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 */
67 #include <semaphore.h>
72 #include <valgrind/memcheck.h>
73 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
74 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
75 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
76 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
77 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
79 #define VGMEMP_CREATE(h,r,z)
80 #define VGMEMP_ALLOC(h,a,s)
81 #define VGMEMP_FREE(h,a)
82 #define VGMEMP_DESTROY(h)
83 #define VGMEMP_DEFINED(a,s)
87 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
88 /* Solaris just defines one or the other */
89 # define LITTLE_ENDIAN 1234
90 # define BIG_ENDIAN 4321
91 # ifdef _LITTLE_ENDIAN
92 # define BYTE_ORDER LITTLE_ENDIAN
94 # define BYTE_ORDER BIG_ENDIAN
97 # define BYTE_ORDER __BYTE_ORDER
101 #ifndef LITTLE_ENDIAN
102 #define LITTLE_ENDIAN __LITTLE_ENDIAN
105 #define BIG_ENDIAN __BIG_ENDIAN
108 #if defined(__i386) || defined(__x86_64)
109 #define MISALIGNED_OK 1
115 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
116 # error "Unknown or unsupported endianness (BYTE_ORDER)"
117 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
118 # error "Two's complement, reasonably sized integer types, please"
121 /** @defgroup internal MDB Internals
124 /** @defgroup compat Windows Compatibility Macros
125 * A bunch of macros to minimize the amount of platform-specific ifdefs
126 * needed throughout the rest of the code. When the features this library
127 * needs are similar enough to POSIX to be hidden in a one-or-two line
128 * replacement, this macro approach is used.
132 #define pthread_t DWORD
133 #define pthread_mutex_t HANDLE
134 #define pthread_key_t DWORD
135 #define pthread_self() GetCurrentThreadId()
136 #define pthread_key_create(x,y) (*(x) = TlsAlloc())
137 #define pthread_key_delete(x) TlsFree(x)
138 #define pthread_getspecific(x) TlsGetValue(x)
139 #define pthread_setspecific(x,y) TlsSetValue(x,y)
140 #define pthread_mutex_unlock(x) ReleaseMutex(x)
141 #define pthread_mutex_lock(x) WaitForSingleObject(x, INFINITE)
142 #define LOCK_MUTEX_R(env) pthread_mutex_lock((env)->me_rmutex)
143 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock((env)->me_rmutex)
144 #define LOCK_MUTEX_W(env) pthread_mutex_lock((env)->me_wmutex)
145 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock((env)->me_wmutex)
146 #define getpid() GetCurrentProcessId()
147 #define fdatasync(fd) (!FlushFileBuffers(fd))
148 #define ErrCode() GetLastError()
149 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
150 #define close(fd) CloseHandle(fd)
151 #define munmap(ptr,len) UnmapViewOfFile(ptr)
154 #define LOCK_MUTEX_R(env) sem_wait((env)->me_rmutex)
155 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
156 #define LOCK_MUTEX_W(env) sem_wait((env)->me_wmutex)
157 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
158 #define fdatasync(fd) fsync(fd)
161 #define fdatasync(fd) fsync(fd)
163 /** Lock the reader mutex.
165 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
166 /** Unlock the reader mutex.
168 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
170 /** Lock the writer mutex.
171 * Only a single write transaction is allowed at a time. Other writers
172 * will block waiting for this mutex.
174 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
175 /** Unlock the writer mutex.
177 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
178 #endif /* __APPLE__ */
180 /** Get the error code for the last failed system function.
182 #define ErrCode() errno
184 /** An abstraction for a file handle.
185 * On POSIX systems file handles are small integers. On Windows
186 * they're opaque pointers.
190 /** A value for an invalid file handle.
191 * Mainly used to initialize file variables and signify that they are
194 #define INVALID_HANDLE_VALUE (-1)
196 /** Get the size of a memory page for the system.
197 * This is the basic size that the platform's memory manager uses, and is
198 * fundamental to the use of memory-mapped files.
200 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
203 #if defined(_WIN32) || defined(__APPLE__)
206 #define MNAME_LEN (sizeof(pthread_mutex_t))
212 /** A flag for opening a file and requesting synchronous data writes.
213 * This is only used when writing a meta page. It's not strictly needed;
214 * we could just do a normal write and then immediately perform a flush.
215 * But if this flag is available it saves us an extra system call.
217 * @note If O_DSYNC is undefined but exists in /usr/include,
218 * preferably set some compiler flag to get the definition.
219 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
222 # define MDB_DSYNC O_DSYNC
226 /** Function for flushing the data of a file. Define this to fsync
227 * if fdatasync() is not supported.
229 #ifndef MDB_FDATASYNC
230 # define MDB_FDATASYNC fdatasync
233 /** A page number in the database.
234 * Note that 64 bit page numbers are overkill, since pages themselves
235 * already represent 12-13 bits of addressable memory, and the OS will
236 * always limit applications to a maximum of 63 bits of address space.
238 * @note In the #MDB_node structure, we only store 48 bits of this value,
239 * which thus limits us to only 60 bits of addressable data.
243 /** A transaction ID.
244 * See struct MDB_txn.mt_txnid for details.
248 /** @defgroup debug Debug Macros
252 /** Enable debug output.
253 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
254 * read from and written to the database (used for free space management).
259 #if !(__STDC_VERSION__ >= 199901L || defined(__GNUC__))
260 # define DPRINTF (void) /* Vararg macros may be unsupported */
262 static int mdb_debug;
263 static txnid_t mdb_debug_start;
265 /** Print a debug message with printf formatting. */
266 # define DPRINTF(fmt, ...) /**< Requires 2 or more args */ \
267 ((void) ((mdb_debug) && \
268 fprintf(stderr, "%s:%d " fmt "\n", __func__, __LINE__, __VA_ARGS__)))
270 # define DPRINTF(fmt, ...) ((void) 0)
272 /** Print a debug string.
273 * The string is printed literally, with no format processing.
275 #define DPUTS(arg) DPRINTF("%s", arg)
278 /** A default memory page size.
279 * The actual size is platform-dependent, but we use this for
280 * boot-strapping. We probably should not be using this any more.
281 * The #GET_PAGESIZE() macro is used to get the actual size.
283 * Note that we don't currently support Huge pages. On Linux,
284 * regular data files cannot use Huge pages, and in general
285 * Huge pages aren't actually pageable. We rely on the OS
286 * demand-pager to read our data and page it out when memory
287 * pressure from other processes is high. So until OSs have
288 * actual paging support for Huge pages, they're not viable.
290 #define MDB_PAGESIZE 4096
292 /** The minimum number of keys required in a database page.
293 * Setting this to a larger value will place a smaller bound on the
294 * maximum size of a data item. Data items larger than this size will
295 * be pushed into overflow pages instead of being stored directly in
296 * the B-tree node. This value used to default to 4. With a page size
297 * of 4096 bytes that meant that any item larger than 1024 bytes would
298 * go into an overflow page. That also meant that on average 2-3KB of
299 * each overflow page was wasted space. The value cannot be lower than
300 * 2 because then there would no longer be a tree structure. With this
301 * value, items larger than 2KB will go into overflow pages, and on
302 * average only 1KB will be wasted.
304 #define MDB_MINKEYS 2
306 /** A stamp that identifies a file as an MDB file.
307 * There's nothing special about this value other than that it is easily
308 * recognizable, and it will reflect any byte order mismatches.
310 #define MDB_MAGIC 0xBEEFC0DE
312 /** The version number for a database's file format. */
313 #define MDB_VERSION 1
315 /** The maximum size of a key in the database.
316 * While data items have essentially unbounded size, we require that
317 * keys all fit onto a regular page. This limit could be raised a bit
318 * further if needed; to something just under #MDB_PAGESIZE / #MDB_MINKEYS.
320 #define MAXKEYSIZE 511
325 * This is used for printing a hex dump of a key's contents.
327 #define DKBUF char kbuf[(MAXKEYSIZE*2+1)]
328 /** Display a key in hex.
330 * Invoke a function to display a key in hex.
332 #define DKEY(x) mdb_dkey(x, kbuf)
334 #define DKBUF typedef int dummy_kbuf /* so we can put ';' after */
338 /** @defgroup lazylock Lazy Locking
339 * Macros for locks that aren't actually needed.
340 * The DB view is always consistent because all writes are wrapped in
341 * the wmutex. Finer-grained locks aren't necessary.
345 /** Use lazy locking. I.e., don't lock these accesses at all. */
349 /** Grab the reader lock */
350 #define LAZY_MUTEX_LOCK(x)
351 /** Release the reader lock */
352 #define LAZY_MUTEX_UNLOCK(x)
353 /** Release the DB table reader/writer lock */
354 #define LAZY_RWLOCK_UNLOCK(x)
355 /** Grab the DB table write lock */
356 #define LAZY_RWLOCK_WRLOCK(x)
357 /** Grab the DB table read lock */
358 #define LAZY_RWLOCK_RDLOCK(x)
359 /** Declare the DB table rwlock. Should not be followed by ';'. */
360 #define LAZY_RWLOCK_DEF(x)
361 /** Initialize the DB table rwlock */
362 #define LAZY_RWLOCK_INIT(x,y)
363 /** Destroy the DB table rwlock */
364 #define LAZY_RWLOCK_DESTROY(x)
366 #define LAZY_MUTEX_LOCK(x) pthread_mutex_lock(x)
367 #define LAZY_MUTEX_UNLOCK(x) pthread_mutex_unlock(x)
368 #define LAZY_RWLOCK_UNLOCK(x) pthread_rwlock_unlock(x)
369 #define LAZY_RWLOCK_WRLOCK(x) pthread_rwlock_wrlock(x)
370 #define LAZY_RWLOCK_RDLOCK(x) pthread_rwlock_rdlock(x)
371 #define LAZY_RWLOCK_DEF(x) pthread_rwlock_t x;
372 #define LAZY_RWLOCK_INIT(x,y) pthread_rwlock_init(x,y)
373 #define LAZY_RWLOCK_DESTROY(x) pthread_rwlock_destroy(x)
377 /** An invalid page number.
378 * Mainly used to denote an empty tree.
380 #define P_INVALID (~0UL)
382 /** Test if a flag \b f is set in a flag word \b w. */
383 #define F_ISSET(w, f) (((w) & (f)) == (f))
385 /** Used for offsets within a single page.
386 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
389 typedef uint16_t indx_t;
391 /** Default size of memory map.
392 * This is certainly too small for any actual applications. Apps should always set
393 * the size explicitly using #mdb_env_set_mapsize().
395 #define DEFAULT_MAPSIZE 1048576
397 /** @defgroup readers Reader Lock Table
398 * Readers don't acquire any locks for their data access. Instead, they
399 * simply record their transaction ID in the reader table. The reader
400 * mutex is needed just to find an empty slot in the reader table. The
401 * slot's address is saved in thread-specific data so that subsequent read
402 * transactions started by the same thread need no further locking to proceed.
404 * Since the database uses multi-version concurrency control, readers don't
405 * actually need any locking. This table is used to keep track of which
406 * readers are using data from which old transactions, so that we'll know
407 * when a particular old transaction is no longer in use. Old transactions
408 * that have discarded any data pages can then have those pages reclaimed
409 * for use by a later write transaction.
411 * The lock table is constructed such that reader slots are aligned with the
412 * processor's cache line size. Any slot is only ever used by one thread.
413 * This alignment guarantees that there will be no contention or cache
414 * thrashing as threads update their own slot info, and also eliminates
415 * any need for locking when accessing a slot.
417 * A writer thread will scan every slot in the table to determine the oldest
418 * outstanding reader transaction. Any freed pages older than this will be
419 * reclaimed by the writer. The writer doesn't use any locks when scanning
420 * this table. This means that there's no guarantee that the writer will
421 * see the most up-to-date reader info, but that's not required for correct
422 * operation - all we need is to know the upper bound on the oldest reader,
423 * we don't care at all about the newest reader. So the only consequence of
424 * reading stale information here is that old pages might hang around a
425 * while longer before being reclaimed. That's actually good anyway, because
426 * the longer we delay reclaiming old pages, the more likely it is that a
427 * string of contiguous pages can be found after coalescing old pages from
428 * many old transactions together.
430 * @todo We don't actually do such coalescing yet, we grab pages from one
431 * old transaction at a time.
434 /** Number of slots in the reader table.
435 * This value was chosen somewhat arbitrarily. 126 readers plus a
436 * couple mutexes fit exactly into 8KB on my development machine.
437 * Applications should set the table size using #mdb_env_set_maxreaders().
439 #define DEFAULT_READERS 126
441 /** The size of a CPU cache line in bytes. We want our lock structures
442 * aligned to this size to avoid false cache line sharing in the
444 * This value works for most CPUs. For Itanium this should be 128.
450 /** The information we store in a single slot of the reader table.
451 * In addition to a transaction ID, we also record the process and
452 * thread ID that owns a slot, so that we can detect stale information,
453 * e.g. threads or processes that went away without cleaning up.
454 * @note We currently don't check for stale records. We simply re-init
455 * the table when we know that we're the only process opening the
458 typedef struct MDB_rxbody {
459 /** The current Transaction ID when this transaction began.
460 * Multiple readers that start at the same time will probably have the
461 * same ID here. Again, it's not important to exclude them from
462 * anything; all we need to know is which version of the DB they
463 * started from so we can avoid overwriting any data used in that
464 * particular version.
467 /** The process ID of the process owning this reader txn. */
469 /** The thread ID of the thread owning this txn. */
473 /** The actual reader record, with cacheline padding. */
474 typedef struct MDB_reader {
477 /** shorthand for mrb_txnid */
478 #define mr_txnid mru.mrx.mrb_txnid
479 #define mr_pid mru.mrx.mrb_pid
480 #define mr_tid mru.mrx.mrb_tid
481 /** cache line alignment */
482 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
486 /** The header for the reader table.
487 * The table resides in a memory-mapped file. (This is a different file
488 * than is used for the main database.)
490 * For POSIX the actual mutexes reside in the shared memory of this
491 * mapped file. On Windows, mutexes are named objects allocated by the
492 * kernel; we store the mutex names in this mapped file so that other
493 * processes can grab them. This same approach is also used on
494 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
495 * process-shared POSIX mutexes. For these cases where a named object
496 * is used, the object name is derived from a 64 bit FNV hash of the
497 * environment pathname. As such, naming collisions are extremely
498 * unlikely. If a collision occurs, the results are unpredictable.
500 typedef struct MDB_txbody {
501 /** Stamp identifying this as an MDB file. It must be set
504 /** Version number of this lock file. Must be set to #MDB_VERSION. */
505 uint32_t mtb_version;
506 #if defined(_WIN32) || defined(__APPLE__)
507 char mtb_rmname[MNAME_LEN];
509 /** Mutex protecting access to this table.
510 * This is the reader lock that #LOCK_MUTEX_R acquires.
512 pthread_mutex_t mtb_mutex;
515 /** The ID of the last transaction committed to the database.
516 * This is recorded here only for convenience; the value can always
517 * be determined by reading the main database meta pages.
519 * Value is unused, but maintained for backwards compatibility.
520 * Drop this or mtb_me_toggle when changing MDB_VERSION.
521 * (Reading both should have been done atomically.)
525 /** The number of slots that have been used in the reader table.
526 * This always records the maximum count, it is not decremented
527 * when readers release their slots.
529 unsigned mtb_numreaders;
530 /** The ID of the most recent meta page in the database.
531 * This is recorded here only for convenience; the value can always
532 * be determined by reading the main database meta pages.
534 uint32_t mtb_me_toggle;
537 /** The actual reader table definition. */
538 typedef struct MDB_txninfo {
541 #define mti_magic mt1.mtb.mtb_magic
542 #define mti_version mt1.mtb.mtb_version
543 #define mti_mutex mt1.mtb.mtb_mutex
544 #define mti_rmname mt1.mtb.mtb_rmname
546 #define mti_txnid mt1.mtb.mtb_txnid
548 #define mti_numreaders mt1.mtb.mtb_numreaders
549 #define mti_me_toggle mt1.mtb.mtb_me_toggle
550 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
553 #if defined(_WIN32) || defined(__APPLE__)
554 char mt2_wmname[MNAME_LEN];
555 #define mti_wmname mt2.mt2_wmname
557 pthread_mutex_t mt2_wmutex;
558 #define mti_wmutex mt2.mt2_wmutex
560 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
562 MDB_reader mti_readers[1];
566 /** Common header for all page types.
567 * Overflow records occupy a number of contiguous pages with no
568 * headers on any page after the first.
570 typedef struct MDB_page {
571 #define mp_pgno mp_p.p_pgno
572 #define mp_next mp_p.p_next
574 pgno_t p_pgno; /**< page number */
575 void * p_next; /**< for in-memory list of freed structs */
578 /** @defgroup mdb_page Page Flags
580 * Flags for the page headers.
583 #define P_BRANCH 0x01 /**< branch page */
584 #define P_LEAF 0x02 /**< leaf page */
585 #define P_OVERFLOW 0x04 /**< overflow page */
586 #define P_META 0x08 /**< meta page */
587 #define P_DIRTY 0x10 /**< dirty page */
588 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
589 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
591 uint16_t mp_flags; /**< @ref mdb_page */
592 #define mp_lower mp_pb.pb.pb_lower
593 #define mp_upper mp_pb.pb.pb_upper
594 #define mp_pages mp_pb.pb_pages
597 indx_t pb_lower; /**< lower bound of free space */
598 indx_t pb_upper; /**< upper bound of free space */
600 uint32_t pb_pages; /**< number of overflow pages */
602 indx_t mp_ptrs[1]; /**< dynamic size */
605 /** Size of the page header, excluding dynamic data at the end */
606 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
608 /** Address of first usable data byte in a page, after the header */
609 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
611 /** Number of nodes on a page */
612 #define NUMKEYS(p) (((p)->mp_lower - PAGEHDRSZ) >> 1)
614 /** The amount of space remaining in the page */
615 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
617 /** The percentage of space used in the page, in tenths of a percent. */
618 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
619 ((env)->me_psize - PAGEHDRSZ))
620 /** The minimum page fill factor, in tenths of a percent.
