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 * No reader table is used if the database is on a read-only filesystem.
411 * Since the database uses multi-version concurrency control, readers don't
412 * actually need any locking. This table is used to keep track of which
413 * readers are using data from which old transactions, so that we'll know
414 * when a particular old transaction is no longer in use. Old transactions
415 * that have discarded any data pages can then have those pages reclaimed
416 * for use by a later write transaction.
418 * The lock table is constructed such that reader slots are aligned with the
419 * processor's cache line size. Any slot is only ever used by one thread.
420 * This alignment guarantees that there will be no contention or cache
421 * thrashing as threads update their own slot info, and also eliminates
422 * any need for locking when accessing a slot.
424 * A writer thread will scan every slot in the table to determine the oldest
425 * outstanding reader transaction. Any freed pages older than this will be
426 * reclaimed by the writer. The writer doesn't use any locks when scanning
427 * this table. This means that there's no guarantee that the writer will
428 * see the most up-to-date reader info, but that's not required for correct
429 * operation - all we need is to know the upper bound on the oldest reader,
430 * we don't care at all about the newest reader. So the only consequence of
431 * reading stale information here is that old pages might hang around a
432 * while longer before being reclaimed. That's actually good anyway, because
433 * the longer we delay reclaiming old pages, the more likely it is that a
434 * string of contiguous pages can be found after coalescing old pages from
435 * many old transactions together.
438 /** Number of slots in the reader table.
439 * This value was chosen somewhat arbitrarily. 126 readers plus a
440 * couple mutexes fit exactly into 8KB on my development machine.
441 * Applications should set the table size using #mdb_env_set_maxreaders().
443 #define DEFAULT_READERS 126
445 /** The size of a CPU cache line in bytes. We want our lock structures
446 * aligned to this size to avoid false cache line sharing in the
448 * This value works for most CPUs. For Itanium this should be 128.
454 /** The information we store in a single slot of the reader table.
455 * In addition to a transaction ID, we also record the process and
456 * thread ID that owns a slot, so that we can detect stale information,
457 * e.g. threads or processes that went away without cleaning up.
458 * @note We currently don't check for stale records. We simply re-init
459 * the table when we know that we're the only process opening the
462 typedef struct MDB_rxbody {
463 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
464 * Multiple readers that start at the same time will probably have the
465 * same ID here. Again, it's not important to exclude them from
466 * anything; all we need to know is which version of the DB they
467 * started from so we can avoid overwriting any data used in that
468 * particular version.
471 /** The process ID of the process owning this reader txn. */
473 /** The thread ID of the thread owning this txn. */
477 /** The actual reader record, with cacheline padding. */
478 typedef struct MDB_reader {
481 /** shorthand for mrb_txnid */
482 #define mr_txnid mru.mrx.mrb_txnid
483 #define mr_pid mru.mrx.mrb_pid
484 #define mr_tid mru.mrx.mrb_tid
485 /** cache line alignment */
486 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
490 /** The header for the reader table.
491 * The table resides in a memory-mapped file. (This is a different file
492 * than is used for the main database.)
494 * For POSIX the actual mutexes reside in the shared memory of this
495 * mapped file. On Windows, mutexes are named objects allocated by the
496 * kernel; we store the mutex names in this mapped file so that other
497 * processes can grab them. This same approach is also used on
498 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
499 * process-shared POSIX mutexes. For these cases where a named object
500 * is used, the object name is derived from a 64 bit FNV hash of the
501 * environment pathname. As such, naming collisions are extremely
502 * unlikely. If a collision occurs, the results are unpredictable.
504 typedef struct MDB_txbody {
505 /** Stamp identifying this as an MDB file. It must be set
508 /** Version number of this lock file. Must be set to #MDB_VERSION. */
509 uint32_t mtb_version;
510 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
511 char mtb_rmname[MNAME_LEN];
513 /** Mutex protecting access to this table.
514 * This is the reader lock that #LOCK_MUTEX_R acquires.
516 pthread_mutex_t mtb_mutex;
518 /** The ID of the last transaction committed to the database.
519 * This is recorded here only for convenience; the value can always
520 * be determined by reading the main database meta pages.
523 /** The number of slots that have been used in the reader table.
524 * This always records the maximum count, it is not decremented
525 * when readers release their slots.
527 unsigned mtb_numreaders;
530 /** The actual reader table definition. */
531 typedef struct MDB_txninfo {
534 #define mti_magic mt1.mtb.mtb_magic
535 #define mti_version mt1.mtb.mtb_version
536 #define mti_mutex mt1.mtb.mtb_mutex
537 #define mti_rmname mt1.mtb.mtb_rmname
538 #define mti_txnid mt1.mtb.mtb_txnid
539 #define mti_numreaders mt1.mtb.mtb_numreaders
540 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
543 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
544 char mt2_wmname[MNAME_LEN];
545 #define mti_wmname mt2.mt2_wmname
547 pthread_mutex_t mt2_wmutex;
548 #define mti_wmutex mt2.mt2_wmutex
550 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
552 MDB_reader mti_readers[1];
556 /** Common header for all page types.
557 * Overflow records occupy a number of contiguous pages with no
558 * headers on any page after the first.
560 typedef struct MDB_page {
561 #define mp_pgno mp_p.p_pgno
562 #define mp_next mp_p.p_next
564 pgno_t p_pgno; /**< page number */
565 void * p_next; /**< for in-memory list of freed structs */
568 /** @defgroup mdb_page Page Flags
570 * Flags for the page headers.
573 #define P_BRANCH 0x01 /**< branch page */
574 #define P_LEAF 0x02 /**< leaf page */
575 #define P_OVERFLOW 0x04 /**< overflow page */
576 #define P_META 0x08 /**< meta page */
577 #define P_DIRTY 0x10 /**< dirty page */
578 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
579 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
581 uint16_t mp_flags; /**< @ref mdb_page */
582 #define mp_lower mp_pb.pb.pb_lower
583 #define mp_upper mp_pb.pb.pb_upper
584 #define mp_pages mp_pb.pb_pages
587 indx_t pb_lower; /**< lower bound of free space */
588 indx_t pb_upper; /**< upper bound of free space */
590 uint32_t pb_pages; /**< number of overflow pages */
592 indx_t mp_ptrs[1]; /**< dynamic size */
595 /** Size of the page header, excluding dynamic data at the end */
596 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
598 /** Address of first usable data byte in a page, after the header */
599 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
601 /** Number of nodes on a page */
602 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
604 /** The amount of space remaining in the page */
605 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
607 /** The percentage of space used in the page, in tenths of a percent. */
608 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
609 ((env)->me_psize - PAGEHDRSZ))
610 /** The minimum page fill factor, in tenths of a percent.
611 * Pages emptier than this are candidates for merging.
613 #define FILL_THRESHOLD 250
615 /** Test if a page is a leaf page */
616 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
617 /** Test if a page is a LEAF2 page */
618 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
619 /** Test if a page is a branch page */
620 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
621 /** Test if a page is an overflow page */
622 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
623 /** Test if a page is a sub page */
624 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
626 /** The number of overflow pages needed to store the given size. */
627 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
629 /** Header for a single key/data pair within a page.
630 * We guarantee 2-byte alignment for nodes.
632 typedef struct MDB_node {
633 /** lo and hi are used for data size on leaf nodes and for
634 * child pgno on branch nodes. On 64 bit platforms, flags
635 * is also used for pgno. (Branch nodes have no flags).
636 * They are in host byte order in case that lets some
637 * accesses be optimized into a 32-bit word access.
639 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
640 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
641 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
642 /** @defgroup mdb_node Node Flags
644 * Flags for node headers.
647 #define F_BIGDATA 0x01 /**< data put on overflow page */
648 #define F_SUBDATA 0x02 /**< data is a sub-database */
649 #define F_DUPDATA 0x04 /**< data has duplicates */
651 /** valid flags for #mdb_node_add() */
652 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
655 unsigned short mn_flags; /**< @ref mdb_node */
656 unsigned short mn_ksize; /**< key size */
657 char mn_data[1]; /**< key and data are appended here */
660 /** Size of the node header, excluding dynamic data at the end */
661 #define NODESIZE offsetof(MDB_node, mn_data)
663 /** Bit position of top word in page number, for shifting mn_flags */
664 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
666 /** Size of a node in a branch page with a given key.
667 * This is just the node header plus the key, there is no data.
669 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
671 /** Size of a node in a leaf page with a given key and data.
672 * This is node header plus key plus data size.
674 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
676 /** Address of node \b i in page \b p */
677 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
679 /** Address of the key for the node */
680 #define NODEKEY(node) (void *)((node)->mn_data)
682 /** Address of the data for a node */
683 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
685 /** Get the page number pointed to by a branch node */
686 #define NODEPGNO(node) \
687 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
688 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
689 /** Set the page number in a branch node */
690 #define SETPGNO(node,pgno) do { \
691 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
692 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
694 /** Get the size of the data in a leaf node */
695 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
696 /** Set the size of the data for a leaf node */
697 #define SETDSZ(node,size) do { \
698 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
699 /** The size of a key in a node */
700 #define NODEKSZ(node) ((node)->mn_ksize)
702 /** Copy a page number from src to dst */
704 #define COPY_PGNO(dst,src) dst = src
706 #if SIZE_MAX > 4294967295UL
707 #define COPY_PGNO(dst,src) do { \
708 unsigned short *s, *d; \
709 s = (unsigned short *)&(src); \
710 d = (unsigned short *)&(dst); \
717 #define COPY_PGNO(dst,src) do { \
718 unsigned short *s, *d; \
719 s = (unsigned short *)&(src); \
720 d = (unsigned short *)&(dst); \
726 /** The address of a key in a LEAF2 page.
727 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
728 * There are no node headers, keys are stored contiguously.
730 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
732 /** Set the \b node's key into \b key, if requested. */
733 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
734 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
736 /** Information about a single database in the environment. */
737 typedef struct MDB_db {
738 uint32_t md_pad; /**< also ksize for LEAF2 pages */
739 uint16_t md_flags; /**< @ref mdb_dbi_open */
740 uint16_t md_depth; /**< depth of this tree */
741 pgno_t md_branch_pages; /**< number of internal pages */
742 pgno_t md_leaf_pages; /**< number of leaf pages */
743 pgno_t md_overflow_pages; /**< number of overflow pages */
744 size_t md_entries; /**< number of data items */
745 pgno_t md_root; /**< the root page of this tree */
748 /** mdb_dbi_open flags */
749 #define PERSISTENT_FLAGS 0x7fff
750 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
751 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
753 /** Handle for the DB used to track free pages. */
755 /** Handle for the default DB. */
758 /** Meta page content. */
759 typedef struct MDB_meta {
760 /** Stamp identifying this as an MDB file. It must be set
763 /** Version number of this lock file. Must be set to #MDB_VERSION. */
765 void *mm_address; /**< address for fixed mapping */
766 size_t mm_mapsize; /**< size of mmap region */
767 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
768 /** The size of pages used in this DB */
769 #define mm_psize mm_dbs[0].md_pad
770 /** Any persistent environment flags. @ref mdb_env */
771 #define mm_flags mm_dbs[0].md_flags
772 pgno_t mm_last_pg; /**< last used page in file */
773 txnid_t mm_txnid; /**< txnid that committed this page */
776 /** Buffer for a stack-allocated dirty page.
777 * The members define size and alignment, and silence type
778 * aliasing warnings. They are not used directly; that could
779 * mean incorrectly using several union members in parallel.
781 typedef union MDB_pagebuf {
782 char mb_raw[MDB_PAGESIZE];
785 char mm_pad[PAGEHDRSZ];
790 /** Auxiliary DB info.
791 * The information here is mostly static/read-only. There is
792 * only a single copy of this record in the environment.
794 typedef struct MDB_dbx {
795 MDB_val md_name; /**< name of the database */
796 MDB_cmp_func *md_cmp; /**< function for comparing keys */
797 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
798 MDB_rel_func *md_rel; /**< user relocate function */
799 void *md_relctx; /**< user-provided context for md_rel */
802 /** A database transaction.
803 * Every operation requires a transaction handle.
806 MDB_txn *mt_parent; /**< parent of a nested txn */
807 MDB_txn *mt_child; /**< nested txn under this txn */
808 pgno_t mt_next_pgno; /**< next unallocated page */
809 /** The ID of this transaction. IDs are integers incrementing from 1.
810 * Only committed write transactions increment the ID. If a transaction
811 * aborts, the ID may be re-used by the next writer.
814 MDB_env *mt_env; /**< the DB environment */
815 /** The list of pages that became unused during this transaction.
819 MDB_ID2L dirty_list; /**< for write txns: modified pages */
820 MDB_reader *reader; /**< this thread's reader table slot or NULL */
822 /** Array of records for each DB known in the environment. */
824 /** Array of MDB_db records for each known DB */
826 /** @defgroup mt_dbflag Transaction DB Flags
830 #define DB_DIRTY 0x01 /**< DB was written in this txn */
831 #define DB_STALE 0x02 /**< DB record is older than txnID */
832 #define DB_NEW 0x04 /**< DB handle opened in this txn */
833 #define DB_VALID 0x08 /**< DB handle is valid */
834 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
836 /** In write txns, array of cursors for each DB */
837 MDB_cursor **mt_cursors;
838 /** Array of flags for each DB */
839 unsigned char *mt_dbflags;
840 /** Number of DB records in use. This number only ever increments;
841 * we don't decrement it when individual DB handles are closed.
845 /** @defgroup mdb_txn Transaction Flags
849 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
850 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
851 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
853 unsigned int mt_flags; /**< @ref mdb_txn */
854 /** dirty_list maxsize - #allocated pages including in parent txns */
855 unsigned int mt_dirty_room;
856 /** Tracks which of the two meta pages was used at the start
857 * of this transaction.
859 unsigned int mt_toggle;
862 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
863 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
864 * raise this on a 64 bit machine.
866 #define CURSOR_STACK 32
870 /** Cursors are used for all DB operations */
872 /** Next cursor on this DB in this txn */
874 /** Original cursor if this is a shadow */
876 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
877 struct MDB_xcursor *mc_xcursor;
878 /** The transaction that owns this cursor */
880 /** The database handle this cursor operates on */
882 /** The database record for this cursor */
884 /** The database auxiliary record for this cursor */
886 /** The @ref mt_dbflag for this database */
887 unsigned char *mc_dbflag;
888 unsigned short mc_snum; /**< number of pushed pages */
889 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
890 /** @defgroup mdb_cursor Cursor Flags
892 * Cursor state flags.
895 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
896 #define C_EOF 0x02 /**< No more data */
897 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
898 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
899 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
900 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
902 unsigned int mc_flags; /**< @ref mdb_cursor */
903 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
904 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
907 /** Context for sorted-dup records.
908 * We could have gone to a fully recursive design, with arbitrarily
909 * deep nesting of sub-databases. But for now we only handle these
910 * levels - main DB, optional sub-DB, sorted-duplicate DB.
912 typedef struct MDB_xcursor {
913 /** A sub-cursor for traversing the Dup DB */
914 MDB_cursor mx_cursor;
915 /** The database record for this Dup DB */
917 /** The auxiliary DB record for this Dup DB */
919 /** The @ref mt_dbflag for this Dup DB */
920 unsigned char mx_dbflag;
923 /** State of FreeDB old pages, stored in the MDB_env */
924 typedef struct MDB_pgstate {
925 txnid_t mf_pglast; /**< ID of last old page record we used */
926 pgno_t *mf_pghead; /**< old pages reclaimed from freelist */
927 pgno_t *mf_pgfree; /**< memory to free when dropping me_pghead */
930 /** The database environment. */
932 HANDLE me_fd; /**< The main data file */
933 HANDLE me_lfd; /**< The lock file */
934 HANDLE me_mfd; /**< just for writing the meta pages */
935 /** Failed to update the meta page. Probably an I/O error. */
936 #define MDB_FATAL_ERROR 0x80000000U
937 /** Read-only Filesystem. Allow read access, no locking. */
938 #define MDB_ROFS 0x40000000U
939 /** Some fields are initialized. */
940 #define MDB_ENV_ACTIVE 0x20000000U
941 uint32_t me_flags; /**< @ref mdb_env */
942 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
943 unsigned int me_maxreaders; /**< size of the reader table */
944 unsigned int me_numreaders; /**< max numreaders set by this env */
945 MDB_dbi me_numdbs; /**< number of DBs opened */
946 MDB_dbi me_maxdbs; /**< size of the DB table */
947 pid_t me_pid; /**< process ID of this env */
948 char *me_path; /**< path to the DB files */
949 char *me_map; /**< the memory map of the data file */
950 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
951 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
952 MDB_txn *me_txn; /**< current write transaction */
953 size_t me_mapsize; /**< size of the data memory map */
954 off_t me_size; /**< current file size */
955 pgno_t me_maxpg; /**< me_mapsize / me_psize */
956 MDB_dbx *me_dbxs; /**< array of static DB info */
957 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
958 pthread_key_t me_txkey; /**< thread-key for readers */
959 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
960 # define me_pglast me_pgstate.mf_pglast
961 # define me_pghead me_pgstate.mf_pghead
962 # define me_pgfree me_pgstate.mf_pgfree
963 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
964 /** IDL of pages that became unused in a write txn */
966 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
967 MDB_ID2L me_dirty_list;
968 /** Max number of freelist items that can fit in a single overflow page */
969 unsigned int me_maxfree_1pg;
970 /** Max size of a node on a page */
971 unsigned int me_nodemax;
973 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
975 #elif defined(MDB_USE_POSIX_SEM)
976 sem_t *me_rmutex; /* Shared mutexes are not supported */
981 /** Nested transaction */
982 typedef struct MDB_ntxn {
983 MDB_txn mnt_txn; /* the transaction */
984 MDB_pgstate mnt_pgstate; /* parent transaction's saved freestate */
987 /** max number of pages to commit in one writev() call */
988 #define MDB_COMMIT_PAGES 64
989 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
990 #undef MDB_COMMIT_PAGES
991 #define MDB_COMMIT_PAGES IOV_MAX
994 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
995 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
996 static int mdb_page_touch(MDB_cursor *mc);
998 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
999 static int mdb_page_search_root(MDB_cursor *mc,
1000 MDB_val *key, int modify);
1001 #define MDB_PS_MODIFY 1
1002 #define MDB_PS_ROOTONLY 2
1003 static int mdb_page_search(MDB_cursor *mc,
1004 MDB_val *key, int flags);
1005 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1007 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1008 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1009 pgno_t newpgno, unsigned int nflags);
1011 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1012 static int mdb_env_pick_meta(const MDB_env *env);
1013 static int mdb_env_write_meta(MDB_txn *txn);
1014 static void mdb_env_close0(MDB_env *env, int excl);
1016 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1017 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1018 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1019 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1020 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1021 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1022 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1023 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1024 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1026 static int mdb_rebalance(MDB_cursor *mc);
1027 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1029 static void mdb_cursor_pop(MDB_cursor *mc);
1030 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1032 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1033 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1034 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1035 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1036 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1038 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1039 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1041 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1042 static void mdb_xcursor_init0(MDB_cursor *mc);
1043 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1045 static int mdb_drop0(MDB_cursor *mc, int subs);
1046 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1049 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1053 static SECURITY_DESCRIPTOR mdb_null_sd;
1054 static SECURITY_ATTRIBUTES mdb_all_sa;
1055 static int mdb_sec_inited;
1058 /** Return the library version info. */
1060 mdb_version(int *major, int *minor, int *patch)
1062 if (major) *major = MDB_VERSION_MAJOR;
1063 if (minor) *minor = MDB_VERSION_MINOR;
1064 if (patch) *patch = MDB_VERSION_PATCH;
1065 return MDB_VERSION_STRING;
1068 /** Table of descriptions for MDB @ref errors */
1069 static char *const mdb_errstr[] = {
1070 "MDB_KEYEXIST: Key/data pair already exists",
1071 "MDB_NOTFOUND: No matching key/data pair found",
1072 "MDB_PAGE_NOTFOUND: Requested page not found",
1073 "MDB_CORRUPTED: Located page was wrong type",
1074 "MDB_PANIC: Update of meta page failed",
1075 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1076 "MDB_INVALID: File is not an MDB file",
1077 "MDB_MAP_FULL: Environment mapsize limit reached",
1078 "MDB_DBS_FULL: Environment maxdbs limit reached",
1079 "MDB_READERS_FULL: Environment maxreaders limit reached",
1080 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1081 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1082 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1083 "MDB_PAGE_FULL: Internal error - page has no more space",
1084 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1085 "MDB_INCOMPATIBLE: Database flags changed or would change",
1089 mdb_strerror(int err)
1093 return ("Successful return: 0");
1095 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1096 i = err - MDB_KEYEXIST;
1097 return mdb_errstr[i];
1100 return strerror(err);
1104 /** Display a key in hexadecimal and return the address of the result.
