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 2011-2012 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__))
289 # define DPRINTF (void) /* Vararg macros may be unsupported */
291 static int mdb_debug;
292 static txnid_t mdb_debug_start;
294 /** Print a debug message with printf formatting. */
295 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
296 ((void) ((mdb_debug) && \
297 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
299 # define DPRINTF(fmt, ...) ((void) 0)
301 /** Print a debug string.
302 * The string is printed literally, with no format processing.
304 #define DPUTS(arg) DPRINTF("%s", arg)
307 /** A default memory page size.
308 * The actual size is platform-dependent, but we use this for
309 * boot-strapping. We probably should not be using this any more.
310 * The #GET_PAGESIZE() macro is used to get the actual size.
312 * Note that we don't currently support Huge pages. On Linux,
313 * regular data files cannot use Huge pages, and in general
314 * Huge pages aren't actually pageable. We rely on the OS
315 * demand-pager to read our data and page it out when memory
316 * pressure from other processes is high. So until OSs have
317 * actual paging support for Huge pages, they're not viable.
319 #define MDB_PAGESIZE 4096
321 /** The minimum number of keys required in a database page.
322 * Setting this to a larger value will place a smaller bound on the
323 * maximum size of a data item. Data items larger than this size will
324 * be pushed into overflow pages instead of being stored directly in
325 * the B-tree node. This value used to default to 4. With a page size
326 * of 4096 bytes that meant that any item larger than 1024 bytes would
327 * go into an overflow page. That also meant that on average 2-3KB of
328 * each overflow page was wasted space. The value cannot be lower than
329 * 2 because then there would no longer be a tree structure. With this
330 * value, items larger than 2KB will go into overflow pages, and on
331 * average only 1KB will be wasted.
333 #define MDB_MINKEYS 2
335 /** A stamp that identifies a file as an MDB file.
336 * There's nothing special about this value other than that it is easily
337 * recognizable, and it will reflect any byte order mismatches.
339 #define MDB_MAGIC 0xBEEFC0DE
341 /** The version number for a database's file format. */
342 #define MDB_VERSION 1
344 /** The maximum size of a key in the database.
345 * While data items have essentially unbounded size, we require that
346 * keys all fit onto a regular page. This limit could be raised a bit
347 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
349 #define MAXKEYSIZE 511
354 * This is used for printing a hex dump of a key's contents.
356 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
357 /** Display a key in hex.
359 * Invoke a function to display a key in hex.
361 #define DKEY(x) mdb_dkey(x, kbuf)
363 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
367 /** An invalid page number.
368 * Mainly used to denote an empty tree.
370 #define P_INVALID (~(pgno_t)0)
372 /** Test if a flag \b f is set in a flag word \b w. */
373 #define F_ISSET(w, f) (((w) & (f)) == (f))
375 /** Used for offsets within a single page.
376 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
379 typedef uint16_t indx_t;
381 /** Default size of memory map.
382 * This is certainly too small for any actual applications. Apps should always set
383 * the size explicitly using #mdb_env_set_mapsize().
385 #define DEFAULT_MAPSIZE 1048576
387 /** @defgroup readers Reader Lock Table
388 * Readers don't acquire any locks for their data access. Instead, they
389 * simply record their transaction ID in the reader table. The reader
390 * mutex is needed just to find an empty slot in the reader table. The
391 * slot's address is saved in thread-specific data so that subsequent read
392 * transactions started by the same thread need no further locking to proceed.
394 * Since the database uses multi-version concurrency control, readers don't
395 * actually need any locking. This table is used to keep track of which
396 * readers are using data from which old transactions, so that we'll know
397 * when a particular old transaction is no longer in use. Old transactions
398 * that have discarded any data pages can then have those pages reclaimed
399 * for use by a later write transaction.
401 * The lock table is constructed such that reader slots are aligned with the
402 * processor's cache line size. Any slot is only ever used by one thread.
403 * This alignment guarantees that there will be no contention or cache
404 * thrashing as threads update their own slot info, and also eliminates
405 * any need for locking when accessing a slot.
407 * A writer thread will scan every slot in the table to determine the oldest
408 * outstanding reader transaction. Any freed pages older than this will be
409 * reclaimed by the writer. The writer doesn't use any locks when scanning
410 * this table. This means that there's no guarantee that the writer will
411 * see the most up-to-date reader info, but that's not required for correct
412 * operation - all we need is to know the upper bound on the oldest reader,
413 * we don't care at all about the newest reader. So the only consequence of
414 * reading stale information here is that old pages might hang around a
415 * while longer before being reclaimed. That's actually good anyway, because
416 * the longer we delay reclaiming old pages, the more likely it is that a
417 * string of contiguous pages can be found after coalescing old pages from
418 * many old transactions together.
420 * @todo We don't actually do such coalescing yet, we grab pages from one
421 * old transaction at a time.
424 /** Number of slots in the reader table.
425 * This value was chosen somewhat arbitrarily. 126 readers plus a
426 * couple mutexes fit exactly into 8KB on my development machine.
427 * Applications should set the table size using #mdb_env_set_maxreaders().
429 #define DEFAULT_READERS 126
431 /** The size of a CPU cache line in bytes. We want our lock structures
432 * aligned to this size to avoid false cache line sharing in the
434 * This value works for most CPUs. For Itanium this should be 128.
440 /** The information we store in a single slot of the reader table.
441 * In addition to a transaction ID, we also record the process and
442 * thread ID that owns a slot, so that we can detect stale information,
443 * e.g. threads or processes that went away without cleaning up.
444 * @note We currently don't check for stale records. We simply re-init
445 * the table when we know that we're the only process opening the
448 typedef struct MDB_rxbody {
449 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
450 * Multiple readers that start at the same time will probably have the
451 * same ID here. Again, it's not important to exclude them from
452 * anything; all we need to know is which version of the DB they
453 * started from so we can avoid overwriting any data used in that
454 * particular version.
457 /** The process ID of the process owning this reader txn. */
459 /** The thread ID of the thread owning this txn. */
463 /** The actual reader record, with cacheline padding. */
464 typedef struct MDB_reader {
467 /** shorthand for mrb_txnid */
468 #define mr_txnid mru.mrx.mrb_txnid
469 #define mr_pid mru.mrx.mrb_pid
470 #define mr_tid mru.mrx.mrb_tid
471 /** cache line alignment */
472 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
476 /** The header for the reader table.
477 * The table resides in a memory-mapped file. (This is a different file
478 * than is used for the main database.)
480 * For POSIX the actual mutexes reside in the shared memory of this
481 * mapped file. On Windows, mutexes are named objects allocated by the
482 * kernel; we store the mutex names in this mapped file so that other
483 * processes can grab them. This same approach is also used on
484 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
485 * process-shared POSIX mutexes. For these cases where a named object
486 * is used, the object name is derived from a 64 bit FNV hash of the
487 * environment pathname. As such, naming collisions are extremely
488 * unlikely. If a collision occurs, the results are unpredictable.
490 typedef struct MDB_txbody {
491 /** Stamp identifying this as an MDB file. It must be set
494 /** Version number of this lock file. Must be set to #MDB_VERSION. */
495 uint32_t mtb_version;
496 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
497 char mtb_rmname[MNAME_LEN];
499 /** Mutex protecting access to this table.
500 * This is the reader lock that #LOCK_MUTEX_R acquires.
502 pthread_mutex_t mtb_mutex;
504 /** The ID of the last transaction committed to the database.
505 * This is recorded here only for convenience; the value can always
506 * be determined by reading the main database meta pages.
509 /** The number of slots that have been used in the reader table.
510 * This always records the maximum count, it is not decremented
511 * when readers release their slots.
513 unsigned mtb_numreaders;
516 /** The actual reader table definition. */
517 typedef struct MDB_txninfo {
520 #define mti_magic mt1.mtb.mtb_magic
521 #define mti_version mt1.mtb.mtb_version
522 #define mti_mutex mt1.mtb.mtb_mutex
523 #define mti_rmname mt1.mtb.mtb_rmname
524 #define mti_txnid mt1.mtb.mtb_txnid
525 #define mti_numreaders mt1.mtb.mtb_numreaders
526 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
529 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
530 char mt2_wmname[MNAME_LEN];
531 #define mti_wmname mt2.mt2_wmname
533 pthread_mutex_t mt2_wmutex;
534 #define mti_wmutex mt2.mt2_wmutex
536 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
538 MDB_reader mti_readers[1];
542 /** Common header for all page types.
543 * Overflow records occupy a number of contiguous pages with no
544 * headers on any page after the first.
546 typedef struct MDB_page {
547 #define mp_pgno mp_p.p_pgno
548 #define mp_next mp_p.p_next
550 pgno_t p_pgno; /**< page number */
551 void * p_next; /**< for in-memory list of freed structs */
554 /** @defgroup mdb_page Page Flags
556 * Flags for the page headers.
559 #define P_BRANCH 0x01 /**< branch page */
560 #define P_LEAF 0x02 /**< leaf page */
561 #define P_OVERFLOW 0x04 /**< overflow page */
562 #define P_META 0x08 /**< meta page */
563 #define P_DIRTY 0x10 /**< dirty page */
564 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
565 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
567 uint16_t mp_flags; /**< @ref mdb_page */
568 #define mp_lower mp_pb.pb.pb_lower
569 #define mp_upper mp_pb.pb.pb_upper
570 #define mp_pages mp_pb.pb_pages
573 indx_t pb_lower; /**< lower bound of free space */
574 indx_t pb_upper; /**< upper bound of free space */
576 uint32_t pb_pages; /**< number of overflow pages */
578 indx_t mp_ptrs[1]; /**< dynamic size */
581 /** Size of the page header, excluding dynamic data at the end */
582 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
584 /** Address of first usable data byte in a page, after the header */
585 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
587 /** Number of nodes on a page */
588 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
590 /** The amount of space remaining in the page */
591 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
593 /** The percentage of space used in the page, in tenths of a percent. */
594 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
595 ((env)->me_psize - PAGEHDRSZ))
596 /** The minimum page fill factor, in tenths of a percent.
597 * Pages emptier than this are candidates for merging.
599 #define FILL_THRESHOLD 250
601 /** Test if a page is a leaf page */
602 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
603 /** Test if a page is a LEAF2 page */
604 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
605 /** Test if a page is a branch page */
606 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
607 /** Test if a page is an overflow page */
608 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
609 /** Test if a page is a sub page */
610 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
612 /** The number of overflow pages needed to store the given size. */
613 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
615 /** Header for a single key/data pair within a page.
616 * We guarantee 2-byte alignment for nodes.
618 typedef struct MDB_node {
619 /** lo and hi are used for data size on leaf nodes and for
620 * child pgno on branch nodes. On 64 bit platforms, flags
621 * is also used for pgno. (Branch nodes have no flags).
622 * They are in host byte order in case that lets some
623 * accesses be optimized into a 32-bit word access.
625 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
626 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
627 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
628 /** @defgroup mdb_node Node Flags
630 * Flags for node headers.
633 #define F_BIGDATA 0x01 /**< data put on overflow page */
634 #define F_SUBDATA 0x02 /**< data is a sub-database */
635 #define F_DUPDATA 0x04 /**< data has duplicates */
637 /** valid flags for #mdb_node_add() */
638 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
641 unsigned short mn_flags; /**< @ref mdb_node */
642 unsigned short mn_ksize; /**< key size */
643 char mn_data[1]; /**< key and data are appended here */
646 /** Size of the node header, excluding dynamic data at the end */
647 #define NODESIZE offsetof(MDB_node, mn_data)
649 /** Bit position of top word in page number, for shifting mn_flags */
650 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
652 /** Size of a node in a branch page with a given key.
653 * This is just the node header plus the key, there is no data.
655 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
657 /** Size of a node in a leaf page with a given key and data.
658 * This is node header plus key plus data size.
660 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
662 /** Address of node \b i in page \b p */
663 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
665 /** Address of the key for the node */
666 #define NODEKEY(node) (void *)((node)->mn_data)
668 /** Address of the data for a node */
669 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
671 /** Get the page number pointed to by a branch node */
672 #define NODEPGNO(node) \
673 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
674 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
675 /** Set the page number in a branch node */
676 #define SETPGNO(node,pgno) do { \
677 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
678 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
680 /** Get the size of the data in a leaf node */
681 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
682 /** Set the size of the data for a leaf node */
683 #define SETDSZ(node,size) do { \
684 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
685 /** The size of a key in a node */
686 #define NODEKSZ(node) ((node)->mn_ksize)
688 /** Copy a page number from src to dst */
690 #define COPY_PGNO(dst,src) dst = src
692 #if SIZE_MAX > 4294967295UL
693 #define COPY_PGNO(dst,src) do { \
694 unsigned short *s, *d; \
695 s = (unsigned short *)&(src); \
696 d = (unsigned short *)&(dst); \
703 #define COPY_PGNO(dst,src) do { \
704 unsigned short *s, *d; \
705 s = (unsigned short *)&(src); \
706 d = (unsigned short *)&(dst); \
712 /** The address of a key in a LEAF2 page.
713 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
714 * There are no node headers, keys are stored contiguously.
716 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
718 /** Set the \b node's key into \b key, if requested. */
719 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
720 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
722 /** Information about a single database in the environment. */
723 typedef struct MDB_db {
724 uint32_t md_pad; /**< also ksize for LEAF2 pages */
725 uint16_t md_flags; /**< @ref mdb_open */
726 uint16_t md_depth; /**< depth of this tree */
727 pgno_t md_branch_pages; /**< number of internal pages */
728 pgno_t md_leaf_pages; /**< number of leaf pages */
729 pgno_t md_overflow_pages; /**< number of overflow pages */
730 size_t md_entries; /**< number of data items */
731 pgno_t md_root; /**< the root page of this tree */
734 /** Handle for the DB used to track free pages. */
736 /** Handle for the default DB. */
739 /** Meta page content. */
740 typedef struct MDB_meta {
741 /** Stamp identifying this as an MDB file. It must be set
744 /** Version number of this lock file. Must be set to #MDB_VERSION. */
746 void *mm_address; /**< address for fixed mapping */
747 size_t mm_mapsize; /**< size of mmap region */
748 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
749 /** The size of pages used in this DB */
750 #define mm_psize mm_dbs[0].md_pad
751 /** Any persistent environment flags. @ref mdb_env */
752 #define mm_flags mm_dbs[0].md_flags
753 pgno_t mm_last_pg; /**< last used page in file */
754 txnid_t mm_txnid; /**< txnid that committed this page */
757 /** Buffer for a stack-allocated dirty page.
758 * The members define size and alignment, and silence type
759 * aliasing warnings. They are not used directly; that could
760 * mean incorrectly using several union members in parallel.
762 typedef union MDB_pagebuf {
763 char mb_raw[MDB_PAGESIZE];
766 char mm_pad[PAGEHDRSZ];
771 /** Auxiliary DB info.
772 * The information here is mostly static/read-only. There is
773 * only a single copy of this record in the environment.
775 typedef struct MDB_dbx {
776 MDB_val md_name; /**< name of the database */
777 MDB_cmp_func *md_cmp; /**< function for comparing keys */
778 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
779 MDB_rel_func *md_rel; /**< user relocate function */
780 void *md_relctx; /**< user-provided context for md_rel */
783 /** A database transaction.
784 * Every operation requires a transaction handle.
787 MDB_txn *mt_parent; /**< parent of a nested txn */
788 MDB_txn *mt_child; /**< nested txn under this txn */
789 pgno_t mt_next_pgno; /**< next unallocated page */
790 /** The ID of this transaction. IDs are integers incrementing from 1.
791 * Only committed write transactions increment the ID. If a transaction
792 * aborts, the ID may be re-used by the next writer.
795 MDB_env *mt_env; /**< the DB environment */
796 /** The list of pages that became unused during this transaction.
800 MDB_ID2L dirty_list; /**< modified pages */
801 MDB_reader *reader; /**< this thread's slot in the reader table */
803 /** Array of records for each DB known in the environment. */
805 /** Array of MDB_db records for each known DB */
807 /** @defgroup mt_dbflag Transaction DB Flags
811 #define DB_DIRTY 0x01 /**< DB was written in this txn */
812 #define DB_STALE 0x02 /**< DB record is older than txnID */
814 /** Array of cursors for each DB */
815 MDB_cursor **mt_cursors;
816 /** Array of flags for each DB */
817 unsigned char *mt_dbflags;
818 /** Number of DB records in use. This number only ever increments;
819 * we don't decrement it when individual DB handles are closed.
823 /** @defgroup mdb_txn Transaction Flags
827 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
828 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
829 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
831 unsigned int mt_flags; /**< @ref mdb_txn */
832 /** Tracks which of the two meta pages was used at the start
833 * of this transaction.
835 unsigned int mt_toggle;
838 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
839 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
840 * raise this on a 64 bit machine.
842 #define CURSOR_STACK 32
846 /** Cursors are used for all DB operations */
848 /** Next cursor on this DB in this txn */
850 /** Original cursor if this is a shadow */
852 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
853 struct MDB_xcursor *mc_xcursor;
854 /** The transaction that owns this cursor */
856 /** The database handle this cursor operates on */
858 /** The database record for this cursor */
860 /** The database auxiliary record for this cursor */
862 /** The @ref mt_dbflag for this database */
863 unsigned char *mc_dbflag;
864 unsigned short mc_snum; /**< number of pushed pages */
865 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
866 /** @defgroup mdb_cursor Cursor Flags
868 * Cursor state flags.
871 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
872 #define C_EOF 0x02 /**< No more data */
873 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
874 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
875 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
876 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
878 unsigned int mc_flags; /**< @ref mdb_cursor */
879 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
880 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
883 /** Context for sorted-dup records.
884 * We could have gone to a fully recursive design, with arbitrarily
885 * deep nesting of sub-databases. But for now we only handle these
886 * levels - main DB, optional sub-DB, sorted-duplicate DB.
888 typedef struct MDB_xcursor {
889 /** A sub-cursor for traversing the Dup DB */
890 MDB_cursor mx_cursor;
891 /** The database record for this Dup DB */
893 /** The auxiliary DB record for this Dup DB */
895 /** The @ref mt_dbflag for this Dup DB */
896 unsigned char mx_dbflag;
899 /** A set of pages freed by an earlier transaction. */
900 typedef struct MDB_oldpages {
901 /** Usually we only read one record from the FREEDB at a time, but
902 * in case we read more, this will chain them together.