621 * Pages emptier than this are candidates for merging.
623 #define FILL_THRESHOLD 250
625 /** Test if a page is a leaf page */
626 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
627 /** Test if a page is a LEAF2 page */
628 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
629 /** Test if a page is a branch page */
630 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
631 /** Test if a page is an overflow page */
632 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
633 /** Test if a page is a sub page */
634 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
636 /** The number of overflow pages needed to store the given size. */
637 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
639 /** Header for a single key/data pair within a page.
640 * We guarantee 2-byte alignment for nodes.
642 typedef struct MDB_node {
643 /** lo and hi are used for data size on leaf nodes and for
644 * child pgno on branch nodes. On 64 bit platforms, flags
645 * is also used for pgno. (Branch nodes have no flags).
646 * They are in host byte order in case that lets some
647 * accesses be optimized into a 32-bit word access.
649 #define mn_lo mn_offset[BYTE_ORDER!=LITTLE_ENDIAN]
650 #define mn_hi mn_offset[BYTE_ORDER==LITTLE_ENDIAN] /**< part of dsize or pgno */
651 unsigned short mn_offset[2]; /**< storage for #mn_lo and #mn_hi */
652 /** @defgroup mdb_node Node Flags
654 * Flags for node headers.
657 #define F_BIGDATA 0x01 /**< data put on overflow page */
658 #define F_SUBDATA 0x02 /**< data is a sub-database */
659 #define F_DUPDATA 0x04 /**< data has duplicates */
661 /** valid flags for #mdb_node_add() */
662 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
665 unsigned short mn_flags; /**< @ref mdb_node */
666 unsigned short mn_ksize; /**< key size */
667 char mn_data[1]; /**< key and data are appended here */
670 /** Size of the node header, excluding dynamic data at the end */
671 #define NODESIZE offsetof(MDB_node, mn_data)
673 /** Bit position of top word in page number, for shifting mn_flags */
674 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
676 /** Size of a node in a branch page with a given key.
677 * This is just the node header plus the key, there is no data.
679 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
681 /** Size of a node in a leaf page with a given key and data.
682 * This is node header plus key plus data size.
684 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
686 /** Address of node \b i in page \b p */
687 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i]))
689 /** Address of the key for the node */
690 #define NODEKEY(node) (void *)((node)->mn_data)
692 /** Address of the data for a node */
693 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
695 /** Get the page number pointed to by a branch node */
696 #define NODEPGNO(node) \
697 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
698 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
699 /** Set the page number in a branch node */
700 #define SETPGNO(node,pgno) do { \
701 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
702 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
704 /** Get the size of the data in a leaf node */
705 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
706 /** Set the size of the data for a leaf node */
707 #define SETDSZ(node,size) do { \
708 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
709 /** The size of a key in a node */
710 #define NODEKSZ(node) ((node)->mn_ksize)
712 /** Copy a page number from src to dst */
714 #define COPY_PGNO(dst,src) dst = src
716 #if SIZE_MAX > 4294967295UL
717 #define COPY_PGNO(dst,src) do { \
718 unsigned short *s, *d; \
719 s = (unsigned short *)&(src); \
720 d = (unsigned short *)&(dst); \
727 #define COPY_PGNO(dst,src) do { \
728 unsigned short *s, *d; \
729 s = (unsigned short *)&(src); \
730 d = (unsigned short *)&(dst); \
736 /** The address of a key in a LEAF2 page.
737 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
738 * There are no node headers, keys are stored contiguously.
740 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
742 /** Set the \b node's key into \b key, if requested. */
743 #define MDB_SET_KEY(node, key) { if ((key) != NULL) { \
744 (key)->mv_size = NODEKSZ(node); (key)->mv_data = NODEKEY(node); } }
746 /** Information about a single database in the environment. */
747 typedef struct MDB_db {
748 uint32_t md_pad; /**< also ksize for LEAF2 pages */
749 uint16_t md_flags; /**< @ref mdb_open */
750 uint16_t md_depth; /**< depth of this tree */
751 pgno_t md_branch_pages; /**< number of internal pages */
752 pgno_t md_leaf_pages; /**< number of leaf pages */
753 pgno_t md_overflow_pages; /**< number of overflow pages */
754 size_t md_entries; /**< number of data items */
755 pgno_t md_root; /**< the root page of this tree */
758 /** Handle for the DB used to track free pages. */
760 /** Handle for the default DB. */
763 /** Meta page content. */
764 typedef struct MDB_meta {
765 /** Stamp identifying this as an MDB file. It must be set
768 /** Version number of this lock file. Must be set to #MDB_VERSION. */
770 void *mm_address; /**< address for fixed mapping */
771 size_t mm_mapsize; /**< size of mmap region */
772 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
773 /** The size of pages used in this DB */
774 #define mm_psize mm_dbs[0].md_pad
775 /** Any persistent environment flags. @ref mdb_env */
776 #define mm_flags mm_dbs[0].md_flags
777 pgno_t mm_last_pg; /**< last used page in file */
778 txnid_t mm_txnid; /**< txnid that committed this page */
781 /** Buffer for a stack-allocated dirty page.
782 * The members define size and alignment, and silence type
783 * aliasing warnings. They are not used directly; that could
784 * mean incorrectly using several union members in parallel.
786 typedef union MDB_pagebuf {
787 char mb_raw[MDB_PAGESIZE];
790 char mm_pad[PAGEHDRSZ];
795 /** Auxiliary DB info.
796 * The information here is mostly static/read-only. There is
797 * only a single copy of this record in the environment.
799 typedef struct MDB_dbx {
800 MDB_val md_name; /**< name of the database */
801 MDB_cmp_func *md_cmp; /**< function for comparing keys */
802 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
803 MDB_rel_func *md_rel; /**< user relocate function */
804 void *md_relctx; /**< user-provided context for md_rel */
807 /** A database transaction.
808 * Every operation requires a transaction handle.
811 MDB_txn *mt_parent; /**< parent of a nested txn */
812 MDB_txn *mt_child; /**< nested txn under this txn */
813 pgno_t mt_next_pgno; /**< next unallocated page */
814 /** The ID of this transaction. IDs are integers incrementing from 1.
815 * Only committed write transactions increment the ID. If a transaction
816 * aborts, the ID may be re-used by the next writer.
819 MDB_env *mt_env; /**< the DB environment */
820 /** The list of pages that became unused during this transaction.
824 ID2L dirty_list; /**< modified pages */
825 MDB_reader *reader; /**< this thread's slot in the reader table */
827 /** Array of records for each DB known in the environment. */
829 /** Array of MDB_db records for each known DB */
831 /** @defgroup mt_dbflag Transaction DB Flags
835 #define DB_DIRTY 0x01 /**< DB was written in this txn */
836 #define DB_STALE 0x02 /**< DB record is older than txnID */
838 /** Array of cursors for each DB */
839 MDB_cursor **mt_cursors;
840 /** Array of flags for each DB */
841 unsigned char *mt_dbflags;
842 /** Number of DB records in use. This number only ever increments;
843 * we don't decrement it when individual DB handles are closed.
847 /** @defgroup mdb_txn Transaction Flags
851 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
852 #define MDB_TXN_ERROR 0x02 /**< an error has occurred */
854 unsigned int mt_flags; /**< @ref mdb_txn */
855 /** Tracks which of the two meta pages was used at the start
856 * of this transaction.
858 unsigned int mt_toggle;
861 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
862 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
863 * raise this on a 64 bit machine.
865 #define CURSOR_STACK 32
869 /** Cursors are used for all DB operations */
871 /** Next cursor on this DB in this txn */
873 /** Original cursor if this is a shadow */
875 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
876 struct MDB_xcursor *mc_xcursor;
877 /** The transaction that owns this cursor */
879 /** The database handle this cursor operates on */
881 /** The database record for this cursor */
883 /** The database auxiliary record for this cursor */
885 /** The @ref mt_dbflag for this database */
886 unsigned char *mc_dbflag;
887 unsigned short mc_snum; /**< number of pushed pages */
888 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
889 /** @defgroup mdb_cursor Cursor Flags
891 * Cursor state flags.
894 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
895 #define C_EOF 0x02 /**< No more data */
896 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
897 #define C_SHADOW 0x08 /**< Cursor is a dup from a parent txn */
898 #define C_ALLOCD 0x10 /**< Cursor was malloc'd */
900 unsigned int mc_flags; /**< @ref mdb_cursor */
901 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
902 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
905 /** Context for sorted-dup records.
906 * We could have gone to a fully recursive design, with arbitrarily
907 * deep nesting of sub-databases. But for now we only handle these
908 * levels - main DB, optional sub-DB, sorted-duplicate DB.
910 typedef struct MDB_xcursor {
911 /** A sub-cursor for traversing the Dup DB */
912 MDB_cursor mx_cursor;
913 /** The database record for this Dup DB */
915 /** The auxiliary DB record for this Dup DB */
917 /** The @ref mt_dbflag for this Dup DB */
918 unsigned char mx_dbflag;
921 /** A set of pages freed by an earlier transaction. */
922 typedef struct MDB_oldpages {
923 /** Usually we only read one record from the FREEDB at a time, but
924 * in case we read more, this will chain them together.
926 struct MDB_oldpages *mo_next;
927 /** The ID of the transaction in which these pages were freed. */
929 /** An #IDL of the pages */
930 pgno_t mo_pages[1]; /* dynamic */
933 /** The database environment. */
935 HANDLE me_fd; /**< The main data file */
936 HANDLE me_lfd; /**< The lock file */
937 HANDLE me_mfd; /**< just for writing the meta pages */
938 /** Failed to update the meta page. Probably an I/O error. */
939 #define MDB_FATAL_ERROR 0x80000000U
940 uint32_t me_flags; /**< @ref mdb_env */
941 uint32_t me_extrapad; /**< unused for now */
942 unsigned int me_maxreaders; /**< size of the reader table */
943 MDB_dbi me_numdbs; /**< number of DBs opened */
944 MDB_dbi me_maxdbs; /**< size of the DB table */
945 char *me_path; /**< path to the DB files */
946 char *me_map; /**< the memory map of the data file */
947 MDB_txninfo *me_txns; /**< the memory map of the lock file */
948 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
949 MDB_txn *me_txn; /**< current write transaction */
950 size_t me_mapsize; /**< size of the data memory map */
951 off_t me_size; /**< current file size */
952 pgno_t me_maxpg; /**< me_mapsize / me_psize */
953 unsigned int me_psize; /**< size of a page, from #GET_PAGESIZE */
954 unsigned int me_db_toggle; /**< which DB table is current */
955 txnid_t me_wtxnid; /**< ID of last txn we committed */
956 txnid_t me_pgfirst; /**< ID of first old page record we used */
957 txnid_t me_pglast; /**< ID of last old page record we used */
958 MDB_dbx *me_dbxs; /**< array of static DB info */
959 MDB_db *me_dbs[2]; /**< two arrays of MDB_db info */
960 MDB_oldpages *me_pghead; /**< list of old page records */
961 pthread_key_t me_txkey; /**< thread-key for readers */
962 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
963 /** IDL of pages that became unused in a write txn */
965 /** ID2L of pages that were written during a write txn */
966 ID2 me_dirty_list[MDB_IDL_UM_SIZE];
967 /** rwlock for the DB tables, if #LAZY_LOCKS is false */
968 LAZY_RWLOCK_DEF(me_dblock)
970 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
974 sem_t *me_rmutex; /* Apple doesn't support shared mutexes */
978 /** max number of pages to commit in one writev() call */
979 #define MDB_COMMIT_PAGES 64
980 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
981 #undef MDB_COMMIT_PAGES
982 #define MDB_COMMIT_PAGES IOV_MAX
985 static MDB_page *mdb_page_alloc(MDB_cursor *mc, int num);
986 static MDB_page *mdb_page_new(MDB_cursor *mc, uint32_t flags, int num);
987 static int mdb_page_touch(MDB_cursor *mc);
989 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp);
990 static int mdb_page_search_root(MDB_cursor *mc,
991 MDB_val *key, int modify);
992 static int mdb_page_search(MDB_cursor *mc,
993 MDB_val *key, int modify);
994 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
995 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
996 pgno_t newpgno, unsigned int nflags);
998 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
999 static int mdb_env_read_meta(MDB_env *env, int *which);
1000 static int mdb_env_write_meta(MDB_txn *txn);
1002 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1003 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1004 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1005 static void mdb_node_del(MDB_page *mp, indx_t indx, int ksize);
1006 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1007 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1008 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1009 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1010 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1012 static int mdb_rebalance(MDB_cursor *mc);
1013 static int mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key);
1015 static void mdb_cursor_pop(MDB_cursor *mc);
1016 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1018 static int mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf);
1019 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1020 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1021 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1022 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1024 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1025 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1027 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1028 static void mdb_xcursor_init0(MDB_cursor *mc);
1029 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1031 static int mdb_drop0(MDB_cursor *mc, int subs);
1032 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1035 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1039 static SECURITY_DESCRIPTOR mdb_null_sd;
1040 static SECURITY_ATTRIBUTES mdb_all_sa;
1041 static int mdb_sec_inited;
1044 /** Return the library version info. */
1046 mdb_version(int *major, int *minor, int *patch)
1048 if (major) *major = MDB_VERSION_MAJOR;
1049 if (minor) *minor = MDB_VERSION_MINOR;
1050 if (patch) *patch = MDB_VERSION_PATCH;
1051 return MDB_VERSION_STRING;
1054 /** Table of descriptions for MDB @ref errors */
1055 static char *const mdb_errstr[] = {
1056 "MDB_KEYEXIST: Key/data pair already exists",
1057 "MDB_NOTFOUND: No matching key/data pair found",
1058 "MDB_PAGE_NOTFOUND: Requested page not found",
1059 "MDB_CORRUPTED: Located page was wrong type",
1060 "MDB_PANIC: Update of meta page failed",
1061 "MDB_VERSION_MISMATCH: Database environment version mismatch"
1065 mdb_strerror(int err)
1068 return ("Successful return: 0");
1070 if (err >= MDB_KEYEXIST && err <= MDB_VERSION_MISMATCH)
1071 return mdb_errstr[err - MDB_KEYEXIST];
1073 return strerror(err);
1077 /** Display a key in hexadecimal and return the address of the result.
1078 * @param[in] key the key to display
1079 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1080 * @return The key in hexadecimal form.
1083 mdb_dkey(MDB_val *key, char *buf)
1086 unsigned char *c = key->mv_data;
1088 if (key->mv_size > MAXKEYSIZE)
1089 return "MAXKEYSIZE";
1090 /* may want to make this a dynamic check: if the key is mostly
1091 * printable characters, print it as-is instead of converting to hex.
1095 for (i=0; i<key->mv_size; i++)
1096 ptr += sprintf(ptr, "%02x", *c++);
1098 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1103 /** Display all the keys in the page. */
1105 mdb_page_keys(MDB_page *mp)
1108 unsigned int i, nkeys;
1112 nkeys = NUMKEYS(mp);
1113 DPRINTF("numkeys %d", nkeys);
1114 for (i=0; i<nkeys; i++) {
1115 node = NODEPTR(mp, i);
1116 key.mv_size = node->mn_ksize;
1117 key.mv_data = node->mn_data;
1118 DPRINTF("key %d: %s", i, DKEY(&key));
1124 /** Count all the pages in each DB and in the freelist
1125 * and make sure it matches the actual number of pages
1128 static void mdb_audit(MDB_txn *txn)
1133 ID freecount, count;
1136 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1137 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1138 freecount += *(ID *)data.mv_data;
1139 freecount += txn->mt_dbs[0].md_branch_pages + txn->mt_dbs[0].md_leaf_pages +
1140 txn->mt_dbs[0].md_overflow_pages;
1143 for (i = 0; i<txn->mt_numdbs; i++) {
1144 count += txn->mt_dbs[i].md_branch_pages +
1145 txn->mt_dbs[i].md_leaf_pages +
1146 txn->mt_dbs[i].md_overflow_pages;
1147 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1149 mdb_cursor_init(&mc, txn, i, &mx);
1150 mdb_page_search(&mc, NULL, 0);
1154 mp = mc.mc_pg[mc.mc_top];
1155 for (j=0; j<NUMKEYS(mp); j++) {
1156 MDB_node *leaf = NODEPTR(mp, j);
1157 if (leaf->mn_flags & F_SUBDATA) {
1159 memcpy(&db, NODEDATA(leaf), sizeof(db));
1160 count += db.md_branch_pages + db.md_leaf_pages +
1161 db.md_overflow_pages;
1165 while (mdb_cursor_sibling(&mc, 1) == 0);
1168 assert(freecount + count + 2 >= txn->mt_next_pgno - 1);
1173 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1175 return txn->mt_dbxs[dbi].md_cmp(a, b);
1179 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1181 if (txn->mt_dbxs[dbi].md_dcmp)
1182 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1184 return EINVAL; /* too bad you can't distinguish this from a valid result */
1187 /** Allocate a single page.
1188 * Re-use old malloc'd pages first, otherwise just malloc.
