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.
35 #include <sys/types.h>
37 #include <sys/param.h>
43 #ifdef HAVE_SYS_FILE_H
60 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
61 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
64 #if defined(__APPLE__) || defined (BSD)
65 # define MDB_USE_POSIX_SEM 1
66 # define MDB_FDATASYNC fsync
67 #elif defined(ANDROID)
68 # define MDB_FDATASYNC fsync
73 #ifdef MDB_USE_POSIX_SEM
74 #include <semaphore.h>
79 #include <valgrind/memcheck.h>
80 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
81 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
82 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
83 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
84 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
86 #define VGMEMP_CREATE(h,r,z)
87 #define VGMEMP_ALLOC(h,a,s)
88 #define VGMEMP_FREE(h,a)
89 #define VGMEMP_DESTROY(h)
90 #define VGMEMP_DEFINED(a,s)
94 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
95 /* Solaris just defines one or the other */
96 # define LITTLE_ENDIAN 1234
97 # define BIG_ENDIAN 4321
98 # ifdef _LITTLE_ENDIAN
99 # define BYTE_ORDER LITTLE_ENDIAN
101 # define BYTE_ORDER BIG_ENDIAN
104 # define BYTE_ORDER __BYTE_ORDER
108 #ifndef LITTLE_ENDIAN
109 #define LITTLE_ENDIAN __LITTLE_ENDIAN
112 #define BIG_ENDIAN __BIG_ENDIAN
115 #if defined(__i386) || defined(__x86_64)
116 #define MISALIGNED_OK 1
122 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
123 # error "Unknown or unsupported endianness (BYTE_ORDER)"
124 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
125 # error "Two's complement, reasonably sized integer types, please"
128 /** @defgroup internal MDB Internals
131 /** @defgroup compat Windows Compatibility Macros
132 * A bunch of macros to minimize the amount of platform-specific ifdefs
133 * needed throughout the rest of the code. When the features this library
134 * needs are similar enough to POSIX to be hidden in a one-or-two line
135 * replacement, this macro approach is used.
139 #define pthread_t DWORD
140 #define pthread_mutex_t HANDLE
141 #define pthread_key_t DWORD
142 #define pthread_self() GetCurrentThreadId()
143 #define pthread_key_create(x,y) (*(x) = TlsAlloc())
144 #define pthread_key_delete(x) TlsFree(x)
145 #define pthread_getspecific(x) TlsGetValue(x)
146 #define pthread_setspecific(x,y) TlsSetValue(x,y)
147 #define pthread_mutex_unlock(x) ReleaseMutex(x)
148 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
149 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
150 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
151 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
152 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
153 #define getpid() GetCurrentProcessId()
154 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
155 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
156 #define ErrCode() GetLastError()
157 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
158 #define close(fd) CloseHandle(fd)
159 #define munmap(ptr,len) UnmapViewOfFile(ptr)
161 #ifdef MDB_USE_POSIX_SEM
162 #define LOCK_MUTEX_R(env) sem_wait((env)->me_rmutex)
163 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
164 #define LOCK_MUTEX_W(env) sem_wait((env)->me_wmutex)
165 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
167 /** Lock the reader mutex.
169 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
170 /** Unlock the reader mutex.
172 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
174 /** Lock the writer mutex.
175 * Only a single write transaction is allowed at a time. Other writers
176 * will block waiting for this mutex.
178 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
179 /** Unlock the writer mutex.
181 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
182 #endif /* MDB_USE_POSIX_SEM */
184 /** Get the error code for the last failed system function.
186 #define ErrCode() errno
188 /** An abstraction for a file handle.
189 * On POSIX systems file handles are small integers. On Windows
190 * they're opaque pointers.
194 /** A value for an invalid file handle.
195 * Mainly used to initialize file variables and signify that they are
198 #define INVALID_HANDLE_VALUE (-1)
200 /** Get the size of a memory page for the system.
201 * This is the basic size that the platform's memory manager uses, and is
202 * fundamental to the use of memory-mapped files.
204 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
207 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
210 #define MNAME_LEN (sizeof(pthread_mutex_t))
216 /** A flag for opening a file and requesting synchronous data writes.
217 * This is only used when writing a meta page. It's not strictly needed;
218 * we could just do a normal write and then immediately perform a flush.
219 * But if this flag is available it saves us an extra system call.
221 * @note If O_DSYNC is undefined but exists in /usr/include,
222 * preferably set some compiler flag to get the definition.
223 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
226 # define MDB_DSYNC O_DSYNC
230 /** Function for flushing the data of a file. Define this to fsync
231 * if fdatasync() is not supported.
233 #ifndef MDB_FDATASYNC
234 # define MDB_FDATASYNC fdatasync
238 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
249 /** A page number in the database.
250 * Note that 64 bit page numbers are overkill, since pages themselves
251 * already represent 12-13 bits of addressable memory, and the OS will
252 * always limit applications to a maximum of 63 bits of address space.
254 * @note In the #MDB_node structure, we only store 48 bits of this value,
255 * which thus limits us to only 60 bits of addressable data.
257 typedef MDB_ID pgno_t;
259 /** A transaction ID.
260 * See struct MDB_txn.mt_txnid for details.
262 typedef MDB_ID txnid_t;
264 /** @defgroup debug Debug Macros
268 /** Enable debug output.
269 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
270 * read from and written to the database (used for free space management).
275 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
276 # define DPRINTF (void) /* Vararg macros may be unsupported */
278 static int mdb_debug;
279 static txnid_t mdb_debug_start;
281 /** Print a debug message with printf formatting. */
282 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
283 ((void) ((mdb_debug) && \
284 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
286 # define DPRINTF(fmt, ...) ((void) 0)
288 /** Print a debug string.
289 * The string is printed literally, with no format processing.
291 #define DPUTS(arg) DPRINTF("%s", arg)
294 /** A default memory page size.
295 * The actual size is platform-dependent, but we use this for
296 * boot-strapping. We probably should not be using this any more.
297 * The #GET_PAGESIZE() macro is used to get the actual size.
299 * Note that we don't currently support Huge pages. On Linux,
300 * regular data files cannot use Huge pages, and in general
301 * Huge pages aren't actually pageable. We rely on the OS
302 * demand-pager to read our data and page it out when memory
303 * pressure from other processes is high. So until OSs have
304 * actual paging support for Huge pages, they're not viable.
306 #define MDB_PAGESIZE 4096
308 /** The minimum number of keys required in a database page.
309 * Setting this to a larger value will place a smaller bound on the
310 * maximum size of a data item. Data items larger than this size will
311 * be pushed into overflow pages instead of being stored directly in
312 * the B-tree node. This value used to default to 4. With a page size
313 * of 4096 bytes that meant that any item larger than 1024 bytes would
314 * go into an overflow page. That also meant that on average 2-3KB of
315 * each overflow page was wasted space. The value cannot be lower than
316 * 2 because then there would no longer be a tree structure. With this
317 * value, items larger than 2KB will go into overflow pages, and on
318 * average only 1KB will be wasted.
320 #define MDB_MINKEYS 2
322 /** A stamp that identifies a file as an MDB file.
323 * There's nothing special about this value other than that it is easily
324 * recognizable, and it will reflect any byte order mismatches.
326 #define MDB_MAGIC 0xBEEFC0DE
328 /** The version number for a database's file format. */
329 #define MDB_VERSION 1
331 /** The maximum size of a key in the database.
332 * While data items have essentially unbounded size, we require that
333 * keys all fit onto a regular page. This limit could be raised a bit
334 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
336 #define MAXKEYSIZE 511
341 * This is used for printing a hex dump of a key's contents.
343 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
344 /** Display a key in hex.
346 * Invoke a function to display a key in hex.
348 #define DKEY(x) mdb_dkey(x, kbuf)
350 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
354 /** An invalid page number.
355 * Mainly used to denote an empty tree.
357 #define P_INVALID (~(pgno_t)0)
359 /** Test if a flag \b f is set in a flag word \b w. */
360 #define F_ISSET(w, f) (((w) & (f)) == (f))
362 /** Used for offsets within a single page.
363 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
366 typedef uint16_t indx_t;
368 /** Default size of memory map.
369 * This is certainly too small for any actual applications. Apps should always set
370 * the size explicitly using #mdb_env_set_mapsize().
372 #define DEFAULT_MAPSIZE 1048576
374 /** @defgroup readers Reader Lock Table
375 * Readers don't acquire any locks for their data access. Instead, they
376 * simply record their transaction ID in the reader table. The reader
377 * mutex is needed just to find an empty slot in the reader table. The
378 * slot's address is saved in thread-specific data so that subsequent read
379 * transactions started by the same thread need no further locking to proceed.
381 * Since the database uses multi-version concurrency control, readers don't
382 * actually need any locking. This table is used to keep track of which
383 * readers are using data from which old transactions, so that we'll know
384 * when a particular old transaction is no longer in use. Old transactions
385 * that have discarded any data pages can then have those pages reclaimed
386 * for use by a later write transaction.
388 * The lock table is constructed such that reader slots are aligned with the
389 * processor's cache line size. Any slot is only ever used by one thread.
390 * This alignment guarantees that there will be no contention or cache
391 * thrashing as threads update their own slot info, and also eliminates
392 * any need for locking when accessing a slot.
394 * A writer thread will scan every slot in the table to determine the oldest
395 * outstanding reader transaction. Any freed pages older than this will be
396 * reclaimed by the writer. The writer doesn't use any locks when scanning
397 * this table. This means that there's no guarantee that the writer will
398 * see the most up-to-date reader info, but that's not required for correct
399 * operation - all we need is to know the upper bound on the oldest reader,
400 * we don't care at all about the newest reader. So the only consequence of
401 * reading stale information here is that old pages might hang around a
402 * while longer before being reclaimed. That's actually good anyway, because
403 * the longer we delay reclaiming old pages, the more likely it is that a
404 * string of contiguous pages can be found after coalescing old pages from
405 * many old transactions together.
407 * @todo We don't actually do such coalescing yet, we grab pages from one
408 * old transaction at a time.
411 /** Number of slots in the reader table.
412 * This value was chosen somewhat arbitrarily. 126 readers plus a
413 * couple mutexes fit exactly into 8KB on my development machine.
414 * Applications should set the table size using #mdb_env_set_maxreaders().
416 #define DEFAULT_READERS 126
418 /** The size of a CPU cache line in bytes. We want our lock structures
419 * aligned to this size to avoid false cache line sharing in the
421 * This value works for most CPUs. For Itanium this should be 128.
427 /** The information we store in a single slot of the reader table.
428 * In addition to a transaction ID, we also record the process and
429 * thread ID that owns a slot, so that we can detect stale information,
430 * e.g. threads or processes that went away without cleaning up.
431 * @note We currently don't check for stale records. We simply re-init
432 * the table when we know that we're the only process opening the
435 typedef struct MDB_rxbody {
436 /** The current Transaction ID when this transaction began.
437 * Multiple readers that start at the same time will probably have the
438 * same ID here. Again, it's not important to exclude them from
439 * anything; all we need to know is which version of the DB they
440 * started from so we can avoid overwriting any data used in that
441 * particular version.
444 /** The process ID of the process owning this reader txn. */
446 /** The thread ID of the thread owning this txn. */
450 /** The actual reader record, with cacheline padding. */
451 typedef struct MDB_reader {
454 /** shorthand for mrb_txnid */
455 #define mr_txnid mru.mrx.mrb_txnid
456 #define mr_pid mru.mrx.mrb_pid
457 #define mr_tid mru.mrx.mrb_tid
458 /** cache line alignment */
459 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
463 /** The header for the reader table.
464 * The table resides in a memory-mapped file. (This is a different file
465 * than is used for the main database.)
467 * For POSIX the actual mutexes reside in the shared memory of this
468 * mapped file. On Windows, mutexes are named objects allocated by the
469 * kernel; we store the mutex names in this mapped file so that other
470 * processes can grab them. This same approach is also used on
471 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
472 * process-shared POSIX mutexes. For these cases where a named object
473 * is used, the object name is derived from a 64 bit FNV hash of the
474 * environment pathname. As such, naming collisions are extremely
475 * unlikely. If a collision occurs, the results are unpredictable.
477 typedef struct MDB_txbody {
478 /** Stamp identifying this as an MDB file. It must be set
481 /** Version number of this lock file. Must be set to #MDB_VERSION. */
482 uint32_t mtb_version;
483 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
484 char mtb_rmname[MNAME_LEN];
486 /** Mutex protecting access to this table.
487 * This is the reader lock that #LOCK_MUTEX_R acquires.
489 pthread_mutex_t mtb_mutex;
491 /** The ID of the last transaction committed to the database.
492 * This is recorded here only for convenience; the value can always
493 * be determined by reading the main database meta pages.
496 /** The number of slots that have been used in the reader table.
497 * This always records the maximum count, it is not decremented
498 * when readers release their slots.
500 unsigned mtb_numreaders;
503 /** The actual reader table definition. */
504 typedef struct MDB_txninfo {
507 #define mti_magic mt1.mtb.mtb_magic
508 #define mti_version mt1.mtb.mtb_version
509 #define mti_mutex mt1.mtb.mtb_mutex
510 #define mti_rmname mt1.mtb.mtb_rmname
511 #define mti_txnid mt1.mtb.mtb_txnid
512 #define mti_numreaders mt1.mtb.mtb_numreaders
513 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
516 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
517 char mt2_wmname[MNAME_LEN];
518 #define mti_wmname mt2.mt2_wmname
520 pthread_mutex_t mt2_wmutex;
521 #define mti_wmutex mt2.mt2_wmutex
523 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
525 MDB_reader mti_readers[1];
529 /** Common header for all page types.
530 * Overflow records occupy a number of contiguous pages with no
531 * headers on any page after the first.
533 typedef struct MDB_page {
534 #define mp_pgno mp_p.p_pgno
535 #define mp_next mp_p.p_next
537 pgno_t p_pgno; /**< page number */
538 void * p_next; /**< for in-memory list of freed structs */
541 /** @defgroup mdb_page Page Flags
543 * Flags for the page headers.
546 #define P_BRANCH 0x01 /**< branch page */
547 #define P_LEAF 0x02 /**< leaf page */
548 #define P_OVERFLOW 0x04 /**< overflow page */
549 #define P_META 0x08 /**< meta page */
550 #define P_DIRTY 0x10 /**< dirty page */
551 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
552 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
554 uint16_t mp_flags; /**< @ref mdb_page */
555 #define mp_lower mp_pb.pb.pb_lower
556 #define mp_upper mp_pb.pb.pb_upper
557 #define mp_pages mp_pb.pb_pages
560 indx_t pb_lower; /**< lower bound of free space */
561 indx_t pb_upper; /**< upper bound of free space */
563 uint32_t pb_pages; /**< number of overflow pages */
565 indx_t mp_ptrs[1]; /**< dynamic size */
568 /** Size of the page header, excluding dynamic data at the end */
569 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
571 /** Address of first usable data byte in a page, after the header */
572 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
574 /** Number of nodes on a page */
575 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
577 /** The amount of space remaining in the page */
578 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
580 /** The percentage of space used in the page, in tenths of a percent. */
581 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
582 ((env)->me_psize - PAGEHDRSZ))
583 /** The minimum page fill factor, in tenths of a percent.
584 * Pages emptier than this are candidates for merging.
586 #define FILL_THRESHOLD 250
588 /** Test if a page is a leaf page */
589 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
590 /** Test if a page is a LEAF2 page */
591 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
592 /** Test if a page is a branch page */
593 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
594 /** Test if a page is an overflow page */
595 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
596 /** Test if a page is a sub page */
597 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
599 /** The number of overflow pages needed to store the given size. */
600 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
602 /** Header for a single key/data pair within a page.
603 * We guarantee 2-byte alignment for nodes.
605 typedef struct MDB_node {
606 /** lo and hi are used for data size on leaf nodes and for
607 * child pgno on branch nodes. On 64 bit platforms, flags
608 * is also used for pgno. (Branch nodes have no flags).
609 * They are in host byte order in case that lets some
610 * accesses be optimized into a 32-bit word access.
612 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
613 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
614 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
615 /** @defgroup mdb_node Node Flags
617 * Flags for node headers.
620 #define F_BIGDATA 0x01 /**< data put on overflow page */
621 #define F_SUBDATA 0x02 /**< data is a sub-database */
622 #define F_DUPDATA 0x04 /**< data has duplicates */
624 /** valid flags for #mdb_node_add() */
625 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
628 unsigned short mn_flags; /**< @ref mdb_node */
629 unsigned short mn_ksize; /**< key size */
630 char mn_data[1]; /**< key and data are appended here */
633 /** Size of the node header, excluding dynamic data at the end */
634 #define NODESIZE offsetof(MDB_node, mn_data)
636 /** Bit position of top word in page number, for shifting mn_flags */
637 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
639 /** Size of a node in a branch page with a given key.
640 * This is just the node header plus the key, there is no data.
642 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
644 /** Size of a node in a leaf page with a given key and data.
645 * This is node header plus key plus data size.
647 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
649 /** Address of node \b i in page \b p */
650 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
652 /** Address of the key for the node */
653 #define NODEKEY(node) (void *)((node)->mn_data)
655 /** Address of the data for a node */
656 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
658 /** Get the page number pointed to by a branch node */
659 #define NODEPGNO(node) \
660 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
661 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
662 /** Set the page number in a branch node */
663 #define SETPGNO(node,pgno) do { \
664 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
665 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
667 /** Get the size of the data in a leaf node */
668 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
669 /** Set the size of the data for a leaf node */
670 #define SETDSZ(node,size) do { \
671 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
672 /** The size of a key in a node */
673 #define NODEKSZ(node) ((node)->mn_ksize)
675 /** Copy a page number from src to dst */
677 #define COPY_PGNO(dst,src) dst = src
679 #if SIZE_MAX > 4294967295UL
680 #define COPY_PGNO(dst,src) do { \
681 unsigned short *s, *d; \
682 s = (unsigned short *)&(src); \
683 d = (unsigned short *)&(dst); \
690 #define COPY_PGNO(dst,src) do { \
691 unsigned short *s, *d; \
692 s = (unsigned short *)&(src); \
693 d = (unsigned short *)&(dst); \
699 /** The address of a key in a LEAF2 page.
700 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
701 * There are no node headers, keys are stored contiguously.