904 struct MDB_oldpages *mo_next;
905 /** The ID of the transaction in which these pages were freed. */
907 /** An #MDB_IDL of the pages */
908 pgno_t mo_pages[1]; /* dynamic */
911 /** The database environment. */
913 HANDLE me_fd; /**< The main data file */
914 HANDLE me_lfd; /**< The lock file */
915 HANDLE me_mfd; /**< just for writing the meta pages */
916 /** Failed to update the meta page. Probably an I/O error. */
917 #define MDB_FATAL_ERROR 0x80000000U
918 /** Read-only Filesystem. Allow read access, no locking. */
919 #define MDB_ROFS 0x40000000U
920 /** Some fields are initialized. */
921 #define MDB_ENV_ACTIVE 0x20000000U
922 uint32_t me_flags; /**< @ref mdb_env */
923 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
924 unsigned int me_maxreaders; /**< size of the reader table */
925 unsigned int me_numreaders; /**< max numreaders set by this env */
926 MDB_dbi me_numdbs; /**< number of DBs opened */
927 MDB_dbi me_maxdbs; /**< size of the DB table */
928 pid_t me_pid; /**< process ID of this env */
929 char *me_path; /**< path to the DB files */
930 char *me_map; /**< the memory map of the data file */
931 MDB_txninfo *me_txns; /**< the memory map of the lock file */
932 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
933 MDB_txn *me_txn; /**< current write transaction */
934 size_t me_mapsize; /**< size of the data memory map */
935 off_t me_size; /**< current file size */
936 pgno_t me_maxpg; /**< me_mapsize / me_psize */
937 txnid_t me_pgfirst; /**< ID of first old page record we used */
938 txnid_t me_pglast; /**< ID of last old page record we used */
939 MDB_dbx *me_dbxs; /**< array of static DB info */
940 uint16_t *me_dbflags; /**< array of DB flags */
941 MDB_oldpages *me_pghead; /**< list of old page records */
942 MDB_oldpages *me_pgfree; /**< list of page records to free */
943 pthread_key_t me_txkey; /**< thread-key for readers */
944 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
945 /** IDL of pages that became unused in a write txn */
947 /** ID2L of pages that were written during a write txn */
948 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
950 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
952 #elif defined(MDB_USE_POSIX_SEM)
953 sem_t *me_rmutex; /* Shared mutexes are not supported */
957 /** max number of pages to commit in one writev() call */
958 #define MDB_COMMIT_PAGES 64
959 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
960 #undef MDB_COMMIT_PAGES
961 #define MDB_COMMIT_PAGES IOV_MAX
964 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
965 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
966 static int mdb_page_touch(MDB_cursor *mc);
968 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
969 static int mdb_page_search_root(MDB_cursor *mc,
970 MDB_val *key, int modify);
971 #define MDB_PS_MODIFY 1
972 #define MDB_PS_ROOTONLY 2
973 static int mdb_page_search(MDB_cursor *mc,
974 MDB_val *key, int flags);
975 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
977 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
978 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
979 pgno_t newpgno, unsigned int nflags);
981 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
982 static int mdb_env_pick_meta(const MDB_env *env);
983 static int mdb_env_write_meta(MDB_txn *txn);
984 static void mdb_env_close0(MDB_env *env, int excl);
986 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
987 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
988 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
989 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
990 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
991 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
992 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
993 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
994 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
996 static int mdb_rebalance(MDB_cursor *mc);
997 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
999 static void mdb_cursor_pop(MDB_cursor *mc);
1000 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1002 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1003 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1004 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1005 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1006 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1008 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1009 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1011 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1012 static void mdb_xcursor_init0(MDB_cursor *mc);
1013 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1015 static int mdb_drop0(MDB_cursor *mc, int subs);
1016 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1019 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1023 static SECURITY_DESCRIPTOR mdb_null_sd;
1024 static SECURITY_ATTRIBUTES mdb_all_sa;
1025 static int mdb_sec_inited;
1028 /** Return the library version info. */
1030 mdb_version(int *major, int *minor, int *patch)
1032 if (major) *major = MDB_VERSION_MAJOR;
1033 if (minor) *minor = MDB_VERSION_MINOR;
1034 if (patch) *patch = MDB_VERSION_PATCH;
1035 return MDB_VERSION_STRING;
1038 /** Table of descriptions for MDB @ref errors */
1039 static char *const mdb_errstr[] = {
1040 "MDB_KEYEXIST: Key/data pair already exists",
1041 "MDB_NOTFOUND: No matching key/data pair found",
1042 "MDB_PAGE_NOTFOUND: Requested page not found",
1043 "MDB_CORRUPTED: Located page was wrong type",
1044 "MDB_PANIC: Update of meta page failed",
1045 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1046 "MDB_INVALID: File is not an MDB file",
1047 "MDB_MAP_FULL: Environment mapsize limit reached",
1048 "MDB_DBS_FULL: Environment maxdbs limit reached",
1049 "MDB_READERS_FULL: Environment maxreaders limit reached",
1050 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1051 "MDB_TXN_FULL: Nested transaction has too many dirty pages - transaction too big",
1052 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1053 "MDB_PAGE_FULL: Internal error - page has no more space"
1057 mdb_strerror(int err)
1061 return ("Successful return: 0");
1063 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1064 i = err - MDB_KEYEXIST;
1065 return mdb_errstr[i];
1068 return strerror(err);
1072 /** Display a key in hexadecimal and return the address of the result.
1073 * @param[in] key the key to display
1074 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1075 * @return The key in hexadecimal form.
1078 mdb_dkey(MDB_val *key, char *buf)
1081 unsigned char *c = key->mv_data;
1083 if (key->mv_size > MAXKEYSIZE)
1084 return "MAXKEYSIZE";
1085 /* may want to make this a dynamic check: if the key is mostly
1086 * printable characters, print it as-is instead of converting to hex.
1090 for (i=0; i<key->mv_size; i++)
1091 ptr += sprintf(ptr, "%02x", *c++);
1093 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1098 /** Display all the keys in the page. */
1100 mdb_page_list(MDB_page *mp)
1103 unsigned int i, nkeys, nsize;
1107 nkeys = NUMKEYS(mp);
1108 fprintf(stderr, "numkeys %d\n", nkeys);
1109 for (i=0; i<nkeys; i++) {
1110 node = NODEPTR(mp, i);
1111 key.mv_size = node->mn_ksize;
1112 key.mv_data = node->mn_data;
1113 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1114 if (F_ISSET(node->mn_flags, F_BIGDATA))
1115 nsize += sizeof(pgno_t);
1117 nsize += NODEDSZ(node);
1118 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1123 mdb_cursor_chk(MDB_cursor *mc)
1129 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1130 for (i=0; i<mc->mc_top; i++) {
1132 node = NODEPTR(mp, mc->mc_ki[i]);
1133 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1136 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1142 /** Count all the pages in each DB and in the freelist
1143 * and make sure it matches the actual number of pages
1146 static void mdb_audit(MDB_txn *txn)
1150 MDB_ID freecount, count;
1155 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1156 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1157 freecount += *(MDB_ID *)data.mv_data;
1160 for (i = 0; i<txn->mt_numdbs; i++) {
1161 MDB_xcursor mx, *mxp;
1162 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1163 mdb_cursor_init(&mc, txn, i, mxp);
1164 if (txn->mt_dbs[i].md_root == P_INVALID)
1166 count += txn->mt_dbs[i].md_branch_pages +
1167 txn->mt_dbs[i].md_leaf_pages +
1168 txn->mt_dbs[i].md_overflow_pages;
1169 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1170 mdb_page_search(&mc, NULL, 0);
1174 mp = mc.mc_pg[mc.mc_top];
1175 for (j=0; j<NUMKEYS(mp); j++) {
1176 MDB_node *leaf = NODEPTR(mp, j);
1177 if (leaf->mn_flags & F_SUBDATA) {
1179 memcpy(&db, NODEDATA(leaf), sizeof(db));
1180 count += db.md_branch_pages + db.md_leaf_pages +
1181 db.md_overflow_pages;
1185 while (mdb_cursor_sibling(&mc, 1) == 0);
1188 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1189 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1190 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1196 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1198 return txn->mt_dbxs[dbi].md_cmp(a, b);
1202 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1204 if (txn->mt_dbxs[dbi].md_dcmp)
1205 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1207 return EINVAL; /* too bad you can't distinguish this from a valid result */
1210 /** Allocate a single page.
1211 * Re-use old malloc'd pages first, otherwise just malloc.
1214 mdb_page_malloc(MDB_cursor *mc) {
1216 size_t sz = mc->mc_txn->mt_env->me_psize;
1217 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1218 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1219 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1220 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1221 } else if ((ret = malloc(sz)) != NULL) {
1222 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1227 /** Allocate pages for writing.
1228 * If there are free pages available from older transactions, they
1229 * will be re-used first. Otherwise a new page will be allocated.
1230 * @param[in] mc cursor A cursor handle identifying the transaction and
1231 * database for which we are allocating.
1232 * @param[in] num the number of pages to allocate.
1233 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1234 * will always be satisfied by a single contiguous chunk of memory.
1235 * @return 0 on success, non-zero on failure.
1238 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1240 MDB_txn *txn = mc->mc_txn;
1242 pgno_t pgno = P_INVALID;
1247 /* The free list won't have any content at all until txn 2 has
1248 * committed. The pages freed by txn 2 will be unreferenced
1249 * after txn 3 commits, and so will be safe to re-use in txn 4.
1251 if (txn->mt_txnid > 3) {
1253 if (!txn->mt_env->me_pghead &&
1254 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1255 /* See if there's anything in the free DB */
1261 txnid_t *kptr, last;
1263 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1264 if (!txn->mt_env->me_pgfirst) {
1265 mdb_page_search(&m2, NULL, 0);
1266 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1267 kptr = (txnid_t *)NODEKEY(leaf);
1274 last = txn->mt_env->me_pglast + 1;
1276 key.mv_data = &last;
1277 key.mv_size = sizeof(last);
1278 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1281 last = *(txnid_t *)key.mv_data;
1284 /* Unusable if referred by a meta page or reader... */
1286 if (last < txn->mt_txnid-1) {
1287 j = txn->mt_env->me_txns->mti_numreaders;
1288 r = txn->mt_env->me_txns->mti_readers + j;
1289 for (j = -j; j && (last<r[j].mr_txnid || !r[j].mr_pid); j++) ;
1293 /* It's usable, grab it.
1298 if (!txn->mt_env->me_pgfirst) {
1299 mdb_node_read(txn, leaf, &data);
1301 txn->mt_env->me_pglast = last;
1302 if (!txn->mt_env->me_pgfirst)
1303 txn->mt_env->me_pgfirst = last;
1304 idl = (MDB_ID *) data.mv_data;
1305 /* We might have a zero-length IDL due to freelist growth
1306 * during a prior commit
1308 if (!idl[0]) goto again;
1309 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1312 mop->mo_next = txn->mt_env->me_pghead;
1313 mop->mo_txnid = last;
1314 txn->mt_env->me_pghead = mop;
1315 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1320 DPRINTF("IDL read txn %zu root %zu num %zu",
1321 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1322 for (i=0; i<idl[0]; i++) {
1323 DPRINTF("IDL %zu", idl[i+1]);
1330 if (txn->mt_env->me_pghead) {
1331 MDB_oldpages *mop = txn->mt_env->me_pghead;
1333 /* FIXME: For now, always use fresh pages. We
1334 * really ought to search the free list for a
1339 /* peel pages off tail, so we only have to truncate the list */
1340 pgno = MDB_IDL_LAST(mop->mo_pages);
1341 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1343 if (mop->mo_pages[2] > mop->mo_pages[1])
1344 mop->mo_pages[0] = 0;
1348 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1349 txn->mt_env->me_pghead = mop->mo_next;
1350 if (mc->mc_dbi == FREE_DBI) {
1351 mop->mo_next = txn->mt_env->me_pgfree;
1352 txn->mt_env->me_pgfree = mop;
1361 if (pgno == P_INVALID) {
1362 /* DB size is maxed out */
1363 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1364 DPUTS("DB size maxed out");
1365 return MDB_MAP_FULL;
1368 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1369 if (pgno == P_INVALID) {
1370 pgno = txn->mt_next_pgno;
1371 txn->mt_next_pgno += num;
1373 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1376 if (txn->mt_env->me_dpages && num == 1) {
1377 np = txn->mt_env->me_dpages;
1378 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1379 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1380 txn->mt_env->me_dpages = np->mp_next;
1382 size_t sz = txn->mt_env->me_psize * num;
1383 if ((np = malloc(sz)) == NULL)
1385 VGMEMP_ALLOC(txn->mt_env, np, sz);
1387 if (pgno == P_INVALID) {
1388 np->mp_pgno = txn->mt_next_pgno;
1389 txn->mt_next_pgno += num;
1394 mid.mid = np->mp_pgno;
1396 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1397 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1399 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1406 /** Copy a page: avoid copying unused portions of the page.
1407 * @param[in] dst page to copy into
1408 * @param[in] src page to copy from
1411 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1413 dst->mp_flags = src->mp_flags | P_DIRTY;
1414 dst->mp_pages = src->mp_pages;
1416 if (IS_LEAF2(src)) {
1417 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1419 unsigned int i, nkeys = NUMKEYS(src);
1420 for (i=0; i<nkeys; i++)
1421 dst->mp_ptrs[i] = src->mp_ptrs[i];
1422 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1423 psize - src->mp_upper);
1427 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1428 * @param[in] mc cursor pointing to the page to be touched
1429 * @return 0 on success, non-zero on failure.
1432 mdb_page_touch(MDB_cursor *mc)
1434 MDB_page *mp = mc->mc_pg[mc->mc_top];
1438 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1440 if ((rc = mdb_page_alloc(mc, 1, &np)))
1442 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1443 assert(mp->mp_pgno != np->mp_pgno);
1444 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1446 /* If page isn't full, just copy the used portion */
1447 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1450 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1452 np->mp_flags |= P_DIRTY;
1457 /* Adjust other cursors pointing to mp */
1458 if (mc->mc_flags & C_SUB) {
1459 MDB_cursor *m2, *m3;
1460 MDB_dbi dbi = mc->mc_dbi-1;
1462 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1463 if (m2 == mc) continue;
1464 m3 = &m2->mc_xcursor->mx_cursor;
1465 if (m3->mc_snum < mc->mc_snum) continue;
1466 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1467 m3->mc_pg[mc->mc_top] = mp;
1473 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1474 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1475 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1476 m2->mc_pg[mc->mc_top] = mp;
1480 mc->mc_pg[mc->mc_top] = mp;
1481 /** If this page has a parent, update the parent to point to
1485 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1487 mc->mc_db->md_root = mp->mp_pgno;
1488 } else if (mc->mc_txn->mt_parent) {
1491 /* If txn has a parent, make sure the page is in our
1494 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1495 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1496 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1497 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1498 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1499 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1500 mc->mc_pg[mc->mc_top] = mp;
1506 np = mdb_page_malloc(mc);
1509 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1510 mid.mid = np->mp_pgno;
1512 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1520 mdb_env_sync(MDB_env *env, int force)
1523 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1524 if (env->me_flags & MDB_WRITEMAP) {
1525 int flags = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
1526 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1529 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1533 if (MDB_FDATASYNC(env->me_fd))
1540 /** Make shadow copies of all of parent txn's cursors */
1542 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1544 MDB_cursor *mc, *m2;
1545 unsigned int i, j, size;
1547 for (i=0;i<src->mt_numdbs; i++) {
1548 if (src->mt_cursors[i]) {
1549 size = sizeof(MDB_cursor);
1550 if (src->mt_cursors[i]->mc_xcursor)
1551 size += sizeof(MDB_xcursor);
1552 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1559 mc->mc_db = &dst->mt_dbs[i];
1560 mc->mc_dbx = m2->mc_dbx;
1561 mc->mc_dbflag = &dst->mt_dbflags[i];
1562 mc->mc_snum = m2->mc_snum;
1563 mc->mc_top = m2->mc_top;
1564 mc->mc_flags = m2->mc_flags | C_SHADOW;
1565 for (j=0; j<mc->mc_snum; j++) {
1566 mc->mc_pg[j] = m2->mc_pg[j];
1567 mc->mc_ki[j] = m2->mc_ki[j];
1569 if (m2->mc_xcursor) {
1570 MDB_xcursor *mx, *mx2;
1571 mx = (MDB_xcursor *)(mc+1);
1572 mc->mc_xcursor = mx;
1573 mx2 = m2->mc_xcursor;
1574 mx->mx_db = mx2->mx_db;
1575 mx->mx_dbx = mx2->mx_dbx;
1576 mx->mx_dbflag = mx2->mx_dbflag;
1577 mx->mx_cursor.mc_txn = dst;
1578 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1579 mx->mx_cursor.mc_db = &mx->mx_db;
1580 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1581 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1582 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1583 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1584 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1585 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1586 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1587 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1590 mc->mc_xcursor = NULL;
1592 mc->mc_next = dst->mt_cursors[i];
1593 dst->mt_cursors[i] = mc;
1600 /** Merge shadow cursors back into parent's */
1602 mdb_cursor_merge(MDB_txn *txn)
1605 for (i=0; i<txn->mt_numdbs; i++) {
1606 if (txn->mt_cursors[i]) {
1608 while ((mc = txn->mt_cursors[i])) {
1609 txn->mt_cursors[i] = mc->mc_next;
1610 if (mc->mc_flags & C_SHADOW) {
1611 MDB_cursor *m2 = mc->mc_orig;
1613 m2->mc_snum = mc->mc_snum;
1614 m2->mc_top = mc->mc_top;
1615 for (j=0; j<mc->mc_snum; j++) {
1616 m2->mc_pg[j] = mc->mc_pg[j];
1617 m2->mc_ki[j] = mc->mc_ki[j];
1620 if (mc->mc_flags & C_ALLOCD)
1628 mdb_txn_reset0(MDB_txn *txn);
1630 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1631 * @param[in] txn the transaction handle to initialize
1632 * @return 0 on success, non-zero on failure. This can only
1633 * fail for read-only transactions, and then only if the
1634 * reader table is full.