1105 * @param[in] key the key to display
1106 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1107 * @return The key in hexadecimal form.
1110 mdb_dkey(MDB_val *key, char *buf)
1113 unsigned char *c = key->mv_data;
1119 if (key->mv_size > MDB_MAXKEYSIZE)
1120 return "MDB_MAXKEYSIZE";
1121 /* may want to make this a dynamic check: if the key is mostly
1122 * printable characters, print it as-is instead of converting to hex.
1126 for (i=0; i<key->mv_size; i++)
1127 ptr += sprintf(ptr, "%02x", *c++);
1129 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1134 /** Display all the keys in the page. */
1136 mdb_page_list(MDB_page *mp)
1139 unsigned int i, nkeys, nsize;
1143 nkeys = NUMKEYS(mp);
1144 fprintf(stderr, "Page %zu numkeys %d\n", mp->mp_pgno, nkeys);
1145 for (i=0; i<nkeys; i++) {
1146 node = NODEPTR(mp, i);
1147 key.mv_size = node->mn_ksize;
1148 key.mv_data = node->mn_data;
1149 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1150 if (IS_BRANCH(mp)) {
1151 fprintf(stderr, "key %d: page %zu, %s\n", i, NODEPGNO(node),
1154 if (F_ISSET(node->mn_flags, F_BIGDATA))
1155 nsize += sizeof(pgno_t);
1157 nsize += NODEDSZ(node);
1158 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1164 mdb_cursor_chk(MDB_cursor *mc)
1170 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1171 for (i=0; i<mc->mc_top; i++) {
1173 node = NODEPTR(mp, mc->mc_ki[i]);
1174 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1177 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1183 /** Count all the pages in each DB and in the freelist
1184 * and make sure it matches the actual number of pages
1187 static void mdb_audit(MDB_txn *txn)
1191 MDB_ID freecount, count;
1196 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1197 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1198 freecount += *(MDB_ID *)data.mv_data;
1201 for (i = 0; i<txn->mt_numdbs; i++) {
1202 MDB_xcursor mx, *mxp;
1203 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1204 mdb_cursor_init(&mc, txn, i, mxp);
1205 if (txn->mt_dbs[i].md_root == P_INVALID)
1207 count += txn->mt_dbs[i].md_branch_pages +
1208 txn->mt_dbs[i].md_leaf_pages +
1209 txn->mt_dbs[i].md_overflow_pages;
1210 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1211 mdb_page_search(&mc, NULL, 0);
1215 mp = mc.mc_pg[mc.mc_top];
1216 for (j=0; j<NUMKEYS(mp); j++) {
1217 MDB_node *leaf = NODEPTR(mp, j);
1218 if (leaf->mn_flags & F_SUBDATA) {
1220 memcpy(&db, NODEDATA(leaf), sizeof(db));
1221 count += db.md_branch_pages + db.md_leaf_pages +
1222 db.md_overflow_pages;
1226 while (mdb_cursor_sibling(&mc, 1) == 0);
1229 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1230 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1231 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1237 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1239 return txn->mt_dbxs[dbi].md_cmp(a, b);
1243 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1245 if (txn->mt_dbxs[dbi].md_dcmp)
1246 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1248 return EINVAL; /* too bad you can't distinguish this from a valid result */
1251 /** Allocate a single page.
1252 * Re-use old malloc'd pages first, otherwise just malloc.
1255 mdb_page_malloc(MDB_cursor *mc) {
1257 size_t sz = mc->mc_txn->mt_env->me_psize;
1258 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1259 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1260 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1261 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1262 } else if ((ret = malloc(sz)) != NULL) {
1263 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1269 mdb_page_free(MDB_env *env, MDB_page *mp)
1271 mp->mp_next = env->me_dpages;
1272 VGMEMP_FREE(env, mp);
1273 env->me_dpages = mp;
1276 /** Allocate pages for writing.
1277 * If there are free pages available from older transactions, they
1278 * will be re-used first. Otherwise a new page will be allocated.
1279 * @param[in] mc cursor A cursor handle identifying the transaction and
1280 * database for which we are allocating.
1281 * @param[in] num the number of pages to allocate.
1282 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1283 * will always be satisfied by a single contiguous chunk of memory.
1284 * @return 0 on success, non-zero on failure.
1287 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1289 MDB_txn *txn = mc->mc_txn;
1291 pgno_t pgno = P_INVALID;
1293 txnid_t oldest = 0, last;
1298 /* If our dirty list is already full, we can't do anything */
1299 if (txn->mt_dirty_room == 0)
1300 return MDB_TXN_FULL;
1302 /* The free list won't have any content at all until txn 2 has
1303 * committed. The pages freed by txn 2 will be unreferenced
1304 * after txn 3 commits, and so will be safe to re-use in txn 4.
1306 if (txn->mt_txnid > 3) {
1307 if (!txn->mt_env->me_pghead &&
1308 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1309 /* See if there's anything in the free DB */
1316 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1317 if (!txn->mt_env->me_pglast) {
1318 mdb_page_search(&m2, NULL, 0);
1319 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1320 kptr = (txnid_t *)NODEKEY(leaf);
1325 last = txn->mt_env->me_pglast + 1;
1327 key.mv_data = &last;
1328 key.mv_size = sizeof(last);
1329 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET_RANGE, NULL);
1332 last = *(txnid_t *)key.mv_data;
1338 oldest = txn->mt_txnid - 1;
1339 nr = txn->mt_env->me_txns->mti_numreaders;
1340 r = txn->mt_env->me_txns->mti_readers;
1341 for (i=0; i<nr; i++) {
1342 if (!r[i].mr_pid) continue;
1349 if (oldest > last) {
1350 /* It's usable, grab it.
1354 if (!txn->mt_env->me_pglast) {
1355 mdb_node_read(txn, leaf, &data);
1357 idl = (MDB_ID *) data.mv_data;
1358 /* We might have a zero-length IDL due to freelist growth
1359 * during a prior commit
1362 txn->mt_env->me_pglast = last;
1365 mop = malloc(MDB_IDL_SIZEOF(idl));
1368 txn->mt_env->me_pglast = last;
1369 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop;
1370 memcpy(mop, idl, MDB_IDL_SIZEOF(idl));
1375 DPRINTF("IDL read txn %zu root %zu num %zu",
1376 last, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1377 for (i=0; i<idl[0]; i++) {
1378 DPRINTF("IDL %zu", idl[i+1]);
1385 if (txn->mt_env->me_pghead) {
1386 pgno_t *mop = txn->mt_env->me_pghead;
1389 int retry = 1, readit = 0, n2 = num-1;
1390 unsigned int i, j, k;
1392 /* If current list is too short, must fetch more and coalesce */
1393 if (mop[0] < (unsigned)num)
1396 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1398 /* If on freelist, don't try to read more. If what we have
1399 * right now isn't enough just use new pages.
1400 * TODO: get all of this working. Many circular dependencies...
1402 if (mc->mc_dbi == FREE_DBI) {
1410 last = txn->mt_env->me_pglast + 1;
1412 /* We haven't hit the readers list yet? */
1418 oldest = txn->mt_txnid - 1;
1419 nr = txn->mt_env->me_txns->mti_numreaders;
1420 r = txn->mt_env->me_txns->mti_readers;
1421 for (i=0; i<nr; i++) {
1422 if (!r[i].mr_pid) continue;
1429 /* There's nothing we can use on the freelist */
1430 if (oldest - last < 1)
1433 key.mv_data = &last;
1434 key.mv_size = sizeof(last);
1435 rc = mdb_cursor_set(&m2,&key,&data,MDB_SET_RANGE,NULL);
1437 if (rc == MDB_NOTFOUND)
1441 last = *(txnid_t*)key.mv_data;
1444 idl = (MDB_ID *) data.mv_data;
1445 mop2 = malloc(MDB_IDL_SIZEOF(idl) + MDB_IDL_SIZEOF(mop));
1448 /* merge in sorted order */
1449 i = idl[0]; j = mop[0]; mop2[0] = k = i+j;
1451 while (i>0 || j>0) {
1452 if (i && idl[i] < mop[j])
1453 mop2[k--] = idl[i--];
1455 mop2[k--] = mop[j--];
1457 txn->mt_env->me_pglast = last;
1458 free(txn->mt_env->me_pgfree);
1459 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = mop2;
1461 /* Keep trying to read until we have enough */
1462 if (mop[0] < (unsigned)num) {
1467 /* current list has enough pages, but are they contiguous? */
1468 for (i=mop[0]; i>=(unsigned)num; i--) {
1469 if (mop[i-n2] == mop[i] + n2) {
1472 /* move any stragglers down */
1473 for (j=i+num; j<=mop[0]; j++)
1480 /* Stop if we succeeded, or no retries */
1481 if (!retry || pgno != P_INVALID)
1487 /* peel pages off tail, so we only have to truncate the list */
1488 pgno = MDB_IDL_LAST(mop);
1491 if (MDB_IDL_IS_ZERO(mop)) {
1492 free(txn->mt_env->me_pgfree);
1493 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
1498 if (pgno == P_INVALID) {
1499 /* DB size is maxed out */
1500 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1501 DPUTS("DB size maxed out");
1502 return MDB_MAP_FULL;
1505 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1506 if (pgno == P_INVALID) {
1507 pgno = txn->mt_next_pgno;
1508 txn->mt_next_pgno += num;
1510 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1513 if (txn->mt_env->me_dpages && num == 1) {
1514 np = txn->mt_env->me_dpages;
1515 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1516 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1517 txn->mt_env->me_dpages = np->mp_next;
1519 size_t sz = txn->mt_env->me_psize * num;
1520 if ((np = malloc(sz)) == NULL)
1522 VGMEMP_ALLOC(txn->mt_env, np, sz);
1524 if (pgno == P_INVALID) {
1525 np->mp_pgno = txn->mt_next_pgno;
1526 txn->mt_next_pgno += num;
1531 mid.mid = np->mp_pgno;
1533 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1534 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1536 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1538 txn->mt_dirty_room--;
1544 /** Copy a page: avoid copying unused portions of the page.
1545 * @param[in] dst page to copy into
1546 * @param[in] src page to copy from
1549 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1551 dst->mp_flags = src->mp_flags | P_DIRTY;
1552 dst->mp_pages = src->mp_pages;
1554 if (IS_LEAF2(src)) {
1555 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1557 unsigned int i, nkeys = NUMKEYS(src);
1558 for (i=0; i<nkeys; i++)
1559 dst->mp_ptrs[i] = src->mp_ptrs[i];
1560 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1561 psize - src->mp_upper);
1565 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1566 * @param[in] mc cursor pointing to the page to be touched
1567 * @return 0 on success, non-zero on failure.
1570 mdb_page_touch(MDB_cursor *mc)
1572 MDB_page *mp = mc->mc_pg[mc->mc_top];
1576 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1578 if ((rc = mdb_page_alloc(mc, 1, &np)))
1580 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1581 assert(mp->mp_pgno != np->mp_pgno);
1582 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1584 /* If page isn't full, just copy the used portion */
1585 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1588 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1590 np->mp_flags |= P_DIRTY;
1595 /* Adjust other cursors pointing to mp */
1596 if (mc->mc_flags & C_SUB) {
1597 MDB_cursor *m2, *m3;
1598 MDB_dbi dbi = mc->mc_dbi-1;
1600 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1601 if (m2 == mc) continue;
1602 m3 = &m2->mc_xcursor->mx_cursor;
1603 if (m3->mc_snum < mc->mc_snum) continue;
1604 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1605 m3->mc_pg[mc->mc_top] = mp;
1611 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1612 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1613 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1614 m2->mc_pg[mc->mc_top] = mp;
1618 mc->mc_pg[mc->mc_top] = mp;
1619 /** If this page has a parent, update the parent to point to
1623 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1625 mc->mc_db->md_root = mp->mp_pgno;
1626 } else if (mc->mc_txn->mt_parent) {
1629 /* If txn has a parent, make sure the page is in our
1632 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1633 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1634 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1635 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1636 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1637 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1638 mc->mc_pg[mc->mc_top] = mp;
1643 assert(mc->mc_txn->mt_u.dirty_list[0].mid < MDB_IDL_UM_MAX);
1645 np = mdb_page_malloc(mc);
1648 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1649 mid.mid = np->mp_pgno;
1651 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1659 mdb_env_sync(MDB_env *env, int force)
1662 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1663 if (env->me_flags & MDB_WRITEMAP) {
1664 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1665 ? MS_ASYNC : MS_SYNC;
1666 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1669 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1673 if (MDB_FDATASYNC(env->me_fd))
1680 /** Make shadow copies of all of parent txn's cursors */
1682 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1684 MDB_cursor *mc, *m2;
1685 unsigned int i, j, size;
1687 for (i=0;i<src->mt_numdbs; i++) {
1688 if (src->mt_cursors[i]) {
1689 size = sizeof(MDB_cursor);
1690 if (src->mt_cursors[i]->mc_xcursor)
1691 size += sizeof(MDB_xcursor);
1692 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1699 mc->mc_db = &dst->mt_dbs[i];
1700 mc->mc_dbx = m2->mc_dbx;
1701 mc->mc_dbflag = &dst->mt_dbflags[i];
1702 mc->mc_snum = m2->mc_snum;
1703 mc->mc_top = m2->mc_top;
1704 mc->mc_flags = m2->mc_flags | C_SHADOW;
1705 for (j=0; j<mc->mc_snum; j++) {
1706 mc->mc_pg[j] = m2->mc_pg[j];
1707 mc->mc_ki[j] = m2->mc_ki[j];
1709 if (m2->mc_xcursor) {
1710 MDB_xcursor *mx, *mx2;
1711 mx = (MDB_xcursor *)(mc+1);
1712 mc->mc_xcursor = mx;
1713 mx2 = m2->mc_xcursor;
1714 mx->mx_db = mx2->mx_db;
1715 mx->mx_dbx = mx2->mx_dbx;
1716 mx->mx_dbflag = mx2->mx_dbflag;
1717 mx->mx_cursor.mc_txn = dst;
1718 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1719 mx->mx_cursor.mc_db = &mx->mx_db;
1720 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1721 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1722 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1723 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1724 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1725 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1726 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1727 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1730 mc->mc_xcursor = NULL;
1732 mc->mc_next = dst->mt_cursors[i];
1733 dst->mt_cursors[i] = mc;
1740 /** Merge shadow cursors back into parent's */
1742 mdb_cursor_merge(MDB_txn *txn)
1745 for (i=0; i<txn->mt_numdbs; i++) {
1746 if (txn->mt_cursors[i]) {
1748 while ((mc = txn->mt_cursors[i])) {
1749 txn->mt_cursors[i] = mc->mc_next;
1750 if (mc->mc_flags & C_SHADOW) {
1751 MDB_cursor *m2 = mc->mc_orig;
1753 m2->mc_snum = mc->mc_snum;
1754 m2->mc_top = mc->mc_top;
1755 for (j=0; j<mc->mc_snum; j++) {
1756 m2->mc_pg[j] = mc->mc_pg[j];
1757 m2->mc_ki[j] = mc->mc_ki[j];
1760 if (mc->mc_flags & C_ALLOCD)
1768 mdb_txn_reset0(MDB_txn *txn);
1770 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1771 * @param[in] txn the transaction handle to initialize
1772 * @return 0 on success, non-zero on failure. This can only
1773 * fail for read-only transactions, and then only if the
1774 * reader table is full.