1191 mdb_page_malloc(MDB_cursor *mc) {
1193 size_t sz = mc->mc_txn->mt_env->me_psize;
1194 if ((ret = mc->mc_txn->mt_env->me_dpages) != NULL) {
1195 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1196 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1197 mc->mc_txn->mt_env->me_dpages = ret->mp_next;
1198 } else if ((ret = malloc(sz)) != NULL) {
1199 VGMEMP_ALLOC(mc->mc_txn->mt_env, ret, sz);
1204 /** Allocate pages for writing.
1205 * If there are free pages available from older transactions, they
1206 * will be re-used first. Otherwise a new page will be allocated.
1207 * @param[in] mc cursor A cursor handle identifying the transaction and
1208 * database for which we are allocating.
1209 * @param[in] num the number of pages to allocate.
1210 * @return Address of the allocated page(s). Requests for multiple pages
1211 * will always be satisfied by a single contiguous chunk of memory.
1214 mdb_page_alloc(MDB_cursor *mc, int num)
1216 MDB_txn *txn = mc->mc_txn;
1218 pgno_t pgno = P_INVALID;
1221 if (txn->mt_txnid > 2) {
1223 if (!txn->mt_env->me_pghead &&
1224 txn->mt_dbs[FREE_DBI].md_root != P_INVALID) {
1225 /* See if there's anything in the free DB */
1229 txnid_t *kptr, oldest, last;
1231 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1232 if (!txn->mt_env->me_pgfirst) {
1233 mdb_page_search(&m2, NULL, 0);
1234 leaf = NODEPTR(m2.mc_pg[m2.mc_top], 0);
1235 kptr = (txnid_t *)NODEKEY(leaf);
1242 last = txn->mt_env->me_pglast + 1;
1244 key.mv_data = &last;
1245 key.mv_size = sizeof(last);
1246 rc = mdb_cursor_set(&m2, &key, &data, MDB_SET, &exact);
1249 last = *(txnid_t *)key.mv_data;
1254 oldest = txn->mt_txnid - 1;
1255 for (i=0; i<txn->mt_env->me_txns->mti_numreaders; i++) {
1256 txnid_t mr = txn->mt_env->me_txns->mti_readers[i].mr_txnid;
1257 if (mr && mr < oldest)
1262 if (oldest > last) {
1263 /* It's usable, grab it.
1268 if (!txn->mt_env->me_pgfirst) {
1269 mdb_node_read(txn, leaf, &data);
1271 txn->mt_env->me_pglast = last;
1272 if (!txn->mt_env->me_pgfirst)
1273 txn->mt_env->me_pgfirst = last;
1274 idl = (ID *) data.mv_data;
1275 /* We might have a zero-length IDL due to freelist growth
1276 * during a prior commit
1278 if (!idl[0]) goto again;
1279 mop = malloc(sizeof(MDB_oldpages) + MDB_IDL_SIZEOF(idl) - sizeof(pgno_t));
1280 mop->mo_next = txn->mt_env->me_pghead;
1281 mop->mo_txnid = last;
1282 txn->mt_env->me_pghead = mop;
1283 memcpy(mop->mo_pages, idl, MDB_IDL_SIZEOF(idl));
1288 DPRINTF("IDL read txn %zu root %zu num %zu",
1289 mop->mo_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1290 for (i=0; i<idl[0]; i++) {
1291 DPRINTF("IDL %zu", idl[i+1]);
1298 if (txn->mt_env->me_pghead) {
1299 MDB_oldpages *mop = txn->mt_env->me_pghead;
1301 /* FIXME: For now, always use fresh pages. We
1302 * really ought to search the free list for a
1307 /* peel pages off tail, so we only have to truncate the list */
1308 pgno = MDB_IDL_LAST(mop->mo_pages);
1309 if (MDB_IDL_IS_RANGE(mop->mo_pages)) {
1311 if (mop->mo_pages[2] > mop->mo_pages[1])
1312 mop->mo_pages[0] = 0;
1316 if (MDB_IDL_IS_ZERO(mop->mo_pages)) {
1317 txn->mt_env->me_pghead = mop->mo_next;
1324 if (pgno == P_INVALID) {
1325 /* DB size is maxed out */
1326 if (txn->mt_next_pgno + num >= txn->mt_env->me_maxpg) {
1327 DPUTS("DB size maxed out");
1331 if (txn->mt_env->me_dpages && num == 1) {
1332 np = txn->mt_env->me_dpages;
1333 VGMEMP_ALLOC(txn->mt_env, np, txn->mt_env->me_psize);
1334 VGMEMP_DEFINED(np, sizeof(np->mp_next));
1335 txn->mt_env->me_dpages = np->mp_next;
1337 size_t sz = txn->mt_env->me_psize * num;
1338 if ((np = malloc(sz)) == NULL)
1340 VGMEMP_ALLOC(txn->mt_env, np, sz);
1342 if (pgno == P_INVALID) {
1343 np->mp_pgno = txn->mt_next_pgno;
1344 txn->mt_next_pgno += num;
1348 mid.mid = np->mp_pgno;
1350 mdb_mid2l_insert(txn->mt_u.dirty_list, &mid);
1355 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
1356 * @param[in] mc cursor pointing to the page to be touched
1357 * @return 0 on success, non-zero on failure.
1360 mdb_page_touch(MDB_cursor *mc)
1362 MDB_page *mp = mc->mc_pg[mc->mc_top];
1365 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
1367 if ((np = mdb_page_alloc(mc, 1)) == NULL)
1369 DPRINTF("touched db %u page %zu -> %zu", mc->mc_dbi, mp->mp_pgno, np->mp_pgno);
1370 assert(mp->mp_pgno != np->mp_pgno);
1371 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
1373 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1376 mp->mp_flags |= P_DIRTY;
1379 /* Adjust other cursors pointing to mp */
1380 if (mc->mc_flags & C_SUB) {
1381 MDB_cursor *m2, *m3;
1382 MDB_dbi dbi = mc->mc_dbi-1;
1384 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
1385 if (m2 == mc) continue;
1386 m3 = &m2->mc_xcursor->mx_cursor;
1387 if (m3->mc_snum < mc->mc_snum) continue;
1388 if (m3->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1389 m3->mc_pg[mc->mc_top] = mp;
1395 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
1396 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
1397 if (m2->mc_pg[mc->mc_top] == mc->mc_pg[mc->mc_top]) {
1398 m2->mc_pg[mc->mc_top] = mp;
1402 mc->mc_pg[mc->mc_top] = mp;
1403 /** If this page has a parent, update the parent to point to
1407 SETPGNO(NODEPTR(mc->mc_pg[mc->mc_top-1], mc->mc_ki[mc->mc_top-1]), mp->mp_pgno);
1409 mc->mc_db->md_root = mp->mp_pgno;
1410 } else if (mc->mc_txn->mt_parent) {
1413 /* If txn has a parent, make sure the page is in our
1416 if (mc->mc_txn->mt_u.dirty_list[0].mid) {
1417 unsigned x = mdb_mid2l_search(mc->mc_txn->mt_u.dirty_list, mp->mp_pgno);
1418 if (x <= mc->mc_txn->mt_u.dirty_list[0].mid &&
1419 mc->mc_txn->mt_u.dirty_list[x].mid == mp->mp_pgno) {
1420 if (mc->mc_txn->mt_u.dirty_list[x].mptr != mp) {
1421 mp = mc->mc_txn->mt_u.dirty_list[x].mptr;
1422 mc->mc_pg[mc->mc_top] = mp;
1428 np = mdb_page_malloc(mc);
1429 memcpy(np, mp, mc->mc_txn->mt_env->me_psize);
1430 mid.mid = np->mp_pgno;
1432 mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &mid);
1440 mdb_env_sync(MDB_env *env, int force)
1443 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
1444 if (MDB_FDATASYNC(env->me_fd))
1450 /** Make shadow copies of all of parent txn's cursors */
1452 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
1454 MDB_cursor *mc, *m2;
1455 unsigned int i, j, size;
1457 for (i=0;i<src->mt_numdbs; i++) {
1458 if (src->mt_cursors[i]) {
1459 size = sizeof(MDB_cursor);
1460 if (src->mt_cursors[i]->mc_xcursor)
1461 size += sizeof(MDB_xcursor);
1462 for (m2 = src->mt_cursors[i]; m2; m2=m2->mc_next) {
1469 mc->mc_db = &dst->mt_dbs[i];
1470 mc->mc_dbx = m2->mc_dbx;
1471 mc->mc_dbflag = &dst->mt_dbflags[i];
1472 mc->mc_snum = m2->mc_snum;
1473 mc->mc_top = m2->mc_top;
1474 mc->mc_flags = m2->mc_flags | C_SHADOW;
1475 for (j=0; j<mc->mc_snum; j++) {
1476 mc->mc_pg[j] = m2->mc_pg[j];
1477 mc->mc_ki[j] = m2->mc_ki[j];
1479 if (m2->mc_xcursor) {
1480 MDB_xcursor *mx, *mx2;
1481 mx = (MDB_xcursor *)(mc+1);
1482 mc->mc_xcursor = mx;
1483 mx2 = m2->mc_xcursor;
1484 mx->mx_db = mx2->mx_db;
1485 mx->mx_dbx = mx2->mx_dbx;
1486 mx->mx_dbflag = mx2->mx_dbflag;
1487 mx->mx_cursor.mc_txn = dst;
1488 mx->mx_cursor.mc_dbi = mx2->mx_cursor.mc_dbi;
1489 mx->mx_cursor.mc_db = &mx->mx_db;
1490 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
1491 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
1492 mx->mx_cursor.mc_snum = mx2->mx_cursor.mc_snum;
1493 mx->mx_cursor.mc_top = mx2->mx_cursor.mc_top;
1494 mx->mx_cursor.mc_flags = mx2->mx_cursor.mc_flags | C_SHADOW;
1495 for (j=0; j<mx2->mx_cursor.mc_snum; j++) {
1496 mx->mx_cursor.mc_pg[j] = mx2->mx_cursor.mc_pg[j];
1497 mx->mx_cursor.mc_ki[j] = mx2->mx_cursor.mc_ki[j];
1500 mc->mc_xcursor = NULL;
1502 mc->mc_next = dst->mt_cursors[i];
1503 dst->mt_cursors[i] = mc;
1510 /** Merge shadow cursors back into parent's */
1512 mdb_cursor_merge(MDB_txn *txn)
1515 for (i=0; i<txn->mt_numdbs; i++) {
1516 if (txn->mt_cursors[i]) {
1518 while ((mc = txn->mt_cursors[i])) {
1519 txn->mt_cursors[i] = mc->mc_next;
1520 if (mc->mc_flags & C_SHADOW) {
1521 MDB_cursor *m2 = mc->mc_orig;
1523 m2->mc_snum = mc->mc_snum;
1524 m2->mc_top = mc->mc_top;
1525 for (j=0; j<mc->mc_snum; j++) {
1526 m2->mc_pg[j] = mc->mc_pg[j];
1527 m2->mc_ki[j] = mc->mc_ki[j];
1530 if (mc->mc_flags & C_ALLOCD)
1538 mdb_txn_reset0(MDB_txn *txn);
1540 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
1541 * @param[in] txn the transaction handle to initialize
1542 * @return 0 on success, non-zero on failure. This can only
1543 * fail for read-only transactions, and then only if the
1544 * reader table is full.
1547 mdb_txn_renew0(MDB_txn *txn)
1549 MDB_env *env = txn->mt_env;
1552 if (txn->mt_flags & MDB_TXN_RDONLY) {
1553 MDB_reader *r = pthread_getspecific(env->me_txkey);
1556 pid_t pid = getpid();
1557 pthread_t tid = pthread_self();
1560 for (i=0; i<env->me_txns->mti_numreaders; i++)
1561 if (env->me_txns->mti_readers[i].mr_pid == 0)
1563 if (i == env->me_maxreaders) {
1564 UNLOCK_MUTEX_R(env);
1567 env->me_txns->mti_readers[i].mr_pid = pid;
1568 env->me_txns->mti_readers[i].mr_tid = tid;
1569 if (i >= env->me_txns->mti_numreaders)
1570 env->me_txns->mti_numreaders = i+1;
1571 UNLOCK_MUTEX_R(env);
1572 r = &env->me_txns->mti_readers[i];
1573 pthread_setspecific(env->me_txkey, r);
1575 txn->mt_toggle = env->me_txns->mti_me_toggle;
1576 txn->mt_txnid = r->mr_txnid = env->me_metas[txn->mt_toggle]->mm_txnid;
1578 /* This happens if a different process was the
1579 * last writer to the DB.
1581 if (env->me_wtxnid < txn->mt_txnid)
1582 mt_dbflag = DB_STALE;
1583 txn->mt_u.reader = r;
1587 txn->mt_toggle = env->me_txns->mti_me_toggle;
1588 txn->mt_txnid = env->me_metas[txn->mt_toggle]->mm_txnid;
1589 if (env->me_wtxnid < txn->mt_txnid)
1590 mt_dbflag = DB_STALE;
1593 if (txn->mt_txnid == mdb_debug_start)
1596 txn->mt_u.dirty_list = env->me_dirty_list;
1597 txn->mt_u.dirty_list[0].mid = 0;
1598 txn->mt_free_pgs = env->me_free_pgs;
1599 txn->mt_free_pgs[0] = 0;
1600 txn->mt_next_pgno = env->me_metas[txn->mt_toggle]->mm_last_pg+1;
1604 /* Copy the DB arrays */
1605 LAZY_RWLOCK_RDLOCK(&env->me_dblock);
1606 txn->mt_numdbs = env->me_numdbs;
1607 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
1608 memcpy(txn->mt_dbs, env->me_metas[txn->mt_toggle]->mm_dbs, 2 * sizeof(MDB_db));
1609 if (txn->mt_numdbs > 2)
1610 memcpy(txn->mt_dbs+2, env->me_dbs[env->me_db_toggle]+2,
1611 (txn->mt_numdbs - 2) * sizeof(MDB_db));
1612 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1614 memset(txn->mt_dbflags, mt_dbflag, env->me_numdbs);
1620 mdb_txn_renew(MDB_txn *txn)
1627 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
1628 DPUTS("environment had fatal error, must shutdown!");
1632 rc = mdb_txn_renew0(txn);
1633 if (rc == MDB_SUCCESS) {
1634 DPRINTF("renew txn %zu%c %p on mdbenv %p, root page %zu",
1635 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1636 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1642 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
1647 if (env->me_flags & MDB_FATAL_ERROR) {
1648 DPUTS("environment had fatal error, must shutdown!");
1652 /* parent already has an active child txn */
1653 if (parent->mt_child) {
1657 size = sizeof(MDB_txn) + env->me_maxdbs * (sizeof(MDB_db)+1);
1658 if (!(flags & MDB_RDONLY))
1659 size += env->me_maxdbs * sizeof(MDB_cursor *);
1661 if ((txn = calloc(1, size)) == NULL) {
1662 DPRINTF("calloc: %s", strerror(ErrCode()));
1665 txn->mt_dbs = (MDB_db *)(txn+1);
1666 if (flags & MDB_RDONLY) {
1667 txn->mt_flags |= MDB_TXN_RDONLY;
1668 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
1670 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
1671 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
1676 txn->mt_free_pgs = mdb_midl_alloc();
1677 if (!txn->mt_free_pgs) {
1681 txn->mt_u.dirty_list = malloc(sizeof(ID2)*MDB_IDL_UM_SIZE);
1682 if (!txn->mt_u.dirty_list) {
1683 free(txn->mt_free_pgs);
1687 txn->mt_txnid = parent->mt_txnid;
1688 txn->mt_toggle = parent->mt_toggle;
1689 txn->mt_u.dirty_list[0].mid = 0;
1690 txn->mt_free_pgs[0] = 0;
1691 txn->mt_next_pgno = parent->mt_next_pgno;
1692 parent->mt_child = txn;
1693 txn->mt_parent = parent;
1694 txn->mt_numdbs = parent->mt_numdbs;
1695 txn->mt_dbxs = parent->mt_dbxs;
1696 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
1697 memcpy(txn->mt_dbflags, parent->mt_dbflags, txn->mt_numdbs);
1698 mdb_cursor_shadow(parent, txn);
1701 rc = mdb_txn_renew0(txn);
1707 DPRINTF("begin txn %zu%c %p on mdbenv %p, root page %zu",
1708 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1709 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root);
1715 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
1716 * @param[in] txn the transaction handle to reset
1719 mdb_txn_reset0(MDB_txn *txn)
1721 MDB_env *env = txn->mt_env;
1723 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1724 txn->mt_u.reader->mr_txnid = 0;
1730 /* close(free) all cursors */
1731 for (i=0; i<txn->mt_numdbs; i++) {
1732 if (txn->mt_cursors[i]) {
1734 while ((mc = txn->mt_cursors[i])) {
1735 txn->mt_cursors[i] = mc->mc_next;
1736 if (mc->mc_flags & C_ALLOCD)
1742 /* return all dirty pages to dpage list */
1743 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
1744 dp = txn->mt_u.dirty_list[i].mptr;
1745 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1746 dp->mp_next = txn->mt_env->me_dpages;
1747 VGMEMP_FREE(txn->mt_env, dp);
1748 txn->mt_env->me_dpages = dp;
1750 /* large pages just get freed directly */
1751 VGMEMP_FREE(txn->mt_env, dp);
1756 if (txn->mt_parent) {
1757 txn->mt_parent->mt_child = NULL;
1758 free(txn->mt_free_pgs);
1759 free(txn->mt_u.dirty_list);
1762 if (mdb_midl_shrink(&txn->mt_free_pgs))
1763 env->me_free_pgs = txn->mt_free_pgs;
1766 while ((mop = txn->mt_env->me_pghead)) {
1767 txn->mt_env->me_pghead = mop->mo_next;
1770 txn->mt_env->me_pgfirst = 0;
1771 txn->mt_env->me_pglast = 0;
1774 /* The writer mutex was locked in mdb_txn_begin. */
1775 UNLOCK_MUTEX_W(env);
1780 mdb_txn_reset(MDB_txn *txn)
1785 DPRINTF("reset txn %zu%c %p on mdbenv %p, root page %zu",
1786 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1787 (void *) txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1789 mdb_txn_reset0(txn);
1793 mdb_txn_abort(MDB_txn *txn)
1798 DPRINTF("abort txn %zu%c %p on mdbenv %p, root page %zu",
1799 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
1800 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root);
1803 mdb_txn_abort(txn->mt_child);
1805 mdb_txn_reset0(txn);
1810 mdb_txn_commit(MDB_txn *txn)
1818 pgno_t next, freecnt;
1821 assert(txn != NULL);
1822 assert(txn->mt_env != NULL);
1824 if (txn->mt_child) {
1825 mdb_txn_commit(txn->mt_child);
1826 txn->mt_child = NULL;
1831 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
1832 if (txn->mt_numdbs > env->me_numdbs) {
1833 /* update the DB tables */
1834 int toggle = !env->me_db_toggle;
1838 ip = &env->me_dbs[toggle][env->me_numdbs];
1839 jp = &txn->mt_dbs[env->me_numdbs];
1840 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
1841 for (i = env->me_numdbs; i < txn->mt_numdbs; i++) {
1845 env->me_db_toggle = toggle;
1846 env->me_numdbs = txn->mt_numdbs;
1847 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
1853 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
1854 DPUTS("error flag is set, can't commit");
1856 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
1861 /* Merge (and close) our cursors with parent's */
1862 mdb_cursor_merge(txn);
1864 if (txn->mt_parent) {
1870 /* Update parent's DB table */
1871 ip = &txn->mt_parent->mt_dbs[2];
1872 jp = &txn->mt_dbs[2];
1873 for (i = 2; i < txn->mt_numdbs; i++) {
1874 if (ip->md_root != jp->md_root)
1878 txn->mt_parent->mt_numdbs = txn->mt_numdbs;
1880 /* Append our free list to parent's */
1881 mdb_midl_append_list(&txn->mt_parent->mt_free_pgs,
1883 mdb_midl_free(txn->mt_free_pgs);
1885 /* Merge our dirty list with parent's */
1886 dst = txn->mt_parent->mt_u.dirty_list;
1887 src = txn->mt_u.dirty_list;
1888 x = mdb_mid2l_search(dst, src[1].mid);
1889 for (y=1; y<=src[0].mid; y++) {
1890 while (x <= dst[0].mid && dst[x].mid != src[y].mid) x++;
1894 dst[x].mptr = src[y].mptr;
1897 for (; y<=src[0].mid; y++) {
1898 if (++x >= MDB_IDL_UM_MAX) {
1905 free(txn->mt_u.dirty_list);
1906 txn->mt_parent->mt_child = NULL;
1911 if (txn != env->me_txn) {
1912 DPUTS("attempt to commit unknown transaction");
1917 if (!txn->mt_u.dirty_list[0].mid)
1920 DPRINTF("committing txn %zu %p on mdbenv %p, root page %zu",
1921 txn->mt_txnid, (void *)txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root);
1923 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1925 /* should only be one record now */
1926 if (env->me_pghead) {
1927 /* make sure first page of freeDB is touched and on freelist */
1928 mdb_page_search(&mc, NULL, 1);
1931 /* Delete IDLs we used from the free list */
1932 if (env->me_pgfirst) {
1937 key.mv_size = sizeof(cur);
1938 for (cur = env->me_pgfirst; cur <= env->me_pglast; cur++) {
1941 mdb_cursor_set(&mc, &key, NULL, MDB_SET, &exact);
1942 mdb_cursor_del(&mc, 0);
1944 env->me_pgfirst = 0;
1948 /* save to free list */
1950 freecnt = txn->mt_free_pgs[0];
1951 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
1954 /* make sure last page of freeDB is touched and on freelist */
1955 key.mv_size = MAXKEYSIZE+1;
1957 mdb_page_search(&mc, &key, 1);
1959 mdb_midl_sort(txn->mt_free_pgs);
1963 ID *idl = txn->mt_free_pgs;
1964 DPRINTF("IDL write txn %zu root %zu num %zu",
1965 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, idl[0]);
1966 for (i=0; i<idl[0]; i++) {
1967 DPRINTF("IDL %zu", idl[i+1]);
1971 /* write to last page of freeDB */
1972 key.mv_size = sizeof(pgno_t);
1973 key.mv_data = &txn->mt_txnid;
1974 data.mv_data = txn->mt_free_pgs;
1975 /* The free list can still grow during this call,
1976 * despite the pre-emptive touches above. So check
1977 * and make sure the entire thing got written.
