size = sizeof(MDB_db);
7158 data.mv_data = &dummy;
7159 memset(&dummy, 0, sizeof(dummy));
7160 dummy.md_root = P_INVALID;
7161 dummy.md_flags = flags & PERSISTENT_FLAGS;
7162 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7166 /* OK, got info, add to table */
7167 if (rc == MDB_SUCCESS) {
7168 unsigned int slot = unused ? unused : txn->mt_numdbs;
7169 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7170 txn->mt_dbxs[slot].md_name.mv_size = len;
7171 txn->mt_dbxs[slot].md_rel = NULL;
7172 txn->mt_dbflags[slot] = dbflag;
7173 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7175 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7176 mdb_default_cmp(txn, slot);
7185 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7187 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7190 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7193 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7196 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7198 ptr = env->me_dbxs[dbi].md_name.mv_data;
7199 env->me_dbxs[dbi].md_name.mv_data = NULL;
7200 env->me_dbxs[dbi].md_name.mv_size = 0;
7204 /** Add all the DB's pages to the free list.
7205 * @param[in] mc Cursor on the DB to free.
7206 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7207 * @return 0 on success, non-zero on failure.
7210 mdb_drop0(MDB_cursor *mc, int subs)
7214 rc = mdb_page_search(mc, NULL, 0);
7215 if (rc == MDB_SUCCESS) {
7220 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7221 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7224 mdb_cursor_copy(mc, &mx);
7225 while (mc->mc_snum > 0) {
7226 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7227 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7228 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7229 if (ni->mn_flags & F_BIGDATA) {
7230 int j, ovpages = OVPAGES(NODEDSZ(ni), mc->mc_txn->mt_env->me_psize);
7232 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7233 for (j=0; j<ovpages; j++) {
7234 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7237 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7238 mdb_xcursor_init1(mc, ni);
7239 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7245 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7247 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7250 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7255 mc->mc_ki[mc->mc_top] = i;
7256 rc = mdb_cursor_sibling(mc, 1);
7258 /* no more siblings, go back to beginning
7259 * of previous level.
7263 for (i=1; i<mc->mc_snum; i++) {
7265 mc->mc_pg[i] = mx.mc_pg[i];
7270 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7271 mc->mc_db->md_root);
7276 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7281 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7284 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7287 rc = mdb_cursor_open(txn, dbi, &mc);
7291 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7295 /* Can't delete the main DB */
7296 if (del && dbi > MAIN_DBI) {
7297 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7299 txn->mt_dbflags[dbi] = DB_STALE;
7300 mdb_dbi_close(txn->mt_env, dbi);
7303 /* reset the DB record, mark it dirty */
7304 txn->mt_dbflags[dbi] |= DB_DIRTY;
7305 txn->mt_dbs[dbi].md_depth = 0;
7306 txn->mt_dbs[dbi].md_branch_pages = 0;
7307 txn->mt_dbs[dbi].md_leaf_pages = 0;
7308 txn->mt_dbs[dbi].md_overflow_pages = 0;
7309 txn->mt_dbs[dbi].md_entries = 0;
7310 txn->mt_dbs[dbi].md_root = P_INVALID;
7312 txn->mt_flags |= MDB_TXN_DIRTY;
7315 mdb_cursor_close(mc);
7319 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7321 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7324 txn->mt_dbxs[dbi].md_cmp = cmp;
7328 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7330 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7333 txn->mt_dbxs[dbi].md_dcmp = cmp;
7337 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7339 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7342 txn->mt_dbxs[dbi].md_rel = rel;
7346 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7348 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7351 txn->mt_dbxs[dbi].md_relctx = ctx;