1777 mdb_txn_renew0(MDB_txn *txn)
1779 MDB_env *env = txn->mt_env;
1785 txn->mt_numdbs = env->me_numdbs;
1786 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1788 if (txn->mt_flags & MDB_TXN_RDONLY) {
1789 if (env->me_flags & MDB_ROFS) {
1790 i = mdb_env_pick_meta(env);
1791 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1792 txn->mt_u.reader = NULL;
1794 MDB_reader *r = pthread_getspecific(env->me_txkey);
1796 pid_t pid = env->me_pid;
1797 pthread_t tid = pthread_self();
1800 for (i=0; i<env->me_txns->mti_numreaders; i++)
1801 if (env->me_txns->mti_readers[i].mr_pid == 0)
1803 if (i == env->me_maxreaders) {
1804 UNLOCK_MUTEX_R(env);
1805 return MDB_READERS_FULL;
1807 env->me_txns->mti_readers[i].mr_pid = pid;
1808 env->me_txns->mti_readers[i].mr_tid = tid;
1809 if (i >= env->me_txns->mti_numreaders)
1810 env->me_txns->mti_numreaders = i+1;
1811 /* Save numreaders for un-mutexed mdb_env_close() */
1812 env->me_numreaders = env->me_txns->mti_numreaders;
1813 UNLOCK_MUTEX_R(env);
1814 r = &env->me_txns->mti_readers[i];
1815 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1816 env->me_txns->mti_readers[i].mr_pid = 0;
1820 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1821 txn->mt_u.reader = r;
1823 txn->mt_toggle = txn->mt_txnid & 1;
1824 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1828 txn->mt_txnid = env->me_txns->mti_txnid;
1829 txn->mt_toggle = txn->mt_txnid & 1;
1830 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1833 if (txn->mt_txnid == mdb_debug_start)
1836 txn->mt_dirty_room = MDB_IDL_UM_MAX;
1837 txn->mt_u.dirty_list = env->me_dirty_list;
1838 txn->mt_u.dirty_list[0].mid = 0;
1839 txn->mt_free_pgs = env->me_free_pgs;
1840 txn->mt_free_pgs[0] = 0;
1844 /* Copy the DB info and flags */
1845 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1846 for (i=2; i<txn->mt_numdbs; i++) {
1847 x = env->me_dbflags[i];
1848 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
1849 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
1851 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
1853 if (env->me_maxpg < txn->mt_next_pgno) {
1854 mdb_txn_reset0(txn);
1855 return MDB_MAP_RESIZED;
1862 mdb_txn_renew(MDB_txn *txn)
1866 if (! (txn && (txn->mt_flags & MDB_TXN_RDONLY)))
1869 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1870 DPUTS("environment had fatal error, must shutdown!");
1874 rc = mdb_txn_renew0(txn);
1875 if (rc == MDB_SUCCESS) {
1876 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1877 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1878 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1884 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1888 int rc, size, tsize = sizeof(MDB_txn);
1890 if (env->me_flags & MDB_FATAL_ERROR) {
1891 DPUTS("environment had fatal error, must shutdown!");
1894 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1897 /* Nested transactions: Max 1 child, write txns only, no writemap */
1898 if (parent->mt_child ||
1899 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1900 (env->me_flags & MDB_WRITEMAP))
1904 tsize = sizeof(MDB_ntxn);
1906 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
1907 if (!(flags & MDB_RDONLY))
1908 size += env->me_maxdbs * sizeof(MDB_cursor *);
1910 if ((txn = calloc(1, size)) == NULL) {
1911 DPRINTF("calloc: %s", strerror(ErrCode()));
1914 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
1915 if (flags & MDB_RDONLY) {
1916 txn->mt_flags |= MDB_TXN_RDONLY;
1917 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1919 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1920 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1926 txn->mt_free_pgs = mdb_midl_alloc();
1927 if (!txn->mt_free_pgs) {
1931 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1932 if (!txn->mt_u.dirty_list) {
1933 free(txn->mt_free_pgs);
1937 txn->mt_txnid = parent->mt_txnid;
1938 txn->mt_toggle = parent->mt_toggle;
1939 txn->mt_dirty_room = parent->mt_dirty_room;
1940 txn->mt_u.dirty_list[0].mid = 0;
1941 txn->mt_free_pgs[0] = 0;
1942 txn->mt_next_pgno = parent->mt_next_pgno;
1943 parent->mt_child = txn;
1944 txn->mt_parent = parent;
1945 txn->mt_numdbs = parent->mt_numdbs;
1946 txn->mt_dbxs = parent->mt_dbxs;
1947 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1948 /* Copy parent's mt_dbflags, but clear DB_NEW */
1949 for (i=0; i<txn->mt_numdbs; i++)
1950 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
1952 ntxn = (MDB_ntxn *)txn;
1953 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
1954 if (env->me_pghead) {
1955 size = MDB_IDL_SIZEOF(env->me_pghead);
1956 env->me_pghead = malloc(size);
1958 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
1962 env->me_pgfree = env->me_pghead;
1964 rc = mdb_cursor_shadow(parent, txn);
1966 mdb_txn_reset0(txn);
1968 rc = mdb_txn_renew0(txn);
1974 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1975 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1976 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1982 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1983 * @param[in] txn the transaction handle to reset
1986 mdb_txn_reset0(MDB_txn *txn)
1988 MDB_env *env = txn->mt_env;
1991 /* Close any DBI handles opened in this txn */
1992 for (i=2; i<txn->mt_numdbs; i++) {
1993 if (txn->mt_dbflags[i] & DB_NEW) {
1994 char *ptr = env->me_dbxs[i].md_name.mv_data;
1995 env->me_dbxs[i].md_name.mv_data = NULL;
1996 env->me_dbxs[i].md_name.mv_size = 0;
2001 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2002 if (!(env->me_flags & MDB_ROFS))
2003 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2007 /* close(free) all cursors */
2008 for (i=0; i<txn->mt_numdbs; i++) {
2009 if (txn->mt_cursors[i]) {
2011 while ((mc = txn->mt_cursors[i])) {
2012 txn->mt_cursors[i] = mc->mc_next;
2013 if (mc->mc_flags & C_ALLOCD)
2019 if (!(env->me_flags & MDB_WRITEMAP)) {
2020 /* return all dirty pages to dpage list */
2021 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2022 dp = txn->mt_u.dirty_list[i].mptr;
2023 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2024 mdb_page_free(txn->mt_env, dp);
2026 /* large pages just get freed directly */
2027 VGMEMP_FREE(txn->mt_env, dp);
2033 free(env->me_pgfree);
2035 if (txn->mt_parent) {
2036 txn->mt_parent->mt_child = NULL;
2037 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2038 mdb_midl_free(txn->mt_free_pgs);
2039 free(txn->mt_u.dirty_list);
2042 if (mdb_midl_shrink(&txn->mt_free_pgs))
2043 env->me_free_pgs = txn->mt_free_pgs;
2046 txn->mt_env->me_pghead = txn->mt_env->me_pgfree = NULL;
2047 txn->mt_env->me_pglast = 0;
2050 /* The writer mutex was locked in mdb_txn_begin. */
2051 UNLOCK_MUTEX_W(env);
2056 mdb_txn_reset(MDB_txn *txn)
2061 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
2062 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2063 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
2065 mdb_txn_reset0(txn);
2069 mdb_txn_abort(MDB_txn *txn)
2074 DPRINTF("abort 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);
2079 mdb_txn_abort(txn->mt_child);
2081 mdb_txn_reset0(txn);
2086 mdb_txn_commit(MDB_txn *txn)
2094 pgno_t next, freecnt;
2095 txnid_t oldpg_txnid, id;
2098 assert(txn != NULL);
2099 assert(txn->mt_env != NULL);
2101 if (txn->mt_child) {
2102 mdb_txn_commit(txn->mt_child);
2103 txn->mt_child = NULL;
2108 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2109 /* update the DB flags */
2110 for (i = 2; i<txn->mt_numdbs; i++) {
2111 if (txn->mt_dbflags[i] & DB_NEW)
2112 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2114 if (txn->mt_numdbs > env->me_numdbs)
2115 env->me_numdbs = txn->mt_numdbs;
2116 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
2121 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2122 DPUTS("error flag is set, can't commit");
2124 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2129 if (txn->mt_parent) {
2130 MDB_txn *parent = txn->mt_parent;
2134 /* Append our free list to parent's */
2135 if (mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs)) {
2139 mdb_midl_free(txn->mt_free_pgs);
2141 parent->mt_next_pgno = txn->mt_next_pgno;
2142 parent->mt_flags = txn->mt_flags;
2144 /* Merge (and close) our cursors with parent's */
2145 mdb_cursor_merge(txn);
2147 /* Update parent's DB table. */
2148 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2149 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2150 txn->mt_parent->mt_dbflags[0] = txn->mt_dbflags[0];
2151 txn->mt_parent->mt_dbflags[1] = txn->mt_dbflags[1];
2152 for (i=2; i<txn->mt_numdbs; i++) {
2153 /* preserve parent's DB_NEW status */
2154 x = txn->mt_parent->mt_dbflags[i] & DB_NEW;
2155 txn->mt_parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
2158 dst = txn->mt_parent->mt_u.dirty_list;
2159 src = txn->mt_u.dirty_list;
2160 /* Find len = length of merging our dirty list with parent's */
2162 dst[0].mid = 0; /* simplify loops */
2163 if (parent->mt_parent) {
2164 len = x + src[0].mid;
2165 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
2166 for (i = x; y && i; y--) {
2167 pgno_t yp = src[y].mid;
2168 while (yp < dst[i].mid)
2170 if (yp == dst[i].mid) {
2175 } else { /* Simplify the above for single-ancestor case */
2176 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
2178 /* Merge our dirty list with parent's */
2180 for (i = len; y; dst[i--] = src[y--]) {
2181 pgno_t yp = src[y].mid;
2182 while (yp < dst[x].mid)
2183 dst[i--] = dst[x--];
2184 if (yp == dst[x].mid)
2185 free(dst[x--].mptr);
2189 free(txn->mt_u.dirty_list);
2190 parent->mt_dirty_room = txn->mt_dirty_room;
2192 txn->mt_parent->mt_child = NULL;
2193 free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pgfree);
2198 if (txn != env->me_txn) {
2199 DPUTS("attempt to commit unknown transaction");
2204 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2207 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2208 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2210 /* Update DB root pointers */
2211 if (txn->mt_numdbs > 2) {
2214 data.mv_size = sizeof(MDB_db);
2216 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2217 for (i = 2; i < txn->mt_numdbs; i++) {
2218 if (txn->mt_dbflags[i] & DB_DIRTY) {
2219 data.mv_data = &txn->mt_dbs[i];
2220 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2227 /* Save the freelist as of this transaction to the freeDB. This
2228 * can change the freelist, so keep trying until it stabilizes.
2230 * env->me_pglast and the length of txn->mt_free_pgs cannot decrease,
2231 * except the code below can decrease env->me_pglast to split pghead.
2232 * Page numbers cannot disappear from txn->mt_free_pgs. New pages
2233 * can only appear in env->me_pghead when env->me_pglast increases.
2234 * Until then, the me_pghead pointer won't move but can become NULL.
2237 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2238 oldpg_txnid = id = 0;
2241 /* should only be one record now */
2242 if (env->me_pghead || env->me_pglast) {
2243 /* make sure first page of freeDB is touched and on freelist */
2244 rc = mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2245 if (rc && rc != MDB_NOTFOUND) {
2252 /* Delete IDLs we used from the free list */
2253 if (env->me_pglast) {
2258 rc = mdb_cursor_first(&mc, &key, NULL);
2261 oldpg_txnid = *(txnid_t *)key.mv_data;
2263 assert(oldpg_txnid <= env->me_pglast);
2265 rc = mdb_cursor_del(&mc, 0);
2268 } while (oldpg_txnid < env->me_pglast);
2271 /* Save IDL of pages freed by this txn, to freeDB */
2273 if (freecnt != txn->mt_free_pgs[0]) {
2276 /* make sure last page of freeDB is touched and on freelist */
2277 key.mv_size = MDB_MAXKEYSIZE+1;
2279 rc = mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2280 if (rc && rc != MDB_NOTFOUND)
2286 MDB_IDL idl = txn->mt_free_pgs;
2287 mdb_midl_sort(txn->mt_free_pgs);
2288 DPRINTF("IDL write txn %zu root %zu num %zu",
2289 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2290 for (i=1; i<=idl[0]; i++) {
2291 DPRINTF("IDL %zu", idl[i]);
2295 /* write to last page of freeDB */
2296 key.mv_size = sizeof(pgno_t);
2297 key.mv_data = &txn->mt_txnid;
2298 /* The free list can still grow during this call,
2299 * despite the pre-emptive touches above. So retry
2300 * until the reserved space remains big enough.
2303 assert(freecnt < txn->mt_free_pgs[0]);
2304 freecnt = txn->mt_free_pgs[0];
2305 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2306 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2309 } while (freecnt != txn->mt_free_pgs[0]);
2310 mdb_midl_sort(txn->mt_free_pgs);
2311 memcpy(data.mv_data, txn->mt_free_pgs, data.mv_size);
2312 if (oldpg_txnid < env->me_pglast || (!env->me_pghead && id))
2313 goto free_pgfirst; /* used up freeDB[oldpg_txnid] */
2316 /* Put back page numbers we took from freeDB but did not use */
2317 if (env->me_pghead) {
2322 mop = env->me_pghead;
2323 id = env->me_pglast;
2324 key.mv_size = sizeof(id);
2326 /* These steps may grow the freelist again
2327 * due to freed overflow pages...
2332 if (orig > env->me_maxfree_1pg && id > 4)
2333 orig = env->me_maxfree_1pg; /* Do not use more than 1 page */
2334 data.mv_size = (orig + 1) * sizeof(pgno_t);
2335 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2338 assert(!env->me_pghead || env->me_pglast);
2339 /* mop could have been used again here */
2340 if (id != env->me_pglast || env->me_pghead == NULL)
2341 goto again; /* was completely used up */
2342 assert(mop == env->me_pghead);
2343 } while (mop[0] < orig && --i);
2344 memcpy(data.mv_data, mop, data.mv_size);
2347 *(pgno_t *)data.mv_data = orig;
2348 mop[orig] = mop[0] - orig;
2349 env->me_pghead = mop += orig;
2350 /* Save more oldpages at the previous txnid. */
2351 assert(env->me_pglast == id && id == oldpg_txnid);
2352 env->me_pglast = --oldpg_txnid;
2356 /* Check for growth of freelist again */
2357 if (freecnt != txn->mt_free_pgs[0])
2360 free(env->me_pgfree);
2361 env->me_pghead = env->me_pgfree = NULL;
2363 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2364 if (mdb_midl_shrink(&txn->mt_free_pgs))
2365 env->me_free_pgs = txn->mt_free_pgs;
2372 if (env->me_flags & MDB_WRITEMAP) {
2373 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2374 dp = txn->mt_u.dirty_list[i].mptr;
2375 /* clear dirty flag */
2376 dp->mp_flags &= ~P_DIRTY;
2377 txn->mt_u.dirty_list[i].mid = 0;
2379 txn->mt_u.dirty_list[0].mid = 0;
2383 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2389 /* Windows actually supports scatter/gather I/O, but only on
2390 * unbuffered file handles. Since we're relying on the OS page
2391 * cache for all our data, that's self-defeating. So we just
2392 * write pages one at a time. We use the ov structure to set
2393 * the write offset, to at least save the overhead of a Seek
2397 memset(&ov, 0, sizeof(ov));
2398 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2400 dp = txn->mt_u.dirty_list[i].mptr;
2401 DPRINTF("committing page %zu", dp->mp_pgno);
2402 size = dp->mp_pgno * env->me_psize;
2403 ov.Offset = size & 0xffffffff;
2404 ov.OffsetHigh = size >> 16;
2405 ov.OffsetHigh >>= 16;
2406 /* clear dirty flag */
2407 dp->mp_flags &= ~P_DIRTY;
2408 wsize = env->me_psize;
2409 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2410 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2413 DPRINTF("WriteFile: %d", n);
2420 struct iovec iov[MDB_COMMIT_PAGES];
2424 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2425 dp = txn->mt_u.dirty_list[i].mptr;
2426 if (dp->mp_pgno != next) {
2428 rc = writev(env->me_fd, iov, n);
2432 DPUTS("short write, filesystem full?");
2434 DPRINTF("writev: %s", strerror(n));
2441 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2444 DPRINTF("committing page %zu", dp->mp_pgno);
2445 iov[n].iov_len = env->me_psize;
2446 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2447 iov[n].iov_base = (char *)dp;
2448 size += iov[n].iov_len;
2449 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2450 /* clear dirty flag */
2451 dp->mp_flags &= ~P_DIRTY;
2452 if (++n >= MDB_COMMIT_PAGES) {
2462 rc = writev(env->me_fd, iov, n);
2466 DPUTS("short write, filesystem full?");
2468 DPRINTF("writev: %s", strerror(n));
2475 /* Drop the dirty pages.
2477 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2478 dp = txn->mt_u.dirty_list[i].mptr;
2479 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2480 mdb_page_free(txn->mt_env, dp);
2482 VGMEMP_FREE(txn->mt_env, dp);
2485 txn->mt_u.dirty_list[i].mid = 0;
2487 txn->mt_u.dirty_list[0].mid = 0;
2490 if ((n = mdb_env_sync(env, 0)) != 0 ||
2491 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2499 /* update the DB flags */
2500 for (i = 2; i<txn->mt_numdbs; i++) {
2501 if (txn->mt_dbflags[i] & DB_NEW)
2502 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2504 if (txn->mt_numdbs > env->me_numdbs)
2505 env->me_numdbs = txn->mt_numdbs;
2507 UNLOCK_MUTEX_W(env);
2513 /** Read the environment parameters of a DB environment before
2514 * mapping it into memory.
2515 * @param[in] env the environment handle
2516 * @param[out] meta address of where to store the meta information
2517 * @return 0 on success, non-zero on failure.
2520 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2527 /* We don't know the page size yet, so use a minimum value.
2528 * Read both meta pages so we can use the latest one.
2531 for (i=0; i<2; i++) {
2533 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2535 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2540 else if (rc != MDB_PAGESIZE) {
2544 DPRINTF("read: %s", strerror(err));
2548 p = (MDB_page *)&pbuf;
2550 if (!F_ISSET(p->mp_flags, P_META)) {
2551 DPRINTF("page %zu not a meta page", p->mp_pgno);
2556 if (m->mm_magic != MDB_MAGIC) {
2557 DPUTS("meta has invalid magic");
2561 if (m->mm_version != MDB_VERSION) {
2562 DPRINTF("database is version %u, expected version %u",
2563 m->mm_version, MDB_VERSION);
2564 return MDB_VERSION_MISMATCH;
2568 if (m->mm_txnid > meta->mm_txnid)
2569 memcpy(meta, m, sizeof(*m));
2571 memcpy(meta, m, sizeof(*m));
2573 if (SetFilePointer(env->me_fd, meta->mm_psize, NULL, FILE_BEGIN) != meta->mm_psize)
2575 if (lseek(env->me_fd, meta->mm_psize, SEEK_SET) != meta->mm_psize)
2583 /** Write the environment parameters of a freshly created DB environment.
2584 * @param[in] env the environment handle
2585 * @param[out] meta address of where to store the meta information
2586 * @return 0 on success, non-zero on failure.
2589 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2596 DPUTS("writing new meta page");
2598 GET_PAGESIZE(psize);
2600 meta->mm_magic = MDB_MAGIC;
2601 meta->mm_version = MDB_VERSION;
2602 meta->mm_mapsize = env->me_mapsize;
2603 meta->mm_psize = psize;
2604 meta->mm_last_pg = 1;
2605 meta->mm_flags = env->me_flags & 0xffff;
2606 meta->mm_flags |= MDB_INTEGERKEY;
2607 meta->mm_dbs[0].md_root = P_INVALID;
2608 meta->mm_dbs[1].md_root = P_INVALID;
2610 p = calloc(2, psize);
2612 p->mp_flags = P_META;
2615 memcpy(m, meta, sizeof(*meta));
2617 q = (MDB_page *)((char *)p + psize);
2620 q->mp_flags = P_META;
2623 memcpy(m, meta, sizeof(*meta));
2628 SetFilePointer(env->me_fd, 0, NULL, FILE_BEGIN);
2629 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2630 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2633 lseek(env->me_fd, 0, SEEK_SET);
2634 rc = write(env->me_fd, p, psize * 2);
2635 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2641 /** Update the environment info to commit a transaction.
2642 * @param[in] txn the transaction that's being committed
2643 * @return 0 on success, non-zero on failure.
2646 mdb_env_write_meta(MDB_txn *txn)
2649 MDB_meta meta, metab, *mp;
2651 int rc, len, toggle;
2658 assert(txn != NULL);
2659 assert(txn->mt_env != NULL);
2661 toggle = !txn->mt_toggle;
2662 DPRINTF("writing meta page %d for root page %zu",
2663 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2666 mp = env->me_metas[toggle];
2668 if (env->me_flags & MDB_WRITEMAP) {
2669 /* Persist any increases of mapsize config */
2670 if (env->me_mapsize > mp->mm_mapsize)
2671 mp->mm_mapsize = env->me_mapsize;
2672 mp->mm_dbs[0] = txn->mt_dbs[0];
2673 mp->mm_dbs[1] = txn->mt_dbs[1];
2674 mp->mm_last_pg = txn->mt_next_pgno - 1;
2675 mp->mm_txnid = txn->mt_txnid;
2676 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2677 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2680 ptr += env->me_psize;
2681 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2688 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2689 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2691 ptr = (char *)&meta;
2692 if (env->me_mapsize > mp->mm_mapsize) {
2693 /* Persist any increases of mapsize config */
2694 meta.mm_mapsize = env->me_mapsize;
2695 off = offsetof(MDB_meta, mm_mapsize);
2697 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2699 len = sizeof(MDB_meta) - off;
2702 meta.mm_dbs[0] = txn->mt_dbs[0];
2703 meta.mm_dbs[1] = txn->mt_dbs[1];
2704 meta.mm_last_pg = txn->mt_next_pgno - 1;
2705 meta.mm_txnid = txn->mt_txnid;
2708 off += env->me_psize;
2711 /* Write to the SYNC fd */
2712 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2713 env->me_fd : env->me_mfd;
2716 memset(&ov, 0, sizeof(ov));
2718 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2721 rc = pwrite(mfd, ptr, len, off);
2726 DPUTS("write failed, disk error?");
2727 /* On a failure, the pagecache still contains the new data.