1980 freecnt = txn->mt_free_pgs[0];
1981 data.mv_size = MDB_IDL_SIZEOF(txn->mt_free_pgs);
1982 rc = mdb_cursor_put(&mc, &key, &data, 0);
1987 } while (freecnt != txn->mt_free_pgs[0]);
1989 /* should only be one record now */
1991 if (env->me_pghead) {
1997 mop = env->me_pghead;
1999 key.mv_size = sizeof(id);
2001 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2002 data.mv_data = mop->mo_pages;
2003 orig = mop->mo_pages[0];
2004 /* These steps may grow the freelist again
2005 * due to freed overflow pages...
2007 mdb_cursor_put(&mc, &key, &data, 0);
2008 if (mop == env->me_pghead && env->me_pghead->mo_txnid == id) {
2009 /* could have been used again here */
2010 if (mop->mo_pages[0] != orig) {
2011 data.mv_size = MDB_IDL_SIZEOF(mop->mo_pages);
2012 data.mv_data = mop->mo_pages;
2014 mdb_cursor_put(&mc, &key, &data, 0);
2016 env->me_pghead = NULL;
2019 /* was completely used up */
2020 mdb_cursor_del(&mc, 0);
2024 env->me_pgfirst = 0;
2027 /* Check for growth of freelist again */
2028 if (freecnt != txn->mt_free_pgs[0])
2031 if (!MDB_IDL_IS_ZERO(txn->mt_free_pgs)) {
2032 if (mdb_midl_shrink(&txn->mt_free_pgs))
2033 env->me_free_pgs = txn->mt_free_pgs;
2036 /* Update DB root pointers. Their pages have already been
2037 * touched so this is all in-place and cannot fail.
2042 data.mv_size = sizeof(MDB_db);
2044 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
2045 for (i = 2; i < txn->mt_numdbs; i++) {
2046 if (txn->mt_dbflags[i] & DB_DIRTY) {
2047 data.mv_data = &txn->mt_dbs[i];
2048 mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
2056 /* Commit up to MDB_COMMIT_PAGES dirty pages to disk until done.
2062 /* Windows actually supports scatter/gather I/O, but only on
2063 * unbuffered file handles. Since we're relying on the OS page
2064 * cache for all our data, that's self-defeating. So we just
2065 * write pages one at a time. We use the ov structure to set
2066 * the write offset, to at least save the overhead of a Seek
2070 memset(&ov, 0, sizeof(ov));
2071 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2073 dp = txn->mt_u.dirty_list[i].mptr;
2074 DPRINTF("committing page %zu", dp->mp_pgno);
2075 size = dp->mp_pgno * env->me_psize;
2076 ov.Offset = size & 0xffffffff;
2077 ov.OffsetHigh = size >> 16;
2078 ov.OffsetHigh >>= 16;
2079 /* clear dirty flag */
2080 dp->mp_flags &= ~P_DIRTY;
2081 wsize = env->me_psize;
2082 if (IS_OVERFLOW(dp)) wsize *= dp->mp_pages;
2083 rc = WriteFile(env->me_fd, dp, wsize, NULL, &ov);
2086 DPRINTF("WriteFile: %d", n);
2093 struct iovec iov[MDB_COMMIT_PAGES];
2097 for (; i<=txn->mt_u.dirty_list[0].mid; i++) {
2098 dp = txn->mt_u.dirty_list[i].mptr;
2099 if (dp->mp_pgno != next) {
2101 rc = writev(env->me_fd, iov, n);
2105 DPUTS("short write, filesystem full?");
2107 DPRINTF("writev: %s", strerror(n));
2114 lseek(env->me_fd, dp->mp_pgno * env->me_psize, SEEK_SET);
2117 DPRINTF("committing page %zu", dp->mp_pgno);
2118 iov[n].iov_len = env->me_psize;
2119 if (IS_OVERFLOW(dp)) iov[n].iov_len *= dp->mp_pages;
2120 iov[n].iov_base = (char *)dp;
2121 size += iov[n].iov_len;
2122 next = dp->mp_pgno + (IS_OVERFLOW(dp) ? dp->mp_pages : 1);
2123 /* clear dirty flag */
2124 dp->mp_flags &= ~P_DIRTY;
2125 if (++n >= MDB_COMMIT_PAGES) {
2135 rc = writev(env->me_fd, iov, n);
2139 DPUTS("short write, filesystem full?");
2141 DPRINTF("writev: %s", strerror(n));
2148 /* Drop the dirty pages.
2150 for (i=1; i<=txn->mt_u.dirty_list[0].mid; i++) {
2151 dp = txn->mt_u.dirty_list[i].mptr;
2152 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
2153 dp->mp_next = txn->mt_env->me_dpages;
2154 VGMEMP_FREE(txn->mt_env, dp);
2155 txn->mt_env->me_dpages = dp;
2157 VGMEMP_FREE(txn->mt_env, dp);
2160 txn->mt_u.dirty_list[i].mid = 0;
2162 txn->mt_u.dirty_list[0].mid = 0;
2164 if ((n = mdb_env_sync(env, 0)) != 0 ||
2165 (n = mdb_env_write_meta(txn)) != MDB_SUCCESS) {
2169 env->me_wtxnid = txn->mt_txnid;
2173 /* update the DB tables */
2175 int toggle = !env->me_db_toggle;
2179 ip = &env->me_dbs[toggle][2];
2180 jp = &txn->mt_dbs[2];
2181 LAZY_RWLOCK_WRLOCK(&env->me_dblock);
2182 for (i = 2; i < txn->mt_numdbs; i++) {
2183 if (ip->md_root != jp->md_root)
2188 env->me_db_toggle = toggle;
2189 env->me_numdbs = txn->mt_numdbs;
2190 LAZY_RWLOCK_UNLOCK(&env->me_dblock);
2193 UNLOCK_MUTEX_W(env);
2199 /** Read the environment parameters of a DB environment before
2200 * mapping it into memory.
2201 * @param[in] env the environment handle
2202 * @param[out] meta address of where to store the meta information
2203 * @return 0 on success, non-zero on failure.
2206 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
2213 /* We don't know the page size yet, so use a minimum value.
2217 if (!ReadFile(env->me_fd, &pbuf, MDB_PAGESIZE, (DWORD *)&rc, NULL) || rc == 0)
2219 if ((rc = read(env->me_fd, &pbuf, MDB_PAGESIZE)) == 0)
2224 else if (rc != MDB_PAGESIZE) {
2228 DPRINTF("read: %s", strerror(err));
2232 p = (MDB_page *)&pbuf;
2234 if (!F_ISSET(p->mp_flags, P_META)) {
2235 DPRINTF("page %zu not a meta page", p->mp_pgno);
2240 if (m->mm_magic != MDB_MAGIC) {
2241 DPUTS("meta has invalid magic");
2245 if (m->mm_version != MDB_VERSION) {
2246 DPRINTF("database is version %u, expected version %u",
2247 m->mm_version, MDB_VERSION);
2248 return MDB_VERSION_MISMATCH;
2251 memcpy(meta, m, sizeof(*m));
2255 /** Write the environment parameters of a freshly created DB environment.
2256 * @param[in] env the environment handle
2257 * @param[out] meta address of where to store the meta information
2258 * @return 0 on success, non-zero on failure.
2261 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
2268 DPUTS("writing new meta page");
2270 GET_PAGESIZE(psize);
2272 meta->mm_magic = MDB_MAGIC;
2273 meta->mm_version = MDB_VERSION;
2274 meta->mm_psize = psize;
2275 meta->mm_last_pg = 1;
2276 meta->mm_flags = env->me_flags & 0xffff;
2277 meta->mm_flags |= MDB_INTEGERKEY;
2278 meta->mm_dbs[0].md_root = P_INVALID;
2279 meta->mm_dbs[1].md_root = P_INVALID;
2281 p = calloc(2, psize);
2283 p->mp_flags = P_META;
2286 memcpy(m, meta, sizeof(*meta));
2288 q = (MDB_page *)((char *)p + psize);
2291 q->mp_flags = P_META;
2294 memcpy(m, meta, sizeof(*meta));
2299 rc = WriteFile(env->me_fd, p, psize * 2, &len, NULL);
2300 rc = (len == psize * 2) ? MDB_SUCCESS : ErrCode();
2303 rc = write(env->me_fd, p, psize * 2);
2304 rc = (rc == (int)psize * 2) ? MDB_SUCCESS : ErrCode();
2310 /** Update the environment info to commit a transaction.
2311 * @param[in] txn the transaction that's being committed
2312 * @return 0 on success, non-zero on failure.
2315 mdb_env_write_meta(MDB_txn *txn)
2318 MDB_meta meta, metab;
2320 int rc, len, toggle;
2326 assert(txn != NULL);
2327 assert(txn->mt_env != NULL);
2329 toggle = !txn->mt_toggle;
2330 DPRINTF("writing meta page %d for root page %zu",
2331 toggle, txn->mt_dbs[MAIN_DBI].md_root);
2335 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
2336 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
2338 ptr = (char *)&meta;
2339 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
2340 len = sizeof(MDB_meta) - off;
2343 meta.mm_dbs[0] = txn->mt_dbs[0];
2344 meta.mm_dbs[1] = txn->mt_dbs[1];
2345 meta.mm_last_pg = txn->mt_next_pgno - 1;
2346 meta.mm_txnid = txn->mt_txnid;
2349 off += env->me_psize;
2352 /* Write to the SYNC fd */
2355 memset(&ov, 0, sizeof(ov));
2357 WriteFile(env->me_mfd, ptr, len, (DWORD *)&rc, &ov);
2360 rc = pwrite(env->me_mfd, ptr, len, off);
2365 DPUTS("write failed, disk error?");
2366 /* On a failure, the pagecache still contains the new data.
2367 * Write some old data back, to prevent it from being used.
2368 * Use the non-SYNC fd; we know it will fail anyway.
2370 meta.mm_last_pg = metab.mm_last_pg;
2371 meta.mm_txnid = metab.mm_txnid;
2373 WriteFile(env->me_fd, ptr, len, NULL, &ov);
2375 r2 = pwrite(env->me_fd, ptr, len, off);
2377 env->me_flags |= MDB_FATAL_ERROR;
2380 /* Memory ordering issues are irrelevant; since the entire writer
2381 * is wrapped by wmutex, all of these changes will become visible
2382 * after the wmutex is unlocked. Since the DB is multi-version,
2383 * readers will get consistent data regardless of how fresh or
2384 * how stale their view of these values is.
2386 LAZY_MUTEX_LOCK(&env->me_txns->mti_mutex);
2387 txn->mt_env->me_txns->mti_me_toggle = toggle;
2388 txn->mt_env->me_txns->mti_txnid = txn->mt_txnid;
2389 LAZY_MUTEX_UNLOCK(&env->me_txns->mti_mutex);
2394 /** Check both meta pages to see which one is newer.
2395 * @param[in] env the environment handle
2396 * @param[out] which address of where to store the meta toggle ID
2397 * @return 0 on success, non-zero on failure.
2400 mdb_env_read_meta(MDB_env *env, int *which)
2404 assert(env != NULL);
2406 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2409 DPRINTF("Using meta page %d", toggle);
2416 mdb_env_create(MDB_env **env)
2420 e = calloc(1, sizeof(MDB_env));
2424 e->me_free_pgs = mdb_midl_alloc();
2425 if (!e->me_free_pgs) {
2429 e->me_maxreaders = DEFAULT_READERS;
2431 e->me_fd = INVALID_HANDLE_VALUE;
2432 e->me_lfd = INVALID_HANDLE_VALUE;
2433 e->me_mfd = INVALID_HANDLE_VALUE;
2434 VGMEMP_CREATE(e,0,0);
2440 mdb_env_set_mapsize(MDB_env *env, size_t size)
2444 env->me_mapsize = size;
2446 env->me_maxpg = env->me_mapsize / env->me_psize;
2451 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
2455 env->me_maxdbs = dbs;
2460 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
2462 if (env->me_map || readers < 1)
2464 env->me_maxreaders = readers;
2469 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
2471 if (!env || !readers)
2473 *readers = env->me_maxreaders;
2477 /** Further setup required for opening an MDB environment
2480 mdb_env_open2(MDB_env *env, unsigned int flags)
2482 int i, newenv = 0, toggle;
2486 env->me_flags = flags;
2488 memset(&meta, 0, sizeof(meta));
2490 if ((i = mdb_env_read_header(env, &meta)) != 0) {
2493 DPUTS("new mdbenv");
2497 if (!env->me_mapsize) {
2498 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
2504 LONG sizelo, sizehi;
2505 sizelo = env->me_mapsize & 0xffffffff;
2506 sizehi = env->me_mapsize >> 16; /* pointless on WIN32, only needed on W64 */
2508 /* Windows won't create mappings for zero length files.
2509 * Just allocate the maxsize right now.
