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)
300 # define MDB_DEBUG_SKIP
302 /** Print a debug string.
303 * The string is printed literally, with no format processing.
305 #define DPUTS(arg) DPRINTF("%s", arg)
308 /** A default memory page size.
309 * The actual size is platform-dependent, but we use this for
310 * boot-strapping. We probably should not be using this any more.
311 * The #GET_PAGESIZE() macro is used to get the actual size.
313 * Note that we don't currently support Huge pages. On Linux,
314 * regular data files cannot use Huge pages, and in general
315 * Huge pages aren't actually pageable. We rely on the OS
316 * demand-pager to read our data and page it out when memory
317 * pressure from other processes is high. So until OSs have
318 * actual paging support for Huge pages, they're not viable.
320 #define MDB_PAGESIZE 4096
322 /** The minimum number of keys required in a database page.
323 * Setting this to a larger value will place a smaller bound on the
324 * maximum size of a data item. Data items larger than this size will
325 * be pushed into overflow pages instead of being stored directly in
326 * the B-tree node. This value used to default to 4. With a page size
327 * of 4096 bytes that meant that any item larger than 1024 bytes would
328 * go into an overflow page. That also meant that on average 2-3KB of
329 * each overflow page was wasted space. The value cannot be lower than
330 * 2 because then there would no longer be a tree structure. With this
331 * value, items larger than 2KB will go into overflow pages, and on
332 * average only 1KB will be wasted.
334 #define MDB_MINKEYS 2
336 /** A stamp that identifies a file as an MDB file.
337 * There's nothing special about this value other than that it is easily
338 * recognizable, and it will reflect any byte order mismatches.
340 #define MDB_MAGIC 0xBEEFC0DE
342 /** The version number for a database's file format. */
343 #define MDB_VERSION 1
345 /** The maximum size of a key in the database.
346 * While data items have essentially unbounded size, we require that
347 * keys all fit onto a regular page. This limit could be raised a bit
348 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
350 #define MAXKEYSIZE 511
355 * This is used for printing a hex dump of a key's contents.
357 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
358 /** Display a key in hex.
360 * Invoke a function to display a key in hex.
362 #define DKEY(x) mdb_dkey(x, kbuf)
364 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
368 /** An invalid page number.
369 * Mainly used to denote an empty tree.
371 #define P_INVALID (~(pgno_t)0)
373 /** Test if a flag \b f is set in a flag word \b w. */
374 #define F_ISSET(w, f) (((w) & (f)) == (f))
376 /** Used for offsets within a single page.
377 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
380 typedef uint16_t indx_t;
382 /** Default size of memory map.
383 * This is certainly too small for any actual applications. Apps should always set
384 * the size explicitly using #mdb_env_set_mapsize().
386 #define DEFAULT_MAPSIZE 1048576
388 /** @defgroup readers Reader Lock Table
389 * Readers don't acquire any locks for their data access. Instead, they
390 * simply record their transaction ID in the reader table. The reader
391 * mutex is needed just to find an empty slot in the reader table. The
392 * slot's address is saved in thread-specific data so that subsequent read
393 * transactions started by the same thread need no further locking to proceed.
395 * Since the database uses multi-version concurrency control, readers don't
396 * actually need any locking. This table is used to keep track of which
397 * readers are using data from which old transactions, so that we'll know
398 * when a particular old transaction is no longer in use. Old transactions
399 * that have discarded any data pages can then have those pages reclaimed
400 * for use by a later write transaction.
402 * The lock table is constructed such that reader slots are aligned with the
403 * processor's cache line size. Any slot is only ever used by one thread.
404 * This alignment guarantees that there will be no contention or cache
405 * thrashing as threads update their own slot info, and also eliminates
406 * any need for locking when accessing a slot.
408 * A writer thread will scan every slot in the table to determine the oldest
409 * outstanding reader transaction. Any freed pages older than this will be
410 * reclaimed by the writer. The writer doesn't use any locks when scanning
411 * this table. This means that there's no guarantee that the writer will
412 * see the most up-to-date reader info, but that's not required for correct
413 * operation - all we need is to know the upper bound on the oldest reader,
414 * we don't care at all about the newest reader. So the only consequence of
415 * reading stale information here is that old pages might hang around a
416 * while longer before being reclaimed. That's actually good anyway, because
417 * the longer we delay reclaiming old pages, the more likely it is that a
418 * string of contiguous pages can be found after coalescing old pages from
419 * many old transactions together.
421 * @todo We don't actually do such coalescing yet, we grab pages from one
422 * old transaction at a time.
425 /** Number of slots in the reader table.
426 * This value was chosen somewhat arbitrarily. 126 readers plus a
427 * couple mutexes fit exactly into 8KB on my development machine.
428 * Applications should set the table size using #mdb_env_set_maxreaders().
430 #define DEFAULT_READERS 126
432 /** The size of a CPU cache line in bytes. We want our lock structures
433 * aligned to this size to avoid false cache line sharing in the
435 * This value works for most CPUs. For Itanium this should be 128.
441 /** The information we store in a single slot of the reader table.
442 * In addition to a transaction ID, we also record the process and
443 * thread ID that owns a slot, so that we can detect stale information,
444 * e.g. threads or processes that went away without cleaning up.
445 * @note We currently don't check for stale records. We simply re-init
446 * the table when we know that we're the only process opening the
449 typedef struct MDB_rxbody {
450 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
451 * Multiple readers that start at the same time will probably have the
452 * same ID here. Again, it's not important to exclude them from
453 * anything; all we need to know is which version of the DB they
454 * started from so we can avoid overwriting any data used in that
455 * particular version.
458 /** The process ID of the process owning this reader txn. */
460 /** The thread ID of the thread owning this txn. */
464 /** The actual reader record, with cacheline padding. */
465 typedef struct MDB_reader {
468 /** shorthand for mrb_txnid */
469 #define mr_txnid mru.mrx.mrb_txnid
470 #define mr_pid mru.mrx.mrb_pid
471 #define mr_tid mru.mrx.mrb_tid
472 /** cache line alignment */
473 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
477 /** The header for the reader table.
478 * The table resides in a memory-mapped file. (This is a different file
479 * than is used for the main database.)
481 * For POSIX the actual mutexes reside in the shared memory of this
482 * mapped file. On Windows, mutexes are named objects allocated by the
483 * kernel; we store the mutex names in this mapped file so that other
484 * processes can grab them. This same approach is also used on
485 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
486 * process-shared POSIX mutexes. For these cases where a named object
487 * is used, the object name is derived from a 64 bit FNV hash of the
488 * environment pathname. As such, naming collisions are extremely
489 * unlikely. If a collision occurs, the results are unpredictable.
491 typedef struct MDB_txbody {
492 /** Stamp identifying this as an MDB file. It must be set
495 /** Version number of this lock file. Must be set to #MDB_VERSION. */
496 uint32_t mtb_version;
497 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
498 char mtb_rmname[MNAME_LEN];
500 /** Mutex protecting access to this table.
501 * This is the reader lock that #LOCK_MUTEX_R acquires.
503 pthread_mutex_t mtb_mutex;
505 /** The ID of the last transaction committed to the database.
506 * This is recorded here only for convenience; the value can always
507 * be determined by reading the main database meta pages.
510 /** The number of slots that have been used in the reader table.
511 * This always records the maximum count, it is not decremented
512 * when readers release their slots.
514 unsigned mtb_numreaders;
517 /** The actual reader table definition. */
518 typedef struct MDB_txninfo {
521 #define mti_magic mt1.mtb.mtb_magic
522 #define mti_version mt1.mtb.mtb_version
523 #define mti_mutex mt1.mtb.mtb_mutex
524 #define mti_rmname mt1.mtb.mtb_rmname
525 #define mti_txnid mt1.mtb.mtb_txnid
526 #define mti_numreaders mt1.mtb.mtb_numreaders
527 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
530 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
531 char mt2_wmname[MNAME_LEN];
532 #define mti_wmname mt2.mt2_wmname
534 pthread_mutex_t mt2_wmutex;
535 #define mti_wmutex mt2.mt2_wmutex
537 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
539 MDB_reader mti_readers[1];
543 /** Common header for all page types.
544 * Overflow records occupy a number of contiguous pages with no
545 * headers on any page after the first.
547 typedef struct MDB_page {
548 #define mp_pgno mp_p.p_pgno
549 #define mp_next mp_p.p_next
551 pgno_t p_pgno; /**< page number */
552 void * p_next; /**< for in-memory list of freed structs */
555 /** @defgroup mdb_page Page Flags
557 * Flags for the page headers.
560 #define P_BRANCH 0x01 /**< branch page */
561 #define P_LEAF 0x02 /**< leaf page */
562 #define P_OVERFLOW 0x04 /**< overflow page */
563 #define P_META 0x08 /**< meta page */
564 #define P_DIRTY 0x10 /**< dirty page */
565 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
566 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
568 uint16_t mp_flags; /**< @ref mdb_page */
569 #define mp_lower mp_pb.pb.pb_lower
570 #define mp_upper mp_pb.pb.pb_upper
571 #define mp_pages mp_pb.pb_pages
574 indx_t pb_lower; /**< lower bound of free space */
575 indx_t pb_upper; /**< upper bound of free space */
577 uint32_t pb_pages; /**< number of overflow pages */
579 indx_t mp_ptrs[1]; /**< dynamic size */
582 /** Size of the page header, excluding dynamic data at the end */
583 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
585 /** Address of first usable data byte in a page, after the header */
586 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
588 /** Number of nodes on a page */
589 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
591 /** The amount of space remaining in the page */
592 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
594 /** The percentage of space used in the page, in tenths of a percent. */
595 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
596 ((env)->me_psize - PAGEHDRSZ))
597 /** The minimum page fill factor, in tenths of a percent.
598 * Pages emptier than this are candidates for merging.
600 #define FILL_THRESHOLD 250
602 /** Test if a page is a leaf page */
603 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
604 /** Test if a page is a LEAF2 page */
605 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
606 /** Test if a page is a branch page */
607 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
608 /** Test if a page is an overflow page */
609 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
610 /** Test if a page is a sub page */
611 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
613 /** The number of overflow pages needed to store the given size. */
614 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
616 /** Header for a single key/data pair within a page.
617 * We guarantee 2-byte alignment for nodes.
619 typedef struct MDB_node {
620 /** lo and hi are used for data size on leaf nodes and for
621 * child pgno on branch nodes. On 64 bit platforms, flags
622 * is also used for pgno. (Branch nodes have no flags).
623 * They are in host byte order in case that lets some
624 * accesses be optimized into a 32-bit word access.
626 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
627 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
628 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
629 /** @defgroup mdb_node Node Flags
631 * Flags for node headers.
634 #define F_BIGDATA 0x01 /**< data put on overflow page */
635 #define F_SUBDATA 0x02 /**< data is a sub-database */
636 #define F_DUPDATA 0x04 /**< data has duplicates */
638 /** valid flags for #mdb_node_add() */
639 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
642 unsigned short mn_flags; /**< @ref mdb_node */
643 unsigned short mn_ksize; /**< key size */
644 char mn_data[1]; /**< key and data are appended here */
647 /** Size of the node header, excluding dynamic data at the end */
648 #define NODESIZE offsetof(MDB_node, mn_data)
650 /** Bit position of top word in page number, for shifting mn_flags */
651 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
653 /** Size of a node in a branch page with a given key.
654 * This is just the node header plus the key, there is no data.
656 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
658 /** Size of a node in a leaf page with a given key and data.
659 * This is node header plus key plus data size.
661 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
663 /** Address of node \b i in page \b p */
664 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
666 /** Address of the key for the node */
667 #define NODEKEY(node) (void *)((node)->mn_data)
669 /** Address of the data for a node */
670 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
672 /** Get the page number pointed to by a branch node */
673 #define NODEPGNO(node) \
674 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
675 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
676 /** Set the page number in a branch node */
677 #define SETPGNO(node,pgno) do { \
678 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
679 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
681 /** Get the size of the data in a leaf node */
682 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
683 /** Set the size of the data for a leaf node */
684 #define SETDSZ(node,size) do { \
685 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
686 /** The size of a key in a node */
687 #define NODEKSZ(node) ((node)->mn_ksize)
689 /** Copy a page number from src to dst */
691 #define COPY_PGNO(dst,src) dst = src
693 #if SIZE_MAX > 4294967295UL
694 #define COPY_PGNO(dst,src) do { \
695 unsigned short *s, *d; \
696 s = (unsigned short *)&(src); \
697 d = (unsigned short *)&(dst); \
704 #define COPY_PGNO(dst,src) do { \
705 unsigned short *s, *d; \
706 s = (unsigned short *)&(src); \
707 d = (unsigned short *)&(dst); \
713 /** The address of a key in a LEAF2 page.
714 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
715 * There are no node headers, keys are stored contiguously.
717 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
719 /** Set the \b node's key into \b key, if requested. */
720 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
721 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
723 /** Information about a single database in the environment. */
724 typedef struct MDB_db {
725 uint32_t md_pad; /**< also ksize for LEAF2 pages */
726 uint16_t md_flags; /**< @ref mdb_dbi_open */
727 uint16_t md_depth; /**< depth of this tree */
728 pgno_t md_branch_pages; /**< number of internal pages */
729 pgno_t md_leaf_pages; /**< number of leaf pages */
730 pgno_t md_overflow_pages; /**< number of overflow pages */
731 size_t md_entries; /**< number of data items */
732 pgno_t md_root; /**< the root page of this tree */
735 /** Handle for the DB used to track free pages. */
737 /** Handle for the default DB. */
740 /** Meta page content. */
741 typedef struct MDB_meta {
742 /** Stamp identifying this as an MDB file. It must be set
745 /** Version number of this lock file. Must be set to #MDB_VERSION. */
747 void *mm_address; /**< address for fixed mapping */
748 size_t mm_mapsize; /**< size of mmap region */
749 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
750 /** The size of pages used in this DB */
751 #define mm_psize mm_dbs[0].md_pad
752 /** Any persistent environment flags. @ref mdb_env */
753 #define mm_flags mm_dbs[0].md_flags
754 pgno_t mm_last_pg; /**< last used page in file */
755 txnid_t mm_txnid; /**< txnid that committed this page */
758 /** Buffer for a stack-allocated dirty page.
759 * The members define size and alignment, and silence type
760 * aliasing warnings. They are not used directly; that could
761 * mean incorrectly using several union members in parallel.
763 typedef union MDB_pagebuf {
764 char mb_raw[MDB_PAGESIZE];
767 char mm_pad[PAGEHDRSZ];
772 /** Auxiliary DB info.
773 * The information here is mostly static/read-only. There is
774 * only a single copy of this record in the environment.
776 typedef struct MDB_dbx {
777 MDB_val md_name; /**< name of the database */
778 MDB_cmp_func *md_cmp; /**< function for comparing keys */
779 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
780 MDB_rel_func *md_rel; /**< user relocate function */
781 void *md_relctx; /**< user-provided context for md_rel */
784 /** A database transaction.
785 * Every operation requires a transaction handle.
788 MDB_txn *mt_parent; /**< parent of a nested txn */
789 MDB_txn *mt_child; /**< nested txn under this txn */
790 pgno_t mt_next_pgno; /**< next unallocated page */
791 /** The ID of this transaction. IDs are integers incrementing from 1.
792 * Only committed write transactions increment the ID. If a transaction
793 * aborts, the ID may be re-used by the next writer.
796 MDB_env *mt_env; /**< the DB environment */
797 /** The list of pages that became unused during this transaction.
801 MDB_ID2L dirty_list; /**< modified pages */
802 MDB_reader *reader; /**< this thread's slot in the reader table */
804 /** Array of records for each DB known in the environment. */
806 /** Array of MDB_db records for each known DB */
808 /** @defgroup mt_dbflag Transaction DB Flags
812 #define DB_DIRTY 0x01 /**< DB was written in this txn */
813 #define DB_STALE 0x02 /**< DB record is older than txnID */
815 /** Array of cursors for each DB */
816 MDB_cursor **mt_cursors;
817 /** Array of flags for each DB */
818 unsigned char *mt_dbflags;
819 /** Number of DB records in use. This number only ever increments;
820 * we don't decrement it when individual DB handles are closed.
824 /** @defgroup mdb_txn Transaction Flags
828 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
829 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
830 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
832 unsigned int mt_flags; /**< @ref mdb_txn */
833 /** Tracks which of the two meta pages was used at the start
834 * of this transaction.
836 unsigned int mt_toggle;
839 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
840 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
841 * raise this on a 64 bit machine.
843 #define CURSOR_STACK 32
847 /** Cursors are used for all DB operations */
849 /** Next cursor on this DB in this txn */
851 /** Original cursor if this is a shadow */
853 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
854 struct MDB_xcursor *mc_xcursor;
855 /** The transaction that owns this cursor */
857 /** The database handle this cursor operates on */
859 /** The database record for this cursor */
861 /** The database auxiliary record for this cursor */
863 /** The @ref mt_dbflag for this database */
864 unsigned char *mc_dbflag;
865 unsigned short mc_snum; /**< number of pushed pages */
866 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
867 /** @defgroup mdb_cursor Cursor Flags
869 * Cursor state flags.
872 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
873 #define C_EOF 0x02 /**< No more data */
874 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
875 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
876 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
877 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
879 unsigned int mc_flags; /**< @ref mdb_cursor */
880 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
881 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
884 /** Context for sorted-dup records.
885 * We could have gone to a fully recursive design, with arbitrarily
886 * deep nesting of sub-databases. But for now we only handle these
887 * levels - main DB, optional sub-DB, sorted-duplicate DB.
889 typedef struct MDB_xcursor {
890 /** A sub-cursor for traversing the Dup DB */
891 MDB_cursor mx_cursor;
892 /** The database record for this Dup DB */
894 /** The auxiliary DB record for this Dup DB */
896 /** The @ref mt_dbflag for this Dup DB */
897 unsigned char mx_dbflag;
900 /** A set of pages freed by an earlier transaction. */
901 typedef struct MDB_oldpages {
902 /** Usually we only read one record from the FREEDB at a time, but
903 * in case we read more, this will chain them together.