2728 * Write some old data back, to prevent it from being used.
2729 * Use the non-SYNC fd; we know it will fail anyway.
2731 meta.mm_last_pg = metab.mm_last_pg;
2732 meta.mm_txnid = metab.mm_txnid;
2734 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2736 r2 = pwrite(env->me_fd, ptr, len, off);
2739 env->me_flags |= MDB_FATAL_ERROR;
2743 /* Memory ordering issues are irrelevant; since the entire writer
2744 * is wrapped by wmutex, all of these changes will become visible
2745 * after the wmutex is unlocked. Since the DB is multi-version,
2746 * readers will get consistent data regardless of how fresh or
2747 * how stale their view of these values is.
2749 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2754 /** Check both meta pages to see which one is newer.
2755 * @param[in] env the environment handle
2756 * @return meta toggle (0 or 1).
2759 mdb_env_pick_meta(const MDB_env *env)
2761 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2765 mdb_env_create(MDB_env **env)
2769 e = calloc(1, sizeof(MDB_env));
2773 e->me_maxreaders = DEFAULT_READERS;
2775 e->me_fd = INVALID_HANDLE_VALUE;
2776 e->me_lfd = INVALID_HANDLE_VALUE;
2777 e->me_mfd = INVALID_HANDLE_VALUE;
2778 #ifdef MDB_USE_POSIX_SEM
2779 e->me_rmutex = SEM_FAILED;
2780 e->me_wmutex = SEM_FAILED;
2782 e->me_pid = getpid();
2783 VGMEMP_CREATE(e,0,0);
2789 mdb_env_set_mapsize(MDB_env *env, size_t size)
2793 env->me_mapsize = size;
2795 env->me_maxpg = env->me_mapsize / env->me_psize;
2800 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2804 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2809 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2811 if (env->me_map || readers < 1)
2813 env->me_maxreaders = readers;
2818 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2820 if (!env || !readers)
2822 *readers = env->me_maxreaders;
2826 /** Further setup required for opening an MDB environment
2829 mdb_env_open2(MDB_env *env)
2831 unsigned int flags = env->me_flags;
2832 int i, newenv = 0, prot;
2836 memset(&meta, 0, sizeof(meta));
2838 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2841 DPUTS("new mdbenv");
2845 /* Was a mapsize configured? */
2846 if (!env->me_mapsize) {
2847 /* If this is a new environment, take the default,
2848 * else use the size recorded in the existing env.
2850 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2851 } else if (env->me_mapsize < meta.mm_mapsize) {
2852 /* If the configured size is smaller, make sure it's
2853 * still big enough. Silently round up to minimum if not.
2855 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
2856 if (env->me_mapsize < minsize)
2857 env->me_mapsize = minsize;
2863 LONG sizelo, sizehi;
2864 sizelo = env->me_mapsize & 0xffffffff;
2865 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2867 /* Windows won't create mappings for zero length files.
2868 * Just allocate the maxsize right now.
2871 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2872 if (!SetEndOfFile(env->me_fd))
2874 SetFilePointer(env->me_fd, 0, NULL, 0);
2876 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2877 PAGE_READWRITE : PAGE_READONLY,
2878 sizehi, sizelo, NULL);
2881 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2882 FILE_MAP_WRITE : FILE_MAP_READ,
2883 0, 0, env->me_mapsize, meta.mm_address);
2891 if (flags & MDB_WRITEMAP) {
2893 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2896 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2898 if (env->me_map == MAP_FAILED) {
2902 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
2904 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
2906 #ifdef POSIX_MADV_RANDOM
2907 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
2908 #endif /* POSIX_MADV_RANDOM */
2909 #endif /* MADV_RANDOM */
2913 if (flags & MDB_FIXEDMAP)
2914 meta.mm_address = env->me_map;
2915 i = mdb_env_init_meta(env, &meta);
2916 if (i != MDB_SUCCESS) {
2919 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2920 /* Can happen because the address argument to mmap() is just a
2921 * hint. mmap() can pick another, e.g. if the range is in use.
2922 * The MAP_FIXED flag would prevent that, but then mmap could
2923 * instead unmap existing pages to make room for the new map.
2925 return EBUSY; /* TODO: Make a new MDB_* error code? */
2927 env->me_psize = meta.mm_psize;
2928 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
2929 env->me_nodemax = (env->me_psize - PAGEHDRSZ) / MDB_MINKEYS;
2931 env->me_maxpg = env->me_mapsize / env->me_psize;
2933 p = (MDB_page *)env->me_map;
2934 env->me_metas[0] = METADATA(p);
2935 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2939 int toggle = mdb_env_pick_meta(env);
2940 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2942 DPRINTF("opened database version %u, pagesize %u",
2943 env->me_metas[0]->mm_version, env->me_psize);
2944 DPRINTF("using meta page %d", toggle);
2945 DPRINTF("depth: %u", db->md_depth);
2946 DPRINTF("entries: %zu", db->md_entries);
2947 DPRINTF("branch pages: %zu", db->md_branch_pages);
2948 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2949 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2950 DPRINTF("root: %zu", db->md_root);
2958 /** Release a reader thread's slot in the reader lock table.
2959 * This function is called automatically when a thread exits.
2960 * @param[in] ptr This points to the slot in the reader lock table.
2963 mdb_env_reader_dest(void *ptr)
2965 MDB_reader *reader = ptr;
2971 /** Junk for arranging thread-specific callbacks on Windows. This is
2972 * necessarily platform and compiler-specific. Windows supports up
2973 * to 1088 keys. Let's assume nobody opens more than 64 environments
2974 * in a single process, for now. They can override this if needed.
2976 #ifndef MAX_TLS_KEYS
2977 #define MAX_TLS_KEYS 64
2979 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2980 static int mdb_tls_nkeys;
2982 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2986 case DLL_PROCESS_ATTACH: break;
2987 case DLL_THREAD_ATTACH: break;
2988 case DLL_THREAD_DETACH:
2989 for (i=0; i<mdb_tls_nkeys; i++) {
2990 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2991 mdb_env_reader_dest(r);
2994 case DLL_PROCESS_DETACH: break;
2999 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3001 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3005 /* Force some symbol references.
3006 * _tls_used forces the linker to create the TLS directory if not already done
3007 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3009 #pragma comment(linker, "/INCLUDE:_tls_used")
3010 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3011 #pragma const_seg(".CRT$XLB")
3012 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
3013 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3016 #pragma comment(linker, "/INCLUDE:__tls_used")
3017 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3018 #pragma data_seg(".CRT$XLB")
3019 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3021 #endif /* WIN 32/64 */
3022 #endif /* !__GNUC__ */
3025 /** Downgrade the exclusive lock on the region back to shared */
3027 mdb_env_share_locks(MDB_env *env, int *excl)
3029 int rc = 0, toggle = mdb_env_pick_meta(env);
3031 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3036 /* First acquire a shared lock. The Unlock will
3037 * then release the existing exclusive lock.
3039 memset(&ov, 0, sizeof(ov));
3040 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3043 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3049 struct flock lock_info;
3050 /* The shared lock replaces the existing lock */
3051 memset((void *)&lock_info, 0, sizeof(lock_info));
3052 lock_info.l_type = F_RDLCK;
3053 lock_info.l_whence = SEEK_SET;
3054 lock_info.l_start = 0;
3055 lock_info.l_len = 1;
3056 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3057 (rc = ErrCode()) == EINTR) ;
3058 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3065 /** Try to get exlusive lock, otherwise shared.
3066 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3069 mdb_env_excl_lock(MDB_env *env, int *excl)
3073 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3077 memset(&ov, 0, sizeof(ov));
3078 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3085 struct flock lock_info;
3086 memset((void *)&lock_info, 0, sizeof(lock_info));
3087 lock_info.l_type = F_WRLCK;
3088 lock_info.l_whence = SEEK_SET;
3089 lock_info.l_start = 0;
3090 lock_info.l_len = 1;
3091 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3092 (rc = ErrCode()) == EINTR) ;
3096 # ifdef MDB_USE_POSIX_SEM
3097 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3100 lock_info.l_type = F_RDLCK;
3101 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3102 (rc = ErrCode()) == EINTR) ;
3110 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
3112 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3114 * @(#) $Revision: 5.1 $
3115 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3116 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3118 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3122 * Please do not copyright this code. This code is in the public domain.
3124 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3125 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3126 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3127 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3128 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3129 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3130 * PERFORMANCE OF THIS SOFTWARE.
3133 * chongo <Landon Curt Noll> /\oo/\
3134 * http://www.isthe.com/chongo/
3136 * Share and Enjoy! :-)
3139 typedef unsigned long long mdb_hash_t;
3140 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3142 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3143 * @param[in] str string to hash
3144 * @param[in] hval initial value for hash
3145 * @return 64 bit hash
3147 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3148 * hval arg on the first call.
3151 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3153 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3154 unsigned char *end = s + val->mv_size;
3156 * FNV-1a hash each octet of the string
3159 /* xor the bottom with the current octet */
3160 hval ^= (mdb_hash_t)*s++;
3162 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3163 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3164 (hval << 7) + (hval << 8) + (hval << 40);
3166 /* return our new hash value */
3170 /** Hash the string and output the hash in hex.
3171 * @param[in] str string to hash
3172 * @param[out] hexbuf an array of 17 chars to hold the hash
3175 mdb_hash_hex(MDB_val *val, char *hexbuf)
3178 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3179 for (i=0; i<8; i++) {
3180 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3186 /** Open and/or initialize the lock region for the environment.
3187 * @param[in] env The MDB environment.
3188 * @param[in] lpath The pathname of the file used for the lock region.
3189 * @param[in] mode The Unix permissions for the file, if we create it.
3190 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3191 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
3192 * @return 0 on success, non-zero on failure.
3195 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3198 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
3200 # define MDB_ERRCODE_ROFS EROFS
3201 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
3202 # define MDB_CLOEXEC O_CLOEXEC
3205 # define MDB_CLOEXEC 0
3212 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3213 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3214 FILE_ATTRIBUTE_NORMAL, NULL);
3216 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
3218 if (env->me_lfd == INVALID_HANDLE_VALUE) {
3220 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
3221 env->me_flags |= MDB_ROFS;
3226 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
3227 /* Lose record locks when exec*() */
3228 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3229 fcntl(env->me_lfd, F_SETFD, fdflags);
3232 /* Try to get exclusive lock. If we succeed, then
3233 * nobody is using the lock region and we should initialize it.
3235 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3238 size = GetFileSize(env->me_lfd, NULL);
3240 size = lseek(env->me_lfd, 0, SEEK_END);
3242 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3243 if (size < rsize && *excl > 0) {
3245 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3246 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3248 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3252 size = rsize - sizeof(MDB_txninfo);
3253 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3258 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3260 if (!mh) goto fail_errno;
3261 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3263 if (!env->me_txns) goto fail_errno;
3265 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3267 if (m == MAP_FAILED) goto fail_errno;
3273 BY_HANDLE_FILE_INFORMATION stbuf;
3282 if (!mdb_sec_inited) {
3283 InitializeSecurityDescriptor(&mdb_null_sd,
3284 SECURITY_DESCRIPTOR_REVISION);
3285 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3286 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3287 mdb_all_sa.bInheritHandle = FALSE;
3288 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3291 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3292 idbuf.volume = stbuf.dwVolumeSerialNumber;
3293 idbuf.nhigh = stbuf.nFileIndexHigh;
3294 idbuf.nlow = stbuf.nFileIndexLow;
3295 val.mv_data = &idbuf;
3296 val.mv_size = sizeof(idbuf);
3297 mdb_hash_hex(&val, hexbuf);
3298 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3299 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3300 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3301 if (!env->me_rmutex) goto fail_errno;
3302 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3303 if (!env->me_wmutex) goto fail_errno;
3304 #elif defined(MDB_USE_POSIX_SEM)
3313 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3314 idbuf.dev = stbuf.st_dev;
3315 idbuf.ino = stbuf.st_ino;
3316 val.mv_data = &idbuf;
3317 val.mv_size = sizeof(idbuf);
3318 mdb_hash_hex(&val, hexbuf);
3319 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3320 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3321 /* Clean up after a previous run, if needed: Try to
3322 * remove both semaphores before doing anything else.
3324 sem_unlink(env->me_txns->mti_rmname);
3325 sem_unlink(env->me_txns->mti_wmname);
3326 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3327 O_CREAT|O_EXCL, mode, 1);
3328 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3329 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3330 O_CREAT|O_EXCL, mode, 1);
3331 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3332 #else /* MDB_USE_POSIX_SEM */
3333 pthread_mutexattr_t mattr;
3335 if ((rc = pthread_mutexattr_init(&mattr))
3336 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3337 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3338 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3340 pthread_mutexattr_destroy(&mattr);
3341 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3343 env->me_txns->mti_version = MDB_VERSION;
3344 env->me_txns->mti_magic = MDB_MAGIC;
3345 env->me_txns->mti_txnid = 0;
3346 env->me_txns->mti_numreaders = 0;
3349 if (env->me_txns->mti_magic != MDB_MAGIC) {
3350 DPUTS("lock region has invalid magic");
3354 if (env->me_txns->mti_version != MDB_VERSION) {
3355 DPRINTF("lock region is version %u, expected version %u",
3356 env->me_txns->mti_version, MDB_VERSION);
3357 rc = MDB_VERSION_MISMATCH;
3361 if (rc != EACCES && rc != EAGAIN) {
3365 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3366 if (!env->me_rmutex) goto fail_errno;
3367 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3368 if (!env->me_wmutex) goto fail_errno;
3369 #elif defined(MDB_USE_POSIX_SEM)
3370 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3371 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3372 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3373 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3384 /** The name of the lock file in the DB environment */
3385 #define LOCKNAME "/lock.mdb"
3386 /** The name of the data file in the DB environment */
3387 #define DATANAME "/data.mdb"
3388 /** The suffix of the lock file when no subdir is used */
3389 #define LOCKSUFF "-lock"
3390 /** Only a subset of the @ref mdb_env flags can be changed
3391 * at runtime. Changing other flags requires closing the
3392 * environment and re-opening it with the new flags.
3394 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3395 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3398 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
3400 int oflags, rc, len, excl = -1;
3401 char *lpath, *dpath;
3403 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3407 if (flags & MDB_NOSUBDIR) {
3408 rc = len + sizeof(LOCKSUFF) + len + 1;
3410 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3415 if (flags & MDB_NOSUBDIR) {
3416 dpath = lpath + len + sizeof(LOCKSUFF);
3417 sprintf(lpath, "%s" LOCKSUFF, path);
3418 strcpy(dpath, path);
3420 dpath = lpath + len + sizeof(LOCKNAME);
3421 sprintf(lpath, "%s" LOCKNAME, path);
3422 sprintf(dpath, "%s" DATANAME, path);
3426 flags |= env->me_flags;
3427 if (flags & MDB_RDONLY) {
3428 /* silently ignore WRITEMAP when we're only getting read access */
3429 flags &= ~MDB_WRITEMAP;
3431 if (!((env->me_free_pgs = mdb_midl_alloc()) &&
3432 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
3435 env->me_flags = flags |= MDB_ENV_ACTIVE;
3439 env->me_path = strdup(path);
3440 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3441 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3442 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
3447 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3452 if (F_ISSET(flags, MDB_RDONLY)) {
3453 oflags = GENERIC_READ;
3454 len = OPEN_EXISTING;
3456 oflags = GENERIC_READ|GENERIC_WRITE;
3459 mode = FILE_ATTRIBUTE_NORMAL;
3460 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3461 NULL, len, mode, NULL);
3463 if (F_ISSET(flags, MDB_RDONLY))
3466 oflags = O_RDWR | O_CREAT;
3468 env->me_fd = open(dpath, oflags, mode);
3470 if (env->me_fd == INVALID_HANDLE_VALUE) {
3475 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3476 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3477 env->me_mfd = env->me_fd;
3479 /* Synchronous fd for meta writes. Needed even with
3480 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3483 env->me_mfd = CreateFile(dpath, oflags,
3484 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3485 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3487 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3489 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3494 DPRINTF("opened dbenv %p", (void *) env);
3495 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3498 env->me_numdbs = 2; /* this notes that me_txkey was set */
3500 /* Windows TLS callbacks need help finding their TLS info. */
3501 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3502 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3509 rc = mdb_env_share_locks(env, &excl);
3515 mdb_env_close0(env, excl);
3521 /** Destroy resources from mdb_env_open() and clear our readers */
3523 mdb_env_close0(MDB_env *env, int excl)
3527 if (!(env->me_flags & MDB_ENV_ACTIVE))
3530 free(env->me_dbflags);
3533 free(env->me_dirty_list);
3534 if (env->me_free_pgs)
3535 mdb_midl_free(env->me_free_pgs);
3537 if (env->me_numdbs) {
3538 pthread_key_delete(env->me_txkey);
3540 /* Delete our key from the global list */
3541 for (i=0; i<mdb_tls_nkeys; i++)
3542 if (mdb_tls_keys[i] == env->me_txkey) {
3543 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3551 munmap(env->me_map, env->me_mapsize);
3553 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3555 if (env->me_fd != INVALID_HANDLE_VALUE)
3558 pid_t pid = env->me_pid;
3559 /* Clearing readers is done in this function because
3560 * me_txkey with its destructor must be disabled first.
3562 for (i = env->me_numreaders; --i >= 0; )
3563 if (env->me_txns->mti_readers[i].mr_pid == pid)
3564 env->me_txns->mti_readers[i].mr_pid = 0;
3566 if (env->me_rmutex) {
3567 CloseHandle(env->me_rmutex);
3568 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3570 /* Windows automatically destroys the mutexes when
3571 * the last handle closes.
3573 #elif defined(MDB_USE_POSIX_SEM)
3574 if (env->me_rmutex != SEM_FAILED) {
3575 sem_close(env->me_rmutex);
3576 if (env->me_wmutex != SEM_FAILED)
3577 sem_close(env->me_wmutex);
3578 /* If we have the filelock: If we are the
3579 * only remaining user, clean up semaphores.
3582 mdb_env_excl_lock(env, &excl);
3584 sem_unlink(env->me_txns->mti_rmname);
3585 sem_unlink(env->me_txns->mti_wmname);
3589 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3591 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3594 /* Unlock the lockfile. Windows would have unlocked it
3595 * after closing anyway, but not necessarily at once.
3597 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3603 env->me_flags &= ~MDB_ENV_ACTIVE;
3607 mdb_env_copy(MDB_env *env, const char *path)
3609 MDB_txn *txn = NULL;
3613 HANDLE newfd = INVALID_HANDLE_VALUE;
3615 if (env->me_flags & MDB_NOSUBDIR) {
3616 lpath = (char *)path;
3619 len += sizeof(DATANAME);
3620 lpath = malloc(len);
3623 sprintf(lpath, "%s" DATANAME, path);
3626 /* The destination path must exist, but the destination file must not.