1637 mdb_txn_renew0(MDB_txn *txn)
1639 MDB_env *env = txn->mt_env;
1644 txn->mt_numdbs = env->me_numdbs;
1645 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1647 if (txn->mt_flags & MDB_TXN_RDONLY) {
1648 if (env->me_flags & MDB_ROFS) {
1649 i = mdb_env_pick_meta(env);
1650 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1651 txn->mt_u.reader = NULL;
1653 MDB_reader *r = pthread_getspecific(env->me_txkey);
1655 pid_t pid = env->me_pid;
1656 pthread_t tid = pthread_self();
1659 for (i=0; i<env->me_txns->mti_numreaders; i++)
1660 if (env->me_txns->mti_readers[i].mr_pid == 0)
1662 if (i == env->me_maxreaders) {
1663 UNLOCK_MUTEX_R(env);
1664 return MDB_READERS_FULL;
1666 env->me_txns->mti_readers[i].mr_pid = pid;
1667 env->me_txns->mti_readers[i].mr_tid = tid;
1668 if (i >= env->me_txns->mti_numreaders)
1669 env->me_txns->mti_numreaders = i+1;
1670 /* Save numreaders for un-mutexed mdb_env_close() */
1671 env->me_numreaders = env->me_txns->mti_numreaders;
1672 UNLOCK_MUTEX_R(env);
1673 r = &env->me_txns->mti_readers[i];
1674 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1675 env->me_txns->mti_readers[i].mr_pid = 0;
1679 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1680 txn->mt_u.reader = r;
1682 txn->mt_toggle = txn->mt_txnid & 1;
1683 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1687 txn->mt_txnid = env->me_txns->mti_txnid;
1688 txn->mt_toggle = txn->mt_txnid & 1;
1689 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1692 if (txn->mt_txnid == mdb_debug_start)
1695 txn->mt_u.dirty_list = env->me_dirty_list;
1696 txn->mt_u.dirty_list[0].mid = 0;
1697 txn->mt_free_pgs = env->me_free_pgs;
1698 txn->mt_free_pgs[0] = 0;
1702 /* Copy the DB info and flags */
1703 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1704 for (i=2; i<txn->mt_numdbs; i++)
1705 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1706 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1707 if (txn->mt_numdbs > 2)
1708 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1714 mdb_txn_renew(MDB_txn *txn)
1718 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1721 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1722 DPUTS("environment had fatal error, must shutdown!");
1726 rc = mdb_txn_renew0(txn);
1727 if (rc == MDB_SUCCESS) {
1728 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1729 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1730 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1736 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1741 if (env->me_flags & MDB_FATAL_ERROR) {
1742 DPUTS("environment had fatal error, must shutdown!");
1745 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1748 /* Nested transactions: Max 1 child, write txns only, no writemap */
1749 if (parent->mt_child ||
1750 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1751 (env->me_flags & MDB_WRITEMAP))
1756 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1757 if (!(flags & MDB_RDONLY))
1758 size += env->me_maxdbs * sizeof(MDB_cursor *);
1760 if ((txn = calloc(1, size)) == NULL) {
1761 DPRINTF("calloc: %s", strerror(ErrCode()));
1764 txn->mt_dbs = (MDB_db *)(txn+1);
1765 if (flags & MDB_RDONLY) {
1766 txn->mt_flags |= MDB_TXN_RDONLY;
1767 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1769 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1770 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1775 txn->mt_free_pgs = mdb_midl_alloc();
1776 if (!txn->mt_free_pgs) {
1780 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1781 if (!txn->mt_u.dirty_list) {
1782 free(txn->mt_free_pgs);
1786 txn->mt_txnid = parent->mt_txnid;
1787 txn->mt_toggle = parent->mt_toggle;
1788 txn->mt_u.dirty_list[0].mid = 0;
1789 txn->mt_free_pgs[0] = 0;
1790 txn->mt_next_pgno = parent->mt_next_pgno;
1791 parent->mt_child = txn;
1792 txn->mt_parent = parent;
1793 txn->mt_numdbs = parent->mt_numdbs;
1794 txn->mt_dbxs = parent->mt_dbxs;
1795 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1796 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1797 mdb_cursor_shadow(parent, txn);
1800 rc = mdb_txn_renew0(txn);
1806 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1807 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1808 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1814 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1815 * @param[in] txn the transaction handle to reset
1818 mdb_txn_reset0(MDB_txn *txn)
1820 MDB_env *env = txn->mt_env;
1822 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1823 if (!(env->me_flags & MDB_ROFS))
1824 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1830 /* close(free) all cursors */
1831 for (i=0; i<txn->mt_numdbs; i++) {
1832 if (txn->mt_cursors[i]) {
1834 while ((mc = txn->mt_cursors[i])) {
1835 txn->mt_cursors[i] = mc->mc_next;
1836 if (mc->mc_flags & C_ALLOCD)
1842 if (!(env->me_flags & MDB_WRITEMAP)) {
1843 /* return all dirty pages to dpage list */
1844 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1845 dp = txn->mt_u.dirty_list[i].mptr;
1846 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1847 dp->mp_next = txn->mt_env->me_dpages;
1848 VGMEMP_FREE(txn->mt_env, dp);
1849 txn->mt_env->me_dpages = dp;
1851 /* large pages just get freed directly */
1852 VGMEMP_FREE(txn->mt_env, dp);
1858 if (txn->mt_parent) {
1859 txn->mt_parent->mt_child = NULL;
1860 mdb_midl_free(txn->mt_free_pgs);
1861 free(txn->mt_u.dirty_list);
1864 if (mdb_midl_shrink(&txn->mt_free_pgs))
1865 env->me_free_pgs = txn->mt_free_pgs;
1868 while ((mop = txn->mt_env->me_pghead)) {
1869 txn->mt_env->me_pghead = mop->mo_next;
1872 txn->mt_env->me_pgfirst = 0;
1873 txn->mt_env->me_pglast = 0;
1876 /* The writer mutex was locked in mdb_txn_begin. */
1877 UNLOCK_MUTEX_W(env);
1882 mdb_txn_reset(MDB_txn *txn)
1887 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1888 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1889 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1891 mdb_txn_reset0(txn);
1895 mdb_txn_abort(MDB_txn *txn)
1900 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1901 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1902 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1905 mdb_txn_abort(txn->mt_child);
1907 mdb_txn_reset0(txn);
1912 mdb_txn_commit(MDB_txn *txn)
1920 pgno_t next, freecnt;
1923 assert(txn != NULL);
1924 assert(txn->mt_env != NULL);
1926 if (txn->mt_child) {
1927 mdb_txn_commit(txn->mt_child);
1928 txn->mt_child = NULL;
1933 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1934 if (txn->mt_numdbs > env->me_numdbs) {
1935 /* update the DB flags */
1937 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
1938 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
1945 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1946 DPUTS("error flag is set, can't commit");
1948 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1953 if (txn->mt_parent) {
1959 /* Merge (and close) our cursors with parent's */
1960 mdb_cursor_merge(txn);
1962 /* Update parent's DB table */
1963 ip = &txn->mt_parent->mt_dbs[2];
1964 jp = &txn->mt_dbs[2];
1965 for (i = 2; i < txn->mt_numdbs; i++) {
1966 if (ip->md_root != jp->md_root)
1970 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1972 /* Append our free list to parent's */
1973 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1975 mdb_midl_free(txn->mt_free_pgs);
1977 /* Merge our dirty list with parent's */
1978 dst = txn->mt_parent->mt_u.dirty_list;
1979 src = txn->mt_u.dirty_list;
1980 x = mdb_mid2l_search(dst, src[1].mid);
1981 for (y=1; y<=src[0].mid; y++) {
1982 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1986 dst[x].mptr = src[y].mptr;
1989 for (; y<=src[0].mid; y++) {
1990 if (++x >= MDB_IDL_UM_MAX) {
1992 return MDB_TXN_FULL;
1997 free(txn->mt_u.dirty_list);
1998 txn->mt_parent->mt_child = NULL;
2003 if (txn != env->me_txn) {
2004 DPUTS("attempt to commit unknown transaction");
2009 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2012 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2013 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2015 /* Update DB root pointers. Their pages have already been
2016 * touched so this is all in-place and cannot fail.
2018 if (txn->mt_numdbs > 2) {
2021 data.mv_size = sizeof(MDB_db);
2023 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2024 for (i = 2; i < txn->mt_numdbs; i++) {
2025 if (txn->mt_dbflags[i] & DB_DIRTY) {
2026 data.mv_data = &txn->mt_dbs[i];
2027 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2032 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2034 /* should only be one record now */
2035 if (env->me_pghead) {
2036 /* make sure first page of freeDB is touched and on freelist */
2037 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2040 /* Delete IDLs we used from the free list */
2041 if (env->me_pgfirst) {
2046 key.mv_size = sizeof(cur);
2047 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2050 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2051 rc = mdb_cursor_del(&mc, 0);
2057 env->me_pgfirst = 0;
2061 /* save to free list */
2063 freecnt = txn->mt_free_pgs[0];
2064 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2067 /* make sure last page of freeDB is touched and on freelist */
2068 key.mv_size = MAXKEYSIZE+1;
2070 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2072 mdb_midl_sort(txn->mt_free_pgs);
2076 MDB_IDL idl = txn->mt_free_pgs;
2077 DPRINTF("IDL write txn %zu root %zu num %zu",
2078 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2079 for (i=0; i<idl[0]; i++) {
2080 DPRINTF("IDL %zu", idl[i+1]);
2084 /* write to last page of freeDB */
2085 key.mv_size = sizeof(pgno_t);
2086 key.mv_data = &txn->mt_txnid;
2087 data.mv_data = txn->mt_free_pgs;
2088 /* The free list can still grow during this call,
2089 * despite the pre-emptive touches above. So check
2090 * and make sure the entire thing got written.
2093 freecnt = txn->mt_free_pgs[0];
2094 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2095 rc = mdb_cursor_put(&mc, &key, &data, 0);
2100 } while (freecnt != txn->mt_free_pgs[0]);
2102 /* should only be one record now */
2104 if (env->me_pghead) {
2110 mop = env->me_pghead;
2112 key.mv_size = sizeof(id);
2114 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2115 data.mv_data = mop->mo_pages;
2116 orig = mop->mo_pages[0];
2117 /* These steps may grow the freelist again
2118 * due to freed overflow pages...
2120 mdb_cursor_put(&mc, &key, &data, 0);
2121 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2122 /* could have been used again here */
2123 if (mop->mo_pages[0] != orig) {
2124 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2125 data.mv_data = mop->mo_pages;
2127 mdb_cursor_put(&mc, &key, &data, 0);
2129 env->me_pghead = NULL;
2132 /* was completely used up */
2133 mdb_cursor_del(&mc, 0);
2137 env->me_pgfirst = 0;
2141 while (env->me_pgfree) {
2142 MDB_oldpages *mop = env->me_pgfree;
2143 env->me_pgfree = mop->mo_next;
2147 /* Check for growth of freelist again */
2148 if (freecnt != txn->mt_free_pgs[0])
2151 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2152 if (mdb_midl_shrink(&txn->mt_free_pgs))
2153 env->me_free_pgs = txn->mt_free_pgs;
2160 if (env->me_flags & MDB_WRITEMAP) {
2161 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2162 dp = txn->mt_u.dirty_list[i].mptr;
2163 /* clear dirty flag */
2164 dp->mp_flags &= ~P_DIRTY;
2165 txn->mt_u.dirty_list[i].mid = 0;
2167 txn->mt_u.dirty_list[0].mid = 0;
2171 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2177 /* Windows actually supports scatter/gather I/O, but only on
2178 * unbuffered file handles. Since we're relying on the OS page
2179 * cache for all our data, that's self-defeating. So we just
2180 * write pages one at a time. We use the ov structure to set
2181 * the write offset, to at least save the overhead of a Seek
2185 memset(&ov, 0, sizeof(ov));
2186 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2188 dp = txn->mt_u.dirty_list[i].mptr;
2189 DPRINTF("committing page %zu", dp->mp_pgno);
2190 size = dp->mp_pgno * env->me_psize;
2191 ov.Offset = size & 0xffffffff;
2192 ov.OffsetHigh = size >> 16;
2193 ov.OffsetHigh >>= 16;
2194 /* clear dirty flag */
2195 dp->mp_flags &= ~P_DIRTY;
2196 wsize = env->me_psize;
2197 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2198 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2201 DPRINTF("WriteFile: %d", n);
2208 struct iovec iov[MDB_COMMIT_PAGES];
2212 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2213 dp = txn->mt_u.dirty_list[i].mptr;
2214 if (dp->mp_pgno != next) {
2216 rc = writev(env->me_fd, iov, n);
2220 DPUTS("short write, filesystem full?");
2222 DPRINTF("writev: %s", strerror(n));
2229 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2232 DPRINTF("committing page %zu", dp->mp_pgno);
2233 iov[n].iov_len = env->me_psize;
2234 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2235 iov[n].iov_base = (char *)dp;
2236 size += iov[n].iov_len;
2237 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2238 /* clear dirty flag */
2239 dp->mp_flags &= ~P_DIRTY;
2240 if (++n >= MDB_COMMIT_PAGES) {
2250 rc = writev(env->me_fd, iov, n);
2254 DPUTS("short write, filesystem full?");
2256 DPRINTF("writev: %s", strerror(n));
2263 /* Drop the dirty pages.
2265 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2266 dp = txn->mt_u.dirty_list[i].mptr;
2267 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2268 dp->mp_next = txn->mt_env->me_dpages;
2269 VGMEMP_FREE(txn->mt_env, dp);
2270 txn->mt_env->me_dpages = dp;
2272 VGMEMP_FREE(txn->mt_env, dp);
2275 txn->mt_u.dirty_list[i].mid = 0;
2277 txn->mt_u.dirty_list[0].mid = 0;
2280 if ((n = mdb_env_sync(env, 0)) != 0 ||
2281 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2288 if (txn->mt_numdbs > env->me_numdbs) {
2289 /* update the DB flags */
2291 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2292 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2296 UNLOCK_MUTEX_W(env);
2302 /** Read the environment parameters of a DB environment before
2303 * mapping it into memory.
2304 * @param[in] env the environment handle
2305 * @param[out] meta address of where to store the meta information
2306 * @return 0 on success, non-zero on failure.
2309 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2316 /* We don't know the page size yet, so use a minimum value.
2320 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2322 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2327 else if (rc != MDB_PAGESIZE) {
2331 DPRINTF("read: %s", strerror(err));
2335 p = (MDB_page *)&pbuf;
2337 if (!F_ISSET(p->mp_flags, P_META)) {
2338 DPRINTF("page %zu not a meta page", p->mp_pgno);
2343 if (m->mm_magic != MDB_MAGIC) {
2344 DPUTS("meta has invalid magic");
2348 if (m->mm_version != MDB_VERSION) {
2349 DPRINTF("database is version %u, expected version %u",
2350 m->mm_version, MDB_VERSION);
2351 return MDB_VERSION_MISMATCH;
2354 memcpy(meta, m, sizeof(*m));
2358 /** Write the environment parameters of a freshly created DB environment.
2359 * @param[in] env the environment handle
2360 * @param[out] meta address of where to store the meta information
2361 * @return 0 on success, non-zero on failure.
2364 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2371 DPUTS("writing new meta page");
2373 GET_PAGESIZE(psize);
2375 meta->mm_magic = MDB_MAGIC;
2376 meta->mm_version = MDB_VERSION;
2377 meta->mm_psize = psize;
2378 meta->mm_last_pg = 1;
2379 meta->mm_flags = env->me_flags & 0xffff;
2380 meta->mm_flags |= MDB_INTEGERKEY;
2381 meta->mm_dbs[0].md_root = P_INVALID;
2382 meta->mm_dbs[1].md_root = P_INVALID;
2384 p = calloc(2, psize);
2386 p->mp_flags = P_META;
2389 memcpy(m, meta, sizeof(*meta));
2391 q = (MDB_page *)((char *)p + psize);
2394 q->mp_flags = P_META;
2397 memcpy(m, meta, sizeof(*meta));
2402 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2403 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2406 rc = write(env->me_fd, p, psize * 2);
2407 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2413 /** Update the environment info to commit a transaction.
2414 * @param[in] txn the transaction that's being committed
2415 * @return 0 on success, non-zero on failure.
2418 mdb_env_write_meta(MDB_txn *txn)
2421 MDB_meta meta, metab, *mp;
2423 int rc, len, toggle;
2429 assert(txn != NULL);
2430 assert(txn->mt_env != NULL);
2432 toggle = !txn->mt_toggle;
2433 DPRINTF("writing meta page %d for root page %zu",
2434 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2437 mp = env->me_metas[toggle];
2439 if (env->me_flags & MDB_WRITEMAP) {
2440 /* Persist any increases of mapsize config */
2441 if (env->me_mapsize > mp->mm_mapsize)
2442 mp->mm_mapsize = env->me_mapsize;
2443 mp->mm_dbs[0] = txn->mt_dbs[0];
2444 mp->mm_dbs[1] = txn->mt_dbs[1];
2445 mp->mm_last_pg = txn->mt_next_pgno - 1;
2446 mp->mm_txnid = txn->mt_txnid;
2447 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2448 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2451 ptr += env->me_psize;
2452 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2459 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2460 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2462 ptr = (char *)&meta;
2463 if (env->me_mapsize > mp->mm_mapsize) {
2464 /* Persist any increases of mapsize config */
2465 meta.mm_mapsize = env->me_mapsize;
2466 off = offsetof(MDB_meta, mm_mapsize);
2468 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2470 len = sizeof(MDB_meta) - off;
2473 meta.mm_dbs[0] = txn->mt_dbs[0];
2474 meta.mm_dbs[1] = txn->mt_dbs[1];
2475 meta.mm_last_pg = txn->mt_next_pgno - 1;
2476 meta.mm_txnid = txn->mt_txnid;
2479 off += env->me_psize;
2482 /* Write to the SYNC fd */
2485 memset(&ov, 0, sizeof(ov));
2487 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2490 rc = pwrite(env->me_mfd, ptr, len, off);
2495 DPUTS("write failed, disk error?");
2496 /* On a failure, the pagecache still contains the new data.