2512 SetFilePointer(env->me_fd, sizelo, sizehi ? &sizehi : NULL, 0);
2513 if (!SetEndOfFile(env->me_fd))
2515 SetFilePointer(env->me_fd, 0, NULL, 0);
2517 mh = CreateFileMapping(env->me_fd, NULL, PAGE_READONLY,
2518 sizehi, sizelo, NULL);
2521 env->me_map = MapViewOfFileEx(mh, FILE_MAP_READ, 0, 0, env->me_mapsize,
2529 if (meta.mm_address && (flags & MDB_FIXEDMAP))
2531 env->me_map = mmap(meta.mm_address, env->me_mapsize, PROT_READ, i,
2533 if (env->me_map == MAP_FAILED) {
2540 meta.mm_mapsize = env->me_mapsize;
2541 if (flags & MDB_FIXEDMAP)
2542 meta.mm_address = env->me_map;
2543 i = mdb_env_init_meta(env, &meta);
2544 if (i != MDB_SUCCESS) {
2545 munmap(env->me_map, env->me_mapsize);
2549 env->me_psize = meta.mm_psize;
2551 env->me_maxpg = env->me_mapsize / env->me_psize;
2553 p = (MDB_page *)env->me_map;
2554 env->me_metas[0] = METADATA(p);
2555 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + meta.mm_psize);
2557 if ((i = mdb_env_read_meta(env, &toggle)) != 0)
2560 DPRINTF("opened database version %u, pagesize %u",
2561 env->me_metas[toggle]->mm_version, env->me_psize);
2562 DPRINTF("depth: %u", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_depth);
2563 DPRINTF("entries: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_entries);
2564 DPRINTF("branch pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_branch_pages);
2565 DPRINTF("leaf pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_leaf_pages);
2566 DPRINTF("overflow pages: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_overflow_pages);
2567 DPRINTF("root: %zu", env->me_metas[toggle]->mm_dbs[MAIN_DBI].md_root);
2573 /** Release a reader thread's slot in the reader lock table.
2574 * This function is called automatically when a thread exits.
2575 * Windows doesn't support destructor callbacks for thread-specific storage,
2576 * so this function is not compiled there.
2577 * @param[in] ptr This points to the slot in the reader lock table.
2580 mdb_env_reader_dest(void *ptr)
2582 MDB_reader *reader = ptr;
2584 reader->mr_txnid = 0;
2590 /** Downgrade the exclusive lock on the region back to shared */
2592 mdb_env_share_locks(MDB_env *env)
2596 if (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid)
2598 env->me_txns->mti_me_toggle = toggle;
2599 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
2604 /* First acquire a shared lock. The Unlock will
2605 * then release the existing exclusive lock.
2607 memset(&ov, 0, sizeof(ov));
2608 LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov);
2609 UnlockFile(env->me_lfd, 0, 0, 1, 0);
2613 struct flock lock_info;
2614 /* The shared lock replaces the existing lock */
2615 memset((void *)&lock_info, 0, sizeof(lock_info));
2616 lock_info.l_type = F_RDLCK;
2617 lock_info.l_whence = SEEK_SET;
2618 lock_info.l_start = 0;
2619 lock_info.l_len = 1;
2620 fcntl(env->me_lfd, F_SETLK, &lock_info);
2624 #if defined(_WIN32) || defined(__APPLE__)
2626 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
2628 * @(#) $Revision: 5.1 $
2629 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
2630 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
2632 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
2636 * Please do not copyright this code. This code is in the public domain.
2638 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
2639 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
2640 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
2641 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
2642 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
2643 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
2644 * PERFORMANCE OF THIS SOFTWARE.
2647 * chongo <Landon Curt Noll> /\oo/\
2648 * http://www.isthe.com/chongo/
2650 * Share and Enjoy! :-)
2653 typedef unsigned long long mdb_hash_t;
2654 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
2656 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
2657 * @param[in] str string to hash
2658 * @param[in] hval initial value for hash
2659 * @return 64 bit hash
2661 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
2662 * hval arg on the first call.
2665 mdb_hash_str(char *str, mdb_hash_t hval)
2667 unsigned char *s = (unsigned char *)str; /* unsigned string */
2669 * FNV-1a hash each octet of the string
2672 /* xor the bottom with the current octet */
2673 hval ^= (mdb_hash_t)*s++;
2675 /* multiply by the 64 bit FNV magic prime mod 2^64 */
2676 hval += (hval << 1) + (hval << 4) + (hval << 5) +
2677 (hval << 7) + (hval << 8) + (hval << 40);
2679 /* return our new hash value */
2683 /** Hash the string and output the hash in hex.
2684 * @param[in] str string to hash
2685 * @param[out] hexbuf an array of 17 chars to hold the hash
2688 mdb_hash_hex(char *str, char *hexbuf)
2691 mdb_hash_t h = mdb_hash_str(str, MDB_HASH_INIT);
2692 for (i=0; i<8; i++) {
2693 hexbuf += sprintf(hexbuf, "%02x", (unsigned int)h & 0xff);
2699 /** Open and/or initialize the lock region for the environment.
2700 * @param[in] env The MDB environment.
2701 * @param[in] lpath The pathname of the file used for the lock region.
2702 * @param[in] mode The Unix permissions for the file, if we create it.
2703 * @param[out] excl Set to true if we got an exclusive lock on the region.
2704 * @return 0 on success, non-zero on failure.
2707 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
2715 if ((env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
2716 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
2717 FILE_ATTRIBUTE_NORMAL, NULL)) == INVALID_HANDLE_VALUE) {
2721 /* Try to get exclusive lock. If we succeed, then
2722 * nobody is using the lock region and we should initialize it.
2725 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
2729 memset(&ov, 0, sizeof(ov));
2730 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
2736 size = GetFileSize(env->me_lfd, NULL);
2738 if ((env->me_lfd = open(lpath, O_RDWR|O_CREAT, mode)) == -1) {
2742 /* Try to get exclusive lock. If we succeed, then
2743 * nobody is using the lock region and we should initialize it.
2746 struct flock lock_info;
2747 memset((void *)&lock_info, 0, sizeof(lock_info));
2748 lock_info.l_type = F_WRLCK;
2749 lock_info.l_whence = SEEK_SET;
2750 lock_info.l_start = 0;
2751 lock_info.l_len = 1;
2752 rc = fcntl(env->me_lfd, F_SETLK, &lock_info);
2756 lock_info.l_type = F_RDLCK;
2757 rc = fcntl(env->me_lfd, F_SETLKW, &lock_info);
2764 size = lseek(env->me_lfd, 0, SEEK_END);
2766 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
2767 if (size < rsize && *excl) {
2769 SetFilePointer(env->me_lfd, rsize, NULL, 0);
2770 if (!SetEndOfFile(env->me_lfd)) {
2775 if (ftruncate(env->me_lfd, rsize) != 0) {
2782 size = rsize - sizeof(MDB_txninfo);
2783 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
2788 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
2794 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
2796 if (!env->me_txns) {
2801 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
2803 if (m == MAP_FAILED) {
2804 env->me_txns = NULL;
2814 if (!mdb_sec_inited) {
2815 InitializeSecurityDescriptor(&mdb_null_sd,
2816 SECURITY_DESCRIPTOR_REVISION);
2817 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
2818 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
2819 mdb_all_sa.bInheritHandle = FALSE;
2820 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
2823 mdb_hash_hex(lpath, hexbuf);
2824 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", hexbuf);
2825 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
2826 if (!env->me_rmutex) {
2830 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", hexbuf);
2831 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
2832 if (!env->me_wmutex) {
2839 mdb_hash_hex(lpath, hexbuf);
2840 sprintf(env->me_txns->mti_rmname, "MDBr%s", hexbuf);
2841 if (sem_unlink(env->me_txns->mti_rmname)) {
2843 if (rc != ENOENT && rc != EINVAL)
2846 env->me_rmutex = sem_open(env->me_txns->mti_rmname, O_CREAT, mode, 1);
2847 if (!env->me_rmutex) {
2851 sprintf(env->me_txns->mti_wmname, "MDBw%s", hexbuf);
2852 if (sem_unlink(env->me_txns->mti_wmname)) {
2854 if (rc != ENOENT && rc != EINVAL)
2857 env->me_wmutex = sem_open(env->me_txns->mti_wmname, O_CREAT, mode, 1);
2858 if (!env->me_wmutex) {
2862 #else /* __APPLE__ */
2863 pthread_mutexattr_t mattr;
2865 pthread_mutexattr_init(&mattr);
2866 rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED);
2870 pthread_mutex_init(&env->me_txns->mti_mutex, &mattr);
2871 pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr);
2872 #endif /* __APPLE__ */
2874 env->me_txns->mti_version = MDB_VERSION;
2875 env->me_txns->mti_magic = MDB_MAGIC;
2876 env->me_txns->mti_txnid = 0;
2877 env->me_txns->mti_numreaders = 0;
2878 env->me_txns->mti_me_toggle = 0;
2881 if (env->me_txns->mti_magic != MDB_MAGIC) {
2882 DPUTS("lock region has invalid magic");
2886 if (env->me_txns->mti_version != MDB_VERSION) {
2887 DPRINTF("lock region is version %u, expected version %u",
2888 env->me_txns->mti_version, MDB_VERSION);
2889 rc = MDB_VERSION_MISMATCH;
2893 if (rc != EACCES && rc != EAGAIN) {
2897 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
2898 if (!env->me_rmutex) {
2902 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
2903 if (!env->me_wmutex) {
2909 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
2910 if (!env->me_rmutex) {
2914 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
2915 if (!env->me_wmutex) {
2925 env->me_lfd = INVALID_HANDLE_VALUE;
2930 /** The name of the lock file in the DB environment */
2931 #define LOCKNAME "/lock.mdb"
2932 /** The name of the data file in the DB environment */
2933 #define DATANAME "/data.mdb"
2934 /** The suffix of the lock file when no subdir is used */
2935 #define LOCKSUFF "-lock"
2938 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mode_t mode)
2940 int oflags, rc, len, excl;
2941 char *lpath, *dpath;
2944 if (flags & MDB_NOSUBDIR) {
2945 rc = len + sizeof(LOCKSUFF) + len + 1;
2947 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
2952 if (flags & MDB_NOSUBDIR) {
2953 dpath = lpath + len + sizeof(LOCKSUFF);
2954 sprintf(lpath, "%s" LOCKSUFF, path);
2955 strcpy(dpath, path);
2957 dpath = lpath + len + sizeof(LOCKNAME);
2958 sprintf(lpath, "%s" LOCKNAME, path);
2959 sprintf(dpath, "%s" DATANAME, path);
2962 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
2967 if (F_ISSET(flags, MDB_RDONLY)) {
2968 oflags = GENERIC_READ;
2969 len = OPEN_EXISTING;
2971 oflags = GENERIC_READ|GENERIC_WRITE;
2974 mode = FILE_ATTRIBUTE_NORMAL;
2975 if ((env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2976 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
2981 if (F_ISSET(flags, MDB_RDONLY))
2984 oflags = O_RDWR | O_CREAT;
2986 if ((env->me_fd = open(dpath, oflags, mode)) == -1) {
2992 if ((rc = mdb_env_open2(env, flags)) == MDB_SUCCESS) {
2993 /* synchronous fd for meta writes */
2995 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
2996 mode |= FILE_FLAG_WRITE_THROUGH;
2997 if ((env->me_mfd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
2998 NULL, len, mode, NULL)) == INVALID_HANDLE_VALUE) {
3003 if (!(flags & (MDB_RDONLY|MDB_NOSYNC)))
3004 oflags |= MDB_DSYNC;
3005 if ((env->me_mfd = open(dpath, oflags, mode)) == -1) {
3010 env->me_path = strdup(path);
3011 DPRINTF("opened dbenv %p", (void *) env);
3012 pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
3013 LAZY_RWLOCK_INIT(&env->me_dblock, NULL);
3015 mdb_env_share_locks(env);
3016 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
3017 env->me_dbs[0] = calloc(env->me_maxdbs, sizeof(MDB_db));
3018 env->me_dbs[1] = calloc(env->me_maxdbs, sizeof(MDB_db));
3024 if (env->me_fd != INVALID_HANDLE_VALUE) {
3026 env->me_fd = INVALID_HANDLE_VALUE;
3028 if (env->me_lfd != INVALID_HANDLE_VALUE) {
3030 env->me_lfd = INVALID_HANDLE_VALUE;
3038 mdb_env_close(MDB_env *env)
3045 VGMEMP_DESTROY(env);
3046 while (env->me_dpages) {
3047 dp = env->me_dpages;
3048 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
3049 env->me_dpages = dp->mp_next;
3053 free(env->me_dbs[1]);
3054 free(env->me_dbs[0]);
3058 LAZY_RWLOCK_DESTROY(&env->me_dblock);
3059 pthread_key_delete(env->me_txkey);
3062 munmap(env->me_map, env->me_mapsize);
3067 pid_t pid = getpid();
3069 for (i=0; i<env->me_txns->mti_numreaders; i++)
3070 if (env->me_txns->mti_readers[i].mr_pid == pid)
3071 env->me_txns->mti_readers[i].mr_pid = 0;
3072 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
3075 mdb_midl_free(env->me_free_pgs);
3079 /** Compare two items pointing at aligned size_t's */
3081 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
3083 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
3084 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
3087 /** Compare two items pointing at aligned int's */
3089 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
3091 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
3092 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
3095 /** Compare two items pointing at ints of unknown alignment.
3096 * Nodes and keys are guaranteed to be 2-byte aligned.
3099 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
3101 #if BYTE_ORDER == LITTLE_ENDIAN
3102 unsigned short *u, *c;
3105 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
3106 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
3109 } while(!x && u > (unsigned short *)a->mv_data);
3112 return memcmp(a->mv_data, b->mv_data, a->mv_size);
3116 /** Compare two items lexically */
3118 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
3125 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3131 diff = memcmp(a->mv_data, b->mv_data, len);
3132 return diff ? diff : len_diff<0 ? -1 : len_diff;
3135 /** Compare two items in reverse byte order */
3137 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
3139 const unsigned char *p1, *p2, *p1_lim;
3143 p1_lim = (const unsigned char *)a->mv_data;
3144 p1 = (const unsigned char *)a->mv_data + a->mv_size;
3145 p2 = (const unsigned char *)b->mv_data + b->mv_size;
3147 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
3153 while (p1 > p1_lim) {
3154 diff = *--p1 - *--p2;
3158 return len_diff<0 ? -1 : len_diff;
3161 /** Search for key within a page, using binary search.
3162 * Returns the smallest entry larger or equal to the key.
3163 * If exactp is non-null, stores whether the found entry was an exact match
3164 * in *exactp (1 or 0).
3165 * Updates the cursor index with the index of the found entry.
3166 * If no entry larger or equal to the key is found, returns NULL.
3169 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
3171 unsigned int i = 0, nkeys;
3174 MDB_page *mp = mc->mc_pg[mc->mc_top];
3175 MDB_node *node = NULL;
3180 nkeys = NUMKEYS(mp);
3185 COPY_PGNO(pgno, mp->mp_pgno);
3186 DPRINTF("searching %u keys in %s %spage %zu",
3187 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
3194 low = IS_LEAF(mp) ? 0 : 1;
3196 cmp = mc->mc_dbx->md_cmp;
3198 /* Branch pages have no data, so if using integer keys,
3199 * alignment is guaranteed. Use faster mdb_cmp_int.
3201 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
3202 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
3209 nodekey.mv_size = mc->mc_db->md_pad;
3210 node = NODEPTR(mp, 0); /* fake */
3211 while (low <= high) {
3212 i = (low + high) >> 1;
3213 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
3214 rc = cmp(key, &nodekey);
3215 DPRINTF("found leaf index %u [%s], rc = %i",
3216 i, DKEY(&nodekey), rc);
3225 while (low <= high) {
3226 i = (low + high) >> 1;
3228 node = NODEPTR(mp, i);
3229 nodekey.mv_size = NODEKSZ(node);
3230 nodekey.mv_data = NODEKEY(node);
3232 rc = cmp(key, &nodekey);
3235 DPRINTF("found leaf index %u [%s], rc = %i",
3236 i, DKEY(&nodekey), rc);
3238 DPRINTF("found branch index %u [%s -> %zu], rc = %i",
3239 i, DKEY(&nodekey), NODEPGNO(node), rc);
3250 if (rc > 0) { /* Found entry is less than the key. */
3251 i++; /* Skip to get the smallest entry larger than key. */
3253 node = NODEPTR(mp, i);
3256 *exactp = (rc == 0);
3257 /* store the key index */
3258 mc->mc_ki[mc->mc_top] = i;
3260 /* There is no entry larger or equal to the key. */
3263 /* nodeptr is fake for LEAF2 */
3269 mdb_cursor_adjust(MDB_cursor *mc, func)
3273 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
3274 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
3281 /** Pop a page off the top of the cursor's stack. */
3283 mdb_cursor_pop(MDB_cursor *mc)
3288 top = mc->mc_pg[mc->mc_top];
3293 DPRINTF("popped page %zu off db %u cursor %p", top->mp_pgno,
3294 mc->mc_dbi, (void *) mc);
3298 /** Push a page onto the top of the cursor's stack. */
3300 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
3302 DPRINTF("pushing page %zu on db %u cursor %p", mp->mp_pgno,
3303 mc->mc_dbi, (void *) mc);
3305 if (mc->mc_snum >= CURSOR_STACK) {
3306 assert(mc->mc_snum < CURSOR_STACK);
3310 mc->mc_top = mc->mc_snum++;
3311 mc->mc_pg[mc->mc_top] = mp;
3312 mc->mc_ki[mc->mc_top] = 0;
3317 /** Find the address of the page corresponding to a given page number.
3318 * @param[in] txn the transaction for this access.
3319 * @param[in] pgno the page number for the page to retrieve.