905 struct MDB_oldpages *mo_next;
906 /** The ID of the transaction in which these pages were freed. */
908 /** An #MDB_IDL of the pages */
909 pgno_t mo_pages[1]; /* dynamic */
912 /** The database environment. */
914 HANDLE me_fd; /**< The main data file */
915 HANDLE me_lfd; /**< The lock file */
916 HANDLE me_mfd; /**< just for writing the meta pages */
917 /** Failed to update the meta page. Probably an I/O error. */
918 #define MDB_FATAL_ERROR 0x80000000U
919 /** Read-only Filesystem. Allow read access, no locking. */
920 #define MDB_ROFS 0x40000000U
921 /** Some fields are initialized. */
922 #define MDB_ENV_ACTIVE 0x20000000U
923 uint32_t me_flags; /**< @ref mdb_env */
924 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
925 unsigned int me_maxreaders; /**< size of the reader table */
926 unsigned int me_numreaders; /**< max numreaders set by this env */
927 MDB_dbi me_numdbs; /**< number of DBs opened */
928 MDB_dbi me_maxdbs; /**< size of the DB table */
929 pid_t me_pid; /**< process ID of this env */
930 char *me_path; /**< path to the DB files */
931 char *me_map; /**< the memory map of the data file */
932 MDB_txninfo *me_txns; /**< the memory map of the lock file */
933 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
934 MDB_txn *me_txn; /**< current write transaction */
935 size_t me_mapsize; /**< size of the data memory map */
936 off_t me_size; /**< current file size */
937 pgno_t me_maxpg; /**< me_mapsize / me_psize */
938 txnid_t me_pgfirst; /**< ID of first old page record we used */
939 txnid_t me_pglast; /**< ID of last old page record we used */
940 MDB_dbx *me_dbxs; /**< array of static DB info */
941 uint16_t *me_dbflags; /**< array of DB flags */
942 MDB_oldpages *me_pghead; /**< list of old page records */
943 MDB_oldpages *me_pgfree; /**< list of page records to free */
944 pthread_key_t me_txkey; /**< thread-key for readers */
945 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
946 /** IDL of pages that became unused in a write txn */
948 /** ID2L of pages that were written during a write txn */
949 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
951 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
953 #elif defined(MDB_USE_POSIX_SEM)
954 sem_t *me_rmutex; /* Shared mutexes are not supported */
958 /** max number of pages to commit in one writev() call */
959 #define MDB_COMMIT_PAGES 64
960 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
961 #undef MDB_COMMIT_PAGES
962 #define MDB_COMMIT_PAGES IOV_MAX
965 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
966 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
967 static int mdb_page_touch(MDB_cursor *mc);
969 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
970 static int mdb_page_search_root(MDB_cursor *mc,
971 MDB_val *key, int modify);
972 #define MDB_PS_MODIFY 1
973 #define MDB_PS_ROOTONLY 2
974 static int mdb_page_search(MDB_cursor *mc,
975 MDB_val *key, int flags);
976 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
978 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
979 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
980 pgno_t newpgno, unsigned int nflags);
982 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
983 static int mdb_env_pick_meta(const MDB_env *env);
984 static int mdb_env_write_meta(MDB_txn *txn);
985 static void mdb_env_close0(MDB_env *env, int excl);
987 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
988 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
989 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
990 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
991 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
992 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
993 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
994 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
995 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
997 static int mdb_rebalance(MDB_cursor *mc);
998 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
1000 static void mdb_cursor_pop(MDB_cursor *mc);
1001 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1003 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1004 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1005 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1006 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1007 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1009 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1010 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1012 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1013 static void mdb_xcursor_init0(MDB_cursor *mc);
1014 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1016 static int mdb_drop0(MDB_cursor *mc, int subs);
1017 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1020 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1024 static SECURITY_DESCRIPTOR mdb_null_sd;
1025 static SECURITY_ATTRIBUTES mdb_all_sa;
1026 static int mdb_sec_inited;
1029 /** Return the library version info. */
1031 mdb_version(int *major, int *minor, int *patch)
1033 if (major) *major = MDB_VERSION_MAJOR;
1034 if (minor) *minor = MDB_VERSION_MINOR;
1035 if (patch) *patch = MDB_VERSION_PATCH;
1036 return MDB_VERSION_STRING;
1039 /** Table of descriptions for MDB @ref errors */
1040 static char *const mdb_errstr[] = {
1041 "MDB_KEYEXIST: Key/data pair already exists",
1042 "MDB_NOTFOUND: No matching key/data pair found",
1043 "MDB_PAGE_NOTFOUND: Requested page not found",
1044 "MDB_CORRUPTED: Located page was wrong type",
1045 "MDB_PANIC: Update of meta page failed",
1046 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1047 "MDB_INVALID: File is not an MDB file",
1048 "MDB_MAP_FULL: Environment mapsize limit reached",
1049 "MDB_DBS_FULL: Environment maxdbs limit reached",
1050 "MDB_READERS_FULL: Environment maxreaders limit reached",
1051 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1052 "MDB_TXN_FULL: Nested transaction has too many dirty pages - transaction too big",
1053 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1054 "MDB_PAGE_FULL: Internal error - page has no more space"
1058 mdb_strerror(int err)
1062 return ("Successful return: 0");
1064 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1065 i = err - MDB_KEYEXIST;
1066 return mdb_errstr[i];
1069 return strerror(err);
1073 /** Display a key in hexadecimal and return the address of the result.
1074 * @param[in] key the key to display
1075 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1076 * @return The key in hexadecimal form.
1079 mdb_dkey(MDB_val *key, char *buf)
1082 unsigned char *c = key->mv_data;
1084 if (key->mv_size > MAXKEYSIZE)
1085 return "MAXKEYSIZE";
1086 /* may want to make this a dynamic check: if the key is mostly
1087 * printable characters, print it as-is instead of converting to hex.
1091 for (i=0; i<key->mv_size; i++)
1092 ptr += sprintf(ptr, "%02x", *c++);
1094 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1099 /** Display all the keys in the page. */
1101 mdb_page_list(MDB_page *mp)
1104 unsigned int i, nkeys, nsize;
1108 nkeys = NUMKEYS(mp);
1109 fprintf(stderr, "numkeys %d\n", nkeys);
1110 for (i=0; i<nkeys; i++) {
1111 node = NODEPTR(mp, i);
1112 key.mv_size = node->mn_ksize;
1113 key.mv_data = node->mn_data;
1114 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1115 if (F_ISSET(node->mn_flags, F_BIGDATA))
1116 nsize += sizeof(pgno_t);
1118 nsize += NODEDSZ(node);
1119 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1124 mdb_cursor_chk(MDB_cursor *mc)
1130 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1131 for (i=0; i<mc->mc_top; i++) {
1133 node = NODEPTR(mp, mc->mc_ki[i]);
1134 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1137 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1143 /** Count all the pages in each DB and in the freelist
1144 * and make sure it matches the actual number of pages
1147 static void mdb_audit(MDB_txn *txn)
1151 MDB_ID freecount, count;
1156 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1157 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1158 freecount += *(MDB_ID *)data.mv_data;
1161 for (i = 0; i<txn->mt_numdbs; i++) {
1162 MDB_xcursor mx, *mxp;
1163 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1164 mdb_cursor_init(&mc, txn, i, mxp);
1165 if (txn->mt_dbs[i].md_root == P_INVALID)
1167 count += txn->mt_dbs[i].md_branch_pages +
1168 txn->mt_dbs[i].md_leaf_pages +
1169 txn->mt_dbs[i].md_overflow_pages;
1170 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1171 mdb_page_search(&mc, NULL, 0);
1175 mp = mc.mc_pg[mc.mc_top];
1176 for (j=0; j<NUMKEYS(mp); j++) {
1177 MDB_node *leaf = NODEPTR(mp, j);
1178 if (leaf->mn_flags & F_SUBDATA) {
1180 memcpy(&db, NODEDATA(leaf), sizeof(db));
1181 count += db.md_branch_pages + db.md_leaf_pages +
1182 db.md_overflow_pages;
1186 while (mdb_cursor_sibling(&mc, 1) == 0);
1189 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1190 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1191 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1197 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1199 return txn->mt_dbxs[dbi].md_cmp(a, b);
1203 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1205 if (txn->mt_dbxs[dbi].md_dcmp)
1206 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1208 return EINVAL; /* too bad you can't distinguish this from a valid result */
1211 /** Allocate a single page.
1212 * Re-use old malloc'd pages first, otherwise just malloc.
1215 mdb_page_malloc(MDB_cursor *mc) {
1217 size_t sz = mc->mc_txn->mt_env->me_psize;
1218 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1219 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1220 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1221 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1222 } else if ((ret = malloc(sz)) != NULL) {
1223 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1228 /** Allocate pages for writing.
1229 * If there are free pages available from older transactions, they
1230 * will be re-used first. Otherwise a new page will be allocated.
1231 * @param[in] mc cursor A cursor handle identifying the transaction and
1232 * database for which we are allocating.
1233 * @param[in] num the number of pages to allocate.
1234 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1235 * will always be satisfied by a single contiguous chunk of memory.
1236 * @return 0 on success, non-zero on failure.
1239 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1241 MDB_txn *txn = mc->mc_txn;
1243 pgno_t pgno = P_INVALID;
1245 txnid_t oldest = 0, last;
1249 /* The free list won't have any content at all until txn 2 has
1250 * committed. The pages freed by txn 2 will be unreferenced
1251 * after txn 3 commits, and so will be safe to re-use in txn 4.
1253 if (txn->mt_txnid > 3) {
1255 if (!txn->mt_env->me_pghead &&
1256 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1257 /* See if there's anything in the free DB */
1264 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1265 if (!txn->mt_env->me_pgfirst) {
1266 mdb_page_search(&m2, NULL, 0);
1267 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1268 kptr = (txnid_t *)NODEKEY(leaf);
1275 last = txn->mt_env->me_pglast + 1;
1277 key.mv_data = &last;
1278 key.mv_size = sizeof(last);
1279 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1282 last = *(txnid_t *)key.mv_data;
1288 oldest = txn->mt_txnid - 1;
1289 nr = txn->mt_env->me_txns->mti_numreaders;
1290 r = txn->mt_env->me_txns->mti_readers;
1291 for (i=0; i<nr; i++) {
1292 if (!r[i].mr_pid) continue;
1299 if (oldest > last) {
1300 /* It's usable, grab it.
1305 if (!txn->mt_env->me_pgfirst) {
1306 mdb_node_read(txn, leaf, &data);
1308 txn->mt_env->me_pglast = last;
1309 if (!txn->mt_env->me_pgfirst)
1310 txn->mt_env->me_pgfirst = last;
1311 idl = (MDB_ID *) data.mv_data;
1312 /* We might have a zero-length IDL due to freelist growth
1313 * during a prior commit
1315 if (!idl[0]) goto again;
1316 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1319 mop->mo_next = txn->mt_env->me_pghead;
1320 mop->mo_txnid = last;
1321 txn->mt_env->me_pghead = mop;
1322 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1327 DPRINTF("IDL read txn %zu root %zu num %zu",
1328 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1329 for (i=0; i<idl[0]; i++) {
1330 DPRINTF("IDL %zu", idl[i+1]);
1337 if (txn->mt_env->me_pghead) {
1338 MDB_oldpages *mop = txn->mt_env->me_pghead;
1341 int retry = 60, readit = 0, n2 = num-1;
1342 unsigned int i, j, k;
1344 /* If current list is too short, must fetch more and coalesce */
1345 if (mop->mo_pages[0] < (unsigned)num)
1348 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1356 last = mop->mo_txnid + 1;
1358 /* We haven't hit the readers list yet? */
1364 oldest = txn->mt_txnid - 1;
1365 nr = txn->mt_env->me_txns->mti_numreaders;
1366 r = txn->mt_env->me_txns->mti_readers;
1367 for (i=0; i<nr; i++) {
1368 if (!r[i].mr_pid) continue;
1375 /* There's nothing we can use on the freelist */
1376 if (oldest - last < 1)
1380 key.mv_data = &last;
1381 key.mv_size = sizeof(last);
1382 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1385 idl = (MDB_ID *) data.mv_data;
1386 mop2 = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - 2*sizeof(pgno_t) + MDB_IDL_SIZEOF(mop->mo_pages));
1389 /* merge in sorted order */
1390 i = idl[0]; j = mop->mo_pages[0]; mop2->mo_pages[0] = k = i+j;
1391 mop->mo_pages[0] = P_INVALID;
1392 while (i>0 || j>0) {
1393 if (i && idl[i] < mop->mo_pages[j])
1394 mop2->mo_pages[k--] = idl[i--];
1396 mop2->mo_pages[k--] = mop->mo_pages[j--];
1398 txn->mt_env->me_pglast = last;
1399 mop2->mo_txnid = last;
1400 mop2->mo_next = mop->mo_next;
1401 txn->mt_env->me_pghead = mop2;
1404 /* Keep trying to read until we have enough */
1405 if (mop->mo_pages[0] < (unsigned)num) {
1410 /* current list has enough pages, but are they contiguous? */
1411 for (i=mop->mo_pages[0]; i>=(unsigned)num; i--) {
1412 if (mop->mo_pages[i-n2] == mop->mo_pages[i] + n2) {
1413 pgno = mop->mo_pages[i];
1415 /* move any stragglers down */
1416 for (j=i+num; j<=mop->mo_pages[0]; j++)
1417 mop->mo_pages[i++] = mop->mo_pages[j];
1418 mop->mo_pages[0] -= num;
1423 /* Stop if we succeeded, or no more retries */
1424 if (!retry || pgno != P_INVALID)
1431 /* peel pages off tail, so we only have to truncate the list */
1432 pgno = MDB_IDL_LAST(mop->mo_pages);
1435 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1436 txn->mt_env->me_pghead = mop->mo_next;
1437 if (mc->mc_dbi == FREE_DBI) {
1438 mop->mo_next = txn->mt_env->me_pgfree;
1439 txn->mt_env->me_pgfree = mop;
1447 if (pgno == P_INVALID) {
1448 /* DB size is maxed out */
1449 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1450 DPUTS("DB size maxed out");
1451 return MDB_MAP_FULL;
1454 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1455 if (pgno == P_INVALID) {
1456 pgno = txn->mt_next_pgno;
1457 txn->mt_next_pgno += num;
1459 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1462 if (txn->mt_env->me_dpages && num == 1) {
1463 np = txn->mt_env->me_dpages;
1464 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1465 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1466 txn->mt_env->me_dpages = np->mp_next;
1468 size_t sz = txn->mt_env->me_psize * num;
1469 if ((np = malloc(sz)) == NULL)
1471 VGMEMP_ALLOC(txn->mt_env, np, sz);
1473 if (pgno == P_INVALID) {
1474 np->mp_pgno = txn->mt_next_pgno;
1475 txn->mt_next_pgno += num;
1480 mid.mid = np->mp_pgno;
1482 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1483 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1485 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1492 /** Copy a page: avoid copying unused portions of the page.
1493 * @param[in] dst page to copy into
1494 * @param[in] src page to copy from
1497 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1499 dst->mp_flags = src->mp_flags | P_DIRTY;
1500 dst->mp_pages = src->mp_pages;
1502 if (IS_LEAF2(src)) {
1503 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1505 unsigned int i, nkeys = NUMKEYS(src);
1506 for (i=0; i<nkeys; i++)
1507 dst->mp_ptrs[i] = src->mp_ptrs[i];
1508 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1509 psize - src->mp_upper);
1513 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1514 * @param[in] mc cursor pointing to the page to be touched
1515 * @return 0 on success, non-zero on failure.
1518 mdb_page_touch(MDB_cursor *mc)
1520 MDB_page *mp = mc->mc_pg[mc->mc_top];
1524 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1526 if ((rc = mdb_page_alloc(mc, 1, &np)))
1528 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1529 assert(mp->mp_pgno != np->mp_pgno);
1530 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1532 /* If page isn't full, just copy the used portion */
1533 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1536 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1538 np->mp_flags |= P_DIRTY;
1543 /* Adjust other cursors pointing to mp */
1544 if (mc->mc_flags & C_SUB) {
1545 MDB_cursor *m2, *m3;
1546 MDB_dbi dbi = mc->mc_dbi-1;
1548 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1549 if (m2 == mc) continue;
1550 m3 = &m2->mc_xcursor->mx_cursor;
1551 if (m3->mc_snum < mc->mc_snum) continue;
1552 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1553 m3->mc_pg[mc->mc_top] = mp;
1559 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1560 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1561 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1562 m2->mc_pg[mc->mc_top] = mp;
1566 mc->mc_pg[mc->mc_top] = mp;
1567 /** If this page has a parent, update the parent to point to
1571 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1573 mc->mc_db->md_root = mp->mp_pgno;
1574 } else if (mc->mc_txn->mt_parent) {
1577 /* If txn has a parent, make sure the page is in our
1580 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1581 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1582 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1583 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1584 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1585 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1586 mc->mc_pg[mc->mc_top] = mp;
1592 np = mdb_page_malloc(mc);
1595 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1596 mid.mid = np->mp_pgno;
1598 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1606 mdb_env_sync(MDB_env *env, int force)
1609 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1610 if (env->me_flags & MDB_WRITEMAP) {
1611 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
1612 ? MS_ASYNC : MS_SYNC;
1613 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1616 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1620 if (MDB_FDATASYNC(env->me_fd))
1627 /** Make shadow copies of all of parent txn's cursors */
1629 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1631 MDB_cursor *mc, *m2;
1632 unsigned int i, j, size;
1634 for (i=0;i<src->mt_numdbs; i++) {
1635 if (src->mt_cursors[i]) {
1636 size = sizeof(MDB_cursor);
1637 if (src->mt_cursors[i]->mc_xcursor)
1638 size += sizeof(MDB_xcursor);
1639 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1646 mc->mc_db = &dst->mt_dbs[i];
1647 mc->mc_dbx = m2->mc_dbx;
1648 mc->mc_dbflag = &dst->mt_dbflags[i];
1649 mc->mc_snum = m2->mc_snum;
1650 mc->mc_top = m2->mc_top;
1651 mc->mc_flags = m2->mc_flags | C_SHADOW;
1652 for (j=0; j<mc->mc_snum; j++) {
1653 mc->mc_pg[j] = m2->mc_pg[j];
1654 mc->mc_ki[j] = m2->mc_ki[j];
1656 if (m2->mc_xcursor) {
1657 MDB_xcursor *mx, *mx2;
1658 mx = (MDB_xcursor *)(mc+1);
1659 mc->mc_xcursor = mx;
1660 mx2 = m2->mc_xcursor;
1661 mx->mx_db = mx2->mx_db;
1662 mx->mx_dbx = mx2->mx_dbx;
1663 mx->mx_dbflag = mx2->mx_dbflag;
1664 mx->mx_cursor.mc_txn = dst;
1665 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1666 mx->mx_cursor.mc_db = &mx->mx_db;
1667 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1668 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1669 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1670 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1671 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1672 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1673 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1674 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1677 mc->mc_xcursor = NULL;
1679 mc->mc_next = dst->mt_cursors[i];
1680 dst->mt_cursors[i] = mc;
1687 /** Merge shadow cursors back into parent's */
1689 mdb_cursor_merge(MDB_txn *txn)
1692 for (i=0; i<txn->mt_numdbs; i++) {
1693 if (txn->mt_cursors[i]) {
1695 while ((mc = txn->mt_cursors[i])) {
1696 txn->mt_cursors[i] = mc->mc_next;
1697 if (mc->mc_flags & C_SHADOW) {
1698 MDB_cursor *m2 = mc->mc_orig;
1700 m2->mc_snum = mc->mc_snum;
1701 m2->mc_top = mc->mc_top;
1702 for (j=0; j<mc->mc_snum; j++) {
1703 m2->mc_pg[j] = mc->mc_pg[j];
1704 m2->mc_ki[j] = mc->mc_ki[j];
1707 if (mc->mc_flags & C_ALLOCD)
1715 mdb_txn_reset0(MDB_txn *txn);
1717 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1718 * @param[in] txn the transaction handle to initialize
1719 * @return 0 on success, non-zero on failure. This can only
1720 * fail for read-only transactions, and then only if the
1721 * reader table is full.