703 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
705 /** Set the \b node's key into \b key, if requested. */
706 #define MDB_GET_KEY(node, key) { if ((key) != NULL) { \
707 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
709 /** Information about a single database in the environment. */
710 typedef struct MDB_db {
711 uint32_t md_pad; /**< also ksize for LEAF2 pages */
712 uint16_t md_flags; /**< @ref mdb_open */
713 uint16_t md_depth; /**< depth of this tree */
714 pgno_t md_branch_pages; /**< number of internal pages */
715 pgno_t md_leaf_pages; /**< number of leaf pages */
716 pgno_t md_overflow_pages; /**< number of overflow pages */
717 size_t md_entries; /**< number of data items */
718 pgno_t md_root; /**< the root page of this tree */
721 /** Handle for the DB used to track free pages. */
723 /** Handle for the default DB. */
726 /** Meta page content. */
727 typedef struct MDB_meta {
728 /** Stamp identifying this as an MDB file. It must be set
731 /** Version number of this lock file. Must be set to #MDB_VERSION. */
733 void *mm_address; /**< address for fixed mapping */
734 size_t mm_mapsize; /**< size of mmap region */
735 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
736 /** The size of pages used in this DB */
737 #define mm_psize mm_dbs[0].md_pad
738 /** Any persistent environment flags. @ref mdb_env */
739 #define mm_flags mm_dbs[0].md_flags
740 pgno_t mm_last_pg; /**< last used page in file */
741 txnid_t mm_txnid; /**< txnid that committed this page */
744 /** Buffer for a stack-allocated dirty page.
745 * The members define size and alignment, and silence type
746 * aliasing warnings. They are not used directly; that could
747 * mean incorrectly using several union members in parallel.
749 typedef union MDB_pagebuf {
750 char mb_raw[MDB_PAGESIZE];
753 char mm_pad[PAGEHDRSZ];
758 /** Auxiliary DB info.
759 * The information here is mostly static/read-only. There is
760 * only a single copy of this record in the environment.
762 typedef struct MDB_dbx {
763 MDB_val md_name; /**< name of the database */
764 MDB_cmp_func *md_cmp; /**< function for comparing keys */
765 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
766 MDB_rel_func *md_rel; /**< user relocate function */
767 void *md_relctx; /**< user-provided context for md_rel */
770 /** A database transaction.
771 * Every operation requires a transaction handle.
774 MDB_txn *mt_parent; /**< parent of a nested txn */
775 MDB_txn *mt_child; /**< nested txn under this txn */
776 pgno_t mt_next_pgno; /**< next unallocated page */
777 /** The ID of this transaction. IDs are integers incrementing from 1.
778 * Only committed write transactions increment the ID. If a transaction
779 * aborts, the ID may be re-used by the next writer.
782 MDB_env *mt_env; /**< the DB environment */
783 /** The list of pages that became unused during this transaction.
787 MDB_ID2L dirty_list; /**< modified pages */
788 MDB_reader *reader; /**< this thread's slot in the reader table */
790 /** Array of records for each DB known in the environment. */
792 /** Array of MDB_db records for each known DB */
794 /** @defgroup mt_dbflag Transaction DB Flags
798 #define DB_DIRTY 0x01 /**< DB was written in this txn */
799 #define DB_STALE 0x02 /**< DB record is older than txnID */
801 /** Array of cursors for each DB */
802 MDB_cursor **mt_cursors;
803 /** Array of flags for each DB */
804 unsigned char *mt_dbflags;
805 /** Number of DB records in use. This number only ever increments;
806 * we don't decrement it when individual DB handles are closed.
810 /** @defgroup mdb_txn Transaction Flags
814 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
815 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
817 unsigned int mt_flags; /**< @ref mdb_txn */
818 /** Tracks which of the two meta pages was used at the start
819 * of this transaction.
821 unsigned int mt_toggle;
824 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
825 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
826 * raise this on a 64 bit machine.
828 #define CURSOR_STACK 32
832 /** Cursors are used for all DB operations */
834 /** Next cursor on this DB in this txn */
836 /** Original cursor if this is a shadow */
838 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
839 struct MDB_xcursor *mc_xcursor;
840 /** The transaction that owns this cursor */
842 /** The database handle this cursor operates on */
844 /** The database record for this cursor */
846 /** The database auxiliary record for this cursor */
848 /** The @ref mt_dbflag for this database */
849 unsigned char *mc_dbflag;
850 unsigned short mc_snum; /**< number of pushed pages */
851 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
852 /** @defgroup mdb_cursor Cursor Flags
854 * Cursor state flags.
857 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
858 #define C_EOF 0x02 /**< No more data */
859 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
860 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
861 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
862 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
864 unsigned int mc_flags; /**< @ref mdb_cursor */
865 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
866 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
869 /** Context for sorted-dup records.
870 * We could have gone to a fully recursive design, with arbitrarily
871 * deep nesting of sub-databases. But for now we only handle these
872 * levels - main DB, optional sub-DB, sorted-duplicate DB.
874 typedef struct MDB_xcursor {
875 /** A sub-cursor for traversing the Dup DB */
876 MDB_cursor mx_cursor;
877 /** The database record for this Dup DB */
879 /** The auxiliary DB record for this Dup DB */
881 /** The @ref mt_dbflag for this Dup DB */
882 unsigned char mx_dbflag;
885 /** A set of pages freed by an earlier transaction. */
886 typedef struct MDB_oldpages {
887 /** Usually we only read one record from the FREEDB at a time, but
888 * in case we read more, this will chain them together.
890 struct MDB_oldpages *mo_next;
891 /** The ID of the transaction in which these pages were freed. */
893 /** An #MDB_IDL of the pages */
894 pgno_t mo_pages[1]; /* dynamic */
897 /** The database environment. */
899 HANDLE me_fd; /**< The main data file */
900 HANDLE me_lfd; /**< The lock file */
901 HANDLE me_mfd; /**< just for writing the meta pages */
902 /** Failed to update the meta page. Probably an I/O error. */
903 #define MDB_FATAL_ERROR 0x80000000U
904 uint32_t me_flags; /**< @ref mdb_env */
905 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
906 unsigned int me_maxreaders; /**< size of the reader table */
907 MDB_dbi me_numdbs; /**< number of DBs opened */
908 MDB_dbi me_maxdbs; /**< size of the DB table */
909 char *me_path; /**< path to the DB files */
910 char *me_map; /**< the memory map of the data file */
911 MDB_txninfo *me_txns; /**< the memory map of the lock file */
912 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
913 MDB_txn *me_txn; /**< current write transaction */
914 size_t me_mapsize; /**< size of the data memory map */
915 off_t me_size; /**< current file size */
916 pgno_t me_maxpg; /**< me_mapsize / me_psize */
917 txnid_t me_pgfirst; /**< ID of first old page record we used */
918 txnid_t me_pglast; /**< ID of last old page record we used */
919 MDB_dbx *me_dbxs; /**< array of static DB info */
920 uint16_t *me_dbflags; /**< array of DB flags */
921 MDB_oldpages *me_pghead; /**< list of old page records */
922 MDB_oldpages *me_pgfree; /**< list of page records to free */
923 pthread_key_t me_txkey; /**< thread-key for readers */
924 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
925 /** IDL of pages that became unused in a write txn */
927 /** ID2L of pages that were written during a write txn */
928 MDB_ID2 me_dirty_list[MDB_IDL_UM_SIZE];
930 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
932 #elif defined(MDB_USE_POSIX_SEM)
933 sem_t *me_rmutex; /* Shared mutexes are not supported */
937 /** max number of pages to commit in one writev() call */
938 #define MDB_COMMIT_PAGES 64
939 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
940 #undef MDB_COMMIT_PAGES
941 #define MDB_COMMIT_PAGES IOV_MAX
944 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
945 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
946 static int mdb_page_touch(MDB_cursor *mc);
948 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
949 static int mdb_page_search_root(MDB_cursor *mc,
950 MDB_val *key, int modify);
951 #define MDB_PS_MODIFY 1
952 #define MDB_PS_ROOTONLY 2
953 static int mdb_page_search(MDB_cursor *mc,
954 MDB_val *key, int flags);
955 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
957 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
958 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
959 pgno_t newpgno, unsigned int nflags);
961 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
962 static int mdb_env_pick_meta(const MDB_env *env);
963 static int mdb_env_write_meta(MDB_txn *txn);
965 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
966 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
967 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
968 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
969 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
970 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
971 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
972 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
973 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
975 static int mdb_rebalance(MDB_cursor *mc);
976 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
978 static void mdb_cursor_pop(MDB_cursor *mc);
979 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
981 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
982 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
983 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
984 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
985 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
987 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
988 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
990 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
991 static void mdb_xcursor_init0(MDB_cursor *mc);
992 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
994 static int mdb_drop0(MDB_cursor *mc, int subs);
995 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
998 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1002 static SECURITY_DESCRIPTOR mdb_null_sd;
1003 static SECURITY_ATTRIBUTES mdb_all_sa;
1004 static int mdb_sec_inited;
1007 /** Return the library version info. */
1009 mdb_version(int *major, int *minor, int *patch)
1011 if (major) *major = MDB_VERSION_MAJOR;
1012 if (minor) *minor = MDB_VERSION_MINOR;
1013 if (patch) *patch = MDB_VERSION_PATCH;
1014 return MDB_VERSION_STRING;
1017 /** Table of descriptions for MDB @ref errors */
1018 static char *const mdb_errstr[] = {
1019 "MDB_KEYEXIST: Key/data pair already exists",
1020 "MDB_NOTFOUND: No matching key/data pair found",
1021 "MDB_PAGE_NOTFOUND: Requested page not found",
1022 "MDB_CORRUPTED: Located page was wrong type",
1023 "MDB_PANIC: Update of meta page failed",
1024 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1025 "MDB_INVALID: File is not an MDB file",
1026 "MDB_MAP_FULL: Environment mapsize limit reached",
1027 "MDB_DBS_FULL: Environment maxdbs limit reached",
1028 "MDB_READERS_FULL: Environment maxreaders limit reached",
1029 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1030 "MDB_TXN_FULL: Nested transaction has too many dirty pages - transaction too big",
1031 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1032 "MDB_PAGE_FULL: Internal error - page has no more space"
1036 mdb_strerror(int err)
1040 return ("Successful return: 0");
1042 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1043 i = err - MDB_KEYEXIST;
1044 return mdb_errstr[i];
1047 return strerror(err);
1051 /** Display a key in hexadecimal and return the address of the result.
1052 * @param[in] key the key to display
1053 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1054 * @return The key in hexadecimal form.
1057 mdb_dkey(MDB_val *key, char *buf)
1060 unsigned char *c = key->mv_data;
1062 if (key->mv_size > MAXKEYSIZE)
1063 return "MAXKEYSIZE";
1064 /* may want to make this a dynamic check: if the key is mostly
1065 * printable characters, print it as-is instead of converting to hex.
1069 for (i=0; i<key->mv_size; i++)
1070 ptr += sprintf(ptr, "%02x", *c++);
1072 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1077 /** Display all the keys in the page. */
1079 mdb_page_list(MDB_page *mp)
1082 unsigned int i, nkeys, nsize;
1086 nkeys = NUMKEYS(mp);
1087 fprintf(stderr, "numkeys %d\n", nkeys);
1088 for (i=0; i<nkeys; i++) {
1089 node = NODEPTR(mp, i);
1090 key.mv_size = node->mn_ksize;
1091 key.mv_data = node->mn_data;
1092 nsize = NODESIZE + NODEKSZ(node) + sizeof(indx_t);
1093 if (F_ISSET(node->mn_flags, F_BIGDATA))
1094 nsize += sizeof(pgno_t);
1096 nsize += NODEDSZ(node);
1097 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1102 mdb_cursor_chk(MDB_cursor *mc)
1108 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1109 for (i=0; i<mc->mc_top; i++) {
1111 node = NODEPTR(mp, mc->mc_ki[i]);
1112 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1115 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1121 /** Count all the pages in each DB and in the freelist
1122 * and make sure it matches the actual number of pages
1125 static void mdb_audit(MDB_txn *txn)
1129 MDB_ID freecount, count;
1134 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1135 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1136 freecount += *(MDB_ID *)data.mv_data;
1139 for (i = 0; i<txn->mt_numdbs; i++) {
1140 MDB_xcursor mx, *mxp;
1141 mxp = (txn->mt_dbs[i].md_flags & MDB_DUPSORT) ? &mx : NULL;
1142 mdb_cursor_init(&mc, txn, i, mxp);
1143 if (txn->mt_dbs[i].md_root == P_INVALID)
1145 count += txn->mt_dbs[i].md_branch_pages +
1146 txn->mt_dbs[i].md_leaf_pages +
1147 txn->mt_dbs[i].md_overflow_pages;
1148 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1149 mdb_page_search(&mc, NULL, 0);
1153 mp = mc.mc_pg[mc.mc_top];
1154 for (j=0; j<NUMKEYS(mp); j++) {
1155 MDB_node *leaf = NODEPTR(mp, j);
1156 if (leaf->mn_flags & F_SUBDATA) {
1158 memcpy(&db, NODEDATA(leaf), sizeof(db));
1159 count += db.md_branch_pages + db.md_leaf_pages +
1160 db.md_overflow_pages;
1164 while (mdb_cursor_sibling(&mc, 1) == 0);
1167 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1168 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1169 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1175 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1177 return txn->mt_dbxs[dbi].md_cmp(a, b);
1181 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1183 if (txn->mt_dbxs[dbi].md_dcmp)
1184 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1186 return EINVAL; /* too bad you can't distinguish this from a valid result */
1189 /** Allocate a single page.
1190 * Re-use old malloc'd pages first, otherwise just malloc.
1193 mdb_page_malloc(MDB_cursor *mc) {
1195 size_t sz = mc->mc_txn->mt_env->me_psize;
1196 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1197 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1198 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1199 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1200 } else if ((ret = malloc(sz)) != NULL) {
1201 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1206 /** Allocate pages for writing.
1207 * If there are free pages available from older transactions, they
1208 * will be re-used first. Otherwise a new page will be allocated.
1209 * @param[in] mc cursor A cursor handle identifying the transaction and
1210 * database for which we are allocating.
1211 * @param[in] num the number of pages to allocate.
1212 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1213 * will always be satisfied by a single contiguous chunk of memory.
1214 * @return 0 on success, non-zero on failure.
1217 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1219 MDB_txn *txn = mc->mc_txn;
1221 pgno_t pgno = P_INVALID;
1226 /* The free list won't have any content at all until txn 2 has
1227 * committed. The pages freed by txn 2 will be unreferenced
1228 * after txn 3 commits, and so will be safe to re-use in txn 4.
1230 if (txn->mt_txnid > 3) {
1232 if (!txn->mt_env->me_pghead &&
1233 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1234 /* See if there's anything in the free DB */
1238 txnid_t *kptr, oldest, last;
1240 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1241 if (!txn->mt_env->me_pgfirst) {
1242 mdb_page_search(&m2, NULL, 0);
1243 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1244 kptr = (txnid_t *)NODEKEY(leaf);
1251 last = txn->mt_env->me_pglast + 1;
1253 key.mv_data = &last;
1254 key.mv_size = sizeof(last);
1255 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1258 last = *(txnid_t *)key.mv_data;
1263 oldest = txn->mt_txnid - 1;
1264 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1265 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1266 if (mr && mr < oldest)
1271 if (oldest > last) {
1272 /* It's usable, grab it.
1277 if (!txn->mt_env->me_pgfirst) {
1278 mdb_node_read(txn, leaf, &data);
1280 txn->mt_env->me_pglast = last;
1281 if (!txn->mt_env->me_pgfirst)
1282 txn->mt_env->me_pgfirst = last;
1283 idl = (MDB_ID *) data.mv_data;
1284 /* We might have a zero-length IDL due to freelist growth
1285 * during a prior commit
1287 if (!idl[0]) goto again;
1288 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1291 mop->mo_next = txn->mt_env->me_pghead;
1292 mop->mo_txnid = last;
1293 txn->mt_env->me_pghead = mop;
1294 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1299 DPRINTF("IDL read txn %zu root %zu num %zu",
1300 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1301 for (i=0; i<idl[0]; i++) {
1302 DPRINTF("IDL %zu", idl[i+1]);
1309 if (txn->mt_env->me_pghead) {
1310 MDB_oldpages *mop = txn->mt_env->me_pghead;
1312 /* FIXME: For now, always use fresh pages. We
1313 * really ought to search the free list for a
1318 /* peel pages off tail, so we only have to truncate the list */
1319 pgno = MDB_IDL_LAST(mop->mo_pages);
1320 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1322 if (mop->mo_pages[2] > mop->mo_pages[1])
1323 mop->mo_pages[0] = 0;
1327 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1328 txn->mt_env->me_pghead = mop->mo_next;
1329 if (mc->mc_dbi == FREE_DBI) {
1330 mop->mo_next = txn->mt_env->me_pgfree;
1331 txn->mt_env->me_pgfree = mop;
1340 if (pgno == P_INVALID) {
1341 /* DB size is maxed out */
1342 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1343 DPUTS("DB size maxed out");
1344 return MDB_MAP_FULL;
1347 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1348 if (pgno == P_INVALID) {
1349 pgno = txn->mt_next_pgno;
1350 txn->mt_next_pgno += num;
1352 np = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
1355 if (txn->mt_env->me_dpages && num == 1) {
1356 np = txn->mt_env->me_dpages;
1357 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1358 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1359 txn->mt_env->me_dpages = np->mp_next;
1361 size_t sz = txn->mt_env->me_psize * num;
1362 if ((np = malloc(sz)) == NULL)
1364 VGMEMP_ALLOC(txn->mt_env, np, sz);
1366 if (pgno == P_INVALID) {
1367 np->mp_pgno = txn->mt_next_pgno;
1368 txn->mt_next_pgno += num;
1373 mid.mid = np->mp_pgno;
1375 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1376 mdb_mid2l_append(txn->mt_u.dirty_list, &mid);
1378 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1385 /** Copy a page: avoid copying unused portions of the page.
1386 * @param[in] dst page to copy into
1387 * @param[in] src page to copy from
1390 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
1392 dst->mp_flags = src->mp_flags | P_DIRTY;
1393 dst->mp_pages = src->mp_pages;
1395 if (IS_LEAF2(src)) {
1396 memcpy(dst->mp_ptrs, src->mp_ptrs, psize - PAGEHDRSZ - SIZELEFT(src));
1398 unsigned int i, nkeys = NUMKEYS(src);
1399 for (i=0; i<nkeys; i++)
1400 dst->mp_ptrs[i] = src->mp_ptrs[i];
1401 memcpy((char *)dst+src->mp_upper, (char *)src+src->mp_upper,
1402 psize - src->mp_upper);
1406 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1407 * @param[in] mc cursor pointing to the page to be touched
1408 * @return 0 on success, non-zero on failure.