3320 * @param[out] ret address of a pointer where the page's address will be stored.
3321 * @return 0 on success, non-zero on failure.
3324 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret)
3328 if (!F_ISSET(txn->mt_flags, MDB_TXN_RDONLY) && txn->mt_u.dirty_list[0].mid) {
3330 x = mdb_mid2l_search(txn->mt_u.dirty_list, pgno);
3331 if (x <= txn->mt_u.dirty_list[0].mid && txn->mt_u.dirty_list[x].mid == pgno) {
3332 p = txn->mt_u.dirty_list[x].mptr;
3336 if (pgno <= txn->mt_env->me_metas[txn->mt_toggle]->mm_last_pg)
3337 p = (MDB_page *)(txn->mt_env->me_map + txn->mt_env->me_psize * pgno);
3341 DPRINTF("page %zu not found", pgno);
3344 return (p != NULL) ? MDB_SUCCESS : MDB_PAGE_NOTFOUND;
3347 /** Search for the page a given key should be in.
3348 * Pushes parent pages on the cursor stack. This function continues a
3349 * search on a cursor that has already been initialized. (Usually by
3350 * #mdb_page_search() but also by #mdb_node_move().)
3351 * @param[in,out] mc the cursor for this operation.
3352 * @param[in] key the key to search for. If NULL, search for the lowest
3353 * page. (This is used by #mdb_cursor_first().)
3354 * @param[in] modify If true, visited pages are updated with new page numbers.
3355 * @return 0 on success, non-zero on failure.
3358 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int modify)
3360 MDB_page *mp = mc->mc_pg[mc->mc_top];
3365 while (IS_BRANCH(mp)) {
3369 DPRINTF("branch page %zu has %u keys", mp->mp_pgno, NUMKEYS(mp));
3370 assert(NUMKEYS(mp) > 1);
3371 DPRINTF("found index 0 to page %zu", NODEPGNO(NODEPTR(mp, 0)));
3373 if (key == NULL) /* Initialize cursor to first page. */
3375 else if (key->mv_size > MAXKEYSIZE && key->mv_data == NULL) {
3376 /* cursor to last page */
3380 node = mdb_node_search(mc, key, &exact);
3382 i = NUMKEYS(mp) - 1;
3384 i = mc->mc_ki[mc->mc_top];
3393 DPRINTF("following index %u for key [%s]",
3395 assert(i < NUMKEYS(mp));
3396 node = NODEPTR(mp, i);
3398 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp)))
3401 mc->mc_ki[mc->mc_top] = i;
3402 if ((rc = mdb_cursor_push(mc, mp)))
3406 if ((rc = mdb_page_touch(mc)) != 0)
3408 mp = mc->mc_pg[mc->mc_top];
3413 DPRINTF("internal error, index points to a %02X page!?",
3415 return MDB_CORRUPTED;
3418 DPRINTF("found leaf page %zu for key [%s]", mp->mp_pgno,
3419 key ? DKEY(key) : NULL);
3424 /** Search for the page a given key should be in.
3425 * Pushes parent pages on the cursor stack. This function just sets up
3426 * the search; it finds the root page for \b mc's database and sets this
3427 * as the root of the cursor's stack. Then #mdb_page_search_root() is
3428 * called to complete the search.
3429 * @param[in,out] mc the cursor for this operation.
3430 * @param[in] key the key to search for. If NULL, search for the lowest
3431 * page. (This is used by #mdb_cursor_first().)
3432 * @param[in] modify If true, visited pages are updated with new page numbers.
3433 * @return 0 on success, non-zero on failure.
3436 mdb_page_search(MDB_cursor *mc, MDB_val *key, int modify)
3441 /* Make sure the txn is still viable, then find the root from
3442 * the txn's db table.
3444 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
3445 DPUTS("transaction has failed, must abort");
3448 /* Make sure we're using an up-to-date root */
3449 if (mc->mc_dbi > MAIN_DBI) {
3450 if ((*mc->mc_dbflag & DB_STALE) ||
3451 (modify && !(*mc->mc_dbflag & DB_DIRTY))) {
3453 unsigned char dbflag = 0;
3454 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
3455 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, modify);
3458 if (*mc->mc_dbflag & DB_STALE) {
3461 MDB_node *leaf = mdb_node_search(&mc2,
3462 &mc->mc_dbx->md_name, &exact);
3464 return MDB_NOTFOUND;
3465 mdb_node_read(mc->mc_txn, leaf, &data);
3466 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
3470 *mc->mc_dbflag = dbflag;
3473 root = mc->mc_db->md_root;
3475 if (root == P_INVALID) { /* Tree is empty. */
3476 DPUTS("tree is empty");
3477 return MDB_NOTFOUND;
3482 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0])))
3488 DPRINTF("db %u root page %zu has flags 0x%X",
3489 mc->mc_dbi, root, mc->mc_pg[0]->mp_flags);
3492 if ((rc = mdb_page_touch(mc)))
3496 return mdb_page_search_root(mc, key, modify);
3499 /** Return the data associated with a given node.
3500 * @param[in] txn The transaction for this operation.
3501 * @param[in] leaf The node being read.
3502 * @param[out] data Updated to point to the node's data.
3503 * @return 0 on success, non-zero on failure.
3506 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
3508 MDB_page *omp; /* overflow page */
3512 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
3513 data->mv_size = NODEDSZ(leaf);
3514 data->mv_data = NODEDATA(leaf);
3518 /* Read overflow data.
3520 data->mv_size = NODEDSZ(leaf);
3521 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
3522 if ((rc = mdb_page_get(txn, pgno, &omp))) {
3523 DPRINTF("read overflow page %zu failed", pgno);
3526 data->mv_data = METADATA(omp);
3532 mdb_get(MDB_txn *txn, MDB_dbi dbi,
3533 MDB_val *key, MDB_val *data)
3542 DPRINTF("===> get db %u key [%s]", dbi, DKEY(key));
3544 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
3547 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
3551 mdb_cursor_init(&mc, txn, dbi, &mx);
3552 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
3555 /** Find a sibling for a page.
3556 * Replaces the page at the top of the cursor's stack with the
3557 * specified sibling, if one exists.
3558 * @param[in] mc The cursor for this operation.
3559 * @param[in] move_right Non-zero if the right sibling is requested,
3560 * otherwise the left sibling.
3561 * @return 0 on success, non-zero on failure.
3564 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
3570 if (mc->mc_snum < 2) {
3571 return MDB_NOTFOUND; /* root has no siblings */
3575 DPRINTF("parent page is page %zu, index %u",
3576 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]);
3578 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
3579 : (mc->mc_ki[mc->mc_top] == 0)) {
3580 DPRINTF("no more keys left, moving to %s sibling",
3581 move_right ? "right" : "left");
3582 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS)
3586 mc->mc_ki[mc->mc_top]++;
3588 mc->mc_ki[mc->mc_top]--;
3589 DPRINTF("just moving to %s index key %u",
3590 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]);
3592 assert(IS_BRANCH(mc->mc_pg[mc->mc_top]));
3594 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
3595 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp)))
3598 mdb_cursor_push(mc, mp);
3603 /** Move the cursor to the next data item. */
3605 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3611 if (mc->mc_flags & C_EOF) {
3612 return MDB_NOTFOUND;
3615 assert(mc->mc_flags & C_INITIALIZED);
3617 mp = mc->mc_pg[mc->mc_top];
3619 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3620 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3621 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3622 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
3623 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
3624 if (op != MDB_NEXT || rc == MDB_SUCCESS)
3628 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3629 if (op == MDB_NEXT_DUP)
3630 return MDB_NOTFOUND;
3634 DPRINTF("cursor_next: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3636 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
3637 DPUTS("=====> move to next sibling page");
3638 if (mdb_cursor_sibling(mc, 1) != MDB_SUCCESS) {
3639 mc->mc_flags |= C_EOF;
3640 mc->mc_flags &= ~C_INITIALIZED;
3641 return MDB_NOTFOUND;
3643 mp = mc->mc_pg[mc->mc_top];
3644 DPRINTF("next page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3646 mc->mc_ki[mc->mc_top]++;
3648 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3649 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3652 key->mv_size = mc->mc_db->md_pad;
3653 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3657 assert(IS_LEAF(mp));
3658 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3660 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3661 mdb_xcursor_init1(mc, leaf);
3664 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3667 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3668 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3669 if (rc != MDB_SUCCESS)
3674 MDB_SET_KEY(leaf, key);
3678 /** Move the cursor to the previous data item. */
3680 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
3686 assert(mc->mc_flags & C_INITIALIZED);
3688 mp = mc->mc_pg[mc->mc_top];
3690 if (mc->mc_db->md_flags & MDB_DUPSORT) {
3691 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3692 if (op == MDB_PREV || op == MDB_PREV_DUP) {
3693 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3694 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
3695 if (op != MDB_PREV || rc == MDB_SUCCESS)
3698 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3699 if (op == MDB_PREV_DUP)
3700 return MDB_NOTFOUND;
3705 DPRINTF("cursor_prev: top page is %zu in cursor %p", mp->mp_pgno, (void *) mc);
3707 if (mc->mc_ki[mc->mc_top] == 0) {
3708 DPUTS("=====> move to prev sibling page");
3709 if (mdb_cursor_sibling(mc, 0) != MDB_SUCCESS) {
3710 mc->mc_flags &= ~C_INITIALIZED;
3711 return MDB_NOTFOUND;
3713 mp = mc->mc_pg[mc->mc_top];
3714 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
3715 DPRINTF("prev page is %zu, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]);
3717 mc->mc_ki[mc->mc_top]--;
3719 mc->mc_flags &= ~C_EOF;
3721 DPRINTF("==> cursor points to page %zu with %u keys, key index %u",
3722 mp->mp_pgno, NUMKEYS(mp), mc->mc_ki[mc->mc_top]);
3725 key->mv_size = mc->mc_db->md_pad;
3726 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3730 assert(IS_LEAF(mp));
3731 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
3733 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3734 mdb_xcursor_init1(mc, leaf);
3737 if ((rc = mdb_node_read(mc->mc_txn, leaf, data) != MDB_SUCCESS))
3740 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3741 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3742 if (rc != MDB_SUCCESS)
3747 MDB_SET_KEY(leaf, key);
3751 /** Set the cursor on a specific data item. */
3753 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
3754 MDB_cursor_op op, int *exactp)
3763 assert(key->mv_size > 0);
3765 /* See if we're already on the right page */
3766 if (mc->mc_flags & C_INITIALIZED) {
3769 mp = mc->mc_pg[mc->mc_top];
3771 mc->mc_ki[mc->mc_top] = 0;
3772 return MDB_NOTFOUND;
3774 if (mp->mp_flags & P_LEAF2) {
3775 nodekey.mv_size = mc->mc_db->md_pad;
3776 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
3778 leaf = NODEPTR(mp, 0);
3779 MDB_SET_KEY(leaf, &nodekey);
3781 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3783 /* Probably happens rarely, but first node on the page
3784 * was the one we wanted.
3786 mc->mc_ki[mc->mc_top] = 0;
3787 leaf = NODEPTR(mp, 0);
3794 unsigned int nkeys = NUMKEYS(mp);
3796 if (mp->mp_flags & P_LEAF2) {
3797 nodekey.mv_data = LEAF2KEY(mp,
3798 nkeys-1, nodekey.mv_size);
3800 leaf = NODEPTR(mp, nkeys-1);
3801 MDB_SET_KEY(leaf, &nodekey);
3803 rc = mc->mc_dbx->md_cmp(key, &nodekey);
3805 /* last node was the one we wanted */
3806 mc->mc_ki[mc->mc_top] = nkeys-1;
3807 leaf = NODEPTR(mp, nkeys-1);
3813 /* This is definitely the right page, skip search_page */
3818 /* If any parents have right-sibs, search.
3819 * Otherwise, there's nothing further.
3821 for (i=0; i<mc->mc_top; i++)
3823 NUMKEYS(mc->mc_pg[i])-1)
3825 if (i == mc->mc_top) {
3826 /* There are no other pages */
3827 mc->mc_ki[mc->mc_top] = nkeys;
3828 return MDB_NOTFOUND;
3832 /* There are no other pages */
3833 mc->mc_ki[mc->mc_top] = 0;
3834 return MDB_NOTFOUND;
3838 rc = mdb_page_search(mc, key, 0);
3839 if (rc != MDB_SUCCESS)
3842 mp = mc->mc_pg[mc->mc_top];
3843 assert(IS_LEAF(mp));
3846 leaf = mdb_node_search(mc, key, exactp);
3847 if (exactp != NULL && !*exactp) {
3848 /* MDB_SET specified and not an exact match. */
3849 return MDB_NOTFOUND;
3853 DPUTS("===> inexact leaf not found, goto sibling");
3854 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
3855 return rc; /* no entries matched */
3856 mp = mc->mc_pg[mc->mc_top];
3857 assert(IS_LEAF(mp));
3858 leaf = NODEPTR(mp, 0);
3862 mc->mc_flags |= C_INITIALIZED;
3863 mc->mc_flags &= ~C_EOF;
3866 key->mv_size = mc->mc_db->md_pad;
3867 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
3871 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3872 mdb_xcursor_init1(mc, leaf);
3875 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3876 if (op == MDB_SET || op == MDB_SET_RANGE) {
3877 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3880 if (op == MDB_GET_BOTH) {
3886 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
3887 if (rc != MDB_SUCCESS)
3890 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
3892 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
3894 rc = mc->mc_dbx->md_dcmp(data, &d2);
3896 if (op == MDB_GET_BOTH || rc > 0)
3897 return MDB_NOTFOUND;
3902 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3903 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3908 /* The key already matches in all other cases */
3909 if (op == MDB_SET_RANGE)
3910 MDB_SET_KEY(leaf, key);
3911 DPRINTF("==> cursor placed on key [%s]", DKEY(key));
3916 /** Move the cursor to the first item in the database. */
3918 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3923 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3924 rc = mdb_page_search(mc, NULL, 0);
3925 if (rc != MDB_SUCCESS)
3928 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3930 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
3931 mc->mc_flags |= C_INITIALIZED;
3932 mc->mc_flags &= ~C_EOF;
3934 mc->mc_ki[mc->mc_top] = 0;
3936 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3937 key->mv_size = mc->mc_db->md_pad;
3938 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
3943 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3944 mdb_xcursor_init1(mc, leaf);
3945 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
3950 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3951 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
3955 MDB_SET_KEY(leaf, key);
3959 /** Move the cursor to the last item in the database. */
3961 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
3967 lkey.mv_size = MAXKEYSIZE+1;
3968 lkey.mv_data = NULL;
3970 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
3971 rc = mdb_page_search(mc, &lkey, 0);
3972 if (rc != MDB_SUCCESS)
3975 assert(IS_LEAF(mc->mc_pg[mc->mc_top]));
3977 leaf = NODEPTR(mc->mc_pg[mc->mc_top], NUMKEYS(mc->mc_pg[mc->mc_top])-1);
3978 mc->mc_flags |= C_INITIALIZED;
3979 mc->mc_flags &= ~C_EOF;
3981 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
3983 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
3984 key->mv_size = mc->mc_db->md_pad;
3985 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
3990 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
3991 mdb_xcursor_init1(mc, leaf);
3992 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
3997 mc->mc_xcursor->mx_cursor.mc_flags &= ~C_INITIALIZED;
3998 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
4003 MDB_SET_KEY(leaf, key);
4008 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4018 case MDB_GET_BOTH_RANGE:
4019 if (data == NULL || mc->mc_xcursor == NULL) {
4026 if (key == NULL || key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
4028 } else if (op == MDB_SET_RANGE)
4029 rc = mdb_cursor_set(mc, key, data, op, NULL);
4031 rc = mdb_cursor_set(mc, key, data, op, &exact);
4033 case MDB_GET_MULTIPLE:
4035 !(mc->mc_db->md_flags & MDB_DUPFIXED) ||
4036 !(mc->mc_flags & C_INITIALIZED)) {
4041 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
4042 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
4045 case MDB_NEXT_MULTIPLE:
4047 !(mc->mc_db->md_flags & MDB_DUPFIXED)) {
4051 if (!(mc->mc_flags & C_INITIALIZED))
4052 rc = mdb_cursor_first(mc, key, data);
4054 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
4055 if (rc == MDB_SUCCESS) {
4056 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
4059 mx = &mc->mc_xcursor->mx_cursor;
4060 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
4062 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
4063 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
4071 case MDB_NEXT_NODUP:
4072 if (!(mc->mc_flags & C_INITIALIZED))
4073 rc = mdb_cursor_first(mc, key, data);
4075 rc = mdb_cursor_next(mc, key, data, op);
4079 case MDB_PREV_NODUP:
4080 if (!(mc->mc_flags & C_INITIALIZED) || (mc->mc_flags & C_EOF))
4081 rc = mdb_cursor_last(mc, key, data);
4083 rc = mdb_cursor_prev(mc, key, data, op);
4086 rc = mdb_cursor_first(mc, key, data);
4090 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4091 !(mc->mc_flags & C_INITIALIZED) ||
4092 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4096 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
4099 rc = mdb_cursor_last(mc, key, data);
4103 !(mc->mc_db->md_flags & MDB_DUPSORT) ||
4104 !(mc->mc_flags & C_INITIALIZED) ||
4105 !(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
4109 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
4112 DPRINTF("unhandled/unimplemented cursor operation %u", op);
4120 /** Touch all the pages in the cursor stack.
4121 * Makes sure all the pages are writable, before attempting a write operation.
4122 * @param[in] mc The cursor to operate on.