1724 mdb_txn_renew0(MDB_txn *txn)
1726 MDB_env *env = txn->mt_env;
1731 txn->mt_numdbs = env->me_numdbs;
1732 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1734 if (txn->mt_flags & MDB_TXN_RDONLY) {
1735 if (env->me_flags & MDB_ROFS) {
1736 i = mdb_env_pick_meta(env);
1737 txn->mt_txnid = env->me_metas[i]->mm_txnid;
1738 txn->mt_u.reader = NULL;
1740 MDB_reader *r = pthread_getspecific(env->me_txkey);
1742 pid_t pid = env->me_pid;
1743 pthread_t tid = pthread_self();
1746 for (i=0; i<env->me_txns->mti_numreaders; i++)
1747 if (env->me_txns->mti_readers[i].mr_pid == 0)
1749 if (i == env->me_maxreaders) {
1750 UNLOCK_MUTEX_R(env);
1751 return MDB_READERS_FULL;
1753 env->me_txns->mti_readers[i].mr_pid = pid;
1754 env->me_txns->mti_readers[i].mr_tid = tid;
1755 if (i >= env->me_txns->mti_numreaders)
1756 env->me_txns->mti_numreaders = i+1;
1757 /* Save numreaders for un-mutexed mdb_env_close() */
1758 env->me_numreaders = env->me_txns->mti_numreaders;
1759 UNLOCK_MUTEX_R(env);
1760 r = &env->me_txns->mti_readers[i];
1761 if ((rc = pthread_setspecific(env->me_txkey, r)) != 0) {
1762 env->me_txns->mti_readers[i].mr_pid = 0;
1766 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1767 txn->mt_u.reader = r;
1769 txn->mt_toggle = txn->mt_txnid & 1;
1770 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1774 txn->mt_txnid = env->me_txns->mti_txnid;
1775 txn->mt_toggle = txn->mt_txnid & 1;
1776 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1779 if (txn->mt_txnid == mdb_debug_start)
1782 txn->mt_u.dirty_list = env->me_dirty_list;
1783 txn->mt_u.dirty_list[0].mid = 0;
1784 txn->mt_free_pgs = env->me_free_pgs;
1785 txn->mt_free_pgs[0] = 0;
1789 /* Copy the DB info and flags */
1790 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1791 for (i=2; i<txn->mt_numdbs; i++)
1792 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1793 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1794 if (txn->mt_numdbs > 2)
1795 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1801 mdb_txn_renew(MDB_txn *txn)
1805 if (! (txn && txn->mt_flags & MDB_TXN_RDONLY))
1808 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1809 DPUTS("environment had fatal error, must shutdown!");
1813 rc = mdb_txn_renew0(txn);
1814 if (rc == MDB_SUCCESS) {
1815 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1816 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1817 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1823 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1828 if (env->me_flags & MDB_FATAL_ERROR) {
1829 DPUTS("environment had fatal error, must shutdown!");
1832 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1835 /* Nested transactions: Max 1 child, write txns only, no writemap */
1836 if (parent->mt_child ||
1837 (flags & MDB_RDONLY) || (parent->mt_flags & MDB_TXN_RDONLY) ||
1838 (env->me_flags & MDB_WRITEMAP))
1843 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1844 if (!(flags & MDB_RDONLY))
1845 size += env->me_maxdbs * sizeof(MDB_cursor *);
1847 if ((txn = calloc(1, size)) == NULL) {
1848 DPRINTF("calloc: %s", strerror(ErrCode()));
1851 txn->mt_dbs = (MDB_db *)(txn+1);
1852 if (flags & MDB_RDONLY) {
1853 txn->mt_flags |= MDB_TXN_RDONLY;
1854 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1856 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1857 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1862 txn->mt_free_pgs = mdb_midl_alloc();
1863 if (!txn->mt_free_pgs) {
1867 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1868 if (!txn->mt_u.dirty_list) {
1869 free(txn->mt_free_pgs);
1873 txn->mt_txnid = parent->mt_txnid;
1874 txn->mt_toggle = parent->mt_toggle;
1875 txn->mt_u.dirty_list[0].mid = 0;
1876 txn->mt_free_pgs[0] = 0;
1877 txn->mt_next_pgno = parent->mt_next_pgno;
1878 parent->mt_child = txn;
1879 txn->mt_parent = parent;
1880 txn->mt_numdbs = parent->mt_numdbs;
1881 txn->mt_dbxs = parent->mt_dbxs;
1882 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1883 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1884 mdb_cursor_shadow(parent, txn);
1887 rc = mdb_txn_renew0(txn);
1893 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1894 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1895 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1901 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1902 * @param[in] txn the transaction handle to reset
1905 mdb_txn_reset0(MDB_txn *txn)
1907 MDB_env *env = txn->mt_env;
1909 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1910 if (!(env->me_flags & MDB_ROFS))
1911 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
1917 /* close(free) all cursors */
1918 for (i=0; i<txn->mt_numdbs; i++) {
1919 if (txn->mt_cursors[i]) {
1921 while ((mc = txn->mt_cursors[i])) {
1922 txn->mt_cursors[i] = mc->mc_next;
1923 if (mc->mc_flags & C_ALLOCD)
1929 if (!(env->me_flags & MDB_WRITEMAP)) {
1930 /* return all dirty pages to dpage list */
1931 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1932 dp = txn->mt_u.dirty_list[i].mptr;
1933 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1934 dp->mp_next = txn->mt_env->me_dpages;
1935 VGMEMP_FREE(txn->mt_env, dp);
1936 txn->mt_env->me_dpages = dp;
1938 /* large pages just get freed directly */
1939 VGMEMP_FREE(txn->mt_env, dp);
1945 if (txn->mt_parent) {
1946 txn->mt_parent->mt_child = NULL;
1947 mdb_midl_free(txn->mt_free_pgs);
1948 free(txn->mt_u.dirty_list);
1951 if (mdb_midl_shrink(&txn->mt_free_pgs))
1952 env->me_free_pgs = txn->mt_free_pgs;
1955 while ((mop = txn->mt_env->me_pghead)) {
1956 txn->mt_env->me_pghead = mop->mo_next;
1959 txn->mt_env->me_pgfirst = 0;
1960 txn->mt_env->me_pglast = 0;
1963 /* The writer mutex was locked in mdb_txn_begin. */
1964 UNLOCK_MUTEX_W(env);
1969 mdb_txn_reset(MDB_txn *txn)
1974 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1975 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1976 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1978 mdb_txn_reset0(txn);
1982 mdb_txn_abort(MDB_txn *txn)
1987 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1988 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1989 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1992 mdb_txn_abort(txn->mt_child);
1994 mdb_txn_reset0(txn);
1999 mdb_txn_commit(MDB_txn *txn)
2007 pgno_t next, freecnt;
2010 assert(txn != NULL);
2011 assert(txn->mt_env != NULL);
2013 if (txn->mt_child) {
2014 mdb_txn_commit(txn->mt_child);
2015 txn->mt_child = NULL;
2020 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2021 if (txn->mt_numdbs > env->me_numdbs) {
2022 /* update the DB flags */
2024 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2025 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2032 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
2033 DPUTS("error flag is set, can't commit");
2035 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
2040 if (txn->mt_parent) {
2046 /* Merge (and close) our cursors with parent's */
2047 mdb_cursor_merge(txn);
2049 /* Update parent's DB table */
2050 ip = &txn->mt_parent->mt_dbs[2];
2051 jp = &txn->mt_dbs[2];
2052 for (i = 2; i < txn->mt_numdbs; i++) {
2053 if (ip->md_root != jp->md_root)
2057 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
2059 /* Append our free list to parent's */
2060 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
2062 mdb_midl_free(txn->mt_free_pgs);
2064 /* Merge our dirty list with parent's */
2065 dst = txn->mt_parent->mt_u.dirty_list;
2066 src = txn->mt_u.dirty_list;
2067 x = mdb_mid2l_search(dst, src[1].mid);
2068 for (y=1; y<=src[0].mid; y++) {
2069 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
2073 dst[x].mptr = src[y].mptr;
2076 for (; y<=src[0].mid; y++) {
2077 if (++x >= MDB_IDL_UM_MAX) {
2079 return MDB_TXN_FULL;
2084 free(txn->mt_u.dirty_list);
2085 txn->mt_parent->mt_child = NULL;
2090 if (txn != env->me_txn) {
2091 DPUTS("attempt to commit unknown transaction");
2096 if (!txn->mt_u.dirty_list[0].mid && !(txn->mt_flags & MDB_TXN_DIRTY))
2099 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
2100 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
2102 /* Update DB root pointers. Their pages have already been
2103 * touched so this is all in-place and cannot fail.
2105 if (txn->mt_numdbs > 2) {
2108 data.mv_size = sizeof(MDB_db);
2110 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2111 for (i = 2; i < txn->mt_numdbs; i++) {
2112 if (txn->mt_dbflags[i] & DB_DIRTY) {
2113 data.mv_data = &txn->mt_dbs[i];
2114 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2119 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2121 /* should only be one record now */
2122 if (env->me_pghead) {
2123 /* make sure first page of freeDB is touched and on freelist */
2124 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2127 /* Delete IDLs we used from the free list */
2128 if (env->me_pgfirst) {
2133 key.mv_size = sizeof(cur);
2134 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2137 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2138 rc = mdb_cursor_del(&mc, 0);
2144 env->me_pgfirst = 0;
2148 /* save to free list */
2150 freecnt = txn->mt_free_pgs[0];
2151 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2154 /* make sure last page of freeDB is touched and on freelist */
2155 key.mv_size = MAXKEYSIZE+1;
2157 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2162 MDB_IDL idl = txn->mt_free_pgs;
2163 mdb_midl_sort(txn->mt_free_pgs);
2164 DPRINTF("IDL write txn %zu root %zu num %zu",
2165 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2166 for (i=1; i<=idl[0]; i++) {
2167 DPRINTF("IDL %zu", idl[i]);
2171 /* write to last page of freeDB */
2172 key.mv_size = sizeof(pgno_t);
2173 key.mv_data = &txn->mt_txnid;
2174 data.mv_data = txn->mt_free_pgs;
2175 /* The free list can still grow during this call,
2176 * despite the pre-emptive touches above. So check
2177 * and make sure the entire thing got written.
2180 freecnt = txn->mt_free_pgs[0];
2181 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2182 mdb_midl_sort(txn->mt_free_pgs);
2183 rc = mdb_cursor_put(&mc, &key, &data, 0);
2188 } while (freecnt != txn->mt_free_pgs[0]);
2190 /* should only be one record now */
2192 if (env->me_pghead) {
2198 mop = env->me_pghead;
2200 key.mv_size = sizeof(id);
2202 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2203 data.mv_data = mop->mo_pages;
2204 orig = mop->mo_pages[0];
2205 /* These steps may grow the freelist again
2206 * due to freed overflow pages...
2208 mdb_cursor_put(&mc, &key, &data, 0);
2209 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2210 /* could have been used again here */
2211 if (mop->mo_pages[0] != orig) {
2212 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2213 data.mv_data = mop->mo_pages;
2215 mdb_cursor_put(&mc, &key, &data, 0);
2218 /* was completely used up */
2219 mdb_cursor_del(&mc, 0);
2223 env->me_pgfirst = 0;
2227 /* Check for growth of freelist again */
2228 if (freecnt != txn->mt_free_pgs[0])
2231 if (env->me_pghead) {
2232 free(env->me_pghead);
2233 env->me_pghead = NULL;
2236 while (env->me_pgfree) {
2237 MDB_oldpages *mop = env->me_pgfree;
2238 env->me_pgfree = mop->mo_next;
2242 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2243 if (mdb_midl_shrink(&txn->mt_free_pgs))
2244 env->me_free_pgs = txn->mt_free_pgs;
2251 if (env->me_flags & MDB_WRITEMAP) {
2252 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2253 dp = txn->mt_u.dirty_list[i].mptr;
2254 /* clear dirty flag */
2255 dp->mp_flags &= ~P_DIRTY;
2256 txn->mt_u.dirty_list[i].mid = 0;
2258 txn->mt_u.dirty_list[0].mid = 0;
2262 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2268 /* Windows actually supports scatter/gather I/O, but only on
2269 * unbuffered file handles. Since we're relying on the OS page
2270 * cache for all our data, that's self-defeating. So we just
2271 * write pages one at a time. We use the ov structure to set
2272 * the write offset, to at least save the overhead of a Seek
2276 memset(&ov, 0, sizeof(ov));
2277 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2279 dp = txn->mt_u.dirty_list[i].mptr;
2280 DPRINTF("committing page %zu", dp->mp_pgno);
2281 size = dp->mp_pgno * env->me_psize;
2282 ov.Offset = size & 0xffffffff;
2283 ov.OffsetHigh = size >> 16;
2284 ov.OffsetHigh >>= 16;
2285 /* clear dirty flag */
2286 dp->mp_flags &= ~P_DIRTY;
2287 wsize = env->me_psize;
2288 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2289 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2292 DPRINTF("WriteFile: %d", n);
2299 struct iovec iov[MDB_COMMIT_PAGES];
2303 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2304 dp = txn->mt_u.dirty_list[i].mptr;
2305 if (dp->mp_pgno != next) {
2307 rc = writev(env->me_fd, iov, n);
2311 DPUTS("short write, filesystem full?");
2313 DPRINTF("writev: %s", strerror(n));
2320 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2323 DPRINTF("committing page %zu", dp->mp_pgno);
2324 iov[n].iov_len = env->me_psize;
2325 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2326 iov[n].iov_base = (char *)dp;
2327 size += iov[n].iov_len;
2328 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2329 /* clear dirty flag */
2330 dp->mp_flags &= ~P_DIRTY;
2331 if (++n >= MDB_COMMIT_PAGES) {
2341 rc = writev(env->me_fd, iov, n);
2345 DPUTS("short write, filesystem full?");
2347 DPRINTF("writev: %s", strerror(n));
2354 /* Drop the dirty pages.
2356 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2357 dp = txn->mt_u.dirty_list[i].mptr;
2358 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2359 dp->mp_next = txn->mt_env->me_dpages;
2360 VGMEMP_FREE(txn->mt_env, dp);
2361 txn->mt_env->me_dpages = dp;
2363 VGMEMP_FREE(txn->mt_env, dp);
2366 txn->mt_u.dirty_list[i].mid = 0;
2368 txn->mt_u.dirty_list[0].mid = 0;
2371 if ((n = mdb_env_sync(env, 0)) != 0 ||
2372 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2379 if (txn->mt_numdbs > env->me_numdbs) {
2380 /* update the DB flags */
2382 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2383 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2387 UNLOCK_MUTEX_W(env);
2393 /** Read the environment parameters of a DB environment before
2394 * mapping it into memory.
2395 * @param[in] env the environment handle
2396 * @param[out] meta address of where to store the meta information
2397 * @return 0 on success, non-zero on failure.
2400 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2407 /* We don't know the page size yet, so use a minimum value.
2411 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2413 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2418 else if (rc != MDB_PAGESIZE) {
2422 DPRINTF("read: %s", strerror(err));
2426 p = (MDB_page *)&pbuf;
2428 if (!F_ISSET(p->mp_flags, P_META)) {
2429 DPRINTF("page %zu not a meta page", p->mp_pgno);
2434 if (m->mm_magic != MDB_MAGIC) {
2435 DPUTS("meta has invalid magic");
2439 if (m->mm_version != MDB_VERSION) {
2440 DPRINTF("database is version %u, expected version %u",
2441 m->mm_version, MDB_VERSION);
2442 return MDB_VERSION_MISMATCH;
2445 memcpy(meta, m, sizeof(*m));
2449 /** Write the environment parameters of a freshly created DB environment.
2450 * @param[in] env the environment handle
2451 * @param[out] meta address of where to store the meta information
2452 * @return 0 on success, non-zero on failure.
2455 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2462 DPUTS("writing new meta page");
2464 GET_PAGESIZE(psize);
2466 meta->mm_magic = MDB_MAGIC;
2467 meta->mm_version = MDB_VERSION;
2468 meta->mm_psize = psize;
2469 meta->mm_last_pg = 1;
2470 meta->mm_flags = env->me_flags & 0xffff;
2471 meta->mm_flags |= MDB_INTEGERKEY;
2472 meta->mm_dbs[0].md_root = P_INVALID;
2473 meta->mm_dbs[1].md_root = P_INVALID;
2475 p = calloc(2, psize);
2477 p->mp_flags = P_META;
2480 memcpy(m, meta, sizeof(*meta));
2482 q = (MDB_page *)((char *)p + psize);
2485 q->mp_flags = P_META;
2488 memcpy(m, meta, sizeof(*meta));
2493 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2494 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2497 rc = write(env->me_fd, p, psize * 2);
2498 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2504 /** Update the environment info to commit a transaction.
2505 * @param[in] txn the transaction that's being committed
2506 * @return 0 on success, non-zero on failure.
2509 mdb_env_write_meta(MDB_txn *txn)
2512 MDB_meta meta, metab, *mp;
2514 int rc, len, toggle;
2521 assert(txn != NULL);
2522 assert(txn->mt_env != NULL);
2524 toggle = !txn->mt_toggle;
2525 DPRINTF("writing meta page %d for root page %zu",
2526 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2529 mp = env->me_metas[toggle];
2531 if (env->me_flags & MDB_WRITEMAP) {
2532 /* Persist any increases of mapsize config */
2533 if (env->me_mapsize > mp->mm_mapsize)
2534 mp->mm_mapsize = env->me_mapsize;
2535 mp->mm_dbs[0] = txn->mt_dbs[0];
2536 mp->mm_dbs[1] = txn->mt_dbs[1];
2537 mp->mm_last_pg = txn->mt_next_pgno - 1;
2538 mp->mm_txnid = txn->mt_txnid;
2539 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2540 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2543 ptr += env->me_psize;
2544 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2551 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2552 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2554 ptr = (char *)&meta;
2555 if (env->me_mapsize > mp->mm_mapsize) {
2556 /* Persist any increases of mapsize config */
2557 meta.mm_mapsize = env->me_mapsize;
2558 off = offsetof(MDB_meta, mm_mapsize);
2560 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2562 len = sizeof(MDB_meta) - off;
2565 meta.mm_dbs[0] = txn->mt_dbs[0];
2566 meta.mm_dbs[1] = txn->mt_dbs[1];
2567 meta.mm_last_pg = txn->mt_next_pgno - 1;
2568 meta.mm_txnid = txn->mt_txnid;
2571 off += env->me_psize;
2574 /* Write to the SYNC fd */
2575 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
2576 env->me_fd : env->me_mfd;
2579 memset(&ov, 0, sizeof(ov));
2581 WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov);
2584 rc = pwrite(mfd, ptr, len, off);
2589 DPUTS("write failed, disk error?");
2590 /* On a failure, the pagecache still contains the new data.