1411 mdb_page_touch(MDB_cursor *mc)
1413 MDB_page *mp = mc->mc_pg[mc->mc_top];
1417 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1419 if ((rc = mdb_page_alloc(mc, 1, &np)))
1421 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1422 assert(mp->mp_pgno != np->mp_pgno);
1423 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1425 /* If page isn't full, just copy the used portion */
1426 mdb_page_copy(np, mp, mc->mc_txn->mt_env->me_psize);
1429 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1431 np->mp_flags |= P_DIRTY;
1436 /* Adjust other cursors pointing to mp */
1437 if (mc->mc_flags & C_SUB) {
1438 MDB_cursor *m2, *m3;
1439 MDB_dbi dbi = mc->mc_dbi-1;
1441 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1442 if (m2 == mc) continue;
1443 m3 = &m2->mc_xcursor->mx_cursor;
1444 if (m3->mc_snum < mc->mc_snum) continue;
1445 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1446 m3->mc_pg[mc->mc_top] = mp;
1452 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1453 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1454 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1455 m2->mc_pg[mc->mc_top] = mp;
1459 mc->mc_pg[mc->mc_top] = mp;
1460 /** If this page has a parent, update the parent to point to
1464 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1466 mc->mc_db->md_root = mp->mp_pgno;
1467 } else if (mc->mc_txn->mt_parent) {
1470 /* If txn has a parent, make sure the page is in our
1473 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1474 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1475 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1476 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1477 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1478 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1479 mc->mc_pg[mc->mc_top] = mp;
1485 np = mdb_page_malloc(mc);
1488 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1489 mid.mid = np->mp_pgno;
1491 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1499 mdb_env_sync(MDB_env *env, int force)
1502 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1503 if (env->me_flags & MDB_WRITEMAP) {
1504 int flags = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
1505 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
1508 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
1512 if (MDB_FDATASYNC(env->me_fd))
1519 /** Make shadow copies of all of parent txn's cursors */
1521 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1523 MDB_cursor *mc, *m2;
1524 unsigned int i, j, size;
1526 for (i=0;i<src->mt_numdbs; i++) {
1527 if (src->mt_cursors[i]) {
1528 size = sizeof(MDB_cursor);
1529 if (src->mt_cursors[i]->mc_xcursor)
1530 size += sizeof(MDB_xcursor);
1531 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1538 mc->mc_db = &dst->mt_dbs[i];
1539 mc->mc_dbx = m2->mc_dbx;
1540 mc->mc_dbflag = &dst->mt_dbflags[i];
1541 mc->mc_snum = m2->mc_snum;
1542 mc->mc_top = m2->mc_top;
1543 mc->mc_flags = m2->mc_flags | C_SHADOW;
1544 for (j=0; j<mc->mc_snum; j++) {
1545 mc->mc_pg[j] = m2->mc_pg[j];
1546 mc->mc_ki[j] = m2->mc_ki[j];
1548 if (m2->mc_xcursor) {
1549 MDB_xcursor *mx, *mx2;
1550 mx = (MDB_xcursor *)(mc+1);
1551 mc->mc_xcursor = mx;
1552 mx2 = m2->mc_xcursor;
1553 mx->mx_db = mx2->mx_db;
1554 mx->mx_dbx = mx2->mx_dbx;
1555 mx->mx_dbflag = mx2->mx_dbflag;
1556 mx->mx_cursor.mc_txn = dst;
1557 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1558 mx->mx_cursor.mc_db = &mx->mx_db;
1559 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1560 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1561 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1562 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1563 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1564 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1565 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1566 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1569 mc->mc_xcursor = NULL;
1571 mc->mc_next = dst->mt_cursors[i];
1572 dst->mt_cursors[i] = mc;
1579 /** Merge shadow cursors back into parent's */
1581 mdb_cursor_merge(MDB_txn *txn)
1584 for (i=0; i<txn->mt_numdbs; i++) {
1585 if (txn->mt_cursors[i]) {
1587 while ((mc = txn->mt_cursors[i])) {
1588 txn->mt_cursors[i] = mc->mc_next;
1589 if (mc->mc_flags & C_SHADOW) {
1590 MDB_cursor *m2 = mc->mc_orig;
1592 m2->mc_snum = mc->mc_snum;
1593 m2->mc_top = mc->mc_top;
1594 for (j=0; j<mc->mc_snum; j++) {
1595 m2->mc_pg[j] = mc->mc_pg[j];
1596 m2->mc_ki[j] = mc->mc_ki[j];
1599 if (mc->mc_flags & C_ALLOCD)
1607 mdb_txn_reset0(MDB_txn *txn);
1609 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1610 * @param[in] txn the transaction handle to initialize
1611 * @return 0 on success, non-zero on failure. This can only
1612 * fail for read-only transactions, and then only if the
1613 * reader table is full.
1616 mdb_txn_renew0(MDB_txn *txn)
1618 MDB_env *env = txn->mt_env;
1622 txn->mt_numdbs = env->me_numdbs;
1623 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1625 if (txn->mt_flags & MDB_TXN_RDONLY) {
1626 MDB_reader *r = pthread_getspecific(env->me_txkey);
1628 pid_t pid = getpid();
1629 pthread_t tid = pthread_self();
1632 for (i=0; i<env->me_txns->mti_numreaders; i++)
1633 if (env->me_txns->mti_readers[i].mr_pid == 0)
1635 if (i == env->me_maxreaders) {
1636 UNLOCK_MUTEX_R(env);
1637 return MDB_READERS_FULL;
1639 env->me_txns->mti_readers[i].mr_pid = pid;
1640 env->me_txns->mti_readers[i].mr_tid = tid;
1641 if (i >= env->me_txns->mti_numreaders)
1642 env->me_txns->mti_numreaders = i+1;
1643 UNLOCK_MUTEX_R(env);
1644 r = &env->me_txns->mti_readers[i];
1645 pthread_setspecific(env->me_txkey, r);
1647 txn->mt_txnid = r->mr_txnid = env->me_txns->mti_txnid;
1648 txn->mt_toggle = txn->mt_txnid & 1;
1649 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1650 txn->mt_u.reader = r;
1654 txn->mt_txnid = env->me_txns->mti_txnid;
1655 txn->mt_toggle = txn->mt_txnid & 1;
1656 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1659 if (txn->mt_txnid == mdb_debug_start)
1662 txn->mt_u.dirty_list = env->me_dirty_list;
1663 txn->mt_u.dirty_list[0].mid = 0;
1664 txn->mt_free_pgs = env->me_free_pgs;
1665 txn->mt_free_pgs[0] = 0;
1669 /* Copy the DB info and flags */
1670 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1671 for (i=2; i<txn->mt_numdbs; i++)
1672 txn->mt_dbs[i].md_flags = env->me_dbflags[i];
1673 txn->mt_dbflags[0] = txn->mt_dbflags[1] = 0;
1674 if (txn->mt_numdbs > 2)
1675 memset(txn->mt_dbflags+2, DB_STALE, txn->mt_numdbs-2);
1681 mdb_txn_renew(MDB_txn *txn)
1688 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1689 DPUTS("environment had fatal error, must shutdown!");
1693 rc = mdb_txn_renew0(txn);
1694 if (rc == MDB_SUCCESS) {
1695 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1696 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1697 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1703 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1708 if (env->me_flags & MDB_FATAL_ERROR) {
1709 DPUTS("environment had fatal error, must shutdown!");
1712 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
1715 /* parent already has an active child txn */
1716 if (parent->mt_child) {
1719 /* nested TXNs not supported here */
1720 if (env->me_flags & MDB_WRITEMAP)
1723 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1724 if (!(flags & MDB_RDONLY))
1725 size += env->me_maxdbs * sizeof(MDB_cursor *);
1727 if ((txn = calloc(1, size)) == NULL) {
1728 DPRINTF("calloc: %s", strerror(ErrCode()));
1731 txn->mt_dbs = (MDB_db *)(txn+1);
1732 if (flags & MDB_RDONLY) {
1733 txn->mt_flags |= MDB_TXN_RDONLY;
1734 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1736 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1737 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1742 txn->mt_free_pgs = mdb_midl_alloc();
1743 if (!txn->mt_free_pgs) {
1747 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
1748 if (!txn->mt_u.dirty_list) {
1749 free(txn->mt_free_pgs);
1753 txn->mt_txnid = parent->mt_txnid;
1754 txn->mt_toggle = parent->mt_toggle;
1755 txn->mt_u.dirty_list[0].mid = 0;
1756 txn->mt_free_pgs[0] = 0;
1757 txn->mt_next_pgno = parent->mt_next_pgno;
1758 parent->mt_child = txn;
1759 txn->mt_parent = parent;
1760 txn->mt_numdbs = parent->mt_numdbs;
1761 txn->mt_dbxs = parent->mt_dbxs;
1762 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1763 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1764 mdb_cursor_shadow(parent, txn);
1767 rc = mdb_txn_renew0(txn);
1773 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1774 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1775 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1781 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1782 * @param[in] txn the transaction handle to reset
1785 mdb_txn_reset0(MDB_txn *txn)
1787 MDB_env *env = txn->mt_env;
1789 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1790 txn->mt_u.reader->mr_txnid = 0;
1796 /* close(free) all cursors */
1797 for (i=0; i<txn->mt_numdbs; i++) {
1798 if (txn->mt_cursors[i]) {
1800 while ((mc = txn->mt_cursors[i])) {
1801 txn->mt_cursors[i] = mc->mc_next;
1802 if (mc->mc_flags & C_ALLOCD)
1808 if (!(env->me_flags & MDB_WRITEMAP)) {
1809 /* return all dirty pages to dpage list */
1810 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1811 dp = txn->mt_u.dirty_list[i].mptr;
1812 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1813 dp->mp_next = txn->mt_env->me_dpages;
1814 VGMEMP_FREE(txn->mt_env, dp);
1815 txn->mt_env->me_dpages = dp;
1817 /* large pages just get freed directly */
1818 VGMEMP_FREE(txn->mt_env, dp);
1824 if (txn->mt_parent) {
1825 txn->mt_parent->mt_child = NULL;
1826 mdb_midl_free(txn->mt_free_pgs);
1827 free(txn->mt_u.dirty_list);
1830 if (mdb_midl_shrink(&txn->mt_free_pgs))
1831 env->me_free_pgs = txn->mt_free_pgs;
1834 while ((mop = txn->mt_env->me_pghead)) {
1835 txn->mt_env->me_pghead = mop->mo_next;
1838 txn->mt_env->me_pgfirst = 0;
1839 txn->mt_env->me_pglast = 0;
1842 /* The writer mutex was locked in mdb_txn_begin. */
1843 UNLOCK_MUTEX_W(env);
1848 mdb_txn_reset(MDB_txn *txn)
1853 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1854 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1855 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1857 mdb_txn_reset0(txn);
1861 mdb_txn_abort(MDB_txn *txn)
1866 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1867 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1868 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1871 mdb_txn_abort(txn->mt_child);
1873 mdb_txn_reset0(txn);
1878 mdb_txn_commit(MDB_txn *txn)
1886 pgno_t next, freecnt;
1889 assert(txn != NULL);
1890 assert(txn->mt_env != NULL);
1892 if (txn->mt_child) {
1893 mdb_txn_commit(txn->mt_child);
1894 txn->mt_child = NULL;
1899 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1900 if (txn->mt_numdbs > env->me_numdbs) {
1901 /* update the DB flags */
1903 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
1904 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
1911 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1912 DPUTS("error flag is set, can't commit");
1914 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1919 /* Merge (and close) our cursors with parent's */
1920 mdb_cursor_merge(txn);
1922 if (txn->mt_parent) {
1928 /* Update parent's DB table */
1929 ip = &txn->mt_parent->mt_dbs[2];
1930 jp = &txn->mt_dbs[2];
1931 for (i = 2; i < txn->mt_numdbs; i++) {
1932 if (ip->md_root != jp->md_root)
1936 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1938 /* Append our free list to parent's */
1939 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1941 mdb_midl_free(txn->mt_free_pgs);
1943 /* Merge our dirty list with parent's */
1944 dst = txn->mt_parent->mt_u.dirty_list;
1945 src = txn->mt_u.dirty_list;
1946 x = mdb_mid2l_search(dst, src[1].mid);
1947 for (y=1; y<=src[0].mid; y++) {
1948 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1952 dst[x].mptr = src[y].mptr;
1955 for (; y<=src[0].mid; y++) {
1956 if (++x >= MDB_IDL_UM_MAX) {
1958 return MDB_TXN_FULL;
1963 free(txn->mt_u.dirty_list);
1964 txn->mt_parent->mt_child = NULL;
1969 if (txn != env->me_txn) {
1970 DPUTS("attempt to commit unknown transaction");
1975 if (!txn->mt_u.dirty_list[0].mid)
1978 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1979 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1981 /* Update DB root pointers. Their pages have already been
1982 * touched so this is all in-place and cannot fail.
1984 if (txn->mt_numdbs > 2) {
1987 data.mv_size = sizeof(MDB_db);
1989 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
1990 for (i = 2; i < txn->mt_numdbs; i++) {
1991 if (txn->mt_dbflags[i] & DB_DIRTY) {
1992 data.mv_data = &txn->mt_dbs[i];
1993 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
1998 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2000 /* should only be one record now */
2001 if (env->me_pghead) {
2002 /* make sure first page of freeDB is touched and on freelist */
2003 mdb_page_search(&mc, NULL, MDB_PS_MODIFY);
2006 /* Delete IDLs we used from the free list */
2007 if (env->me_pgfirst) {
2012 key.mv_size = sizeof(cur);
2013 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
2016 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
2017 rc = mdb_cursor_del(&mc, 0);
2023 env->me_pgfirst = 0;
2027 /* save to free list */
2029 freecnt = txn->mt_free_pgs[0];
2030 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2033 /* make sure last page of freeDB is touched and on freelist */
2034 key.mv_size = MAXKEYSIZE+1;
2036 mdb_page_search(&mc, &key, MDB_PS_MODIFY);
2038 mdb_midl_sort(txn->mt_free_pgs);
2042 MDB_IDL idl = txn->mt_free_pgs;
2043 DPRINTF("IDL write txn %zu root %zu num %zu",
2044 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
2045 for (i=0; i<idl[0]; i++) {
2046 DPRINTF("IDL %zu", idl[i+1]);
2050 /* write to last page of freeDB */
2051 key.mv_size = sizeof(pgno_t);
2052 key.mv_data = &txn->mt_txnid;
2053 data.mv_data = txn->mt_free_pgs;
2054 /* The free list can still grow during this call,
2055 * despite the pre-emptive touches above. So check
2056 * and make sure the entire thing got written.
2059 freecnt = txn->mt_free_pgs[0];
2060 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
2061 rc = mdb_cursor_put(&mc, &key, &data, 0);
2066 } while (freecnt != txn->mt_free_pgs[0]);
2068 /* should only be one record now */
2070 if (env->me_pghead) {
2076 mop = env->me_pghead;
2078 key.mv_size = sizeof(id);
2080 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2081 data.mv_data = mop->mo_pages;
2082 orig = mop->mo_pages[0];
2083 /* These steps may grow the freelist again
2084 * due to freed overflow pages...
2086 mdb_cursor_put(&mc, &key, &data, 0);
2087 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2088 /* could have been used again here */
2089 if (mop->mo_pages[0] != orig) {
2090 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2091 data.mv_data = mop->mo_pages;
2093 mdb_cursor_put(&mc, &key, &data, 0);
2095 env->me_pghead = NULL;
2098 /* was completely used up */
2099 mdb_cursor_del(&mc, 0);
2103 env->me_pgfirst = 0;
2107 while (env->me_pgfree) {
2108 MDB_oldpages *mop = env->me_pgfree;
2109 env->me_pgfree = mop->mo_next;
2113 /* Check for growth of freelist again */
2114 if (freecnt != txn->mt_free_pgs[0])
2117 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2118 if (mdb_midl_shrink(&txn->mt_free_pgs))
2119 env->me_free_pgs = txn->mt_free_pgs;
2126 if (env->me_flags & MDB_WRITEMAP) {
2127 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2128 dp = txn->mt_u.dirty_list[i].mptr;
2129 /* clear dirty flag */
2130 dp->mp_flags &= ~P_DIRTY;
2131 txn->mt_u.dirty_list[i].mid = 0;
2133 txn->mt_u.dirty_list[0].mid = 0;
2137 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2143 /* Windows actually supports scatter/gather I/O, but only on
2144 * unbuffered file handles. Since we're relying on the OS page
2145 * cache for all our data, that's self-defeating. So we just
2146 * write pages one at a time. We use the ov structure to set
2147 * the write offset, to at least save the overhead of a Seek
2151 memset(&ov, 0, sizeof(ov));
2152 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2154 dp = txn->mt_u.dirty_list[i].mptr;
2155 DPRINTF("committing page %zu", dp->mp_pgno);
2156 size = dp->mp_pgno * env->me_psize;
2157 ov.Offset = size & 0xffffffff;
2158 ov.OffsetHigh = size >> 16;
2159 ov.OffsetHigh >>= 16;
2160 /* clear dirty flag */
2161 dp->mp_flags &= ~P_DIRTY;
2162 wsize = env->me_psize;
2163 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2164 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2167 DPRINTF("WriteFile: %d", n);
2174 struct iovec iov[MDB_COMMIT_PAGES];
2178 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2179 dp = txn->mt_u.dirty_list[i].mptr;
2180 if (dp->mp_pgno != next) {
2182 rc = writev(env->me_fd, iov, n);
2186 DPUTS("short write, filesystem full?");
2188 DPRINTF("writev: %s", strerror(n));
2195 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2198 DPRINTF("committing page %zu", dp->mp_pgno);
2199 iov[n].iov_len = env->me_psize;
2200 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2201 iov[n].iov_base = (char *)dp;
2202 size += iov[n].iov_len;
2203 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2204 /* clear dirty flag */
2205 dp->mp_flags &= ~P_DIRTY;
2206 if (++n >= MDB_COMMIT_PAGES) {
2216 rc = writev(env->me_fd, iov, n);
2220 DPUTS("short write, filesystem full?");
2222 DPRINTF("writev: %s", strerror(n));
2229 /* Drop the dirty pages.