2497 * Write some old data back, to prevent it from being used.
2498 * Use the non-SYNC fd; we know it will fail anyway.
2500 meta.mm_last_pg = metab.mm_last_pg;
2501 meta.mm_txnid = metab.mm_txnid;
2503 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2505 r2 = pwrite(env->me_fd, ptr, len, off);
2508 env->me_flags |= MDB_FATAL_ERROR;
2512 /* Memory ordering issues are irrelevant; since the entire writer
2513 * is wrapped by wmutex, all of these changes will become visible
2514 * after the wmutex is unlocked. Since the DB is multi-version,
2515 * readers will get consistent data regardless of how fresh or
2516 * how stale their view of these values is.
2518 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2523 /** Check both meta pages to see which one is newer.
2524 * @param[in] env the environment handle
2525 * @return meta toggle (0 or 1).
2528 mdb_env_pick_meta(const MDB_env *env)
2530 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2534 mdb_env_create(MDB_env **env)
2538 e = calloc(1, sizeof(MDB_env));
2542 e->me_free_pgs = mdb_midl_alloc();
2543 if (!e->me_free_pgs) {
2547 e->me_maxreaders = DEFAULT_READERS;
2549 e->me_fd = INVALID_HANDLE_VALUE;
2550 e->me_lfd = INVALID_HANDLE_VALUE;
2551 e->me_mfd = INVALID_HANDLE_VALUE;
2552 #ifdef MDB_USE_POSIX_SEM
2553 e->me_rmutex = SEM_FAILED;
2554 e->me_wmutex = SEM_FAILED;
2556 e->me_pid = getpid();
2557 VGMEMP_CREATE(e,0,0);
2563 mdb_env_set_mapsize(MDB_env *env, size_t size)
2567 env->me_mapsize = size;
2569 env->me_maxpg = env->me_mapsize / env->me_psize;
2574 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2578 env->me_maxdbs = dbs;
2583 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2585 if (env->me_map || readers < 1)
2587 env->me_maxreaders = readers;
2592 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2594 if (!env || !readers)
2596 *readers = env->me_maxreaders;
2600 /** Further setup required for opening an MDB environment
2603 mdb_env_open2(MDB_env *env)
2605 unsigned int flags = env->me_flags;
2606 int i, newenv = 0, prot;
2610 memset(&meta, 0, sizeof(meta));
2612 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2615 DPUTS("new mdbenv");
2617 meta.mm_mapsize = env->me_mapsize > DEFAULT_MAPSIZE ? env->me_mapsize : DEFAULT_MAPSIZE;
2620 if (env->me_mapsize < meta.mm_mapsize)
2621 env->me_mapsize = meta.mm_mapsize;
2626 LONG sizelo, sizehi;
2627 sizelo = env->me_mapsize & 0xffffffff;
2628 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2630 /* Windows won't create mappings for zero length files.
2631 * Just allocate the maxsize right now.
2634 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2635 if (!SetEndOfFile(env->me_fd))
2637 SetFilePointer(env->me_fd, 0, NULL, 0);
2639 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2640 PAGE_READWRITE : PAGE_READONLY,
2641 sizehi, sizelo, NULL);
2644 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2645 FILE_MAP_WRITE : FILE_MAP_READ,
2646 0, 0, env->me_mapsize, meta.mm_address);
2653 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2656 if (flags & MDB_WRITEMAP) {
2658 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2661 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2663 if (env->me_map == MAP_FAILED) {
2670 if (flags & MDB_FIXEDMAP)
2671 meta.mm_address = env->me_map;
2672 i = mdb_env_init_meta(env, &meta);
2673 if (i != MDB_SUCCESS) {
2677 env->me_psize = meta.mm_psize;
2679 env->me_maxpg = env->me_mapsize / env->me_psize;
2681 p = (MDB_page *)env->me_map;
2682 env->me_metas[0] = METADATA(p);
2683 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2687 int toggle = mdb_env_pick_meta(env);
2688 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2690 DPRINTF("opened database version %u, pagesize %u",
2691 env->me_metas[0]->mm_version, env->me_psize);
2692 DPRINTF("using meta page %d", toggle);
2693 DPRINTF("depth: %u", db->md_depth);
2694 DPRINTF("entries: %zu", db->md_entries);
2695 DPRINTF("branch pages: %zu", db->md_branch_pages);
2696 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2697 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2698 DPRINTF("root: %zu", db->md_root);
2706 /** Release a reader thread's slot in the reader lock table.
2707 * This function is called automatically when a thread exits.
2708 * @param[in] ptr This points to the slot in the reader lock table.
2711 mdb_env_reader_dest(void *ptr)
2713 MDB_reader *reader = ptr;
2719 /** Junk for arranging thread-specific callbacks on Windows. This is
2720 * necessarily platform and compiler-specific. Windows supports up
2721 * to 1088 keys. Let's assume nobody opens more than 64 environments
2722 * in a single process, for now. They can override this if needed.
2724 #ifndef MAX_TLS_KEYS
2725 #define MAX_TLS_KEYS 64
2727 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2728 static int mdb_tls_nkeys;
2730 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2734 case DLL_PROCESS_ATTACH: break;
2735 case DLL_THREAD_ATTACH: break;
2736 case DLL_THREAD_DETACH:
2737 for (i=0; i<mdb_tls_nkeys; i++) {
2738 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2739 mdb_env_reader_dest(r);
2742 case DLL_PROCESS_DETACH: break;
2747 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2749 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2753 /* Force some symbol references.
2754 * _tls_used forces the linker to create the TLS directory if not already done
2755 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2757 #pragma comment(linker, "/INCLUDE:_tls_used")
2758 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2759 #pragma const_seg(".CRT$XLB")
2760 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2761 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2764 #pragma comment(linker, "/INCLUDE:__tls_used")
2765 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2766 #pragma data_seg(".CRT$XLB")
2767 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2769 #endif /* WIN 32/64 */
2770 #endif /* !__GNUC__ */
2773 /** Downgrade the exclusive lock on the region back to shared */
2775 mdb_env_share_locks(MDB_env *env, int *excl)
2777 int rc = 0, toggle = mdb_env_pick_meta(env);
2779 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2784 /* First acquire a shared lock. The Unlock will
2785 * then release the existing exclusive lock.
2787 memset(&ov, 0, sizeof(ov));
2788 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2791 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2797 struct flock lock_info;
2798 /* The shared lock replaces the existing lock */
2799 memset((void *)&lock_info, 0, sizeof(lock_info));
2800 lock_info.l_type = F_RDLCK;
2801 lock_info.l_whence = SEEK_SET;
2802 lock_info.l_start = 0;
2803 lock_info.l_len = 1;
2804 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2805 (rc = ErrCode()) == EINTR) ;
2806 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2813 /** Try to get exlusive lock, otherwise shared.
2814 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2817 mdb_env_excl_lock(MDB_env *env, int *excl)
2821 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2825 memset(&ov, 0, sizeof(ov));
2826 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2833 struct flock lock_info;
2834 memset((void *)&lock_info, 0, sizeof(lock_info));
2835 lock_info.l_type = F_WRLCK;
2836 lock_info.l_whence = SEEK_SET;
2837 lock_info.l_start = 0;
2838 lock_info.l_len = 1;
2839 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2840 (rc = ErrCode()) == EINTR) ;
2844 # ifdef MDB_USE_POSIX_SEM
2845 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
2848 lock_info.l_type = F_RDLCK;
2849 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
2850 (rc = ErrCode()) == EINTR) ;
2858 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
2860 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2862 * @(#) $Revision: 5.1 $
2863 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2864 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2866 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2870 * Please do not copyright this code. This code is in the public domain.
2872 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2873 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2874 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2875 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2876 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2877 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2878 * PERFORMANCE OF THIS SOFTWARE.
2881 * chongo <Landon Curt Noll> /\oo/\
2882 * http://www.isthe.com/chongo/
2884 * Share and Enjoy! :-)
2887 typedef unsigned long long mdb_hash_t;
2888 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2890 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2891 * @param[in] str string to hash
2892 * @param[in] hval initial value for hash
2893 * @return 64 bit hash
2895 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2896 * hval arg on the first call.
2899 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
2901 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
2902 unsigned char *end = s + val->mv_size;
2904 * FNV-1a hash each octet of the string
2907 /* xor the bottom with the current octet */
2908 hval ^= (mdb_hash_t)*s++;
2910 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2911 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2912 (hval << 7) + (hval << 8) + (hval << 40);
2914 /* return our new hash value */
2918 /** Hash the string and output the hash in hex.
2919 * @param[in] str string to hash
2920 * @param[out] hexbuf an array of 17 chars to hold the hash
2923 mdb_hash_hex(MDB_val *val, char *hexbuf)
2926 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
2927 for (i=0; i<8; i++) {
2928 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2934 /** Open and/or initialize the lock region for the environment.
2935 * @param[in] env The MDB environment.
2936 * @param[in] lpath The pathname of the file used for the lock region.
2937 * @param[in] mode The Unix permissions for the file, if we create it.
2938 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
2939 * @return 0 on success, non-zero on failure.
2942 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2950 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2951 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2952 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2954 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
2955 env->me_flags |= MDB_ROFS;
2960 /* Try to get exclusive lock. If we succeed, then
2961 * nobody is using the lock region and we should initialize it.
2963 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2964 size = GetFileSize(env->me_lfd, NULL);
2970 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2972 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
2973 env->me_flags |= MDB_ROFS;
2978 /* Lose record locks when exec*() */
2979 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
2980 fcntl(env->me_lfd, F_SETFD, fdflags);
2982 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
2983 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
2985 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
2986 env->me_flags |= MDB_ROFS;
2993 /* Try to get exclusive lock. If we succeed, then
2994 * nobody is using the lock region and we should initialize it.
2996 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2998 size = lseek(env->me_lfd, 0, SEEK_END);
3000 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3001 if (size < rsize && *excl > 0) {
3003 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3004 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3006 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3010 size = rsize - sizeof(MDB_txninfo);
3011 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3016 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3018 if (!mh) goto fail_errno;
3019 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3021 if (!env->me_txns) goto fail_errno;
3023 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3025 if (m == MAP_FAILED) goto fail_errno;
3031 BY_HANDLE_FILE_INFORMATION stbuf;
3040 if (!mdb_sec_inited) {
3041 InitializeSecurityDescriptor(&mdb_null_sd,
3042 SECURITY_DESCRIPTOR_REVISION);
3043 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3044 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3045 mdb_all_sa.bInheritHandle = FALSE;
3046 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3049 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3050 idbuf.volume = stbuf.dwVolumeSerialNumber;
3051 idbuf.nhigh = stbuf.nFileIndexHigh;
3052 idbuf.nlow = stbuf.nFileIndexLow;
3053 val.mv_data = &idbuf;
3054 val.mv_size = sizeof(idbuf);
3055 mdb_hash_hex(&val, hexbuf);
3056 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3057 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3058 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3059 if (!env->me_rmutex) goto fail_errno;
3060 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3061 if (!env->me_wmutex) goto fail_errno;
3062 #elif defined(MDB_USE_POSIX_SEM)
3071 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3072 idbuf.dev = stbuf.st_dev;
3073 idbuf.ino = stbuf.st_ino;
3074 val.mv_data = &idbuf;
3075 val.mv_size = sizeof(idbuf);
3076 mdb_hash_hex(&val, hexbuf);
3077 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3078 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3079 /* Clean up after a previous run, if needed: Try to
3080 * remove both semaphores before doing anything else.
3082 sem_unlink(env->me_txns->mti_rmname);
3083 sem_unlink(env->me_txns->mti_wmname);
3084 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3085 O_CREAT|O_EXCL, mode, 1);
3086 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3087 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3088 O_CREAT|O_EXCL, mode, 1);
3089 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3090 #else /* MDB_USE_POSIX_SEM */
3091 pthread_mutexattr_t mattr;
3093 if ((rc = pthread_mutexattr_init(&mattr))
3094 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3095 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3096 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3098 pthread_mutexattr_destroy(&mattr);
3099 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3101 env->me_txns->mti_version = MDB_VERSION;
3102 env->me_txns->mti_magic = MDB_MAGIC;
3103 env->me_txns->mti_txnid = 0;
3104 env->me_txns->mti_numreaders = 0;
3107 if (env->me_txns->mti_magic != MDB_MAGIC) {
3108 DPUTS("lock region has invalid magic");
3112 if (env->me_txns->mti_version != MDB_VERSION) {
3113 DPRINTF("lock region is version %u, expected version %u",
3114 env->me_txns->mti_version, MDB_VERSION);
3115 rc = MDB_VERSION_MISMATCH;
3119 if (rc != EACCES && rc != EAGAIN) {
3123 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3124 if (!env->me_rmutex) goto fail_errno;
3125 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3126 if (!env->me_wmutex) goto fail_errno;
3127 #elif defined(MDB_USE_POSIX_SEM)
3128 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3129 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3130 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3131 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3142 /** The name of the lock file in the DB environment */
3143 #define LOCKNAME "/lock.mdb"
3144 /** The name of the data file in the DB environment */
3145 #define DATANAME "/data.mdb"
3146 /** The suffix of the lock file when no subdir is used */
3147 #define LOCKSUFF "-lock"
3150 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3152 int oflags, rc, len, excl;
3153 char *lpath, *dpath;
3155 if (env->me_fd != INVALID_HANDLE_VALUE)
3159 if (flags & MDB_NOSUBDIR) {
3160 rc = len + sizeof(LOCKSUFF) + len + 1;
3162 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3167 if (flags & MDB_NOSUBDIR) {
3168 dpath = lpath + len + sizeof(LOCKSUFF);
3169 sprintf(lpath, "%s" LOCKSUFF, path);
3170 strcpy(dpath, path);
3172 dpath = lpath + len + sizeof(LOCKNAME);
3173 sprintf(lpath, "%s" LOCKNAME, path);
3174 sprintf(dpath, "%s" DATANAME, path);
3177 flags |= env->me_flags;
3178 /* silently ignore WRITEMAP if we're only getting read access */
3179 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3180 flags ^= MDB_WRITEMAP;
3181 env->me_flags = flags |= MDB_ENV_ACTIVE;
3183 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3188 if (F_ISSET(flags, MDB_RDONLY)) {
3189 oflags = GENERIC_READ;
3190 len = OPEN_EXISTING;
3192 oflags = GENERIC_READ|GENERIC_WRITE;
3195 mode = FILE_ATTRIBUTE_NORMAL;
3196 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3197 NULL, len, mode, NULL);
3199 if (F_ISSET(flags, MDB_RDONLY))
3202 oflags = O_RDWR | O_CREAT;
3204 env->me_fd = open(dpath, oflags, mode);
3206 if (env->me_fd == INVALID_HANDLE_VALUE) {
3211 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3212 if (flags & (MDB_RDONLY|MDB_NOSYNC|MDB_NOMETASYNC|MDB_WRITEMAP)) {
3213 env->me_mfd = env->me_fd;
3215 /* synchronous fd for meta writes */
3217 env->me_mfd = CreateFile(dpath, oflags,
3218 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3219 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3221 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3223 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3228 DPRINTF("opened dbenv %p", (void *) env);
3229 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3232 env->me_numdbs = 2; /* this notes that me_txkey was set */
3234 /* Windows TLS callbacks need help finding their TLS info. */
3235 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3236 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3243 rc = mdb_env_share_locks(env, &excl);
3247 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3248 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3249 env->me_path = strdup(path);
3250 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3256 mdb_env_close0(env, excl);
3262 /** Destroy resources from mdb_env_open() and clear our readers */
3264 mdb_env_close0(MDB_env *env, int excl)
3268 if (!(env->me_flags & MDB_ENV_ACTIVE))
3271 free(env->me_dbflags);
3275 if (env->me_numdbs) {
3276 pthread_key_delete(env->me_txkey);
3278 /* Delete our key from the global list */
3279 for (i=0; i<mdb_tls_nkeys; i++)
3280 if (mdb_tls_keys[i] == env->me_txkey) {
3281 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3289 munmap(env->me_map, env->me_mapsize);
3291 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3293 if (env->me_fd != INVALID_HANDLE_VALUE)
3296 pid_t pid = env->me_pid;
3297 /* Clearing readers is done in this function because
3298 * me_txkey with its destructor must be disabled first.
3300 for (i = env->me_numreaders; --i >= 0; )
3301 if (env->me_txns->mti_readers[i].mr_pid == pid)
3302 env->me_txns->mti_readers[i].mr_pid = 0;
3304 if (env->me_rmutex) {
3305 CloseHandle(env->me_rmutex);
3306 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3308 /* Windows automatically destroys the mutexes when
3309 * the last handle closes.
3311 #elif defined(MDB_USE_POSIX_SEM)
3312 if (env->me_rmutex != SEM_FAILED) {
3313 sem_close(env->me_rmutex);
3314 if (env->me_wmutex != SEM_FAILED)
3315 sem_close(env->me_wmutex);
3316 /* If we have the filelock: If we are the
3317 * only remaining user, clean up semaphores.
3320 mdb_env_excl_lock(env, &excl);
3322 sem_unlink(env->me_txns->mti_rmname);
3323 sem_unlink(env->me_txns->mti_wmname);
3327 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3329 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3332 /* Unlock the lockfile. Windows would have unlocked it
3333 * after closing anyway, but not necessarily at once.
3335 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3341 env->me_flags &= ~MDB_ENV_ACTIVE;
3345 mdb_env_copy(MDB_env *env, const char *path)
3347 MDB_txn *txn = NULL;
3351 HANDLE newfd = INVALID_HANDLE_VALUE;
3353 if (env->me_flags & MDB_NOSUBDIR) {
3354 lpath = (char *)path;
3357 len += sizeof(DATANAME);
3358 lpath = malloc(len);
3361 sprintf(lpath, "%s" DATANAME, path);
3364 /* The destination path must exist, but the destination file must not.
3365 * We don't want the OS to cache the writes, since the source data is
3366 * already in the OS cache.
3369 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3370 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3372 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3378 if (!(env->me_flags & MDB_NOSUBDIR))
3380 if (newfd == INVALID_HANDLE_VALUE) {
3385 #ifdef F_NOCACHE /* __APPLE__ */
3386 rc = fcntl(newfd, F_NOCACHE, 1);
3393 /* Do the lock/unlock of the reader mutex before starting the
3394 * write txn. Otherwise other read txns could block writers.