2591 * Write some old data back, to prevent it from being used.
2592 * Use the non-SYNC fd; we know it will fail anyway.
2594 meta.mm_last_pg = metab.mm_last_pg;
2595 meta.mm_txnid = metab.mm_txnid;
2597 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2599 r2 = pwrite(env->me_fd, ptr, len, off);
2602 env->me_flags |= MDB_FATAL_ERROR;
2606 /* Memory ordering issues are irrelevant; since the entire writer
2607 * is wrapped by wmutex, all of these changes will become visible
2608 * after the wmutex is unlocked. Since the DB is multi-version,
2609 * readers will get consistent data regardless of how fresh or
2610 * how stale their view of these values is.
2612 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2617 /** Check both meta pages to see which one is newer.
2618 * @param[in] env the environment handle
2619 * @return meta toggle (0 or 1).
2622 mdb_env_pick_meta(const MDB_env *env)
2624 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2628 mdb_env_create(MDB_env **env)
2632 e = calloc(1, sizeof(MDB_env));
2636 e->me_free_pgs = mdb_midl_alloc();
2637 if (!e->me_free_pgs) {
2641 e->me_maxreaders = DEFAULT_READERS;
2643 e->me_fd = INVALID_HANDLE_VALUE;
2644 e->me_lfd = INVALID_HANDLE_VALUE;
2645 e->me_mfd = INVALID_HANDLE_VALUE;
2646 #ifdef MDB_USE_POSIX_SEM
2647 e->me_rmutex = SEM_FAILED;
2648 e->me_wmutex = SEM_FAILED;
2650 e->me_pid = getpid();
2651 VGMEMP_CREATE(e,0,0);
2657 mdb_env_set_mapsize(MDB_env *env, size_t size)
2661 env->me_mapsize = size;
2663 env->me_maxpg = env->me_mapsize / env->me_psize;
2668 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2672 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
2677 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2679 if (env->me_map || readers < 1)
2681 env->me_maxreaders = readers;
2686 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2688 if (!env || !readers)
2690 *readers = env->me_maxreaders;
2694 /** Further setup required for opening an MDB environment
2697 mdb_env_open2(MDB_env *env)
2699 unsigned int flags = env->me_flags;
2700 int i, newenv = 0, prot;
2704 memset(&meta, 0, sizeof(meta));
2706 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2709 DPUTS("new mdbenv");
2711 meta.mm_mapsize = env->me_mapsize > DEFAULT_MAPSIZE ? env->me_mapsize : DEFAULT_MAPSIZE;
2714 if (env->me_mapsize < meta.mm_mapsize)
2715 env->me_mapsize = meta.mm_mapsize;
2720 LONG sizelo, sizehi;
2721 sizelo = env->me_mapsize & 0xffffffff;
2722 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2724 /* Windows won't create mappings for zero length files.
2725 * Just allocate the maxsize right now.
2728 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2729 if (!SetEndOfFile(env->me_fd))
2731 SetFilePointer(env->me_fd, 0, NULL, 0);
2733 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2734 PAGE_READWRITE : PAGE_READONLY,
2735 sizehi, sizelo, NULL);
2738 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2739 FILE_MAP_WRITE : FILE_MAP_READ,
2740 0, 0, env->me_mapsize, meta.mm_address);
2748 if (flags & MDB_WRITEMAP) {
2750 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
2753 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2755 if (env->me_map == MAP_FAILED) {
2762 if (flags & MDB_FIXEDMAP)
2763 meta.mm_address = env->me_map;
2764 i = mdb_env_init_meta(env, &meta);
2765 if (i != MDB_SUCCESS) {
2768 } else if (meta.mm_address && env->me_map != meta.mm_address) {
2769 /* Can happen because the address argument to mmap() is just a
2770 * hint. mmap() can pick another, e.g. if the range is in use.
2771 * The MAP_FIXED flag would prevent that, but then mmap could
2772 * instead unmap existing pages to make room for the new map.
2774 return EBUSY; /* TODO: Make a new MDB_* error code? */
2776 env->me_psize = meta.mm_psize;
2778 env->me_maxpg = env->me_mapsize / env->me_psize;
2780 p = (MDB_page *)env->me_map;
2781 env->me_metas[0] = METADATA(p);
2782 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2786 int toggle = mdb_env_pick_meta(env);
2787 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2789 DPRINTF("opened database version %u, pagesize %u",
2790 env->me_metas[0]->mm_version, env->me_psize);
2791 DPRINTF("using meta page %d", toggle);
2792 DPRINTF("depth: %u", db->md_depth);
2793 DPRINTF("entries: %zu", db->md_entries);
2794 DPRINTF("branch pages: %zu", db->md_branch_pages);
2795 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2796 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2797 DPRINTF("root: %zu", db->md_root);
2805 /** Release a reader thread's slot in the reader lock table.
2806 * This function is called automatically when a thread exits.
2807 * @param[in] ptr This points to the slot in the reader lock table.
2810 mdb_env_reader_dest(void *ptr)
2812 MDB_reader *reader = ptr;
2818 /** Junk for arranging thread-specific callbacks on Windows. This is
2819 * necessarily platform and compiler-specific. Windows supports up
2820 * to 1088 keys. Let's assume nobody opens more than 64 environments
2821 * in a single process, for now. They can override this if needed.
2823 #ifndef MAX_TLS_KEYS
2824 #define MAX_TLS_KEYS 64
2826 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2827 static int mdb_tls_nkeys;
2829 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2833 case DLL_PROCESS_ATTACH: break;
2834 case DLL_THREAD_ATTACH: break;
2835 case DLL_THREAD_DETACH:
2836 for (i=0; i<mdb_tls_nkeys; i++) {
2837 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2838 mdb_env_reader_dest(r);
2841 case DLL_PROCESS_DETACH: break;
2846 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2848 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2852 /* Force some symbol references.
2853 * _tls_used forces the linker to create the TLS directory if not already done
2854 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2856 #pragma comment(linker, "/INCLUDE:_tls_used")
2857 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2858 #pragma const_seg(".CRT$XLB")
2859 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2860 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2863 #pragma comment(linker, "/INCLUDE:__tls_used")
2864 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2865 #pragma data_seg(".CRT$XLB")
2866 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2868 #endif /* WIN 32/64 */
2869 #endif /* !__GNUC__ */
2872 /** Downgrade the exclusive lock on the region back to shared */
2874 mdb_env_share_locks(MDB_env *env, int *excl)
2876 int rc = 0, toggle = mdb_env_pick_meta(env);
2878 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2883 /* First acquire a shared lock. The Unlock will
2884 * then release the existing exclusive lock.
2886 memset(&ov, 0, sizeof(ov));
2887 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2890 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2896 struct flock lock_info;
2897 /* The shared lock replaces the existing lock */
2898 memset((void *)&lock_info, 0, sizeof(lock_info));
2899 lock_info.l_type = F_RDLCK;
2900 lock_info.l_whence = SEEK_SET;
2901 lock_info.l_start = 0;
2902 lock_info.l_len = 1;
2903 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2904 (rc = ErrCode()) == EINTR) ;
2905 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
2912 /** Try to get exlusive lock, otherwise shared.
2913 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
2916 mdb_env_excl_lock(MDB_env *env, int *excl)
2920 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2924 memset(&ov, 0, sizeof(ov));
2925 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2932 struct flock lock_info;
2933 memset((void *)&lock_info, 0, sizeof(lock_info));
2934 lock_info.l_type = F_WRLCK;
2935 lock_info.l_whence = SEEK_SET;
2936 lock_info.l_start = 0;
2937 lock_info.l_len = 1;
2938 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
2939 (rc = ErrCode()) == EINTR) ;
2943 # ifdef MDB_USE_POSIX_SEM
2944 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
2947 lock_info.l_type = F_RDLCK;
2948 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
2949 (rc = ErrCode()) == EINTR) ;
2957 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
2959 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2961 * @(#) $Revision: 5.1 $
2962 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2963 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2965 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2969 * Please do not copyright this code. This code is in the public domain.
2971 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2972 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2973 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2974 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2975 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2976 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2977 * PERFORMANCE OF THIS SOFTWARE.
2980 * chongo <Landon Curt Noll> /\oo/\
2981 * http://www.isthe.com/chongo/
2983 * Share and Enjoy! :-)
2986 typedef unsigned long long mdb_hash_t;
2987 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2989 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2990 * @param[in] str string to hash
2991 * @param[in] hval initial value for hash
2992 * @return 64 bit hash
2994 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2995 * hval arg on the first call.
2998 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3000 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3001 unsigned char *end = s + val->mv_size;
3003 * FNV-1a hash each octet of the string
3006 /* xor the bottom with the current octet */
3007 hval ^= (mdb_hash_t)*s++;
3009 /* multiply by the 64 bit FNV magic prime mod 2^64 */
3010 hval += (hval << 1) + (hval << 4) + (hval << 5) +
3011 (hval << 7) + (hval << 8) + (hval << 40);
3013 /* return our new hash value */
3017 /** Hash the string and output the hash in hex.
3018 * @param[in] str string to hash
3019 * @param[out] hexbuf an array of 17 chars to hold the hash
3022 mdb_hash_hex(MDB_val *val, char *hexbuf)
3025 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
3026 for (i=0; i<8; i++) {
3027 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
3033 /** Open and/or initialize the lock region for the environment.
3034 * @param[in] env The MDB environment.
3035 * @param[in] lpath The pathname of the file used for the lock region.
3036 * @param[in] mode The Unix permissions for the file, if we create it.
3037 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
3038 * @return 0 on success, non-zero on failure.
3041 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
3049 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
3050 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
3051 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
3053 if (rc == ERROR_WRITE_PROTECT && (env->me_flags & MDB_RDONLY)) {
3054 env->me_flags |= MDB_ROFS;
3059 /* Try to get exclusive lock. If we succeed, then
3060 * nobody is using the lock region and we should initialize it.
3062 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3063 size = GetFileSize(env->me_lfd, NULL);
3069 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
3071 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3072 env->me_flags |= MDB_ROFS;
3077 /* Lose record locks when exec*() */
3078 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
3079 fcntl(env->me_lfd, F_SETFD, fdflags);
3081 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
3082 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1) {
3084 if (rc == EROFS && (env->me_flags & MDB_RDONLY)) {
3085 env->me_flags |= MDB_ROFS;
3092 /* Try to get exclusive lock. If we succeed, then
3093 * nobody is using the lock region and we should initialize it.
3095 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
3097 size = lseek(env->me_lfd, 0, SEEK_END);
3099 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
3100 if (size < rsize && *excl > 0) {
3102 SetFilePointer(env->me_lfd, rsize, NULL, 0);
3103 if (!SetEndOfFile(env->me_lfd)) goto fail_errno;
3105 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
3109 size = rsize - sizeof(MDB_txninfo);
3110 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
3115 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
3117 if (!mh) goto fail_errno;
3118 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
3120 if (!env->me_txns) goto fail_errno;
3122 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
3124 if (m == MAP_FAILED) goto fail_errno;
3130 BY_HANDLE_FILE_INFORMATION stbuf;
3139 if (!mdb_sec_inited) {
3140 InitializeSecurityDescriptor(&mdb_null_sd,
3141 SECURITY_DESCRIPTOR_REVISION);
3142 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
3143 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
3144 mdb_all_sa.bInheritHandle = FALSE;
3145 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
3148 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
3149 idbuf.volume = stbuf.dwVolumeSerialNumber;
3150 idbuf.nhigh = stbuf.nFileIndexHigh;
3151 idbuf.nlow = stbuf.nFileIndexLow;
3152 val.mv_data = &idbuf;
3153 val.mv_size = sizeof(idbuf);
3154 mdb_hash_hex(&val, hexbuf);
3155 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
3156 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
3157 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
3158 if (!env->me_rmutex) goto fail_errno;
3159 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
3160 if (!env->me_wmutex) goto fail_errno;
3161 #elif defined(MDB_USE_POSIX_SEM)
3170 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
3171 idbuf.dev = stbuf.st_dev;
3172 idbuf.ino = stbuf.st_ino;
3173 val.mv_data = &idbuf;
3174 val.mv_size = sizeof(idbuf);
3175 mdb_hash_hex(&val, hexbuf);
3176 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3177 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3178 /* Clean up after a previous run, if needed: Try to
3179 * remove both semaphores before doing anything else.
3181 sem_unlink(env->me_txns->mti_rmname);
3182 sem_unlink(env->me_txns->mti_wmname);
3183 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
3184 O_CREAT|O_EXCL, mode, 1);
3185 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3186 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
3187 O_CREAT|O_EXCL, mode, 1);
3188 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3189 #else /* MDB_USE_POSIX_SEM */
3190 pthread_mutexattr_t mattr;
3192 if ((rc = pthread_mutexattr_init(&mattr))
3193 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
3194 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
3195 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
3197 pthread_mutexattr_destroy(&mattr);
3198 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3200 env->me_txns->mti_version = MDB_VERSION;
3201 env->me_txns->mti_magic = MDB_MAGIC;
3202 env->me_txns->mti_txnid = 0;
3203 env->me_txns->mti_numreaders = 0;
3206 if (env->me_txns->mti_magic != MDB_MAGIC) {
3207 DPUTS("lock region has invalid magic");
3211 if (env->me_txns->mti_version != MDB_VERSION) {
3212 DPRINTF("lock region is version %u, expected version %u",
3213 env->me_txns->mti_version, MDB_VERSION);
3214 rc = MDB_VERSION_MISMATCH;
3218 if (rc != EACCES && rc != EAGAIN) {
3222 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3223 if (!env->me_rmutex) goto fail_errno;
3224 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3225 if (!env->me_wmutex) goto fail_errno;
3226 #elif defined(MDB_USE_POSIX_SEM)
3227 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3228 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
3229 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3230 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
3241 /** The name of the lock file in the DB environment */
3242 #define LOCKNAME "/lock.mdb"
3243 /** The name of the data file in the DB environment */
3244 #define DATANAME "/data.mdb"
3245 /** The suffix of the lock file when no subdir is used */
3246 #define LOCKSUFF "-lock"
3247 /** Only a subset of the @ref mdb_env flags can be changed
3248 * at runtime. Changing other flags requires closing the
3249 * environment and re-opening it with the new flags.
3251 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
3252 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP)
3255 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3257 int oflags, rc, len, excl;
3258 char *lpath, *dpath;
3260 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
3264 if (flags & MDB_NOSUBDIR) {
3265 rc = len + sizeof(LOCKSUFF) + len + 1;
3267 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3272 if (flags & MDB_NOSUBDIR) {
3273 dpath = lpath + len + sizeof(LOCKSUFF);
3274 sprintf(lpath, "%s" LOCKSUFF, path);
3275 strcpy(dpath, path);
3277 dpath = lpath + len + sizeof(LOCKNAME);
3278 sprintf(lpath, "%s" LOCKNAME, path);
3279 sprintf(dpath, "%s" DATANAME, path);
3282 flags |= env->me_flags;
3283 /* silently ignore WRITEMAP if we're only getting read access */
3284 if (F_ISSET(flags, MDB_RDONLY|MDB_WRITEMAP))
3285 flags ^= MDB_WRITEMAP;
3286 env->me_flags = flags |= MDB_ENV_ACTIVE;
3288 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3293 if (F_ISSET(flags, MDB_RDONLY)) {
3294 oflags = GENERIC_READ;
3295 len = OPEN_EXISTING;
3297 oflags = GENERIC_READ|GENERIC_WRITE;
3300 mode = FILE_ATTRIBUTE_NORMAL;
3301 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3302 NULL, len, mode, NULL);
3304 if (F_ISSET(flags, MDB_RDONLY))
3307 oflags = O_RDWR | O_CREAT;
3309 env->me_fd = open(dpath, oflags, mode);
3311 if (env->me_fd == INVALID_HANDLE_VALUE) {
3316 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
3317 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
3318 env->me_mfd = env->me_fd;
3320 /* Synchronous fd for meta writes. Needed even with
3321 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
3324 env->me_mfd = CreateFile(dpath, oflags,
3325 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3326 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3328 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3330 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3335 DPRINTF("opened dbenv %p", (void *) env);
3336 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3339 env->me_numdbs = 2; /* this notes that me_txkey was set */
3341 /* Windows TLS callbacks need help finding their TLS info. */
3342 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3343 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3350 rc = mdb_env_share_locks(env, &excl);
3354 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3355 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3356 env->me_path = strdup(path);
3357 if (!env->me_dbxs || !env->me_dbflags || !env->me_path)
3363 mdb_env_close0(env, excl);
3369 /** Destroy resources from mdb_env_open() and clear our readers */
3371 mdb_env_close0(MDB_env *env, int excl)
3375 if (!(env->me_flags & MDB_ENV_ACTIVE))
3378 free(env->me_dbflags);
3382 if (env->me_numdbs) {
3383 pthread_key_delete(env->me_txkey);
3385 /* Delete our key from the global list */
3386 for (i=0; i<mdb_tls_nkeys; i++)
3387 if (mdb_tls_keys[i] == env->me_txkey) {
3388 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3396 munmap(env->me_map, env->me_mapsize);
3398 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
3400 if (env->me_fd != INVALID_HANDLE_VALUE)
3403 pid_t pid = env->me_pid;
3404 /* Clearing readers is done in this function because
3405 * me_txkey with its destructor must be disabled first.
3407 for (i = env->me_numreaders; --i >= 0; )
3408 if (env->me_txns->mti_readers[i].mr_pid == pid)
3409 env->me_txns->mti_readers[i].mr_pid = 0;
3411 if (env->me_rmutex) {
3412 CloseHandle(env->me_rmutex);
3413 if (env->me_wmutex) CloseHandle(env->me_wmutex);
3415 /* Windows automatically destroys the mutexes when
3416 * the last handle closes.
3418 #elif defined(MDB_USE_POSIX_SEM)
3419 if (env->me_rmutex != SEM_FAILED) {
3420 sem_close(env->me_rmutex);
3421 if (env->me_wmutex != SEM_FAILED)
3422 sem_close(env->me_wmutex);
3423 /* If we have the filelock: If we are the
3424 * only remaining user, clean up semaphores.
3427 mdb_env_excl_lock(env, &excl);
3429 sem_unlink(env->me_txns->mti_rmname);
3430 sem_unlink(env->me_txns->mti_wmname);
3434 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3436 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3439 /* Unlock the lockfile. Windows would have unlocked it
3440 * after closing anyway, but not necessarily at once.
