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 */
830 unsigned int mt_flags; /**< @ref mdb_txn */
831 /** Tracks which of the two meta pages was used at the start
832 * of this transaction.
834 unsigned int mt_toggle;
837 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
838 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
839 * raise this on a 64 bit machine.
841 #define CURSOR_STACK 32
845 /** Cursors are used for all DB operations */
847 /** Next cursor on this DB in this txn */
849 /** Original cursor if this is a shadow */
851 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
852 struct MDB_xcursor *mc_xcursor;
853 /** The transaction that owns this cursor */
855 /** The database handle this cursor operates on */
857 /** The database record for this cursor */
859 /** The database auxiliary record for this cursor */
861 /** The @ref mt_dbflag for this database */
862 unsigned char *mc_dbflag;
863 unsigned short mc_snum; /**< number of pushed pages */
864 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
865 /** @defgroup mdb_cursor Cursor Flags
867 * Cursor state flags.
870 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
871 #define C_EOF 0x02 /**< No more data */
872 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
873 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
874 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
875 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
877 unsigned int mc_flags; /**< @ref mdb_cursor */
878 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
879 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
882 /** Context for sorted-dup records.
883 * We could have gone to a fully recursive design, with arbitrarily
884 * deep nesting of sub-databases. But for now we only handle these
885 * levels - main DB, optional sub-DB, sorted-duplicate DB.
887 typedef struct MDB_xcursor {
888 /** A sub-cursor for traversing the Dup DB */
889 MDB_cursor mx_cursor;
890 /** The database record for this Dup DB */
892 /** The auxiliary DB record for this Dup DB */
894 /** The @ref mt_dbflag for this Dup DB */
895 unsigned char mx_dbflag;
898 /** A set of pages freed by an earlier transaction. */
899 typedef struct MDB_oldpages {
900 /** Usually we only read one record from the FREEDB at a time, but
901 * in case we read more, this will chain them together.
903 struct MDB_oldpages *mo_next;
904 /** The ID of the transaction in which these pages were freed. */
906 /** An #MDB_IDL of the pages */
907 pgno_t mo_pages[1]; /* dynamic */
910 /** The database environment. */
912 HANDLE me_fd; /**< The main data file */
913 HANDLE me_lfd; /**< The lock file */
914 HANDLE me_mfd; /**< just for writing the meta pages */
915 /** Failed to update the meta page. Probably an I/O error. */
916 #define MDB_FATAL_ERROR 0x80000000U
917 /** Read-only Filesystem. Allow read access, no locking. */
918 #define MDB_ROFS 0x40000000U
919 /** Some fields are initialized. */
920 #define MDB_ENV_ACTIVE 0x20000000U
921 uint32_t me_flags; /**< @ref mdb_env */
922 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
923 unsigned int me_maxreaders; /**< size of the reader table */
924 unsigned int me_numreaders; /**< max numreaders set by this env */
925 MDB_dbi me_numdbs; /**< number of DBs opened */
926 MDB_dbi me_maxdbs; /**< size of the DB table */
927 pid_t me_pid; /**< process ID of this env */
928 char *me_path; /**< path to the DB files */
929 char *me_map; /**< the memory map of the data file */
930 MDB_txninfo *me_txns; /**< the memory map of the lock file */
931 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
932 MDB_txn *me_txn; /**< current write transaction */
933 size_t me_mapsize; /**< size of the data memory map */
934 off_t me_size; /**< current file size */
935 pgno_t me_maxpg; /**< me_mapsize / me_psize */
936 txnid_t me_pgfirst; /**< ID of first old page record we used */
937 txnid_t me_pglast; /**< ID of last old page record we used */
938 MDB_dbx *me_dbxs; /**< array of static DB info */
939 uint16_t *me_dbflags; /**< array of DB flags */
940 MDB_oldpages *me_pghead; /**< list of old page records */
941 MDB_oldpages *me_pgfree; /**< list of page records to free */
942 pthread_key_t me_txkey; /**< thread-key for readers */
943 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
944 /** IDL of pages that became unused in a write txn */
946 /** ID2L of pages that were written during a write txn */
947 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
949 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
951 #elif defined(MDB_USE_POSIX_SEM)
952 sem_t *me_rmutex; /* Shared mutexes are not supported */
956 /** max number of pages to commit in one writev() call */
957 #define MDB_COMMIT_PAGES 64
958 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
959 #undef MDB_COMMIT_PAGES
960 #define MDB_COMMIT_PAGES IOV_MAX
963 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
964 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
965 static int mdb_page_touch(MDB_cursor *mc);
967 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
968 static int mdb_page_search_root(MDB_cursor *mc,
969 MDB_val *key, int modify);
970 #define MDB_PS_MODIFY 1
971 #define MDB_PS_ROOTONLY 2
972 static int mdb_page_search(MDB_cursor *mc,
973 MDB_val *key, int flags);
974 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
976 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
977 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
978 pgno_t newpgno, unsigned int nflags);
980 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
981 static int mdb_env_pick_meta(const MDB_env *env);
982 static int mdb_env_write_meta(MDB_txn *txn);
983 static void mdb_env_close0(MDB_env *env, int excl);
985 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
986 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
987 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
988 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
989 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
990 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
991 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
992 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
993 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
995 static int mdb_rebalance(MDB_cursor *mc);
996 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
998 static void mdb_cursor_pop(MDB_cursor *mc);
999 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1001 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1002 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1003 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1004 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1005 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1007 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1008 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1010 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1011 static void mdb_xcursor_init0(MDB_cursor *mc);
1012 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1014 static int mdb_drop0(MDB_cursor *mc, int subs);
1015 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1018 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1022 static SECURITY_DESCRIPTOR mdb_null_sd;
1023 static SECURITY_ATTRIBUTES mdb_all_sa;
1024 static int mdb_sec_inited;
1027 /** Return the library version info. */
1029 mdb_version(int *major, int *minor, int *patch)
1031 if (major) *major = MDB_VERSION_MAJOR;
1032 if (minor) *minor = MDB_VERSION_MINOR;
1033 if (patch) *patch = MDB_VERSION_PATCH;
1034 return MDB_VERSION_STRING;
1037 /** Table of descriptions for MDB @ref errors */
1038 static char *const mdb_errstr[] = {
1039 "MDB_KEYEXIST: Key/data pair already exists",
1040 "MDB_NOTFOUND: No matching key/data pair found",
1041 "MDB_PAGE_NOTFOUND: Requested page not found",
1042 "MDB_CORRUPTED: Located page was wrong type",
1043 "MDB_PANIC: Update of meta page failed",
1044 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1045 "MDB_INVALID: File is not an MDB file",
1046 "MDB_MAP_FULL: Environment mapsize limit reached",
1047 "MDB_DBS_FULL: Environment maxdbs limit reached",
1048 "MDB_READERS_FULL: Environment maxreaders limit reached",
1049 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1050 "MDB_TXN_FULL: Nested transaction has too many dirty pages - transaction too big",
1051 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1052 "MDB_PAGE_FULL: Internal error - page has no more space"
1056 mdb_strerror(int err)
1060 return ("Successful return: 0");
1062 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1063 i = err - MDB_KEYEXIST;
1064 return mdb_errstr[i];
1067 return strerror(err);
1071 /** Display a key in hexadecimal and return the address of the result.
1072 * @param[in] key the key to display
1073 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1074 * @return The key in hexadecimal form.
1077 mdb_dkey(MDB_val *key, char *buf)
1080 unsigned char *c = key->mv_data;
1082 if (key->mv_size > MAXKEYSIZE)
1083 return "MAXKEYSIZE";
1084 /* may want to make this a dynamic check: if the key is mostly
1085 * printable characters, print it as-is instead of converting to hex.
1089 for (i=0; i<key->mv_size; i++)
1090 ptr += sprintf(ptr, "%02x", *c++);
1092 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1097 /** Display all the keys in the page. */
1099 mdb_page_list(MDB_page *mp)
1102 unsigned int i, nkeys, nsize;
1106 nkeys = NUMKEYS(mp);
1107 fprintf(stderr, "numkeys %d\n", nkeys);
1108 for (i=0; i<nkeys; i++) {
1109 node = NODEPTR(mp, i);
1110 key.mv_size = node->mn_ksize;
1111 key.mv_data = node->mn_data;
1112 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1113 if (F_ISSET(node->mn_flags, F_BIGDATA))
1114 nsize += sizeof(pgno_t);
1116 nsize += NODEDSZ(node);
1117 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1122 mdb_cursor_chk(MDB_cursor *mc)
1128 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1129 for (i=0; i<mc->mc_top; i++) {
1131 node = NODEPTR(mp, mc->mc_ki[i]);
1132 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1135 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1141 /** Count all the pages in each DB and in the freelist
1142 * and make sure it matches the actual number of pages
1145 static void mdb_audit(MDB_txn *txn)
1149 MDB_ID freecount, count;
1154 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1155 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1156 freecount += *(MDB_ID *)data.mv_data;
1159 for (i = 0; i<txn->mt_numdbs; i++) {
1160 MDB_xcursor mx, *mxp;
1161 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1162 mdb_cursor_init(&mc, txn, i, mxp);
1163 if (txn->mt_dbs[i].md_root == P_INVALID)
1165 count += txn->mt_dbs[i].md_branch_pages +
1166 txn->mt_dbs[i].md_leaf_pages +
1167 txn->mt_dbs[i].md_overflow_pages;
1168 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1169 mdb_page_search(&mc, NULL, 0);
1173 mp = mc.mc_pg[mc.mc_top];
1174 for (j=0; j<NUMKEYS(mp); j++) {
1175 MDB_node *leaf = NODEPTR(mp, j);
1176 if (leaf->mn_flags & F_SUBDATA) {
1178 memcpy(&db, NODEDATA(leaf), sizeof(db));
1179 count += db.md_branch_pages + db.md_leaf_pages +
1180 db.md_overflow_pages;
1184 while (mdb_cursor_sibling(&mc, 1) == 0);
1187 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1188 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1189 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1195 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1197 return txn->mt_dbxs[dbi].md_cmp(a, b);
1201 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1203 if (txn->mt_dbxs[dbi].md_dcmp)
1204 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1206 return EINVAL; /* too bad you can't distinguish this from a valid result */
1209 /** Allocate a single page.
1210 * Re-use old malloc'd pages first, otherwise just malloc.
1213 mdb_page_malloc(MDB_cursor *mc) {
1215 size_t sz = mc->mc_txn->mt_env->me_psize;
1216 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1217 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1218 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1219 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1220 } else if ((ret = malloc(sz)) != NULL) {
1221 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1226 /** Allocate pages for writing.
1227 * If there are free pages available from older transactions, they
1228 * will be re-used first. Otherwise a new page will be allocated.
1229 * @param[in] mc cursor A cursor handle identifying the transaction and
1230 * database for which we are allocating.
1231 * @param[in] num the number of pages to allocate.
1232 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1233 * will always be satisfied by a single contiguous chunk of memory.
1234 * @return 0 on success, non-zero on failure.
1237 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1239 MDB_txn *txn = mc->mc_txn;
1241 pgno_t pgno = P_INVALID;
1246 /* The free list won't have any content at all until txn 2 has
1247 * committed. The pages freed by txn 2 will be unreferenced
1248 * after txn 3 commits, and so will be safe to re-use in txn 4.
1250 if (txn->mt_txnid > 3) {
1252 if (!txn->mt_env->me_pghead &&
1253 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1254 /* See if there's anything in the free DB */
1260 txnid_t *kptr, last;
1262 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1263 if (!txn->mt_env->me_pgfirst) {
1264 mdb_page_search(&m2, NULL, 0);
1265 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1266 kptr = (txnid_t *)NODEKEY(leaf);
1273 last = txn->mt_env->me_pglast + 1;
1275 key.mv_data = &last;
1276 key.mv_size = sizeof(last);
1277 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1280 last = *(txnid_t *)key.mv_data;
1283 /* Unusable if referred by a meta page or reader... */
1285 if (last < txn->mt_txnid-1) {
1286 j = txn->mt_env->me_txns->mti_numreaders;
1287 r = txn->mt_env->me_txns->mti_readers + j;
1288 for (j = -j; j && (last<r[j].mr_txnid || !r[j].mr_pid); j++) ;
1292 /* It's usable, grab it.
1297 if (!txn->mt_env->me_pgfirst) {
1298 mdb_node_read(txn, leaf, &data);
1300 txn->mt_env->me_pglast = last;
1301 if (!txn->mt_env->me_pgfirst)
1302 txn->mt_env->me_pgfirst = last;
1303 idl = (MDB_ID *) data.mv_data;
1304 /* We might have a zero-length IDL due to freelist growth
1305 * during a prior commit
1307 if (!idl[0]) goto again;
1308 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1311 mop->mo_next = txn->mt_env->me_pghead;
1312 mop->mo_txnid = last;
1313 txn->mt_env->me_pghead = mop;
1314 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1319 DPRINTF("IDL read txn %zu root %zu num %zu",
1320 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1321 for (i=0; i<idl[0]; i++) {
1322 DPRINTF("IDL %zu", idl[i+1]);
1329 if (txn->mt_env->me_pghead) {
1330 MDB_oldpages *mop = txn->mt_env->me_pghead;
1332 /* FIXME: For now, always use fresh pages. We
1333 * really ought to search the free list for a
1338 /* peel pages off tail, so we only have to truncate the list */
1339 pgno = MDB_IDL_LAST(mop->mo_pages);
1340 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1342 if (mop->mo_pages[2] > mop->mo_pages[1])
1343 mop->mo_pages[0] = 0;
1347 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1348 txn->mt_env->me_pghead = mop->mo_next;
1349 if (mc->mc_dbi == FREE_DBI) {
1350 mop->mo_next = txn->mt_env->me_pgfree;
1351 txn->mt_env->me_pgfree = mop;
1360 if (pgno == P_INVALID) {
1361 /* DB size is maxed out */
1362 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1363 DPUTS("DB size maxed out");
1364 return MDB_MAP_FULL;
1367 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1368 if (pgno == P_INVALID) {
1369 pgno = txn->mt_next_pgno;
1370 txn->mt_next_pgno += num;
1372 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1375 if (txn->mt_env->me_dpages && num == 1) {
1376 np = txn->mt_env->me_dpages;
1377 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1378 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1379 txn->mt_env->me_dpages = np->mp_next;
1381 size_t sz = txn->mt_env->me_psize * num;
1382 if ((np = malloc(sz)) == NULL)
1384 VGMEMP_ALLOC(txn->mt_env, np, sz);
1386 if (pgno == P_INVALID) {
1387 np->mp_pgno = txn->mt_next_pgno;
1388 txn->mt_next_pgno += num;
1393 mid.mid = np->mp_pgno;
1395 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1396 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1398 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1405 /** Copy a page: avoid copying unused portions of the page.
1406 * @param[in] dst page to copy into
1407 * @param[in] src page to copy from
1410 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1412 dst->mp_flags = src->mp_flags | P_DIRTY;
1413 dst->mp_pages = src->mp_pages;
1415 if (IS_LEAF2(src)) {
1416 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1418 unsigned int i, nkeys = NUMKEYS(src);
1419 for (i=0; i<nkeys; i++)
1420 dst->mp_ptrs[i] = src->mp_ptrs[i];
1421 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1422 psize - src->mp_upper);
1426 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1427 * @param[in] mc cursor pointing to the page to be touched
1428 * @return 0 on success, non-zero on failure.