2231 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2232 dp = txn->mt_u.dirty_list[i].mptr;
2233 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2234 dp->mp_next = txn->mt_env->me_dpages;
2235 VGMEMP_FREE(txn->mt_env, dp);
2236 txn->mt_env->me_dpages = dp;
2238 VGMEMP_FREE(txn->mt_env, dp);
2241 txn->mt_u.dirty_list[i].mid = 0;
2243 txn->mt_u.dirty_list[0].mid = 0;
2246 if ((n = mdb_env_sync(env, 0)) != 0 ||
2247 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2254 if (txn->mt_numdbs > env->me_numdbs) {
2255 /* update the DB flags */
2257 for (i = env->me_numdbs; i<txn->mt_numdbs; i++)
2258 env->me_dbflags[i] = txn->mt_dbs[i].md_flags;
2262 UNLOCK_MUTEX_W(env);
2268 /** Read the environment parameters of a DB environment before
2269 * mapping it into memory.
2270 * @param[in] env the environment handle
2271 * @param[out] meta address of where to store the meta information
2272 * @return 0 on success, non-zero on failure.
2275 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2282 /* We don't know the page size yet, so use a minimum value.
2286 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2288 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2293 else if (rc != MDB_PAGESIZE) {
2297 DPRINTF("read: %s", strerror(err));
2301 p = (MDB_page *)&pbuf;
2303 if (!F_ISSET(p->mp_flags, P_META)) {
2304 DPRINTF("page %zu not a meta page", p->mp_pgno);
2309 if (m->mm_magic != MDB_MAGIC) {
2310 DPUTS("meta has invalid magic");
2314 if (m->mm_version != MDB_VERSION) {
2315 DPRINTF("database is version %u, expected version %u",
2316 m->mm_version, MDB_VERSION);
2317 return MDB_VERSION_MISMATCH;
2320 memcpy(meta, m, sizeof(*m));
2324 /** Write the environment parameters of a freshly created DB environment.
2325 * @param[in] env the environment handle
2326 * @param[out] meta address of where to store the meta information
2327 * @return 0 on success, non-zero on failure.
2330 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2337 DPUTS("writing new meta page");
2339 GET_PAGESIZE(psize);
2341 meta->mm_magic = MDB_MAGIC;
2342 meta->mm_version = MDB_VERSION;
2343 meta->mm_psize = psize;
2344 meta->mm_last_pg = 1;
2345 meta->mm_flags = env->me_flags & 0xffff;
2346 meta->mm_flags |= MDB_INTEGERKEY;
2347 meta->mm_dbs[0].md_root = P_INVALID;
2348 meta->mm_dbs[1].md_root = P_INVALID;
2350 p = calloc(2, psize);
2352 p->mp_flags = P_META;
2355 memcpy(m, meta, sizeof(*meta));
2357 q = (MDB_page *)((char *)p + psize);
2360 q->mp_flags = P_META;
2363 memcpy(m, meta, sizeof(*meta));
2368 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2369 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2372 rc = write(env->me_fd, p, psize * 2);
2373 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2379 /** Update the environment info to commit a transaction.
2380 * @param[in] txn the transaction that's being committed
2381 * @return 0 on success, non-zero on failure.
2384 mdb_env_write_meta(MDB_txn *txn)
2387 MDB_meta meta, metab;
2389 int rc, len, toggle;
2395 assert(txn != NULL);
2396 assert(txn->mt_env != NULL);
2398 toggle = !txn->mt_toggle;
2399 DPRINTF("writing meta page %d for root page %zu",
2400 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2404 if (env->me_flags & MDB_WRITEMAP) {
2405 MDB_meta *mp = env->me_metas[toggle];
2406 mp->mm_dbs[0] = txn->mt_dbs[0];
2407 mp->mm_dbs[1] = txn->mt_dbs[1];
2408 mp->mm_last_pg = txn->mt_next_pgno - 1;
2409 mp->mm_txnid = txn->mt_txnid;
2410 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
2411 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
2414 ptr += env->me_psize;
2415 if (MDB_MSYNC(ptr, env->me_psize, rc)) {
2422 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2423 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2425 ptr = (char *)&meta;
2426 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2427 len = sizeof(MDB_meta) - off;
2430 meta.mm_dbs[0] = txn->mt_dbs[0];
2431 meta.mm_dbs[1] = txn->mt_dbs[1];
2432 meta.mm_last_pg = txn->mt_next_pgno - 1;
2433 meta.mm_txnid = txn->mt_txnid;
2436 off += env->me_psize;
2439 /* Write to the SYNC fd */
2442 memset(&ov, 0, sizeof(ov));
2444 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2447 rc = pwrite(env->me_mfd, ptr, len, off);
2452 DPUTS("write failed, disk error?");
2453 /* On a failure, the pagecache still contains the new data.
2454 * Write some old data back, to prevent it from being used.
2455 * Use the non-SYNC fd; we know it will fail anyway.
2457 meta.mm_last_pg = metab.mm_last_pg;
2458 meta.mm_txnid = metab.mm_txnid;
2460 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2462 r2 = pwrite(env->me_fd, ptr, len, off);
2465 env->me_flags |= MDB_FATAL_ERROR;
2469 /* Memory ordering issues are irrelevant; since the entire writer
2470 * is wrapped by wmutex, all of these changes will become visible
2471 * after the wmutex is unlocked. Since the DB is multi-version,
2472 * readers will get consistent data regardless of how fresh or
2473 * how stale their view of these values is.
2475 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2480 /** Check both meta pages to see which one is newer.
2481 * @param[in] env the environment handle
2482 * @return meta toggle (0 or 1).
2485 mdb_env_pick_meta(const MDB_env *env)
2487 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
2491 mdb_env_create(MDB_env **env)
2495 e = calloc(1, sizeof(MDB_env));
2499 e->me_free_pgs = mdb_midl_alloc();
2500 if (!e->me_free_pgs) {
2504 e->me_maxreaders = DEFAULT_READERS;
2506 e->me_fd = INVALID_HANDLE_VALUE;
2507 e->me_lfd = INVALID_HANDLE_VALUE;
2508 e->me_mfd = INVALID_HANDLE_VALUE;
2509 VGMEMP_CREATE(e,0,0);
2515 mdb_env_set_mapsize(MDB_env *env, size_t size)
2519 env->me_mapsize = size;
2521 env->me_maxpg = env->me_mapsize / env->me_psize;
2526 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2530 env->me_maxdbs = dbs;
2535 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2537 if (env->me_map || readers < 1)
2539 env->me_maxreaders = readers;
2544 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2546 if (!env || !readers)
2548 *readers = env->me_maxreaders;
2552 /** Further setup required for opening an MDB environment
2555 mdb_env_open2(MDB_env *env, unsigned int flags)
2557 int i, newenv = 0, prot;
2561 env->me_flags = flags;
2563 memset(&meta, 0, sizeof(meta));
2565 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2568 DPUTS("new mdbenv");
2572 if (!env->me_mapsize) {
2573 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2579 LONG sizelo, sizehi;
2580 sizelo = env->me_mapsize & 0xffffffff;
2581 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2583 /* Windows won't create mappings for zero length files.
2584 * Just allocate the maxsize right now.
2587 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2588 if (!SetEndOfFile(env->me_fd))
2590 SetFilePointer(env->me_fd, 0, NULL, 0);
2592 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
2593 PAGE_READWRITE : PAGE_READONLY,
2594 sizehi, sizelo, NULL);
2597 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
2598 FILE_MAP_WRITE : FILE_MAP_READ,
2599 0, 0, env->me_mapsize, meta.mm_address);
2606 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2609 if (flags & MDB_WRITEMAP) {
2611 ftruncate(env->me_fd, env->me_mapsize);
2613 env->me_map = mmap(meta.mm_address, env->me_mapsize, prot, i,
2615 if (env->me_map == MAP_FAILED) {
2622 meta.mm_mapsize = env->me_mapsize;
2623 if (flags & MDB_FIXEDMAP)
2624 meta.mm_address = env->me_map;
2625 i = mdb_env_init_meta(env, &meta);
2626 if (i != MDB_SUCCESS) {
2627 munmap(env->me_map, env->me_mapsize);
2631 env->me_psize = meta.mm_psize;
2633 env->me_maxpg = env->me_mapsize / env->me_psize;
2635 p = (MDB_page *)env->me_map;
2636 env->me_metas[0] = METADATA(p);
2637 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2641 int toggle = mdb_env_pick_meta(env);
2642 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
2644 DPRINTF("opened database version %u, pagesize %u",
2645 env->me_metas[0]->mm_version, env->me_psize);
2646 DPRINTF("using meta page %d", toggle);
2647 DPRINTF("depth: %u", db->md_depth);
2648 DPRINTF("entries: %zu", db->md_entries);
2649 DPRINTF("branch pages: %zu", db->md_branch_pages);
2650 DPRINTF("leaf pages: %zu", db->md_leaf_pages);
2651 DPRINTF("overflow pages: %zu", db->md_overflow_pages);
2652 DPRINTF("root: %zu", db->md_root);
2660 /** Release a reader thread's slot in the reader lock table.
2661 * This function is called automatically when a thread exits.
2662 * @param[in] ptr This points to the slot in the reader lock table.
2665 mdb_env_reader_dest(void *ptr)
2667 MDB_reader *reader = ptr;
2669 reader->mr_txnid = 0;
2675 /** Junk for arranging thread-specific callbacks on Windows. This is
2676 * necessarily platform and compiler-specific. Windows supports up
2677 * to 1088 keys. Let's assume nobody opens more than 64 environments
2678 * in a single process, for now. They can override this if needed.
2680 #ifndef MAX_TLS_KEYS
2681 #define MAX_TLS_KEYS 64
2683 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
2684 static int mdb_tls_nkeys;
2686 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
2690 case DLL_PROCESS_ATTACH: break;
2691 case DLL_THREAD_ATTACH: break;
2692 case DLL_THREAD_DETACH:
2693 for (i=0; i<mdb_tls_nkeys; i++) {
2694 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
2695 mdb_env_reader_dest(r);
2698 case DLL_PROCESS_DETACH: break;
2703 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2705 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
2709 /* Force some symbol references.
2710 * _tls_used forces the linker to create the TLS directory if not already done
2711 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
2713 #pragma comment(linker, "/INCLUDE:_tls_used")
2714 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
2715 #pragma const_seg(".CRT$XLB")
2716 extern const PIMAGE_TLS_CALLBACK mdb_tls_callback;
2717 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2720 #pragma comment(linker, "/INCLUDE:__tls_used")
2721 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
2722 #pragma data_seg(".CRT$XLB")
2723 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
2725 #endif /* WIN 32/64 */
2726 #endif /* !__GNUC__ */
2729 /** Downgrade the exclusive lock on the region back to shared */
2731 mdb_env_share_locks(MDB_env *env)
2733 int toggle = mdb_env_pick_meta(env);
2735 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2740 /* First acquire a shared lock. The Unlock will
2741 * then release the existing exclusive lock.
2743 memset(&ov, 0, sizeof(ov));
2744 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2745 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2749 struct flock lock_info;
2750 /* The shared lock replaces the existing lock */
2751 memset((void *)&lock_info, 0, sizeof(lock_info));
2752 lock_info.l_type = F_RDLCK;
2753 lock_info.l_whence = SEEK_SET;
2754 lock_info.l_start = 0;
2755 lock_info.l_len = 1;
2756 fcntl(env->me_lfd, F_SETLK, &lock_info);
2762 mdb_env_excl_lock(MDB_env *env, int *excl)
2765 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2769 memset(&ov, 0, sizeof(ov));
2770 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2775 struct flock lock_info;
2776 memset((void *)&lock_info, 0, sizeof(lock_info));
2777 lock_info.l_type = F_WRLCK;
2778 lock_info.l_whence = SEEK_SET;
2779 lock_info.l_start = 0;
2780 lock_info.l_len = 1;
2781 if (!fcntl(env->me_lfd, F_SETLK, &lock_info)) {
2784 lock_info.l_type = F_RDLCK;
2785 if (fcntl(env->me_lfd, F_SETLKW, &lock_info)) {
2793 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
2795 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2797 * @(#) $Revision: 5.1 $
2798 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2799 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2801 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2805 * Please do not copyright this code. This code is in the public domain.
2807 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2808 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2809 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2810 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2811 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2812 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2813 * PERFORMANCE OF THIS SOFTWARE.
2816 * chongo <Landon Curt Noll> /\oo/\
2817 * http://www.isthe.com/chongo/
2819 * Share and Enjoy! :-)
2822 typedef unsigned long long mdb_hash_t;
2823 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2825 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2826 * @param[in] str string to hash
2827 * @param[in] hval initial value for hash
2828 * @return 64 bit hash
2830 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2831 * hval arg on the first call.
2834 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
2836 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
2837 unsigned char *end = s + val->mv_size;
2839 * FNV-1a hash each octet of the string
2842 /* xor the bottom with the current octet */
2843 hval ^= (mdb_hash_t)*s++;
2845 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2846 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2847 (hval << 7) + (hval << 8) + (hval << 40);
2849 /* return our new hash value */
2853 /** Hash the string and output the hash in hex.
2854 * @param[in] str string to hash
2855 * @param[out] hexbuf an array of 17 chars to hold the hash
2858 mdb_hash_hex(MDB_val *val, char *hexbuf)
2861 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
2862 for (i=0; i<8; i++) {
2863 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2869 /** Open and/or initialize the lock region for the environment.
2870 * @param[in] env The MDB environment.
2871 * @param[in] lpath The pathname of the file used for the lock region.
2872 * @param[in] mode The Unix permissions for the file, if we create it.
2873 * @param[out] excl Set to true if we got an exclusive lock on the region.
2874 * @return 0 on success, non-zero on failure.
2877 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2885 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2886 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2887 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2891 /* Try to get exclusive lock. If we succeed, then
2892 * nobody is using the lock region and we should initialize it.
2894 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2895 size = GetFileSize(env->me_lfd, NULL);
2901 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1)
2903 /* Lose record locks when exec*() */
2904 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
2905 fcntl(env->me_lfd, F_SETFD, fdflags);
2907 #else /* O_CLOEXEC on Linux: Open file and set FD_CLOEXEC atomically */
2908 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT|O_CLOEXEC, mode)) == -1)
2912 /* Try to get exclusive lock. If we succeed, then
2913 * nobody is using the lock region and we should initialize it.
2915 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
2917 size = lseek(env->me_lfd, 0, SEEK_END);
2919 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2920 if (size < rsize && *excl) {
2922 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2923 if (!SetEndOfFile(env->me_lfd)) {
2928 if (ftruncate(env->me_lfd, rsize) != 0) {
2935 size = rsize - sizeof(MDB_txninfo);
2936 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2941 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2947 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2949 if (!env->me_txns) {
2954 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2956 if (m == MAP_FAILED) {
2957 env->me_txns = NULL;
2966 BY_HANDLE_FILE_INFORMATION stbuf;
2975 if (!mdb_sec_inited) {
2976 InitializeSecurityDescriptor(&mdb_null_sd,
2977 SECURITY_DESCRIPTOR_REVISION);
2978 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2979 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2980 mdb_all_sa.bInheritHandle = FALSE;
2981 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2984 GetFileInformationByHandle(env->me_lfd, &stbuf);
2985 idbuf.volume = stbuf.dwVolumeSerialNumber;
2986 idbuf.nhigh = stbuf.nFileIndexHigh;
2987 idbuf.nlow = stbuf.nFileIndexLow;
2988 val.mv_data = &idbuf;
2989 val.mv_size = sizeof(idbuf);
2990 mdb_hash_hex(&val, hexbuf);
2991 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2992 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2993 if (!env->me_rmutex) {
2997 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2998 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2999 if (!env->me_wmutex) {
3003 #elif defined(MDB_USE_POSIX_SEM)
3012 fstat(env->me_lfd, &stbuf);
3013 idbuf.dev = stbuf.st_dev;
3014 idbuf.ino = stbuf.st_ino;
3015 val.mv_data = &idbuf;
3016 val.mv_size = sizeof(idbuf);
3017 mdb_hash_hex(&val, hexbuf);
3018 sprintf(env->me_txns->mti_rmname, "/MDBr%s", hexbuf);
3019 if (sem_unlink(env->me_txns->mti_rmname)) {
3021 if (rc != ENOENT && rc != EINVAL)
3024 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
3025 if (env->me_rmutex == SEM_FAILED) {
3029 sprintf(env->me_txns->mti_wmname, "/MDBw%s", hexbuf);
3030 if (sem_unlink(env->me_txns->mti_wmname)) {
3032 if (rc != ENOENT && rc != EINVAL)
3035 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
3036 if (env->me_wmutex == SEM_FAILED) {
3040 #else /* MDB_USE_POSIX_SEM */
3041 pthread_mutexattr_t mattr;
3043 pthread_mutexattr_init(&mattr);
3044 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
3048 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
3049 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
3050 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
3051 env->me_txns->mti_version = MDB_VERSION;
3052 env->me_txns->mti_magic = MDB_MAGIC;
3053 env->me_txns->mti_txnid = 0;
3054 env->me_txns->mti_numreaders = 0;
3057 if (env->me_txns->mti_magic != MDB_MAGIC) {
3058 DPUTS("lock region has invalid magic");
3062 if (env->me_txns->mti_version != MDB_VERSION) {
3063 DPRINTF("lock region is version %u, expected version %u",
3064 env->me_txns->mti_version, MDB_VERSION);
3065 rc = MDB_VERSION_MISMATCH;
3069 if (rc != EACCES && rc != EAGAIN) {
3073 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
3074 if (!env->me_rmutex) {
3078 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
3079 if (!env->me_wmutex) {
3083 #elif defined(MDB_USE_POSIX_SEM)
3084 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
3085 if (env->me_rmutex == SEM_FAILED) {
3089 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
3090 if (env->me_wmutex == SEM_FAILED) {
3100 env->me_lfd = INVALID_HANDLE_VALUE;
3105 /** The name of the lock file in the DB environment */
3106 #define LOCKNAME "/lock.mdb"
3107 /** The name of the data file in the DB environment */
3108 #define DATANAME "/data.mdb"
3109 /** The suffix of the lock file when no subdir is used */
3110 #define LOCKSUFF "-lock"
3113 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
3115 int oflags, rc, len, excl;
3116 char *lpath, *dpath;
3119 if (flags & MDB_NOSUBDIR) {
3120 rc = len + sizeof(LOCKSUFF) + len + 1;
3122 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
3127 if (flags & MDB_NOSUBDIR) {
3128 dpath = lpath + len + sizeof(LOCKSUFF);
3129 sprintf(lpath, "%s" LOCKSUFF, path);
3130 strcpy(dpath, path);
3132 dpath = lpath + len + sizeof(LOCKNAME);
3133 sprintf(lpath, "%s" LOCKNAME, path);
3134 sprintf(dpath, "%s" DATANAME, path);
3137 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
3142 if (F_ISSET(flags, MDB_RDONLY)) {
3143 oflags = GENERIC_READ;
3144 len = OPEN_EXISTING;
3146 oflags = GENERIC_READ|GENERIC_WRITE;
3149 mode = FILE_ATTRIBUTE_NORMAL;
3150 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
3151 NULL, len, mode, NULL);
3153 if (F_ISSET(flags, MDB_RDONLY))
3156 oflags = O_RDWR | O_CREAT;
3158 env->me_fd = open(dpath, oflags, mode);
3160 if (env->me_fd == INVALID_HANDLE_VALUE) {
3165 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
3166 if (flags & (MDB_RDONLY|MDB_NOSYNC|MDB_NOMETASYNC|MDB_WRITEMAP)) {
3167 env->me_mfd = env->me_fd;
3169 /* synchronous fd for meta writes */
3171 env->me_mfd = CreateFile(dpath, oflags,
3172 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
3173 mode | FILE_FLAG_WRITE_THROUGH, NULL);
3175 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
3177 if (env->me_mfd == INVALID_HANDLE_VALUE) {
3182 env->me_path = strdup(path);
3183 DPRINTF("opened dbenv %p", (void *) env);
3184 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3186 /* Windows TLS callbacks need help finding their TLS info. */
3187 if (mdb_tls_nkeys < MAX_TLS_KEYS)
3188 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
3195 mdb_env_share_locks(env);
3197 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3198 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
3199 if (!env->me_dbxs || !env->me_dbflags)
3205 if (env->me_fd != INVALID_HANDLE_VALUE) {
3207 env->me_fd = INVALID_HANDLE_VALUE;
3209 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3211 env->me_lfd = INVALID_HANDLE_VALUE;
3219 mdb_env_close(MDB_env *env)
3226 VGMEMP_DESTROY(env);
3227 while (env->me_dpages) {
3228 dp = env->me_dpages;
3229 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3230 env->me_dpages = dp->mp_next;
3234 free(env->me_dbflags);
3238 pthread_key_delete(env->me_txkey);
3240 /* Delete our key from the global list */
3242 for (i=0; i<mdb_tls_nkeys; i++)
3243 if (mdb_tls_keys[i] == env->me_txkey) {
3244 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
3252 munmap(env->me_map, env->me_mapsize);
3254 if (env->me_mfd != env->me_fd)
3258 pid_t pid = getpid();
3260 for (i=0; i<env->me_txns->mti_numreaders; i++)
3261 if (env->me_txns->mti_readers[i].mr_pid == pid)
3262 env->me_txns->mti_readers[i].mr_pid = 0;
3264 CloseHandle(env->me_rmutex);
3265 CloseHandle(env->me_wmutex);
3266 /* Windows automatically destroys the mutexes when
3267 * the last handle closes.