3442 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3448 env->me_flags &= ~MDB_ENV_ACTIVE;
3452 mdb_env_copy(MDB_env *env, const char *path)
3454 MDB_txn *txn = NULL;
3458 HANDLE newfd = INVALID_HANDLE_VALUE;
3460 if (env->me_flags & MDB_NOSUBDIR) {
3461 lpath = (char *)path;
3464 len += sizeof(DATANAME);
3465 lpath = malloc(len);
3468 sprintf(lpath, "%s" DATANAME, path);
3471 /* The destination path must exist, but the destination file must not.
3472 * We don't want the OS to cache the writes, since the source data is
3473 * already in the OS cache.
3476 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
3477 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
3479 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL
3485 if (!(env->me_flags & MDB_NOSUBDIR))
3487 if (newfd == INVALID_HANDLE_VALUE) {
3492 #ifdef F_NOCACHE /* __APPLE__ */
3493 rc = fcntl(newfd, F_NOCACHE, 1);
3500 /* Do the lock/unlock of the reader mutex before starting the
3501 * write txn. Otherwise other read txns could block writers.
3503 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
3507 if (!(env->me_flags & MDB_ROFS)) {
3508 /* We must start the actual read txn after blocking writers */
3509 mdb_txn_reset0(txn);
3511 /* Temporarily block writers until we snapshot the meta pages */
3514 rc = mdb_txn_renew0(txn);
3516 UNLOCK_MUTEX_W(env);
3521 wsize = env->me_psize * 2;
3525 rc = WriteFile(newfd, env->me_map, wsize, &len, NULL);
3526 rc = (len == wsize) ? MDB_SUCCESS : ErrCode();
3529 rc = write(newfd, env->me_map, wsize);
3530 rc = (rc == (int)wsize) ? MDB_SUCCESS : ErrCode();
3532 if (! (env->me_flags & MDB_ROFS))
3533 UNLOCK_MUTEX_W(env);
3538 ptr = env->me_map + wsize;
3539 wsize = txn->mt_next_pgno * env->me_psize - wsize;
3540 #define MAX_WRITE 2147483648U
3544 if (wsize > MAX_WRITE)
3548 rc = WriteFile(newfd, ptr, w2, &len, NULL);
3549 rc = (len == w2) ? MDB_SUCCESS : ErrCode();
3558 if (wsize > MAX_WRITE)
3562 wres = write(newfd, ptr, w2);
3563 rc = (wres > 0) ? MDB_SUCCESS : ErrCode();
3572 if (newfd != INVALID_HANDLE_VALUE)
3579 mdb_env_close(MDB_env *env)
3586 VGMEMP_DESTROY(env);
3587 while ((dp = env->me_dpages) != NULL) {
3588 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3589 env->me_dpages = dp->mp_next;
3593 mdb_env_close0(env, 0);
3594 mdb_midl_free(env->me_free_pgs);
3598 /** Compare two items pointing at aligned size_t's */
3600 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3602 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3603 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3606 /** Compare two items pointing at aligned int's */
3608 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3610 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3611 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3614 /** Compare two items pointing at ints of unknown alignment.
3615 * Nodes and keys are guaranteed to be 2-byte aligned.
3618 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3620 #if BYTE_ORDER == LITTLE_ENDIAN
3621 unsigned short *u, *c;
3624 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3625 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3628 } while(!x && u > (unsigned short *)a->mv_data);
3631 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3635 /** Compare two items lexically */
3637 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3644 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3650 diff = memcmp(a->mv_data, b->mv_data, len);
3651 return diff ? diff : len_diff<0 ? -1 : len_diff;
3654 /** Compare two items in reverse byte order */
3656 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3658 const unsigned char *p1, *p2, *p1_lim;
3662 p1_lim = (const unsigned char *)a->mv_data;
3663 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3664 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3666 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3672 while (p1 > p1_lim) {
3673 diff = *--p1 - *--p2;
3677 return len_diff<0 ? -1 : len_diff;
3680 /** Search for key within a page, using binary search.
3681 * Returns the smallest entry larger or equal to the key.
3682 * If exactp is non-null, stores whether the found entry was an exact match
3683 * in *exactp (1 or 0).
3684 * Updates the cursor index with the index of the found entry.
3685 * If no entry larger or equal to the key is found, returns NULL.
3688 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3690 unsigned int i = 0, nkeys;
3693 MDB_page *mp = mc->mc_pg[mc->mc_top];
3694 MDB_node *node = NULL;
3699 nkeys = NUMKEYS(mp);
3704 COPY_PGNO(pgno, mp->mp_pgno);
3705 DPRINTF("searching %u keys in %s %spage %zu",
3706 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3713 low = IS_LEAF(mp) ? 0 : 1;
3715 cmp = mc->mc_dbx->md_cmp;
3717 /* Branch pages have no data, so if using integer keys,
3718 * alignment is guaranteed. Use faster mdb_cmp_int.
3720 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3721 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3728 nodekey.mv_size = mc->mc_db->md_pad;
3729 node = NODEPTR(mp, 0); /* fake */
3730 while (low <= high) {
3731 i = (low + high) >> 1;
3732 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3733 rc = cmp(key, &nodekey);
3734 DPRINTF("found leaf index %u [%s], rc = %i",
3735 i, DKEY(&nodekey), rc);
3744 while (low <= high) {
3745 i = (low + high) >> 1;
3747 node = NODEPTR(mp, i);
3748 nodekey.mv_size = NODEKSZ(node);
3749 nodekey.mv_data = NODEKEY(node);
3751 rc = cmp(key, &nodekey);
3754 DPRINTF("found leaf index %u [%s], rc = %i",
3755 i, DKEY(&nodekey), rc);
3757 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3758 i, DKEY(&nodekey), NODEPGNO(node), rc);
3769 if (rc > 0) { /* Found entry is less than the key. */
3770 i++; /* Skip to get the smallest entry larger than key. */
3772 node = NODEPTR(mp, i);
3775 *exactp = (rc == 0);
3776 /* store the key index */
3777 mc->mc_ki[mc->mc_top] = i;
3779 /* There is no entry larger or equal to the key. */
3782 /* nodeptr is fake for LEAF2 */
3788 mdb_cursor_adjust(MDB_cursor *mc, func)
3792 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3793 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3800 /** Pop a page off the top of the cursor's stack. */
3802 mdb_cursor_pop(MDB_cursor *mc)
3805 #ifndef MDB_DEBUG_SKIP
3806 MDB_page *top = mc->mc_pg[mc->mc_top];
3812 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3813 mc->mc_dbi, (void *) mc);
3817 /** Push a page onto the top of the cursor's stack. */
3819 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3821 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3822 mc->mc_dbi, (void *) mc);
3824 if (mc->mc_snum >= CURSOR_STACK) {
3825 assert(mc->mc_snum < CURSOR_STACK);
3826 return MDB_CURSOR_FULL;
3829 mc->mc_top = mc->mc_snum++;
3830 mc->mc_pg[mc->mc_top] = mp;
3831 mc->mc_ki[mc->mc_top] = 0;
3836 /** Find the address of the page corresponding to a given page number.
3837 * @param[in] txn the transaction for this access.
3838 * @param[in] pgno the page number for the page to retrieve.
3839 * @param[out] ret address of a pointer where the page's address will be stored.
3840 * @return 0 on success, non-zero on failure.
3843 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3847 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3848 if (pgno < txn->mt_next_pgno)
3849 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3852 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3854 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3855 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3856 p = txn->mt_u.dirty_list[x].mptr;
3860 if (pgno < txn->mt_next_pgno)
3861 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3866 DPRINTF("page %zu not found", pgno);
3869 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3872 /** Search for the page a given key should be in.
3873 * Pushes parent pages on the cursor stack. This function continues a
3874 * search on a cursor that has already been initialized. (Usually by
3875 * #mdb_page_search() but also by #mdb_node_move().)
3876 * @param[in,out] mc the cursor for this operation.
3877 * @param[in] key the key to search for. If NULL, search for the lowest
3878 * page. (This is used by #mdb_cursor_first().)
3879 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3880 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3881 * @return 0 on success, non-zero on failure.
3884 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3886 MDB_page *mp = mc->mc_pg[mc->mc_top];
3891 while (IS_BRANCH(mp)) {
3895 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3896 assert(NUMKEYS(mp) > 1);
3897 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3899 if (key == NULL) /* Initialize cursor to first page. */
3901 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3902 /* cursor to last page */
3906 node = mdb_node_search(mc, key, &exact);
3908 i = NUMKEYS(mp) - 1;
3910 i = mc->mc_ki[mc->mc_top];
3919 DPRINTF("following index %u for key [%s]",
3921 assert(i < NUMKEYS(mp));
3922 node = NODEPTR(mp, i);
3924 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3927 mc->mc_ki[mc->mc_top] = i;
3928 if ((rc = mdb_cursor_push(mc, mp)))
3932 if ((rc = mdb_page_touch(mc)) != 0)
3934 mp = mc->mc_pg[mc->mc_top];
3939 DPRINTF("internal error, index points to a %02X page!?",
3941 return MDB_CORRUPTED;
3944 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3945 key ? DKEY(key) : NULL);
3950 /** Search for the page a given key should be in.
3951 * Pushes parent pages on the cursor stack. This function just sets up
3952 * the search; it finds the root page for \b mc's database and sets this
3953 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3954 * called to complete the search.
3955 * @param[in,out] mc the cursor for this operation.
3956 * @param[in] key the key to search for. If NULL, search for the lowest
3957 * page. (This is used by #mdb_cursor_first().)
3958 * @param[in] modify If true, visited pages are updated with new page numbers.
3959 * @return 0 on success, non-zero on failure.
3962 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3967 /* Make sure the txn is still viable, then find the root from
3968 * the txn's db table.
3970 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3971 DPUTS("transaction has failed, must abort");
3974 /* Make sure we're using an up-to-date root */
3975 if (mc->mc_dbi > MAIN_DBI) {
3976 if ((*mc->mc_dbflag & DB_STALE) ||
3977 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3979 unsigned char dbflag = 0;
3980 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3981 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3984 if (*mc->mc_dbflag & DB_STALE) {
3987 MDB_node *leaf = mdb_node_search(&mc2,
3988 &mc->mc_dbx->md_name, &exact);
3990 return MDB_NOTFOUND;
3991 mdb_node_read(mc->mc_txn, leaf, &data);
3992 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3994 if (flags & MDB_PS_MODIFY)
3996 *mc->mc_dbflag = dbflag;
3999 root = mc->mc_db->md_root;
4001 if (root == P_INVALID) { /* Tree is empty. */
4002 DPUTS("tree is empty");
4003 return MDB_NOTFOUND;
4008 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4009 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
4015 DPRINTF("db %u root page %zu has flags 0x%X",
4016 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
4018 if (flags & MDB_PS_MODIFY) {
4019 if ((rc = mdb_page_touch(mc)))
4023 if (flags & MDB_PS_ROOTONLY)
4026 return mdb_page_search_root(mc, key, flags);
4029 /** Return the data associated with a given node.
4030 * @param[in] txn The transaction for this operation.
4031 * @param[in] leaf The node being read.
4032 * @param[out] data Updated to point to the node's data.
4033 * @return 0 on success, non-zero on failure.
4036 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
4038 MDB_page *omp; /* overflow page */
4042 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4043 data->mv_size = NODEDSZ(leaf);
4044 data->mv_data = NODEDATA(leaf);
4048 /* Read overflow data.
4050 data->mv_size = NODEDSZ(leaf);
4051 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
4052 if ((rc = mdb_page_get(txn, pgno, &omp))) {
4053 DPRINTF("read overflow page %zu failed", pgno);
4056 data->mv_data = METADATA(omp);
4062 mdb_get(MDB_txn *txn, MDB_dbi dbi,
4063 MDB_val *key, MDB_val *data)
4072 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
4074 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
4077 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4081 mdb_cursor_init(&mc, txn, dbi, &mx);
4082 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
4085 /** Find a sibling for a page.
4086 * Replaces the page at the top of the cursor's stack with the
4087 * specified sibling, if one exists.
4088 * @param[in] mc The cursor for this operation.
4089 * @param[in] move_right Non-zero if the right sibling is requested,
4090 * otherwise the left sibling.
4091 * @return 0 on success, non-zero on failure.
4094 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
4100 if (mc->mc_snum < 2) {
4101 return MDB_NOTFOUND; /* root has no siblings */
4105 DPRINTF("parent page is page %zu, index %u",
4106 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
4108 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
4109 : (mc->mc_ki[mc->mc_top] == 0)) {
4110 DPRINTF("no more keys left, moving to %s sibling",
4111 move_right ? "right" : "left");
4112 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
4113 /* undo cursor_pop before returning */
4120 mc->mc_ki[mc->mc_top]++;
4122 mc->mc_ki[mc->mc_top]--;
4123 DPRINTF("just moving to %s index key %u",
4124 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
4126 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
4128 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4129 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
4132 mdb_cursor_push(mc, mp);
4137 /** Move the cursor to the next data item. */
4139 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4145 if (mc->mc_flags & C_EOF) {
4146 return MDB_NOTFOUND;
4149 assert(mc->mc_flags & C_INITIALIZED);
4151 mp = mc->mc_pg[mc->mc_top];
4153 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4154 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4155 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4156 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
4157 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
4158 if (op != MDB_NEXT || rc == MDB_SUCCESS)
4162 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4163 if (op == MDB_NEXT_DUP)
4164 return MDB_NOTFOUND;
4168 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4170 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
4171 DPUTS("=====> move to next sibling page");
4172 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
4173 mc->mc_flags |= C_EOF;
4174 mc->mc_flags &= ~C_INITIALIZED;
4175 return MDB_NOTFOUND;
4177 mp = mc->mc_pg[mc->mc_top];
4178 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4180 mc->mc_ki[mc->mc_top]++;
4182 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4183 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4186 key->mv_size = mc->mc_db->md_pad;
4187 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4191 assert(IS_LEAF(mp));
4192 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4194 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4195 mdb_xcursor_init1(mc, leaf);
4198 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4201 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4202 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4203 if (rc != MDB_SUCCESS)
4208 MDB_GET_KEY(leaf, key);
4212 /** Move the cursor to the previous data item. */
4214 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
4220 assert(mc->mc_flags & C_INITIALIZED);
4222 mp = mc->mc_pg[mc->mc_top];
4224 if (mc->mc_db->md_flags & MDB_DUPSORT) {
4225 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4226 if (op == MDB_PREV || op == MDB_PREV_DUP) {
4227 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4228 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
4229 if (op != MDB_PREV || rc == MDB_SUCCESS)
4232 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4233 if (op == MDB_PREV_DUP)
4234 return MDB_NOTFOUND;
4239 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
4241 if (mc->mc_ki[mc->mc_top] == 0) {
4242 DPUTS("=====> move to prev sibling page");
4243 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
4244 mc->mc_flags &= ~C_INITIALIZED;
4245 return MDB_NOTFOUND;
4247 mp = mc->mc_pg[mc->mc_top];
4248 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
4249 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
4251 mc->mc_ki[mc->mc_top]--;
4253 mc->mc_flags &= ~C_EOF;
4255 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
4256 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
4259 key->mv_size = mc->mc_db->md_pad;
4260 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4264 assert(IS_LEAF(mp));
4265 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4267 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4268 mdb_xcursor_init1(mc, leaf);
4271 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
4274 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4275 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4276 if (rc != MDB_SUCCESS)
4281 MDB_GET_KEY(leaf, key);
4285 /** Set the cursor on a specific data item. */
4287 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4288 MDB_cursor_op op, int *exactp)
4292 MDB_node *leaf = NULL;
4297 assert(key->mv_size > 0);
4299 /* See if we're already on the right page */
4300 if (mc->mc_flags & C_INITIALIZED) {
4303 mp = mc->mc_pg[mc->mc_top];
4305 mc->mc_ki[mc->mc_top] = 0;
4306 return MDB_NOTFOUND;
4308 if (mp->mp_flags & P_LEAF2) {
4309 nodekey.mv_size = mc->mc_db->md_pad;
4310 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
4312 leaf = NODEPTR(mp, 0);
4313 MDB_GET_KEY(leaf, &nodekey);
4315 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4317 /* Probably happens rarely, but first node on the page
4318 * was the one we wanted.
4320 mc->mc_ki[mc->mc_top] = 0;
4327 unsigned int nkeys = NUMKEYS(mp);
4329 if (mp->mp_flags & P_LEAF2) {
4330 nodekey.mv_data = LEAF2KEY(mp,
4331 nkeys-1, nodekey.mv_size);
4333 leaf = NODEPTR(mp, nkeys-1);
4334 MDB_GET_KEY(leaf, &nodekey);
4336 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4338 /* last node was the one we wanted */
4339 mc->mc_ki[mc->mc_top] = nkeys-1;
4345 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4346 /* This is definitely the right page, skip search_page */
4347 if (mp->mp_flags & P_LEAF2) {
4348 nodekey.mv_data = LEAF2KEY(mp,
4349 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4351 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4352 MDB_GET_KEY(leaf, &nodekey);
4354 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4356 /* current node was the one we wanted */
4366 /* If any parents have right-sibs, search.
4367 * Otherwise, there's nothing further.