1431 mdb_page_touch(MDB_cursor *mc)
1433 MDB_page *mp = mc->mc_pg[mc->mc_top];
1437 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1439 if ((rc = mdb_page_alloc(mc, 1, &np)))
1441 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1442 assert(mp->mp_pgno != np->mp_pgno);
1443 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1445 /* If page isn't full, just copy the used portion */
1446 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1449 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1451 np->mp_flags |= P_DIRTY;
1456 /* Adjust other cursors pointing to mp */
1457 if (mc->mc_flags & C_SUB) {
1458 MDB_cursor *m2, *m3;
1459 MDB_dbi dbi = mc->mc_dbi-1;
1461 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1462 if (m2 == mc) continue;
1463 m3 = &m2->mc_xcursor->mx_cursor;
1464 if (m3->mc_snum < mc->mc_snum) continue;
1465 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1466 m3->mc_pg[mc->mc_top] = mp;
1472 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1473 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1474 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1475 m2->mc_pg[mc->mc_top] = mp;
1479 mc->mc_pg[mc->mc_top] = mp;
1480 /** If this page has a parent, update the parent to point to
1484 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1486 mc->mc_db->md_root = mp->mp_pgno;
1487 } else if (mc->mc_txn->mt_parent) {
1490 /* If txn has a parent, make sure the page is in our
1493 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1494 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1495 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1496 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1497 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1498 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1499 mc->mc_pg[mc->mc_top] = mp;
1505 np = mdb_page_malloc(mc);
1508 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1509 mid.mid = np->mp_pgno;
1511 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1519 mdb_env_sync(MDB_env *env, int force)
1522 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1523 if (env->me_flags & MDB_WRITEMAP) {
1524 int flags = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
1525 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1528 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1532 if (MDB_FDATASYNC(env->me_fd))
1539 /** Make shadow copies of all of parent txn's cursors */
1541 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1543 MDB_cursor *mc, *m2;
1544 unsigned int i, j, size;
1546 for (i=0;i<src->mt_numdbs; i++) {
1547 if (src->mt_cursors[i]) {
1548 size = sizeof(MDB_cursor);
1549 if (src->mt_cursors[i]->mc_xcursor)
1550 size += sizeof(MDB_xcursor);
1551 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1558 mc->mc_db = &dst->mt_dbs[i];
1559 mc->mc_dbx = m2->mc_dbx;
1560 mc->mc_dbflag = &dst->mt_dbflags[i];
1561 mc->mc_snum = m2->mc_snum;
1562 mc->mc_top = m2->mc_top;
1563 mc->mc_flags = m2->mc_flags | C_SHADOW;
1564 for (j=0; j<mc->mc_snum; j++) {
1565 mc->mc_pg[j] = m2->mc_pg[j];
1566 mc->mc_ki[j] = m2->mc_ki[j];
1568 if (m2->mc_xcursor) {
1569 MDB_xcursor *mx, *mx2;
1570 mx = (MDB_xcursor *)(mc+1);
1571 mc->mc_xcursor = mx;
1572 mx2 = m2->mc_xcursor;
1573 mx->mx_db = mx2->mx_db;
1574 mx->mx_dbx = mx2->mx_dbx;
1575 mx->mx_dbflag = mx2->mx_dbflag;
1576 mx->mx_cursor.mc_txn = dst;
1577 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1578 mx->mx_cursor.mc_db = &mx->mx_db;
1579 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1580 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1581 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1582 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1583 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1584 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1585 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1586 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1589 mc->mc_xcursor = NULL;
1591 mc->mc_next = dst->mt_cursors[i];
1592 dst->mt_cursors[i] = mc;
1599 /** Merge shadow cursors back into parent's */
1601 mdb_cursor_merge(MDB_txn *txn)
1604 for (i=0; i<txn->mt_numdbs; i++) {
1605 if (txn->mt_cursors[i]) {
1607 while ((mc = txn->mt_cursors[i])) {
1608 txn->mt_cursors[i] = mc->mc_next;
1609 if (mc->mc_flags & C_SHADOW) {
1610 MDB_cursor *m2 = mc->mc_orig;
1612 m2->mc_snum = mc->mc_snum;
1613 m2->mc_top = mc->mc_top;
1614 for (j=0; j<mc->mc_snum; j++) {
1615 m2->mc_pg[j] = mc->mc_pg[j];
1616 m2->mc_ki[j] = mc->mc_ki[j];
1619 if (mc->mc_flags & C_ALLOCD)
1627 mdb_txn_reset0(MDB_txn *txn);
1629 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1630 * @param[in] txn the transaction handle to initialize
1631 * @return 0 on success, non-zero on failure. This can only
1632 * fail for read-only transactions, and then only if the
1633 * reader table is full.
1636 mdb_txn_renew0(MDB_txn *txn)
1638 MDB_env *env = txn->mt_env;
1643 txn->mt_numdbs = env->me_numdbs;
1644 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1646 if (txn->mt_flags & MDB_TXN_RDONLY) {
1647 if (env->me_flags & MDB_ROFS) {
1648 i = mdb_env_pick_meta(env);
1649 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1650 txn->mt_u.reader = NULL;
1652 MDB_reader *r = pthread_getspecific(env->me_txkey);
1654 pid_t pid = env->me_pid;
1655 pthread_t tid = pthread_self();
1658 for (i=0; i<env->me_txns->mti_numreaders; i++)
1659 if (env->me_txns->mti_readers[i].mr_pid == 0)
1661 if (i == env->me_maxreaders) {
1662 UNLOCK_MUTEX_R(env);
1663 return MDB_READERS_FULL;
1665 env->me_txns->mti_readers[i].mr_pid = pid;
1666 env->me_txns->mti_readers[i].mr_tid = tid;
1667 if (i >= env->me_txns->mti_numreaders)
1668 env->me_txns->mti_numreaders = i+1;
1669 /* Save numreaders for un-mutexed mdb_env_close() */
1670 env->me_numreaders = env->me_txns->mti_numreaders;
1671 UNLOCK_MUTEX_R(env);
1672 r = &env->me_txns->mti_readers[i];
1673 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1674 env->me_txns->mti_readers[i].mr_pid = 0;
1678 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1679 txn->mt_u.reader = r;
1681 txn->mt_toggle = txn->mt_txnid & 1;
1682 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1686 txn->mt_txnid = env->me_txns->mti_txnid;
1687 txn->mt_toggle = txn->mt_txnid & 1;
1688 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1691 if (txn->mt_txnid == mdb_debug_start)
1694 txn->mt_u.dirty_list = env->me_dirty_list;
1695 txn->mt_u.dirty_list[0].mid = 0;
1696 txn->mt_free_pgs = env->me_free_pgs;
1697 txn->mt_free_pgs[0] = 0;
1701 /* Copy the DB info and flags */
1702 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1703 for (i=2; i<txn->mt_numdbs; i++)
1704 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1705 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1706 if (txn->mt_numdbs > 2)
1707 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1713 mdb_txn_renew(MDB_txn *txn)
1717 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1720 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1721 DPUTS("environment had fatal error, must shutdown!");
1725 rc = mdb_txn_renew0(txn);
1726 if (rc == MDB_SUCCESS) {
1727 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1728 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1729 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1735 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1740 if (env->me_flags & MDB_FATAL_ERROR) {
1741 DPUTS("environment had fatal error, must shutdown!");
1744 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1747 /* Nested transactions: Max 1 child, write txns only, no writemap */
1748 if (parent->mt_child ||
1749 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1750 (env->me_flags & MDB_WRITEMAP))
1755 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1756 if (!(flags & MDB_RDONLY))
1757 size += env->me_maxdbs * sizeof(MDB_cursor *);
1759 if ((txn = calloc(1, size)) == NULL) {
1760 DPRINTF("calloc: %s", strerror(ErrCode()));
1763 txn->mt_dbs = (MDB_db *)(txn+1);
1764 if (flags & MDB_RDONLY) {
1765 txn->mt_flags |= MDB_TXN_RDONLY;
1766 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1768 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1769 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1774 txn->mt_free_pgs = mdb_midl_alloc();
1775 if (!txn->mt_free_pgs) {
1779 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1780 if (!txn->mt_u.dirty_list) {
1781 free(txn->mt_free_pgs);
1785 txn->mt_txnid = parent->mt_txnid;
1786 txn->mt_toggle = parent->mt_toggle;
1787 txn->mt_u.dirty_list[0].mid = 0;
1788 txn->mt_free_pgs[0] = 0;
1789 txn->mt_next_pgno = parent->mt_next_pgno;
1790 parent->mt_child = txn;
1791 txn->mt_parent = parent;
1792 txn->mt_numdbs = parent->mt_numdbs;
1793 txn->mt_dbxs = parent->mt_dbxs;
1794 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1795 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1796 mdb_cursor_shadow(parent, txn);
1799 rc = mdb_txn_renew0(txn);
1805 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1806 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1807 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1813 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1814 * @param[in] txn the transaction handle to reset
1817 mdb_txn_reset0(MDB_txn *txn)
1819 MDB_env *env = txn->mt_env;
1821 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1822 if (!(env->me_flags & MDB_ROFS))
1823 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1829 /* close(free) all cursors */
1830 for (i=0; i<txn->mt_numdbs; i++) {
1831 if (txn->mt_cursors[i]) {
1833 while ((mc = txn->mt_cursors[i])) {
1834 txn->mt_cursors[i] = mc->mc_next;
1835 if (mc->mc_flags & C_ALLOCD)
1841 if (!(env->me_flags & MDB_WRITEMAP)) {
1842 /* return all dirty pages to dpage list */
1843 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1844 dp = txn->mt_u.dirty_list[i].mptr;
1845 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1846 dp->mp_next = txn->mt_env->me_dpages;
1847 VGMEMP_FREE(txn->mt_env, dp);
1848 txn->mt_env->me_dpages = dp;
1850 /* large pages just get freed directly */
1851 VGMEMP_FREE(txn->mt_env, dp);
1857 if (txn->mt_parent) {
1858 txn->mt_parent->mt_child = NULL;
1859 mdb_midl_free(txn->mt_free_pgs);
1860 free(txn->mt_u.dirty_list);
1863 if (mdb_midl_shrink(&txn->mt_free_pgs))
1864 env->me_free_pgs = txn->mt_free_pgs;
1867 while ((mop = txn->mt_env->me_pghead)) {
1868 txn->mt_env->me_pghead = mop->mo_next;
1871 txn->mt_env->me_pgfirst = 0;
1872 txn->mt_env->me_pglast = 0;
1875 /* The writer mutex was locked in mdb_txn_begin. */
1876 UNLOCK_MUTEX_W(env);
1881 mdb_txn_reset(MDB_txn *txn)
1886 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1887 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1888 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1890 mdb_txn_reset0(txn);
1894 mdb_txn_abort(MDB_txn *txn)
1899 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1900 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1901 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1904 mdb_txn_abort(txn->mt_child);
1906 mdb_txn_reset0(txn);
1911 mdb_txn_commit(MDB_txn *txn)
1919 pgno_t next, freecnt;
1922 assert(txn != NULL);
1923 assert(txn->mt_env != NULL);
1925 if (txn->mt_child) {
1926 mdb_txn_commit(txn->mt_child);
1927 txn->mt_child = NULL;
1932 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1933 if (txn->mt_numdbs > env->me_numdbs) {
1934 /* update the DB flags */
1936 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
1937 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
1944 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1945 DPUTS("error flag is set, can't commit");
1947 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1952 if (txn->mt_parent) {
1958 /* Merge (and close) our cursors with parent's */
1959 mdb_cursor_merge(txn);
1961 /* Update parent's DB table */
1962 ip = &txn->mt_parent->mt_dbs[2];
1963 jp = &txn->mt_dbs[2];
1964 for (i = 2; i < txn->mt_numdbs; i++) {
1965 if (ip->md_root != jp->md_root)
1969 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1971 /* Append our free list to parent's */
1972 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1974 mdb_midl_free(txn->mt_free_pgs);
1976 /* Merge our dirty list with parent's */
1977 dst = txn->mt_parent->mt_u.dirty_list;
1978 src = txn->mt_u.dirty_list;
1979 x = mdb_mid2l_search(dst, src[1].mid);
1980 for (y=1; y<=src[0].mid; y++) {
1981 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1985 dst[x].mptr = src[y].mptr;
1988 for (; y<=src[0].mid; y++) {
1989 if (++x >= MDB_IDL_UM_MAX) {
1991 return MDB_TXN_FULL;
1996 free(txn->mt_u.dirty_list);
1997 txn->mt_parent->mt_child = NULL;
2002 if (txn != env->me_txn) {
2003 DPUTS("attempt to commit unknown transaction");
2008 if (!txn->mt_u.dirty_list[0].mid)
2011 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2012 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2014 /* Update DB root pointers. Their pages have already been
2015 * touched so this is all in-place and cannot fail.
2017 if (txn->mt_numdbs > 2) {
2020 data.mv_size = sizeof(MDB_db);
2022 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2023 for (i = 2; i < txn->mt_numdbs; i++) {
2024 if (txn->mt_dbflags[i] & DB_DIRTY) {
2025 data.mv_data = &txn->mt_dbs[i];
2026 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2031 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2033 /* should only be one record now */
2034 if (env->me_pghead) {
2035 /* make sure first page of freeDB is touched and on freelist */
2036 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2039 /* Delete IDLs we used from the free list */
2040 if (env->me_pgfirst) {
2045 key.mv_size = sizeof(cur);
2046 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2049 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2050 rc = mdb_cursor_del(&mc, 0);
2056 env->me_pgfirst = 0;
2060 /* save to free list */
2062 freecnt = txn->mt_free_pgs[0];
2063 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2066 /* make sure last page of freeDB is touched and on freelist */
2067 key.mv_size = MAXKEYSIZE+1;
2069 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2071 mdb_midl_sort(txn->mt_free_pgs);
2075 MDB_IDL idl = txn->mt_free_pgs;
2076 DPRINTF("IDL write txn %zu root %zu num %zu",
2077 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2078 for (i=0; i<idl[0]; i++) {
2079 DPRINTF("IDL %zu", idl[i+1]);
2083 /* write to last page of freeDB */
2084 key.mv_size = sizeof(pgno_t);
2085 key.mv_data = &txn->mt_txnid;
2086 data.mv_data = txn->mt_free_pgs;
2087 /* The free list can still grow during this call,
2088 * despite the pre-emptive touches above. So check
2089 * and make sure the entire thing got written.
2092 freecnt = txn->mt_free_pgs[0];
2093 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2094 rc = mdb_cursor_put(&mc, &key, &data, 0);
2099 } while (freecnt != txn->mt_free_pgs[0]);
2101 /* should only be one record now */
2103 if (env->me_pghead) {
2109 mop = env->me_pghead;
2111 key.mv_size = sizeof(id);
2113 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2114 data.mv_data = mop->mo_pages;
2115 orig = mop->mo_pages[0];
2116 /* These steps may grow the freelist again
2117 * due to freed overflow pages...
2119 mdb_cursor_put(&mc, &key, &data, 0);
2120 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2121 /* could have been used again here */
2122 if (mop->mo_pages[0] != orig) {
2123 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2124 data.mv_data = mop->mo_pages;
2126 mdb_cursor_put(&mc, &key, &data, 0);
2128 env->me_pghead = NULL;
2131 /* was completely used up */
2132 mdb_cursor_del(&mc, 0);
2136 env->me_pgfirst = 0;
2140 while (env->me_pgfree) {
2141 MDB_oldpages *mop = env->me_pgfree;
2142 env->me_pgfree = mop->mo_next;
2146 /* Check for growth of freelist again */
2147 if (freecnt != txn->mt_free_pgs[0])
2150 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2151 if (mdb_midl_shrink(&txn->mt_free_pgs))
2152 env->me_free_pgs = txn->mt_free_pgs;
2159 if (env->me_flags & MDB_WRITEMAP) {
2160 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2161 dp = txn->mt_u.dirty_list[i].mptr;
2162 /* clear dirty flag */
2163 dp->mp_flags &= ~P_DIRTY;
2164 txn->mt_u.dirty_list[i].mid = 0;
2166 txn->mt_u.dirty_list[0].mid = 0;
2170 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2176 /* Windows actually supports scatter/gather I/O, but only on
2177 * unbuffered file handles. Since we're relying on the OS page
2178 * cache for all our data, that's self-defeating. So we just
2179 * write pages one at a time. We use the ov structure to set
2180 * the write offset, to at least save the overhead of a Seek
2184 memset(&ov, 0, sizeof(ov));
2185 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2187 dp = txn->mt_u.dirty_list[i].mptr;
2188 DPRINTF("committing page %zu", dp->mp_pgno);
2189 size = dp->mp_pgno * env->me_psize;
2190 ov.Offset = size & 0xffffffff;
2191 ov.OffsetHigh = size >> 16;
2192 ov.OffsetHigh >>= 16;
2193 /* clear dirty flag */
2194 dp->mp_flags &= ~P_DIRTY;
2195 wsize = env->me_psize;
2196 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2197 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2200 DPRINTF("WriteFile: %d", n);
2207 struct iovec iov[MDB_COMMIT_PAGES];
2211 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2212 dp = txn->mt_u.dirty_list[i].mptr;
2213 if (dp->mp_pgno != next) {
2215 rc = writev(env->me_fd, iov, n);
2219 DPUTS("short write, filesystem full?");
2221 DPRINTF("writev: %s", strerror(n));
2228 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2231 DPRINTF("committing page %zu", dp->mp_pgno);
2232 iov[n].iov_len = env->me_psize;
2233 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2234 iov[n].iov_base = (char *)dp;
2235 size += iov[n].iov_len;
2236 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2237 /* clear dirty flag */
2238 dp->mp_flags &= ~P_DIRTY;
2239 if (++n >= MDB_COMMIT_PAGES) {
2249 rc = writev(env->me_fd, iov, n);
2253 DPUTS("short write, filesystem full?");
2255 DPRINTF("writev: %s", strerror(n));
2262 /* Drop the dirty pages.