3627 * We don't want the OS to cache the writes, since the source data is
3628 * already in the OS cache.
3631 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3632 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3634 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3640 if (!(env->me_flags & MDB_NOSUBDIR))
3642 if (newfd == INVALID_HANDLE_VALUE) {
3647 #ifdef F_NOCACHE /* __APPLE__ */
3648 rc = fcntl(newfd, F_NOCACHE, 1);
3655 /* Do the lock/unlock of the reader mutex before starting the
3656 * write txn. Otherwise other read txns could block writers.
3658 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3662 if (!(env->me_flags & MDB_ROFS)) {
3663 /* We must start the actual read txn after blocking writers */
3664 mdb_txn_reset0(txn);
3666 /* Temporarily block writers until we snapshot the meta pages */
3669 rc = mdb_txn_renew0(txn);
3671 UNLOCK_MUTEX_W(env);
3676 wsize = env->me_psize * 2;
3680 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3681 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3684 rc = write(newfd, env->me_map, wsize);
3685 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3687 if (! (env->me_flags & MDB_ROFS))
3688 UNLOCK_MUTEX_W(env);
3693 ptr = env->me_map + wsize;
3694 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3695 #define MAX_WRITE 2147483648U
3699 if (wsize > MAX_WRITE)
3703 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3704 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3713 if (wsize > MAX_WRITE)
3717 wres = write(newfd, ptr, w2);
3718 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3727 if (newfd != INVALID_HANDLE_VALUE)
3734 mdb_env_close(MDB_env *env)
3742 for (i = env->me_numdbs; --i > MAIN_DBI; )
3743 free(env->me_dbxs[i].md_name.mv_data);
3745 VGMEMP_DESTROY(env);
3746 while ((dp = env->me_dpages) != NULL) {
3747 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3748 env->me_dpages = dp->mp_next;
3752 mdb_env_close0(env, 0);
3756 /** Compare two items pointing at aligned size_t's */
3758 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3760 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3761 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3764 /** Compare two items pointing at aligned int's */
3766 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3768 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3769 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3772 /** Compare two items pointing at ints of unknown alignment.
3773 * Nodes and keys are guaranteed to be 2-byte aligned.
3776 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3778 #if BYTE_ORDER == LITTLE_ENDIAN
3779 unsigned short *u, *c;
3782 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3783 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3786 } while(!x && u > (unsigned short *)a->mv_data);
3789 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3793 /** Compare two items lexically */
3795 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3802 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3808 diff = memcmp(a->mv_data, b->mv_data, len);
3809 return diff ? diff : len_diff<0 ? -1 : len_diff;
3812 /** Compare two items in reverse byte order */
3814 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3816 const unsigned char *p1, *p2, *p1_lim;
3820 p1_lim = (const unsigned char *)a->mv_data;
3821 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3822 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3824 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3830 while (p1 > p1_lim) {
3831 diff = *--p1 - *--p2;
3835 return len_diff<0 ? -1 : len_diff;
3838 /** Search for key within a page, using binary search.
3839 * Returns the smallest entry larger or equal to the key.
3840 * If exactp is non-null, stores whether the found entry was an exact match
3841 * in *exactp (1 or 0).
3842 * Updates the cursor index with the index of the found entry.
3843 * If no entry larger or equal to the key is found, returns NULL.
3846 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3848 unsigned int i = 0, nkeys;
3851 MDB_page *mp = mc->mc_pg[mc->mc_top];
3852 MDB_node *node = NULL;
3857 nkeys = NUMKEYS(mp);
3862 COPY_PGNO(pgno, mp->mp_pgno);
3863 DPRINTF("searching %u keys in %s %spage %zu",
3864 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3871 low = IS_LEAF(mp) ? 0 : 1;
3873 cmp = mc->mc_dbx->md_cmp;
3875 /* Branch pages have no data, so if using integer keys,
3876 * alignment is guaranteed. Use faster mdb_cmp_int.
3878 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3879 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3886 nodekey.mv_size = mc->mc_db->md_pad;
3887 node = NODEPTR(mp, 0); /* fake */
3888 while (low <= high) {
3889 i = (low + high) >> 1;
3890 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3891 rc = cmp(key, &nodekey);
3892 DPRINTF("found leaf index %u [%s], rc = %i",
3893 i, DKEY(&nodekey), rc);
3902 while (low <= high) {
3903 i = (low + high) >> 1;
3905 node = NODEPTR(mp, i);
3906 nodekey.mv_size = NODEKSZ(node);
3907 nodekey.mv_data = NODEKEY(node);
3909 rc = cmp(key, &nodekey);
3912 DPRINTF("found leaf index %u [%s], rc = %i",
3913 i, DKEY(&nodekey), rc);
3915 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3916 i, DKEY(&nodekey), NODEPGNO(node), rc);
3927 if (rc > 0) { /* Found entry is less than the key. */
3928 i++; /* Skip to get the smallest entry larger than key. */
3930 node = NODEPTR(mp, i);
3933 *exactp = (rc == 0);
3934 /* store the key index */
3935 mc->mc_ki[mc->mc_top] = i;
3937 /* There is no entry larger or equal to the key. */
3940 /* nodeptr is fake for LEAF2 */
3946 mdb_cursor_adjust(MDB_cursor *mc, func)
3950 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3951 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3958 /** Pop a page off the top of the cursor's stack. */
3960 mdb_cursor_pop(MDB_cursor *mc)
3963 #ifndef MDB_DEBUG_SKIP
3964 MDB_page *top = mc->mc_pg[mc->mc_top];
3970 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3971 mc->mc_dbi, (void *) mc);
3975 /** Push a page onto the top of the cursor's stack. */
3977 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3979 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3980 mc->mc_dbi, (void *) mc);
3982 if (mc->mc_snum >= CURSOR_STACK) {
3983 assert(mc->mc_snum < CURSOR_STACK);
3984 return MDB_CURSOR_FULL;
3987 mc->mc_top = mc->mc_snum++;
3988 mc->mc_pg[mc->mc_top] = mp;
3989 mc->mc_ki[mc->mc_top] = 0;
3994 /** Find the address of the page corresponding to a given page number.
3995 * @param[in] txn the transaction for this access.
3996 * @param[in] pgno the page number for the page to retrieve.
3997 * @param[out] ret address of a pointer where the page's address will be stored.
3998 * @return 0 on success, non-zero on failure.
4001 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
4005 if (!((txn->mt_flags & MDB_TXN_RDONLY) |
4006 (txn->mt_env->me_flags & MDB_WRITEMAP)))
4010 MDB_ID2L dl = tx2->mt_u.dirty_list;
4012 unsigned x = mdb_mid2l_search(dl, pgno);
4013 if (x <= dl[0].mid && dl[x].mid == pgno) {
4018 } while ((tx2 = tx2->mt_parent) != NULL);
4021 if (pgno < txn->mt_next_pgno) {
4022 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
4024 DPRINTF("page %zu not found", pgno);
4030 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
4033 /** Search for the page a given key should be in.
4034 * Pushes parent pages on the cursor stack. This function continues a
4035 * search on a cursor that has already been initialized. (Usually by
4036 * #mdb_page_search() but also by #mdb_node_move().)
4037 * @param[in,out] mc the cursor for this operation.
4038 * @param[in] key the key to search for. If NULL, search for the lowest
4039 * page. (This is used by #mdb_cursor_first().)
4040 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
4041 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4042 * @return 0 on success, non-zero on failure.
4045 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
4047 MDB_page *mp = mc->mc_pg[mc->mc_top];
4052 while (IS_BRANCH(mp)) {
4056 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
4057 assert(NUMKEYS(mp) > 1);
4058 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
4060 if (key == NULL) /* Initialize cursor to first page. */
4062 else if (key->mv_size > MDB_MAXKEYSIZE && key->mv_data == NULL) {
4063 /* cursor to last page */
4067 node = mdb_node_search(mc, key, &exact);
4069 i = NUMKEYS(mp) - 1;
4071 i = mc->mc_ki[mc->mc_top];
4080 DPRINTF("following index %u for key [%s]",
4082 assert(i < NUMKEYS(mp));
4083 node = NODEPTR(mp, i);
4085 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
4088 mc->mc_ki[mc->mc_top] = i;
4089 if ((rc = mdb_cursor_push(mc, mp)))
4093 if ((rc = mdb_page_touch(mc)) != 0)
4095 mp = mc->mc_pg[mc->mc_top];
4100 DPRINTF("internal error, index points to a %02X page!?",
4102 return MDB_CORRUPTED;
4105 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
4106 key ? DKEY(key) : NULL);
4111 /** Search for the page a given key should be in.
4112 * Pushes parent pages on the cursor stack. This function just sets up
4113 * the search; it finds the root page for \b mc's database and sets this
4114 * as the root of the cursor's stack. Then #mdb_page_search_root() is
4115 * called to complete the search.
4116 * @param[in,out] mc the cursor for this operation.
4117 * @param[in] key the key to search for. If NULL, search for the lowest
4118 * page. (This is used by #mdb_cursor_first().)
4119 * @param[in] modify If true, visited pages are updated with new page numbers.
4120 * @return 0 on success, non-zero on failure.
4123 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4128 /* Make sure the txn is still viable, then find the root from
4129 * the txn's db table.
4131 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4132 DPUTS("transaction has failed, must abort");
4135 /* Make sure we're using an up-to-date root */
4136 if (mc->mc_dbi > MAIN_DBI) {
4137 if ((*mc->mc_dbflag & DB_STALE) ||
4138 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
4140 unsigned char dbflag = 0;
4141 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4142 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
4145 if (*mc->mc_dbflag & DB_STALE) {
4149 MDB_node *leaf = mdb_node_search(&mc2,
4150 &mc->mc_dbx->md_name, &exact);
4152 return MDB_NOTFOUND;
4153 mdb_node_read(mc->mc_txn, leaf, &data);
4154 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4156 /* The txn may not know this DBI, or another process may
4157 * have dropped and recreated the DB with other flags.
4159 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4160 return MDB_INCOMPATIBLE;
4161 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4163 if (flags & MDB_PS_MODIFY)
4165 *mc->mc_dbflag &= ~DB_STALE;
4166 *mc->mc_dbflag |= dbflag;
4169 root = mc->mc_db->md_root;
4171 if (root == P_INVALID) { /* Tree is empty. */
4172 DPUTS("tree is empty");
4173 return MDB_NOTFOUND;
4178 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4179 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4185 DPRINTF("db %u root page %zu has flags 0x%X",
4186 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4188 if (flags & MDB_PS_MODIFY) {
4189 if ((rc = mdb_page_touch(mc)))
4193 if (flags & MDB_PS_ROOTONLY)
4196 return mdb_page_search_root(mc, key, flags);
4199 /** Return the data associated with a given node.
4200 * @param[in] txn The transaction for this operation.
4201 * @param[in] leaf The node being read.
4202 * @param[out] data Updated to point to the node's data.
4203 * @return 0 on success, non-zero on failure.
4206 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4208 MDB_page *omp; /* overflow page */
4212 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4213 data->mv_size = NODEDSZ(leaf);
4214 data->mv_data = NODEDATA(leaf);
4218 /* Read overflow data.
4220 data->mv_size = NODEDSZ(leaf);
4221 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4222 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4223 DPRINTF("read overflow page %zu failed", pgno);
4226 data->mv_data = METADATA(omp);
4232 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4233 MDB_val *key, MDB_val *data)
4242 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4244 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
4247 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4251 mdb_cursor_init(&mc, txn, dbi, &mx);
4252 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4255 /** Find a sibling for a page.
4256 * Replaces the page at the top of the cursor's stack with the
4257 * specified sibling, if one exists.
4258 * @param[in] mc The cursor for this operation.
4259 * @param[in] move_right Non-zero if the right sibling is requested,
4260 * otherwise the left sibling.
4261 * @return 0 on success, non-zero on failure.
4264 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4270 if (mc->mc_snum < 2) {
4271 return MDB_NOTFOUND; /* root has no siblings */
4275 DPRINTF("parent page is page %zu, index %u",
4276 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4278 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4279 : (mc->mc_ki[mc->mc_top] == 0)) {
4280 DPRINTF("no more keys left, moving to %s sibling",
4281 move_right ? "right" : "left");
4282 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4283 /* undo cursor_pop before returning */
4290 mc->mc_ki[mc->mc_top]++;
4292 mc->mc_ki[mc->mc_top]--;
4293 DPRINTF("just moving to %s index key %u",
4294 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4296 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4298 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4299 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4302 mdb_cursor_push(mc, mp);
4304 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
4309 /** Move the cursor to the next data item. */
4311 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4317 if (mc->mc_flags & C_EOF) {
4318 return MDB_NOTFOUND;
4321 assert(mc->mc_flags & C_INITIALIZED);
4323 mp = mc->mc_pg[mc->mc_top];
4325 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4326 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4327 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4328 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4329 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4330 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4334 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4335 if (op == MDB_NEXT_DUP)
4336 return MDB_NOTFOUND;
4340 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4342 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4343 DPUTS("=====> move to next sibling page");
4344 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4345 mc->mc_flags |= C_EOF;
4346 mc->mc_flags &= ~C_INITIALIZED;
4347 return MDB_NOTFOUND;
4349 mp = mc->mc_pg[mc->mc_top];
4350 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4352 mc->mc_ki[mc->mc_top]++;
4354 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4355 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4358 key->mv_size = mc->mc_db->md_pad;
4359 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4363 assert(IS_LEAF(mp));
4364 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4366 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4367 mdb_xcursor_init1(mc, leaf);
4370 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4373 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4374 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4375 if (rc != MDB_SUCCESS)
4380 MDB_GET_KEY(leaf, key);
4384 /** Move the cursor to the previous data item. */
4386 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4392 assert(mc->mc_flags & C_INITIALIZED);
4394 mp = mc->mc_pg[mc->mc_top];
4396 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4397 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4398 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4399 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4400 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4401 if (op != MDB_PREV || rc == MDB_SUCCESS)
4404 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4405 if (op == MDB_PREV_DUP)
4406 return MDB_NOTFOUND;
4411 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4413 if (mc->mc_ki[mc->mc_top] == 0) {
4414 DPUTS("=====> move to prev sibling page");
4415 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4416 mc->mc_flags &= ~C_INITIALIZED;
4417 return MDB_NOTFOUND;
4419 mp = mc->mc_pg[mc->mc_top];
4420 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4421 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4423 mc->mc_ki[mc->mc_top]--;
4425 mc->mc_flags &= ~C_EOF;
4427 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4428 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4431 key->mv_size = mc->mc_db->md_pad;
4432 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4436 assert(IS_LEAF(mp));
4437 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4439 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4440 mdb_xcursor_init1(mc, leaf);
4443 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4446 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4447 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4448 if (rc != MDB_SUCCESS)
4453 MDB_GET_KEY(leaf, key);
4457 /** Set the cursor on a specific data item. */
4459 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4460 MDB_cursor_op op, int *exactp)
4464 MDB_node *leaf = NULL;
4469 assert(key->mv_size > 0);
4471 /* See if we're already on the right page */
4472 if (mc->mc_flags & C_INITIALIZED) {
4475 mp = mc->mc_pg[mc->mc_top];
4477 mc->mc_ki[mc->mc_top] = 0;
4478 return MDB_NOTFOUND;
4480 if (mp->mp_flags & P_LEAF2) {
4481 nodekey.mv_size = mc->mc_db->md_pad;
4482 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4484 leaf = NODEPTR(mp, 0);
4485 MDB_GET_KEY(leaf, &nodekey);
4487 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4489 /* Probably happens rarely, but first node on the page
4490 * was the one we wanted.
4492 mc->mc_ki[mc->mc_top] = 0;
4499 unsigned int nkeys = NUMKEYS(mp);
4501 if (mp->mp_flags & P_LEAF2) {
4502 nodekey.mv_data = LEAF2KEY(mp,
4503 nkeys-1, nodekey.mv_size);
4505 leaf = NODEPTR(mp, nkeys-1);
4506 MDB_GET_KEY(leaf, &nodekey);
4508 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4510 /* last node was the one we wanted */
4511 mc->mc_ki[mc->mc_top] = nkeys-1;
4517 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4518 /* This is definitely the right page, skip search_page */
4519 if (mp->mp_flags & P_LEAF2) {
4520 nodekey.mv_data = LEAF2KEY(mp,
4521 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4523 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4524 MDB_GET_KEY(leaf, &nodekey);
4526 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4528 /* current node was the one we wanted */
4538 /* If any parents have right-sibs, search.
4539 * Otherwise, there's nothing further.