4369 for (i=0; i<mc->mc_top; i++)
4371 NUMKEYS(mc->mc_pg[i])-1)
4373 if (i == mc->mc_top) {
4374 /* There are no other pages */
4375 mc->mc_ki[mc->mc_top] = nkeys;
4376 return MDB_NOTFOUND;
4380 /* There are no other pages */
4381 mc->mc_ki[mc->mc_top] = 0;
4382 return MDB_NOTFOUND;
4386 rc = mdb_page_search(mc, key, 0);
4387 if (rc != MDB_SUCCESS)
4390 mp = mc->mc_pg[mc->mc_top];
4391 assert(IS_LEAF(mp));
4394 leaf = mdb_node_search(mc, key, exactp);
4395 if (exactp != NULL && !*exactp) {
4396 /* MDB_SET specified and not an exact match. */
4397 return MDB_NOTFOUND;
4401 DPUTS("===> inexact leaf not found, goto sibling");
4402 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4403 return rc; /* no entries matched */
4404 mp = mc->mc_pg[mc->mc_top];
4405 assert(IS_LEAF(mp));
4406 leaf = NODEPTR(mp, 0);
4410 mc->mc_flags |= C_INITIALIZED;
4411 mc->mc_flags &= ~C_EOF;
4414 key->mv_size = mc->mc_db->md_pad;
4415 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4419 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4420 mdb_xcursor_init1(mc, leaf);
4423 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4424 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4425 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4428 if (op == MDB_GET_BOTH) {
4434 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4435 if (rc != MDB_SUCCESS)
4438 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4440 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4442 rc = mc->mc_dbx->md_dcmp(data, &d2);
4444 if (op == MDB_GET_BOTH || rc > 0)
4445 return MDB_NOTFOUND;
4450 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4451 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4456 /* The key already matches in all other cases */
4457 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4458 MDB_GET_KEY(leaf, key);
4459 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4464 /** Move the cursor to the first item in the database. */
4466 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4471 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4472 rc = mdb_page_search(mc, NULL, 0);
4473 if (rc != MDB_SUCCESS)
4476 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4478 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4479 mc->mc_flags |= C_INITIALIZED;
4480 mc->mc_flags &= ~C_EOF;
4482 mc->mc_ki[mc->mc_top] = 0;
4484 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4485 key->mv_size = mc->mc_db->md_pad;
4486 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4491 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4492 mdb_xcursor_init1(mc, leaf);
4493 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4498 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4499 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4503 MDB_GET_KEY(leaf, key);
4507 /** Move the cursor to the last item in the database. */
4509 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4514 if (!(mc->mc_flags & C_EOF)) {
4516 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4519 lkey.mv_size = MAXKEYSIZE+1;
4520 lkey.mv_data = NULL;
4521 rc = mdb_page_search(mc, &lkey, 0);
4522 if (rc != MDB_SUCCESS)
4525 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4527 mc->mc_flags |= C_INITIALIZED|C_EOF;
4528 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4530 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4532 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4533 key->mv_size = mc->mc_db->md_pad;
4534 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4539 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4540 mdb_xcursor_init1(mc, leaf);
4541 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4546 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4547 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4552 MDB_GET_KEY(leaf, key);
4557 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4566 case MDB_GET_CURRENT:
4567 if (!mc->mc_flags & C_INITIALIZED) {
4570 MDB_page *mp = mc->mc_pg[mc->mc_top];
4572 mc->mc_ki[mc->mc_top] = 0;
4578 key->mv_size = mc->mc_db->md_pad;
4579 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4581 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4582 MDB_GET_KEY(leaf, key);
4584 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4585 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
4587 rc = mdb_node_read(mc->mc_txn, leaf, data);
4594 case MDB_GET_BOTH_RANGE:
4595 if (data == NULL || mc->mc_xcursor == NULL) {
4603 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4605 } else if (op == MDB_SET_RANGE)
4606 rc = mdb_cursor_set(mc, key, data, op, NULL);
4608 rc = mdb_cursor_set(mc, key, data, op, &exact);
4610 case MDB_GET_MULTIPLE:
4612 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4613 !(mc->mc_flags & C_INITIALIZED)) {
4618 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4619 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4622 case MDB_NEXT_MULTIPLE:
4624 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4628 if (!(mc->mc_flags & C_INITIALIZED))
4629 rc = mdb_cursor_first(mc, key, data);
4631 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4632 if (rc == MDB_SUCCESS) {
4633 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4636 mx = &mc->mc_xcursor->mx_cursor;
4637 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4639 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4640 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4648 case MDB_NEXT_NODUP:
4649 if (!(mc->mc_flags & C_INITIALIZED))
4650 rc = mdb_cursor_first(mc, key, data);
4652 rc = mdb_cursor_next(mc, key, data, op);
4656 case MDB_PREV_NODUP:
4657 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4658 rc = mdb_cursor_last(mc, key, data);
4659 mc->mc_flags |= C_INITIALIZED;
4660 mc->mc_ki[mc->mc_top]++;
4662 rc = mdb_cursor_prev(mc, key, data, op);
4665 rc = mdb_cursor_first(mc, key, data);
4669 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4670 !(mc->mc_flags & C_INITIALIZED) ||
4671 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4675 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4678 rc = mdb_cursor_last(mc, key, data);
4682 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4683 !(mc->mc_flags & C_INITIALIZED) ||
4684 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4688 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4691 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4699 /** Touch all the pages in the cursor stack.
4700 * Makes sure all the pages are writable, before attempting a write operation.
4701 * @param[in] mc The cursor to operate on.
4704 mdb_cursor_touch(MDB_cursor *mc)
4708 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4711 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI,
4712 mc->mc_txn->mt_dbs[MAIN_DBI].md_flags & MDB_DUPSORT ? &mcx : NULL);
4713 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4716 *mc->mc_dbflag = DB_DIRTY;
4718 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4719 rc = mdb_page_touch(mc);
4723 mc->mc_top = mc->mc_snum-1;
4728 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4731 MDB_node *leaf = NULL;
4732 MDB_val xdata, *rdata, dkey;
4735 int do_sub = 0, insert = 0;
4736 unsigned int mcount = 0;
4740 char dbuf[MAXKEYSIZE+1];
4741 unsigned int nflags;
4744 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4747 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4748 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4752 if (flags == MDB_CURRENT) {
4753 if (!(mc->mc_flags & C_INITIALIZED))
4756 } else if (mc->mc_db->md_root == P_INVALID) {
4758 /* new database, write a root leaf page */
4759 DPUTS("allocating new root leaf page");
4760 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4764 mdb_cursor_push(mc, np);
4765 mc->mc_db->md_root = np->mp_pgno;
4766 mc->mc_db->md_depth++;
4767 *mc->mc_dbflag = DB_DIRTY;
4768 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4770 np->mp_flags |= P_LEAF2;
4771 mc->mc_flags |= C_INITIALIZED;
4777 if (flags & MDB_APPEND) {
4779 rc = mdb_cursor_last(mc, &k2, &d2);
4781 rc = mc->mc_dbx->md_cmp(key, &k2);
4784 mc->mc_ki[mc->mc_top]++;
4790 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4792 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4793 DPRINTF("duplicate key [%s]", DKEY(key));
4795 return MDB_KEYEXIST;
4797 if (rc && rc != MDB_NOTFOUND)
4801 /* Cursor is positioned, now make sure all pages are writable */
4802 rc2 = mdb_cursor_touch(mc);
4807 /* The key already exists */
4808 if (rc == MDB_SUCCESS) {
4809 /* there's only a key anyway, so this is a no-op */
4810 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4811 unsigned int ksize = mc->mc_db->md_pad;
4812 if (key->mv_size != ksize)
4814 if (flags == MDB_CURRENT) {
4815 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4816 memcpy(ptr, key->mv_data, ksize);
4821 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4824 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4825 /* Was a single item before, must convert now */
4827 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4828 /* Just overwrite the current item */
4829 if (flags == MDB_CURRENT)
4832 dkey.mv_size = NODEDSZ(leaf);
4833 dkey.mv_data = NODEDATA(leaf);
4834 #if UINT_MAX < SIZE_MAX
4835 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4836 #ifdef MISALIGNED_OK
4837 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4839 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4842 /* if data matches, ignore it */
4843 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4844 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4846 /* create a fake page for the dup items */
4847 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4848 dkey.mv_data = dbuf;
4849 fp = (MDB_page *)&pbuf;
4850 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4851 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4852 fp->mp_lower = PAGEHDRSZ;
4853 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4854 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4855 fp->mp_flags |= P_LEAF2;
4856 fp->mp_pad = data->mv_size;
4857 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4859 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4860 (dkey.mv_size & 1) + (data->mv_size & 1);
4862 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4865 xdata.mv_size = fp->mp_upper;
4870 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4871 /* See if we need to convert from fake page to subDB */
4873 unsigned int offset;
4876 fp = NODEDATA(leaf);
4877 if (flags == MDB_CURRENT) {
4879 fp->mp_flags |= P_DIRTY;
4880 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4881 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4885 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4886 offset = fp->mp_pad;
4887 if (SIZELEFT(fp) >= offset)
4889 offset *= 4; /* space for 4 more */
4891 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4893 offset += offset & 1;
4894 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4895 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4897 /* yes, convert it */
4899 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4900 dummy.md_pad = fp->mp_pad;
4901 dummy.md_flags = MDB_DUPFIXED;
4902 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4903 dummy.md_flags |= MDB_INTEGERKEY;
4906 dummy.md_branch_pages = 0;
4907 dummy.md_leaf_pages = 1;
4908 dummy.md_overflow_pages = 0;
4909 dummy.md_entries = NUMKEYS(fp);
4911 xdata.mv_size = sizeof(MDB_db);
4912 xdata.mv_data = &dummy;
4913 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4915 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4916 flags |= F_DUPDATA|F_SUBDATA;
4917 dummy.md_root = mp->mp_pgno;
4919 /* no, just grow it */
4921 xdata.mv_size = NODEDSZ(leaf) + offset;
4922 xdata.mv_data = &pbuf;
4923 mp = (MDB_page *)&pbuf;
4924 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4927 mp->mp_flags = fp->mp_flags | P_DIRTY;
4928 mp->mp_pad = fp->mp_pad;
4929 mp->mp_lower = fp->mp_lower;
4930 mp->mp_upper = fp->mp_upper + offset;
4932 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4934 nsize = NODEDSZ(leaf) - fp->mp_upper;
4935 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4936 for (i=0; i<NUMKEYS(fp); i++)
4937 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4939 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4943 /* data is on sub-DB, just store it */
4944 flags |= F_DUPDATA|F_SUBDATA;
4948 /* overflow page overwrites need special handling */
4949 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4952 int ovpages, dpages;
4954 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4955 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4956 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4957 mdb_page_get(mc->mc_txn, pg, &omp);
4958 /* Is the ov page writable and large enough? */
4959 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4960 /* yes, overwrite it. Note in this case we don't
4961 * bother to try shrinking the node if the new data
4962 * is smaller than the overflow threshold.
4964 if (F_ISSET(flags, MDB_RESERVE))
4965 data->mv_data = METADATA(omp);
4967 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4970 /* no, free ovpages */
4972 mc->mc_db->md_overflow_pages -= ovpages;
4973 for (i=0; i<ovpages; i++) {
4974 DPRINTF("freed ov page %zu", pg);
4975 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4979 } else if (NODEDSZ(leaf) == data->mv_size) {
4980 /* same size, just replace it. Note that we could
4981 * also reuse this node if the new data is smaller,
4982 * but instead we opt to shrink the node in that case.
4984 if (F_ISSET(flags, MDB_RESERVE))
4985 data->mv_data = NODEDATA(leaf);
4987 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4990 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4991 mc->mc_db->md_entries--;
4993 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
5000 nflags = flags & NODE_ADD_FLAGS;
5001 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
5002 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
5003 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
5004 nflags &= ~MDB_APPEND;
5006 nflags |= MDB_SPLIT_REPLACE;
5007 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
5009 /* There is room already in this leaf page. */
5010 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
5011 if (rc == 0 && !do_sub && insert) {
5012 /* Adjust other cursors pointing to mp */
5013 MDB_cursor *m2, *m3;
5014 MDB_dbi dbi = mc->mc_dbi;
5015 unsigned i = mc->mc_top;
5016 MDB_page *mp = mc->mc_pg[i];
5018 if (mc->mc_flags & C_SUB)
5021 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5022 if (mc->mc_flags & C_SUB)
5023 m3 = &m2->mc_xcursor->mx_cursor;
5026 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
5027 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
5034 if (rc != MDB_SUCCESS)
5035 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5037 /* Now store the actual data in the child DB. Note that we're
5038 * storing the user data in the keys field, so there are strict
5039 * size limits on dupdata. The actual data fields of the child
5040 * DB are all zero size.
5047 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5048 if (flags & MDB_CURRENT) {
5049 xflags = MDB_CURRENT;
5051 mdb_xcursor_init1(mc, leaf);
5052 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
5054 /* converted, write the original data first */
5056 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
5060 /* Adjust other cursors pointing to mp */
5062 unsigned i = mc->mc_top;
5063 MDB_page *mp = mc->mc_pg[i];
5065 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5066 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
5067 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
5068 mdb_xcursor_init1(m2, leaf);
5073 if (flags & MDB_APPENDDUP)
5074 xflags |= MDB_APPEND;
5075 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
5076 if (flags & F_SUBDATA) {
5077 void *db = NODEDATA(leaf);
5078 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5081 /* sub-writes might have failed so check rc again.
5082 * Don't increment count if we just replaced an existing item.
5084 if (!rc && !(flags & MDB_CURRENT))
5085 mc->mc_db->md_entries++;
5086 if (flags & MDB_MULTIPLE) {
5088 if (mcount < data[1].mv_size) {
5089 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
5090 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5100 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
5105 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
5108 if (!mc->mc_flags & C_INITIALIZED)
5111 rc = mdb_cursor_touch(mc);
5115 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5117 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5118 if (flags != MDB_NODUPDATA) {
5119 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
5120 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
5122 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
5123 /* If sub-DB still has entries, we're done */
5124 if (mc->mc_xcursor->mx_db.md_entries) {
5125 if (leaf->mn_flags & F_SUBDATA) {
5126 /* update subDB info */
5127 void *db = NODEDATA(leaf);
5128 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
5130 /* shrink fake page */
5131 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5133 mc->mc_db->md_entries--;
5136 /* otherwise fall thru and delete the sub-DB */
5139 if (leaf->mn_flags & F_SUBDATA) {
5140 /* add all the child DB's pages to the free list */
5141 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
5142 if (rc == MDB_SUCCESS) {
5143 mc->mc_db->md_entries -=
5144 mc->mc_xcursor->mx_db.md_entries;
5149 return mdb_cursor_del0(mc, leaf);
5152 /** Allocate and initialize new pages for a database.
5153 * @param[in] mc a cursor on the database being added to.
5154 * @param[in] flags flags defining what type of page is being allocated.
5155 * @param[in] num the number of pages to allocate. This is usually 1,
5156 * unless allocating overflow pages for a large record.
5157 * @param[out] mp Address of a page, or NULL on failure.
5158 * @return 0 on success, non-zero on failure.
5161 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
5166 if ((rc = mdb_page_alloc(mc, num, &np)))
5168 DPRINTF("allocated new mpage %zu, page size %u",
5169 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
5170 np->mp_flags = flags | P_DIRTY;
5171 np->mp_lower = PAGEHDRSZ;
5172 np->mp_upper = mc->mc_txn->mt_env->me_psize;
5175 mc->mc_db->md_branch_pages++;
5176 else if (IS_LEAF(np))
5177 mc->mc_db->md_leaf_pages++;
5178 else if (IS_OVERFLOW(np)) {
5179 mc->mc_db->md_overflow_pages += num;
5187 /** Calculate the size of a leaf node.
5188 * The size depends on the environment's page size; if a data item
5189 * is too large it will be put onto an overflow page and the node
5190 * size will only include the key and not the data. Sizes are always
5191 * rounded up to an even number of bytes, to guarantee 2-byte alignment
5192 * of the #MDB_node headers.
5193 * @param[in] env The environment handle.
5194 * @param[in] key The key for the node.
5195 * @param[in] data The data for the node.
5196 * @return The number of bytes needed to store the node.
5199 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
5203 sz = LEAFSIZE(key, data);
5204 if (sz >= env->me_psize / MDB_MINKEYS) {
5205 /* put on overflow page */
5206 sz -= data->mv_size - sizeof(pgno_t);
5210 return sz + sizeof(indx_t);
5213 /** Calculate the size of a branch node.
5214 * The size should depend on the environment's page size but since
5215 * we currently don't support spilling large keys onto overflow
5216 * pages, it's simply the size of the #MDB_node header plus the
5217 * size of the key. Sizes are always rounded up to an even number
5218 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
5219 * @param[in] env The environment handle.
5220 * @param[in] key The key for the node.
5221 * @return The number of bytes needed to store the node.
5224 mdb_branch_size(MDB_env *env, MDB_val *key)
5229 if (sz >= env->me_psize / MDB_MINKEYS) {
5230 /* put on overflow page */
5231 /* not implemented */
5232 /* sz -= key->size - sizeof(pgno_t); */
5235 return sz + sizeof(indx_t);
5238 /** Add a node to the page pointed to by the cursor.
5239 * @param[in] mc The cursor for this operation.
5240 * @param[in] indx The index on the page where the new node should be added.
5241 * @param[in] key The key for the new node.
5242 * @param[in] data The data for the new node, if any.
5243 * @param[in] pgno The page number, if adding a branch node.
5244 * @param[in] flags Flags for the node.
5245 * @return 0 on success, non-zero on failure. Possible errors are:
5247 * <li>ENOMEM - failed to allocate overflow pages for the node.
5248 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
5249 * should never happen since all callers already calculate the
5250 * page's free space before calling this function.
5254 mdb_node_add(MDB_cursor *mc, indx_t indx,
5255 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
5258 size_t node_size = NODESIZE;
5261 MDB_page *mp = mc->mc_pg[mc->mc_top];
5262 MDB_page *ofp = NULL; /* overflow page */
5265 assert(mp->mp_upper >= mp->mp_lower);
5267 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
5268 IS_LEAF(mp) ? "leaf" : "branch",
5269 IS_SUBP(mp) ? "sub-" : "",
5270 mp->mp_pgno, indx, data ? data->mv_size : 0,
5271 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
5274 /* Move higher keys up one slot. */
5275 int ksize = mc->mc_db->md_pad, dif;
5276 char *ptr = LEAF2KEY(mp, indx, ksize);
5277 dif = NUMKEYS(mp) - indx;
5279 memmove(ptr+ksize, ptr, dif*ksize);
5280 /* insert new key */
5281 memcpy(ptr, key->mv_data, ksize);
5283 /* Just using these for counting */
5284 mp->mp_lower += sizeof(indx_t);
5285 mp->mp_upper -= ksize - sizeof(indx_t);
5290 node_size += key->mv_size;
5294 if (F_ISSET(flags, F_BIGDATA)) {
5295 /* Data already on overflow page. */
5296 node_size += sizeof(pgno_t);
5297 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
5298 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
5300 /* Put data on overflow page. */
5301 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
5302 data->mv_size, node_size+data->mv_size);
5303 node_size += sizeof(pgno_t);
5304 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
5306 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
5309 node_size += data->mv_size;
5312 node_size += node_size & 1;
5314 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
5315 DPRINTF("not enough room in page %zu, got %u ptrs",
5316 mp->mp_pgno, NUMKEYS(mp));
5317 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
5318 mp->mp_upper - mp->mp_lower);
5319 DPRINTF("node size = %zu", node_size);
5320 return MDB_PAGE_FULL;
5323 /* Move higher pointers up one slot. */
5324 for (i = NUMKEYS(mp); i > indx; i--)
5325 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
5327 /* Adjust free space offsets. */
5328 ofs = mp->mp_upper - node_size;
5329 assert(ofs >= mp->mp_lower + sizeof(indx_t));
5330 mp->mp_ptrs[indx] = ofs;
5332 mp->mp_lower += sizeof(indx_t);
5334 /* Write the node data. */
5335 node = NODEPTR(mp, indx);
5336 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
5337 node->mn_flags = flags;
5339 SETDSZ(node,data->mv_size);
5344 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5349 if (F_ISSET(flags, F_BIGDATA))
5350 memcpy(node->mn_data + key->mv_size, data->mv_data,
5352 else if (F_ISSET(flags, MDB_RESERVE))
5353 data->mv_data = node->mn_data + key->mv_size;
5355 memcpy(node->mn_data + key->mv_size, data->mv_data,
5358 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5360 if (F_ISSET(flags, MDB_RESERVE))
5361 data->mv_data = METADATA(ofp);
5363 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5370 /** Delete the specified node from a page.