4125 mdb_cursor_touch(MDB_cursor *mc)
4129 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
4131 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4132 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 1);
4135 *mc->mc_dbflag = DB_DIRTY;
4137 for (mc->mc_top = 0; mc->mc_top < mc->mc_snum; mc->mc_top++) {
4138 rc = mdb_page_touch(mc);
4142 mc->mc_top = mc->mc_snum-1;
4147 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
4150 MDB_node *leaf = NULL;
4151 MDB_val xdata, *rdata, dkey;
4155 unsigned int mcount = 0;
4159 char dbuf[MAXKEYSIZE+1];
4160 unsigned int nflags;
4163 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4166 DPRINTF("==> put db %u key [%s], size %zu, data size %zu",
4167 mc->mc_dbi, DKEY(key), key ? key->mv_size:0, data->mv_size);
4171 if (flags == MDB_CURRENT) {
4172 if (!(mc->mc_flags & C_INITIALIZED))
4175 } else if (mc->mc_db->md_root == P_INVALID) {
4177 /* new database, write a root leaf page */
4178 DPUTS("allocating new root leaf page");
4179 if ((np = mdb_page_new(mc, P_LEAF, 1)) == NULL) {
4183 mdb_cursor_push(mc, np);
4184 mc->mc_db->md_root = np->mp_pgno;
4185 mc->mc_db->md_depth++;
4186 *mc->mc_dbflag = DB_DIRTY;
4187 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
4189 np->mp_flags |= P_LEAF2;
4190 mc->mc_flags |= C_INITIALIZED;
4196 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
4197 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
4198 DPRINTF("duplicate key [%s]", DKEY(key));
4200 return MDB_KEYEXIST;
4202 if (rc && rc != MDB_NOTFOUND)
4206 /* Cursor is positioned, now make sure all pages are writable */
4207 rc2 = mdb_cursor_touch(mc);
4212 /* The key already exists */
4213 if (rc == MDB_SUCCESS) {
4214 /* there's only a key anyway, so this is a no-op */
4215 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
4216 unsigned int ksize = mc->mc_db->md_pad;
4217 if (key->mv_size != ksize)
4219 if (flags == MDB_CURRENT) {
4220 char *ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
4221 memcpy(ptr, key->mv_data, ksize);
4226 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4229 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
4230 /* Was a single item before, must convert now */
4232 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4233 /* Just overwrite the current item */
4234 if (flags == MDB_CURRENT)
4237 dkey.mv_size = NODEDSZ(leaf);
4238 dkey.mv_data = NODEDATA(leaf);
4239 #if UINT_MAX < SIZE_MAX
4240 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && dkey.mv_size == sizeof(size_t))
4241 #ifdef MISALIGNED_OK
4242 mc->mc_dbx->md_dcmp = mdb_cmp_long;
4244 mc->mc_dbx->md_dcmp = mdb_cmp_cint;
4247 /* if data matches, ignore it */
4248 if (!mc->mc_dbx->md_dcmp(data, &dkey))
4249 return (flags == MDB_NODUPDATA) ? MDB_KEYEXIST : MDB_SUCCESS;
4251 /* create a fake page for the dup items */
4252 memcpy(dbuf, dkey.mv_data, dkey.mv_size);
4253 dkey.mv_data = dbuf;
4254 fp = (MDB_page *)&pbuf;
4255 fp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4256 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
4257 fp->mp_lower = PAGEHDRSZ;
4258 fp->mp_upper = PAGEHDRSZ + dkey.mv_size + data->mv_size;
4259 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4260 fp->mp_flags |= P_LEAF2;
4261 fp->mp_pad = data->mv_size;
4263 fp->mp_upper += 2 * sizeof(indx_t) + 2 * NODESIZE +
4264 (dkey.mv_size & 1) + (data->mv_size & 1);
4266 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4269 xdata.mv_size = fp->mp_upper;
4274 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4275 /* See if we need to convert from fake page to subDB */
4277 unsigned int offset;
4280 fp = NODEDATA(leaf);
4281 if (flags == MDB_CURRENT) {
4282 fp->mp_flags |= P_DIRTY;
4283 COPY_PGNO(fp->mp_pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
4284 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
4288 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4289 offset = fp->mp_pad;
4291 offset = NODESIZE + sizeof(indx_t) + data->mv_size;
4293 offset += offset & 1;
4294 if (NODESIZE + sizeof(indx_t) + NODEKSZ(leaf) + NODEDSZ(leaf) +
4295 offset >= (mc->mc_txn->mt_env->me_psize - PAGEHDRSZ) /
4297 /* yes, convert it */
4299 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4300 dummy.md_pad = fp->mp_pad;
4301 dummy.md_flags = MDB_DUPFIXED;
4302 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4303 dummy.md_flags |= MDB_INTEGERKEY;
4306 dummy.md_branch_pages = 0;
4307 dummy.md_leaf_pages = 1;
4308 dummy.md_overflow_pages = 0;
4309 dummy.md_entries = NUMKEYS(fp);
4311 xdata.mv_size = sizeof(MDB_db);
4312 xdata.mv_data = &dummy;
4313 mp = mdb_page_alloc(mc, 1);
4316 offset = mc->mc_txn->mt_env->me_psize - NODEDSZ(leaf);
4317 flags |= F_DUPDATA|F_SUBDATA;
4318 dummy.md_root = mp->mp_pgno;
4320 /* no, just grow it */
4322 xdata.mv_size = NODEDSZ(leaf) + offset;
4323 xdata.mv_data = &pbuf;
4324 mp = (MDB_page *)&pbuf;
4325 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
4328 mp->mp_flags = fp->mp_flags | P_DIRTY;
4329 mp->mp_pad = fp->mp_pad;
4330 mp->mp_lower = fp->mp_lower;
4331 mp->mp_upper = fp->mp_upper + offset;
4333 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
4335 nsize = NODEDSZ(leaf) - fp->mp_upper;
4336 memcpy((char *)mp + mp->mp_upper, (char *)fp + fp->mp_upper, nsize);
4337 for (i=0; i<NUMKEYS(fp); i++)
4338 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
4340 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4344 /* data is on sub-DB, just store it */
4345 flags |= F_DUPDATA|F_SUBDATA;
4349 /* overflow page overwrites need special handling */
4350 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
4353 int ovpages, dpages;
4355 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
4356 dpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4357 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
4358 mdb_page_get(mc->mc_txn, pg, &omp);
4359 /* Is the ov page writable and large enough? */
4360 if ((omp->mp_flags & P_DIRTY) && ovpages >= dpages) {
4361 /* yes, overwrite it. Note in this case we don't
4362 * bother to try shrinking the node if the new data
4363 * is smaller than the overflow threshold.
4365 if (F_ISSET(flags, MDB_RESERVE))
4366 data->mv_data = METADATA(omp);
4368 memcpy(METADATA(omp), data->mv_data, data->mv_size);
4371 /* no, free ovpages */
4373 mc->mc_db->md_overflow_pages -= ovpages;
4374 for (i=0; i<ovpages; i++) {
4375 DPRINTF("freed ov page %zu", pg);
4376 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
4380 } else if (NODEDSZ(leaf) == data->mv_size) {
4381 /* same size, just replace it. Note that we could
4382 * also reuse this node if the new data is smaller,
4383 * but instead we opt to shrink the node in that case.
4385 if (F_ISSET(flags, MDB_RESERVE))
4386 data->mv_data = NODEDATA(leaf);
4388 memcpy(NODEDATA(leaf), data->mv_data, data->mv_size);
4391 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], 0);
4392 mc->mc_db->md_entries--;
4394 DPRINTF("inserting key at index %i", mc->mc_ki[mc->mc_top]);
4400 nflags = flags & NODE_ADD_FLAGS;
4401 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(mc->mc_txn->mt_env, key, rdata);
4402 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
4403 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
4404 nflags &= ~MDB_APPEND;
4405 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
4407 /* There is room already in this leaf page. */
4408 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
4409 if (rc == 0 && !do_sub) {
4410 /* Adjust other cursors pointing to mp */
4411 MDB_cursor *m2, *m3;
4412 MDB_dbi dbi = mc->mc_dbi;
4413 unsigned i = mc->mc_top;
4414 MDB_page *mp = mc->mc_pg[i];
4416 if (mc->mc_flags & C_SUB)
4419 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
4420 if (mc->mc_flags & C_SUB)
4421 m3 = &m2->mc_xcursor->mx_cursor;
4424 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
4425 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
4432 if (rc != MDB_SUCCESS)
4433 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4435 /* Now store the actual data in the child DB. Note that we're
4436 * storing the user data in the keys field, so there are strict
4437 * size limits on dupdata. The actual data fields of the child
4438 * DB are all zero size.
4445 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4446 if (flags & MDB_CURRENT) {
4447 xflags = MDB_CURRENT;
4449 mdb_xcursor_init1(mc, leaf);
4450 xflags = (flags & MDB_NODUPDATA) ? MDB_NOOVERWRITE : 0;
4452 /* converted, write the original data first */
4454 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
4458 /* Adjust other cursors pointing to mp */
4460 unsigned i = mc->mc_top;
4461 MDB_page *mp = mc->mc_pg[i];
4463 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4464 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
4465 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
4466 mdb_xcursor_init1(m2, leaf);
4471 xflags |= (flags & MDB_APPEND);
4472 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
4473 if (flags & F_SUBDATA) {
4474 void *db = NODEDATA(leaf);
4475 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4478 /* sub-writes might have failed so check rc again.
4479 * Don't increment count if we just replaced an existing item.
4481 if (!rc && !(flags & MDB_CURRENT))
4482 mc->mc_db->md_entries++;
4483 if (flags & MDB_MULTIPLE) {
4485 if (mcount < data[1].mv_size) {
4486 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
4487 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4497 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
4502 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_RDONLY))
4505 if (!mc->mc_flags & C_INITIALIZED)
4508 rc = mdb_cursor_touch(mc);
4512 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4514 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_DUPDATA)) {
4515 if (flags != MDB_NODUPDATA) {
4516 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
4517 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
4519 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, 0);
4520 /* If sub-DB still has entries, we're done */
4521 if (mc->mc_xcursor->mx_db.md_entries) {
4522 if (leaf->mn_flags & F_SUBDATA) {
4523 /* update subDB info */
4524 void *db = NODEDATA(leaf);
4525 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
4527 /* shrink fake page */
4528 mdb_node_shrink(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
4530 mc->mc_db->md_entries--;
4533 /* otherwise fall thru and delete the sub-DB */
4536 if (leaf->mn_flags & F_SUBDATA) {
4537 /* add all the child DB's pages to the free list */
4538 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
4539 if (rc == MDB_SUCCESS) {
4540 mc->mc_db->md_entries -=
4541 mc->mc_xcursor->mx_db.md_entries;
4546 return mdb_cursor_del0(mc, leaf);
4549 /** Allocate and initialize new pages for a database.
4550 * @param[in] mc a cursor on the database being added to.
4551 * @param[in] flags flags defining what type of page is being allocated.
4552 * @param[in] num the number of pages to allocate. This is usually 1,
4553 * unless allocating overflow pages for a large record.
4554 * @return Address of a page, or NULL on failure.
4557 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num)
4561 if ((np = mdb_page_alloc(mc, num)) == NULL)
4563 DPRINTF("allocated new mpage %zu, page size %u",
4564 np->mp_pgno, mc->mc_txn->mt_env->me_psize);
4565 np->mp_flags = flags | P_DIRTY;
4566 np->mp_lower = PAGEHDRSZ;
4567 np->mp_upper = mc->mc_txn->mt_env->me_psize;
4570 mc->mc_db->md_branch_pages++;
4571 else if (IS_LEAF(np))
4572 mc->mc_db->md_leaf_pages++;
4573 else if (IS_OVERFLOW(np)) {
4574 mc->mc_db->md_overflow_pages += num;
4581 /** Calculate the size of a leaf node.
4582 * The size depends on the environment's page size; if a data item
4583 * is too large it will be put onto an overflow page and the node
4584 * size will only include the key and not the data. Sizes are always
4585 * rounded up to an even number of bytes, to guarantee 2-byte alignment
4586 * of the #MDB_node headers.
4587 * @param[in] env The environment handle.
4588 * @param[in] key The key for the node.
4589 * @param[in] data The data for the node.
4590 * @return The number of bytes needed to store the node.
4593 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
4597 sz = LEAFSIZE(key, data);
4598 if (sz >= env->me_psize / MDB_MINKEYS) {
4599 /* put on overflow page */
4600 sz -= data->mv_size - sizeof(pgno_t);
4604 return sz + sizeof(indx_t);
4607 /** Calculate the size of a branch node.
4608 * The size should depend on the environment's page size but since
4609 * we currently don't support spilling large keys onto overflow
4610 * pages, it's simply the size of the #MDB_node header plus the
4611 * size of the key. Sizes are always rounded up to an even number
4612 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
4613 * @param[in] env The environment handle.
4614 * @param[in] key The key for the node.
4615 * @return The number of bytes needed to store the node.
4618 mdb_branch_size(MDB_env *env, MDB_val *key)
4623 if (sz >= env->me_psize / MDB_MINKEYS) {
4624 /* put on overflow page */
4625 /* not implemented */
4626 /* sz -= key->size - sizeof(pgno_t); */
4629 return sz + sizeof(indx_t);
4632 /** Add a node to the page pointed to by the cursor.
4633 * @param[in] mc The cursor for this operation.
4634 * @param[in] indx The index on the page where the new node should be added.
4635 * @param[in] key The key for the new node.
4636 * @param[in] data The data for the new node, if any.
4637 * @param[in] pgno The page number, if adding a branch node.
4638 * @param[in] flags Flags for the node.
4639 * @return 0 on success, non-zero on failure. Possible errors are:
4641 * <li>ENOMEM - failed to allocate overflow pages for the node.
4642 * <li>ENOSPC - there is insufficient room in the page. This error
4643 * should never happen since all callers already calculate the
4644 * page's free space before calling this function.
4648 mdb_node_add(MDB_cursor *mc, indx_t indx,
4649 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
4652 size_t node_size = NODESIZE;
4655 MDB_page *mp = mc->mc_pg[mc->mc_top];
4656 MDB_page *ofp = NULL; /* overflow page */
4659 assert(mp->mp_upper >= mp->mp_lower);
4661 DPRINTF("add to %s %spage %zu index %i, data size %zu key size %zu [%s]",
4662 IS_LEAF(mp) ? "leaf" : "branch",
4663 IS_SUBP(mp) ? "sub-" : "",
4664 mp->mp_pgno, indx, data ? data->mv_size : 0,
4665 key ? key->mv_size : 0, key ? DKEY(key) : NULL);
4668 /* Move higher keys up one slot. */
4669 int ksize = mc->mc_db->md_pad, dif;
4670 char *ptr = LEAF2KEY(mp, indx, ksize);
4671 dif = NUMKEYS(mp) - indx;
4673 memmove(ptr+ksize, ptr, dif*ksize);
4674 /* insert new key */
4675 memcpy(ptr, key->mv_data, ksize);
4677 /* Just using these for counting */
4678 mp->mp_lower += sizeof(indx_t);
4679 mp->mp_upper -= ksize - sizeof(indx_t);
4684 node_size += key->mv_size;
4688 if (F_ISSET(flags, F_BIGDATA)) {
4689 /* Data already on overflow page. */
4690 node_size += sizeof(pgno_t);
4691 } else if (node_size + data->mv_size >= mc->mc_txn->mt_env->me_psize / MDB_MINKEYS) {
4692 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
4693 /* Put data on overflow page. */
4694 DPRINTF("data size is %zu, node would be %zu, put data on overflow page",
4695 data->mv_size, node_size+data->mv_size);
4696 node_size += sizeof(pgno_t);
4697 if ((ofp = mdb_page_new(mc, P_OVERFLOW, ovpages)) == NULL)
4699 DPRINTF("allocated overflow page %zu", ofp->mp_pgno);
4702 node_size += data->mv_size;
4705 node_size += node_size & 1;
4707 if (node_size + sizeof(indx_t) > SIZELEFT(mp)) {
4708 DPRINTF("not enough room in page %zu, got %u ptrs",
4709 mp->mp_pgno, NUMKEYS(mp));
4710 DPRINTF("upper - lower = %u - %u = %u", mp->mp_upper, mp->mp_lower,
4711 mp->mp_upper - mp->mp_lower);
4712 DPRINTF("node size = %zu", node_size);
4716 /* Move higher pointers up one slot. */
4717 for (i = NUMKEYS(mp); i > indx; i--)
4718 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
4720 /* Adjust free space offsets. */
4721 ofs = mp->mp_upper - node_size;
4722 assert(ofs >= mp->mp_lower + sizeof(indx_t));
4723 mp->mp_ptrs[indx] = ofs;
4725 mp->mp_lower += sizeof(indx_t);
4727 /* Write the node data. */
4728 node = NODEPTR(mp, indx);
4729 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
4730 node->mn_flags = flags;
4732 SETDSZ(node,data->mv_size);
4737 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
4742 if (F_ISSET(flags, F_BIGDATA))
4743 memcpy(node->mn_data + key->mv_size, data->mv_data,
4745 else if (F_ISSET(flags, MDB_RESERVE))
4746 data->mv_data = node->mn_data + key->mv_size;
4748 memcpy(node->mn_data + key->mv_size, data->mv_data,
4751 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
4753 if (F_ISSET(flags, MDB_RESERVE))
4754 data->mv_data = METADATA(ofp);
4756 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
4763 /** Delete the specified node from a page.
4764 * @param[in] mp The page to operate on.
4765 * @param[in] indx The index of the node to delete.
4766 * @param[in] ksize The size of a node. Only used if the page is
4767 * part of a #MDB_DUPFIXED database.