4541 for (i=0; i<mc->mc_top; i++)
4543 NUMKEYS(mc->mc_pg[i])-1)
4545 if (i == mc->mc_top) {
4546 /* There are no other pages */
4547 mc->mc_ki[mc->mc_top] = nkeys;
4548 return MDB_NOTFOUND;
4552 /* There are no other pages */
4553 mc->mc_ki[mc->mc_top] = 0;
4554 return MDB_NOTFOUND;
4558 rc = mdb_page_search(mc, key, 0);
4559 if (rc != MDB_SUCCESS)
4562 mp = mc->mc_pg[mc->mc_top];
4563 assert(IS_LEAF(mp));
4566 leaf = mdb_node_search(mc, key, exactp);
4567 if (exactp != NULL && !*exactp) {
4568 /* MDB_SET specified and not an exact match. */
4569 return MDB_NOTFOUND;
4573 DPUTS("===> inexact leaf not found, goto sibling");
4574 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4575 return rc; /* no entries matched */
4576 mp = mc->mc_pg[mc->mc_top];
4577 assert(IS_LEAF(mp));
4578 leaf = NODEPTR(mp, 0);
4582 mc->mc_flags |= C_INITIALIZED;
4583 mc->mc_flags &= ~C_EOF;
4586 key->mv_size = mc->mc_db->md_pad;
4587 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4591 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4592 mdb_xcursor_init1(mc, leaf);
4595 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4596 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4597 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4600 if (op == MDB_GET_BOTH) {
4606 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4607 if (rc != MDB_SUCCESS)
4610 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4612 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4614 rc = mc->mc_dbx->md_dcmp(data, &d2);
4616 if (op == MDB_GET_BOTH || rc > 0)
4617 return MDB_NOTFOUND;
4622 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4623 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4628 /* The key already matches in all other cases */
4629 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4630 MDB_GET_KEY(leaf, key);
4631 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4636 /** Move the cursor to the first item in the database. */
4638 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4643 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4644 rc = mdb_page_search(mc, NULL, 0);
4645 if (rc != MDB_SUCCESS)
4648 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4650 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4651 mc->mc_flags |= C_INITIALIZED;
4652 mc->mc_flags &= ~C_EOF;
4654 mc->mc_ki[mc->mc_top] = 0;
4656 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4657 key->mv_size = mc->mc_db->md_pad;
4658 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4663 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4664 mdb_xcursor_init1(mc, leaf);
4665 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4670 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4671 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4675 MDB_GET_KEY(leaf, key);
4679 /** Move the cursor to the last item in the database. */
4681 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4686 if (!(mc->mc_flags & C_EOF)) {
4688 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4691 lkey.mv_size = MDB_MAXKEYSIZE+1;
4692 lkey.mv_data = NULL;
4693 rc = mdb_page_search(mc, &lkey, 0);
4694 if (rc != MDB_SUCCESS)
4697 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4699 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4701 mc->mc_flags |= C_INITIALIZED|C_EOF;
4702 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4704 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4705 key->mv_size = mc->mc_db->md_pad;
4706 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4711 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4712 mdb_xcursor_init1(mc, leaf);
4713 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4718 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4719 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4724 MDB_GET_KEY(leaf, key);
4729 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4738 case MDB_GET_CURRENT:
4739 if (!(mc->mc_flags & C_INITIALIZED)) {
4742 MDB_page *mp = mc->mc_pg[mc->mc_top];
4744 mc->mc_ki[mc->mc_top] = 0;
4750 key->mv_size = mc->mc_db->md_pad;
4751 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4753 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4754 MDB_GET_KEY(leaf, key);
4756 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4757 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4759 rc = mdb_node_read(mc->mc_txn, leaf, data);
4766 case MDB_GET_BOTH_RANGE:
4767 if (data == NULL || mc->mc_xcursor == NULL) {
4775 if (key == NULL || key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
4777 } else if (op == MDB_SET_RANGE)
4778 rc = mdb_cursor_set(mc, key, data, op, NULL);
4780 rc = mdb_cursor_set(mc, key, data, op, &exact);
4782 case MDB_GET_MULTIPLE:
4784 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4785 !(mc->mc_flags & C_INITIALIZED)) {
4790 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4791 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4794 case MDB_NEXT_MULTIPLE:
4796 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4800 if (!(mc->mc_flags & C_INITIALIZED))
4801 rc = mdb_cursor_first(mc, key, data);
4803 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4804 if (rc == MDB_SUCCESS) {
4805 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4808 mx = &mc->mc_xcursor->mx_cursor;
4809 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4811 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4812 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4820 case MDB_NEXT_NODUP:
4821 if (!(mc->mc_flags & C_INITIALIZED))
4822 rc = mdb_cursor_first(mc, key, data);
4824 rc = mdb_cursor_next(mc, key, data, op);
4828 case MDB_PREV_NODUP:
4829 if (!(mc->mc_flags & C_INITIALIZED)) {
4830 rc = mdb_cursor_last(mc, key, data);
4831 mc->mc_flags |= C_INITIALIZED;
4832 mc->mc_ki[mc->mc_top]++;
4834 rc = mdb_cursor_prev(mc, key, data, op);
4837 rc = mdb_cursor_first(mc, key, data);
4841 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4842 !(mc->mc_flags & C_INITIALIZED) ||
4843 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4847 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4850 rc = mdb_cursor_last(mc, key, data);
4854 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4855 !(mc->mc_flags & C_INITIALIZED) ||
4856 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4860 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4863 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4871 /** Touch all the pages in the cursor stack.
4872 * Makes sure all the pages are writable, before attempting a write operation.
4873 * @param[in] mc The cursor to operate on.
4876 mdb_cursor_touch(MDB_cursor *mc)
4880 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4883 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4884 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4885 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4888 *mc->mc_dbflag |= DB_DIRTY;
4890 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4891 rc = mdb_page_touch(mc);
4895 mc->mc_top = mc->mc_snum-1;
4900 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4903 MDB_node *leaf = NULL;
4904 MDB_val xdata, *rdata, dkey;
4907 int do_sub = 0, insert = 0;
4908 unsigned int mcount = 0;
4912 char dbuf[MDB_MAXKEYSIZE+1];
4913 unsigned int nflags;
4916 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4919 if (flags != MDB_CURRENT && (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE))
4922 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT) && data->mv_size > MDB_MAXKEYSIZE)
4925 #if SIZE_MAX > MAXDATASIZE
4926 if (data->mv_size > MAXDATASIZE)
4930 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4931 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4935 if (flags == MDB_CURRENT) {
4936 if (!(mc->mc_flags & C_INITIALIZED))
4939 } else if (mc->mc_db->md_root == P_INVALID) {
4941 /* new database, write a root leaf page */
4942 DPUTS("allocating new root leaf page");
4943 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4947 mdb_cursor_push(mc, np);
4948 mc->mc_db->md_root = np->mp_pgno;
4949 mc->mc_db->md_depth++;
4950 *mc->mc_dbflag |= DB_DIRTY;
4951 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4953 np->mp_flags |= P_LEAF2;
4954 mc->mc_flags |= C_INITIALIZED;
4960 if (flags & MDB_APPEND) {
4962 rc = mdb_cursor_last(mc, &k2, &d2);
4964 rc = mc->mc_dbx->md_cmp(key, &k2);
4967 mc->mc_ki[mc->mc_top]++;
4969 /* new key is <= last key */
4974 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4976 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4977 DPRINTF("duplicate key [%s]", DKEY(key));
4979 return MDB_KEYEXIST;
4981 if (rc && rc != MDB_NOTFOUND)
4985 /* Cursor is positioned, now make sure all pages are writable */
4986 rc2 = mdb_cursor_touch(mc);
4991 /* The key already exists */
4992 if (rc == MDB_SUCCESS) {
4993 /* there's only a key anyway, so this is a no-op */
4994 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4995 unsigned int ksize = mc->mc_db->md_pad;
4996 if (key->mv_size != ksize)
4998 if (flags == MDB_CURRENT) {
4999 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5000 memcpy(ptr, key->mv_data, ksize);
5005 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5008 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5009 /* Was a single item before, must convert now */
5011 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5012 /* Just overwrite the current item */
5013 if (flags == MDB_CURRENT)
5016 dkey.mv_size = NODEDSZ(leaf);
5017 dkey.mv_data = NODEDATA(leaf);
5018 #if UINT_MAX < SIZE_MAX
5019 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
5020 #ifdef MISALIGNED_OK
5021 mc->mc_dbx->md_dcmp = mdb_cmp_long;
5023 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
5026 /* if data matches, ignore it */
5027 if (!mc->mc_dbx->md_dcmp(data, &dkey))
5028 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
5030 /* create a fake page for the dup items */
5031 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
5032 dkey.mv_data = dbuf;
5033 fp = (MDB_page *)&pbuf;
5034 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5035 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
5036 fp->mp_lower = PAGEHDRSZ;
5037 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
5038 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5039 fp->mp_flags |= P_LEAF2;
5040 fp->mp_pad = data->mv_size;
5041 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
5043 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
5044 (dkey.mv_size & 1) + (data->mv_size & 1);
5046 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5049 xdata.mv_size = fp->mp_upper;
5054 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5055 /* See if we need to convert from fake page to subDB */
5057 unsigned int offset;
5060 fp = NODEDATA(leaf);
5061 if (flags == MDB_CURRENT) {
5063 fp->mp_flags |= P_DIRTY;
5064 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5065 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
5069 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5070 offset = fp->mp_pad;
5071 if (SIZELEFT(fp) >= offset)
5073 offset *= 4; /* space for 4 more */
5075 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
5077 offset += offset & 1;
5078 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
5079 offset >= mc->mc_txn->mt_env->me_nodemax) {
5080 /* yes, convert it */
5082 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5083 dummy.md_pad = fp->mp_pad;
5084 dummy.md_flags = MDB_DUPFIXED;
5085 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5086 dummy.md_flags |= MDB_INTEGERKEY;
5089 dummy.md_branch_pages = 0;
5090 dummy.md_leaf_pages = 1;
5091 dummy.md_overflow_pages = 0;
5092 dummy.md_entries = NUMKEYS(fp);
5094 xdata.mv_size = sizeof(MDB_db);
5095 xdata.mv_data = &dummy;
5096 if ((rc = mdb_page_alloc(mc, 1, &mp)))
5098 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
5099 flags |= F_DUPDATA|F_SUBDATA;
5100 dummy.md_root = mp->mp_pgno;
5102 /* no, just grow it */
5104 xdata.mv_size = NODEDSZ(leaf) + offset;
5105 xdata.mv_data = &pbuf;
5106 mp = (MDB_page *)&pbuf;
5107 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5110 mp->mp_flags = fp->mp_flags | P_DIRTY;
5111 mp->mp_pad = fp->mp_pad;
5112 mp->mp_lower = fp->mp_lower;
5113 mp->mp_upper = fp->mp_upper + offset;
5115 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
5117 nsize = NODEDSZ(leaf) - fp->mp_upper;
5118 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
5119 for (i=0; i<NUMKEYS(fp); i++)
5120 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
5122 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5126 /* data is on sub-DB, just store it */
5127 flags |= F_DUPDATA|F_SUBDATA;
5131 /* overflow page overwrites need special handling */
5132 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5135 int ovpages, dpages;
5137 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5138 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5139 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5140 mdb_page_get(mc->mc_txn, pg, &omp);
5141 /* Is the ov page writable and large enough? */
5142 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
5143 /* yes, overwrite it. Note in this case we don't
5144 * bother to try shrinking the node if the new data
5145 * is smaller than the overflow threshold.
5147 if (F_ISSET(flags, MDB_RESERVE))
5148 data->mv_data = METADATA(omp);
5150 memcpy(METADATA(omp), data->mv_data, data->mv_size);
5153 /* no, free ovpages */
5155 mc->mc_db->md_overflow_pages -= ovpages;
5156 for (i=0; i<ovpages; i++) {
5157 DPRINTF("freed ov page %zu", pg);
5158 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5162 } else if (NODEDSZ(leaf) == data->mv_size) {
5163 /* same size, just replace it. Note that we could
5164 * also reuse this node if the new data is smaller,
5165 * but instead we opt to shrink the node in that case.
5167 if (F_ISSET(flags, MDB_RESERVE))
5168 data->mv_data = NODEDATA(leaf);
5169 else if (data->mv_size)
5170 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
5172 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
5175 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5176 mc->mc_db->md_entries--;
5178 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5185 nflags = flags & NODE_ADD_FLAGS;
5186 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5187 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5188 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5189 nflags &= ~MDB_APPEND;
5191 nflags |= MDB_SPLIT_REPLACE;
5192 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5194 /* There is room already in this leaf page. */
5195 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5196 if (rc == 0 && !do_sub && insert) {
5197 /* Adjust other cursors pointing to mp */
5198 MDB_cursor *m2, *m3;
5199 MDB_dbi dbi = mc->mc_dbi;
5200 unsigned i = mc->mc_top;
5201 MDB_page *mp = mc->mc_pg[i];
5203 if (mc->mc_flags & C_SUB)
5206 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5207 if (mc->mc_flags & C_SUB)
5208 m3 = &m2->mc_xcursor->mx_cursor;
5211 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5212 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5219 if (rc != MDB_SUCCESS)
5220 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5222 /* Now store the actual data in the child DB. Note that we're
5223 * storing the user data in the keys field, so there are strict
5224 * size limits on dupdata. The actual data fields of the child
5225 * DB are all zero size.
5232 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5233 if (flags & MDB_CURRENT) {
5234 xflags = MDB_CURRENT;
5236 mdb_xcursor_init1(mc, leaf);
5237 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5239 /* converted, write the original data first */
5241 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5245 /* Adjust other cursors pointing to mp */
5247 unsigned i = mc->mc_top;
5248 MDB_page *mp = mc->mc_pg[i];
5250 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5251 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5252 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5253 mdb_xcursor_init1(m2, leaf);
5257 /* we've done our job */
5260 if (flags & MDB_APPENDDUP)
5261 xflags |= MDB_APPEND;
5262 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5263 if (flags & F_SUBDATA) {
5264 void *db = NODEDATA(leaf);
5265 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5268 /* sub-writes might have failed so check rc again.
5269 * Don't increment count if we just replaced an existing item.
5271 if (!rc && !(flags & MDB_CURRENT))
5272 mc->mc_db->md_entries++;
5273 if (flags & MDB_MULTIPLE) {
5275 if (mcount < data[1].mv_size) {
5276 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5277 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5283 /* If we succeeded and the key didn't exist before, make sure
5284 * the cursor is marked valid.
5287 mc->mc_flags |= C_INITIALIZED;
5292 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5297 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5300 if (!(mc->mc_flags & C_INITIALIZED))
5303 rc = mdb_cursor_touch(mc);
5307 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5309 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5310 if (flags != MDB_NODUPDATA) {
5311 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5312 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5314 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5315 /* If sub-DB still has entries, we're done */
5316 if (mc->mc_xcursor->mx_db.md_entries) {
5317 if (leaf->mn_flags & F_SUBDATA) {
5318 /* update subDB info */
5319 void *db = NODEDATA(leaf);
5320 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5322 /* shrink fake page */
5323 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5325 mc->mc_db->md_entries--;
5328 /* otherwise fall thru and delete the sub-DB */
5331 if (leaf->mn_flags & F_SUBDATA) {
5332 /* add all the child DB's pages to the free list */
5333 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5334 if (rc == MDB_SUCCESS) {
5335 mc->mc_db->md_entries -=
5336 mc->mc_xcursor->mx_db.md_entries;
5341 return mdb_cursor_del0(mc, leaf);
5344 /** Allocate and initialize new pages for a database.
5345 * @param[in] mc a cursor on the database being added to.
5346 * @param[in] flags flags defining what type of page is being allocated.
5347 * @param[in] num the number of pages to allocate. This is usually 1,
5348 * unless allocating overflow pages for a large record.
5349 * @param[out] mp Address of a page, or NULL on failure.
5350 * @return 0 on success, non-zero on failure.
5353 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5358 if ((rc = mdb_page_alloc(mc, num, &np)))
5360 DPRINTF("allocated new mpage %zu, page size %u",
5361 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5362 np->mp_flags = flags | P_DIRTY;
5363 np->mp_lower = PAGEHDRSZ;
5364 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5367 mc->mc_db->md_branch_pages++;
5368 else if (IS_LEAF(np))
5369 mc->mc_db->md_leaf_pages++;
5370 else if (IS_OVERFLOW(np)) {
5371 mc->mc_db->md_overflow_pages += num;
5379 /** Calculate the size of a leaf node.
5380 * The size depends on the environment's page size; if a data item
5381 * is too large it will be put onto an overflow page and the node
5382 * size will only include the key and not the data. Sizes are always
5383 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5384 * of the #MDB_node headers.
5385 * @param[in] env The environment handle.
5386 * @param[in] key The key for the node.
5387 * @param[in] data The data for the node.
5388 * @return The number of bytes needed to store the node.
5391 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5395 sz = LEAFSIZE(key, data);
5396 if (sz >= env->me_nodemax) {
5397 /* put on overflow page */
5398 sz -= data->mv_size - sizeof(pgno_t);
5402 return sz + sizeof(indx_t);
5405 /** Calculate the size of a branch node.
5406 * The size should depend on the environment's page size but since
5407 * we currently don't support spilling large keys onto overflow
5408 * pages, it's simply the size of the #MDB_node header plus the
5409 * size of the key. Sizes are always rounded up to an even number
5410 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5411 * @param[in] env The environment handle.
5412 * @param[in] key The key for the node.
5413 * @return The number of bytes needed to store the node.
5416 mdb_branch_size(MDB_env *env, MDB_val *key)
5421 if (sz >= env->me_nodemax) {
5422 /* put on overflow page */
5423 /* not implemented */
5424 /* sz -= key->size - sizeof(pgno_t); */
5427 return sz + sizeof(indx_t);
5430 /** Add a node to the page pointed to by the cursor.
5431 * @param[in] mc The cursor for this operation.
5432 * @param[in] indx The index on the page where the new node should be added.
5433 * @param[in] key The key for the new node.
5434 * @param[in] data The data for the new node, if any.
5435 * @param[in] pgno The page number, if adding a branch node.
5436 * @param[in] flags Flags for the node.
5437 * @return 0 on success, non-zero on failure. Possible errors are:
5439 * <li>ENOMEM - failed to allocate overflow pages for the node.
5440 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5441 * should never happen since all callers already calculate the
5442 * page's free space before calling this function.
5446 mdb_node_add(MDB_cursor *mc, indx_t indx,
5447 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5450 size_t node_size = NODESIZE;
5453 MDB_page *mp = mc->mc_pg[mc->mc_top];
5454 MDB_page *ofp = NULL; /* overflow page */
5457 assert(mp->mp_upper >= mp->mp_lower);
5459 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5460 IS_LEAF(mp) ? "leaf" : "branch",
5461 IS_SUBP(mp) ? "sub-" : "",
5462 mp->mp_pgno, indx, data ? data->mv_size : 0,
5463 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5466 /* Move higher keys up one slot. */
5467 int ksize = mc->mc_db->md_pad, dif;
5468 char *ptr = LEAF2KEY(mp, indx, ksize);
5469 dif = NUMKEYS(mp) - indx;
5471 memmove(ptr+ksize, ptr, dif*ksize);
5472 /* insert new key */
5473 memcpy(ptr, key->mv_data, ksize);
5475 /* Just using these for counting */
5476 mp->mp_lower += sizeof(indx_t);
5477 mp->mp_upper -= ksize - sizeof(indx_t);
5482 node_size += key->mv_size;
5486 if (F_ISSET(flags, F_BIGDATA)) {
5487 /* Data already on overflow page. */
5488 node_size += sizeof(pgno_t);
5489 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_nodemax) {
5490 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5492 /* Put data on overflow page. */
5493 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5494 data->mv_size, node_size+data->mv_size);
5495 node_size += sizeof(pgno_t);
5496 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5498 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5501 node_size += data->mv_size;
5504 node_size += node_size & 1;
5506 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5507 DPRINTF("not enough room in page %zu, got %u ptrs",
5508 mp->mp_pgno, NUMKEYS(mp));
5509 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5510 mp->mp_upper - mp->mp_lower);
5511 DPRINTF("node size = %zu", node_size);
5512 return MDB_PAGE_FULL;
5515 /* Move higher pointers up one slot. */
5516 for (i = NUMKEYS(mp); i > indx; i--)
5517 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5519 /* Adjust free space offsets. */
5520 ofs = mp->mp_upper - node_size;
5521 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5522 mp->mp_ptrs[indx] = ofs;
5524 mp->mp_lower += sizeof(indx_t);
5526 /* Write the node data. */
5527 node = NODEPTR(mp, indx);
5528 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5529 node->mn_flags = flags;
5531 SETDSZ(node,data->mv_size);
5536 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5541 if (F_ISSET(flags, F_BIGDATA))
5542 memcpy(node->mn_data + key->mv_size, data->mv_data,
5544 else if (F_ISSET(flags, MDB_RESERVE))
5545 data->mv_data = node->mn_data + key->mv_size;
5547 memcpy(node->mn_data + key->mv_size, data->mv_data,
5550 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5552 if (F_ISSET(flags, MDB_RESERVE))
5553 data->mv_data = METADATA(ofp);
5555 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5562 /** Delete the specified node from a page.
5563 * @param[in] mp The page to operate on.
5564 * @param[in] indx The index of the node to delete.
5565 * @param[in] ksize The size of a node. Only used if the page is
5566 * part of a #MDB_DUPFIXED database.