3269 #elif defined(MDB_USE_POSIX_SEM)
3270 sem_close(env->me_rmutex);
3271 sem_close(env->me_wmutex);
3273 if (!mdb_env_excl_lock(env, &excl) && excl) {
3274 /* we are the only remaining user of the environment.
3275 clean up semaphores. */
3276 sem_unlink(env->me_txns->mti_rmname);
3277 sem_unlink(env->me_txns->mti_wmname);
3281 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3284 mdb_midl_free(env->me_free_pgs);
3288 /** Compare two items pointing at aligned size_t's */
3290 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3292 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3293 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3296 /** Compare two items pointing at aligned int's */
3298 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3300 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3301 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3304 /** Compare two items pointing at ints of unknown alignment.
3305 * Nodes and keys are guaranteed to be 2-byte aligned.
3308 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3310 #if BYTE_ORDER == LITTLE_ENDIAN
3311 unsigned short *u, *c;
3314 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3315 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3318 } while(!x && u > (unsigned short *)a->mv_data);
3321 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3325 /** Compare two items lexically */
3327 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3334 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3340 diff = memcmp(a->mv_data, b->mv_data, len);
3341 return diff ? diff : len_diff<0 ? -1 : len_diff;
3344 /** Compare two items in reverse byte order */
3346 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3348 const unsigned char *p1, *p2, *p1_lim;
3352 p1_lim = (const unsigned char *)a->mv_data;
3353 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3354 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3356 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3362 while (p1 > p1_lim) {
3363 diff = *--p1 - *--p2;
3367 return len_diff<0 ? -1 : len_diff;
3370 /** Search for key within a page, using binary search.
3371 * Returns the smallest entry larger or equal to the key.
3372 * If exactp is non-null, stores whether the found entry was an exact match
3373 * in *exactp (1 or 0).
3374 * Updates the cursor index with the index of the found entry.
3375 * If no entry larger or equal to the key is found, returns NULL.
3378 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3380 unsigned int i = 0, nkeys;
3383 MDB_page *mp = mc->mc_pg[mc->mc_top];
3384 MDB_node *node = NULL;
3389 nkeys = NUMKEYS(mp);
3394 COPY_PGNO(pgno, mp->mp_pgno);
3395 DPRINTF("searching %u keys in %s %spage %zu",
3396 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3403 low = IS_LEAF(mp) ? 0 : 1;
3405 cmp = mc->mc_dbx->md_cmp;
3407 /* Branch pages have no data, so if using integer keys,
3408 * alignment is guaranteed. Use faster mdb_cmp_int.
3410 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3411 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3418 nodekey.mv_size = mc->mc_db->md_pad;
3419 node = NODEPTR(mp, 0); /* fake */
3420 while (low <= high) {
3421 i = (low + high) >> 1;
3422 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3423 rc = cmp(key, &nodekey);
3424 DPRINTF("found leaf index %u [%s], rc = %i",
3425 i, DKEY(&nodekey), rc);
3434 while (low <= high) {
3435 i = (low + high) >> 1;
3437 node = NODEPTR(mp, i);
3438 nodekey.mv_size = NODEKSZ(node);
3439 nodekey.mv_data = NODEKEY(node);
3441 rc = cmp(key, &nodekey);
3444 DPRINTF("found leaf index %u [%s], rc = %i",
3445 i, DKEY(&nodekey), rc);
3447 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3448 i, DKEY(&nodekey), NODEPGNO(node), rc);
3459 if (rc > 0) { /* Found entry is less than the key. */
3460 i++; /* Skip to get the smallest entry larger than key. */
3462 node = NODEPTR(mp, i);
3465 *exactp = (rc == 0);
3466 /* store the key index */
3467 mc->mc_ki[mc->mc_top] = i;
3469 /* There is no entry larger or equal to the key. */
3472 /* nodeptr is fake for LEAF2 */
3478 mdb_cursor_adjust(MDB_cursor *mc, func)
3482 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3483 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3490 /** Pop a page off the top of the cursor's stack. */
3492 mdb_cursor_pop(MDB_cursor *mc)
3496 MDB_page *top = mc->mc_pg[mc->mc_top];
3502 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3503 mc->mc_dbi, (void *) mc);
3507 /** Push a page onto the top of the cursor's stack. */
3509 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3511 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3512 mc->mc_dbi, (void *) mc);
3514 if (mc->mc_snum >= CURSOR_STACK) {
3515 assert(mc->mc_snum < CURSOR_STACK);
3516 return MDB_CURSOR_FULL;
3519 mc->mc_top = mc->mc_snum++;
3520 mc->mc_pg[mc->mc_top] = mp;
3521 mc->mc_ki[mc->mc_top] = 0;
3526 /** Find the address of the page corresponding to a given page number.
3527 * @param[in] txn the transaction for this access.
3528 * @param[in] pgno the page number for the page to retrieve.
3529 * @param[out] ret address of a pointer where the page's address will be stored.
3530 * @return 0 on success, non-zero on failure.
3533 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3537 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
3538 if (pgno < txn->mt_next_pgno)
3539 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3542 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3544 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3545 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3546 p = txn->mt_u.dirty_list[x].mptr;
3550 if (pgno < txn->mt_next_pgno)
3551 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3556 DPRINTF("page %zu not found", pgno);
3559 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3562 /** Search for the page a given key should be in.
3563 * Pushes parent pages on the cursor stack. This function continues a
3564 * search on a cursor that has already been initialized. (Usually by
3565 * #mdb_page_search() but also by #mdb_node_move().)
3566 * @param[in,out] mc the cursor for this operation.
3567 * @param[in] key the key to search for. If NULL, search for the lowest
3568 * page. (This is used by #mdb_cursor_first().)
3569 * @param[in] flags If MDB_PS_MODIFY set, visited pages are updated with new page numbers.
3570 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
3571 * @return 0 on success, non-zero on failure.
3574 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3576 MDB_page *mp = mc->mc_pg[mc->mc_top];
3581 while (IS_BRANCH(mp)) {
3585 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3586 assert(NUMKEYS(mp) > 1);
3587 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3589 if (key == NULL) /* Initialize cursor to first page. */
3591 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3592 /* cursor to last page */
3596 node = mdb_node_search(mc, key, &exact);
3598 i = NUMKEYS(mp) - 1;
3600 i = mc->mc_ki[mc->mc_top];
3609 DPRINTF("following index %u for key [%s]",
3611 assert(i < NUMKEYS(mp));
3612 node = NODEPTR(mp, i);
3614 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3617 mc->mc_ki[mc->mc_top] = i;
3618 if ((rc = mdb_cursor_push(mc, mp)))
3622 if ((rc = mdb_page_touch(mc)) != 0)
3624 mp = mc->mc_pg[mc->mc_top];
3629 DPRINTF("internal error, index points to a %02X page!?",
3631 return MDB_CORRUPTED;
3634 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3635 key ? DKEY(key) : NULL);
3640 /** Search for the page a given key should be in.
3641 * Pushes parent pages on the cursor stack. This function just sets up
3642 * the search; it finds the root page for \b mc's database and sets this
3643 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3644 * called to complete the search.
3645 * @param[in,out] mc the cursor for this operation.
3646 * @param[in] key the key to search for. If NULL, search for the lowest
3647 * page. (This is used by #mdb_cursor_first().)
3648 * @param[in] modify If true, visited pages are updated with new page numbers.
3649 * @return 0 on success, non-zero on failure.
3652 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
3657 /* Make sure the txn is still viable, then find the root from
3658 * the txn's db table.
3660 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3661 DPUTS("transaction has failed, must abort");
3664 /* Make sure we're using an up-to-date root */
3665 if (mc->mc_dbi > MAIN_DBI) {
3666 if ((*mc->mc_dbflag & DB_STALE) ||
3667 ((flags & MDB_PS_MODIFY) && !(*mc->mc_dbflag & DB_DIRTY))) {
3669 unsigned char dbflag = 0;
3670 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3671 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, flags & MDB_PS_MODIFY);
3674 if (*mc->mc_dbflag & DB_STALE) {
3677 MDB_node *leaf = mdb_node_search(&mc2,
3678 &mc->mc_dbx->md_name, &exact);
3680 return MDB_NOTFOUND;
3681 mdb_node_read(mc->mc_txn, leaf, &data);
3682 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3684 if (flags & MDB_PS_MODIFY)
3686 *mc->mc_dbflag = dbflag;
3689 root = mc->mc_db->md_root;
3691 if (root == P_INVALID) { /* Tree is empty. */
3692 DPUTS("tree is empty");
3693 return MDB_NOTFOUND;
3698 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
3699 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3705 DPRINTF("db %u root page %zu has flags 0x%X",
3706 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3708 if (flags & MDB_PS_MODIFY) {
3709 if ((rc = mdb_page_touch(mc)))
3713 if (flags & MDB_PS_ROOTONLY)
3716 return mdb_page_search_root(mc, key, flags);
3719 /** Return the data associated with a given node.
3720 * @param[in] txn The transaction for this operation.
3721 * @param[in] leaf The node being read.
3722 * @param[out] data Updated to point to the node's data.
3723 * @return 0 on success, non-zero on failure.
3726 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3728 MDB_page *omp; /* overflow page */
3732 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3733 data->mv_size = NODEDSZ(leaf);
3734 data->mv_data = NODEDATA(leaf);
3738 /* Read overflow data.
3740 data->mv_size = NODEDSZ(leaf);
3741 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3742 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3743 DPRINTF("read overflow page %zu failed", pgno);
3746 data->mv_data = METADATA(omp);
3752 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3753 MDB_val *key, MDB_val *data)
3762 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3764 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3767 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3771 mdb_cursor_init(&mc, txn, dbi, &mx);
3772 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3775 /** Find a sibling for a page.
3776 * Replaces the page at the top of the cursor's stack with the
3777 * specified sibling, if one exists.
3778 * @param[in] mc The cursor for this operation.
3779 * @param[in] move_right Non-zero if the right sibling is requested,
3780 * otherwise the left sibling.
3781 * @return 0 on success, non-zero on failure.
3784 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3790 if (mc->mc_snum < 2) {
3791 return MDB_NOTFOUND; /* root has no siblings */
3795 DPRINTF("parent page is page %zu, index %u",
3796 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3798 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3799 : (mc->mc_ki[mc->mc_top] == 0)) {
3800 DPRINTF("no more keys left, moving to %s sibling",
3801 move_right ? "right" : "left");
3802 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3806 mc->mc_ki[mc->mc_top]++;
3808 mc->mc_ki[mc->mc_top]--;
3809 DPRINTF("just moving to %s index key %u",
3810 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3812 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3814 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3815 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3818 mdb_cursor_push(mc, mp);
3823 /** Move the cursor to the next data item. */
3825 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3831 if (mc->mc_flags & C_EOF) {
3832 return MDB_NOTFOUND;
3835 assert(mc->mc_flags & C_INITIALIZED);
3837 mp = mc->mc_pg[mc->mc_top];
3839 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3840 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3841 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3842 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3843 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3844 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3848 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3849 if (op == MDB_NEXT_DUP)
3850 return MDB_NOTFOUND;
3854 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3856 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3857 DPUTS("=====> move to next sibling page");
3858 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3859 mc->mc_flags |= C_EOF;
3860 mc->mc_flags &= ~C_INITIALIZED;
3861 return MDB_NOTFOUND;
3863 mp = mc->mc_pg[mc->mc_top];
3864 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3866 mc->mc_ki[mc->mc_top]++;
3868 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3869 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3872 key->mv_size = mc->mc_db->md_pad;
3873 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3877 assert(IS_LEAF(mp));
3878 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3880 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3881 mdb_xcursor_init1(mc, leaf);
3884 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3887 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3888 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3889 if (rc != MDB_SUCCESS)
3894 MDB_GET_KEY(leaf, key);
3898 /** Move the cursor to the previous data item. */
3900 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3906 assert(mc->mc_flags & C_INITIALIZED);
3908 mp = mc->mc_pg[mc->mc_top];
3910 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3911 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3912 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3913 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3914 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3915 if (op != MDB_PREV || rc == MDB_SUCCESS)
3918 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3919 if (op == MDB_PREV_DUP)
3920 return MDB_NOTFOUND;
3925 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3927 if (mc->mc_ki[mc->mc_top] == 0) {
3928 DPUTS("=====> move to prev sibling page");
3929 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3930 mc->mc_flags &= ~C_INITIALIZED;
3931 return MDB_NOTFOUND;
3933 mp = mc->mc_pg[mc->mc_top];
3934 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3935 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3937 mc->mc_ki[mc->mc_top]--;
3939 mc->mc_flags &= ~C_EOF;
3941 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3942 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3945 key->mv_size = mc->mc_db->md_pad;
3946 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3950 assert(IS_LEAF(mp));
3951 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3953 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3954 mdb_xcursor_init1(mc, leaf);
3957 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3960 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3961 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3962 if (rc != MDB_SUCCESS)
3967 MDB_GET_KEY(leaf, key);
3971 /** Set the cursor on a specific data item. */
3973 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3974 MDB_cursor_op op, int *exactp)
3978 MDB_node *leaf = NULL;
3983 assert(key->mv_size > 0);
3985 /* See if we're already on the right page */
3986 if (mc->mc_flags & C_INITIALIZED) {
3989 mp = mc->mc_pg[mc->mc_top];
3991 mc->mc_ki[mc->mc_top] = 0;
3992 return MDB_NOTFOUND;
3994 if (mp->mp_flags & P_LEAF2) {
3995 nodekey.mv_size = mc->mc_db->md_pad;
3996 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3998 leaf = NODEPTR(mp, 0);
3999 MDB_GET_KEY(leaf, &nodekey);
4001 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4003 /* Probably happens rarely, but first node on the page
4004 * was the one we wanted.
4006 mc->mc_ki[mc->mc_top] = 0;
4013 unsigned int nkeys = NUMKEYS(mp);
4015 if (mp->mp_flags & P_LEAF2) {
4016 nodekey.mv_data = LEAF2KEY(mp,
4017 nkeys-1, nodekey.mv_size);
4019 leaf = NODEPTR(mp, nkeys-1);
4020 MDB_GET_KEY(leaf, &nodekey);
4022 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4024 /* last node was the one we wanted */
4025 mc->mc_ki[mc->mc_top] = nkeys-1;
4031 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
4032 /* This is definitely the right page, skip search_page */
4033 if (mp->mp_flags & P_LEAF2) {
4034 nodekey.mv_data = LEAF2KEY(mp,
4035 mc->mc_ki[mc->mc_top], nodekey.mv_size);
4037 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
4038 MDB_GET_KEY(leaf, &nodekey);
4040 rc = mc->mc_dbx->md_cmp(key, &nodekey);
4042 /* current node was the one we wanted */
4052 /* If any parents have right-sibs, search.
4053 * Otherwise, there's nothing further.