2264 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2265 dp = txn->mt_u.dirty_list[i].mptr;
2266 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2267 dp->mp_next = txn->mt_env->me_dpages;
2268 VGMEMP_FREE(txn->mt_env, dp);
2269 txn->mt_env->me_dpages = dp;
2271 VGMEMP_FREE(txn->mt_env, dp);
2274 txn->mt_u.dirty_list[i].mid = 0;
2276 txn->mt_u.dirty_list[0].mid = 0;
2279 if ((n = mdb_env_sync(env, 0)) != 0 ||
2280 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2287 if (txn->mt_numdbs > env->me_numdbs) {
2288 /* update the DB flags */
2290 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2291 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2295 UNLOCK_MUTEX_W(env);
2301 /** Read the environment parameters of a DB environment before
2302 * mapping it into memory.
2303 * @param[in] env the environment handle
2304 * @param[out] meta address of where to store the meta information
2305 * @return 0 on success, non-zero on failure.
2308 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2315 /* We don't know the page size yet, so use a minimum value.
2319 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2321 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2326 else if (rc != MDB_PAGESIZE) {
2330 DPRINTF("read: %s", strerror(err));
2334 p = (MDB_page *)&pbuf;
2336 if (!F_ISSET(p->mp_flags, P_META)) {
2337 DPRINTF("page %zu not a meta page", p->mp_pgno);
2342 if (m->mm_magic != MDB_MAGIC) {
2343 DPUTS("meta has invalid magic");
2347 if (m->mm_version != MDB_VERSION) {
2348 DPRINTF("database is version %u, expected version %u",
2349 m->mm_version, MDB_VERSION);
2350 return MDB_VERSION_MISMATCH;
2353 memcpy(meta, m, sizeof(*m));
2357 /** Write the environment parameters of a freshly created DB environment.
2358 * @param[in] env the environment handle
2359 * @param[out] meta address of where to store the meta information
2360 * @return 0 on success, non-zero on failure.
2363 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2370 DPUTS("writing new meta page");
2372 GET_PAGESIZE(psize);
2374 meta->mm_magic = MDB_MAGIC;
2375 meta->mm_version = MDB_VERSION;
2376 meta->mm_psize = psize;
2377 meta->mm_last_pg = 1;
2378 meta->mm_flags = env->me_flags & 0xffff;
2379 meta->mm_flags |= MDB_INTEGERKEY;
2380 meta->mm_dbs[0].md_root = P_INVALID;
2381 meta->mm_dbs[1].md_root = P_INVALID;
2383 p = calloc(2, psize);
2385 p->mp_flags = P_META;
2388 memcpy(m, meta, sizeof(*meta));
2390 q = (MDB_page *)((char *)p + psize);
2393 q->mp_flags = P_META;
2396 memcpy(m, meta, sizeof(*meta));
2401 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2402 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2405 rc = write(env->me_fd, p, psize * 2);
2406 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2412 /** Update the environment info to commit a transaction.
2413 * @param[in] txn the transaction that's being committed
2414 * @return 0 on success, non-zero on failure.
2417 mdb_env_write_meta(MDB_txn *txn)
2420 MDB_meta meta, metab, *mp;
2422 int rc, len, toggle;
2428 assert(txn != NULL);
2429 assert(txn->mt_env != NULL);
2431 toggle = !txn->mt_toggle;
2432 DPRINTF("writing meta page %d for root page %zu",
2433 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2436 mp = env->me_metas[toggle];
2438 if (env->me_flags & MDB_WRITEMAP) {
2439 /* Persist any increases of mapsize config */
2440 if (env->me_mapsize > mp->mm_mapsize)
2441 mp->mm_mapsize = env->me_mapsize;
2442 mp->mm_dbs[0] = txn->mt_dbs[0];
2443 mp->mm_dbs[1] = txn->mt_dbs[1];
2444 mp->mm_last_pg = txn->mt_next_pgno - 1;
2445 mp->mm_txnid = txn->mt_txnid;
2446 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2447 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2450 ptr += env->me_psize;
2451 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2458 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2459 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2461 ptr = (char *)&meta;
2462 if (env->me_mapsize > mp->mm_mapsize) {
2463 /* Persist any increases of mapsize config */
2464 meta.mm_mapsize = env->me_mapsize;
2465 off = offsetof(MDB_meta, mm_mapsize);
2467 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2469 len = sizeof(MDB_meta) - off;
2472 meta.mm_dbs[0] = txn->mt_dbs[0];
2473 meta.mm_dbs[1] = txn->mt_dbs[1];
2474 meta.mm_last_pg = txn->mt_next_pgno - 1;
2475 meta.mm_txnid = txn->mt_txnid;
2478 off += env->me_psize;
2481 /* Write to the SYNC fd */
2484 memset(&ov, 0, sizeof(ov));
2486 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2489 rc = pwrite(env->me_mfd, ptr, len, off);
2494 DPUTS("write failed, disk error?");
2495 /* On a failure, the pagecache still contains the new data.
2496 * Write some old data back, to prevent it from being used.
2497 * Use the non-SYNC fd; we know it will fail anyway.
2499 meta.mm_last_pg = metab.mm_last_pg;
2500 meta.mm_txnid = metab.mm_txnid;
2502 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2504 r2 = pwrite(env->me_fd, ptr, len, off);
2507 env->me_flags |= MDB_FATAL_ERROR;
2511 /* Memory ordering issues are irrelevant; since the entire writer
2512 * is wrapped by wmutex, all of these changes will become visible
2513 * after the wmutex is unlocked. Since the DB is multi-version,
2514 * readers will get consistent data regardless of how fresh or
2515 * how stale their view of these values is.
2517 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2522 /** Check both meta pages to see which one is newer.
2523 * @param[in] env the environment handle
2524 * @return meta toggle (0 or 1).
2527 mdb_env_pick_meta(const MDB_env *env)
2529 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2533 mdb_env_create(MDB_env **env)
2537 e = calloc(1, sizeof(MDB_env));
2541 e->me_free_pgs = mdb_midl_alloc();
2542 if (!e->me_free_pgs) {
2546 e->me_maxreaders = DEFAULT_READERS;
2548 e->me_fd = INVALID_HANDLE_VALUE;
2549 e->me_lfd = INVALID_HANDLE_VALUE;
2550 e->me_mfd = INVALID_HANDLE_VALUE;
2551 #ifdef MDB_USE_POSIX_SEM
2552 e->me_rmutex = SEM_FAILED;
2553 e->me_wmutex = SEM_FAILED;
2555 e->me_pid = getpid();
2556 VGMEMP_CREATE(e,0,0);
2562 mdb_env_set_mapsize(MDB_env *env, size_t size)
2566 env->me_mapsize = size;
2568 env->me_maxpg = env->me_mapsize / env->me_psize;
2573 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2577 env->me_maxdbs = dbs;
2582 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2584 if (env->me_map || readers < 1)
2586 env->me_maxreaders = readers;
2591 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2593 if (!env || !readers)
2595 *readers = env->me_maxreaders;
2599 /** Further setup required for opening an MDB environment
2602 mdb_env_open2(MDB_env *env)
2604 unsigned int flags = env->me_flags;
2605 int i, newenv = 0, prot;
2609 memset(&meta, 0, sizeof(meta));
2611 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2614 DPUTS("new mdbenv");
2616 meta.mm_mapsize = env->me_mapsize > DEFAULT_MAPSIZE ? env->me_mapsize : DEFAULT_MAPSIZE;
2619 if (env->me_mapsize < meta.mm_mapsize)
2620 env->me_mapsize = meta.mm_mapsize;
2625 LONG sizelo, sizehi;
2626 sizelo = env->me_mapsize & 0xffffffff;
2627 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2629 /* Windows won't create mappings for zero length files.
2630 * Just allocate the maxsize right now.
2633 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2634 if (!SetEndOfFile(env->me_fd))
2636 SetFilePointer(env->me_fd, 0, NULL, 0);
2638 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2639 PAGE_READWRITE : PAGE_READONLY,
2640 sizehi, sizelo, NULL);
2643 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2644 FILE_MAP_WRITE : FILE_MAP_READ,
2645 0, 0, env->me_mapsize, meta.mm_address);
2652 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2655 if (flags & MDB_WRITEMAP) {
2657 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2660 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2662 if (env->me_map == MAP_FAILED) {
2669 if (flags & MDB_FIXEDMAP)
2670 meta.mm_address = env->me_map;
2671 i = mdb_env_init_meta(env, &meta);
2672 if (i != MDB_SUCCESS) {
2676 env->me_psize = meta.mm_psize;
2678 env->me_maxpg = env->me_mapsize / env->me_psize;
2680 p = (MDB_page *)env->me_map;
2681 env->me_metas[0] = METADATA(p);
2682 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2686 int toggle = mdb_env_pick_meta(env);
2687 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2689 DPRINTF("opened database version %u, pagesize %u",
2690 env->me_metas[0]->mm_version, env->me_psize);
2691 DPRINTF("using meta page %d", toggle);
2692 DPRINTF("depth: %u", db->md_depth);
2693 DPRINTF("entries: %zu", db->md_entries);
2694 DPRINTF("branch pages: %zu", db->md_branch_pages);
2695 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2696 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2697 DPRINTF("root: %zu", db->md_root);
2705 /** Release a reader thread's slot in the reader lock table.
2706 * This function is called automatically when a thread exits.
2707 * @param[in] ptr This points to the slot in the reader lock table.
2710 mdb_env_reader_dest(void *ptr)
2712 MDB_reader *reader = ptr;
2718 /** Junk for arranging thread-specific callbacks on Windows. This is
2719 * necessarily platform and compiler-specific. Windows supports up
2720 * to 1088 keys. Let's assume nobody opens more than 64 environments
2721 * in a single process, for now. They can override this if needed.
2723 #ifndef MAX_TLS_KEYS
2724 #define MAX_TLS_KEYS 64
2726 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2727 static int mdb_tls_nkeys;
2729 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2733 case DLL_PROCESS_ATTACH: break;
2734 case DLL_THREAD_ATTACH: break;
2735 case DLL_THREAD_DETACH:
2736 for (i=0; i<mdb_tls_nkeys; i++) {
2737 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2738 mdb_env_reader_dest(r);
2741 case DLL_PROCESS_DETACH: break;
2746 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2748 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2752 /* Force some symbol references.
2753 * _tls_used forces the linker to create the TLS directory if not already done
2754 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2756 #pragma comment(linker, "/INCLUDE:_tls_used")
2757 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2758 #pragma const_seg(".CRT$XLB")
2759 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2760 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2763 #pragma comment(linker, "/INCLUDE:__tls_used")
2764 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2765 #pragma data_seg(".CRT$XLB")
2766 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2768 #endif /* WIN 32/64 */
2769 #endif /* !__GNUC__ */
2772 /** Downgrade the exclusive lock on the region back to shared */
2774 mdb_env_share_locks(MDB_env *env, int *excl)
2776 int rc = 0, toggle = mdb_env_pick_meta(env);
2778 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2783 /* First acquire a shared lock. The Unlock will
2784 * then release the existing exclusive lock.
2786 memset(&ov, 0, sizeof(ov));
2787 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2790 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2796 struct flock lock_info;
2797 /* The shared lock replaces the existing lock */
2798 memset((void *)&lock_info, 0, sizeof(lock_info));
2799 lock_info.l_type = F_RDLCK;
2800 lock_info.l_whence = SEEK_SET;
2801 lock_info.l_start = 0;
2802 lock_info.l_len = 1;
2803 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2804 (rc = ErrCode()) == EINTR) ;
2805 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2812 /** Try to get exlusive lock, otherwise shared.
2813 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2816 mdb_env_excl_lock(MDB_env *env, int *excl)
2820 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2824 memset(&ov, 0, sizeof(ov));
2825 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2832 struct flock lock_info;
2833 memset((void *)&lock_info, 0, sizeof(lock_info));
2834 lock_info.l_type = F_WRLCK;
2835 lock_info.l_whence = SEEK_SET;
2836 lock_info.l_start = 0;
2837 lock_info.l_len = 1;
2838 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2839 (rc = ErrCode()) == EINTR) ;
2843 # ifdef MDB_USE_POSIX_SEM
2844 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
2847 lock_info.l_type = F_RDLCK;
2848 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
2849 (rc = ErrCode()) == EINTR) ;
2857 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
2859 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2861 * @(#) $Revision: 5.1 $
2862 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2863 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2865 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2869 * Please do not copyright this code. This code is in the public domain.
2871 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2872 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2873 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2874 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2875 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2876 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2877 * PERFORMANCE OF THIS SOFTWARE.
2880 * chongo <Landon Curt Noll> /\oo/\
2881 * http://www.isthe.com/chongo/
2883 * Share and Enjoy! :-)
2886 typedef unsigned long long mdb_hash_t;
2887 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2889 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2890 * @param[in] str string to hash
2891 * @param[in] hval initial value for hash
2892 * @return 64 bit hash
2894 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2895 * hval arg on the first call.
2898 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
2900 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
2901 unsigned char *end = s + val->mv_size;
2903 * FNV-1a hash each octet of the string
2906 /* xor the bottom with the current octet */
2907 hval ^= (mdb_hash_t)*s++;
2909 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2910 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2911 (hval << 7) + (hval << 8) + (hval << 40);
2913 /* return our new hash value */
2917 /** Hash the string and output the hash in hex.
2918 * @param[in] str string to hash
2919 * @param[out] hexbuf an array of 17 chars to hold the hash
2922 mdb_hash_hex(MDB_val *val, char *hexbuf)
2925 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
2926 for (i=0; i<8; i++) {
2927 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2933 /** Open and/or initialize the lock region for the environment.
2934 * @param[in] env The MDB environment.
2935 * @param[in] lpath The pathname of the file used for the lock region.
2936 * @param[in] mode The Unix permissions for the file, if we create it.
2937 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
2938 * @return 0 on success, non-zero on failure.
2941 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2949 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2950 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2951 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2953 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
2954 env->me_flags |= MDB_ROFS;
2959 /* Try to get exclusive lock. If we succeed, then
2960 * nobody is using the lock region and we should initialize it.
2962 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2963 size = GetFileSize(env->me_lfd, NULL);
2969 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2971 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
2972 env->me_flags |= MDB_ROFS;
2977 /* Lose record locks when exec*() */
2978 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
2979 fcntl(env->me_lfd, F_SETFD, fdflags);
2981 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
2982 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
2984 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
2985 env->me_flags |= MDB_ROFS;
2992 /* Try to get exclusive lock. If we succeed, then
2993 * nobody is using the lock region and we should initialize it.