5569 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5572 indx_t i, j, numkeys, ptr;
5579 COPY_PGNO(pgno, mp->mp_pgno);
5580 DPRINTF("delete node %u on %s page %zu", indx,
5581 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5584 assert(indx < NUMKEYS(mp));
5587 int x = NUMKEYS(mp) - 1 - indx;
5588 base = LEAF2KEY(mp, indx, ksize);
5590 memmove(base, base + ksize, x * ksize);
5591 mp->mp_lower -= sizeof(indx_t);
5592 mp->mp_upper += ksize - sizeof(indx_t);
5596 node = NODEPTR(mp, indx);
5597 sz = NODESIZE + node->mn_ksize;
5599 if (F_ISSET(node->mn_flags, F_BIGDATA))
5600 sz += sizeof(pgno_t);
5602 sz += NODEDSZ(node);
5606 ptr = mp->mp_ptrs[indx];
5607 numkeys = NUMKEYS(mp);
5608 for (i = j = 0; i < numkeys; i++) {
5610 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5611 if (mp->mp_ptrs[i] < ptr)
5612 mp->mp_ptrs[j] += sz;
5617 base = (char *)mp + mp->mp_upper;
5618 memmove(base + sz, base, ptr - mp->mp_upper);
5620 mp->mp_lower -= sizeof(indx_t);
5624 /** Compact the main page after deleting a node on a subpage.
5625 * @param[in] mp The main page to operate on.
5626 * @param[in] indx The index of the subpage on the main page.
5629 mdb_node_shrink(MDB_page *mp, indx_t indx)
5636 indx_t i, numkeys, ptr;
5638 node = NODEPTR(mp, indx);
5639 sp = (MDB_page *)NODEDATA(node);
5640 osize = NODEDSZ(node);
5642 delta = sp->mp_upper - sp->mp_lower;
5643 SETDSZ(node, osize - delta);
5644 xp = (MDB_page *)((char *)sp + delta);
5646 /* shift subpage upward */
5648 nsize = NUMKEYS(sp) * sp->mp_pad;
5649 memmove(METADATA(xp), METADATA(sp), nsize);
5652 nsize = osize - sp->mp_upper;
5653 numkeys = NUMKEYS(sp);
5654 for (i=numkeys-1; i>=0; i--)
5655 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5657 xp->mp_upper = sp->mp_lower;
5658 xp->mp_lower = sp->mp_lower;
5659 xp->mp_flags = sp->mp_flags;
5660 xp->mp_pad = sp->mp_pad;
5661 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5663 /* shift lower nodes upward */
5664 ptr = mp->mp_ptrs[indx];
5665 numkeys = NUMKEYS(mp);
5666 for (i = 0; i < numkeys; i++) {
5667 if (mp->mp_ptrs[i] <= ptr)
5668 mp->mp_ptrs[i] += delta;
5671 base = (char *)mp + mp->mp_upper;
5672 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5673 mp->mp_upper += delta;
5676 /** Initial setup of a sorted-dups cursor.
5677 * Sorted duplicates are implemented as a sub-database for the given key.
5678 * The duplicate data items are actually keys of the sub-database.
5679 * Operations on the duplicate data items are performed using a sub-cursor
5680 * initialized when the sub-database is first accessed. This function does
5681 * the preliminary setup of the sub-cursor, filling in the fields that
5682 * depend only on the parent DB.
5683 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5686 mdb_xcursor_init0(MDB_cursor *mc)
5688 MDB_xcursor *mx = mc->mc_xcursor;
5690 mx->mx_cursor.mc_xcursor = NULL;
5691 mx->mx_cursor.mc_txn = mc->mc_txn;
5692 mx->mx_cursor.mc_db = &mx->mx_db;
5693 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5694 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5695 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5696 mx->mx_cursor.mc_snum = 0;
5697 mx->mx_cursor.mc_top = 0;
5698 mx->mx_cursor.mc_flags = C_SUB;
5699 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5700 mx->mx_dbx.md_dcmp = NULL;
5701 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5704 /** Final setup of a sorted-dups cursor.
5705 * Sets up the fields that depend on the data from the main cursor.
5706 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5707 * @param[in] node The data containing the #MDB_db record for the
5708 * sorted-dup database.
5711 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5713 MDB_xcursor *mx = mc->mc_xcursor;
5715 if (node->mn_flags & F_SUBDATA) {
5716 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5717 mx->mx_cursor.mc_pg[0] = 0;
5718 mx->mx_cursor.mc_snum = 0;
5719 mx->mx_cursor.mc_flags = C_SUB;
5721 MDB_page *fp = NODEDATA(node);
5722 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5723 mx->mx_db.md_flags = 0;
5724 mx->mx_db.md_depth = 1;
5725 mx->mx_db.md_branch_pages = 0;
5726 mx->mx_db.md_leaf_pages = 1;
5727 mx->mx_db.md_overflow_pages = 0;
5728 mx->mx_db.md_entries = NUMKEYS(fp);
5729 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5730 mx->mx_cursor.mc_snum = 1;
5731 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5732 mx->mx_cursor.mc_top = 0;
5733 mx->mx_cursor.mc_pg[0] = fp;
5734 mx->mx_cursor.mc_ki[0] = 0;
5735 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5736 mx->mx_db.md_flags = MDB_DUPFIXED;
5737 mx->mx_db.md_pad = fp->mp_pad;
5738 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5739 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5742 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5744 mx->mx_dbflag = DB_VALID | (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY) ?
5746 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5747 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5748 #if UINT_MAX < SIZE_MAX
5749 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5750 #ifdef MISALIGNED_OK
5751 mx->mx_dbx.md_cmp = mdb_cmp_long;
5753 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5758 /** Initialize a cursor for a given transaction and database. */
5760 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5765 mc->mc_db = &txn->mt_dbs[dbi];
5766 mc->mc_dbx = &txn->mt_dbxs[dbi];
5767 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5772 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5774 mc->mc_xcursor = mx;
5775 mdb_xcursor_init0(mc);
5777 mc->mc_xcursor = NULL;
5779 if (*mc->mc_dbflag & DB_STALE) {
5780 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5785 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5788 MDB_xcursor *mx = NULL;
5789 size_t size = sizeof(MDB_cursor);
5791 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
5794 /* Allow read access to the freelist */
5795 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5798 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5799 size += sizeof(MDB_xcursor);
5801 if ((mc = malloc(size)) != NULL) {
5802 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5803 mx = (MDB_xcursor *)(mc + 1);
5805 mdb_cursor_init(mc, txn, dbi, mx);
5806 if (txn->mt_cursors) {
5807 mc->mc_next = txn->mt_cursors[dbi];
5808 txn->mt_cursors[dbi] = mc;
5810 mc->mc_flags |= C_ALLOCD;
5821 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5823 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5826 if (txn->mt_cursors)
5829 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5833 /* Return the count of duplicate data items for the current key */
5835 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5839 if (mc == NULL || countp == NULL)
5842 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5845 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5846 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5849 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5852 *countp = mc->mc_xcursor->mx_db.md_entries;
5858 mdb_cursor_close(MDB_cursor *mc)
5861 /* remove from txn, if tracked */
5862 if (mc->mc_txn->mt_cursors) {
5863 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5864 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5866 *prev = mc->mc_next;
5868 if (mc->mc_flags & C_ALLOCD)
5874 mdb_cursor_txn(MDB_cursor *mc)
5876 if (!mc) return NULL;
5881 mdb_cursor_dbi(MDB_cursor *mc)
5887 /** Replace the key for a node with a new key.
5888 * @param[in] mp The page containing the node to operate on.
5889 * @param[in] indx The index of the node to operate on.
5890 * @param[in] key The new key to use.
5891 * @return 0 on success, non-zero on failure.
5894 mdb_update_key(MDB_cursor *mc, MDB_val *key)
5901 indx_t ptr, i, numkeys, indx;
5904 indx = mc->mc_ki[mc->mc_top];
5905 mp = mc->mc_pg[mc->mc_top];
5906 node = NODEPTR(mp, indx);
5907 ptr = mp->mp_ptrs[indx];
5911 char kbuf2[(MDB_MAXKEYSIZE*2+1)];
5912 k2.mv_data = NODEKEY(node);
5913 k2.mv_size = node->mn_ksize;
5914 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5916 mdb_dkey(&k2, kbuf2),
5922 delta0 = delta = key->mv_size - node->mn_ksize;
5924 /* Must be 2-byte aligned. If new key is
5925 * shorter by 1, the shift will be skipped.
5927 delta += (delta & 1);
5929 if (delta > 0 && SIZELEFT(mp) < delta) {
5931 /* not enough space left, do a delete and split */
5932 DPRINTF("Not enough room, delta = %d, splitting...", delta);
5933 pgno = NODEPGNO(node);
5934 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
5935 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
5938 numkeys = NUMKEYS(mp);
5939 for (i = 0; i < numkeys; i++) {
5940 if (mp->mp_ptrs[i] <= ptr)
5941 mp->mp_ptrs[i] -= delta;
5944 base = (char *)mp + mp->mp_upper;
5945 len = ptr - mp->mp_upper + NODESIZE;
5946 memmove(base - delta, base, len);
5947 mp->mp_upper -= delta;
5949 node = NODEPTR(mp, indx);
5952 /* But even if no shift was needed, update ksize */
5954 node->mn_ksize = key->mv_size;
5957 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5963 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
5965 /** Move a node from csrc to cdst.
5968 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5975 unsigned short flags;
5979 /* Mark src and dst as dirty. */
5980 if ((rc = mdb_page_touch(csrc)) ||
5981 (rc = mdb_page_touch(cdst)))
5984 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5985 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5986 key.mv_size = csrc->mc_db->md_pad;
5987 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5989 data.mv_data = NULL;
5993 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5994 assert(!((long)srcnode&1));
5995 srcpg = NODEPGNO(srcnode);
5996 flags = srcnode->mn_flags;
5997 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5998 unsigned int snum = csrc->mc_snum;
6000 /* must find the lowest key below src */
6001 mdb_page_search_root(csrc, NULL, 0);
6002 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6003 key.mv_size = csrc->mc_db->md_pad;
6004 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6006 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6007 key.mv_size = NODEKSZ(s2);
6008 key.mv_data = NODEKEY(s2);
6010 csrc->mc_snum = snum--;
6011 csrc->mc_top = snum;
6013 key.mv_size = NODEKSZ(srcnode);
6014 key.mv_data = NODEKEY(srcnode);
6016 data.mv_size = NODEDSZ(srcnode);
6017 data.mv_data = NODEDATA(srcnode);
6019 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
6020 unsigned int snum = cdst->mc_snum;
6023 /* must find the lowest key below dst */
6024 mdb_page_search_root(cdst, NULL, 0);
6025 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
6026 bkey.mv_size = cdst->mc_db->md_pad;
6027 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
6029 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6030 bkey.mv_size = NODEKSZ(s2);
6031 bkey.mv_data = NODEKEY(s2);
6033 cdst->mc_snum = snum--;
6034 cdst->mc_top = snum;
6035 mdb_cursor_copy(cdst, &mn);
6037 rc = mdb_update_key(&mn, &bkey);
6042 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
6043 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
6044 csrc->mc_ki[csrc->mc_top],
6046 csrc->mc_pg[csrc->mc_top]->mp_pgno,
6047 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
6049 /* Add the node to the destination page.
6051 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
6052 if (rc != MDB_SUCCESS)
6055 /* Delete the node from the source page.
6057 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
6060 /* Adjust other cursors pointing to mp */
6061 MDB_cursor *m2, *m3;
6062 MDB_dbi dbi = csrc->mc_dbi;
6063 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
6065 if (csrc->mc_flags & C_SUB)
6068 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6069 if (m2 == csrc) continue;
6070 if (csrc->mc_flags & C_SUB)
6071 m3 = &m2->mc_xcursor->mx_cursor;
6074 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
6075 csrc->mc_ki[csrc->mc_top]) {
6076 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
6077 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
6082 /* Update the parent separators.
6084 if (csrc->mc_ki[csrc->mc_top] == 0) {
6085 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
6086 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6087 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6089 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6090 key.mv_size = NODEKSZ(srcnode);
6091 key.mv_data = NODEKEY(srcnode);
6093 DPRINTF("update separator for source page %zu to [%s]",
6094 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
6095 mdb_cursor_copy(csrc, &mn);
6098 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6101 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6103 indx_t ix = csrc->mc_ki[csrc->mc_top];
6104 nullkey.mv_size = 0;
6105 csrc->mc_ki[csrc->mc_top] = 0;
6106 rc = mdb_update_key(csrc, &nullkey);
6107 csrc->mc_ki[csrc->mc_top] = ix;
6108 assert(rc == MDB_SUCCESS);
6112 if (cdst->mc_ki[cdst->mc_top] == 0) {
6113 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
6114 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6115 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
6117 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
6118 key.mv_size = NODEKSZ(srcnode);
6119 key.mv_data = NODEKEY(srcnode);
6121 DPRINTF("update separator for destination page %zu to [%s]",
6122 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
6123 mdb_cursor_copy(cdst, &mn);
6126 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
6129 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
6131 indx_t ix = cdst->mc_ki[cdst->mc_top];
6132 nullkey.mv_size = 0;
6133 cdst->mc_ki[cdst->mc_top] = 0;
6134 rc = mdb_update_key(cdst, &nullkey);
6135 cdst->mc_ki[cdst->mc_top] = ix;
6136 assert(rc == MDB_SUCCESS);
6143 /** Merge one page into another.
6144 * The nodes from the page pointed to by \b csrc will
6145 * be copied to the page pointed to by \b cdst and then
6146 * the \b csrc page will be freed.
6147 * @param[in] csrc Cursor pointing to the source page.
6148 * @param[in] cdst Cursor pointing to the destination page.
6151 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
6159 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
6160 cdst->mc_pg[cdst->mc_top]->mp_pgno);
6162 assert(csrc->mc_snum > 1); /* can't merge root page */
6163 assert(cdst->mc_snum > 1);
6165 /* Mark dst as dirty. */
6166 if ((rc = mdb_page_touch(cdst)))
6169 /* Move all nodes from src to dst.
6171 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
6172 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6173 key.mv_size = csrc->mc_db->md_pad;
6174 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
6175 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6176 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
6177 if (rc != MDB_SUCCESS)
6179 key.mv_data = (char *)key.mv_data + key.mv_size;
6182 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
6183 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
6184 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
6185 unsigned int snum = csrc->mc_snum;
6187 /* must find the lowest key below src */
6188 mdb_page_search_root(csrc, NULL, 0);
6189 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
6190 key.mv_size = csrc->mc_db->md_pad;
6191 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
6193 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
6194 key.mv_size = NODEKSZ(s2);
6195 key.mv_data = NODEKEY(s2);
6197 csrc->mc_snum = snum--;
6198 csrc->mc_top = snum;
6200 key.mv_size = srcnode->mn_ksize;
6201 key.mv_data = NODEKEY(srcnode);
6204 data.mv_size = NODEDSZ(srcnode);
6205 data.mv_data = NODEDATA(srcnode);
6206 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
6207 if (rc != MDB_SUCCESS)
6212 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
6213 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);
6215 /* Unlink the src page from parent and add to free list.
6217 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6218 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6221 rc = mdb_update_key(csrc, &key);
6227 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6228 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6229 csrc->mc_db->md_leaf_pages--;
6231 csrc->mc_db->md_branch_pages--;
6233 /* Adjust other cursors pointing to mp */
6234 MDB_cursor *m2, *m3;
6235 MDB_dbi dbi = csrc->mc_dbi;
6236 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6238 if (csrc->mc_flags & C_SUB)
6241 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6242 if (csrc->mc_flags & C_SUB)
6243 m3 = &m2->mc_xcursor->mx_cursor;
6246 if (m3 == csrc) continue;
6247 if (m3->mc_snum < csrc->mc_snum) continue;
6248 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6249 m3->mc_pg[csrc->mc_top] = mp;
6250 m3->mc_ki[csrc->mc_top] += nkeys;
6254 mdb_cursor_pop(csrc);
6256 return mdb_rebalance(csrc);
6259 /** Copy the contents of a cursor.
6260 * @param[in] csrc The cursor to copy from.
6261 * @param[out] cdst The cursor to copy to.
6264 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6268 cdst->mc_txn = csrc->mc_txn;
6269 cdst->mc_dbi = csrc->mc_dbi;
6270 cdst->mc_db = csrc->mc_db;
6271 cdst->mc_dbx = csrc->mc_dbx;
6272 cdst->mc_snum = csrc->mc_snum;
6273 cdst->mc_top = csrc->mc_top;
6274 cdst->mc_flags = csrc->mc_flags;
6276 for (i=0; i<csrc->mc_snum; i++) {
6277 cdst->mc_pg[i] = csrc->mc_pg[i];
6278 cdst->mc_ki[i] = csrc->mc_ki[i];
6282 /** Rebalance the tree after a delete operation.
6283 * @param[in] mc Cursor pointing to the page where rebalancing
6285 * @return 0 on success, non-zero on failure.
6288 mdb_rebalance(MDB_cursor *mc)
6292 unsigned int ptop, minkeys;
6295 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
6299 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6300 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6301 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6302 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6306 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
6307 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
6310 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6311 DPRINTF("no need to rebalance page %zu, above fill threshold",
6317 if (mc->mc_snum < 2) {
6318 MDB_page *mp = mc->mc_pg[0];
6320 DPUTS("Can't rebalance a subpage, ignoring");
6323 if (NUMKEYS(mp) == 0) {
6324 DPUTS("tree is completely empty");
6325 mc->mc_db->md_root = P_INVALID;
6326 mc->mc_db->md_depth = 0;
6327 mc->mc_db->md_leaf_pages = 0;
6328 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6332 /* Adjust other cursors pointing to mp */
6333 MDB_cursor *m2, *m3;
6334 MDB_dbi dbi = mc->mc_dbi;
6336 if (mc->mc_flags & C_SUB)
6339 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6340 if (m2 == mc) continue;
6341 if (mc->mc_flags & C_SUB)
6342 m3 = &m2->mc_xcursor->mx_cursor;
6345 if (m3->mc_snum < mc->mc_snum) continue;
6346 if (m3->mc_pg[0] == mp) {
6352 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6353 DPUTS("collapsing root page!");
6354 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6355 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6356 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6359 mc->mc_db->md_depth--;
6360 mc->mc_db->md_branch_pages--;
6362 /* Adjust other cursors pointing to mp */
6363 MDB_cursor *m2, *m3;
6364 MDB_dbi dbi = mc->mc_dbi;
6366 if (mc->mc_flags & C_SUB)
6369 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6370 if (m2 == mc) continue;
6371 if (mc->mc_flags & C_SUB)
6372 m3 = &m2->mc_xcursor->mx_cursor;
6375 if (m3->mc_snum < mc->mc_snum) continue;
6376 if (m3->mc_pg[0] == mp) {
6377 m3->mc_pg[0] = mc->mc_pg[0];
6384 DPUTS("root page doesn't need rebalancing");
6388 /* The parent (branch page) must have at least 2 pointers,
6389 * otherwise the tree is invalid.
6391 ptop = mc->mc_top-1;
6392 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6394 /* Leaf page fill factor is below the threshold.
6395 * Try to move keys from left or right neighbor, or
6396 * merge with a neighbor page.
6401 mdb_cursor_copy(mc, &mn);
6402 mn.mc_xcursor = NULL;
6404 if (mc->mc_ki[ptop] == 0) {
6405 /* We're the leftmost leaf in our parent.
6407 DPUTS("reading right neighbor");
6409 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6410 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6412 mn.mc_ki[mn.mc_top] = 0;
6413 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6415 /* There is at least one neighbor to the left.
6417 DPUTS("reading left neighbor");
6419 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6420 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6422 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6423 mc->mc_ki[mc->mc_top] = 0;
6426 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6427 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);
6429 /* If the neighbor page is above threshold and has enough keys,
6430 * move one key from it. Otherwise we should try to merge them.