4055 for (i=0; i<mc->mc_top; i++)
4057 NUMKEYS(mc->mc_pg[i])-1)
4059 if (i == mc->mc_top) {
4060 /* There are no other pages */
4061 mc->mc_ki[mc->mc_top] = nkeys;
4062 return MDB_NOTFOUND;
4066 /* There are no other pages */
4067 mc->mc_ki[mc->mc_top] = 0;
4068 return MDB_NOTFOUND;
4072 rc = mdb_page_search(mc, key, 0);
4073 if (rc != MDB_SUCCESS)
4076 mp = mc->mc_pg[mc->mc_top];
4077 assert(IS_LEAF(mp));
4080 leaf = mdb_node_search(mc, key, exactp);
4081 if (exactp != NULL && !*exactp) {
4082 /* MDB_SET specified and not an exact match. */
4083 return MDB_NOTFOUND;
4087 DPUTS("===> inexact leaf not found, goto sibling");
4088 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
4089 return rc; /* no entries matched */
4090 mp = mc->mc_pg[mc->mc_top];
4091 assert(IS_LEAF(mp));
4092 leaf = NODEPTR(mp, 0);
4096 mc->mc_flags |= C_INITIALIZED;
4097 mc->mc_flags &= ~C_EOF;
4100 key->mv_size = mc->mc_db->md_pad;
4101 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
4105 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4106 mdb_xcursor_init1(mc, leaf);
4109 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4110 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
4111 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4114 if (op == MDB_GET_BOTH) {
4120 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
4121 if (rc != MDB_SUCCESS)
4124 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
4126 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
4128 rc = mc->mc_dbx->md_dcmp(data, &d2);
4130 if (op == MDB_GET_BOTH || rc > 0)
4131 return MDB_NOTFOUND;
4136 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4137 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4142 /* The key already matches in all other cases */
4143 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
4144 MDB_GET_KEY(leaf, key);
4145 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
4150 /** Move the cursor to the first item in the database. */
4152 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4157 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4158 rc = mdb_page_search(mc, NULL, 0);
4159 if (rc != MDB_SUCCESS)
4162 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4164 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
4165 mc->mc_flags |= C_INITIALIZED;
4166 mc->mc_flags &= ~C_EOF;
4168 mc->mc_ki[mc->mc_top] = 0;
4170 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4171 key->mv_size = mc->mc_db->md_pad;
4172 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
4177 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4178 mdb_xcursor_init1(mc, leaf);
4179 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4184 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4185 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4189 MDB_GET_KEY(leaf, key);
4193 /** Move the cursor to the last item in the database. */
4195 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
4200 if (!(mc->mc_flags & C_EOF)) {
4202 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
4205 lkey.mv_size = MAXKEYSIZE+1;
4206 lkey.mv_data = NULL;
4207 rc = mdb_page_search(mc, &lkey, 0);
4208 if (rc != MDB_SUCCESS)
4211 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
4213 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
4214 mc->mc_flags |= C_INITIALIZED|C_EOF;
4216 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4218 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4219 key->mv_size = mc->mc_db->md_pad;
4220 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
4225 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4226 mdb_xcursor_init1(mc, leaf);
4227 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4232 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
4233 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4238 MDB_GET_KEY(leaf, key);
4243 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4253 case MDB_GET_BOTH_RANGE:
4254 if (data == NULL || mc->mc_xcursor == NULL) {
4262 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4264 } else if (op == MDB_SET_RANGE)
4265 rc = mdb_cursor_set(mc, key, data, op, NULL);
4267 rc = mdb_cursor_set(mc, key, data, op, &exact);
4269 case MDB_GET_MULTIPLE:
4271 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4272 !(mc->mc_flags & C_INITIALIZED)) {
4277 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4278 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4281 case MDB_NEXT_MULTIPLE:
4283 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4287 if (!(mc->mc_flags & C_INITIALIZED))
4288 rc = mdb_cursor_first(mc, key, data);
4290 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4291 if (rc == MDB_SUCCESS) {
4292 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4295 mx = &mc->mc_xcursor->mx_cursor;
4296 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4298 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4299 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4307 case MDB_NEXT_NODUP:
4308 if (!(mc->mc_flags & C_INITIALIZED))
4309 rc = mdb_cursor_first(mc, key, data);
4311 rc = mdb_cursor_next(mc, key, data, op);
4315 case MDB_PREV_NODUP:
4316 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF)) {
4317 rc = mdb_cursor_last(mc, key, data);
4318 mc->mc_flags &= ~C_EOF;
4320 rc = mdb_cursor_prev(mc, key, data, op);
4323 rc = mdb_cursor_first(mc, key, data);
4327 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4328 !(mc->mc_flags & C_INITIALIZED) ||
4329 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4333 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4336 rc = mdb_cursor_last(mc, key, data);
4340 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4341 !(mc->mc_flags & C_INITIALIZED) ||
4342 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4346 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4349 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4357 /** Touch all the pages in the cursor stack.
4358 * Makes sure all the pages are writable, before attempting a write operation.
4359 * @param[in] mc The cursor to operate on.
4362 mdb_cursor_touch(MDB_cursor *mc)
4366 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4368 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4369 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
4372 *mc->mc_dbflag = DB_DIRTY;
4374 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4375 rc = mdb_page_touch(mc);
4379 mc->mc_top = mc->mc_snum-1;
4384 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4387 MDB_node *leaf = NULL;
4388 MDB_val xdata, *rdata, dkey;
4391 int do_sub = 0, insert = 0;
4392 unsigned int mcount = 0;
4396 char dbuf[MAXKEYSIZE+1];
4397 unsigned int nflags;
4400 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4403 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4404 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4408 if (flags == MDB_CURRENT) {
4409 if (!(mc->mc_flags & C_INITIALIZED))
4412 } else if (mc->mc_db->md_root == P_INVALID) {
4414 /* new database, write a root leaf page */
4415 DPUTS("allocating new root leaf page");
4416 if ((rc = mdb_page_new(mc, P_LEAF, 1, &np))) {
4420 mdb_cursor_push(mc, np);
4421 mc->mc_db->md_root = np->mp_pgno;
4422 mc->mc_db->md_depth++;
4423 *mc->mc_dbflag = DB_DIRTY;
4424 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4426 np->mp_flags |= P_LEAF2;
4427 mc->mc_flags |= C_INITIALIZED;
4433 if (flags & MDB_APPEND) {
4435 rc = mdb_cursor_last(mc, &k2, &d2);
4437 rc = mc->mc_dbx->md_cmp(key, &k2);
4440 mc->mc_ki[mc->mc_top]++;
4446 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4448 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4449 DPRINTF("duplicate key [%s]", DKEY(key));
4451 return MDB_KEYEXIST;
4453 if (rc && rc != MDB_NOTFOUND)
4457 /* Cursor is positioned, now make sure all pages are writable */
4458 rc2 = mdb_cursor_touch(mc);
4463 /* The key already exists */
4464 if (rc == MDB_SUCCESS) {
4465 /* there's only a key anyway, so this is a no-op */
4466 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4467 unsigned int ksize = mc->mc_db->md_pad;
4468 if (key->mv_size != ksize)
4470 if (flags == MDB_CURRENT) {
4471 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4472 memcpy(ptr, key->mv_data, ksize);
4477 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4480 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4481 /* Was a single item before, must convert now */
4483 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4484 /* Just overwrite the current item */
4485 if (flags == MDB_CURRENT)
4488 dkey.mv_size = NODEDSZ(leaf);
4489 dkey.mv_data = NODEDATA(leaf);
4490 #if UINT_MAX < SIZE_MAX
4491 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4492 #ifdef MISALIGNED_OK
4493 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4495 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4498 /* if data matches, ignore it */
4499 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4500 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4502 /* create a fake page for the dup items */
4503 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4504 dkey.mv_data = dbuf;
4505 fp = (MDB_page *)&pbuf;
4506 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4507 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4508 fp->mp_lower = PAGEHDRSZ;
4509 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4510 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4511 fp->mp_flags |= P_LEAF2;
4512 fp->mp_pad = data->mv_size;
4513 fp->mp_upper += 2 * data->mv_size; /* leave space for 2 more */
4515 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4516 (dkey.mv_size & 1) + (data->mv_size & 1);
4518 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4521 xdata.mv_size = fp->mp_upper;
4526 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4527 /* See if we need to convert from fake page to subDB */
4529 unsigned int offset;
4532 fp = NODEDATA(leaf);
4533 if (flags == MDB_CURRENT) {
4535 fp->mp_flags |= P_DIRTY;
4536 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4537 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4541 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4542 offset = fp->mp_pad;
4543 if (SIZELEFT(fp) >= offset)
4545 offset *= 4; /* space for 4 more */
4547 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4549 offset += offset & 1;
4550 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4551 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4553 /* yes, convert it */
4555 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4556 dummy.md_pad = fp->mp_pad;
4557 dummy.md_flags = MDB_DUPFIXED;
4558 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4559 dummy.md_flags |= MDB_INTEGERKEY;
4562 dummy.md_branch_pages = 0;
4563 dummy.md_leaf_pages = 1;
4564 dummy.md_overflow_pages = 0;
4565 dummy.md_entries = NUMKEYS(fp);
4567 xdata.mv_size = sizeof(MDB_db);
4568 xdata.mv_data = &dummy;
4569 if ((rc = mdb_page_alloc(mc, 1, &mp)))
4571 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4572 flags |= F_DUPDATA|F_SUBDATA;
4573 dummy.md_root = mp->mp_pgno;
4575 /* no, just grow it */
4577 xdata.mv_size = NODEDSZ(leaf) + offset;
4578 xdata.mv_data = &pbuf;
4579 mp = (MDB_page *)&pbuf;
4580 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4583 mp->mp_flags = fp->mp_flags | P_DIRTY;
4584 mp->mp_pad = fp->mp_pad;
4585 mp->mp_lower = fp->mp_lower;
4586 mp->mp_upper = fp->mp_upper + offset;
4588 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4590 nsize = NODEDSZ(leaf) - fp->mp_upper;
4591 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4592 for (i=0; i<NUMKEYS(fp); i++)
4593 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4595 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4599 /* data is on sub-DB, just store it */
4600 flags |= F_DUPDATA|F_SUBDATA;
4604 /* overflow page overwrites need special handling */
4605 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4608 int ovpages, dpages;
4610 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4611 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4612 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4613 mdb_page_get(mc->mc_txn, pg, &omp);
4614 /* Is the ov page writable and large enough? */
4615 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4616 /* yes, overwrite it. Note in this case we don't
4617 * bother to try shrinking the node if the new data
4618 * is smaller than the overflow threshold.
4620 if (F_ISSET(flags, MDB_RESERVE))
4621 data->mv_data = METADATA(omp);
4623 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4626 /* no, free ovpages */
4628 mc->mc_db->md_overflow_pages -= ovpages;
4629 for (i=0; i<ovpages; i++) {
4630 DPRINTF("freed ov page %zu", pg);
4631 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4635 } else if (NODEDSZ(leaf) == data->mv_size) {
4636 /* same size, just replace it. Note that we could
4637 * also reuse this node if the new data is smaller,
4638 * but instead we opt to shrink the node in that case.
4640 if (F_ISSET(flags, MDB_RESERVE))
4641 data->mv_data = NODEDATA(leaf);
4643 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4646 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4647 mc->mc_db->md_entries--;
4649 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4656 nflags = flags & NODE_ADD_FLAGS;
4657 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4658 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4659 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4660 nflags &= ~MDB_APPEND;
4662 nflags |= MDB_SPLIT_REPLACE;
4663 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4665 /* There is room already in this leaf page. */
4666 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4667 if (rc == 0 && !do_sub && insert) {
4668 /* Adjust other cursors pointing to mp */
4669 MDB_cursor *m2, *m3;
4670 MDB_dbi dbi = mc->mc_dbi;
4671 unsigned i = mc->mc_top;
4672 MDB_page *mp = mc->mc_pg[i];
4674 if (mc->mc_flags & C_SUB)
4677 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4678 if (mc->mc_flags & C_SUB)
4679 m3 = &m2->mc_xcursor->mx_cursor;
4682 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4683 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4690 if (rc != MDB_SUCCESS)
4691 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4693 /* Now store the actual data in the child DB. Note that we're
4694 * storing the user data in the keys field, so there are strict
4695 * size limits on dupdata. The actual data fields of the child
4696 * DB are all zero size.
4703 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4704 if (flags & MDB_CURRENT) {
4705 xflags = MDB_CURRENT;
4707 mdb_xcursor_init1(mc, leaf);
4708 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4710 /* converted, write the original data first */
4712 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4716 /* Adjust other cursors pointing to mp */
4718 unsigned i = mc->mc_top;
4719 MDB_page *mp = mc->mc_pg[i];
4721 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4722 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4723 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4724 mdb_xcursor_init1(m2, leaf);
4729 if (flags & MDB_APPENDDUP)
4730 xflags |= MDB_APPEND;
4731 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4732 if (flags & F_SUBDATA) {
4733 void *db = NODEDATA(leaf);
4734 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4737 /* sub-writes might have failed so check rc again.
4738 * Don't increment count if we just replaced an existing item.
4740 if (!rc && !(flags & MDB_CURRENT))
4741 mc->mc_db->md_entries++;
4742 if (flags & MDB_MULTIPLE) {
4744 if (mcount < data[1].mv_size) {
4745 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4746 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4756 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4761 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4764 if (!mc->mc_flags & C_INITIALIZED)
4767 rc = mdb_cursor_touch(mc);
4771 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4773 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4774 if (flags != MDB_NODUPDATA) {
4775 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4776 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4778 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4779 /* If sub-DB still has entries, we're done */
4780 if (mc->mc_xcursor->mx_db.md_entries) {
4781 if (leaf->mn_flags & F_SUBDATA) {
4782 /* update subDB info */
4783 void *db = NODEDATA(leaf);
4784 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4786 /* shrink fake page */
4787 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4789 mc->mc_db->md_entries--;
4792 /* otherwise fall thru and delete the sub-DB */
4795 if (leaf->mn_flags & F_SUBDATA) {
4796 /* add all the child DB's pages to the free list */
4797 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4798 if (rc == MDB_SUCCESS) {
4799 mc->mc_db->md_entries -=
4800 mc->mc_xcursor->mx_db.md_entries;
4805 return mdb_cursor_del0(mc, leaf);
4808 /** Allocate and initialize new pages for a database.
4809 * @param[in] mc a cursor on the database being added to.
4810 * @param[in] flags flags defining what type of page is being allocated.
4811 * @param[in] num the number of pages to allocate. This is usually 1,
4812 * unless allocating overflow pages for a large record.
4813 * @param[out] mp Address of a page, or NULL on failure.
4814 * @return 0 on success, non-zero on failure.
4817 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
4822 if ((rc = mdb_page_alloc(mc, num, &np)))
4824 DPRINTF("allocated new mpage %zu, page size %u",
4825 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4826 np->mp_flags = flags | P_DIRTY;
4827 np->mp_lower = PAGEHDRSZ;
4828 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4831 mc->mc_db->md_branch_pages++;
4832 else if (IS_LEAF(np))
4833 mc->mc_db->md_leaf_pages++;
4834 else if (IS_OVERFLOW(np)) {
4835 mc->mc_db->md_overflow_pages += num;
4843 /** Calculate the size of a leaf node.
4844 * The size depends on the environment's page size; if a data item
4845 * is too large it will be put onto an overflow page and the node
4846 * size will only include the key and not the data. Sizes are always
4847 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4848 * of the #MDB_node headers.
4849 * @param[in] env The environment handle.
4850 * @param[in] key The key for the node.
4851 * @param[in] data The data for the node.
4852 * @return The number of bytes needed to store the node.
4855 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4859 sz = LEAFSIZE(key, data);
4860 if (sz >= env->me_psize / MDB_MINKEYS) {
4861 /* put on overflow page */
4862 sz -= data->mv_size - sizeof(pgno_t);
4866 return sz + sizeof(indx_t);
4869 /** Calculate the size of a branch node.
4870 * The size should depend on the environment's page size but since
4871 * we currently don't support spilling large keys onto overflow
4872 * pages, it's simply the size of the #MDB_node header plus the
4873 * size of the key. Sizes are always rounded up to an even number
4874 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4875 * @param[in] env The environment handle.
4876 * @param[in] key The key for the node.
4877 * @return The number of bytes needed to store the node.
4880 mdb_branch_size(MDB_env *env, MDB_val *key)
4885 if (sz >= env->me_psize / MDB_MINKEYS) {
4886 /* put on overflow page */
4887 /* not implemented */
4888 /* sz -= key->size - sizeof(pgno_t); */
4891 return sz + sizeof(indx_t);
4894 /** Add a node to the page pointed to by the cursor.
4895 * @param[in] mc The cursor for this operation.
4896 * @param[in] indx The index on the page where the new node should be added.
4897 * @param[in] key The key for the new node.
4898 * @param[in] data The data for the new node, if any.
4899 * @param[in] pgno The page number, if adding a branch node.
4900 * @param[in] flags Flags for the node.
4901 * @return 0 on success, non-zero on failure. Possible errors are:
4903 * <li>ENOMEM - failed to allocate overflow pages for the node.
4904 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
4905 * should never happen since all callers already calculate the
4906 * page's free space before calling this function.
4910 mdb_node_add(MDB_cursor *mc, indx_t indx,
4911 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4914 size_t node_size = NODESIZE;
4917 MDB_page *mp = mc->mc_pg[mc->mc_top];
4918 MDB_page *ofp = NULL; /* overflow page */
4921 assert(mp->mp_upper >= mp->mp_lower);
4923 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4924 IS_LEAF(mp) ? "leaf" : "branch",
4925 IS_SUBP(mp) ? "sub-" : "",
4926 mp->mp_pgno, indx, data ? data->mv_size : 0,
4927 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4930 /* Move higher keys up one slot. */
4931 int ksize = mc->mc_db->md_pad, dif;
4932 char *ptr = LEAF2KEY(mp, indx, ksize);
4933 dif = NUMKEYS(mp) - indx;
4935 memmove(ptr+ksize, ptr, dif*ksize);
4936 /* insert new key */
4937 memcpy(ptr, key->mv_data, ksize);
4939 /* Just using these for counting */
4940 mp->mp_lower += sizeof(indx_t);
4941 mp->mp_upper -= ksize - sizeof(indx_t);
4946 node_size += key->mv_size;
4950 if (F_ISSET(flags, F_BIGDATA)) {
4951 /* Data already on overflow page. */
4952 node_size += sizeof(pgno_t);
4953 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4954 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4956 /* Put data on overflow page. */
4957 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4958 data->mv_size, node_size+data->mv_size);
4959 node_size += sizeof(pgno_t);
4960 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
4962 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4965 node_size += data->mv_size;
4968 node_size += node_size & 1;
4970 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4971 DPRINTF("not enough room in page %zu, got %u ptrs",
4972 mp->mp_pgno, NUMKEYS(mp));
4973 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4974 mp->mp_upper - mp->mp_lower);
4975 DPRINTF("node size = %zu", node_size);
4976 return MDB_PAGE_FULL;
4979 /* Move higher pointers up one slot. */
4980 for (i = NUMKEYS(mp); i > indx; i--)
4981 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4983 /* Adjust free space offsets. */
4984 ofs = mp->mp_upper - node_size;
4985 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4986 mp->mp_ptrs[indx] = ofs;
4988 mp->mp_lower += sizeof(indx_t);
4990 /* Write the node data. */
4991 node = NODEPTR(mp, indx);
4992 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4993 node->mn_flags = flags;
4995 SETDSZ(node,data->mv_size);
5000 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5005 if (F_ISSET(flags, F_BIGDATA))
5006 memcpy(node->mn_data + key->mv_size, data->mv_data,
5008 else if (F_ISSET(flags, MDB_RESERVE))
5009 data->mv_data = node->mn_data + key->mv_size;
5011 memcpy(node->mn_data + key->mv_size, data->mv_data,
5014 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
5016 if (F_ISSET(flags, MDB_RESERVE))
5017 data->mv_data = METADATA(ofp);
5019 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
5026 /** Delete the specified node from a page.