2995 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2997 size = lseek(env->me_lfd, 0, SEEK_END);
2999 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3000 if (size < rsize && *excl > 0) {
3002 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3003 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3005 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3009 size = rsize - sizeof(MDB_txninfo);
3010 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3015 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3017 if (!mh) goto fail_errno;
3018 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3020 if (!env->me_txns) goto fail_errno;
3022 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3024 if (m == MAP_FAILED) goto fail_errno;
3030 BY_HANDLE_FILE_INFORMATION stbuf;
3039 if (!mdb_sec_inited) {
3040 InitializeSecurityDescriptor(&mdb_null_sd,
3041 SECURITY_DESCRIPTOR_REVISION);
3042 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3043 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3044 mdb_all_sa.bInheritHandle = FALSE;
3045 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3048 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3049 idbuf.volume = stbuf.dwVolumeSerialNumber;
3050 idbuf.nhigh = stbuf.nFileIndexHigh;
3051 idbuf.nlow = stbuf.nFileIndexLow;
3052 val.mv_data = &idbuf;
3053 val.mv_size = sizeof(idbuf);
3054 mdb_hash_hex(&val, hexbuf);
3055 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3056 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3057 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3058 if (!env->me_rmutex) goto fail_errno;
3059 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3060 if (!env->me_wmutex) goto fail_errno;
3061 #elif defined(MDB_USE_POSIX_SEM)
3070 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3071 idbuf.dev = stbuf.st_dev;
3072 idbuf.ino = stbuf.st_ino;
3073 val.mv_data = &idbuf;
3074 val.mv_size = sizeof(idbuf);
3075 mdb_hash_hex(&val, hexbuf);
3076 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3077 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3078 /* Clean up after a previous run, if needed: Try to
3079 * remove both semaphores before doing anything else.
3081 sem_unlink(env->me_txns->mti_rmname);
3082 sem_unlink(env->me_txns->mti_wmname);
3083 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3084 O_CREAT|O_EXCL, mode, 1);
3085 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3086 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3087 O_CREAT|O_EXCL, mode, 1);
3088 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3089 #else /* MDB_USE_POSIX_SEM */
3090 pthread_mutexattr_t mattr;
3092 if ((rc = pthread_mutexattr_init(&mattr))
3093 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3094 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3095 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3097 pthread_mutexattr_destroy(&mattr);
3098 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3100 env->me_txns->mti_version = MDB_VERSION;
3101 env->me_txns->mti_magic = MDB_MAGIC;
3102 env->me_txns->mti_txnid = 0;
3103 env->me_txns->mti_numreaders = 0;
3106 if (env->me_txns->mti_magic != MDB_MAGIC) {
3107 DPUTS("lock region has invalid magic");
3111 if (env->me_txns->mti_version != MDB_VERSION) {
3112 DPRINTF("lock region is version %u, expected version %u",
3113 env->me_txns->mti_version, MDB_VERSION);
3114 rc = MDB_VERSION_MISMATCH;
3118 if (rc != EACCES && rc != EAGAIN) {
3122 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3123 if (!env->me_rmutex) goto fail_errno;
3124 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3125 if (!env->me_wmutex) goto fail_errno;
3126 #elif defined(MDB_USE_POSIX_SEM)
3127 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3128 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3129 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3130 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3141 /** The name of the lock file in the DB environment */
3142 #define LOCKNAME "/lock.mdb"
3143 /** The name of the data file in the DB environment */
3144 #define DATANAME "/data.mdb"
3145 /** The suffix of the lock file when no subdir is used */
3146 #define LOCKSUFF "-lock"
3149 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3151 int oflags, rc, len, excl;
3152 char *lpath, *dpath;
3154 if (env->me_fd != INVALID_HANDLE_VALUE)
3158 if (flags & MDB_NOSUBDIR) {
3159 rc = len + sizeof(LOCKSUFF) + len + 1;
3161 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3166 if (flags & MDB_NOSUBDIR) {
3167 dpath = lpath + len + sizeof(LOCKSUFF);
3168 sprintf(lpath, "%s" LOCKSUFF, path);
3169 strcpy(dpath, path);
3171 dpath = lpath + len + sizeof(LOCKNAME);
3172 sprintf(lpath, "%s" LOCKNAME, path);
3173 sprintf(dpath, "%s" DATANAME, path);
3176 flags |= env->me_flags;
3177 /* silently ignore WRITEMAP if we're only getting read access */
3178 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3179 flags ^= MDB_WRITEMAP;
3180 env->me_flags = flags |= MDB_ENV_ACTIVE;
3182 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3187 if (F_ISSET(flags, MDB_RDONLY)) {
3188 oflags = GENERIC_READ;
3189 len = OPEN_EXISTING;
3191 oflags = GENERIC_READ|GENERIC_WRITE;
3194 mode = FILE_ATTRIBUTE_NORMAL;
3195 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3196 NULL, len, mode, NULL);
3198 if (F_ISSET(flags, MDB_RDONLY))
3201 oflags = O_RDWR | O_CREAT;
3203 env->me_fd = open(dpath, oflags, mode);
3205 if (env->me_fd == INVALID_HANDLE_VALUE) {
3210 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3211 if (flags & (MDB_RDONLY|MDB_NOSYNC|MDB_NOMETASYNC|MDB_WRITEMAP)) {
3212 env->me_mfd = env->me_fd;
3214 /* synchronous fd for meta writes */
3216 env->me_mfd = CreateFile(dpath, oflags,
3217 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3218 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3220 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3222 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3227 DPRINTF("opened dbenv %p", (void *) env);
3228 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3231 env->me_numdbs = 2; /* this notes that me_txkey was set */
3233 /* Windows TLS callbacks need help finding their TLS info. */
3234 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3235 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3242 rc = mdb_env_share_locks(env, &excl);
3246 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3247 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3248 env->me_path = strdup(path);
3249 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3255 mdb_env_close0(env, excl);
3261 /** Destroy resources from mdb_env_open() and clear our readers */
3263 mdb_env_close0(MDB_env *env, int excl)
3267 if (!(env->me_flags & MDB_ENV_ACTIVE))
3270 free(env->me_dbflags);
3274 if (env->me_numdbs) {
3275 pthread_key_delete(env->me_txkey);
3277 /* Delete our key from the global list */
3278 for (i=0; i<mdb_tls_nkeys; i++)
3279 if (mdb_tls_keys[i] == env->me_txkey) {
3280 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3288 munmap(env->me_map, env->me_mapsize);
3290 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3292 if (env->me_fd != INVALID_HANDLE_VALUE)
3295 pid_t pid = env->me_pid;
3296 /* Clearing readers is done in this function because
3297 * me_txkey with its destructor must be disabled first.
3299 for (i = env->me_numreaders; --i >= 0; )
3300 if (env->me_txns->mti_readers[i].mr_pid == pid)
3301 env->me_txns->mti_readers[i].mr_pid = 0;
3303 if (env->me_rmutex) {
3304 CloseHandle(env->me_rmutex);
3305 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3307 /* Windows automatically destroys the mutexes when
3308 * the last handle closes.
3310 #elif defined(MDB_USE_POSIX_SEM)
3311 if (env->me_rmutex != SEM_FAILED) {
3312 sem_close(env->me_rmutex);
3313 if (env->me_wmutex != SEM_FAILED)
3314 sem_close(env->me_wmutex);
3315 /* If we have the filelock: If we are the
3316 * only remaining user, clean up semaphores.
3319 mdb_env_excl_lock(env, &excl);
3321 sem_unlink(env->me_txns->mti_rmname);
3322 sem_unlink(env->me_txns->mti_wmname);
3326 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3328 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3331 /* Unlock the lockfile. Windows would have unlocked it
3332 * after closing anyway, but not necessarily at once.
3334 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3340 env->me_flags &= ~MDB_ENV_ACTIVE;
3344 mdb_env_copy(MDB_env *env, const char *path)
3346 MDB_txn *txn = NULL;
3350 HANDLE newfd = INVALID_HANDLE_VALUE;
3352 if (env->me_flags & MDB_NOSUBDIR) {
3353 lpath = (char *)path;
3356 len += sizeof(DATANAME);
3357 lpath = malloc(len);
3360 sprintf(lpath, "%s" DATANAME, path);
3363 /* The destination path must exist, but the destination file must not.
3364 * We don't want the OS to cache the writes, since the source data is
3365 * already in the OS cache.
3368 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3369 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3371 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3377 if (!(env->me_flags & MDB_NOSUBDIR))
3379 if (newfd == INVALID_HANDLE_VALUE) {
3384 #ifdef F_NOCACHE /* __APPLE__ */
3385 rc = fcntl(newfd, F_NOCACHE, 1);
3392 /* Do the lock/unlock of the reader mutex before starting the
3393 * write txn. Otherwise other read txns could block writers.
3395 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3399 if (!(env->me_flags & MDB_ROFS)) {
3400 /* We must start the actual read txn after blocking writers */
3401 mdb_txn_reset0(txn);
3403 /* Temporarily block writers until we snapshot the meta pages */
3406 rc = mdb_txn_renew0(txn);
3408 UNLOCK_MUTEX_W(env);
3413 wsize = env->me_psize * 2;
3417 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3418 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3421 rc = write(newfd, env->me_map, wsize);
3422 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3424 if (! (env->me_flags & MDB_ROFS))
3425 UNLOCK_MUTEX_W(env);
3430 ptr = env->me_map + wsize;
3431 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3435 rc = WriteFile(newfd, ptr, wsize, &len, NULL);
3436 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3439 rc = write(newfd, ptr, wsize);
3440 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3445 if (newfd != INVALID_HANDLE_VALUE)
3452 mdb_env_close(MDB_env *env)
3459 VGMEMP_DESTROY(env);
3460 while ((dp = env->me_dpages) != NULL) {
3461 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3462 env->me_dpages = dp->mp_next;
3466 mdb_env_close0(env, 0);
3467 mdb_midl_free(env->me_free_pgs);
3471 /** Compare two items pointing at aligned size_t's */
3473 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3475 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3476 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3479 /** Compare two items pointing at aligned int's */
3481 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3483 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3484 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3487 /** Compare two items pointing at ints of unknown alignment.
3488 * Nodes and keys are guaranteed to be 2-byte aligned.
3491 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3493 #if BYTE_ORDER == LITTLE_ENDIAN
3494 unsigned short *u, *c;
3497 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3498 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3501 } while(!x && u > (unsigned short *)a->mv_data);
3504 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3508 /** Compare two items lexically */
3510 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3517 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3523 diff = memcmp(a->mv_data, b->mv_data, len);
3524 return diff ? diff : len_diff<0 ? -1 : len_diff;
3527 /** Compare two items in reverse byte order */
3529 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3531 const unsigned char *p1, *p2, *p1_lim;
3535 p1_lim = (const unsigned char *)a->mv_data;
3536 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3537 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3539 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3545 while (p1 > p1_lim) {
3546 diff = *--p1 - *--p2;
3550 return len_diff<0 ? -1 : len_diff;
3553 /** Search for key within a page, using binary search.
3554 * Returns the smallest entry larger or equal to the key.
3555 * If exactp is non-null, stores whether the found entry was an exact match
3556 * in *exactp (1 or 0).
3557 * Updates the cursor index with the index of the found entry.
3558 * If no entry larger or equal to the key is found, returns NULL.
3561 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3563 unsigned int i = 0, nkeys;
3566 MDB_page *mp = mc->mc_pg[mc->mc_top];
3567 MDB_node *node = NULL;
3572 nkeys = NUMKEYS(mp);
3577 COPY_PGNO(pgno, mp->mp_pgno);
3578 DPRINTF("searching %u keys in %s %spage %zu",
3579 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3586 low = IS_LEAF(mp) ? 0 : 1;
3588 cmp = mc->mc_dbx->md_cmp;
3590 /* Branch pages have no data, so if using integer keys,
3591 * alignment is guaranteed. Use faster mdb_cmp_int.
3593 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3594 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3601 nodekey.mv_size = mc->mc_db->md_pad;
3602 node = NODEPTR(mp, 0); /* fake */
3603 while (low <= high) {
3604 i = (low + high) >> 1;
3605 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3606 rc = cmp(key, &nodekey);
3607 DPRINTF("found leaf index %u [%s], rc = %i",
3608 i, DKEY(&nodekey), rc);
3617 while (low <= high) {
3618 i = (low + high) >> 1;
3620 node = NODEPTR(mp, i);
3621 nodekey.mv_size = NODEKSZ(node);
3622 nodekey.mv_data = NODEKEY(node);
3624 rc = cmp(key, &nodekey);
3627 DPRINTF("found leaf index %u [%s], rc = %i",
3628 i, DKEY(&nodekey), rc);
3630 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3631 i, DKEY(&nodekey), NODEPGNO(node), rc);
3642 if (rc > 0) { /* Found entry is less than the key. */
3643 i++; /* Skip to get the smallest entry larger than key. */
3645 node = NODEPTR(mp, i);
3648 *exactp = (rc == 0);
3649 /* store the key index */
3650 mc->mc_ki[mc->mc_top] = i;
3652 /* There is no entry larger or equal to the key. */
3655 /* nodeptr is fake for LEAF2 */
3661 mdb_cursor_adjust(MDB_cursor *mc, func)
3665 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3666 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3673 /** Pop a page off the top of the cursor's stack. */
3675 mdb_cursor_pop(MDB_cursor *mc)
3679 MDB_page *top = mc->mc_pg[mc->mc_top];
3685 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3686 mc->mc_dbi, (void *) mc);
3690 /** Push a page onto the top of the cursor's stack. */
3692 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3694 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3695 mc->mc_dbi, (void *) mc);
3697 if (mc->mc_snum >= CURSOR_STACK) {
3698 assert(mc->mc_snum < CURSOR_STACK);
3699 return MDB_CURSOR_FULL;
3702 mc->mc_top = mc->mc_snum++;
3703 mc->mc_pg[mc->mc_top] = mp;
3704 mc->mc_ki[mc->mc_top] = 0;
3709 /** Find the address of the page corresponding to a given page number.
3710 * @param[in] txn the transaction for this access.
3711 * @param[in] pgno the page number for the page to retrieve.
3712 * @param[out] ret address of a pointer where the page's address will be stored.
3713 * @return 0 on success, non-zero on failure.
3716 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3720 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3721 if (pgno < txn->mt_next_pgno)
3722 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3725 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3727 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3728 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3729 p = txn->mt_u.dirty_list[x].mptr;
3733 if (pgno < txn->mt_next_pgno)
3734 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3739 DPRINTF("page %zu not found", pgno);
3742 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3745 /** Search for the page a given key should be in.
3746 * Pushes parent pages on the cursor stack. This function continues a
3747 * search on a cursor that has already been initialized. (Usually by
3748 * #mdb_page_search() but also by #mdb_node_move().)
3749 * @param[in,out] mc the cursor for this operation.
3750 * @param[in] key the key to search for. If NULL, search for the lowest
3751 * page. (This is used by #mdb_cursor_first().)
3752 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3753 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3754 * @return 0 on success, non-zero on failure.
3757 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3759 MDB_page *mp = mc->mc_pg[mc->mc_top];
3764 while (IS_BRANCH(mp)) {
3768 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3769 assert(NUMKEYS(mp) > 1);
3770 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3772 if (key == NULL) /* Initialize cursor to first page. */
3774 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3775 /* cursor to last page */
3779 node = mdb_node_search(mc, key, &exact);
3781 i = NUMKEYS(mp) - 1;
3783 i = mc->mc_ki[mc->mc_top];
3792 DPRINTF("following index %u for key [%s]",
3794 assert(i < NUMKEYS(mp));
3795 node = NODEPTR(mp, i);
3797 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3800 mc->mc_ki[mc->mc_top] = i;
3801 if ((rc = mdb_cursor_push(mc, mp)))
3805 if ((rc = mdb_page_touch(mc)) != 0)
3807 mp = mc->mc_pg[mc->mc_top];
3812 DPRINTF("internal error, index points to a %02X page!?",
3814 return MDB_CORRUPTED;
3817 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3818 key ? DKEY(key) : NULL);
3823 /** Search for the page a given key should be in.
3824 * Pushes parent pages on the cursor stack. This function just sets up
3825 * the search; it finds the root page for \b mc's database and sets this
3826 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3827 * called to complete the search.
3828 * @param[in,out] mc the cursor for this operation.
3829 * @param[in] key the key to search for. If NULL, search for the lowest
3830 * page. (This is used by #mdb_cursor_first().)
3831 * @param[in] modify If true, visited pages are updated with new page numbers.