3396 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3400 if (!(env->me_flags & MDB_ROFS)) {
3401 /* We must start the actual read txn after blocking writers */
3402 mdb_txn_reset0(txn);
3404 /* Temporarily block writers until we snapshot the meta pages */
3407 rc = mdb_txn_renew0(txn);
3409 UNLOCK_MUTEX_W(env);
3414 wsize = env->me_psize * 2;
3418 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3419 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3422 rc = write(newfd, env->me_map, wsize);
3423 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3425 if (! (env->me_flags & MDB_ROFS))
3426 UNLOCK_MUTEX_W(env);
3431 ptr = env->me_map + wsize;
3432 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3434 #define MAX_UINT32 4294967295U
3437 if (wsize > MAX_UINT32)
3438 w2 = MAX_UINT32 - env->me_psize + 1; /* write in pagesize chunks */
3441 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3442 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3447 rc = write(newfd, ptr, wsize);
3448 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3453 if (newfd != INVALID_HANDLE_VALUE)
3460 mdb_env_close(MDB_env *env)
3467 VGMEMP_DESTROY(env);
3468 while ((dp = env->me_dpages) != NULL) {
3469 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3470 env->me_dpages = dp->mp_next;
3474 mdb_env_close0(env, 0);
3475 mdb_midl_free(env->me_free_pgs);
3479 /** Compare two items pointing at aligned size_t's */
3481 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3483 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3484 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3487 /** Compare two items pointing at aligned int's */
3489 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3491 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3492 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3495 /** Compare two items pointing at ints of unknown alignment.
3496 * Nodes and keys are guaranteed to be 2-byte aligned.
3499 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3501 #if BYTE_ORDER == LITTLE_ENDIAN
3502 unsigned short *u, *c;
3505 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3506 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3509 } while(!x && u > (unsigned short *)a->mv_data);
3512 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3516 /** Compare two items lexically */
3518 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3525 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3531 diff = memcmp(a->mv_data, b->mv_data, len);
3532 return diff ? diff : len_diff<0 ? -1 : len_diff;
3535 /** Compare two items in reverse byte order */
3537 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3539 const unsigned char *p1, *p2, *p1_lim;
3543 p1_lim = (const unsigned char *)a->mv_data;
3544 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3545 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3547 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3553 while (p1 > p1_lim) {
3554 diff = *--p1 - *--p2;
3558 return len_diff<0 ? -1 : len_diff;
3561 /** Search for key within a page, using binary search.
3562 * Returns the smallest entry larger or equal to the key.
3563 * If exactp is non-null, stores whether the found entry was an exact match
3564 * in *exactp (1 or 0).
3565 * Updates the cursor index with the index of the found entry.
3566 * If no entry larger or equal to the key is found, returns NULL.
3569 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3571 unsigned int i = 0, nkeys;
3574 MDB_page *mp = mc->mc_pg[mc->mc_top];
3575 MDB_node *node = NULL;
3580 nkeys = NUMKEYS(mp);
3585 COPY_PGNO(pgno, mp->mp_pgno);
3586 DPRINTF("searching %u keys in %s %spage %zu",
3587 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3594 low = IS_LEAF(mp) ? 0 : 1;
3596 cmp = mc->mc_dbx->md_cmp;
3598 /* Branch pages have no data, so if using integer keys,
3599 * alignment is guaranteed. Use faster mdb_cmp_int.
3601 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3602 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3609 nodekey.mv_size = mc->mc_db->md_pad;
3610 node = NODEPTR(mp, 0); /* fake */
3611 while (low <= high) {
3612 i = (low + high) >> 1;
3613 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3614 rc = cmp(key, &nodekey);
3615 DPRINTF("found leaf index %u [%s], rc = %i",
3616 i, DKEY(&nodekey), rc);
3625 while (low <= high) {
3626 i = (low + high) >> 1;
3628 node = NODEPTR(mp, i);
3629 nodekey.mv_size = NODEKSZ(node);
3630 nodekey.mv_data = NODEKEY(node);
3632 rc = cmp(key, &nodekey);
3635 DPRINTF("found leaf index %u [%s], rc = %i",
3636 i, DKEY(&nodekey), rc);
3638 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3639 i, DKEY(&nodekey), NODEPGNO(node), rc);
3650 if (rc > 0) { /* Found entry is less than the key. */
3651 i++; /* Skip to get the smallest entry larger than key. */
3653 node = NODEPTR(mp, i);
3656 *exactp = (rc == 0);
3657 /* store the key index */
3658 mc->mc_ki[mc->mc_top] = i;
3660 /* There is no entry larger or equal to the key. */
3663 /* nodeptr is fake for LEAF2 */
3669 mdb_cursor_adjust(MDB_cursor *mc, func)
3673 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3674 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3681 /** Pop a page off the top of the cursor's stack. */
3683 mdb_cursor_pop(MDB_cursor *mc)
3691 MDB_page *top = mc->mc_pg[mc->mc_top];
3692 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3693 mc->mc_dbi, (void *) mc);
3699 /** Push a page onto the top of the cursor's stack. */
3701 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3703 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3704 mc->mc_dbi, (void *) mc);
3706 if (mc->mc_snum >= CURSOR_STACK) {
3707 assert(mc->mc_snum < CURSOR_STACK);
3708 return MDB_CURSOR_FULL;
3711 mc->mc_top = mc->mc_snum++;
3712 mc->mc_pg[mc->mc_top] = mp;
3713 mc->mc_ki[mc->mc_top] = 0;
3718 /** Find the address of the page corresponding to a given page number.
3719 * @param[in] txn the transaction for this access.
3720 * @param[in] pgno the page number for the page to retrieve.
3721 * @param[out] ret address of a pointer where the page's address will be stored.
3722 * @return 0 on success, non-zero on failure.
3725 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3729 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3730 if (pgno < txn->mt_next_pgno)
3731 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3734 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3736 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3737 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3738 p = txn->mt_u.dirty_list[x].mptr;
3742 if (pgno < txn->mt_next_pgno)
3743 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3748 DPRINTF("page %zu not found", pgno);
3751 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3754 /** Search for the page a given key should be in.
3755 * Pushes parent pages on the cursor stack. This function continues a
3756 * search on a cursor that has already been initialized. (Usually by
3757 * #mdb_page_search() but also by #mdb_node_move().)
3758 * @param[in,out] mc the cursor for this operation.
3759 * @param[in] key the key to search for. If NULL, search for the lowest
3760 * page. (This is used by #mdb_cursor_first().)
3761 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3762 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3763 * @return 0 on success, non-zero on failure.
3766 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3768 MDB_page *mp = mc->mc_pg[mc->mc_top];
3773 while (IS_BRANCH(mp)) {
3777 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3778 assert(NUMKEYS(mp) > 1);
3779 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3781 if (key == NULL) /* Initialize cursor to first page. */
3783 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3784 /* cursor to last page */
3788 node = mdb_node_search(mc, key, &exact);
3790 i = NUMKEYS(mp) - 1;
3792 i = mc->mc_ki[mc->mc_top];
3801 DPRINTF("following index %u for key [%s]",
3803 assert(i < NUMKEYS(mp));
3804 node = NODEPTR(mp, i);
3806 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3809 mc->mc_ki[mc->mc_top] = i;
3810 if ((rc = mdb_cursor_push(mc, mp)))
3814 if ((rc = mdb_page_touch(mc)) != 0)
3816 mp = mc->mc_pg[mc->mc_top];
3821 DPRINTF("internal error, index points to a %02X page!?",
3823 return MDB_CORRUPTED;
3826 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3827 key ? DKEY(key) : NULL);
3832 /** Search for the page a given key should be in.
3833 * Pushes parent pages on the cursor stack. This function just sets up
3834 * the search; it finds the root page for \b mc's database and sets this
3835 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3836 * called to complete the search.
3837 * @param[in,out] mc the cursor for this operation.
3838 * @param[in] key the key to search for. If NULL, search for the lowest
3839 * page. (This is used by #mdb_cursor_first().)
3840 * @param[in] modify If true, visited pages are updated with new page numbers.
3841 * @return 0 on success, non-zero on failure.
3844 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3849 /* Make sure the txn is still viable, then find the root from
3850 * the txn's db table.
3852 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3853 DPUTS("transaction has failed, must abort");
3856 /* Make sure we're using an up-to-date root */
3857 if (mc->mc_dbi > MAIN_DBI) {
3858 if ((*mc->mc_dbflag & DB_STALE) ||
3859 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3861 unsigned char dbflag = 0;
3862 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3863 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3866 if (*mc->mc_dbflag & DB_STALE) {
3869 MDB_node *leaf = mdb_node_search(&mc2,
3870 &mc->mc_dbx->md_name, &exact);
3872 return MDB_NOTFOUND;
3873 mdb_node_read(mc->mc_txn, leaf, &data);
3874 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3876 if (flags & MDB_PS_MODIFY)
3878 *mc->mc_dbflag = dbflag;
3881 root = mc->mc_db->md_root;
3883 if (root == P_INVALID) { /* Tree is empty. */
3884 DPUTS("tree is empty");
3885 return MDB_NOTFOUND;
3890 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
3891 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3897 DPRINTF("db %u root page %zu has flags 0x%X",
3898 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3900 if (flags & MDB_PS_MODIFY) {
3901 if ((rc = mdb_page_touch(mc)))
3905 if (flags & MDB_PS_ROOTONLY)
3908 return mdb_page_search_root(mc, key, flags);
3911 /** Return the data associated with a given node.
3912 * @param[in] txn The transaction for this operation.
3913 * @param[in] leaf The node being read.
3914 * @param[out] data Updated to point to the node's data.
3915 * @return 0 on success, non-zero on failure.
3918 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3920 MDB_page *omp; /* overflow page */
3924 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3925 data->mv_size = NODEDSZ(leaf);
3926 data->mv_data = NODEDATA(leaf);
3930 /* Read overflow data.
3932 data->mv_size = NODEDSZ(leaf);
3933 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3934 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3935 DPRINTF("read overflow page %zu failed", pgno);
3938 data->mv_data = METADATA(omp);
3944 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3945 MDB_val *key, MDB_val *data)
3954 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3956 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3959 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3963 mdb_cursor_init(&mc, txn, dbi, &mx);
3964 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3967 /** Find a sibling for a page.
3968 * Replaces the page at the top of the cursor's stack with the
3969 * specified sibling, if one exists.
3970 * @param[in] mc The cursor for this operation.
3971 * @param[in] move_right Non-zero if the right sibling is requested,
3972 * otherwise the left sibling.
3973 * @return 0 on success, non-zero on failure.
3976 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3982 if (mc->mc_snum < 2) {
3983 return MDB_NOTFOUND; /* root has no siblings */
3987 DPRINTF("parent page is page %zu, index %u",
3988 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3990 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3991 : (mc->mc_ki[mc->mc_top] == 0)) {
3992 DPRINTF("no more keys left, moving to %s sibling",
3993 move_right ? "right" : "left");
3994 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
3995 /* undo cursor_pop before returning */
4002 mc->mc_ki[mc->mc_top]++;
4004 mc->mc_ki[mc->mc_top]--;
4005 DPRINTF("just moving to %s index key %u",
4006 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4008 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4010 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4011 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4014 mdb_cursor_push(mc, mp);
4019 /** Move the cursor to the next data item. */
4021 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4027 if (mc->mc_flags & C_EOF) {
4028 return MDB_NOTFOUND;
4031 assert(mc->mc_flags & C_INITIALIZED);
4033 mp = mc->mc_pg[mc->mc_top];
4035 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4036 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4037 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4038 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4039 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4040 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4044 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4045 if (op == MDB_NEXT_DUP)
4046 return MDB_NOTFOUND;
4050 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4052 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4053 DPUTS("=====> move to next sibling page");
4054 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4055 mc->mc_flags |= C_EOF;
4056 mc->mc_flags &= ~C_INITIALIZED;
4057 return MDB_NOTFOUND;
4059 mp = mc->mc_pg[mc->mc_top];
4060 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4062 mc->mc_ki[mc->mc_top]++;
4064 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4065 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4068 key->mv_size = mc->mc_db->md_pad;
4069 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4073 assert(IS_LEAF(mp));
4074 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4076 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4077 mdb_xcursor_init1(mc, leaf);
4080 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4083 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4084 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4085 if (rc != MDB_SUCCESS)
4090 MDB_GET_KEY(leaf, key);
4094 /** Move the cursor to the previous data item. */
4096 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4102 assert(mc->mc_flags & C_INITIALIZED);
4104 mp = mc->mc_pg[mc->mc_top];
4106 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4107 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4108 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4109 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4110 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4111 if (op != MDB_PREV || rc == MDB_SUCCESS)
4114 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4115 if (op == MDB_PREV_DUP)
4116 return MDB_NOTFOUND;
4121 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4123 if (mc->mc_ki[mc->mc_top] == 0) {
4124 DPUTS("=====> move to prev sibling page");
4125 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4126 mc->mc_flags &= ~C_INITIALIZED;
4127 return MDB_NOTFOUND;
4129 mp = mc->mc_pg[mc->mc_top];
4130 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4131 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4133 mc->mc_ki[mc->mc_top]--;
4135 mc->mc_flags &= ~C_EOF;
4137 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4138 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4141 key->mv_size = mc->mc_db->md_pad;
4142 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4146 assert(IS_LEAF(mp));
4147 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4149 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4150 mdb_xcursor_init1(mc, leaf);
4153 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4156 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4157 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4158 if (rc != MDB_SUCCESS)
4163 MDB_GET_KEY(leaf, key);
4167 /** Set the cursor on a specific data item. */
4169 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4170 MDB_cursor_op op, int *exactp)
4174 MDB_node *leaf = NULL;
4179 assert(key->mv_size > 0);
4181 /* See if we're already on the right page */
4182 if (mc->mc_flags & C_INITIALIZED) {
4185 mp = mc->mc_pg[mc->mc_top];
4187 mc->mc_ki[mc->mc_top] = 0;
4188 return MDB_NOTFOUND;
4190 if (mp->mp_flags & P_LEAF2) {
4191 nodekey.mv_size = mc->mc_db->md_pad;
4192 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4194 leaf = NODEPTR(mp, 0);
4195 MDB_GET_KEY(leaf, &nodekey);
4197 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4199 /* Probably happens rarely, but first node on the page
4200 * was the one we wanted.
4202 mc->mc_ki[mc->mc_top] = 0;
4209 unsigned int nkeys = NUMKEYS(mp);
4211 if (mp->mp_flags & P_LEAF2) {
4212 nodekey.mv_data = LEAF2KEY(mp,
4213 nkeys-1, nodekey.mv_size);
4215 leaf = NODEPTR(mp, nkeys-1);
4216 MDB_GET_KEY(leaf, &nodekey);
4218 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4220 /* last node was the one we wanted */
4221 mc->mc_ki[mc->mc_top] = nkeys-1;
4227 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4228 /* This is definitely the right page, skip search_page */
4229 if (mp->mp_flags & P_LEAF2) {
4230 nodekey.mv_data = LEAF2KEY(mp,
4231 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4233 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4234 MDB_GET_KEY(leaf, &nodekey);
4236 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4238 /* current node was the one we wanted */
4248 /* If any parents have right-sibs, search.
4249 * Otherwise, there's nothing further.
4251 for (i=0; i<mc->mc_top; i++)
4253 NUMKEYS(mc->mc_pg[i])-1)
4255 if (i == mc->mc_top) {
4256 /* There are no other pages */
4257 mc->mc_ki[mc->mc_top] = nkeys;
4258 return MDB_NOTFOUND;
4262 /* There are no other pages */
4263 mc->mc_ki[mc->mc_top] = 0;
4264 return MDB_NOTFOUND;
4268 rc = mdb_page_search(mc, key, 0);
4269 if (rc != MDB_SUCCESS)
4272 mp = mc->mc_pg[mc->mc_top];
4273 assert(IS_LEAF(mp));
4276 leaf = mdb_node_search(mc, key, exactp);
4277 if (exactp != NULL && !*exactp) {
4278 /* MDB_SET specified and not an exact match. */
4279 return MDB_NOTFOUND;
4283 DPUTS("===> inexact leaf not found, goto sibling");
4284 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4285 return rc; /* no entries matched */
4286 mp = mc->mc_pg[mc->mc_top];
4287 assert(IS_LEAF(mp));
4288 leaf = NODEPTR(mp, 0);
4292 mc->mc_flags |= C_INITIALIZED;
4293 mc->mc_flags &= ~C_EOF;
4296 key->mv_size = mc->mc_db->md_pad;
4297 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4301 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4302 mdb_xcursor_init1(mc, leaf);
4305 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4306 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4307 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4310 if (op == MDB_GET_BOTH) {
4316 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4317 if (rc != MDB_SUCCESS)
4320 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4322 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4324 rc = mc->mc_dbx->md_dcmp(data, &d2);
4326 if (op == MDB_GET_BOTH || rc > 0)
4327 return MDB_NOTFOUND;
4332 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4333 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4338 /* The key already matches in all other cases */
4339 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4340 MDB_GET_KEY(leaf, key);
4341 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4346 /** Move the cursor to the first item in the database. */
4348 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4353 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4354 rc = mdb_page_search(mc, NULL, 0);
4355 if (rc != MDB_SUCCESS)
4358 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4360 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4361 mc->mc_flags |= C_INITIALIZED;
4362 mc->mc_flags &= ~C_EOF;
4364 mc->mc_ki[mc->mc_top] = 0;
4366 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4367 key->mv_size = mc->mc_db->md_pad;
4368 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4373 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4374 mdb_xcursor_init1(mc, leaf);
4375 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4380 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4381 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4385 MDB_GET_KEY(leaf, key);
4389 /** Move the cursor to the last item in the database. */
4391 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4396 if (!(mc->mc_flags & C_EOF)) {
4398 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4401 lkey.mv_size = MAXKEYSIZE+1;
4402 lkey.mv_data = NULL;
4403 rc = mdb_page_search(mc, &lkey, 0);
4404 if (rc != MDB_SUCCESS)
4407 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4409 mc->mc_flags |= C_INITIALIZED|C_EOF;
4410 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4412 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4414 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4415 key->mv_size = mc->mc_db->md_pad;
4416 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4421 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4422 mdb_xcursor_init1(mc, leaf);
4423 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4428 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4429 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4434 MDB_GET_KEY(leaf, key);
4439 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4448 case MDB_GET_CURRENT:
4449 if (!mc->mc_flags & C_INITIALIZED) {
4452 MDB_page *mp = mc->mc_pg[mc->mc_top];
4454 mc->mc_ki[mc->mc_top] = 0;
4460 key->mv_size = mc->mc_db->md_pad;
4461 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4463 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4464 MDB_GET_KEY(leaf, key);
4466 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4467 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4469 rc = mdb_node_read(mc->mc_txn, leaf, data);
4476 case MDB_GET_BOTH_RANGE:
4477 if (data == NULL || mc->mc_xcursor == NULL) {
4485 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4487 } else if (op == MDB_SET_RANGE)
4488 rc = mdb_cursor_set(mc, key, data, op, NULL);
4490 rc = mdb_cursor_set(mc, key, data, op, &exact);
4492 case MDB_GET_MULTIPLE:
4494 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4495 !(mc->mc_flags & C_INITIALIZED)) {
4500 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4501 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4504 case MDB_NEXT_MULTIPLE:
4506 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4510 if (!(mc->mc_flags & C_INITIALIZED))
4511 rc = mdb_cursor_first(mc, key, data);
4513 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4514 if (rc == MDB_SUCCESS) {
4515 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4518 mx = &mc->mc_xcursor->mx_cursor;
4519 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4521 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4522 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4530 case MDB_NEXT_NODUP:
4531 if (!(mc->mc_flags & C_INITIALIZED))
4532 rc = mdb_cursor_first(mc, key, data);
4534 rc = mdb_cursor_next(mc, key, data, op);
4538 case MDB_PREV_NODUP:
4539 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4540 rc = mdb_cursor_last(mc, key, data);
4541 mc->mc_flags |= C_INITIALIZED;
4542 mc->mc_ki[mc->mc_top]++;
4544 rc = mdb_cursor_prev(mc, key, data, op);
4547 rc = mdb_cursor_first(mc, key, data);
4551 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4552 !(mc->mc_flags & C_INITIALIZED) ||
4553 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4557 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4560 rc = mdb_cursor_last(mc, key, data);
4564 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4565 !(mc->mc_flags & C_INITIALIZED) ||
4566 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4570 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4573 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4581 /** Touch all the pages in the cursor stack.