5027 * @param[in] mp The page to operate on.
5028 * @param[in] indx The index of the node to delete.
5029 * @param[in] ksize The size of a node. Only used if the page is
5030 * part of a #MDB_DUPFIXED database.
5033 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
5036 indx_t i, j, numkeys, ptr;
5043 COPY_PGNO(pgno, mp->mp_pgno);
5044 DPRINTF("delete node %u on %s page %zu", indx,
5045 IS_LEAF(mp) ? "leaf" : "branch", pgno);
5048 assert(indx < NUMKEYS(mp));
5051 int x = NUMKEYS(mp) - 1 - indx;
5052 base = LEAF2KEY(mp, indx, ksize);
5054 memmove(base, base + ksize, x * ksize);
5055 mp->mp_lower -= sizeof(indx_t);
5056 mp->mp_upper += ksize - sizeof(indx_t);
5060 node = NODEPTR(mp, indx);
5061 sz = NODESIZE + node->mn_ksize;
5063 if (F_ISSET(node->mn_flags, F_BIGDATA))
5064 sz += sizeof(pgno_t);
5066 sz += NODEDSZ(node);
5070 ptr = mp->mp_ptrs[indx];
5071 numkeys = NUMKEYS(mp);
5072 for (i = j = 0; i < numkeys; i++) {
5074 mp->mp_ptrs[j] = mp->mp_ptrs[i];
5075 if (mp->mp_ptrs[i] < ptr)
5076 mp->mp_ptrs[j] += sz;
5081 base = (char *)mp + mp->mp_upper;
5082 memmove(base + sz, base, ptr - mp->mp_upper);
5084 mp->mp_lower -= sizeof(indx_t);
5088 /** Compact the main page after deleting a node on a subpage.
5089 * @param[in] mp The main page to operate on.
5090 * @param[in] indx The index of the subpage on the main page.
5093 mdb_node_shrink(MDB_page *mp, indx_t indx)
5100 indx_t i, numkeys, ptr;
5102 node = NODEPTR(mp, indx);
5103 sp = (MDB_page *)NODEDATA(node);
5104 osize = NODEDSZ(node);
5106 delta = sp->mp_upper - sp->mp_lower;
5107 SETDSZ(node, osize - delta);
5108 xp = (MDB_page *)((char *)sp + delta);
5110 /* shift subpage upward */
5112 nsize = NUMKEYS(sp) * sp->mp_pad;
5113 memmove(METADATA(xp), METADATA(sp), nsize);
5116 nsize = osize - sp->mp_upper;
5117 numkeys = NUMKEYS(sp);
5118 for (i=numkeys-1; i>=0; i--)
5119 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
5121 xp->mp_upper = sp->mp_lower;
5122 xp->mp_lower = sp->mp_lower;
5123 xp->mp_flags = sp->mp_flags;
5124 xp->mp_pad = sp->mp_pad;
5125 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
5127 /* shift lower nodes upward */
5128 ptr = mp->mp_ptrs[indx];
5129 numkeys = NUMKEYS(mp);
5130 for (i = 0; i < numkeys; i++) {
5131 if (mp->mp_ptrs[i] <= ptr)
5132 mp->mp_ptrs[i] += delta;
5135 base = (char *)mp + mp->mp_upper;
5136 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
5137 mp->mp_upper += delta;
5140 /** Initial setup of a sorted-dups cursor.
5141 * Sorted duplicates are implemented as a sub-database for the given key.
5142 * The duplicate data items are actually keys of the sub-database.
5143 * Operations on the duplicate data items are performed using a sub-cursor
5144 * initialized when the sub-database is first accessed. This function does
5145 * the preliminary setup of the sub-cursor, filling in the fields that
5146 * depend only on the parent DB.
5147 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5150 mdb_xcursor_init0(MDB_cursor *mc)
5152 MDB_xcursor *mx = mc->mc_xcursor;
5154 mx->mx_cursor.mc_xcursor = NULL;
5155 mx->mx_cursor.mc_txn = mc->mc_txn;
5156 mx->mx_cursor.mc_db = &mx->mx_db;
5157 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
5158 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
5159 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
5160 mx->mx_cursor.mc_snum = 0;
5161 mx->mx_cursor.mc_top = 0;
5162 mx->mx_cursor.mc_flags = C_SUB;
5163 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
5164 mx->mx_dbx.md_dcmp = NULL;
5165 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
5168 /** Final setup of a sorted-dups cursor.
5169 * Sets up the fields that depend on the data from the main cursor.
5170 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
5171 * @param[in] node The data containing the #MDB_db record for the
5172 * sorted-dup database.
5175 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
5177 MDB_xcursor *mx = mc->mc_xcursor;
5179 if (node->mn_flags & F_SUBDATA) {
5180 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
5181 mx->mx_cursor.mc_pg[0] = 0;
5182 mx->mx_cursor.mc_snum = 0;
5183 mx->mx_cursor.mc_flags = C_SUB;
5185 MDB_page *fp = NODEDATA(node);
5186 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
5187 mx->mx_db.md_flags = 0;
5188 mx->mx_db.md_depth = 1;
5189 mx->mx_db.md_branch_pages = 0;
5190 mx->mx_db.md_leaf_pages = 1;
5191 mx->mx_db.md_overflow_pages = 0;
5192 mx->mx_db.md_entries = NUMKEYS(fp);
5193 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
5194 mx->mx_cursor.mc_snum = 1;
5195 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
5196 mx->mx_cursor.mc_top = 0;
5197 mx->mx_cursor.mc_pg[0] = fp;
5198 mx->mx_cursor.mc_ki[0] = 0;
5199 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
5200 mx->mx_db.md_flags = MDB_DUPFIXED;
5201 mx->mx_db.md_pad = fp->mp_pad;
5202 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
5203 mx->mx_db.md_flags |= MDB_INTEGERKEY;
5206 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
5208 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
5210 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
5211 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
5212 #if UINT_MAX < SIZE_MAX
5213 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
5214 #ifdef MISALIGNED_OK
5215 mx->mx_dbx.md_cmp = mdb_cmp_long;
5217 mx->mx_dbx.md_cmp = mdb_cmp_cint;
5222 /** Initialize a cursor for a given transaction and database. */
5224 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
5229 mc->mc_db = &txn->mt_dbs[dbi];
5230 mc->mc_dbx = &txn->mt_dbxs[dbi];
5231 mc->mc_dbflag = &txn->mt_dbflags[dbi];
5236 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5238 mc->mc_xcursor = mx;
5239 mdb_xcursor_init0(mc);
5241 mc->mc_xcursor = NULL;
5243 if (*mc->mc_dbflag & DB_STALE) {
5244 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
5249 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
5252 MDB_xcursor *mx = NULL;
5253 size_t size = sizeof(MDB_cursor);
5255 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
5258 /* Allow read access to the freelist */
5259 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
5262 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
5263 size += sizeof(MDB_xcursor);
5265 if ((mc = malloc(size)) != NULL) {
5266 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
5267 mx = (MDB_xcursor *)(mc + 1);
5269 mdb_cursor_init(mc, txn, dbi, mx);
5270 if (txn->mt_cursors) {
5271 mc->mc_next = txn->mt_cursors[dbi];
5272 txn->mt_cursors[dbi] = mc;
5274 mc->mc_flags |= C_ALLOCD;
5285 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
5287 if (txn == NULL || mc == NULL || mc->mc_dbi >= txn->mt_numdbs)
5290 if (txn->mt_cursors)
5293 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
5297 /* Return the count of duplicate data items for the current key */
5299 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5303 if (mc == NULL || countp == NULL)
5306 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5309 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5310 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5313 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5316 *countp = mc->mc_xcursor->mx_db.md_entries;
5322 mdb_cursor_close(MDB_cursor *mc)
5325 /* remove from txn, if tracked */
5326 if (mc->mc_txn->mt_cursors) {
5327 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5328 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5330 *prev = mc->mc_next;
5332 if (mc->mc_flags & C_ALLOCD)
5338 mdb_cursor_txn(MDB_cursor *mc)
5340 if (!mc) return NULL;
5345 mdb_cursor_dbi(MDB_cursor *mc)
5351 /** Replace the key for a node with a new key.
5352 * @param[in] mp The page containing the node to operate on.
5353 * @param[in] indx The index of the node to operate on.
5354 * @param[in] key The new key to use.
5355 * @return 0 on success, non-zero on failure.
5358 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5364 indx_t ptr, i, numkeys;
5367 node = NODEPTR(mp, indx);
5368 ptr = mp->mp_ptrs[indx];
5372 char kbuf2[(MAXKEYSIZE*2+1)];
5373 k2.mv_data = NODEKEY(node);
5374 k2.mv_size = node->mn_ksize;
5375 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5377 mdb_dkey(&k2, kbuf2),
5383 delta0 = delta = key->mv_size - node->mn_ksize;
5385 /* Must be 2-byte aligned. If new key is
5386 * shorter by 1, the shift will be skipped.
5388 delta += (delta & 1);
5390 if (delta > 0 && SIZELEFT(mp) < delta) {
5391 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5392 return MDB_PAGE_FULL;
5395 numkeys = NUMKEYS(mp);
5396 for (i = 0; i < numkeys; i++) {
5397 if (mp->mp_ptrs[i] <= ptr)
5398 mp->mp_ptrs[i] -= delta;
5401 base = (char *)mp + mp->mp_upper;
5402 len = ptr - mp->mp_upper + NODESIZE;
5403 memmove(base - delta, base, len);
5404 mp->mp_upper -= delta;
5406 node = NODEPTR(mp, indx);
5409 /* But even if no shift was needed, update ksize */
5411 node->mn_ksize = key->mv_size;
5414 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5419 /** Move a node from csrc to cdst.
5422 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5428 unsigned short flags;
5432 /* Mark src and dst as dirty. */
5433 if ((rc = mdb_page_touch(csrc)) ||
5434 (rc = mdb_page_touch(cdst)))
5437 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5438 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5439 key.mv_size = csrc->mc_db->md_pad;
5440 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5442 data.mv_data = NULL;
5446 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5447 assert(!((long)srcnode&1));
5448 srcpg = NODEPGNO(srcnode);
5449 flags = srcnode->mn_flags;
5450 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5451 unsigned int snum = csrc->mc_snum;
5453 /* must find the lowest key below src */
5454 mdb_page_search_root(csrc, NULL, 0);
5455 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5456 key.mv_size = csrc->mc_db->md_pad;
5457 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5459 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5460 key.mv_size = NODEKSZ(s2);
5461 key.mv_data = NODEKEY(s2);
5463 csrc->mc_snum = snum--;
5464 csrc->mc_top = snum;
5466 key.mv_size = NODEKSZ(srcnode);
5467 key.mv_data = NODEKEY(srcnode);
5469 data.mv_size = NODEDSZ(srcnode);
5470 data.mv_data = NODEDATA(srcnode);
5472 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5473 unsigned int snum = cdst->mc_snum;
5476 /* must find the lowest key below dst */
5477 mdb_page_search_root(cdst, NULL, 0);
5478 if (IS_LEAF2(cdst->mc_pg[cdst->mc_top])) {
5479 bkey.mv_size = cdst->mc_db->md_pad;
5480 bkey.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, bkey.mv_size);
5482 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5483 bkey.mv_size = NODEKSZ(s2);
5484 bkey.mv_data = NODEKEY(s2);
5486 cdst->mc_snum = snum--;
5487 cdst->mc_top = snum;
5488 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5491 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5492 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5493 csrc->mc_ki[csrc->mc_top],
5495 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5496 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5498 /* Add the node to the destination page.
5500 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5501 if (rc != MDB_SUCCESS)
5504 /* Delete the node from the source page.
5506 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5509 /* Adjust other cursors pointing to mp */
5510 MDB_cursor *m2, *m3;
5511 MDB_dbi dbi = csrc->mc_dbi;
5512 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5514 if (csrc->mc_flags & C_SUB)
5517 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5518 if (m2 == csrc) continue;
5519 if (csrc->mc_flags & C_SUB)
5520 m3 = &m2->mc_xcursor->mx_cursor;
5523 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5524 csrc->mc_ki[csrc->mc_top]) {
5525 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5526 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5531 /* Update the parent separators.
5533 if (csrc->mc_ki[csrc->mc_top] == 0) {
5534 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5535 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5536 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5538 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5539 key.mv_size = NODEKSZ(srcnode);
5540 key.mv_data = NODEKEY(srcnode);
5542 DPRINTF("update separator for source page %zu to [%s]",
5543 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5544 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5545 &key)) != MDB_SUCCESS)
5548 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5550 nullkey.mv_size = 0;
5551 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5552 assert(rc == MDB_SUCCESS);
5556 if (cdst->mc_ki[cdst->mc_top] == 0) {
5557 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5558 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5559 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5561 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5562 key.mv_size = NODEKSZ(srcnode);
5563 key.mv_data = NODEKEY(srcnode);
5565 DPRINTF("update separator for destination page %zu to [%s]",
5566 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5567 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5568 &key)) != MDB_SUCCESS)
5571 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5573 nullkey.mv_size = 0;
5574 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5575 assert(rc == MDB_SUCCESS);
5582 /** Merge one page into another.
5583 * The nodes from the page pointed to by \b csrc will
5584 * be copied to the page pointed to by \b cdst and then
5585 * the \b csrc page will be freed.
5586 * @param[in] csrc Cursor pointing to the source page.
5587 * @param[in] cdst Cursor pointing to the destination page.
5590 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5598 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5599 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5601 assert(csrc->mc_snum > 1); /* can't merge root page */
5602 assert(cdst->mc_snum > 1);
5604 /* Mark dst as dirty. */
5605 if ((rc = mdb_page_touch(cdst)))
5608 /* Move all nodes from src to dst.
5610 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5611 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5612 key.mv_size = csrc->mc_db->md_pad;
5613 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5614 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5615 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5616 if (rc != MDB_SUCCESS)
5618 key.mv_data = (char *)key.mv_data + key.mv_size;
5621 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5622 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5623 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5624 unsigned int snum = csrc->mc_snum;
5626 /* must find the lowest key below src */
5627 mdb_page_search_root(csrc, NULL, 0);
5628 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5629 key.mv_size = csrc->mc_db->md_pad;
5630 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5632 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5633 key.mv_size = NODEKSZ(s2);
5634 key.mv_data = NODEKEY(s2);
5636 csrc->mc_snum = snum--;
5637 csrc->mc_top = snum;
5639 key.mv_size = srcnode->mn_ksize;
5640 key.mv_data = NODEKEY(srcnode);
5643 data.mv_size = NODEDSZ(srcnode);
5644 data.mv_data = NODEDATA(srcnode);
5645 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5646 if (rc != MDB_SUCCESS)
5651 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5652 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);
5654 /* Unlink the src page from parent and add to free list.
5656 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5657 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5659 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5663 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5664 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5665 csrc->mc_db->md_leaf_pages--;
5667 csrc->mc_db->md_branch_pages--;
5669 /* Adjust other cursors pointing to mp */
5670 MDB_cursor *m2, *m3;
5671 MDB_dbi dbi = csrc->mc_dbi;
5672 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5674 if (csrc->mc_flags & C_SUB)
5677 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5678 if (csrc->mc_flags & C_SUB)
5679 m3 = &m2->mc_xcursor->mx_cursor;
5682 if (m3 == csrc) continue;
5683 if (m3->mc_snum < csrc->mc_snum) continue;
5684 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5685 m3->mc_pg[csrc->mc_top] = mp;
5686 m3->mc_ki[csrc->mc_top] += nkeys;
5690 mdb_cursor_pop(csrc);
5692 return mdb_rebalance(csrc);
5695 /** Copy the contents of a cursor.
5696 * @param[in] csrc The cursor to copy from.
5697 * @param[out] cdst The cursor to copy to.
5700 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5704 cdst->mc_txn = csrc->mc_txn;
5705 cdst->mc_dbi = csrc->mc_dbi;
5706 cdst->mc_db = csrc->mc_db;
5707 cdst->mc_dbx = csrc->mc_dbx;
5708 cdst->mc_snum = csrc->mc_snum;
5709 cdst->mc_top = csrc->mc_top;
5710 cdst->mc_flags = csrc->mc_flags;
5712 for (i=0; i<csrc->mc_snum; i++) {
5713 cdst->mc_pg[i] = csrc->mc_pg[i];
5714 cdst->mc_ki[i] = csrc->mc_ki[i];
5718 /** Rebalance the tree after a delete operation.
5719 * @param[in] mc Cursor pointing to the page where rebalancing
5721 * @return 0 on success, non-zero on failure.
5724 mdb_rebalance(MDB_cursor *mc)
5734 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5735 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5736 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5737 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5741 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5744 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5745 DPRINTF("no need to rebalance page %zu, above fill threshold",
5751 if (mc->mc_snum < 2) {
5752 MDB_page *mp = mc->mc_pg[0];
5753 if (NUMKEYS(mp) == 0) {
5754 DPUTS("tree is completely empty");
5755 mc->mc_db->md_root = P_INVALID;
5756 mc->mc_db->md_depth = 0;
5757 mc->mc_db->md_leaf_pages = 0;
5758 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5762 /* Adjust other cursors pointing to mp */
5763 MDB_cursor *m2, *m3;
5764 MDB_dbi dbi = mc->mc_dbi;
5766 if (mc->mc_flags & C_SUB)
5769 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5770 if (m2 == mc) continue;
5771 if (mc->mc_flags & C_SUB)
5772 m3 = &m2->mc_xcursor->mx_cursor;
5775 if (m3->mc_snum < mc->mc_snum) continue;
5776 if (m3->mc_pg[0] == mp) {
5782 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5783 DPUTS("collapsing root page!");
5784 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5785 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5786 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5789 mc->mc_db->md_depth--;
5790 mc->mc_db->md_branch_pages--;
5792 /* Adjust other cursors pointing to mp */
5793 MDB_cursor *m2, *m3;
5794 MDB_dbi dbi = mc->mc_dbi;
5796 if (mc->mc_flags & C_SUB)
5799 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5800 if (m2 == mc) continue;
5801 if (mc->mc_flags & C_SUB)
5802 m3 = &m2->mc_xcursor->mx_cursor;
5805 if (m3->mc_snum < mc->mc_snum) continue;
5806 if (m3->mc_pg[0] == mp) {
5807 m3->mc_pg[0] = mc->mc_pg[0];
5812 DPUTS("root page doesn't need rebalancing");
5816 /* The parent (branch page) must have at least 2 pointers,
5817 * otherwise the tree is invalid.