3832 * @return 0 on success, non-zero on failure.
3835 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3840 /* Make sure the txn is still viable, then find the root from
3841 * the txn's db table.
3843 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3844 DPUTS("transaction has failed, must abort");
3847 /* Make sure we're using an up-to-date root */
3848 if (mc->mc_dbi > MAIN_DBI) {
3849 if ((*mc->mc_dbflag & DB_STALE) ||
3850 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3852 unsigned char dbflag = 0;
3853 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3854 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3857 if (*mc->mc_dbflag & DB_STALE) {
3860 MDB_node *leaf = mdb_node_search(&mc2,
3861 &mc->mc_dbx->md_name, &exact);
3863 return MDB_NOTFOUND;
3864 mdb_node_read(mc->mc_txn, leaf, &data);
3865 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3867 if (flags & MDB_PS_MODIFY)
3869 *mc->mc_dbflag = dbflag;
3872 root = mc->mc_db->md_root;
3874 if (root == P_INVALID) { /* Tree is empty. */
3875 DPUTS("tree is empty");
3876 return MDB_NOTFOUND;
3881 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
3882 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3888 DPRINTF("db %u root page %zu has flags 0x%X",
3889 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3891 if (flags & MDB_PS_MODIFY) {
3892 if ((rc = mdb_page_touch(mc)))
3896 if (flags & MDB_PS_ROOTONLY)
3899 return mdb_page_search_root(mc, key, flags);
3902 /** Return the data associated with a given node.
3903 * @param[in] txn The transaction for this operation.
3904 * @param[in] leaf The node being read.
3905 * @param[out] data Updated to point to the node's data.
3906 * @return 0 on success, non-zero on failure.
3909 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3911 MDB_page *omp; /* overflow page */
3915 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3916 data->mv_size = NODEDSZ(leaf);
3917 data->mv_data = NODEDATA(leaf);
3921 /* Read overflow data.
3923 data->mv_size = NODEDSZ(leaf);
3924 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3925 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3926 DPRINTF("read overflow page %zu failed", pgno);
3929 data->mv_data = METADATA(omp);
3935 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3936 MDB_val *key, MDB_val *data)
3945 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3947 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3950 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3954 mdb_cursor_init(&mc, txn, dbi, &mx);
3955 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3958 /** Find a sibling for a page.
3959 * Replaces the page at the top of the cursor's stack with the
3960 * specified sibling, if one exists.
3961 * @param[in] mc The cursor for this operation.
3962 * @param[in] move_right Non-zero if the right sibling is requested,
3963 * otherwise the left sibling.
3964 * @return 0 on success, non-zero on failure.
3967 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3973 if (mc->mc_snum < 2) {
3974 return MDB_NOTFOUND; /* root has no siblings */
3978 DPRINTF("parent page is page %zu, index %u",
3979 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3981 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3982 : (mc->mc_ki[mc->mc_top] == 0)) {
3983 DPRINTF("no more keys left, moving to %s sibling",
3984 move_right ? "right" : "left");
3985 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
3986 /* undo cursor_pop before returning */
3993 mc->mc_ki[mc->mc_top]++;
3995 mc->mc_ki[mc->mc_top]--;
3996 DPRINTF("just moving to %s index key %u",
3997 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3999 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4001 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4002 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4005 mdb_cursor_push(mc, mp);
4010 /** Move the cursor to the next data item. */
4012 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4018 if (mc->mc_flags & C_EOF) {
4019 return MDB_NOTFOUND;
4022 assert(mc->mc_flags & C_INITIALIZED);
4024 mp = mc->mc_pg[mc->mc_top];
4026 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4027 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4028 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4029 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4030 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4031 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4035 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4036 if (op == MDB_NEXT_DUP)
4037 return MDB_NOTFOUND;
4041 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4043 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4044 DPUTS("=====> move to next sibling page");
4045 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4046 mc->mc_flags |= C_EOF;
4047 mc->mc_flags &= ~C_INITIALIZED;
4048 return MDB_NOTFOUND;
4050 mp = mc->mc_pg[mc->mc_top];
4051 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4053 mc->mc_ki[mc->mc_top]++;
4055 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4056 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4059 key->mv_size = mc->mc_db->md_pad;
4060 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4064 assert(IS_LEAF(mp));
4065 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4067 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4068 mdb_xcursor_init1(mc, leaf);
4071 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4074 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4075 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4076 if (rc != MDB_SUCCESS)
4081 MDB_GET_KEY(leaf, key);
4085 /** Move the cursor to the previous data item. */
4087 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4093 assert(mc->mc_flags & C_INITIALIZED);
4095 mp = mc->mc_pg[mc->mc_top];
4097 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4098 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4099 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4100 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4101 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4102 if (op != MDB_PREV || rc == MDB_SUCCESS)
4105 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4106 if (op == MDB_PREV_DUP)
4107 return MDB_NOTFOUND;
4112 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4114 if (mc->mc_ki[mc->mc_top] == 0) {
4115 DPUTS("=====> move to prev sibling page");
4116 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4117 mc->mc_flags &= ~C_INITIALIZED;
4118 return MDB_NOTFOUND;
4120 mp = mc->mc_pg[mc->mc_top];
4121 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4122 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4124 mc->mc_ki[mc->mc_top]--;
4126 mc->mc_flags &= ~C_EOF;
4128 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4129 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4132 key->mv_size = mc->mc_db->md_pad;
4133 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4137 assert(IS_LEAF(mp));
4138 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4140 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4141 mdb_xcursor_init1(mc, leaf);
4144 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4147 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4148 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4149 if (rc != MDB_SUCCESS)
4154 MDB_GET_KEY(leaf, key);
4158 /** Set the cursor on a specific data item. */
4160 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4161 MDB_cursor_op op, int *exactp)
4165 MDB_node *leaf = NULL;
4170 assert(key->mv_size > 0);
4172 /* See if we're already on the right page */
4173 if (mc->mc_flags & C_INITIALIZED) {
4176 mp = mc->mc_pg[mc->mc_top];
4178 mc->mc_ki[mc->mc_top] = 0;
4179 return MDB_NOTFOUND;
4181 if (mp->mp_flags & P_LEAF2) {
4182 nodekey.mv_size = mc->mc_db->md_pad;
4183 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4185 leaf = NODEPTR(mp, 0);
4186 MDB_GET_KEY(leaf, &nodekey);
4188 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4190 /* Probably happens rarely, but first node on the page
4191 * was the one we wanted.
4193 mc->mc_ki[mc->mc_top] = 0;
4200 unsigned int nkeys = NUMKEYS(mp);
4202 if (mp->mp_flags & P_LEAF2) {
4203 nodekey.mv_data = LEAF2KEY(mp,
4204 nkeys-1, nodekey.mv_size);
4206 leaf = NODEPTR(mp, nkeys-1);
4207 MDB_GET_KEY(leaf, &nodekey);
4209 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4211 /* last node was the one we wanted */
4212 mc->mc_ki[mc->mc_top] = nkeys-1;
4218 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4219 /* This is definitely the right page, skip search_page */
4220 if (mp->mp_flags & P_LEAF2) {
4221 nodekey.mv_data = LEAF2KEY(mp,
4222 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4224 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4225 MDB_GET_KEY(leaf, &nodekey);
4227 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4229 /* current node was the one we wanted */
4239 /* If any parents have right-sibs, search.
4240 * Otherwise, there's nothing further.
4242 for (i=0; i<mc->mc_top; i++)
4244 NUMKEYS(mc->mc_pg[i])-1)
4246 if (i == mc->mc_top) {
4247 /* There are no other pages */
4248 mc->mc_ki[mc->mc_top] = nkeys;
4249 return MDB_NOTFOUND;
4253 /* There are no other pages */
4254 mc->mc_ki[mc->mc_top] = 0;
4255 return MDB_NOTFOUND;
4259 rc = mdb_page_search(mc, key, 0);
4260 if (rc != MDB_SUCCESS)
4263 mp = mc->mc_pg[mc->mc_top];
4264 assert(IS_LEAF(mp));
4267 leaf = mdb_node_search(mc, key, exactp);
4268 if (exactp != NULL && !*exactp) {
4269 /* MDB_SET specified and not an exact match. */
4270 return MDB_NOTFOUND;
4274 DPUTS("===> inexact leaf not found, goto sibling");
4275 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4276 return rc; /* no entries matched */
4277 mp = mc->mc_pg[mc->mc_top];
4278 assert(IS_LEAF(mp));
4279 leaf = NODEPTR(mp, 0);
4283 mc->mc_flags |= C_INITIALIZED;
4284 mc->mc_flags &= ~C_EOF;
4287 key->mv_size = mc->mc_db->md_pad;
4288 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4292 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4293 mdb_xcursor_init1(mc, leaf);
4296 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4297 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4298 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4301 if (op == MDB_GET_BOTH) {
4307 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4308 if (rc != MDB_SUCCESS)
4311 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4313 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4315 rc = mc->mc_dbx->md_dcmp(data, &d2);
4317 if (op == MDB_GET_BOTH || rc > 0)
4318 return MDB_NOTFOUND;
4323 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4324 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4329 /* The key already matches in all other cases */
4330 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4331 MDB_GET_KEY(leaf, key);
4332 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4337 /** Move the cursor to the first item in the database. */
4339 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4344 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4345 rc = mdb_page_search(mc, NULL, 0);
4346 if (rc != MDB_SUCCESS)
4349 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4351 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4352 mc->mc_flags |= C_INITIALIZED;
4353 mc->mc_flags &= ~C_EOF;
4355 mc->mc_ki[mc->mc_top] = 0;
4357 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4358 key->mv_size = mc->mc_db->md_pad;
4359 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4364 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4365 mdb_xcursor_init1(mc, leaf);
4366 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4371 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4372 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4376 MDB_GET_KEY(leaf, key);
4380 /** Move the cursor to the last item in the database. */
4382 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4387 if (!(mc->mc_flags & C_EOF)) {
4389 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4392 lkey.mv_size = MAXKEYSIZE+1;
4393 lkey.mv_data = NULL;
4394 rc = mdb_page_search(mc, &lkey, 0);
4395 if (rc != MDB_SUCCESS)
4398 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4400 mc->mc_flags |= C_INITIALIZED|C_EOF;
4401 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4403 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4405 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4406 key->mv_size = mc->mc_db->md_pad;
4407 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4412 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4413 mdb_xcursor_init1(mc, leaf);
4414 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4419 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4420 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4425 MDB_GET_KEY(leaf, key);
4430 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4439 case MDB_GET_CURRENT:
4440 if (!mc->mc_flags & C_INITIALIZED) {
4443 MDB_page *mp = mc->mc_pg[mc->mc_top];
4445 mc->mc_ki[mc->mc_top] = 0;
4451 key->mv_size = mc->mc_db->md_pad;
4452 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4454 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4455 MDB_GET_KEY(leaf, key);
4457 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4458 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4460 rc = mdb_node_read(mc->mc_txn, leaf, data);
4467 case MDB_GET_BOTH_RANGE:
4468 if (data == NULL || mc->mc_xcursor == NULL) {
4476 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4478 } else if (op == MDB_SET_RANGE)
4479 rc = mdb_cursor_set(mc, key, data, op, NULL);
4481 rc = mdb_cursor_set(mc, key, data, op, &exact);
4483 case MDB_GET_MULTIPLE:
4485 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4486 !(mc->mc_flags & C_INITIALIZED)) {
4491 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4492 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4495 case MDB_NEXT_MULTIPLE:
4497 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4501 if (!(mc->mc_flags & C_INITIALIZED))
4502 rc = mdb_cursor_first(mc, key, data);
4504 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4505 if (rc == MDB_SUCCESS) {
4506 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4509 mx = &mc->mc_xcursor->mx_cursor;
4510 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4512 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4513 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4521 case MDB_NEXT_NODUP:
4522 if (!(mc->mc_flags & C_INITIALIZED))
4523 rc = mdb_cursor_first(mc, key, data);
4525 rc = mdb_cursor_next(mc, key, data, op);
4529 case MDB_PREV_NODUP:
4530 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4531 rc = mdb_cursor_last(mc, key, data);
4532 mc->mc_flags |= C_INITIALIZED;
4533 mc->mc_ki[mc->mc_top]++;
4535 rc = mdb_cursor_prev(mc, key, data, op);
4538 rc = mdb_cursor_first(mc, key, data);
4542 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4543 !(mc->mc_flags & C_INITIALIZED) ||
4544 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4548 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4551 rc = mdb_cursor_last(mc, key, data);
4555 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4556 !(mc->mc_flags & C_INITIALIZED) ||
4557 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4561 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4564 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4572 /** Touch all the pages in the cursor stack.
4573 * Makes sure all the pages are writable, before attempting a write operation.
4574 * @param[in] mc The cursor to operate on.
4577 mdb_cursor_touch(MDB_cursor *mc)
4581 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4584 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4585 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4586 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4589 *mc->mc_dbflag = DB_DIRTY;
4591 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4592 rc = mdb_page_touch(mc);
4596 mc->mc_top = mc->mc_snum-1;
4601 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4604 MDB_node *leaf = NULL;
4605 MDB_val xdata, *rdata, dkey;
4608 int do_sub = 0, insert = 0;
4609 unsigned int mcount = 0;
4613 char dbuf[MAXKEYSIZE+1];
4614 unsigned int nflags;
4617 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4620 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4621 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4625 if (flags == MDB_CURRENT) {
4626 if (!(mc->mc_flags & C_INITIALIZED))
4629 } else if (mc->mc_db->md_root == P_INVALID) {
4631 /* new database, write a root leaf page */
4632 DPUTS("allocating new root leaf page");
4633 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4637 mdb_cursor_push(mc, np);
4638 mc->mc_db->md_root = np->mp_pgno;
4639 mc->mc_db->md_depth++;
4640 *mc->mc_dbflag = DB_DIRTY;
4641 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4643 np->mp_flags |= P_LEAF2;
4644 mc->mc_flags |= C_INITIALIZED;
4650 if (flags & MDB_APPEND) {
4652 rc = mdb_cursor_last(mc, &k2, &d2);
4654 rc = mc->mc_dbx->md_cmp(key, &k2);
4657 mc->mc_ki[mc->mc_top]++;
4663 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4665 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4666 DPRINTF("duplicate key [%s]", DKEY(key));
4668 return MDB_KEYEXIST;
4670 if (rc && rc != MDB_NOTFOUND)
4674 /* Cursor is positioned, now make sure all pages are writable */
4675 rc2 = mdb_cursor_touch(mc);
4680 /* The key already exists */
4681 if (rc == MDB_SUCCESS) {
4682 /* there's only a key anyway, so this is a no-op */
4683 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4684 unsigned int ksize = mc->mc_db->md_pad;
4685 if (key->mv_size != ksize)
4687 if (flags == MDB_CURRENT) {
4688 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4689 memcpy(ptr, key->mv_data, ksize);
4694 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4697 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4698 /* Was a single item before, must convert now */
4700 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4701 /* Just overwrite the current item */
4702 if (flags == MDB_CURRENT)
4705 dkey.mv_size = NODEDSZ(leaf);
4706 dkey.mv_data = NODEDATA(leaf);
4707 #if UINT_MAX < SIZE_MAX
4708 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4709 #ifdef MISALIGNED_OK
4710 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4712 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4715 /* if data matches, ignore it */
4716 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4717 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4719 /* create a fake page for the dup items */
4720 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4721 dkey.mv_data = dbuf;
4722 fp = (MDB_page *)&pbuf;
4723 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4724 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4725 fp->mp_lower = PAGEHDRSZ;
4726 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4727 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4728 fp->mp_flags |= P_LEAF2;
4729 fp->mp_pad = data->mv_size;
4730 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4732 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4733 (dkey.mv_size & 1) + (data->mv_size & 1);
4735 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4738 xdata.mv_size = fp->mp_upper;
4743 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4744 /* See if we need to convert from fake page to subDB */
4746 unsigned int offset;
4749 fp = NODEDATA(leaf);
4750 if (flags == MDB_CURRENT) {
4752 fp->mp_flags |= P_DIRTY;
4753 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4754 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4758 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4759 offset = fp->mp_pad;
4760 if (SIZELEFT(fp) >= offset)
4762 offset *= 4; /* space for 4 more */
4764 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4766 offset += offset & 1;
4767 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4768 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4770 /* yes, convert it */
4772 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4773 dummy.md_pad = fp->mp_pad;
4774 dummy.md_flags = MDB_DUPFIXED;
4775 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4776 dummy.md_flags |= MDB_INTEGERKEY;
4779 dummy.md_branch_pages = 0;
4780 dummy.md_leaf_pages = 1;
4781 dummy.md_overflow_pages = 0;
4782 dummy.md_entries = NUMKEYS(fp);
4784 xdata.mv_size = sizeof(MDB_db);
4785 xdata.mv_data = &dummy;
4786 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4788 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4789 flags |= F_DUPDATA|F_SUBDATA;
4790 dummy.md_root = mp->mp_pgno;
4792 /* no, just grow it */
4794 xdata.mv_size = NODEDSZ(leaf) + offset;
4795 xdata.mv_data = &pbuf;
4796 mp = (MDB_page *)&pbuf;
4797 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4800 mp->mp_flags = fp->mp_flags | P_DIRTY;
4801 mp->mp_pad = fp->mp_pad;
4802 mp->mp_lower = fp->mp_lower;
4803 mp->mp_upper = fp->mp_upper + offset;
4805 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4807 nsize = NODEDSZ(leaf) - fp->mp_upper;
4808 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4809 for (i=0; i<NUMKEYS(fp); i++)
4810 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4812 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4816 /* data is on sub-DB, just store it */
4817 flags |= F_DUPDATA|F_SUBDATA;
4821 /* overflow page overwrites need special handling */
4822 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4825 int ovpages, dpages;
4827 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4828 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4829 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4830 mdb_page_get(mc->mc_txn, pg, &omp);
4831 /* Is the ov page writable and large enough? */
4832 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4833 /* yes, overwrite it. Note in this case we don't
4834 * bother to try shrinking the node if the new data
4835 * is smaller than the overflow threshold.