4582 * Makes sure all the pages are writable, before attempting a write operation.
4583 * @param[in] mc The cursor to operate on.
4586 mdb_cursor_touch(MDB_cursor *mc)
4590 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4593 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4594 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4595 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4598 *mc->mc_dbflag = DB_DIRTY;
4600 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4601 rc = mdb_page_touch(mc);
4605 mc->mc_top = mc->mc_snum-1;
4610 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4613 MDB_node *leaf = NULL;
4614 MDB_val xdata, *rdata, dkey;
4617 int do_sub = 0, insert = 0;
4618 unsigned int mcount = 0;
4622 char dbuf[MAXKEYSIZE+1];
4623 unsigned int nflags;
4626 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4629 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4630 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4634 if (flags == MDB_CURRENT) {
4635 if (!(mc->mc_flags & C_INITIALIZED))
4638 } else if (mc->mc_db->md_root == P_INVALID) {
4640 /* new database, write a root leaf page */
4641 DPUTS("allocating new root leaf page");
4642 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4646 mdb_cursor_push(mc, np);
4647 mc->mc_db->md_root = np->mp_pgno;
4648 mc->mc_db->md_depth++;
4649 *mc->mc_dbflag = DB_DIRTY;
4650 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4652 np->mp_flags |= P_LEAF2;
4653 mc->mc_flags |= C_INITIALIZED;
4659 if (flags & MDB_APPEND) {
4661 rc = mdb_cursor_last(mc, &k2, &d2);
4663 rc = mc->mc_dbx->md_cmp(key, &k2);
4666 mc->mc_ki[mc->mc_top]++;
4672 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4674 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4675 DPRINTF("duplicate key [%s]", DKEY(key));
4677 return MDB_KEYEXIST;
4679 if (rc && rc != MDB_NOTFOUND)
4683 /* Cursor is positioned, now make sure all pages are writable */
4684 rc2 = mdb_cursor_touch(mc);
4689 /* The key already exists */
4690 if (rc == MDB_SUCCESS) {
4691 /* there's only a key anyway, so this is a no-op */
4692 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4693 unsigned int ksize = mc->mc_db->md_pad;
4694 if (key->mv_size != ksize)
4696 if (flags == MDB_CURRENT) {
4697 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4698 memcpy(ptr, key->mv_data, ksize);
4703 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4706 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4707 /* Was a single item before, must convert now */
4709 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4710 /* Just overwrite the current item */
4711 if (flags == MDB_CURRENT)
4714 dkey.mv_size = NODEDSZ(leaf);
4715 dkey.mv_data = NODEDATA(leaf);
4716 #if UINT_MAX < SIZE_MAX
4717 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4718 #ifdef MISALIGNED_OK
4719 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4721 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4724 /* if data matches, ignore it */
4725 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4726 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4728 /* create a fake page for the dup items */
4729 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4730 dkey.mv_data = dbuf;
4731 fp = (MDB_page *)&pbuf;
4732 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4733 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4734 fp->mp_lower = PAGEHDRSZ;
4735 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4736 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4737 fp->mp_flags |= P_LEAF2;
4738 fp->mp_pad = data->mv_size;
4739 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4741 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4742 (dkey.mv_size & 1) + (data->mv_size & 1);
4744 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4747 xdata.mv_size = fp->mp_upper;
4752 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4753 /* See if we need to convert from fake page to subDB */
4755 unsigned int offset;
4758 fp = NODEDATA(leaf);
4759 if (flags == MDB_CURRENT) {
4761 fp->mp_flags |= P_DIRTY;
4762 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4763 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4767 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4768 offset = fp->mp_pad;
4769 if (SIZELEFT(fp) >= offset)
4771 offset *= 4; /* space for 4 more */
4773 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4775 offset += offset & 1;
4776 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4777 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4779 /* yes, convert it */
4781 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4782 dummy.md_pad = fp->mp_pad;
4783 dummy.md_flags = MDB_DUPFIXED;
4784 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4785 dummy.md_flags |= MDB_INTEGERKEY;
4788 dummy.md_branch_pages = 0;
4789 dummy.md_leaf_pages = 1;
4790 dummy.md_overflow_pages = 0;
4791 dummy.md_entries = NUMKEYS(fp);
4793 xdata.mv_size = sizeof(MDB_db);
4794 xdata.mv_data = &dummy;
4795 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4797 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4798 flags |= F_DUPDATA|F_SUBDATA;
4799 dummy.md_root = mp->mp_pgno;
4801 /* no, just grow it */
4803 xdata.mv_size = NODEDSZ(leaf) + offset;
4804 xdata.mv_data = &pbuf;
4805 mp = (MDB_page *)&pbuf;
4806 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4809 mp->mp_flags = fp->mp_flags | P_DIRTY;
4810 mp->mp_pad = fp->mp_pad;
4811 mp->mp_lower = fp->mp_lower;
4812 mp->mp_upper = fp->mp_upper + offset;
4814 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4816 nsize = NODEDSZ(leaf) - fp->mp_upper;
4817 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4818 for (i=0; i<NUMKEYS(fp); i++)
4819 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4821 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4825 /* data is on sub-DB, just store it */
4826 flags |= F_DUPDATA|F_SUBDATA;
4830 /* overflow page overwrites need special handling */
4831 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4834 int ovpages, dpages;
4836 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4837 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4838 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4839 mdb_page_get(mc->mc_txn, pg, &omp);
4840 /* Is the ov page writable and large enough? */
4841 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4842 /* yes, overwrite it. Note in this case we don't
4843 * bother to try shrinking the node if the new data
4844 * is smaller than the overflow threshold.
4846 if (F_ISSET(flags, MDB_RESERVE))
4847 data->mv_data = METADATA(omp);
4849 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4852 /* no, free ovpages */
4854 mc->mc_db->md_overflow_pages -= ovpages;
4855 for (i=0; i<ovpages; i++) {
4856 DPRINTF("freed ov page %zu", pg);
4857 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4861 } else if (NODEDSZ(leaf) == data->mv_size) {
4862 /* same size, just replace it. Note that we could
4863 * also reuse this node if the new data is smaller,
4864 * but instead we opt to shrink the node in that case.
4866 if (F_ISSET(flags, MDB_RESERVE))
4867 data->mv_data = NODEDATA(leaf);
4869 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4872 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4873 mc->mc_db->md_entries--;
4875 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4882 nflags = flags & NODE_ADD_FLAGS;
4883 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4884 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4885 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4886 nflags &= ~MDB_APPEND;
4888 nflags |= MDB_SPLIT_REPLACE;
4889 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4891 /* There is room already in this leaf page. */
4892 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4893 if (rc == 0 && !do_sub && insert) {
4894 /* Adjust other cursors pointing to mp */
4895 MDB_cursor *m2, *m3;
4896 MDB_dbi dbi = mc->mc_dbi;
4897 unsigned i = mc->mc_top;
4898 MDB_page *mp = mc->mc_pg[i];
4900 if (mc->mc_flags & C_SUB)
4903 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4904 if (mc->mc_flags & C_SUB)
4905 m3 = &m2->mc_xcursor->mx_cursor;
4908 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4909 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4916 if (rc != MDB_SUCCESS)
4917 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4919 /* Now store the actual data in the child DB. Note that we're
4920 * storing the user data in the keys field, so there are strict
4921 * size limits on dupdata. The actual data fields of the child
4922 * DB are all zero size.
4929 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4930 if (flags & MDB_CURRENT) {
4931 xflags = MDB_CURRENT;
4933 mdb_xcursor_init1(mc, leaf);
4934 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4936 /* converted, write the original data first */
4938 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4942 /* Adjust other cursors pointing to mp */
4944 unsigned i = mc->mc_top;
4945 MDB_page *mp = mc->mc_pg[i];
4947 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4948 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4949 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4950 mdb_xcursor_init1(m2, leaf);
4955 if (flags & MDB_APPENDDUP)
4956 xflags |= MDB_APPEND;
4957 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4958 if (flags & F_SUBDATA) {
4959 void *db = NODEDATA(leaf);
4960 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4963 /* sub-writes might have failed so check rc again.
4964 * Don't increment count if we just replaced an existing item.
4966 if (!rc && !(flags & MDB_CURRENT))
4967 mc->mc_db->md_entries++;
4968 if (flags & MDB_MULTIPLE) {
4970 if (mcount < data[1].mv_size) {
4971 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4972 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4982 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4987 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4990 if (!mc->mc_flags & C_INITIALIZED)
4993 rc = mdb_cursor_touch(mc);
4997 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4999 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5000 if (flags != MDB_NODUPDATA) {
5001 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5002 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5004 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5005 /* If sub-DB still has entries, we're done */
5006 if (mc->mc_xcursor->mx_db.md_entries) {
5007 if (leaf->mn_flags & F_SUBDATA) {
5008 /* update subDB info */
5009 void *db = NODEDATA(leaf);
5010 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5012 /* shrink fake page */
5013 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5015 mc->mc_db->md_entries--;
5018 /* otherwise fall thru and delete the sub-DB */
5021 if (leaf->mn_flags & F_SUBDATA) {
5022 /* add all the child DB's pages to the free list */
5023 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5024 if (rc == MDB_SUCCESS) {
5025 mc->mc_db->md_entries -=
5026 mc->mc_xcursor->mx_db.md_entries;
5031 return mdb_cursor_del0(mc, leaf);
5034 /** Allocate and initialize new pages for a database.
5035 * @param[in] mc a cursor on the database being added to.
5036 * @param[in] flags flags defining what type of page is being allocated.
5037 * @param[in] num the number of pages to allocate. This is usually 1,
5038 * unless allocating overflow pages for a large record.
5039 * @param[out] mp Address of a page, or NULL on failure.
5040 * @return 0 on success, non-zero on failure.
5043 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5048 if ((rc = mdb_page_alloc(mc, num, &np)))
5050 DPRINTF("allocated new mpage %zu, page size %u",
5051 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5052 np->mp_flags = flags | P_DIRTY;
5053 np->mp_lower = PAGEHDRSZ;
5054 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5057 mc->mc_db->md_branch_pages++;
5058 else if (IS_LEAF(np))
5059 mc->mc_db->md_leaf_pages++;
5060 else if (IS_OVERFLOW(np)) {
5061 mc->mc_db->md_overflow_pages += num;
5069 /** Calculate the size of a leaf node.
5070 * The size depends on the environment's page size; if a data item
5071 * is too large it will be put onto an overflow page and the node
5072 * size will only include the key and not the data. Sizes are always
5073 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5074 * of the #MDB_node headers.
5075 * @param[in] env The environment handle.
5076 * @param[in] key The key for the node.
5077 * @param[in] data The data for the node.
5078 * @return The number of bytes needed to store the node.
5081 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5085 sz = LEAFSIZE(key, data);
5086 if (sz >= env->me_psize / MDB_MINKEYS) {
5087 /* put on overflow page */
5088 sz -= data->mv_size - sizeof(pgno_t);
5092 return sz + sizeof(indx_t);
5095 /** Calculate the size of a branch node.
5096 * The size should depend on the environment's page size but since
5097 * we currently don't support spilling large keys onto overflow
5098 * pages, it's simply the size of the #MDB_node header plus the
5099 * size of the key. Sizes are always rounded up to an even number
5100 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5101 * @param[in] env The environment handle.
5102 * @param[in] key The key for the node.
5103 * @return The number of bytes needed to store the node.
5106 mdb_branch_size(MDB_env *env, MDB_val *key)
5111 if (sz >= env->me_psize / MDB_MINKEYS) {
5112 /* put on overflow page */
5113 /* not implemented */
5114 /* sz -= key->size - sizeof(pgno_t); */
5117 return sz + sizeof(indx_t);
5120 /** Add a node to the page pointed to by the cursor.
5121 * @param[in] mc The cursor for this operation.
5122 * @param[in] indx The index on the page where the new node should be added.
5123 * @param[in] key The key for the new node.
5124 * @param[in] data The data for the new node, if any.
5125 * @param[in] pgno The page number, if adding a branch node.
5126 * @param[in] flags Flags for the node.
5127 * @return 0 on success, non-zero on failure. Possible errors are:
5129 * <li>ENOMEM - failed to allocate overflow pages for the node.
5130 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5131 * should never happen since all callers already calculate the
5132 * page's free space before calling this function.
5136 mdb_node_add(MDB_cursor *mc, indx_t indx,
5137 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5140 size_t node_size = NODESIZE;
5143 MDB_page *mp = mc->mc_pg[mc->mc_top];
5144 MDB_page *ofp = NULL; /* overflow page */
5147 assert(mp->mp_upper >= mp->mp_lower);
5149 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5150 IS_LEAF(mp) ? "leaf" : "branch",
5151 IS_SUBP(mp) ? "sub-" : "",
5152 mp->mp_pgno, indx, data ? data->mv_size : 0,
5153 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5156 /* Move higher keys up one slot. */
5157 int ksize = mc->mc_db->md_pad, dif;
5158 char *ptr = LEAF2KEY(mp, indx, ksize);
5159 dif = NUMKEYS(mp) - indx;
5161 memmove(ptr+ksize, ptr, dif*ksize);
5162 /* insert new key */
5163 memcpy(ptr, key->mv_data, ksize);
5165 /* Just using these for counting */
5166 mp->mp_lower += sizeof(indx_t);
5167 mp->mp_upper -= ksize - sizeof(indx_t);
5172 node_size += key->mv_size;
5176 if (F_ISSET(flags, F_BIGDATA)) {
5177 /* Data already on overflow page. */
5178 node_size += sizeof(pgno_t);
5179 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5180 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5182 /* Put data on overflow page. */
5183 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5184 data->mv_size, node_size+data->mv_size);
5185 node_size += sizeof(pgno_t);
5186 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5188 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5191 node_size += data->mv_size;
5194 node_size += node_size & 1;
5196 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5197 DPRINTF("not enough room in page %zu, got %u ptrs",
5198 mp->mp_pgno, NUMKEYS(mp));
5199 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5200 mp->mp_upper - mp->mp_lower);
5201 DPRINTF("node size = %zu", node_size);
5202 return MDB_PAGE_FULL;
5205 /* Move higher pointers up one slot. */
5206 for (i = NUMKEYS(mp); i > indx; i--)
5207 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5209 /* Adjust free space offsets. */
5210 ofs = mp->mp_upper - node_size;
5211 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5212 mp->mp_ptrs[indx] = ofs;
5214 mp->mp_lower += sizeof(indx_t);
5216 /* Write the node data. */
5217 node = NODEPTR(mp, indx);
5218 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5219 node->mn_flags = flags;
5221 SETDSZ(node,data->mv_size);
5226 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5231 if (F_ISSET(flags, F_BIGDATA))
5232 memcpy(node->mn_data + key->mv_size, data->mv_data,
5234 else if (F_ISSET(flags, MDB_RESERVE))
5235 data->mv_data = node->mn_data + key->mv_size;
5237 memcpy(node->mn_data + key->mv_size, data->mv_data,
5240 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5242 if (F_ISSET(flags, MDB_RESERVE))
5243 data->mv_data = METADATA(ofp);
5245 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5252 /** Delete the specified node from a page.
5253 * @param[in] mp The page to operate on.
5254 * @param[in] indx The index of the node to delete.
5255 * @param[in] ksize The size of a node. Only used if the page is
5256 * part of a #MDB_DUPFIXED database.