5819 ptop = mc->mc_top-1;
5820 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5822 /* Leaf page fill factor is below the threshold.
5823 * Try to move keys from left or right neighbor, or
5824 * merge with a neighbor page.
5829 mdb_cursor_copy(mc, &mn);
5830 mn.mc_xcursor = NULL;
5832 if (mc->mc_ki[ptop] == 0) {
5833 /* We're the leftmost leaf in our parent.
5835 DPUTS("reading right neighbor");
5837 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5838 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5840 mn.mc_ki[mn.mc_top] = 0;
5841 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5843 /* There is at least one neighbor to the left.
5845 DPUTS("reading left neighbor");
5847 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5848 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5850 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5851 mc->mc_ki[mc->mc_top] = 0;
5854 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5855 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);
5857 /* If the neighbor page is above threshold and has at least two
5858 * keys, move one key from it.
5860 * Otherwise we should try to merge them.
5862 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5863 return mdb_node_move(&mn, mc);
5864 else { /* FIXME: if (has_enough_room()) */
5865 mc->mc_flags &= ~C_INITIALIZED;
5866 if (mc->mc_ki[ptop] == 0)
5867 return mdb_page_merge(&mn, mc);
5869 return mdb_page_merge(mc, &mn);
5873 /** Complete a delete operation started by #mdb_cursor_del(). */
5875 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5879 /* add overflow pages to free list */
5880 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5884 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5885 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5886 mc->mc_db->md_overflow_pages -= ovpages;
5887 for (i=0; i<ovpages; i++) {
5888 DPRINTF("freed ov page %zu", pg);
5889 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5893 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5894 mc->mc_db->md_entries--;
5895 rc = mdb_rebalance(mc);
5896 if (rc != MDB_SUCCESS)
5897 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5903 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5904 MDB_val *key, MDB_val *data)
5909 MDB_val rdata, *xdata;
5913 assert(key != NULL);
5915 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5917 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5920 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5924 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5928 mdb_cursor_init(&mc, txn, dbi, &mx);
5939 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5941 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5945 /** Split a page and insert a new node.
5946 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5947 * The cursor will be updated to point to the actual page and index where
5948 * the node got inserted after the split.
5949 * @param[in] newkey The key for the newly inserted node.
5950 * @param[in] newdata The data for the newly inserted node.
5951 * @param[in] newpgno The page number, if the new node is a branch node.
5952 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
5953 * @return 0 on success, non-zero on failure.
5956 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5957 unsigned int nflags)
5960 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1, did_split = 0;
5963 unsigned int i, j, split_indx, nkeys, pmax;
5965 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5967 MDB_page *mp, *rp, *pp;
5972 mp = mc->mc_pg[mc->mc_top];
5973 newindx = mc->mc_ki[mc->mc_top];
5975 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5976 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5977 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5979 /* Create a right sibling. */
5980 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
5982 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5984 if (mc->mc_snum < 2) {
5985 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
5987 /* shift current top to make room for new parent */
5988 mc->mc_pg[1] = mc->mc_pg[0];
5989 mc->mc_ki[1] = mc->mc_ki[0];
5992 mc->mc_db->md_root = pp->mp_pgno;
5993 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5994 mc->mc_db->md_depth++;
5997 /* Add left (implicit) pointer. */
5998 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5999 /* undo the pre-push */
6000 mc->mc_pg[0] = mc->mc_pg[1];
6001 mc->mc_ki[0] = mc->mc_ki[1];
6002 mc->mc_db->md_root = mp->mp_pgno;
6003 mc->mc_db->md_depth--;
6010 ptop = mc->mc_top-1;
6011 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
6014 mc->mc_flags |= C_SPLITTING;
6015 mdb_cursor_copy(mc, &mn);
6016 mn.mc_pg[mn.mc_top] = rp;
6017 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
6019 if (nflags & MDB_APPEND) {
6020 mn.mc_ki[mn.mc_top] = 0;
6022 split_indx = newindx;
6027 nkeys = NUMKEYS(mp);
6028 split_indx = nkeys / 2;
6029 if (newindx < split_indx)
6035 unsigned int lsize, rsize, ksize;
6036 /* Move half of the keys to the right sibling */
6038 x = mc->mc_ki[mc->mc_top] - split_indx;
6039 ksize = mc->mc_db->md_pad;
6040 split = LEAF2KEY(mp, split_indx, ksize);
6041 rsize = (nkeys - split_indx) * ksize;
6042 lsize = (nkeys - split_indx) * sizeof(indx_t);
6043 mp->mp_lower -= lsize;
6044 rp->mp_lower += lsize;
6045 mp->mp_upper += rsize - lsize;
6046 rp->mp_upper -= rsize - lsize;
6047 sepkey.mv_size = ksize;
6048 if (newindx == split_indx) {
6049 sepkey.mv_data = newkey->mv_data;
6051 sepkey.mv_data = split;
6054 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
6055 memcpy(rp->mp_ptrs, split, rsize);
6056 sepkey.mv_data = rp->mp_ptrs;
6057 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
6058 memcpy(ins, newkey->mv_data, ksize);
6059 mp->mp_lower += sizeof(indx_t);
6060 mp->mp_upper -= ksize - sizeof(indx_t);
6063 memcpy(rp->mp_ptrs, split, x * ksize);
6064 ins = LEAF2KEY(rp, x, ksize);
6065 memcpy(ins, newkey->mv_data, ksize);
6066 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
6067 rp->mp_lower += sizeof(indx_t);
6068 rp->mp_upper -= ksize - sizeof(indx_t);
6069 mc->mc_ki[mc->mc_top] = x;
6070 mc->mc_pg[mc->mc_top] = rp;
6075 /* For leaf pages, check the split point based on what
6076 * fits where, since otherwise mdb_node_add can fail.
6078 * This check is only needed when the data items are
6079 * relatively large, such that being off by one will
6080 * make the difference between success or failure.
6081 * When the size of the data items is much smaller than
6082 * one-half of a page, this check is irrelevant.
6085 unsigned int psize, nsize;
6086 /* Maximum free space in an empty page */
6087 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
6088 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
6089 if ((nkeys < 20) || (nsize > pmax/4)) {
6090 if (newindx <= split_indx) {
6093 for (i=0; i<split_indx; i++) {
6094 node = NODEPTR(mp, i);
6095 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6096 if (F_ISSET(node->mn_flags, F_BIGDATA))
6097 psize += sizeof(pgno_t);
6099 psize += NODEDSZ(node);
6103 split_indx = newindx;
6114 for (i=nkeys-1; i>=split_indx; i--) {
6115 node = NODEPTR(mp, i);
6116 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
6117 if (F_ISSET(node->mn_flags, F_BIGDATA))
6118 psize += sizeof(pgno_t);
6120 psize += NODEDSZ(node);
6124 split_indx = newindx;
6134 /* First find the separating key between the split pages.
6135 * The case where newindx == split_indx is ambiguous; the
6136 * new item could go to the new page or stay on the original
6137 * page. If newpos == 1 it goes to the new page.
6139 if (newindx == split_indx && newpos) {
6140 sepkey.mv_size = newkey->mv_size;
6141 sepkey.mv_data = newkey->mv_data;
6143 node = NODEPTR(mp, split_indx);
6144 sepkey.mv_size = node->mn_ksize;
6145 sepkey.mv_data = NODEKEY(node);
6149 DPRINTF("separator is [%s]", DKEY(&sepkey));
6151 /* Copy separator key to the parent.
6153 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
6157 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
6160 if (mn.mc_snum == mc->mc_snum) {
6161 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
6162 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
6163 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
6164 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
6169 /* Right page might now have changed parent.
6170 * Check if left page also changed parent.
6172 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
6173 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
6174 for (i=0; i<ptop; i++) {
6175 mc->mc_pg[i] = mn.mc_pg[i];
6176 mc->mc_ki[i] = mn.mc_ki[i];
6178 mc->mc_pg[ptop] = mn.mc_pg[ptop];
6179 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
6183 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
6186 mc->mc_flags ^= C_SPLITTING;
6187 if (rc != MDB_SUCCESS) {
6190 if (nflags & MDB_APPEND) {
6191 mc->mc_pg[mc->mc_top] = rp;
6192 mc->mc_ki[mc->mc_top] = 0;
6193 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
6196 for (i=0; i<mc->mc_top; i++)
6197 mc->mc_ki[i] = mn.mc_ki[i];
6204 /* Move half of the keys to the right sibling. */
6206 /* grab a page to hold a temporary copy */
6207 copy = mdb_page_malloc(mc);
6211 copy->mp_pgno = mp->mp_pgno;
6212 copy->mp_flags = mp->mp_flags;
6213 copy->mp_lower = PAGEHDRSZ;
6214 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
6215 mc->mc_pg[mc->mc_top] = copy;
6216 for (i = j = 0; i <= nkeys; j++) {
6217 if (i == split_indx) {
6218 /* Insert in right sibling. */
6219 /* Reset insert index for right sibling. */
6220 if (i != newindx || (newpos ^ ins_new)) {
6222 mc->mc_pg[mc->mc_top] = rp;
6226 if (i == newindx && !ins_new) {
6227 /* Insert the original entry that caused the split. */
6228 rkey.mv_data = newkey->mv_data;
6229 rkey.mv_size = newkey->mv_size;
6238 /* Update index for the new key. */
6239 mc->mc_ki[mc->mc_top] = j;
6240 } else if (i == nkeys) {
6243 node = NODEPTR(mp, i);
6244 rkey.mv_data = NODEKEY(node);
6245 rkey.mv_size = node->mn_ksize;
6247 xdata.mv_data = NODEDATA(node);
6248 xdata.mv_size = NODEDSZ(node);
6251 pgno = NODEPGNO(node);
6252 flags = node->mn_flags;
6257 if (!IS_LEAF(mp) && j == 0) {
6258 /* First branch index doesn't need key data. */
6262 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
6266 nkeys = NUMKEYS(copy);
6267 for (i=0; i<nkeys; i++)
6268 mp->mp_ptrs[i] = copy->mp_ptrs[i];
6269 mp->mp_lower = copy->mp_lower;
6270 mp->mp_upper = copy->mp_upper;
6271 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
6272 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
6274 /* reset back to original page */
6275 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
6276 mc->mc_pg[mc->mc_top] = mp;
6277 if (nflags & MDB_RESERVE) {
6278 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6279 if (!(node->mn_flags & F_BIGDATA))
6280 newdata->mv_data = NODEDATA(node);
6286 /* return tmp page to freelist */
6287 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
6288 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
6289 mc->mc_txn->mt_env->me_dpages = copy;
6292 /* Adjust other cursors pointing to mp */
6293 MDB_cursor *m2, *m3;
6294 MDB_dbi dbi = mc->mc_dbi;
6295 int fixup = NUMKEYS(mp);
6297 if (mc->mc_flags & C_SUB)
6300 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6301 if (m2 == mc) continue;
6302 if (mc->mc_flags & C_SUB)
6303 m3 = &m2->mc_xcursor->mx_cursor;
6306 if (!(m3->mc_flags & C_INITIALIZED))
6308 if (m3->mc_flags & C_SPLITTING)
6313 for (k=m3->mc_top; k>=0; k--) {
6314 m3->mc_ki[k+1] = m3->mc_ki[k];
6315 m3->mc_pg[k+1] = m3->mc_pg[k];
6317 if (m3->mc_ki[0] >= split_indx) {
6322 m3->mc_pg[0] = mc->mc_pg[0];
6326 if (m3->mc_pg[mc->mc_top] == mp) {
6327 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
6328 m3->mc_ki[mc->mc_top]++;
6329 if (m3->mc_ki[mc->mc_top] >= fixup) {
6330 m3->mc_pg[mc->mc_top] = rp;
6331 m3->mc_ki[mc->mc_top] -= fixup;
6332 m3->mc_ki[ptop] = mn.mc_ki[ptop];
6334 } else if (!did_split && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
6335 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
6344 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6345 MDB_val *key, MDB_val *data, unsigned int flags)
6350 assert(key != NULL);
6351 assert(data != NULL);
6353 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6356 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6360 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6364 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6367 mdb_cursor_init(&mc, txn, dbi, &mx);
6368 return mdb_cursor_put(&mc, key, data, flags);
6371 /** Only a subset of the @ref mdb_env flags can be changed
6372 * at runtime. Changing other flags requires closing the environment
6373 * and re-opening it with the new flags.
6375 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC)
6377 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6379 if ((flag & CHANGEABLE) != flag)
6382 env->me_flags |= flag;
6384 env->me_flags &= ~flag;
6389 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6394 *arg = env->me_flags;
6399 mdb_env_get_path(MDB_env *env, const char **arg)
6404 *arg = env->me_path;
6408 /** Common code for #mdb_stat() and #mdb_env_stat().
6409 * @param[in] env the environment to operate in.
6410 * @param[in] db the #MDB_db record containing the stats to return.
6411 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6412 * @return 0, this function always succeeds.
6415 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6417 arg->ms_psize = env->me_psize;
6418 arg->ms_depth = db->md_depth;
6419 arg->ms_branch_pages = db->md_branch_pages;
6420 arg->ms_leaf_pages = db->md_leaf_pages;
6421 arg->ms_overflow_pages = db->md_overflow_pages;
6422 arg->ms_entries = db->md_entries;
6427 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6431 if (env == NULL || arg == NULL)
6434 toggle = mdb_env_pick_meta(env);
6436 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6439 /** Set the default comparison functions for a database.
6440 * Called immediately after a database is opened to set the defaults.
6441 * The user can then override them with #mdb_set_compare() or
6442 * #mdb_set_dupsort().
6443 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6444 * @param[in] dbi A database handle returned by #mdb_open()
6447 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6449 uint16_t f = txn->mt_dbs[dbi].md_flags;
6451 txn->mt_dbxs[dbi].md_cmp =
6452 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
6453 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
6455 txn->mt_dbxs[dbi].md_dcmp =
6456 !(f & MDB_DUPSORT) ? 0 :
6457 ((f & MDB_INTEGERDUP)
6458 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
6459 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
6462 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6467 int rc, dbflag, exact;
6468 unsigned int unused = 0;
6471 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6472 mdb_default_cmp(txn, FREE_DBI);
6478 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6479 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6480 mdb_default_cmp(txn, MAIN_DBI);
6484 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6485 mdb_default_cmp(txn, MAIN_DBI);
6488 /* Is the DB already open? */
6490 for (i=2; i<txn->mt_numdbs; i++) {
6491 if (!txn->mt_dbxs[i].md_name.mv_size) {
6492 /* Remember this free slot */
6493 if (!unused) unused = i;
6496 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6497 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6503 /* If no free slot and max hit, fail */
6504 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
6505 return MDB_DBS_FULL;
6507 /* Find the DB info */
6511 key.mv_data = (void *)name;
6512 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6513 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6514 if (rc == MDB_SUCCESS) {
6515 /* make sure this is actually a DB */
6516 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6517 if (!(node->mn_flags & F_SUBDATA))
6519 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6520 /* Create if requested */
6522 data.mv_size = sizeof(MDB_db);
6523 data.mv_data = &dummy;
6524 memset(&dummy, 0, sizeof(dummy));
6525 dummy.md_root = P_INVALID;
6526 dummy.md_flags = flags & 0xffff;
6527 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6531 /* OK, got info, add to table */
6532 if (rc == MDB_SUCCESS) {
6533 unsigned int slot = unused ? unused : txn->mt_numdbs;
6534 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
6535 txn->mt_dbxs[slot].md_name.mv_size = len;
6536 txn->mt_dbxs[slot].md_rel = NULL;
6537 txn->mt_dbflags[slot] = dbflag;
6538 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
6540 txn->mt_env->me_dbflags[slot] = txn->mt_dbs[slot].md_flags;
6541 mdb_default_cmp(txn, slot);
6544 txn->mt_env->me_numdbs++;
6551 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6553 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6556 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6559 void mdb_close(MDB_env *env, MDB_dbi dbi)
6562 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6564 ptr = env->me_dbxs[dbi].md_name.mv_data;
6565 env->me_dbxs[dbi].md_name.mv_data = NULL;
6566 env->me_dbxs[dbi].md_name.mv_size = 0;
6570 /** Add all the DB's pages to the free list.
6571 * @param[in] mc Cursor on the DB to free.
6572 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6573 * @return 0 on success, non-zero on failure.
6576 mdb_drop0(MDB_cursor *mc, int subs)
6580 rc = mdb_page_search(mc, NULL, 0);
6581 if (rc == MDB_SUCCESS) {
6586 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6587 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6590 mdb_cursor_copy(mc, &mx);
6591 while (mc->mc_snum > 0) {
6592 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6593 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6594 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6595 if (ni->mn_flags & F_SUBDATA) {
6596 mdb_xcursor_init1(mc, ni);
6597 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6603 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6605 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6608 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6613 rc = mdb_cursor_sibling(mc, 1);
6615 /* no more siblings, go back to beginning
6616 * of previous level. (stack was already popped
6617 * by mdb_cursor_sibling)
6619 for (i=1; i<mc->mc_top; i++)
6620 mc->mc_pg[i] = mx.mc_pg[i];
6624 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6625 mc->mc_db->md_root);
6630 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6635 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6638 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6641 rc = mdb_cursor_open(txn, dbi, &mc);
6645 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6649 /* Can't delete the main DB */
6650 if (del && dbi > MAIN_DBI) {
6651 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6653 mdb_close(txn->mt_env, dbi);
6655 txn->mt_dbflags[dbi] |= DB_DIRTY;
6656 txn->mt_dbs[dbi].md_depth = 0;
6657 txn->mt_dbs[dbi].md_branch_pages = 0;
6658 txn->mt_dbs[dbi].md_leaf_pages = 0;
6659 txn->mt_dbs[dbi].md_overflow_pages = 0;
6660 txn->mt_dbs[dbi].md_entries = 0;
6661 txn->mt_dbs[dbi].md_root = P_INVALID;
6664 mdb_cursor_close(mc);
6668 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6670 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6673 txn->mt_dbxs[dbi].md_cmp = cmp;
6677 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6679 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6682 txn->mt_dbxs[dbi].md_dcmp = cmp;
6686 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6688 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6691 txn->mt_dbxs[dbi].md_rel = rel;
6695 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6697 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6700 txn->mt_dbxs[dbi].md_relctx = ctx;