4837 if (F_ISSET(flags, MDB_RESERVE))
4838 data->mv_data = METADATA(omp);
4840 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4843 /* no, free ovpages */
4845 mc->mc_db->md_overflow_pages -= ovpages;
4846 for (i=0; i<ovpages; i++) {
4847 DPRINTF("freed ov page %zu", pg);
4848 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4852 } else if (NODEDSZ(leaf) == data->mv_size) {
4853 /* same size, just replace it. Note that we could
4854 * also reuse this node if the new data is smaller,
4855 * but instead we opt to shrink the node in that case.
4857 if (F_ISSET(flags, MDB_RESERVE))
4858 data->mv_data = NODEDATA(leaf);
4860 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4863 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4864 mc->mc_db->md_entries--;
4866 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4873 nflags = flags & NODE_ADD_FLAGS;
4874 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4875 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4876 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4877 nflags &= ~MDB_APPEND;
4879 nflags |= MDB_SPLIT_REPLACE;
4880 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4882 /* There is room already in this leaf page. */
4883 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4884 if (rc == 0 && !do_sub && insert) {
4885 /* Adjust other cursors pointing to mp */
4886 MDB_cursor *m2, *m3;
4887 MDB_dbi dbi = mc->mc_dbi;
4888 unsigned i = mc->mc_top;
4889 MDB_page *mp = mc->mc_pg[i];
4891 if (mc->mc_flags & C_SUB)
4894 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4895 if (mc->mc_flags & C_SUB)
4896 m3 = &m2->mc_xcursor->mx_cursor;
4899 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4900 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4907 if (rc != MDB_SUCCESS)
4908 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4910 /* Now store the actual data in the child DB. Note that we're
4911 * storing the user data in the keys field, so there are strict
4912 * size limits on dupdata. The actual data fields of the child
4913 * DB are all zero size.
4920 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4921 if (flags & MDB_CURRENT) {
4922 xflags = MDB_CURRENT;
4924 mdb_xcursor_init1(mc, leaf);
4925 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4927 /* converted, write the original data first */
4929 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4933 /* Adjust other cursors pointing to mp */
4935 unsigned i = mc->mc_top;
4936 MDB_page *mp = mc->mc_pg[i];
4938 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4939 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4940 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4941 mdb_xcursor_init1(m2, leaf);
4946 if (flags & MDB_APPENDDUP)
4947 xflags |= MDB_APPEND;
4948 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4949 if (flags & F_SUBDATA) {
4950 void *db = NODEDATA(leaf);
4951 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4954 /* sub-writes might have failed so check rc again.
4955 * Don't increment count if we just replaced an existing item.
4957 if (!rc && !(flags & MDB_CURRENT))
4958 mc->mc_db->md_entries++;
4959 if (flags & MDB_MULTIPLE) {
4961 if (mcount < data[1].mv_size) {
4962 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4963 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4973 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4978 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4981 if (!mc->mc_flags & C_INITIALIZED)
4984 rc = mdb_cursor_touch(mc);
4988 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4990 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4991 if (flags != MDB_NODUPDATA) {
4992 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4993 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4995 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4996 /* If sub-DB still has entries, we're done */
4997 if (mc->mc_xcursor->mx_db.md_entries) {
4998 if (leaf->mn_flags & F_SUBDATA) {
4999 /* update subDB info */
5000 void *db = NODEDATA(leaf);
5001 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5003 /* shrink fake page */
5004 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5006 mc->mc_db->md_entries--;
5009 /* otherwise fall thru and delete the sub-DB */
5012 if (leaf->mn_flags & F_SUBDATA) {
5013 /* add all the child DB's pages to the free list */
5014 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5015 if (rc == MDB_SUCCESS) {
5016 mc->mc_db->md_entries -=
5017 mc->mc_xcursor->mx_db.md_entries;
5022 return mdb_cursor_del0(mc, leaf);
5025 /** Allocate and initialize new pages for a database.
5026 * @param[in] mc a cursor on the database being added to.
5027 * @param[in] flags flags defining what type of page is being allocated.
5028 * @param[in] num the number of pages to allocate. This is usually 1,
5029 * unless allocating overflow pages for a large record.
5030 * @param[out] mp Address of a page, or NULL on failure.
5031 * @return 0 on success, non-zero on failure.
5034 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5039 if ((rc = mdb_page_alloc(mc, num, &np)))
5041 DPRINTF("allocated new mpage %zu, page size %u",
5042 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5043 np->mp_flags = flags | P_DIRTY;
5044 np->mp_lower = PAGEHDRSZ;
5045 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5048 mc->mc_db->md_branch_pages++;
5049 else if (IS_LEAF(np))
5050 mc->mc_db->md_leaf_pages++;
5051 else if (IS_OVERFLOW(np)) {
5052 mc->mc_db->md_overflow_pages += num;
5060 /** Calculate the size of a leaf node.
5061 * The size depends on the environment's page size; if a data item
5062 * is too large it will be put onto an overflow page and the node
5063 * size will only include the key and not the data. Sizes are always
5064 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5065 * of the #MDB_node headers.
5066 * @param[in] env The environment handle.
5067 * @param[in] key The key for the node.
5068 * @param[in] data The data for the node.
5069 * @return The number of bytes needed to store the node.
5072 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5076 sz = LEAFSIZE(key, data);
5077 if (sz >= env->me_psize / MDB_MINKEYS) {
5078 /* put on overflow page */
5079 sz -= data->mv_size - sizeof(pgno_t);
5083 return sz + sizeof(indx_t);
5086 /** Calculate the size of a branch node.
5087 * The size should depend on the environment's page size but since
5088 * we currently don't support spilling large keys onto overflow
5089 * pages, it's simply the size of the #MDB_node header plus the
5090 * size of the key. Sizes are always rounded up to an even number
5091 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5092 * @param[in] env The environment handle.
5093 * @param[in] key The key for the node.
5094 * @return The number of bytes needed to store the node.
5097 mdb_branch_size(MDB_env *env, MDB_val *key)
5102 if (sz >= env->me_psize / MDB_MINKEYS) {
5103 /* put on overflow page */
5104 /* not implemented */
5105 /* sz -= key->size - sizeof(pgno_t); */
5108 return sz + sizeof(indx_t);
5111 /** Add a node to the page pointed to by the cursor.
5112 * @param[in] mc The cursor for this operation.
5113 * @param[in] indx The index on the page where the new node should be added.
5114 * @param[in] key The key for the new node.
5115 * @param[in] data The data for the new node, if any.
5116 * @param[in] pgno The page number, if adding a branch node.
5117 * @param[in] flags Flags for the node.
5118 * @return 0 on success, non-zero on failure. Possible errors are:
5120 * <li>ENOMEM - failed to allocate overflow pages for the node.
5121 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5122 * should never happen since all callers already calculate the
5123 * page's free space before calling this function.
5127 mdb_node_add(MDB_cursor *mc, indx_t indx,
5128 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5131 size_t node_size = NODESIZE;
5134 MDB_page *mp = mc->mc_pg[mc->mc_top];
5135 MDB_page *ofp = NULL; /* overflow page */
5138 assert(mp->mp_upper >= mp->mp_lower);
5140 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5141 IS_LEAF(mp) ? "leaf" : "branch",
5142 IS_SUBP(mp) ? "sub-" : "",
5143 mp->mp_pgno, indx, data ? data->mv_size : 0,
5144 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5147 /* Move higher keys up one slot. */
5148 int ksize = mc->mc_db->md_pad, dif;
5149 char *ptr = LEAF2KEY(mp, indx, ksize);
5150 dif = NUMKEYS(mp) - indx;
5152 memmove(ptr+ksize, ptr, dif*ksize);
5153 /* insert new key */
5154 memcpy(ptr, key->mv_data, ksize);
5156 /* Just using these for counting */
5157 mp->mp_lower += sizeof(indx_t);
5158 mp->mp_upper -= ksize - sizeof(indx_t);
5163 node_size += key->mv_size;
5167 if (F_ISSET(flags, F_BIGDATA)) {
5168 /* Data already on overflow page. */
5169 node_size += sizeof(pgno_t);
5170 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5171 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5173 /* Put data on overflow page. */
5174 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5175 data->mv_size, node_size+data->mv_size);
5176 node_size += sizeof(pgno_t);
5177 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5179 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5182 node_size += data->mv_size;
5185 node_size += node_size & 1;
5187 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5188 DPRINTF("not enough room in page %zu, got %u ptrs",
5189 mp->mp_pgno, NUMKEYS(mp));
5190 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5191 mp->mp_upper - mp->mp_lower);
5192 DPRINTF("node size = %zu", node_size);
5193 return MDB_PAGE_FULL;
5196 /* Move higher pointers up one slot. */
5197 for (i = NUMKEYS(mp); i > indx; i--)
5198 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5200 /* Adjust free space offsets. */
5201 ofs = mp->mp_upper - node_size;
5202 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5203 mp->mp_ptrs[indx] = ofs;
5205 mp->mp_lower += sizeof(indx_t);
5207 /* Write the node data. */
5208 node = NODEPTR(mp, indx);
5209 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5210 node->mn_flags = flags;
5212 SETDSZ(node,data->mv_size);
5217 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5222 if (F_ISSET(flags, F_BIGDATA))
5223 memcpy(node->mn_data + key->mv_size, data->mv_data,
5225 else if (F_ISSET(flags, MDB_RESERVE))
5226 data->mv_data = node->mn_data + key->mv_size;
5228 memcpy(node->mn_data + key->mv_size, data->mv_data,
5231 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5233 if (F_ISSET(flags, MDB_RESERVE))
5234 data->mv_data = METADATA(ofp);
5236 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5243 /** Delete the specified node from a page.
5244 * @param[in] mp The page to operate on.
5245 * @param[in] indx The index of the node to delete.
5246 * @param[in] ksize The size of a node. Only used if the page is
5247 * part of a #MDB_DUPFIXED database.
5250 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5253 indx_t i, j, numkeys, ptr;
5260 COPY_PGNO(pgno, mp->mp_pgno);
5261 DPRINTF("delete node %u on %s page %zu", indx,
5262 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5265 assert(indx < NUMKEYS(mp));
5268 int x = NUMKEYS(mp) - 1 - indx;
5269 base = LEAF2KEY(mp, indx, ksize);
5271 memmove(base, base + ksize, x * ksize);
5272 mp->mp_lower -= sizeof(indx_t);
5273 mp->mp_upper += ksize - sizeof(indx_t);
5277 node = NODEPTR(mp, indx);
5278 sz = NODESIZE + node->mn_ksize;
5280 if (F_ISSET(node->mn_flags, F_BIGDATA))
5281 sz += sizeof(pgno_t);
5283 sz += NODEDSZ(node);
5287 ptr = mp->mp_ptrs[indx];
5288 numkeys = NUMKEYS(mp);
5289 for (i = j = 0; i < numkeys; i++) {
5291 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5292 if (mp->mp_ptrs[i] < ptr)
5293 mp->mp_ptrs[j] += sz;
5298 base = (char *)mp + mp->mp_upper;
5299 memmove(base + sz, base, ptr - mp->mp_upper);
5301 mp->mp_lower -= sizeof(indx_t);
5305 /** Compact the main page after deleting a node on a subpage.
5306 * @param[in] mp The main page to operate on.
5307 * @param[in] indx The index of the subpage on the main page.
5310 mdb_node_shrink(MDB_page *mp, indx_t indx)
5317 indx_t i, numkeys, ptr;
5319 node = NODEPTR(mp, indx);
5320 sp = (MDB_page *)NODEDATA(node);
5321 osize = NODEDSZ(node);
5323 delta = sp->mp_upper - sp->mp_lower;
5324 SETDSZ(node, osize - delta);
5325 xp = (MDB_page *)((char *)sp + delta);
5327 /* shift subpage upward */
5329 nsize = NUMKEYS(sp) * sp->mp_pad;
5330 memmove(METADATA(xp), METADATA(sp), nsize);
5333 nsize = osize - sp->mp_upper;
5334 numkeys = NUMKEYS(sp);
5335 for (i=numkeys-1; i>=0; i--)
5336 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5338 xp->mp_upper = sp->mp_lower;
5339 xp->mp_lower = sp->mp_lower;
5340 xp->mp_flags = sp->mp_flags;
5341 xp->mp_pad = sp->mp_pad;
5342 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5344 /* shift lower nodes upward */
5345 ptr = mp->mp_ptrs[indx];
5346 numkeys = NUMKEYS(mp);
5347 for (i = 0; i < numkeys; i++) {
5348 if (mp->mp_ptrs[i] <= ptr)
5349 mp->mp_ptrs[i] += delta;
5352 base = (char *)mp + mp->mp_upper;
5353 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5354 mp->mp_upper += delta;
5357 /** Initial setup of a sorted-dups cursor.
5358 * Sorted duplicates are implemented as a sub-database for the given key.
5359 * The duplicate data items are actually keys of the sub-database.
5360 * Operations on the duplicate data items are performed using a sub-cursor
5361 * initialized when the sub-database is first accessed. This function does
5362 * the preliminary setup of the sub-cursor, filling in the fields that
5363 * depend only on the parent DB.
5364 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5367 mdb_xcursor_init0(MDB_cursor *mc)
5369 MDB_xcursor *mx = mc->mc_xcursor;
5371 mx->mx_cursor.mc_xcursor = NULL;
5372 mx->mx_cursor.mc_txn = mc->mc_txn;
5373 mx->mx_cursor.mc_db = &mx->mx_db;
5374 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5375 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5376 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5377 mx->mx_cursor.mc_snum = 0;
5378 mx->mx_cursor.mc_top = 0;
5379 mx->mx_cursor.mc_flags = C_SUB;
5380 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5381 mx->mx_dbx.md_dcmp = NULL;
5382 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5385 /** Final setup of a sorted-dups cursor.
5386 * Sets up the fields that depend on the data from the main cursor.
5387 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5388 * @param[in] node The data containing the #MDB_db record for the
5389 * sorted-dup database.