6431 * (A branch page must never have less than 2 keys.)
6433 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
6434 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys)
6435 return mdb_node_move(&mn, mc);
6437 if (mc->mc_ki[ptop] == 0)
6438 rc = mdb_page_merge(&mn, mc);
6440 rc = mdb_page_merge(mc, &mn);
6441 mc->mc_flags &= ~C_INITIALIZED;
6446 /** Complete a delete operation started by #mdb_cursor_del(). */
6448 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6452 /* add overflow pages to free list */
6453 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6457 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6458 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6459 mc->mc_db->md_overflow_pages -= ovpages;
6460 for (i=0; i<ovpages; i++) {
6461 DPRINTF("freed ov page %zu", pg);
6462 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6466 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6467 mc->mc_db->md_entries--;
6468 rc = mdb_rebalance(mc);
6469 if (rc != MDB_SUCCESS)
6470 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6471 /* if mc points past last node in page, invalidate */
6472 else if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6473 mc->mc_flags &= ~C_INITIALIZED;
6479 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6480 MDB_val *key, MDB_val *data)
6485 MDB_val rdata, *xdata;
6489 assert(key != NULL);
6491 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6493 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6496 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6500 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6504 mdb_cursor_init(&mc, txn, dbi, &mx);
6515 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6517 /* let mdb_page_split know about this cursor if needed:
6518 * delete will trigger a rebalance; if it needs to move
6519 * a node from one page to another, it will have to
6520 * update the parent's separator key(s). If the new sepkey
6521 * is larger than the current one, the parent page may
6522 * run out of space, triggering a split. We need this
6523 * cursor to be consistent until the end of the rebalance.
6525 mc.mc_next = txn->mt_cursors[dbi];
6526 txn->mt_cursors[dbi] = &mc;
6527 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6528 txn->mt_cursors[dbi] = mc.mc_next;
6533 /** Split a page and insert a new node.
6534 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6535 * The cursor will be updated to point to the actual page and index where
6536 * the node got inserted after the split.
6537 * @param[in] newkey The key for the newly inserted node.
6538 * @param[in] newdata The data for the newly inserted node.
6539 * @param[in] newpgno The page number, if the new node is a branch node.
6540 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6541 * @return 0 on success, non-zero on failure.
6544 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6545 unsigned int nflags)
6548 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6551 unsigned int i, j, split_indx, nkeys, pmax;
6553 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6555 MDB_page *mp, *rp, *pp;
6560 mp = mc->mc_pg[mc->mc_top];
6561 newindx = mc->mc_ki[mc->mc_top];
6563 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6564 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6565 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6567 /* Create a right sibling. */
6568 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6570 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6572 if (mc->mc_snum < 2) {
6573 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6575 /* shift current top to make room for new parent */
6576 mc->mc_pg[1] = mc->mc_pg[0];
6577 mc->mc_ki[1] = mc->mc_ki[0];
6580 mc->mc_db->md_root = pp->mp_pgno;
6581 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6582 mc->mc_db->md_depth++;
6585 /* Add left (implicit) pointer. */
6586 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6587 /* undo the pre-push */
6588 mc->mc_pg[0] = mc->mc_pg[1];
6589 mc->mc_ki[0] = mc->mc_ki[1];
6590 mc->mc_db->md_root = mp->mp_pgno;
6591 mc->mc_db->md_depth--;
6598 ptop = mc->mc_top-1;
6599 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6602 mc->mc_flags |= C_SPLITTING;
6603 mdb_cursor_copy(mc, &mn);
6604 mn.mc_pg[mn.mc_top] = rp;
6605 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6607 if (nflags & MDB_APPEND) {
6608 mn.mc_ki[mn.mc_top] = 0;
6610 split_indx = newindx;
6615 nkeys = NUMKEYS(mp);
6616 split_indx = nkeys / 2;
6617 if (newindx < split_indx)
6623 unsigned int lsize, rsize, ksize;
6624 /* Move half of the keys to the right sibling */
6626 x = mc->mc_ki[mc->mc_top] - split_indx;
6627 ksize = mc->mc_db->md_pad;
6628 split = LEAF2KEY(mp, split_indx, ksize);
6629 rsize = (nkeys - split_indx) * ksize;
6630 lsize = (nkeys - split_indx) * sizeof(indx_t);
6631 mp->mp_lower -= lsize;
6632 rp->mp_lower += lsize;
6633 mp->mp_upper += rsize - lsize;
6634 rp->mp_upper -= rsize - lsize;
6635 sepkey.mv_size = ksize;
6636 if (newindx == split_indx) {
6637 sepkey.mv_data = newkey->mv_data;
6639 sepkey.mv_data = split;
6642 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6643 memcpy(rp->mp_ptrs, split, rsize);
6644 sepkey.mv_data = rp->mp_ptrs;
6645 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6646 memcpy(ins, newkey->mv_data, ksize);
6647 mp->mp_lower += sizeof(indx_t);
6648 mp->mp_upper -= ksize - sizeof(indx_t);
6651 memcpy(rp->mp_ptrs, split, x * ksize);
6652 ins = LEAF2KEY(rp, x, ksize);
6653 memcpy(ins, newkey->mv_data, ksize);
6654 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6655 rp->mp_lower += sizeof(indx_t);
6656 rp->mp_upper -= ksize - sizeof(indx_t);
6657 mc->mc_ki[mc->mc_top] = x;
6658 mc->mc_pg[mc->mc_top] = rp;
6663 /* For leaf pages, check the split point based on what
6664 * fits where, since otherwise mdb_node_add can fail.
6666 * This check is only needed when the data items are
6667 * relatively large, such that being off by one will
6668 * make the difference between success or failure.
6670 * It's also relevant if a page happens to be laid out
6671 * such that one half of its nodes are all "small" and
6672 * the other half of its nodes are "large." If the new
6673 * item is also "large" and falls on the half with
6674 * "large" nodes, it also may not fit.
6677 unsigned int psize, nsize;
6678 /* Maximum free space in an empty page */
6679 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6680 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6681 if ((nkeys < 20) || (nsize > pmax/16)) {
6682 if (newindx <= split_indx) {
6685 for (i=0; i<split_indx; i++) {
6686 node = NODEPTR(mp, i);
6687 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6688 if (F_ISSET(node->mn_flags, F_BIGDATA))
6689 psize += sizeof(pgno_t);
6691 psize += NODEDSZ(node);
6695 split_indx = newindx;
6706 for (i=nkeys-1; i>=split_indx; i--) {
6707 node = NODEPTR(mp, i);
6708 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6709 if (F_ISSET(node->mn_flags, F_BIGDATA))
6710 psize += sizeof(pgno_t);
6712 psize += NODEDSZ(node);
6716 split_indx = newindx;
6727 /* First find the separating key between the split pages.
6728 * The case where newindx == split_indx is ambiguous; the
6729 * new item could go to the new page or stay on the original
6730 * page. If newpos == 1 it goes to the new page.
6732 if (newindx == split_indx && newpos) {
6733 sepkey.mv_size = newkey->mv_size;
6734 sepkey.mv_data = newkey->mv_data;
6736 node = NODEPTR(mp, split_indx);
6737 sepkey.mv_size = node->mn_ksize;
6738 sepkey.mv_data = NODEKEY(node);
6742 DPRINTF("separator is [%s]", DKEY(&sepkey));
6744 /* Copy separator key to the parent.
6746 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6750 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6753 if (mn.mc_snum == mc->mc_snum) {
6754 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6755 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6756 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6757 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6762 /* Right page might now have changed parent.
6763 * Check if left page also changed parent.
6765 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6766 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6767 for (i=0; i<ptop; i++) {
6768 mc->mc_pg[i] = mn.mc_pg[i];
6769 mc->mc_ki[i] = mn.mc_ki[i];
6771 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6772 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6776 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6779 mc->mc_flags ^= C_SPLITTING;
6780 if (rc != MDB_SUCCESS) {
6783 if (nflags & MDB_APPEND) {
6784 mc->mc_pg[mc->mc_top] = rp;
6785 mc->mc_ki[mc->mc_top] = 0;
6786 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6789 for (i=0; i<mc->mc_top; i++)
6790 mc->mc_ki[i] = mn.mc_ki[i];
6797 /* Move half of the keys to the right sibling. */
6799 /* grab a page to hold a temporary copy */
6800 copy = mdb_page_malloc(mc);
6804 copy->mp_pgno = mp->mp_pgno;
6805 copy->mp_flags = mp->mp_flags;
6806 copy->mp_lower = PAGEHDRSZ;
6807 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6808 mc->mc_pg[mc->mc_top] = copy;
6809 for (i = j = 0; i <= nkeys; j++) {
6810 if (i == split_indx) {
6811 /* Insert in right sibling. */
6812 /* Reset insert index for right sibling. */
6813 if (i != newindx || (newpos ^ ins_new)) {
6815 mc->mc_pg[mc->mc_top] = rp;
6819 if (i == newindx && !ins_new) {
6820 /* Insert the original entry that caused the split. */
6821 rkey.mv_data = newkey->mv_data;
6822 rkey.mv_size = newkey->mv_size;
6831 /* Update index for the new key. */
6832 mc->mc_ki[mc->mc_top] = j;
6833 } else if (i == nkeys) {
6836 node = NODEPTR(mp, i);
6837 rkey.mv_data = NODEKEY(node);
6838 rkey.mv_size = node->mn_ksize;
6840 xdata.mv_data = NODEDATA(node);
6841 xdata.mv_size = NODEDSZ(node);
6844 pgno = NODEPGNO(node);
6845 flags = node->mn_flags;
6850 if (!IS_LEAF(mp) && j == 0) {
6851 /* First branch index doesn't need key data. */
6855 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6859 nkeys = NUMKEYS(copy);
6860 for (i=0; i<nkeys; i++)
6861 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6862 mp->mp_lower = copy->mp_lower;
6863 mp->mp_upper = copy->mp_upper;
6864 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6865 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6867 /* reset back to original page */
6868 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6869 mc->mc_pg[mc->mc_top] = mp;
6870 if (nflags & MDB_RESERVE) {
6871 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6872 if (!(node->mn_flags & F_BIGDATA))
6873 newdata->mv_data = NODEDATA(node);
6879 /* return tmp page to freelist */
6880 mdb_page_free(mc->mc_txn->mt_env, copy);
6883 /* Adjust other cursors pointing to mp */
6884 MDB_cursor *m2, *m3;
6885 MDB_dbi dbi = mc->mc_dbi;
6886 int fixup = NUMKEYS(mp);
6888 if (mc->mc_flags & C_SUB)
6891 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6892 if (m2 == mc) continue;
6893 if (mc->mc_flags & C_SUB)
6894 m3 = &m2->mc_xcursor->mx_cursor;
6897 if (!(m3->mc_flags & C_INITIALIZED))
6899 if (m3->mc_flags & C_SPLITTING)
6904 for (k=m3->mc_top; k>=0; k--) {
6905 m3->mc_ki[k+1] = m3->mc_ki[k];
6906 m3->mc_pg[k+1] = m3->mc_pg[k];
6908 if (m3->mc_ki[0] >= split_indx) {
6913 m3->mc_pg[0] = mc->mc_pg[0];
6917 if (m3->mc_pg[mc->mc_top] == mp) {
6918 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6919 m3->mc_ki[mc->mc_top]++;
6920 if (m3->mc_ki[mc->mc_top] >= fixup) {
6921 m3->mc_pg[mc->mc_top] = rp;
6922 m3->mc_ki[mc->mc_top] -= fixup;
6923 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6925 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6926 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6935 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6936 MDB_val *key, MDB_val *data, unsigned int flags)
6941 assert(key != NULL);
6942 assert(data != NULL);
6944 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
6947 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6951 if (key->mv_size == 0 || key->mv_size > MDB_MAXKEYSIZE) {
6955 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
6958 mdb_cursor_init(&mc, txn, dbi, &mx);
6959 return mdb_cursor_put(&mc, key, data, flags);
6963 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6965 if ((flag & CHANGEABLE) != flag)
6968 env->me_flags |= flag;
6970 env->me_flags &= ~flag;
6975 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6980 *arg = env->me_flags;
6985 mdb_env_get_path(MDB_env *env, const char **arg)
6990 *arg = env->me_path;
6994 /** Common code for #mdb_stat() and #mdb_env_stat().
6995 * @param[in] env the environment to operate in.
6996 * @param[in] db the #MDB_db record containing the stats to return.
6997 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6998 * @return 0, this function always succeeds.
7001 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
7003 arg->ms_psize = env->me_psize;
7004 arg->ms_depth = db->md_depth;
7005 arg->ms_branch_pages = db->md_branch_pages;
7006 arg->ms_leaf_pages = db->md_leaf_pages;
7007 arg->ms_overflow_pages = db->md_overflow_pages;
7008 arg->ms_entries = db->md_entries;
7013 mdb_env_stat(MDB_env *env, MDB_stat *arg)
7017 if (env == NULL || arg == NULL)
7020 toggle = mdb_env_pick_meta(env);
7022 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
7026 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
7030 if (env == NULL || arg == NULL)
7033 toggle = mdb_env_pick_meta(env);
7034 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
7035 arg->me_mapsize = env->me_mapsize;
7036 arg->me_maxreaders = env->me_maxreaders;
7037 arg->me_numreaders = env->me_numreaders;
7038 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
7039 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
7043 /** Set the default comparison functions for a database.
7044 * Called immediately after a database is opened to set the defaults.
7045 * The user can then override them with #mdb_set_compare() or
7046 * #mdb_set_dupsort().
7047 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
7048 * @param[in] dbi A database handle returned by #mdb_dbi_open()
7051 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
7053 uint16_t f = txn->mt_dbs[dbi].md_flags;
7055 txn->mt_dbxs[dbi].md_cmp =
7056 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
7057 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
7059 txn->mt_dbxs[dbi].md_dcmp =
7060 !(f & MDB_DUPSORT) ? 0 :
7061 ((f & MDB_INTEGERDUP)
7062 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
7063 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
7066 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
7071 int rc, dbflag, exact;
7072 unsigned int unused = 0;
7075 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
7076 mdb_default_cmp(txn, FREE_DBI);
7079 if ((flags & VALID_FLAGS) != flags)
7085 if (flags & PERSISTENT_FLAGS) {
7086 uint16_t f2 = flags & PERSISTENT_FLAGS;
7087 /* make sure flag changes get committed */
7088 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
7089 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
7090 txn->mt_flags |= MDB_TXN_DIRTY;
7093 mdb_default_cmp(txn, MAIN_DBI);
7097 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
7098 mdb_default_cmp(txn, MAIN_DBI);
7101 /* Is the DB already open? */
7103 for (i=2; i<txn->mt_numdbs; i++) {
7104 if (!txn->mt_dbxs[i].md_name.mv_size) {
7105 /* Remember this free slot */
7106 if (!unused) unused = i;
7109 if (len == txn->mt_dbxs[i].md_name.mv_size &&
7110 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
7116 /* If no free slot and max hit, fail */
7117 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
7118 return MDB_DBS_FULL;
7120 /* Find the DB info */
7121 dbflag = DB_NEW|DB_VALID;
7124 key.mv_data = (void *)name;
7125 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
7126 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
7127 if (rc == MDB_SUCCESS) {
7128 /* make sure this is actually a DB */
7129 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
7130 if (!(node->mn_flags & F_SUBDATA))
7132 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
7133 /* Create if requested */
7135 data.mv_size = sizeof(MDB_db);
7136 data.mv_data = &dummy;
7137 memset(&dummy, 0, sizeof(dummy));
7138 dummy.md_root = P_INVALID;
7139 dummy.md_flags = flags & PERSISTENT_FLAGS;
7140 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
7144 /* OK, got info, add to table */
7145 if (rc == MDB_SUCCESS) {
7146 unsigned int slot = unused ? unused : txn->mt_numdbs;
7147 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
7148 txn->mt_dbxs[slot].md_name.mv_size = len;
7149 txn->mt_dbxs[slot].md_rel = NULL;
7150 txn->mt_dbflags[slot] = dbflag;
7151 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
7153 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
7154 mdb_default_cmp(txn, slot);
7163 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
7165 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
7168 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
7171 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
7174 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
7176 ptr = env->me_dbxs[dbi].md_name.mv_data;
7177 env->me_dbxs[dbi].md_name.mv_data = NULL;
7178 env->me_dbxs[dbi].md_name.mv_size = 0;
7182 /** Add all the DB's pages to the free list.
7183 * @param[in] mc Cursor on the DB to free.
7184 * @param[in] subs non-Zero to check for sub-DBs in this DB.
7185 * @return 0 on success, non-zero on failure.
7188 mdb_drop0(MDB_cursor *mc, int subs)
7192 rc = mdb_page_search(mc, NULL, 0);
7193 if (rc == MDB_SUCCESS) {
7198 /* LEAF2 pages have no nodes, cannot have sub-DBs */
7199 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
7202 mdb_cursor_copy(mc, &mx);
7203 while (mc->mc_snum > 0) {
7204 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
7205 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7206 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7207 if (ni->mn_flags & F_BIGDATA) {
7208 int j, ovpages = OVPAGES(NODEDSZ(ni), mc->mc_txn->mt_env->me_psize);
7210 memcpy(&pg, NODEDATA(ni), sizeof(pg));
7211 for (j=0; j<ovpages; j++) {
7212 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7215 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
7216 mdb_xcursor_init1(mc, ni);
7217 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
7223 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
7225 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
7228 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
7233 mc->mc_ki[mc->mc_top] = i;
7234 rc = mdb_cursor_sibling(mc, 1);
7236 /* no more siblings, go back to beginning
7237 * of previous level.
7241 for (i=1; i<mc->mc_snum; i++) {
7243 mc->mc_pg[i] = mx.mc_pg[i];
7248 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
7249 mc->mc_db->md_root);
7254 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7259 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1 || !(txn->mt_dbflags[dbi] & DB_VALID))
7262 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7265 rc = mdb_cursor_open(txn, dbi, &mc);
7269 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7273 /* Can't delete the main DB */
7274 if (del && dbi > MAIN_DBI) {
7275 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7277 txn->mt_dbflags[dbi] = DB_STALE;
7278 mdb_dbi_close(txn->mt_env, dbi);
7281 /* reset the DB record, mark it dirty */
7282 txn->mt_dbflags[dbi] |= DB_DIRTY;
7283 txn->mt_dbs[dbi].md_depth = 0;
7284 txn->mt_dbs[dbi].md_branch_pages = 0;
7285 txn->mt_dbs[dbi].md_leaf_pages = 0;
7286 txn->mt_dbs[dbi].md_overflow_pages = 0;
7287 txn->mt_dbs[dbi].md_entries = 0;
7288 txn->mt_dbs[dbi].md_root = P_INVALID;
7290 txn->mt_flags |= MDB_TXN_DIRTY;
7293 mdb_cursor_close(mc);
7297 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7299 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7302 txn->mt_dbxs[dbi].md_cmp = cmp;
7306 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7308 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7311 txn->mt_dbxs[dbi].md_dcmp = cmp;
7315 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7317 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7320 txn->mt_dbxs[dbi].md_rel = rel;
7324 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7326 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs || !(txn->mt_dbflags[dbi] & DB_VALID))
7329 txn->mt_dbxs[dbi].md_relctx = ctx;