5259 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5262 indx_t i, j, numkeys, ptr;
5269 COPY_PGNO(pgno, mp->mp_pgno);
5270 DPRINTF("delete node %u on %s page %zu", indx,
5271 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5274 assert(indx < NUMKEYS(mp));
5277 int x = NUMKEYS(mp) - 1 - indx;
5278 base = LEAF2KEY(mp, indx, ksize);
5280 memmove(base, base + ksize, x * ksize);
5281 mp->mp_lower -= sizeof(indx_t);
5282 mp->mp_upper += ksize - sizeof(indx_t);
5286 node = NODEPTR(mp, indx);
5287 sz = NODESIZE + node->mn_ksize;
5289 if (F_ISSET(node->mn_flags, F_BIGDATA))
5290 sz += sizeof(pgno_t);
5292 sz += NODEDSZ(node);
5296 ptr = mp->mp_ptrs[indx];
5297 numkeys = NUMKEYS(mp);
5298 for (i = j = 0; i < numkeys; i++) {
5300 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5301 if (mp->mp_ptrs[i] < ptr)
5302 mp->mp_ptrs[j] += sz;
5307 base = (char *)mp + mp->mp_upper;
5308 memmove(base + sz, base, ptr - mp->mp_upper);
5310 mp->mp_lower -= sizeof(indx_t);
5314 /** Compact the main page after deleting a node on a subpage.
5315 * @param[in] mp The main page to operate on.
5316 * @param[in] indx The index of the subpage on the main page.
5319 mdb_node_shrink(MDB_page *mp, indx_t indx)
5326 indx_t i, numkeys, ptr;
5328 node = NODEPTR(mp, indx);
5329 sp = (MDB_page *)NODEDATA(node);
5330 osize = NODEDSZ(node);
5332 delta = sp->mp_upper - sp->mp_lower;
5333 SETDSZ(node, osize - delta);
5334 xp = (MDB_page *)((char *)sp + delta);
5336 /* shift subpage upward */
5338 nsize = NUMKEYS(sp) * sp->mp_pad;
5339 memmove(METADATA(xp), METADATA(sp), nsize);
5342 nsize = osize - sp->mp_upper;
5343 numkeys = NUMKEYS(sp);
5344 for (i=numkeys-1; i>=0; i--)
5345 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5347 xp->mp_upper = sp->mp_lower;
5348 xp->mp_lower = sp->mp_lower;
5349 xp->mp_flags = sp->mp_flags;
5350 xp->mp_pad = sp->mp_pad;
5351 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5353 /* shift lower nodes upward */
5354 ptr = mp->mp_ptrs[indx];
5355 numkeys = NUMKEYS(mp);
5356 for (i = 0; i < numkeys; i++) {
5357 if (mp->mp_ptrs[i] <= ptr)
5358 mp->mp_ptrs[i] += delta;
5361 base = (char *)mp + mp->mp_upper;
5362 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5363 mp->mp_upper += delta;
5366 /** Initial setup of a sorted-dups cursor.
5367 * Sorted duplicates are implemented as a sub-database for the given key.
5368 * The duplicate data items are actually keys of the sub-database.
5369 * Operations on the duplicate data items are performed using a sub-cursor
5370 * initialized when the sub-database is first accessed. This function does
5371 * the preliminary setup of the sub-cursor, filling in the fields that
5372 * depend only on the parent DB.
5373 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5376 mdb_xcursor_init0(MDB_cursor *mc)
5378 MDB_xcursor *mx = mc->mc_xcursor;
5380 mx->mx_cursor.mc_xcursor = NULL;
5381 mx->mx_cursor.mc_txn = mc->mc_txn;
5382 mx->mx_cursor.mc_db = &mx->mx_db;
5383 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5384 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5385 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5386 mx->mx_cursor.mc_snum = 0;
5387 mx->mx_cursor.mc_top = 0;
5388 mx->mx_cursor.mc_flags = C_SUB;
5389 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5390 mx->mx_dbx.md_dcmp = NULL;
5391 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5394 /** Final setup of a sorted-dups cursor.
5395 * Sets up the fields that depend on the data from the main cursor.
5396 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5397 * @param[in] node The data containing the #MDB_db record for the
5398 * sorted-dup database.
5401 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5403 MDB_xcursor *mx = mc->mc_xcursor;
5405 if (node->mn_flags & F_SUBDATA) {
5406 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5407 mx->mx_cursor.mc_pg[0] = 0;
5408 mx->mx_cursor.mc_snum = 0;
5409 mx->mx_cursor.mc_flags = C_SUB;
5411 MDB_page *fp = NODEDATA(node);
5412 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5413 mx->mx_db.md_flags = 0;
5414 mx->mx_db.md_depth = 1;
5415 mx->mx_db.md_branch_pages = 0;
5416 mx->mx_db.md_leaf_pages = 1;
5417 mx->mx_db.md_overflow_pages = 0;
5418 mx->mx_db.md_entries = NUMKEYS(fp);
5419 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5420 mx->mx_cursor.mc_snum = 1;
5421 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5422 mx->mx_cursor.mc_top = 0;
5423 mx->mx_cursor.mc_pg[0] = fp;
5424 mx->mx_cursor.mc_ki[0] = 0;
5425 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5426 mx->mx_db.md_flags = MDB_DUPFIXED;
5427 mx->mx_db.md_pad = fp->mp_pad;
5428 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5429 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5432 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5434 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5436 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5437 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5438 #if UINT_MAX < SIZE_MAX
5439 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5440 #ifdef MISALIGNED_OK
5441 mx->mx_dbx.md_cmp = mdb_cmp_long;
5443 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5448 /** Initialize a cursor for a given transaction and database. */
5450 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5455 mc->mc_db = &txn->mt_dbs[dbi];
5456 mc->mc_dbx = &txn->mt_dbxs[dbi];
5457 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5462 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5464 mc->mc_xcursor = mx;
5465 mdb_xcursor_init0(mc);
5467 mc->mc_xcursor = NULL;
5469 if (*mc->mc_dbflag & DB_STALE) {
5470 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5475 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5478 MDB_xcursor *mx = NULL;
5479 size_t size = sizeof(MDB_cursor);
5481 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5484 /* Allow read access to the freelist */
5485 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5488 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5489 size += sizeof(MDB_xcursor);
5491 if ((mc = malloc(size)) != NULL) {
5492 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5493 mx = (MDB_xcursor *)(mc + 1);
5495 mdb_cursor_init(mc, txn, dbi, mx);
5496 if (txn->mt_cursors) {
5497 mc->mc_next = txn->mt_cursors[dbi];
5498 txn->mt_cursors[dbi] = mc;
5500 mc->mc_flags |= C_ALLOCD;
5511 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5513 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5516 if (txn->mt_cursors)
5519 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5523 /* Return the count of duplicate data items for the current key */
5525 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5529 if (mc == NULL || countp == NULL)
5532 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5535 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5536 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5539 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5542 *countp = mc->mc_xcursor->mx_db.md_entries;
5548 mdb_cursor_close(MDB_cursor *mc)
5551 /* remove from txn, if tracked */
5552 if (mc->mc_txn->mt_cursors) {
5553 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5554 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5556 *prev = mc->mc_next;
5558 if (mc->mc_flags & C_ALLOCD)
5564 mdb_cursor_txn(MDB_cursor *mc)
5566 if (!mc) return NULL;
5571 mdb_cursor_dbi(MDB_cursor *mc)
5577 /** Replace the key for a node with a new key.
5578 * @param[in] mp The page containing the node to operate on.
5579 * @param[in] indx The index of the node to operate on.
5580 * @param[in] key The new key to use.
5581 * @return 0 on success, non-zero on failure.
5584 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5590 indx_t ptr, i, numkeys;
5593 node = NODEPTR(mp, indx);
5594 ptr = mp->mp_ptrs[indx];
5598 char kbuf2[(MAXKEYSIZE*2+1)];
5599 k2.mv_data = NODEKEY(node);
5600 k2.mv_size = node->mn_ksize;
5601 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5603 mdb_dkey(&k2, kbuf2),
5609 delta0 = delta = key->mv_size - node->mn_ksize;
5611 /* Must be 2-byte aligned. If new key is
5612 * shorter by 1, the shift will be skipped.
5614 delta += (delta & 1);
5616 if (delta > 0 && SIZELEFT(mp) < delta) {
5617 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5618 return MDB_PAGE_FULL;
5621 numkeys = NUMKEYS(mp);
5622 for (i = 0; i < numkeys; i++) {
5623 if (mp->mp_ptrs[i] <= ptr)
5624 mp->mp_ptrs[i] -= delta;
5627 base = (char *)mp + mp->mp_upper;
5628 len = ptr - mp->mp_upper + NODESIZE;
5629 memmove(base - delta, base, len);
5630 mp->mp_upper -= delta;
5632 node = NODEPTR(mp, indx);
5635 /* But even if no shift was needed, update ksize */
5637 node->mn_ksize = key->mv_size;
5640 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5645 /** Move a node from csrc to cdst.
5648 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5654 unsigned short flags;
5658 /* Mark src and dst as dirty. */
5659 if ((rc = mdb_page_touch(csrc)) ||
5660 (rc = mdb_page_touch(cdst)))
5663 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5664 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5665 key.mv_size = csrc->mc_db->md_pad;
5666 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5668 data.mv_data = NULL;
5672 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5673 assert(!((long)srcnode&1));
5674 srcpg = NODEPGNO(srcnode);
5675 flags = srcnode->mn_flags;
5676 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5677 unsigned int snum = csrc->mc_snum;
5679 /* must find the lowest key below src */
5680 mdb_page_search_root(csrc, NULL, 0);
5681 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5682 key.mv_size = csrc->mc_db->md_pad;
5683 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5685 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5686 key.mv_size = NODEKSZ(s2);
5687 key.mv_data = NODEKEY(s2);
5689 csrc->mc_snum = snum--;
5690 csrc->mc_top = snum;
5692 key.mv_size = NODEKSZ(srcnode);
5693 key.mv_data = NODEKEY(srcnode);
5695 data.mv_size = NODEDSZ(srcnode);
5696 data.mv_data = NODEDATA(srcnode);
5698 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5699 unsigned int snum = cdst->mc_snum;
5702 /* must find the lowest key below dst */
5703 mdb_page_search_root(cdst, NULL, 0);
5704 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5705 bkey.mv_size = cdst->mc_db->md_pad;
5706 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5708 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5709 bkey.mv_size = NODEKSZ(s2);
5710 bkey.mv_data = NODEKEY(s2);
5712 cdst->mc_snum = snum--;
5713 cdst->mc_top = snum;
5714 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5717 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5718 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5719 csrc->mc_ki[csrc->mc_top],
5721 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5722 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5724 /* Add the node to the destination page.
5726 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5727 if (rc != MDB_SUCCESS)
5730 /* Delete the node from the source page.
5732 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5735 /* Adjust other cursors pointing to mp */
5736 MDB_cursor *m2, *m3;
5737 MDB_dbi dbi = csrc->mc_dbi;
5738 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5740 if (csrc->mc_flags & C_SUB)
5743 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5744 if (m2 == csrc) continue;
5745 if (csrc->mc_flags & C_SUB)
5746 m3 = &m2->mc_xcursor->mx_cursor;
5749 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5750 csrc->mc_ki[csrc->mc_top]) {
5751 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5752 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5757 /* Update the parent separators.
5759 if (csrc->mc_ki[csrc->mc_top] == 0) {
5760 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5761 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5762 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5764 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5765 key.mv_size = NODEKSZ(srcnode);
5766 key.mv_data = NODEKEY(srcnode);
5768 DPRINTF("update separator for source page %zu to [%s]",
5769 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5770 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5771 &key)) != MDB_SUCCESS)
5774 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5776 nullkey.mv_size = 0;
5777 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5778 assert(rc == MDB_SUCCESS);
5782 if (cdst->mc_ki[cdst->mc_top] == 0) {
5783 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5784 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5785 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5787 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5788 key.mv_size = NODEKSZ(srcnode);
5789 key.mv_data = NODEKEY(srcnode);
5791 DPRINTF("update separator for destination page %zu to [%s]",
5792 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5793 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5794 &key)) != MDB_SUCCESS)
5797 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5799 nullkey.mv_size = 0;
5800 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5801 assert(rc == MDB_SUCCESS);
5808 /** Merge one page into another.
5809 * The nodes from the page pointed to by \b csrc will
5810 * be copied to the page pointed to by \b cdst and then
5811 * the \b csrc page will be freed.
5812 * @param[in] csrc Cursor pointing to the source page.
5813 * @param[in] cdst Cursor pointing to the destination page.
5816 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5824 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5825 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5827 assert(csrc->mc_snum > 1); /* can't merge root page */
5828 assert(cdst->mc_snum > 1);
5830 /* Mark dst as dirty. */
5831 if ((rc = mdb_page_touch(cdst)))
5834 /* Move all nodes from src to dst.
5836 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5837 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5838 key.mv_size = csrc->mc_db->md_pad;
5839 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5840 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5841 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5842 if (rc != MDB_SUCCESS)
5844 key.mv_data = (char *)key.mv_data + key.mv_size;
5847 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5848 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5849 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5850 unsigned int snum = csrc->mc_snum;
5852 /* must find the lowest key below src */
5853 mdb_page_search_root(csrc, NULL, 0);
5854 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5855 key.mv_size = csrc->mc_db->md_pad;
5856 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5858 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5859 key.mv_size = NODEKSZ(s2);
5860 key.mv_data = NODEKEY(s2);
5862 csrc->mc_snum = snum--;
5863 csrc->mc_top = snum;
5865 key.mv_size = srcnode->mn_ksize;
5866 key.mv_data = NODEKEY(srcnode);
5869 data.mv_size = NODEDSZ(srcnode);
5870 data.mv_data = NODEDATA(srcnode);
5871 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5872 if (rc != MDB_SUCCESS)
5877 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5878 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);
5880 /* Unlink the src page from parent and add to free list.
5882 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5883 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5885 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5889 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5890 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5891 csrc->mc_db->md_leaf_pages--;
5893 csrc->mc_db->md_branch_pages--;
5895 /* Adjust other cursors pointing to mp */
5896 MDB_cursor *m2, *m3;
5897 MDB_dbi dbi = csrc->mc_dbi;
5898 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5900 if (csrc->mc_flags & C_SUB)
5903 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5904 if (csrc->mc_flags & C_SUB)
5905 m3 = &m2->mc_xcursor->mx_cursor;
5908 if (m3 == csrc) continue;
5909 if (m3->mc_snum < csrc->mc_snum) continue;
5910 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5911 m3->mc_pg[csrc->mc_top] = mp;
5912 m3->mc_ki[csrc->mc_top] += nkeys;
5916 mdb_cursor_pop(csrc);
5918 return mdb_rebalance(csrc);
5921 /** Copy the contents of a cursor.
5922 * @param[in] csrc The cursor to copy from.
5923 * @param[out] cdst The cursor to copy to.
5926 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5930 cdst->mc_txn = csrc->mc_txn;
5931 cdst->mc_dbi = csrc->mc_dbi;
5932 cdst->mc_db = csrc->mc_db;
5933 cdst->mc_dbx = csrc->mc_dbx;
5934 cdst->mc_snum = csrc->mc_snum;
5935 cdst->mc_top = csrc->mc_top;
5936 cdst->mc_flags = csrc->mc_flags;
5938 for (i=0; i<csrc->mc_snum; i++) {
5939 cdst->mc_pg[i] = csrc->mc_pg[i];
5940 cdst->mc_ki[i] = csrc->mc_ki[i];
5944 /** Rebalance the tree after a delete operation.
5945 * @param[in] mc Cursor pointing to the page where rebalancing
5947 * @return 0 on success, non-zero on failure.
5950 mdb_rebalance(MDB_cursor *mc)
5960 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5961 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5962 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5963 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5967 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5970 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5971 DPRINTF("no need to rebalance page %zu, above fill threshold",
5977 if (mc->mc_snum < 2) {
5978 MDB_page *mp = mc->mc_pg[0];
5979 if (NUMKEYS(mp) == 0) {
5980 DPUTS("tree is completely empty");
5981 mc->mc_db->md_root = P_INVALID;
5982 mc->mc_db->md_depth = 0;
5983 mc->mc_db->md_leaf_pages = 0;
5984 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5988 /* Adjust other cursors pointing to mp */
5989 MDB_cursor *m2, *m3;
5990 MDB_dbi dbi = mc->mc_dbi;
5992 if (mc->mc_flags & C_SUB)
5995 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5996 if (m2 == mc) continue;
5997 if (mc->mc_flags & C_SUB)
5998 m3 = &m2->mc_xcursor->mx_cursor;
6001 if (m3->mc_snum < mc->mc_snum) continue;
6002 if (m3->mc_pg[0] == mp) {
6008 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6009 DPUTS("collapsing root page!");
6010 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6011 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6012 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6015 mc->mc_db->md_depth--;
6016 mc->mc_db->md_branch_pages--;
6018 /* Adjust other cursors pointing to mp */
6019 MDB_cursor *m2, *m3;
6020 MDB_dbi dbi = mc->mc_dbi;
6022 if (mc->mc_flags & C_SUB)
6025 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6026 if (m2 == mc) continue;
6027 if (mc->mc_flags & C_SUB)
6028 m3 = &m2->mc_xcursor->mx_cursor;
6031 if (m3->mc_snum < mc->mc_snum) continue;
6032 if (m3->mc_pg[0] == mp) {
6033 m3->mc_pg[0] = mc->mc_pg[0];
6038 DPUTS("root page doesn't need rebalancing");
6042 /* The parent (branch page) must have at least 2 pointers,
6043 * otherwise the tree is invalid.
6045 ptop = mc->mc_top-1;
6046 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6048 /* Leaf page fill factor is below the threshold.
6049 * Try to move keys from left or right neighbor, or
6050 * merge with a neighbor page.
6055 mdb_cursor_copy(mc, &mn);
6056 mn.mc_xcursor = NULL;
6058 if (mc->mc_ki[ptop] == 0) {
6059 /* We're the leftmost leaf in our parent.
6061 DPUTS("reading right neighbor");
6063 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6064 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6066 mn.mc_ki[mn.mc_top] = 0;
6067 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6069 /* There is at least one neighbor to the left.
6071 DPUTS("reading left neighbor");
6073 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6074 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6076 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6077 mc->mc_ki[mc->mc_top] = 0;
6080 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6081 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);
6083 /* If the neighbor page is above threshold and has at least two
6084 * keys, move one key from it.
6086 * Otherwise we should try to merge them.