5392 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5394 MDB_xcursor *mx = mc->mc_xcursor;
5396 if (node->mn_flags & F_SUBDATA) {
5397 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5398 mx->mx_cursor.mc_pg[0] = 0;
5399 mx->mx_cursor.mc_snum = 0;
5400 mx->mx_cursor.mc_flags = C_SUB;
5402 MDB_page *fp = NODEDATA(node);
5403 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5404 mx->mx_db.md_flags = 0;
5405 mx->mx_db.md_depth = 1;
5406 mx->mx_db.md_branch_pages = 0;
5407 mx->mx_db.md_leaf_pages = 1;
5408 mx->mx_db.md_overflow_pages = 0;
5409 mx->mx_db.md_entries = NUMKEYS(fp);
5410 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5411 mx->mx_cursor.mc_snum = 1;
5412 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5413 mx->mx_cursor.mc_top = 0;
5414 mx->mx_cursor.mc_pg[0] = fp;
5415 mx->mx_cursor.mc_ki[0] = 0;
5416 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5417 mx->mx_db.md_flags = MDB_DUPFIXED;
5418 mx->mx_db.md_pad = fp->mp_pad;
5419 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5420 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5423 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5425 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5427 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5428 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5429 #if UINT_MAX < SIZE_MAX
5430 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5431 #ifdef MISALIGNED_OK
5432 mx->mx_dbx.md_cmp = mdb_cmp_long;
5434 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5439 /** Initialize a cursor for a given transaction and database. */
5441 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5446 mc->mc_db = &txn->mt_dbs[dbi];
5447 mc->mc_dbx = &txn->mt_dbxs[dbi];
5448 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5453 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5455 mc->mc_xcursor = mx;
5456 mdb_xcursor_init0(mc);
5458 mc->mc_xcursor = NULL;
5460 if (*mc->mc_dbflag & DB_STALE) {
5461 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5466 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5469 MDB_xcursor *mx = NULL;
5470 size_t size = sizeof(MDB_cursor);
5472 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5475 /* Allow read access to the freelist */
5476 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5479 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5480 size += sizeof(MDB_xcursor);
5482 if ((mc = malloc(size)) != NULL) {
5483 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5484 mx = (MDB_xcursor *)(mc + 1);
5486 mdb_cursor_init(mc, txn, dbi, mx);
5487 if (txn->mt_cursors) {
5488 mc->mc_next = txn->mt_cursors[dbi];
5489 txn->mt_cursors[dbi] = mc;
5491 mc->mc_flags |= C_ALLOCD;
5502 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5504 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5507 if (txn->mt_cursors)
5510 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5514 /* Return the count of duplicate data items for the current key */
5516 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5520 if (mc == NULL || countp == NULL)
5523 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5526 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5527 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5530 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5533 *countp = mc->mc_xcursor->mx_db.md_entries;
5539 mdb_cursor_close(MDB_cursor *mc)
5542 /* remove from txn, if tracked */
5543 if (mc->mc_txn->mt_cursors) {
5544 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5545 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5547 *prev = mc->mc_next;
5549 if (mc->mc_flags & C_ALLOCD)
5555 mdb_cursor_txn(MDB_cursor *mc)
5557 if (!mc) return NULL;
5562 mdb_cursor_dbi(MDB_cursor *mc)
5568 /** Replace the key for a node with a new key.
5569 * @param[in] mp The page containing the node to operate on.
5570 * @param[in] indx The index of the node to operate on.
5571 * @param[in] key The new key to use.
5572 * @return 0 on success, non-zero on failure.
5575 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5581 indx_t ptr, i, numkeys;
5584 node = NODEPTR(mp, indx);
5585 ptr = mp->mp_ptrs[indx];
5589 char kbuf2[(MAXKEYSIZE*2+1)];
5590 k2.mv_data = NODEKEY(node);
5591 k2.mv_size = node->mn_ksize;
5592 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5594 mdb_dkey(&k2, kbuf2),
5600 delta0 = delta = key->mv_size - node->mn_ksize;
5602 /* Must be 2-byte aligned. If new key is
5603 * shorter by 1, the shift will be skipped.
5605 delta += (delta & 1);
5607 if (delta > 0 && SIZELEFT(mp) < delta) {
5608 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5609 return MDB_PAGE_FULL;
5612 numkeys = NUMKEYS(mp);
5613 for (i = 0; i < numkeys; i++) {
5614 if (mp->mp_ptrs[i] <= ptr)
5615 mp->mp_ptrs[i] -= delta;
5618 base = (char *)mp + mp->mp_upper;
5619 len = ptr - mp->mp_upper + NODESIZE;
5620 memmove(base - delta, base, len);
5621 mp->mp_upper -= delta;
5623 node = NODEPTR(mp, indx);
5626 /* But even if no shift was needed, update ksize */
5628 node->mn_ksize = key->mv_size;
5631 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5636 /** Move a node from csrc to cdst.
5639 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5645 unsigned short flags;
5649 /* Mark src and dst as dirty. */
5650 if ((rc = mdb_page_touch(csrc)) ||
5651 (rc = mdb_page_touch(cdst)))
5654 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5655 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5656 key.mv_size = csrc->mc_db->md_pad;
5657 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5659 data.mv_data = NULL;
5663 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5664 assert(!((long)srcnode&1));
5665 srcpg = NODEPGNO(srcnode);
5666 flags = srcnode->mn_flags;
5667 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5668 unsigned int snum = csrc->mc_snum;
5670 /* must find the lowest key below src */
5671 mdb_page_search_root(csrc, NULL, 0);
5672 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5673 key.mv_size = csrc->mc_db->md_pad;
5674 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5676 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5677 key.mv_size = NODEKSZ(s2);
5678 key.mv_data = NODEKEY(s2);
5680 csrc->mc_snum = snum--;
5681 csrc->mc_top = snum;
5683 key.mv_size = NODEKSZ(srcnode);
5684 key.mv_data = NODEKEY(srcnode);
5686 data.mv_size = NODEDSZ(srcnode);
5687 data.mv_data = NODEDATA(srcnode);
5689 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5690 unsigned int snum = cdst->mc_snum;
5693 /* must find the lowest key below dst */
5694 mdb_page_search_root(cdst, NULL, 0);
5695 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5696 bkey.mv_size = cdst->mc_db->md_pad;
5697 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5699 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5700 bkey.mv_size = NODEKSZ(s2);
5701 bkey.mv_data = NODEKEY(s2);
5703 cdst->mc_snum = snum--;
5704 cdst->mc_top = snum;
5705 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5708 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5709 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5710 csrc->mc_ki[csrc->mc_top],
5712 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5713 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5715 /* Add the node to the destination page.
5717 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5718 if (rc != MDB_SUCCESS)
5721 /* Delete the node from the source page.
5723 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5726 /* Adjust other cursors pointing to mp */
5727 MDB_cursor *m2, *m3;
5728 MDB_dbi dbi = csrc->mc_dbi;
5729 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5731 if (csrc->mc_flags & C_SUB)
5734 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5735 if (m2 == csrc) continue;
5736 if (csrc->mc_flags & C_SUB)
5737 m3 = &m2->mc_xcursor->mx_cursor;
5740 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5741 csrc->mc_ki[csrc->mc_top]) {
5742 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5743 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5748 /* Update the parent separators.
5750 if (csrc->mc_ki[csrc->mc_top] == 0) {
5751 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5752 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5753 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5755 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5756 key.mv_size = NODEKSZ(srcnode);
5757 key.mv_data = NODEKEY(srcnode);
5759 DPRINTF("update separator for source page %zu to [%s]",
5760 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5761 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5762 &key)) != MDB_SUCCESS)
5765 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5767 nullkey.mv_size = 0;
5768 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5769 assert(rc == MDB_SUCCESS);
5773 if (cdst->mc_ki[cdst->mc_top] == 0) {
5774 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5775 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5776 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5778 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5779 key.mv_size = NODEKSZ(srcnode);
5780 key.mv_data = NODEKEY(srcnode);
5782 DPRINTF("update separator for destination page %zu to [%s]",
5783 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5784 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5785 &key)) != MDB_SUCCESS)
5788 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5790 nullkey.mv_size = 0;
5791 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5792 assert(rc == MDB_SUCCESS);
5799 /** Merge one page into another.
5800 * The nodes from the page pointed to by \b csrc will
5801 * be copied to the page pointed to by \b cdst and then
5802 * the \b csrc page will be freed.
5803 * @param[in] csrc Cursor pointing to the source page.
5804 * @param[in] cdst Cursor pointing to the destination page.
5807 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5815 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5816 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5818 assert(csrc->mc_snum > 1); /* can't merge root page */
5819 assert(cdst->mc_snum > 1);
5821 /* Mark dst as dirty. */
5822 if ((rc = mdb_page_touch(cdst)))
5825 /* Move all nodes from src to dst.
5827 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5828 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5829 key.mv_size = csrc->mc_db->md_pad;
5830 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5831 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5832 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5833 if (rc != MDB_SUCCESS)
5835 key.mv_data = (char *)key.mv_data + key.mv_size;
5838 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5839 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5840 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5841 unsigned int snum = csrc->mc_snum;
5843 /* must find the lowest key below src */
5844 mdb_page_search_root(csrc, NULL, 0);
5845 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5846 key.mv_size = csrc->mc_db->md_pad;
5847 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5849 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5850 key.mv_size = NODEKSZ(s2);
5851 key.mv_data = NODEKEY(s2);
5853 csrc->mc_snum = snum--;
5854 csrc->mc_top = snum;
5856 key.mv_size = srcnode->mn_ksize;
5857 key.mv_data = NODEKEY(srcnode);
5860 data.mv_size = NODEDSZ(srcnode);
5861 data.mv_data = NODEDATA(srcnode);
5862 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5863 if (rc != MDB_SUCCESS)
5868 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5869 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);
5871 /* Unlink the src page from parent and add to free list.
5873 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5874 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5876 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5880 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5881 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5882 csrc->mc_db->md_leaf_pages--;
5884 csrc->mc_db->md_branch_pages--;
5886 /* Adjust other cursors pointing to mp */
5887 MDB_cursor *m2, *m3;
5888 MDB_dbi dbi = csrc->mc_dbi;
5889 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5891 if (csrc->mc_flags & C_SUB)
5894 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5895 if (csrc->mc_flags & C_SUB)
5896 m3 = &m2->mc_xcursor->mx_cursor;
5899 if (m3 == csrc) continue;
5900 if (m3->mc_snum < csrc->mc_snum) continue;
5901 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5902 m3->mc_pg[csrc->mc_top] = mp;
5903 m3->mc_ki[csrc->mc_top] += nkeys;
5907 mdb_cursor_pop(csrc);
5909 return mdb_rebalance(csrc);
5912 /** Copy the contents of a cursor.
5913 * @param[in] csrc The cursor to copy from.
5914 * @param[out] cdst The cursor to copy to.
5917 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5921 cdst->mc_txn = csrc->mc_txn;
5922 cdst->mc_dbi = csrc->mc_dbi;
5923 cdst->mc_db = csrc->mc_db;
5924 cdst->mc_dbx = csrc->mc_dbx;
5925 cdst->mc_snum = csrc->mc_snum;
5926 cdst->mc_top = csrc->mc_top;
5927 cdst->mc_flags = csrc->mc_flags;
5929 for (i=0; i<csrc->mc_snum; i++) {
5930 cdst->mc_pg[i] = csrc->mc_pg[i];
5931 cdst->mc_ki[i] = csrc->mc_ki[i];
5935 /** Rebalance the tree after a delete operation.
5936 * @param[in] mc Cursor pointing to the page where rebalancing
5938 * @return 0 on success, non-zero on failure.
5941 mdb_rebalance(MDB_cursor *mc)
5951 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5952 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5953 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5954 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5958 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5961 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5962 DPRINTF("no need to rebalance page %zu, above fill threshold",
5968 if (mc->mc_snum < 2) {
5969 MDB_page *mp = mc->mc_pg[0];
5970 if (NUMKEYS(mp) == 0) {
5971 DPUTS("tree is completely empty");
5972 mc->mc_db->md_root = P_INVALID;
5973 mc->mc_db->md_depth = 0;
5974 mc->mc_db->md_leaf_pages = 0;
5975 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5979 /* Adjust other cursors pointing to mp */
5980 MDB_cursor *m2, *m3;
5981 MDB_dbi dbi = mc->mc_dbi;
5983 if (mc->mc_flags & C_SUB)
5986 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5987 if (m2 == mc) continue;
5988 if (mc->mc_flags & C_SUB)
5989 m3 = &m2->mc_xcursor->mx_cursor;
5992 if (m3->mc_snum < mc->mc_snum) continue;
5993 if (m3->mc_pg[0] == mp) {
5999 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6000 DPUTS("collapsing root page!");
6001 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6002 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6003 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6006 mc->mc_db->md_depth--;
6007 mc->mc_db->md_branch_pages--;
6009 /* Adjust other cursors pointing to mp */
6010 MDB_cursor *m2, *m3;
6011 MDB_dbi dbi = mc->mc_dbi;
6013 if (mc->mc_flags & C_SUB)
6016 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6017 if (m2 == mc) continue;
6018 if (mc->mc_flags & C_SUB)
6019 m3 = &m2->mc_xcursor->mx_cursor;
6022 if (m3->mc_snum < mc->mc_snum) continue;
6023 if (m3->mc_pg[0] == mp) {
6024 m3->mc_pg[0] = mc->mc_pg[0];
6029 DPUTS("root page doesn't need rebalancing");
6033 /* The parent (branch page) must have at least 2 pointers,
6034 * otherwise the tree is invalid.
6036 ptop = mc->mc_top-1;
6037 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6039 /* Leaf page fill factor is below the threshold.
6040 * Try to move keys from left or right neighbor, or
6041 * merge with a neighbor page.
6046 mdb_cursor_copy(mc, &mn);
6047 mn.mc_xcursor = NULL;
6049 if (mc->mc_ki[ptop] == 0) {
6050 /* We're the leftmost leaf in our parent.
6052 DPUTS("reading right neighbor");
6054 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6055 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6057 mn.mc_ki[mn.mc_top] = 0;
6058 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6060 /* There is at least one neighbor to the left.
6062 DPUTS("reading left neighbor");
6064 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6065 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6067 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6068 mc->mc_ki[mc->mc_top] = 0;
6071 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6072 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);
6074 /* If the neighbor page is above threshold and has at least two
6075 * keys, move one key from it.
6077 * Otherwise we should try to merge them.
6079 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6080 return mdb_node_move(&mn, mc);
6081 else { /* FIXME: if (has_enough_room()) */
6082 mc->mc_flags &= ~C_INITIALIZED;
6083 if (mc->mc_ki[ptop] == 0)
6084 return mdb_page_merge(&mn, mc);
6086 return mdb_page_merge(mc, &mn);
6090 /** Complete a delete operation started by #mdb_cursor_del(). */
6092 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6096 /* add overflow pages to free list */
6097 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6101 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6102 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6103 mc->mc_db->md_overflow_pages -= ovpages;
6104 for (i=0; i<ovpages; i++) {
6105 DPRINTF("freed ov page %zu", pg);
6106 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6110 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6111 mc->mc_db->md_entries--;
6112 rc = mdb_rebalance(mc);
6113 if (rc != MDB_SUCCESS)
6114 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6120 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6121 MDB_val *key, MDB_val *data)
6126 MDB_val rdata, *xdata;
6130 assert(key != NULL);
6132 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6134 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6137 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6141 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6145 mdb_cursor_init(&mc, txn, dbi, &mx);
6156 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6158 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6162 /** Split a page and insert a new node.
6163 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6164 * The cursor will be updated to point to the actual page and index where
6165 * the node got inserted after the split.
6166 * @param[in] newkey The key for the newly inserted node.
6167 * @param[in] newdata The data for the newly inserted node.
6168 * @param[in] newpgno The page number, if the new node is a branch node.
6169 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6170 * @return 0 on success, non-zero on failure.