5371 * @param[in] mp The page to operate on.
5372 * @param[in] indx The index of the node to delete.
5373 * @param[in] ksize The size of a node. Only used if the page is
5374 * part of a #MDB_DUPFIXED database.
5377 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5380 indx_t i, j, numkeys, ptr;
5387 COPY_PGNO(pgno, mp->mp_pgno);
5388 DPRINTF("delete node %u on %s page %zu", indx,
5389 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5392 assert(indx < NUMKEYS(mp));
5395 int x = NUMKEYS(mp) - 1 - indx;
5396 base = LEAF2KEY(mp, indx, ksize);
5398 memmove(base, base + ksize, x * ksize);
5399 mp->mp_lower -= sizeof(indx_t);
5400 mp->mp_upper += ksize - sizeof(indx_t);
5404 node = NODEPTR(mp, indx);
5405 sz = NODESIZE + node->mn_ksize;
5407 if (F_ISSET(node->mn_flags, F_BIGDATA))
5408 sz += sizeof(pgno_t);
5410 sz += NODEDSZ(node);
5414 ptr = mp->mp_ptrs[indx];
5415 numkeys = NUMKEYS(mp);
5416 for (i = j = 0; i < numkeys; i++) {
5418 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5419 if (mp->mp_ptrs[i] < ptr)
5420 mp->mp_ptrs[j] += sz;
5425 base = (char *)mp + mp->mp_upper;
5426 memmove(base + sz, base, ptr - mp->mp_upper);
5428 mp->mp_lower -= sizeof(indx_t);
5432 /** Compact the main page after deleting a node on a subpage.
5433 * @param[in] mp The main page to operate on.
5434 * @param[in] indx The index of the subpage on the main page.
5437 mdb_node_shrink(MDB_page *mp, indx_t indx)
5444 indx_t i, numkeys, ptr;
5446 node = NODEPTR(mp, indx);
5447 sp = (MDB_page *)NODEDATA(node);
5448 osize = NODEDSZ(node);
5450 delta = sp->mp_upper - sp->mp_lower;
5451 SETDSZ(node, osize - delta);
5452 xp = (MDB_page *)((char *)sp + delta);
5454 /* shift subpage upward */
5456 nsize = NUMKEYS(sp) * sp->mp_pad;
5457 memmove(METADATA(xp), METADATA(sp), nsize);
5460 nsize = osize - sp->mp_upper;
5461 numkeys = NUMKEYS(sp);
5462 for (i=numkeys-1; i>=0; i--)
5463 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5465 xp->mp_upper = sp->mp_lower;
5466 xp->mp_lower = sp->mp_lower;
5467 xp->mp_flags = sp->mp_flags;
5468 xp->mp_pad = sp->mp_pad;
5469 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5471 /* shift lower nodes upward */
5472 ptr = mp->mp_ptrs[indx];
5473 numkeys = NUMKEYS(mp);
5474 for (i = 0; i < numkeys; i++) {
5475 if (mp->mp_ptrs[i] <= ptr)
5476 mp->mp_ptrs[i] += delta;
5479 base = (char *)mp + mp->mp_upper;
5480 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5481 mp->mp_upper += delta;
5484 /** Initial setup of a sorted-dups cursor.
5485 * Sorted duplicates are implemented as a sub-database for the given key.
5486 * The duplicate data items are actually keys of the sub-database.
5487 * Operations on the duplicate data items are performed using a sub-cursor
5488 * initialized when the sub-database is first accessed. This function does
5489 * the preliminary setup of the sub-cursor, filling in the fields that
5490 * depend only on the parent DB.
5491 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5494 mdb_xcursor_init0(MDB_cursor *mc)
5496 MDB_xcursor *mx = mc->mc_xcursor;
5498 mx->mx_cursor.mc_xcursor = NULL;
5499 mx->mx_cursor.mc_txn = mc->mc_txn;
5500 mx->mx_cursor.mc_db = &mx->mx_db;
5501 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5502 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5503 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5504 mx->mx_cursor.mc_snum = 0;
5505 mx->mx_cursor.mc_top = 0;
5506 mx->mx_cursor.mc_flags = C_SUB;
5507 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5508 mx->mx_dbx.md_dcmp = NULL;
5509 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5512 /** Final setup of a sorted-dups cursor.
5513 * Sets up the fields that depend on the data from the main cursor.
5514 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5515 * @param[in] node The data containing the #MDB_db record for the
5516 * sorted-dup database.
5519 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5521 MDB_xcursor *mx = mc->mc_xcursor;
5523 if (node->mn_flags & F_SUBDATA) {
5524 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5525 mx->mx_cursor.mc_pg[0] = 0;
5526 mx->mx_cursor.mc_snum = 0;
5527 mx->mx_cursor.mc_flags = C_SUB;
5529 MDB_page *fp = NODEDATA(node);
5530 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5531 mx->mx_db.md_flags = 0;
5532 mx->mx_db.md_depth = 1;
5533 mx->mx_db.md_branch_pages = 0;
5534 mx->mx_db.md_leaf_pages = 1;
5535 mx->mx_db.md_overflow_pages = 0;
5536 mx->mx_db.md_entries = NUMKEYS(fp);
5537 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5538 mx->mx_cursor.mc_snum = 1;
5539 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5540 mx->mx_cursor.mc_top = 0;
5541 mx->mx_cursor.mc_pg[0] = fp;
5542 mx->mx_cursor.mc_ki[0] = 0;
5543 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5544 mx->mx_db.md_flags = MDB_DUPFIXED;
5545 mx->mx_db.md_pad = fp->mp_pad;
5546 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5547 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5550 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5552 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5554 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5555 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5556 #if UINT_MAX < SIZE_MAX
5557 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5558 #ifdef MISALIGNED_OK
5559 mx->mx_dbx.md_cmp = mdb_cmp_long;
5561 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5566 /** Initialize a cursor for a given transaction and database. */
5568 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5573 mc->mc_db = &txn->mt_dbs[dbi];
5574 mc->mc_dbx = &txn->mt_dbxs[dbi];
5575 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5580 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5582 mc->mc_xcursor = mx;
5583 mdb_xcursor_init0(mc);
5585 mc->mc_xcursor = NULL;
5587 if (*mc->mc_dbflag & DB_STALE) {
5588 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5593 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5596 MDB_xcursor *mx = NULL;
5597 size_t size = sizeof(MDB_cursor);
5599 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5602 /* Allow read access to the freelist */
5603 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5606 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5607 size += sizeof(MDB_xcursor);
5609 if ((mc = malloc(size)) != NULL) {
5610 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5611 mx = (MDB_xcursor *)(mc + 1);
5613 mdb_cursor_init(mc, txn, dbi, mx);
5614 if (txn->mt_cursors) {
5615 mc->mc_next = txn->mt_cursors[dbi];
5616 txn->mt_cursors[dbi] = mc;
5618 mc->mc_flags |= C_ALLOCD;
5629 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5631 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5634 if (txn->mt_cursors)
5637 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5641 /* Return the count of duplicate data items for the current key */
5643 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5647 if (mc == NULL || countp == NULL)
5650 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5653 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5654 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5657 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5660 *countp = mc->mc_xcursor->mx_db.md_entries;
5666 mdb_cursor_close(MDB_cursor *mc)
5669 /* remove from txn, if tracked */
5670 if (mc->mc_txn->mt_cursors) {
5671 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5672 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5674 *prev = mc->mc_next;
5676 if (mc->mc_flags & C_ALLOCD)
5682 mdb_cursor_txn(MDB_cursor *mc)
5684 if (!mc) return NULL;
5689 mdb_cursor_dbi(MDB_cursor *mc)
5695 /** Replace the key for a node with a new key.
5696 * @param[in] mp The page containing the node to operate on.
5697 * @param[in] indx The index of the node to operate on.
5698 * @param[in] key The new key to use.
5699 * @return 0 on success, non-zero on failure.
5702 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5708 indx_t ptr, i, numkeys;
5711 node = NODEPTR(mp, indx);
5712 ptr = mp->mp_ptrs[indx];
5716 char kbuf2[(MAXKEYSIZE*2+1)];
5717 k2.mv_data = NODEKEY(node);
5718 k2.mv_size = node->mn_ksize;
5719 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5721 mdb_dkey(&k2, kbuf2),
5727 delta0 = delta = key->mv_size - node->mn_ksize;
5729 /* Must be 2-byte aligned. If new key is
5730 * shorter by 1, the shift will be skipped.
5732 delta += (delta & 1);
5734 if (delta > 0 && SIZELEFT(mp) < delta) {
5735 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5736 return MDB_PAGE_FULL;
5739 numkeys = NUMKEYS(mp);
5740 for (i = 0; i < numkeys; i++) {
5741 if (mp->mp_ptrs[i] <= ptr)
5742 mp->mp_ptrs[i] -= delta;
5745 base = (char *)mp + mp->mp_upper;
5746 len = ptr - mp->mp_upper + NODESIZE;
5747 memmove(base - delta, base, len);
5748 mp->mp_upper -= delta;
5750 node = NODEPTR(mp, indx);
5753 /* But even if no shift was needed, update ksize */
5755 node->mn_ksize = key->mv_size;
5758 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5763 /** Move a node from csrc to cdst.
5766 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5772 unsigned short flags;
5776 /* Mark src and dst as dirty. */
5777 if ((rc = mdb_page_touch(csrc)) ||
5778 (rc = mdb_page_touch(cdst)))
5781 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5782 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5783 key.mv_size = csrc->mc_db->md_pad;
5784 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5786 data.mv_data = NULL;
5790 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5791 assert(!((long)srcnode&1));
5792 srcpg = NODEPGNO(srcnode);
5793 flags = srcnode->mn_flags;
5794 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5795 unsigned int snum = csrc->mc_snum;
5797 /* must find the lowest key below src */
5798 mdb_page_search_root(csrc, NULL, 0);
5799 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5800 key.mv_size = csrc->mc_db->md_pad;
5801 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5803 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5804 key.mv_size = NODEKSZ(s2);
5805 key.mv_data = NODEKEY(s2);
5807 csrc->mc_snum = snum--;
5808 csrc->mc_top = snum;
5810 key.mv_size = NODEKSZ(srcnode);
5811 key.mv_data = NODEKEY(srcnode);
5813 data.mv_size = NODEDSZ(srcnode);
5814 data.mv_data = NODEDATA(srcnode);
5816 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5817 unsigned int snum = cdst->mc_snum;
5820 /* must find the lowest key below dst */
5821 mdb_page_search_root(cdst, NULL, 0);
5822 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5823 bkey.mv_size = cdst->mc_db->md_pad;
5824 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5826 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5827 bkey.mv_size = NODEKSZ(s2);
5828 bkey.mv_data = NODEKEY(s2);
5830 cdst->mc_snum = snum--;
5831 cdst->mc_top = snum;
5832 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5835 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5836 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5837 csrc->mc_ki[csrc->mc_top],
5839 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5840 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5842 /* Add the node to the destination page.
5844 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5845 if (rc != MDB_SUCCESS)
5848 /* Delete the node from the source page.
5850 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5853 /* Adjust other cursors pointing to mp */
5854 MDB_cursor *m2, *m3;
5855 MDB_dbi dbi = csrc->mc_dbi;
5856 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5858 if (csrc->mc_flags & C_SUB)
5861 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5862 if (m2 == csrc) continue;
5863 if (csrc->mc_flags & C_SUB)
5864 m3 = &m2->mc_xcursor->mx_cursor;
5867 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5868 csrc->mc_ki[csrc->mc_top]) {
5869 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5870 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5875 /* Update the parent separators.
5877 if (csrc->mc_ki[csrc->mc_top] == 0) {
5878 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5879 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5880 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5882 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5883 key.mv_size = NODEKSZ(srcnode);
5884 key.mv_data = NODEKEY(srcnode);
5886 DPRINTF("update separator for source page %zu to [%s]",
5887 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5888 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5889 &key)) != MDB_SUCCESS)
5892 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5894 nullkey.mv_size = 0;
5895 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5896 assert(rc == MDB_SUCCESS);
5900 if (cdst->mc_ki[cdst->mc_top] == 0) {
5901 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5902 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5903 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5905 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5906 key.mv_size = NODEKSZ(srcnode);
5907 key.mv_data = NODEKEY(srcnode);
5909 DPRINTF("update separator for destination page %zu to [%s]",
5910 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5911 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5912 &key)) != MDB_SUCCESS)
5915 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5917 nullkey.mv_size = 0;
5918 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5919 assert(rc == MDB_SUCCESS);
5926 /** Merge one page into another.
5927 * The nodes from the page pointed to by \b csrc will
5928 * be copied to the page pointed to by \b cdst and then
5929 * the \b csrc page will be freed.
5930 * @param[in] csrc Cursor pointing to the source page.
5931 * @param[in] cdst Cursor pointing to the destination page.
5934 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5942 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5943 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5945 assert(csrc->mc_snum > 1); /* can't merge root page */
5946 assert(cdst->mc_snum > 1);
5948 /* Mark dst as dirty. */
5949 if ((rc = mdb_page_touch(cdst)))
5952 /* Move all nodes from src to dst.
5954 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5955 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5956 key.mv_size = csrc->mc_db->md_pad;
5957 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5958 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5959 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5960 if (rc != MDB_SUCCESS)
5962 key.mv_data = (char *)key.mv_data + key.mv_size;
5965 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5966 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5967 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5968 unsigned int snum = csrc->mc_snum;
5970 /* must find the lowest key below src */
5971 mdb_page_search_root(csrc, NULL, 0);
5972 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5973 key.mv_size = csrc->mc_db->md_pad;
5974 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5976 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5977 key.mv_size = NODEKSZ(s2);
5978 key.mv_data = NODEKEY(s2);
5980 csrc->mc_snum = snum--;
5981 csrc->mc_top = snum;
5983 key.mv_size = srcnode->mn_ksize;
5984 key.mv_data = NODEKEY(srcnode);
5987 data.mv_size = NODEDSZ(srcnode);
5988 data.mv_data = NODEDATA(srcnode);
5989 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5990 if (rc != MDB_SUCCESS)
5995 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5996 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);
5998 /* Unlink the src page from parent and add to free list.
6000 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
6001 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
6003 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
6007 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
6008 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
6009 csrc->mc_db->md_leaf_pages--;
6011 csrc->mc_db->md_branch_pages--;
6013 /* Adjust other cursors pointing to mp */
6014 MDB_cursor *m2, *m3;
6015 MDB_dbi dbi = csrc->mc_dbi;
6016 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
6018 if (csrc->mc_flags & C_SUB)
6021 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6022 if (csrc->mc_flags & C_SUB)
6023 m3 = &m2->mc_xcursor->mx_cursor;
6026 if (m3 == csrc) continue;
6027 if (m3->mc_snum < csrc->mc_snum) continue;
6028 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
6029 m3->mc_pg[csrc->mc_top] = mp;
6030 m3->mc_ki[csrc->mc_top] += nkeys;
6034 mdb_cursor_pop(csrc);
6036 return mdb_rebalance(csrc);
6039 /** Copy the contents of a cursor.
6040 * @param[in] csrc The cursor to copy from.
6041 * @param[out] cdst The cursor to copy to.
6044 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
6048 cdst->mc_txn = csrc->mc_txn;
6049 cdst->mc_dbi = csrc->mc_dbi;
6050 cdst->mc_db = csrc->mc_db;
6051 cdst->mc_dbx = csrc->mc_dbx;
6052 cdst->mc_snum = csrc->mc_snum;
6053 cdst->mc_top = csrc->mc_top;
6054 cdst->mc_flags = csrc->mc_flags;
6056 for (i=0; i<csrc->mc_snum; i++) {
6057 cdst->mc_pg[i] = csrc->mc_pg[i];
6058 cdst->mc_ki[i] = csrc->mc_ki[i];
6062 /** Rebalance the tree after a delete operation.
6063 * @param[in] mc Cursor pointing to the page where rebalancing
6065 * @return 0 on success, non-zero on failure.
6068 mdb_rebalance(MDB_cursor *mc)
6078 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6079 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
6080 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
6081 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
6085 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
6088 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
6089 DPRINTF("no need to rebalance page %zu, above fill threshold",
6095 if (mc->mc_snum < 2) {
6096 MDB_page *mp = mc->mc_pg[0];
6097 if (NUMKEYS(mp) == 0) {
6098 DPUTS("tree is completely empty");
6099 mc->mc_db->md_root = P_INVALID;
6100 mc->mc_db->md_depth = 0;
6101 mc->mc_db->md_leaf_pages = 0;
6102 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6106 /* Adjust other cursors pointing to mp */
6107 MDB_cursor *m2, *m3;
6108 MDB_dbi dbi = mc->mc_dbi;
6110 if (mc->mc_flags & C_SUB)
6113 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6114 if (m2 == mc) continue;
6115 if (mc->mc_flags & C_SUB)
6116 m3 = &m2->mc_xcursor->mx_cursor;
6119 if (m3->mc_snum < mc->mc_snum) continue;
6120 if (m3->mc_pg[0] == mp) {
6126 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
6127 DPUTS("collapsing root page!");
6128 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
6129 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
6130 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
6133 mc->mc_db->md_depth--;
6134 mc->mc_db->md_branch_pages--;
6136 /* Adjust other cursors pointing to mp */
6137 MDB_cursor *m2, *m3;
6138 MDB_dbi dbi = mc->mc_dbi;
6140 if (mc->mc_flags & C_SUB)
6143 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6144 if (m2 == mc) continue;
6145 if (mc->mc_flags & C_SUB)
6146 m3 = &m2->mc_xcursor->mx_cursor;
6149 if (m3->mc_snum < mc->mc_snum) continue;
6150 if (m3->mc_pg[0] == mp) {
6151 m3->mc_pg[0] = mc->mc_pg[0];
6156 DPUTS("root page doesn't need rebalancing");
6160 /* The parent (branch page) must have at least 2 pointers,
6161 * otherwise the tree is invalid.
6163 ptop = mc->mc_top-1;
6164 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
6166 /* Leaf page fill factor is below the threshold.
6167 * Try to move keys from left or right neighbor, or
6168 * merge with a neighbor page.
6173 mdb_cursor_copy(mc, &mn);
6174 mn.mc_xcursor = NULL;
6176 if (mc->mc_ki[ptop] == 0) {
6177 /* We're the leftmost leaf in our parent.
6179 DPUTS("reading right neighbor");
6181 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6182 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6184 mn.mc_ki[mn.mc_top] = 0;
6185 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
6187 /* There is at least one neighbor to the left.