6088 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6089 return mdb_node_move(&mn, mc);
6090 else { /* FIXME: if (has_enough_room()) */
6091 mc->mc_flags &= ~C_INITIALIZED;
6092 if (mc->mc_ki[ptop] == 0)
6093 return mdb_page_merge(&mn, mc);
6095 return mdb_page_merge(mc, &mn);
6099 /** Complete a delete operation started by #mdb_cursor_del(). */
6101 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6105 /* add overflow pages to free list */
6106 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6110 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6111 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6112 mc->mc_db->md_overflow_pages -= ovpages;
6113 for (i=0; i<ovpages; i++) {
6114 DPRINTF("freed ov page %zu", pg);
6115 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6119 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6120 mc->mc_db->md_entries--;
6121 rc = mdb_rebalance(mc);
6122 if (rc != MDB_SUCCESS)
6123 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6129 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6130 MDB_val *key, MDB_val *data)
6135 MDB_val rdata, *xdata;
6139 assert(key != NULL);
6141 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6143 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6146 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6150 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6154 mdb_cursor_init(&mc, txn, dbi, &mx);
6165 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6167 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6171 /** Split a page and insert a new node.
6172 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6173 * The cursor will be updated to point to the actual page and index where
6174 * the node got inserted after the split.
6175 * @param[in] newkey The key for the newly inserted node.
6176 * @param[in] newdata The data for the newly inserted node.
6177 * @param[in] newpgno The page number, if the new node is a branch node.
6178 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6179 * @return 0 on success, non-zero on failure.
6182 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6183 unsigned int nflags)
6186 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6189 unsigned int i, j, split_indx, nkeys, pmax;
6191 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6193 MDB_page *mp, *rp, *pp;
6198 mp = mc->mc_pg[mc->mc_top];
6199 newindx = mc->mc_ki[mc->mc_top];
6201 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6202 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6203 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6205 /* Create a right sibling. */
6206 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6208 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6210 if (mc->mc_snum < 2) {
6211 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6213 /* shift current top to make room for new parent */
6214 mc->mc_pg[1] = mc->mc_pg[0];
6215 mc->mc_ki[1] = mc->mc_ki[0];
6218 mc->mc_db->md_root = pp->mp_pgno;
6219 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6220 mc->mc_db->md_depth++;
6223 /* Add left (implicit) pointer. */
6224 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6225 /* undo the pre-push */
6226 mc->mc_pg[0] = mc->mc_pg[1];
6227 mc->mc_ki[0] = mc->mc_ki[1];
6228 mc->mc_db->md_root = mp->mp_pgno;
6229 mc->mc_db->md_depth--;
6236 ptop = mc->mc_top-1;
6237 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6240 mc->mc_flags |= C_SPLITTING;
6241 mdb_cursor_copy(mc, &mn);
6242 mn.mc_pg[mn.mc_top] = rp;
6243 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6245 if (nflags & MDB_APPEND) {
6246 mn.mc_ki[mn.mc_top] = 0;
6248 split_indx = newindx;
6253 nkeys = NUMKEYS(mp);
6254 split_indx = nkeys / 2;
6255 if (newindx < split_indx)
6261 unsigned int lsize, rsize, ksize;
6262 /* Move half of the keys to the right sibling */
6264 x = mc->mc_ki[mc->mc_top] - split_indx;
6265 ksize = mc->mc_db->md_pad;
6266 split = LEAF2KEY(mp, split_indx, ksize);
6267 rsize = (nkeys - split_indx) * ksize;
6268 lsize = (nkeys - split_indx) * sizeof(indx_t);
6269 mp->mp_lower -= lsize;
6270 rp->mp_lower += lsize;
6271 mp->mp_upper += rsize - lsize;
6272 rp->mp_upper -= rsize - lsize;
6273 sepkey.mv_size = ksize;
6274 if (newindx == split_indx) {
6275 sepkey.mv_data = newkey->mv_data;
6277 sepkey.mv_data = split;
6280 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6281 memcpy(rp->mp_ptrs, split, rsize);
6282 sepkey.mv_data = rp->mp_ptrs;
6283 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6284 memcpy(ins, newkey->mv_data, ksize);
6285 mp->mp_lower += sizeof(indx_t);
6286 mp->mp_upper -= ksize - sizeof(indx_t);
6289 memcpy(rp->mp_ptrs, split, x * ksize);
6290 ins = LEAF2KEY(rp, x, ksize);
6291 memcpy(ins, newkey->mv_data, ksize);
6292 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6293 rp->mp_lower += sizeof(indx_t);
6294 rp->mp_upper -= ksize - sizeof(indx_t);
6295 mc->mc_ki[mc->mc_top] = x;
6296 mc->mc_pg[mc->mc_top] = rp;
6301 /* For leaf pages, check the split point based on what
6302 * fits where, since otherwise mdb_node_add can fail.
6304 * This check is only needed when the data items are
6305 * relatively large, such that being off by one will
6306 * make the difference between success or failure.
6308 * It's also relevant if a page happens to be laid out
6309 * such that one half of its nodes are all "small" and
6310 * the other half of its nodes are "large." If the new
6311 * item is also "large" and falls on the half with
6312 * "large" nodes, it also may not fit.
6315 unsigned int psize, nsize;
6316 /* Maximum free space in an empty page */
6317 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6318 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6319 if ((nkeys < 20) || (nsize > pmax/16)) {
6320 if (newindx <= split_indx) {
6323 for (i=0; i<split_indx; i++) {
6324 node = NODEPTR(mp, i);
6325 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6326 if (F_ISSET(node->mn_flags, F_BIGDATA))
6327 psize += sizeof(pgno_t);
6329 psize += NODEDSZ(node);
6333 split_indx = newindx;
6344 for (i=nkeys-1; i>=split_indx; i--) {
6345 node = NODEPTR(mp, i);
6346 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6347 if (F_ISSET(node->mn_flags, F_BIGDATA))
6348 psize += sizeof(pgno_t);
6350 psize += NODEDSZ(node);
6354 split_indx = newindx;
6365 /* First find the separating key between the split pages.
6366 * The case where newindx == split_indx is ambiguous; the
6367 * new item could go to the new page or stay on the original
6368 * page. If newpos == 1 it goes to the new page.
6370 if (newindx == split_indx && newpos) {
6371 sepkey.mv_size = newkey->mv_size;
6372 sepkey.mv_data = newkey->mv_data;
6374 node = NODEPTR(mp, split_indx);
6375 sepkey.mv_size = node->mn_ksize;
6376 sepkey.mv_data = NODEKEY(node);
6380 DPRINTF("separator is [%s]", DKEY(&sepkey));
6382 /* Copy separator key to the parent.
6384 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6388 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6391 if (mn.mc_snum == mc->mc_snum) {
6392 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6393 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6394 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6395 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6400 /* Right page might now have changed parent.
6401 * Check if left page also changed parent.
6403 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6404 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6405 for (i=0; i<ptop; i++) {
6406 mc->mc_pg[i] = mn.mc_pg[i];
6407 mc->mc_ki[i] = mn.mc_ki[i];
6409 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6410 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6414 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6417 mc->mc_flags ^= C_SPLITTING;
6418 if (rc != MDB_SUCCESS) {
6421 if (nflags & MDB_APPEND) {
6422 mc->mc_pg[mc->mc_top] = rp;
6423 mc->mc_ki[mc->mc_top] = 0;
6424 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6427 for (i=0; i<mc->mc_top; i++)
6428 mc->mc_ki[i] = mn.mc_ki[i];
6435 /* Move half of the keys to the right sibling. */
6437 /* grab a page to hold a temporary copy */
6438 copy = mdb_page_malloc(mc);
6442 copy->mp_pgno = mp->mp_pgno;
6443 copy->mp_flags = mp->mp_flags;
6444 copy->mp_lower = PAGEHDRSZ;
6445 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6446 mc->mc_pg[mc->mc_top] = copy;
6447 for (i = j = 0; i <= nkeys; j++) {
6448 if (i == split_indx) {
6449 /* Insert in right sibling. */
6450 /* Reset insert index for right sibling. */
6451 if (i != newindx || (newpos ^ ins_new)) {
6453 mc->mc_pg[mc->mc_top] = rp;
6457 if (i == newindx && !ins_new) {
6458 /* Insert the original entry that caused the split. */
6459 rkey.mv_data = newkey->mv_data;
6460 rkey.mv_size = newkey->mv_size;
6469 /* Update index for the new key. */
6470 mc->mc_ki[mc->mc_top] = j;
6471 } else if (i == nkeys) {
6474 node = NODEPTR(mp, i);
6475 rkey.mv_data = NODEKEY(node);
6476 rkey.mv_size = node->mn_ksize;
6478 xdata.mv_data = NODEDATA(node);
6479 xdata.mv_size = NODEDSZ(node);
6482 pgno = NODEPGNO(node);
6483 flags = node->mn_flags;
6488 if (!IS_LEAF(mp) && j == 0) {
6489 /* First branch index doesn't need key data. */
6493 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6497 nkeys = NUMKEYS(copy);
6498 for (i=0; i<nkeys; i++)
6499 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6500 mp->mp_lower = copy->mp_lower;
6501 mp->mp_upper = copy->mp_upper;
6502 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6503 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6505 /* reset back to original page */
6506 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6507 mc->mc_pg[mc->mc_top] = mp;
6508 if (nflags & MDB_RESERVE) {
6509 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6510 if (!(node->mn_flags & F_BIGDATA))
6511 newdata->mv_data = NODEDATA(node);
6517 /* return tmp page to freelist */
6518 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6519 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6520 mc->mc_txn->mt_env->me_dpages = copy;
6523 /* Adjust other cursors pointing to mp */
6524 MDB_cursor *m2, *m3;
6525 MDB_dbi dbi = mc->mc_dbi;
6526 int fixup = NUMKEYS(mp);
6528 if (mc->mc_flags & C_SUB)
6531 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6532 if (m2 == mc) continue;
6533 if (mc->mc_flags & C_SUB)
6534 m3 = &m2->mc_xcursor->mx_cursor;
6537 if (!(m3->mc_flags & C_INITIALIZED))
6539 if (m3->mc_flags & C_SPLITTING)
6544 for (k=m3->mc_top; k>=0; k--) {
6545 m3->mc_ki[k+1] = m3->mc_ki[k];
6546 m3->mc_pg[k+1] = m3->mc_pg[k];
6548 if (m3->mc_ki[0] >= split_indx) {
6553 m3->mc_pg[0] = mc->mc_pg[0];
6557 if (m3->mc_pg[mc->mc_top] == mp) {
6558 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6559 m3->mc_ki[mc->mc_top]++;
6560 if (m3->mc_ki[mc->mc_top] >= fixup) {
6561 m3->mc_pg[mc->mc_top] = rp;
6562 m3->mc_ki[mc->mc_top] -= fixup;
6563 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6565 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6566 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6575 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6576 MDB_val *key, MDB_val *data, unsigned int flags)
6581 assert(key != NULL);
6582 assert(data != NULL);
6584 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6587 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6591 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6595 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6598 mdb_cursor_init(&mc, txn, dbi, &mx);
6599 return mdb_cursor_put(&mc, key, data, flags);
6602 /** Only a subset of the @ref mdb_env flags can be changed
6603 * at runtime. Changing other flags requires closing the environment
6604 * and re-opening it with the new flags.
6606 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
6608 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6610 if ((flag & CHANGEABLE) != flag)
6613 env->me_flags |= flag;
6615 env->me_flags &= ~flag;
6620 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6625 *arg = env->me_flags;
6630 mdb_env_get_path(MDB_env *env, const char **arg)
6635 *arg = env->me_path;
6639 /** Common code for #mdb_stat() and #mdb_env_stat().
6640 * @param[in] env the environment to operate in.
6641 * @param[in] db the #MDB_db record containing the stats to return.
6642 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6643 * @return 0, this function always succeeds.
6646 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6648 arg->ms_psize = env->me_psize;
6649 arg->ms_depth = db->md_depth;
6650 arg->ms_branch_pages = db->md_branch_pages;
6651 arg->ms_leaf_pages = db->md_leaf_pages;
6652 arg->ms_overflow_pages = db->md_overflow_pages;
6653 arg->ms_entries = db->md_entries;
6658 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6662 if (env == NULL || arg == NULL)
6665 toggle = mdb_env_pick_meta(env);
6667 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6671 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6675 if (env == NULL || arg == NULL)
6678 toggle = mdb_env_pick_meta(env);
6679 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6680 arg->me_mapsize = env->me_mapsize;
6681 arg->me_maxreaders = env->me_maxreaders;
6682 arg->me_numreaders = env->me_numreaders;
6683 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6684 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6688 /** Set the default comparison functions for a database.
6689 * Called immediately after a database is opened to set the defaults.
6690 * The user can then override them with #mdb_set_compare() or
6691 * #mdb_set_dupsort().
6692 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6693 * @param[in] dbi A database handle returned by #mdb_open()
6696 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6698 uint16_t f = txn->mt_dbs[dbi].md_flags;
6700 txn->mt_dbxs[dbi].md_cmp =
6701 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6702 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6704 txn->mt_dbxs[dbi].md_dcmp =
6705 !(f & MDB_DUPSORT) ? 0 :
6706 ((f & MDB_INTEGERDUP)
6707 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6708 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6711 #define PERSISTENT_FLAGS 0xffff
6712 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
6713 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
6714 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6719 int rc, dbflag, exact;
6720 unsigned int unused = 0;
6723 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6724 mdb_default_cmp(txn, FREE_DBI);
6727 if ((flags & VALID_FLAGS) != flags)
6733 if (flags & PERSISTENT_FLAGS) {
6734 uint16_t f2 = flags & PERSISTENT_FLAGS;
6735 /* make sure flag changes get committed */
6736 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
6737 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
6738 txn->mt_flags |= MDB_TXN_DIRTY;
6741 mdb_default_cmp(txn, MAIN_DBI);
6745 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6746 mdb_default_cmp(txn, MAIN_DBI);
6749 /* Is the DB already open? */
6751 for (i=2; i<txn->mt_numdbs; i++) {
6752 if (!txn->mt_dbxs[i].md_name.mv_size) {
6753 /* Remember this free slot */
6754 if (!unused) unused = i;
6757 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6758 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6764 /* If no free slot and max hit, fail */
6765 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
6766 return MDB_DBS_FULL;
6768 /* Find the DB info */
6772 key.mv_data = (void *)name;
6773 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6774 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6775 if (rc == MDB_SUCCESS) {
6776 /* make sure this is actually a DB */
6777 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6778 if (!(node->mn_flags & F_SUBDATA))
6780 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6781 /* Create if requested */
6783 data.mv_size = sizeof(MDB_db);
6784 data.mv_data = &dummy;
6785 memset(&dummy, 0, sizeof(dummy));
6786 dummy.md_root = P_INVALID;
6787 dummy.md_flags = flags & PERSISTENT_FLAGS;
6788 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6792 /* OK, got info, add to table */
6793 if (rc == MDB_SUCCESS) {
6794 unsigned int slot = unused ? unused : txn->mt_numdbs;
6795 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6796 txn->mt_dbxs[slot].md_name.mv_size = len;
6797 txn->mt_dbxs[slot].md_rel = NULL;
6798 txn->mt_dbflags[slot] = dbflag;
6799 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6801 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6802 mdb_default_cmp(txn, slot);
6805 txn->mt_env->me_numdbs++;
6812 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6814 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6817 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6820 void mdb_close(MDB_env *env, MDB_dbi dbi)
6823 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6825 ptr = env->me_dbxs[dbi].md_name.mv_data;
6826 env->me_dbxs[dbi].md_name.mv_data = NULL;
6827 env->me_dbxs[dbi].md_name.mv_size = 0;
6831 /** Add all the DB's pages to the free list.
6832 * @param[in] mc Cursor on the DB to free.
6833 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6834 * @return 0 on success, non-zero on failure.
6837 mdb_drop0(MDB_cursor *mc, int subs)
6841 rc = mdb_page_search(mc, NULL, 0);
6842 if (rc == MDB_SUCCESS) {
6847 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6848 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6851 mdb_cursor_copy(mc, &mx);
6852 while (mc->mc_snum > 0) {
6853 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6854 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6855 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6856 if (ni->mn_flags & F_SUBDATA) {
6857 mdb_xcursor_init1(mc, ni);
6858 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6864 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6866 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6869 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6874 rc = mdb_cursor_sibling(mc, 1);
6876 /* no more siblings, go back to beginning
6877 * of previous level. (stack was already popped
6878 * by mdb_cursor_sibling)
6880 for (i=1; i<mc->mc_top; i++)
6881 mc->mc_pg[i] = mx.mc_pg[i];
6885 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6886 mc->mc_db->md_root);
6891 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6896 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6899 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6902 rc = mdb_cursor_open(txn, dbi, &mc);
6906 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6910 /* Can't delete the main DB */
6911 if (del && dbi > MAIN_DBI) {
6912 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6914 mdb_close(txn->mt_env, dbi);
6916 /* reset the DB record, mark it dirty */
6917 txn->mt_dbflags[dbi] |= DB_DIRTY;
6918 txn->mt_dbs[dbi].md_depth = 0;
6919 txn->mt_dbs[dbi].md_branch_pages = 0;
6920 txn->mt_dbs[dbi].md_leaf_pages = 0;
6921 txn->mt_dbs[dbi].md_overflow_pages = 0;
6922 txn->mt_dbs[dbi].md_entries = 0;
6923 txn->mt_dbs[dbi].md_root = P_INVALID;
6925 if (!txn->mt_u.dirty_list[0].mid) {
6928 /* make sure we have at least one dirty page in this txn
6929 * otherwise these changes will be ignored.
6931 key.mv_size = sizeof(txnid_t);
6932 key.mv_data = &txn->mt_txnid;
6933 data.mv_size = sizeof(MDB_ID);
6934 data.mv_data = txn->mt_free_pgs;
6935 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
6936 rc = mdb_cursor_put(&m2, &key, &data, 0);
6940 mdb_cursor_close(mc);
6944 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6946 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6949 txn->mt_dbxs[dbi].md_cmp = cmp;
6953 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6955 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6958 txn->mt_dbxs[dbi].md_dcmp = cmp;
6962 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6964 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6967 txn->mt_dbxs[dbi].md_rel = rel;
6971 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6973 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6976 txn->mt_dbxs[dbi].md_relctx = ctx;