6173 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6174 unsigned int nflags)
6177 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6180 unsigned int i, j, split_indx, nkeys, pmax;
6182 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6184 MDB_page *mp, *rp, *pp;
6189 mp = mc->mc_pg[mc->mc_top];
6190 newindx = mc->mc_ki[mc->mc_top];
6192 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6193 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6194 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6196 /* Create a right sibling. */
6197 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6199 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6201 if (mc->mc_snum < 2) {
6202 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6204 /* shift current top to make room for new parent */
6205 mc->mc_pg[1] = mc->mc_pg[0];
6206 mc->mc_ki[1] = mc->mc_ki[0];
6209 mc->mc_db->md_root = pp->mp_pgno;
6210 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6211 mc->mc_db->md_depth++;
6214 /* Add left (implicit) pointer. */
6215 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6216 /* undo the pre-push */
6217 mc->mc_pg[0] = mc->mc_pg[1];
6218 mc->mc_ki[0] = mc->mc_ki[1];
6219 mc->mc_db->md_root = mp->mp_pgno;
6220 mc->mc_db->md_depth--;
6227 ptop = mc->mc_top-1;
6228 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6231 mc->mc_flags |= C_SPLITTING;
6232 mdb_cursor_copy(mc, &mn);
6233 mn.mc_pg[mn.mc_top] = rp;
6234 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6236 if (nflags & MDB_APPEND) {
6237 mn.mc_ki[mn.mc_top] = 0;
6239 split_indx = newindx;
6244 nkeys = NUMKEYS(mp);
6245 split_indx = nkeys / 2;
6246 if (newindx < split_indx)
6252 unsigned int lsize, rsize, ksize;
6253 /* Move half of the keys to the right sibling */
6255 x = mc->mc_ki[mc->mc_top] - split_indx;
6256 ksize = mc->mc_db->md_pad;
6257 split = LEAF2KEY(mp, split_indx, ksize);
6258 rsize = (nkeys - split_indx) * ksize;
6259 lsize = (nkeys - split_indx) * sizeof(indx_t);
6260 mp->mp_lower -= lsize;
6261 rp->mp_lower += lsize;
6262 mp->mp_upper += rsize - lsize;
6263 rp->mp_upper -= rsize - lsize;
6264 sepkey.mv_size = ksize;
6265 if (newindx == split_indx) {
6266 sepkey.mv_data = newkey->mv_data;
6268 sepkey.mv_data = split;
6271 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6272 memcpy(rp->mp_ptrs, split, rsize);
6273 sepkey.mv_data = rp->mp_ptrs;
6274 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6275 memcpy(ins, newkey->mv_data, ksize);
6276 mp->mp_lower += sizeof(indx_t);
6277 mp->mp_upper -= ksize - sizeof(indx_t);
6280 memcpy(rp->mp_ptrs, split, x * ksize);
6281 ins = LEAF2KEY(rp, x, ksize);
6282 memcpy(ins, newkey->mv_data, ksize);
6283 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6284 rp->mp_lower += sizeof(indx_t);
6285 rp->mp_upper -= ksize - sizeof(indx_t);
6286 mc->mc_ki[mc->mc_top] = x;
6287 mc->mc_pg[mc->mc_top] = rp;
6292 /* For leaf pages, check the split point based on what
6293 * fits where, since otherwise mdb_node_add can fail.
6295 * This check is only needed when the data items are
6296 * relatively large, such that being off by one will
6297 * make the difference between success or failure.
6299 * It's also relevant if a page happens to be laid out
6300 * such that one half of its nodes are all "small" and
6301 * the other half of its nodes are "large." If the new
6302 * item is also "large" and falls on the half with
6303 * "large" nodes, it also may not fit.
6306 unsigned int psize, nsize;
6307 /* Maximum free space in an empty page */
6308 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6309 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6310 if ((nkeys < 20) || (nsize > pmax/16)) {
6311 if (newindx <= split_indx) {
6314 for (i=0; i<split_indx; i++) {
6315 node = NODEPTR(mp, i);
6316 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6317 if (F_ISSET(node->mn_flags, F_BIGDATA))
6318 psize += sizeof(pgno_t);
6320 psize += NODEDSZ(node);
6324 split_indx = newindx;
6335 for (i=nkeys-1; i>=split_indx; i--) {
6336 node = NODEPTR(mp, i);
6337 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6338 if (F_ISSET(node->mn_flags, F_BIGDATA))
6339 psize += sizeof(pgno_t);
6341 psize += NODEDSZ(node);
6345 split_indx = newindx;
6356 /* First find the separating key between the split pages.
6357 * The case where newindx == split_indx is ambiguous; the
6358 * new item could go to the new page or stay on the original
6359 * page. If newpos == 1 it goes to the new page.
6361 if (newindx == split_indx && newpos) {
6362 sepkey.mv_size = newkey->mv_size;
6363 sepkey.mv_data = newkey->mv_data;
6365 node = NODEPTR(mp, split_indx);
6366 sepkey.mv_size = node->mn_ksize;
6367 sepkey.mv_data = NODEKEY(node);
6371 DPRINTF("separator is [%s]", DKEY(&sepkey));
6373 /* Copy separator key to the parent.
6375 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6379 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6382 if (mn.mc_snum == mc->mc_snum) {
6383 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6384 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6385 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6386 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6391 /* Right page might now have changed parent.
6392 * Check if left page also changed parent.
6394 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6395 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6396 for (i=0; i<ptop; i++) {
6397 mc->mc_pg[i] = mn.mc_pg[i];
6398 mc->mc_ki[i] = mn.mc_ki[i];
6400 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6401 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6405 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6408 mc->mc_flags ^= C_SPLITTING;
6409 if (rc != MDB_SUCCESS) {
6412 if (nflags & MDB_APPEND) {
6413 mc->mc_pg[mc->mc_top] = rp;
6414 mc->mc_ki[mc->mc_top] = 0;
6415 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6418 for (i=0; i<mc->mc_top; i++)
6419 mc->mc_ki[i] = mn.mc_ki[i];
6426 /* Move half of the keys to the right sibling. */
6428 /* grab a page to hold a temporary copy */
6429 copy = mdb_page_malloc(mc);
6433 copy->mp_pgno = mp->mp_pgno;
6434 copy->mp_flags = mp->mp_flags;
6435 copy->mp_lower = PAGEHDRSZ;
6436 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6437 mc->mc_pg[mc->mc_top] = copy;
6438 for (i = j = 0; i <= nkeys; j++) {
6439 if (i == split_indx) {
6440 /* Insert in right sibling. */
6441 /* Reset insert index for right sibling. */
6442 if (i != newindx || (newpos ^ ins_new)) {
6444 mc->mc_pg[mc->mc_top] = rp;
6448 if (i == newindx && !ins_new) {
6449 /* Insert the original entry that caused the split. */
6450 rkey.mv_data = newkey->mv_data;
6451 rkey.mv_size = newkey->mv_size;
6460 /* Update index for the new key. */
6461 mc->mc_ki[mc->mc_top] = j;
6462 } else if (i == nkeys) {
6465 node = NODEPTR(mp, i);
6466 rkey.mv_data = NODEKEY(node);
6467 rkey.mv_size = node->mn_ksize;
6469 xdata.mv_data = NODEDATA(node);
6470 xdata.mv_size = NODEDSZ(node);
6473 pgno = NODEPGNO(node);
6474 flags = node->mn_flags;
6479 if (!IS_LEAF(mp) && j == 0) {
6480 /* First branch index doesn't need key data. */
6484 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6488 nkeys = NUMKEYS(copy);
6489 for (i=0; i<nkeys; i++)
6490 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6491 mp->mp_lower = copy->mp_lower;
6492 mp->mp_upper = copy->mp_upper;
6493 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6494 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6496 /* reset back to original page */
6497 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6498 mc->mc_pg[mc->mc_top] = mp;
6499 if (nflags & MDB_RESERVE) {
6500 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6501 if (!(node->mn_flags & F_BIGDATA))
6502 newdata->mv_data = NODEDATA(node);
6508 /* return tmp page to freelist */
6509 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6510 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6511 mc->mc_txn->mt_env->me_dpages = copy;
6514 /* Adjust other cursors pointing to mp */
6515 MDB_cursor *m2, *m3;
6516 MDB_dbi dbi = mc->mc_dbi;
6517 int fixup = NUMKEYS(mp);
6519 if (mc->mc_flags & C_SUB)
6522 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6523 if (m2 == mc) continue;
6524 if (mc->mc_flags & C_SUB)
6525 m3 = &m2->mc_xcursor->mx_cursor;
6528 if (!(m3->mc_flags & C_INITIALIZED))
6530 if (m3->mc_flags & C_SPLITTING)
6535 for (k=m3->mc_top; k>=0; k--) {
6536 m3->mc_ki[k+1] = m3->mc_ki[k];
6537 m3->mc_pg[k+1] = m3->mc_pg[k];
6539 if (m3->mc_ki[0] >= split_indx) {
6544 m3->mc_pg[0] = mc->mc_pg[0];
6548 if (m3->mc_pg[mc->mc_top] == mp) {
6549 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6550 m3->mc_ki[mc->mc_top]++;
6551 if (m3->mc_ki[mc->mc_top] >= fixup) {
6552 m3->mc_pg[mc->mc_top] = rp;
6553 m3->mc_ki[mc->mc_top] -= fixup;
6554 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6556 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6557 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6566 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6567 MDB_val *key, MDB_val *data, unsigned int flags)
6572 assert(key != NULL);
6573 assert(data != NULL);
6575 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6578 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6582 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6586 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6589 mdb_cursor_init(&mc, txn, dbi, &mx);
6590 return mdb_cursor_put(&mc, key, data, flags);
6593 /** Only a subset of the @ref mdb_env flags can be changed
6594 * at runtime. Changing other flags requires closing the environment
6595 * and re-opening it with the new flags.
6597 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
6599 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6601 if ((flag & CHANGEABLE) != flag)
6604 env->me_flags |= flag;
6606 env->me_flags &= ~flag;
6611 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6616 *arg = env->me_flags;
6621 mdb_env_get_path(MDB_env *env, const char **arg)
6626 *arg = env->me_path;
6630 /** Common code for #mdb_stat() and #mdb_env_stat().
6631 * @param[in] env the environment to operate in.
6632 * @param[in] db the #MDB_db record containing the stats to return.
6633 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6634 * @return 0, this function always succeeds.
6637 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6639 arg->ms_psize = env->me_psize;
6640 arg->ms_depth = db->md_depth;
6641 arg->ms_branch_pages = db->md_branch_pages;
6642 arg->ms_leaf_pages = db->md_leaf_pages;
6643 arg->ms_overflow_pages = db->md_overflow_pages;
6644 arg->ms_entries = db->md_entries;
6649 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6653 if (env == NULL || arg == NULL)
6656 toggle = mdb_env_pick_meta(env);
6658 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6662 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6666 if (env == NULL || arg == NULL)
6669 toggle = mdb_env_pick_meta(env);
6670 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6671 arg->me_mapsize = env->me_mapsize;
6672 arg->me_maxreaders = env->me_maxreaders;
6673 arg->me_numreaders = env->me_numreaders;
6674 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6675 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6679 /** Set the default comparison functions for a database.
6680 * Called immediately after a database is opened to set the defaults.
6681 * The user can then override them with #mdb_set_compare() or
6682 * #mdb_set_dupsort().
6683 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6684 * @param[in] dbi A database handle returned by #mdb_open()
6687 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6689 uint16_t f = txn->mt_dbs[dbi].md_flags;
6691 txn->mt_dbxs[dbi].md_cmp =
6692 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6693 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6695 txn->mt_dbxs[dbi].md_dcmp =
6696 !(f & MDB_DUPSORT) ? 0 :
6697 ((f & MDB_INTEGERDUP)
6698 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6699 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6702 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6707 int rc, dbflag, exact;
6708 unsigned int unused = 0;
6711 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6712 mdb_default_cmp(txn, FREE_DBI);
6718 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6719 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6720 mdb_default_cmp(txn, MAIN_DBI);
6724 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6725 mdb_default_cmp(txn, MAIN_DBI);
6728 /* Is the DB already open? */
6730 for (i=2; i<txn->mt_numdbs; i++) {
6731 if (!txn->mt_dbxs[i].md_name.mv_size) {
6732 /* Remember this free slot */
6733 if (!unused) unused = i;
6736 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6737 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6743 /* If no free slot and max hit, fail */
6744 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
6745 return MDB_DBS_FULL;
6747 /* Find the DB info */
6751 key.mv_data = (void *)name;
6752 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6753 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6754 if (rc == MDB_SUCCESS) {
6755 /* make sure this is actually a DB */
6756 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6757 if (!(node->mn_flags & F_SUBDATA))
6759 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6760 /* Create if requested */
6762 data.mv_size = sizeof(MDB_db);
6763 data.mv_data = &dummy;
6764 memset(&dummy, 0, sizeof(dummy));
6765 dummy.md_root = P_INVALID;
6766 dummy.md_flags = flags & 0xffff;
6767 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6771 /* OK, got info, add to table */
6772 if (rc == MDB_SUCCESS) {
6773 unsigned int slot = unused ? unused : txn->mt_numdbs;
6774 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6775 txn->mt_dbxs[slot].md_name.mv_size = len;
6776 txn->mt_dbxs[slot].md_rel = NULL;
6777 txn->mt_dbflags[slot] = dbflag;
6778 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6780 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6781 mdb_default_cmp(txn, slot);
6784 txn->mt_env->me_numdbs++;
6791 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6793 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6796 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6799 void mdb_close(MDB_env *env, MDB_dbi dbi)
6802 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6804 ptr = env->me_dbxs[dbi].md_name.mv_data;
6805 env->me_dbxs[dbi].md_name.mv_data = NULL;
6806 env->me_dbxs[dbi].md_name.mv_size = 0;
6810 /** Add all the DB's pages to the free list.
6811 * @param[in] mc Cursor on the DB to free.
6812 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6813 * @return 0 on success, non-zero on failure.
6816 mdb_drop0(MDB_cursor *mc, int subs)
6820 rc = mdb_page_search(mc, NULL, 0);
6821 if (rc == MDB_SUCCESS) {
6826 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6827 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6830 mdb_cursor_copy(mc, &mx);
6831 while (mc->mc_snum > 0) {
6832 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6833 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6834 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6835 if (ni->mn_flags & F_SUBDATA) {
6836 mdb_xcursor_init1(mc, ni);
6837 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6843 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6845 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6848 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6853 rc = mdb_cursor_sibling(mc, 1);
6855 /* no more siblings, go back to beginning
6856 * of previous level. (stack was already popped
6857 * by mdb_cursor_sibling)
6859 for (i=1; i<mc->mc_top; i++)
6860 mc->mc_pg[i] = mx.mc_pg[i];
6864 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6865 mc->mc_db->md_root);
6870 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6875 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6878 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6881 rc = mdb_cursor_open(txn, dbi, &mc);
6885 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6889 /* Can't delete the main DB */
6890 if (del && dbi > MAIN_DBI) {
6891 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6893 mdb_close(txn->mt_env, dbi);
6895 /* reset the DB record, mark it dirty */
6896 txn->mt_dbflags[dbi] |= DB_DIRTY;
6897 txn->mt_dbs[dbi].md_depth = 0;
6898 txn->mt_dbs[dbi].md_branch_pages = 0;
6899 txn->mt_dbs[dbi].md_leaf_pages = 0;
6900 txn->mt_dbs[dbi].md_overflow_pages = 0;
6901 txn->mt_dbs[dbi].md_entries = 0;
6902 txn->mt_dbs[dbi].md_root = P_INVALID;
6904 if (!txn->mt_u.dirty_list[0].mid) {
6907 /* make sure we have at least one dirty page in this txn
6908 * otherwise these changes will be ignored.
6910 key.mv_size = sizeof(txnid_t);
6911 key.mv_data = &txn->mt_txnid;
6912 data.mv_size = sizeof(MDB_ID);
6913 data.mv_data = txn->mt_free_pgs;
6914 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
6915 rc = mdb_cursor_put(&m2, &key, &data, 0);
6919 mdb_cursor_close(mc);
6923 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6925 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6928 txn->mt_dbxs[dbi].md_cmp = cmp;
6932 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6934 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6937 txn->mt_dbxs[dbi].md_dcmp = cmp;
6941 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6943 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6946 txn->mt_dbxs[dbi].md_rel = rel;
6950 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6952 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6955 txn->mt_dbxs[dbi].md_relctx = ctx;