6189 DPUTS("reading left neighbor");
6191 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
6192 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
6194 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
6195 mc->mc_ki[mc->mc_top] = 0;
6198 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
6199 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);
6201 /* If the neighbor page is above threshold and has at least two
6202 * keys, move one key from it.
6204 * Otherwise we should try to merge them.
6206 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
6207 return mdb_node_move(&mn, mc);
6208 else { /* FIXME: if (has_enough_room()) */
6209 mc->mc_flags &= ~C_INITIALIZED;
6210 if (mc->mc_ki[ptop] == 0)
6211 return mdb_page_merge(&mn, mc);
6213 return mdb_page_merge(mc, &mn);
6217 /** Complete a delete operation started by #mdb_cursor_del(). */
6219 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
6223 /* add overflow pages to free list */
6224 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6228 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6229 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
6230 mc->mc_db->md_overflow_pages -= ovpages;
6231 for (i=0; i<ovpages; i++) {
6232 DPRINTF("freed ov page %zu", pg);
6233 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6237 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
6238 mc->mc_db->md_entries--;
6239 rc = mdb_rebalance(mc);
6240 if (rc != MDB_SUCCESS)
6241 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6247 mdb_del(MDB_txn *txn, MDB_dbi dbi,
6248 MDB_val *key, MDB_val *data)
6253 MDB_val rdata, *xdata;
6257 assert(key != NULL);
6259 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
6261 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6264 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6268 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6272 mdb_cursor_init(&mc, txn, dbi, &mx);
6283 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
6285 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
6289 /** Split a page and insert a new node.
6290 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
6291 * The cursor will be updated to point to the actual page and index where
6292 * the node got inserted after the split.
6293 * @param[in] newkey The key for the newly inserted node.
6294 * @param[in] newdata The data for the newly inserted node.
6295 * @param[in] newpgno The page number, if the new node is a branch node.
6296 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
6297 * @return 0 on success, non-zero on failure.
6300 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
6301 unsigned int nflags)
6304 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
6307 unsigned int i, j, split_indx, nkeys, pmax;
6309 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
6311 MDB_page *mp, *rp, *pp;
6316 mp = mc->mc_pg[mc->mc_top];
6317 newindx = mc->mc_ki[mc->mc_top];
6319 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
6320 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
6321 DKEY(newkey), mc->mc_ki[mc->mc_top]);
6323 /* Create a right sibling. */
6324 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
6326 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
6328 if (mc->mc_snum < 2) {
6329 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
6331 /* shift current top to make room for new parent */
6332 mc->mc_pg[1] = mc->mc_pg[0];
6333 mc->mc_ki[1] = mc->mc_ki[0];
6336 mc->mc_db->md_root = pp->mp_pgno;
6337 DPRINTF("root split! new root = %zu", pp->mp_pgno);
6338 mc->mc_db->md_depth++;
6341 /* Add left (implicit) pointer. */
6342 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
6343 /* undo the pre-push */
6344 mc->mc_pg[0] = mc->mc_pg[1];
6345 mc->mc_ki[0] = mc->mc_ki[1];
6346 mc->mc_db->md_root = mp->mp_pgno;
6347 mc->mc_db->md_depth--;
6354 ptop = mc->mc_top-1;
6355 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6358 mc->mc_flags |= C_SPLITTING;
6359 mdb_cursor_copy(mc, &mn);
6360 mn.mc_pg[mn.mc_top] = rp;
6361 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6363 if (nflags & MDB_APPEND) {
6364 mn.mc_ki[mn.mc_top] = 0;
6366 split_indx = newindx;
6371 nkeys = NUMKEYS(mp);
6372 split_indx = nkeys / 2;
6373 if (newindx < split_indx)
6379 unsigned int lsize, rsize, ksize;
6380 /* Move half of the keys to the right sibling */
6382 x = mc->mc_ki[mc->mc_top] - split_indx;
6383 ksize = mc->mc_db->md_pad;
6384 split = LEAF2KEY(mp, split_indx, ksize);
6385 rsize = (nkeys - split_indx) * ksize;
6386 lsize = (nkeys - split_indx) * sizeof(indx_t);
6387 mp->mp_lower -= lsize;
6388 rp->mp_lower += lsize;
6389 mp->mp_upper += rsize - lsize;
6390 rp->mp_upper -= rsize - lsize;
6391 sepkey.mv_size = ksize;
6392 if (newindx == split_indx) {
6393 sepkey.mv_data = newkey->mv_data;
6395 sepkey.mv_data = split;
6398 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6399 memcpy(rp->mp_ptrs, split, rsize);
6400 sepkey.mv_data = rp->mp_ptrs;
6401 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6402 memcpy(ins, newkey->mv_data, ksize);
6403 mp->mp_lower += sizeof(indx_t);
6404 mp->mp_upper -= ksize - sizeof(indx_t);
6407 memcpy(rp->mp_ptrs, split, x * ksize);
6408 ins = LEAF2KEY(rp, x, ksize);
6409 memcpy(ins, newkey->mv_data, ksize);
6410 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6411 rp->mp_lower += sizeof(indx_t);
6412 rp->mp_upper -= ksize - sizeof(indx_t);
6413 mc->mc_ki[mc->mc_top] = x;
6414 mc->mc_pg[mc->mc_top] = rp;
6419 /* For leaf pages, check the split point based on what
6420 * fits where, since otherwise mdb_node_add can fail.
6422 * This check is only needed when the data items are
6423 * relatively large, such that being off by one will
6424 * make the difference between success or failure.
6426 * It's also relevant if a page happens to be laid out
6427 * such that one half of its nodes are all "small" and
6428 * the other half of its nodes are "large." If the new
6429 * item is also "large" and falls on the half with
6430 * "large" nodes, it also may not fit.
6433 unsigned int psize, nsize;
6434 /* Maximum free space in an empty page */
6435 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6436 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6437 if ((nkeys < 20) || (nsize > pmax/16)) {
6438 if (newindx <= split_indx) {
6441 for (i=0; i<split_indx; i++) {
6442 node = NODEPTR(mp, i);
6443 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6444 if (F_ISSET(node->mn_flags, F_BIGDATA))
6445 psize += sizeof(pgno_t);
6447 psize += NODEDSZ(node);
6451 split_indx = newindx;
6462 for (i=nkeys-1; i>=split_indx; i--) {
6463 node = NODEPTR(mp, i);
6464 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6465 if (F_ISSET(node->mn_flags, F_BIGDATA))
6466 psize += sizeof(pgno_t);
6468 psize += NODEDSZ(node);
6472 split_indx = newindx;
6483 /* First find the separating key between the split pages.
6484 * The case where newindx == split_indx is ambiguous; the
6485 * new item could go to the new page or stay on the original
6486 * page. If newpos == 1 it goes to the new page.
6488 if (newindx == split_indx && newpos) {
6489 sepkey.mv_size = newkey->mv_size;
6490 sepkey.mv_data = newkey->mv_data;
6492 node = NODEPTR(mp, split_indx);
6493 sepkey.mv_size = node->mn_ksize;
6494 sepkey.mv_data = NODEKEY(node);
6498 DPRINTF("separator is [%s]", DKEY(&sepkey));
6500 /* Copy separator key to the parent.
6502 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6506 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6509 if (mn.mc_snum == mc->mc_snum) {
6510 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6511 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6512 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6513 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6518 /* Right page might now have changed parent.
6519 * Check if left page also changed parent.
6521 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6522 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6523 for (i=0; i<ptop; i++) {
6524 mc->mc_pg[i] = mn.mc_pg[i];
6525 mc->mc_ki[i] = mn.mc_ki[i];
6527 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6528 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6532 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6535 mc->mc_flags ^= C_SPLITTING;
6536 if (rc != MDB_SUCCESS) {
6539 if (nflags & MDB_APPEND) {
6540 mc->mc_pg[mc->mc_top] = rp;
6541 mc->mc_ki[mc->mc_top] = 0;
6542 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6545 for (i=0; i<mc->mc_top; i++)
6546 mc->mc_ki[i] = mn.mc_ki[i];
6553 /* Move half of the keys to the right sibling. */
6555 /* grab a page to hold a temporary copy */
6556 copy = mdb_page_malloc(mc);
6560 copy->mp_pgno = mp->mp_pgno;
6561 copy->mp_flags = mp->mp_flags;
6562 copy->mp_lower = PAGEHDRSZ;
6563 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6564 mc->mc_pg[mc->mc_top] = copy;
6565 for (i = j = 0; i <= nkeys; j++) {
6566 if (i == split_indx) {
6567 /* Insert in right sibling. */
6568 /* Reset insert index for right sibling. */
6569 if (i != newindx || (newpos ^ ins_new)) {
6571 mc->mc_pg[mc->mc_top] = rp;
6575 if (i == newindx && !ins_new) {
6576 /* Insert the original entry that caused the split. */
6577 rkey.mv_data = newkey->mv_data;
6578 rkey.mv_size = newkey->mv_size;
6587 /* Update index for the new key. */
6588 mc->mc_ki[mc->mc_top] = j;
6589 } else if (i == nkeys) {
6592 node = NODEPTR(mp, i);
6593 rkey.mv_data = NODEKEY(node);
6594 rkey.mv_size = node->mn_ksize;
6596 xdata.mv_data = NODEDATA(node);
6597 xdata.mv_size = NODEDSZ(node);
6600 pgno = NODEPGNO(node);
6601 flags = node->mn_flags;
6606 if (!IS_LEAF(mp) && j == 0) {
6607 /* First branch index doesn't need key data. */
6611 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6615 nkeys = NUMKEYS(copy);
6616 for (i=0; i<nkeys; i++)
6617 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6618 mp->mp_lower = copy->mp_lower;
6619 mp->mp_upper = copy->mp_upper;
6620 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6621 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6623 /* reset back to original page */
6624 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6625 mc->mc_pg[mc->mc_top] = mp;
6626 if (nflags & MDB_RESERVE) {
6627 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6628 if (!(node->mn_flags & F_BIGDATA))
6629 newdata->mv_data = NODEDATA(node);
6635 /* return tmp page to freelist */
6636 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6637 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6638 mc->mc_txn->mt_env->me_dpages = copy;
6641 /* Adjust other cursors pointing to mp */
6642 MDB_cursor *m2, *m3;
6643 MDB_dbi dbi = mc->mc_dbi;
6644 int fixup = NUMKEYS(mp);
6646 if (mc->mc_flags & C_SUB)
6649 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6650 if (m2 == mc) continue;
6651 if (mc->mc_flags & C_SUB)
6652 m3 = &m2->mc_xcursor->mx_cursor;
6655 if (!(m3->mc_flags & C_INITIALIZED))
6657 if (m3->mc_flags & C_SPLITTING)
6662 for (k=m3->mc_top; k>=0; k--) {
6663 m3->mc_ki[k+1] = m3->mc_ki[k];
6664 m3->mc_pg[k+1] = m3->mc_pg[k];
6666 if (m3->mc_ki[0] >= split_indx) {
6671 m3->mc_pg[0] = mc->mc_pg[0];
6675 if (m3->mc_pg[mc->mc_top] == mp) {
6676 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6677 m3->mc_ki[mc->mc_top]++;
6678 if (m3->mc_ki[mc->mc_top] >= fixup) {
6679 m3->mc_pg[mc->mc_top] = rp;
6680 m3->mc_ki[mc->mc_top] -= fixup;
6681 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6683 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6684 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6693 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6694 MDB_val *key, MDB_val *data, unsigned int flags)
6699 assert(key != NULL);
6700 assert(data != NULL);
6702 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6705 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6709 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6713 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6716 mdb_cursor_init(&mc, txn, dbi, &mx);
6717 return mdb_cursor_put(&mc, key, data, flags);
6721 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6723 if ((flag & CHANGEABLE) != flag)
6726 env->me_flags |= flag;
6728 env->me_flags &= ~flag;
6733 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6738 *arg = env->me_flags;
6743 mdb_env_get_path(MDB_env *env, const char **arg)
6748 *arg = env->me_path;
6752 /** Common code for #mdb_stat() and #mdb_env_stat().
6753 * @param[in] env the environment to operate in.
6754 * @param[in] db the #MDB_db record containing the stats to return.
6755 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6756 * @return 0, this function always succeeds.
6759 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6761 arg->ms_psize = env->me_psize;
6762 arg->ms_depth = db->md_depth;
6763 arg->ms_branch_pages = db->md_branch_pages;
6764 arg->ms_leaf_pages = db->md_leaf_pages;
6765 arg->ms_overflow_pages = db->md_overflow_pages;
6766 arg->ms_entries = db->md_entries;
6771 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6775 if (env == NULL || arg == NULL)
6778 toggle = mdb_env_pick_meta(env);
6780 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6784 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
6788 if (env == NULL || arg == NULL)
6791 toggle = mdb_env_pick_meta(env);
6792 arg->me_mapaddr = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : 0;
6793 arg->me_mapsize = env->me_mapsize;
6794 arg->me_maxreaders = env->me_maxreaders;
6795 arg->me_numreaders = env->me_numreaders;
6796 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
6797 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
6801 /** Set the default comparison functions for a database.
6802 * Called immediately after a database is opened to set the defaults.
6803 * The user can then override them with #mdb_set_compare() or
6804 * #mdb_set_dupsort().
6805 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6806 * @param[in] dbi A database handle returned by #mdb_dbi_open()
6809 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6811 uint16_t f = txn->mt_dbs[dbi].md_flags;
6813 txn->mt_dbxs[dbi].md_cmp =
6814 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6815 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6817 txn->mt_dbxs[dbi].md_dcmp =
6818 !(f & MDB_DUPSORT) ? 0 :
6819 ((f & MDB_INTEGERDUP)
6820 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6821 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6824 #define PERSISTENT_FLAGS 0xffff
6825 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
6826 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
6827 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6832 int rc, dbflag, exact;
6833 unsigned int unused = 0;
6836 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6837 mdb_default_cmp(txn, FREE_DBI);
6840 if ((flags & VALID_FLAGS) != flags)
6846 if (flags & PERSISTENT_FLAGS) {
6847 uint16_t f2 = flags & PERSISTENT_FLAGS;
6848 /* make sure flag changes get committed */
6849 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
6850 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
6851 txn->mt_flags |= MDB_TXN_DIRTY;
6854 mdb_default_cmp(txn, MAIN_DBI);
6858 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6859 mdb_default_cmp(txn, MAIN_DBI);
6862 /* Is the DB already open? */
6864 for (i=2; i<txn->mt_numdbs; i++) {
6865 if (!txn->mt_dbxs[i].md_name.mv_size) {
6866 /* Remember this free slot */
6867 if (!unused) unused = i;
6870 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6871 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6877 /* If no free slot and max hit, fail */
6878 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
6879 return MDB_DBS_FULL;
6881 /* Find the DB info */
6885 key.mv_data = (void *)name;
6886 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6887 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6888 if (rc == MDB_SUCCESS) {
6889 /* make sure this is actually a DB */
6890 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6891 if (!(node->mn_flags & F_SUBDATA))
6893 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6894 /* Create if requested */
6896 data.mv_size = sizeof(MDB_db);
6897 data.mv_data = &dummy;
6898 memset(&dummy, 0, sizeof(dummy));
6899 dummy.md_root = P_INVALID;
6900 dummy.md_flags = flags & PERSISTENT_FLAGS;
6901 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6905 /* OK, got info, add to table */
6906 if (rc == MDB_SUCCESS) {
6907 unsigned int slot = unused ? unused : txn->mt_numdbs;
6908 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6909 txn->mt_dbxs[slot].md_name.mv_size = len;
6910 txn->mt_dbxs[slot].md_rel = NULL;
6911 txn->mt_dbflags[slot] = dbflag;
6912 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6914 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6915 mdb_default_cmp(txn, slot);
6918 txn->mt_env->me_numdbs++;
6925 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6927 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6930 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6933 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
6936 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6938 ptr = env->me_dbxs[dbi].md_name.mv_data;
6939 env->me_dbxs[dbi].md_name.mv_data = NULL;
6940 env->me_dbxs[dbi].md_name.mv_size = 0;
6944 /** Add all the DB's pages to the free list.
6945 * @param[in] mc Cursor on the DB to free.
6946 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6947 * @return 0 on success, non-zero on failure.
6950 mdb_drop0(MDB_cursor *mc, int subs)
6954 rc = mdb_page_search(mc, NULL, 0);
6955 if (rc == MDB_SUCCESS) {
6960 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6961 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6964 mdb_cursor_copy(mc, &mx);
6965 while (mc->mc_snum > 0) {
6966 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6967 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6968 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6969 if (ni->mn_flags & F_SUBDATA) {
6970 mdb_xcursor_init1(mc, ni);
6971 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6977 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6979 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6982 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6987 rc = mdb_cursor_sibling(mc, 1);
6989 /* no more siblings, go back to beginning
6990 * of previous level.
6993 for (i=1; i<mc->mc_top; i++)
6994 mc->mc_pg[i] = mx.mc_pg[i];
6998 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6999 mc->mc_db->md_root);
7004 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
7009 if (!txn || !dbi || dbi >= txn->mt_numdbs || (unsigned)del > 1)
7012 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7015 rc = mdb_cursor_open(txn, dbi, &mc);
7019 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
7023 /* Can't delete the main DB */
7024 if (del && dbi > MAIN_DBI) {
7025 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
7027 mdb_dbi_close(txn->mt_env, dbi);
7029 /* reset the DB record, mark it dirty */
7030 txn->mt_dbflags[dbi] |= DB_DIRTY;
7031 txn->mt_dbs[dbi].md_depth = 0;
7032 txn->mt_dbs[dbi].md_branch_pages = 0;
7033 txn->mt_dbs[dbi].md_leaf_pages = 0;
7034 txn->mt_dbs[dbi].md_overflow_pages = 0;
7035 txn->mt_dbs[dbi].md_entries = 0;
7036 txn->mt_dbs[dbi].md_root = P_INVALID;
7038 if (!txn->mt_u.dirty_list[0].mid) {
7041 /* make sure we have at least one dirty page in this txn
7042 * otherwise these changes will be ignored.
7044 key.mv_size = sizeof(txnid_t);
7045 key.mv_data = &txn->mt_txnid;
7046 data.mv_size = sizeof(MDB_ID);
7047 data.mv_data = txn->mt_free_pgs;
7048 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
7049 rc = mdb_cursor_put(&m2, &key, &data, 0);
7053 mdb_cursor_close(mc);
7057 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7059 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7062 txn->mt_dbxs[dbi].md_cmp = cmp;
7066 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
7068 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7071 txn->mt_dbxs[dbi].md_dcmp = cmp;
7075 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
7077 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7080 txn->mt_dbxs[dbi].md_rel = rel;
7084 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
7086 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
7089 txn->mt_dbxs[dbi].md_relctx = ctx;