4770 mdb_node_del(MDB_page *mp, indx_t indx, int ksize)
4773 indx_t i, j, numkeys, ptr;
4780 COPY_PGNO(pgno, mp->mp_pgno);
4781 DPRINTF("delete node %u on %s page %zu", indx,
4782 IS_LEAF(mp) ? "leaf" : "branch", pgno);
4785 assert(indx < NUMKEYS(mp));
4788 int x = NUMKEYS(mp) - 1 - indx;
4789 base = LEAF2KEY(mp, indx, ksize);
4791 memmove(base, base + ksize, x * ksize);
4792 mp->mp_lower -= sizeof(indx_t);
4793 mp->mp_upper += ksize - sizeof(indx_t);
4797 node = NODEPTR(mp, indx);
4798 sz = NODESIZE + node->mn_ksize;
4800 if (F_ISSET(node->mn_flags, F_BIGDATA))
4801 sz += sizeof(pgno_t);
4803 sz += NODEDSZ(node);
4807 ptr = mp->mp_ptrs[indx];
4808 numkeys = NUMKEYS(mp);
4809 for (i = j = 0; i < numkeys; i++) {
4811 mp->mp_ptrs[j] = mp->mp_ptrs[i];
4812 if (mp->mp_ptrs[i] < ptr)
4813 mp->mp_ptrs[j] += sz;
4818 base = (char *)mp + mp->mp_upper;
4819 memmove(base + sz, base, ptr - mp->mp_upper);
4821 mp->mp_lower -= sizeof(indx_t);
4825 /** Compact the main page after deleting a node on a subpage.
4826 * @param[in] mp The main page to operate on.
4827 * @param[in] indx The index of the subpage on the main page.
4830 mdb_node_shrink(MDB_page *mp, indx_t indx)
4837 indx_t i, numkeys, ptr;
4839 node = NODEPTR(mp, indx);
4840 sp = (MDB_page *)NODEDATA(node);
4841 osize = NODEDSZ(node);
4843 delta = sp->mp_upper - sp->mp_lower;
4844 SETDSZ(node, osize - delta);
4845 xp = (MDB_page *)((char *)sp + delta);
4847 /* shift subpage upward */
4849 nsize = NUMKEYS(sp) * sp->mp_pad;
4850 memmove(METADATA(xp), METADATA(sp), nsize);
4853 nsize = osize - sp->mp_upper;
4854 numkeys = NUMKEYS(sp);
4855 for (i=numkeys-1; i>=0; i--)
4856 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
4858 xp->mp_upper = sp->mp_lower;
4859 xp->mp_lower = sp->mp_lower;
4860 xp->mp_flags = sp->mp_flags;
4861 xp->mp_pad = sp->mp_pad;
4862 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
4864 /* shift lower nodes upward */
4865 ptr = mp->mp_ptrs[indx];
4866 numkeys = NUMKEYS(mp);
4867 for (i = 0; i < numkeys; i++) {
4868 if (mp->mp_ptrs[i] <= ptr)
4869 mp->mp_ptrs[i] += delta;
4872 base = (char *)mp + mp->mp_upper;
4873 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
4874 mp->mp_upper += delta;
4877 /** Initial setup of a sorted-dups cursor.
4878 * Sorted duplicates are implemented as a sub-database for the given key.
4879 * The duplicate data items are actually keys of the sub-database.
4880 * Operations on the duplicate data items are performed using a sub-cursor
4881 * initialized when the sub-database is first accessed. This function does
4882 * the preliminary setup of the sub-cursor, filling in the fields that
4883 * depend only on the parent DB.
4884 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4887 mdb_xcursor_init0(MDB_cursor *mc)
4889 MDB_xcursor *mx = mc->mc_xcursor;
4891 mx->mx_cursor.mc_xcursor = NULL;
4892 mx->mx_cursor.mc_txn = mc->mc_txn;
4893 mx->mx_cursor.mc_db = &mx->mx_db;
4894 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
4895 mx->mx_cursor.mc_dbi = mc->mc_dbi+1;
4896 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
4897 mx->mx_cursor.mc_snum = 0;
4898 mx->mx_cursor.mc_top = 0;
4899 mx->mx_cursor.mc_flags = C_SUB;
4900 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
4901 mx->mx_dbx.md_dcmp = NULL;
4902 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
4905 /** Final setup of a sorted-dups cursor.
4906 * Sets up the fields that depend on the data from the main cursor.
4907 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
4908 * @param[in] node The data containing the #MDB_db record for the
4909 * sorted-dup database.
4912 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
4914 MDB_xcursor *mx = mc->mc_xcursor;
4916 if (node->mn_flags & F_SUBDATA) {
4917 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
4918 mx->mx_cursor.mc_snum = 0;
4919 mx->mx_cursor.mc_flags = C_SUB;
4921 MDB_page *fp = NODEDATA(node);
4922 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
4923 mx->mx_db.md_flags = 0;
4924 mx->mx_db.md_depth = 1;
4925 mx->mx_db.md_branch_pages = 0;
4926 mx->mx_db.md_leaf_pages = 1;
4927 mx->mx_db.md_overflow_pages = 0;
4928 mx->mx_db.md_entries = NUMKEYS(fp);
4929 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
4930 mx->mx_cursor.mc_snum = 1;
4931 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
4932 mx->mx_cursor.mc_top = 0;
4933 mx->mx_cursor.mc_pg[0] = fp;
4934 mx->mx_cursor.mc_ki[0] = 0;
4935 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
4936 mx->mx_db.md_flags = MDB_DUPFIXED;
4937 mx->mx_db.md_pad = fp->mp_pad;
4938 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
4939 mx->mx_db.md_flags |= MDB_INTEGERKEY;
4942 DPRINTF("Sub-db %u for db %u root page %zu", mx->mx_cursor.mc_dbi, mc->mc_dbi,
4944 mx->mx_dbflag = (F_ISSET(mc->mc_pg[mc->mc_top]->mp_flags, P_DIRTY)) ?
4946 mx->mx_dbx.md_name.mv_data = NODEKEY(node);
4947 mx->mx_dbx.md_name.mv_size = node->mn_ksize;
4948 #if UINT_MAX < SIZE_MAX
4949 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
4950 #ifdef MISALIGNED_OK
4951 mx->mx_dbx.md_cmp = mdb_cmp_long;
4953 mx->mx_dbx.md_cmp = mdb_cmp_cint;
4958 /** Initialize a cursor for a given transaction and database. */
4960 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
4965 mc->mc_db = &txn->mt_dbs[dbi];
4966 mc->mc_dbx = &txn->mt_dbxs[dbi];
4967 mc->mc_dbflag = &txn->mt_dbflags[dbi];
4971 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4973 mc->mc_xcursor = mx;
4974 mdb_xcursor_init0(mc);
4976 mc->mc_xcursor = NULL;
4981 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
4984 MDB_xcursor *mx = NULL;
4985 size_t size = sizeof(MDB_cursor);
4987 if (txn == NULL || ret == NULL || dbi >= txn->mt_numdbs)
4990 /* Allow read access to the freelist */
4991 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
4994 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
4995 size += sizeof(MDB_xcursor);
4997 if ((mc = malloc(size)) != NULL) {
4998 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
4999 mx = (MDB_xcursor *)(mc + 1);
5001 mdb_cursor_init(mc, txn, dbi, mx);
5002 if (txn->mt_cursors) {
5003 mc->mc_next = txn->mt_cursors[dbi];
5004 txn->mt_cursors[dbi] = mc;
5006 mc->mc_flags |= C_ALLOCD;
5016 /* Return the count of duplicate data items for the current key */
5018 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
5022 if (mc == NULL || countp == NULL)
5025 if (!(mc->mc_db->md_flags & MDB_DUPSORT))
5028 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5029 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5032 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
5035 *countp = mc->mc_xcursor->mx_db.md_entries;
5041 mdb_cursor_close(MDB_cursor *mc)
5044 /* remove from txn, if tracked */
5045 if (mc->mc_txn->mt_cursors) {
5046 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
5047 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
5049 *prev = mc->mc_next;
5051 if (mc->mc_flags & C_ALLOCD)
5057 mdb_cursor_txn(MDB_cursor *mc)
5059 if (!mc) return NULL;
5064 mdb_cursor_dbi(MDB_cursor *mc)
5070 /** Replace the key for a node with a new key.
5071 * @param[in] mp The page containing the node to operate on.
5072 * @param[in] indx The index of the node to operate on.
5073 * @param[in] key The new key to use.
5074 * @return 0 on success, non-zero on failure.
5077 mdb_update_key(MDB_page *mp, indx_t indx, MDB_val *key)
5083 indx_t ptr, i, numkeys;
5086 node = NODEPTR(mp, indx);
5087 ptr = mp->mp_ptrs[indx];
5091 char kbuf2[(MAXKEYSIZE*2+1)];
5092 k2.mv_data = NODEKEY(node);
5093 k2.mv_size = node->mn_ksize;
5094 DPRINTF("update key %u (ofs %u) [%s] to [%s] on page %zu",
5096 mdb_dkey(&k2, kbuf2),
5102 delta0 = delta = key->mv_size - node->mn_ksize;
5104 /* Must be 2-byte aligned. If new key is
5105 * shorter by 1, the shift will be skipped.
5107 delta += (delta & 1);
5109 if (delta > 0 && SIZELEFT(mp) < delta) {
5110 DPRINTF("OUCH! Not enough room, delta = %d", delta);
5114 numkeys = NUMKEYS(mp);
5115 for (i = 0; i < numkeys; i++) {
5116 if (mp->mp_ptrs[i] <= ptr)
5117 mp->mp_ptrs[i] -= delta;
5120 base = (char *)mp + mp->mp_upper;
5121 len = ptr - mp->mp_upper + NODESIZE;
5122 memmove(base - delta, base, len);
5123 mp->mp_upper -= delta;
5125 node = NODEPTR(mp, indx);
5128 /* But even if no shift was needed, update ksize */
5130 node->mn_ksize = key->mv_size;
5133 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
5138 /** Move a node from csrc to cdst.
5141 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
5147 unsigned short flags;
5151 /* Mark src and dst as dirty. */
5152 if ((rc = mdb_page_touch(csrc)) ||
5153 (rc = mdb_page_touch(cdst)))
5156 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5157 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0); /* fake */
5158 key.mv_size = csrc->mc_db->md_pad;
5159 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5161 data.mv_data = NULL;
5165 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
5166 assert(!((long)srcnode&1));
5167 srcpg = NODEPGNO(srcnode);
5168 flags = srcnode->mn_flags;
5169 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5170 unsigned int snum = csrc->mc_snum;
5172 /* must find the lowest key below src */
5173 mdb_page_search_root(csrc, NULL, 0);
5174 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5175 key.mv_size = NODEKSZ(s2);
5176 key.mv_data = NODEKEY(s2);
5177 csrc->mc_snum = snum--;
5178 csrc->mc_top = snum;
5180 key.mv_size = NODEKSZ(srcnode);
5181 key.mv_data = NODEKEY(srcnode);
5183 data.mv_size = NODEDSZ(srcnode);
5184 data.mv_data = NODEDATA(srcnode);
5186 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
5187 unsigned int snum = cdst->mc_snum;
5190 /* must find the lowest key below dst */
5191 mdb_page_search_root(cdst, NULL, 0);
5192 s2 = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5193 bkey.mv_size = NODEKSZ(s2);
5194 bkey.mv_data = NODEKEY(s2);
5195 cdst->mc_snum = snum--;
5196 cdst->mc_top = snum;
5197 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &bkey);
5200 DPRINTF("moving %s node %u [%s] on page %zu to node %u on page %zu",
5201 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
5202 csrc->mc_ki[csrc->mc_top],
5204 csrc->mc_pg[csrc->mc_top]->mp_pgno,
5205 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno);
5207 /* Add the node to the destination page.
5209 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
5210 if (rc != MDB_SUCCESS)
5213 /* Delete the node from the source page.
5215 mdb_node_del(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
5218 /* Adjust other cursors pointing to mp */
5219 MDB_cursor *m2, *m3;
5220 MDB_dbi dbi = csrc->mc_dbi;
5221 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
5223 if (csrc->mc_flags & C_SUB)
5226 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5227 if (m2 == csrc) continue;
5228 if (csrc->mc_flags & C_SUB)
5229 m3 = &m2->mc_xcursor->mx_cursor;
5232 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
5233 csrc->mc_ki[csrc->mc_top]) {
5234 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
5235 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
5240 /* Update the parent separators.
5242 if (csrc->mc_ki[csrc->mc_top] == 0) {
5243 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
5244 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5245 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
5247 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5248 key.mv_size = NODEKSZ(srcnode);
5249 key.mv_data = NODEKEY(srcnode);
5251 DPRINTF("update separator for source page %zu to [%s]",
5252 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key));
5253 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1],
5254 &key)) != MDB_SUCCESS)
5257 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5259 nullkey.mv_size = 0;
5260 rc = mdb_update_key(csrc->mc_pg[csrc->mc_top], 0, &nullkey);
5261 assert(rc == MDB_SUCCESS);
5265 if (cdst->mc_ki[cdst->mc_top] == 0) {
5266 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
5267 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5268 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
5270 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
5271 key.mv_size = NODEKSZ(srcnode);
5272 key.mv_data = NODEKEY(srcnode);
5274 DPRINTF("update separator for destination page %zu to [%s]",
5275 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key));
5276 if ((rc = mdb_update_key(cdst->mc_pg[cdst->mc_top-1], cdst->mc_ki[cdst->mc_top-1],
5277 &key)) != MDB_SUCCESS)
5280 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
5282 nullkey.mv_size = 0;
5283 rc = mdb_update_key(cdst->mc_pg[cdst->mc_top], 0, &nullkey);
5284 assert(rc == MDB_SUCCESS);
5291 /** Merge one page into another.
5292 * The nodes from the page pointed to by \b csrc will
5293 * be copied to the page pointed to by \b cdst and then
5294 * the \b csrc page will be freed.
5295 * @param[in] csrc Cursor pointing to the source page.
5296 * @param[in] cdst Cursor pointing to the destination page.
5299 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
5307 DPRINTF("merging page %zu into %zu", csrc->mc_pg[csrc->mc_top]->mp_pgno,
5308 cdst->mc_pg[cdst->mc_top]->mp_pgno);
5310 assert(csrc->mc_snum > 1); /* can't merge root page */
5311 assert(cdst->mc_snum > 1);
5313 /* Mark dst as dirty. */
5314 if ((rc = mdb_page_touch(cdst)))
5317 /* Move all nodes from src to dst.
5319 j = nkeys = NUMKEYS(cdst->mc_pg[cdst->mc_top]);
5320 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
5321 key.mv_size = csrc->mc_db->md_pad;
5322 key.mv_data = METADATA(csrc->mc_pg[csrc->mc_top]);
5323 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5324 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
5325 if (rc != MDB_SUCCESS)
5327 key.mv_data = (char *)key.mv_data + key.mv_size;
5330 for (i = 0; i < NUMKEYS(csrc->mc_pg[csrc->mc_top]); i++, j++) {
5331 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], i);
5332 if (i == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
5333 unsigned int snum = csrc->mc_snum;
5335 /* must find the lowest key below src */
5336 mdb_page_search_root(csrc, NULL, 0);
5337 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
5338 key.mv_size = NODEKSZ(s2);
5339 key.mv_data = NODEKEY(s2);
5340 csrc->mc_snum = snum--;
5341 csrc->mc_top = snum;
5343 key.mv_size = srcnode->mn_ksize;
5344 key.mv_data = NODEKEY(srcnode);
5347 data.mv_size = NODEDSZ(srcnode);
5348 data.mv_data = NODEDATA(srcnode);
5349 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
5350 if (rc != MDB_SUCCESS)
5355 DPRINTF("dst page %zu now has %u keys (%.1f%% filled)",
5356 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);
5358 /* Unlink the src page from parent and add to free list.
5360 mdb_node_del(csrc->mc_pg[csrc->mc_top-1], csrc->mc_ki[csrc->mc_top-1], 0);
5361 if (csrc->mc_ki[csrc->mc_top-1] == 0) {
5363 if ((rc = mdb_update_key(csrc->mc_pg[csrc->mc_top-1], 0, &key)) != MDB_SUCCESS)
5367 mdb_midl_append(&csrc->mc_txn->mt_free_pgs, csrc->mc_pg[csrc->mc_top]->mp_pgno);
5368 if (IS_LEAF(csrc->mc_pg[csrc->mc_top]))
5369 csrc->mc_db->md_leaf_pages--;
5371 csrc->mc_db->md_branch_pages--;
5373 /* Adjust other cursors pointing to mp */
5374 MDB_cursor *m2, *m3;
5375 MDB_dbi dbi = csrc->mc_dbi;
5376 MDB_page *mp = cdst->mc_pg[cdst->mc_top];
5378 if (csrc->mc_flags & C_SUB)
5381 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5382 if (csrc->mc_flags & C_SUB)
5383 m3 = &m2->mc_xcursor->mx_cursor;
5386 if (m3 == csrc) continue;
5387 if (m3->mc_snum < csrc->mc_snum) continue;
5388 if (m3->mc_pg[csrc->mc_top] == csrc->mc_pg[csrc->mc_top]) {
5389 m3->mc_pg[csrc->mc_top] = mp;
5390 m3->mc_ki[csrc->mc_top] += nkeys;
5394 mdb_cursor_pop(csrc);
5396 return mdb_rebalance(csrc);
5399 /** Copy the contents of a cursor.
5400 * @param[in] csrc The cursor to copy from.
5401 * @param[out] cdst The cursor to copy to.
5404 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
5408 cdst->mc_txn = csrc->mc_txn;
5409 cdst->mc_dbi = csrc->mc_dbi;
5410 cdst->mc_db = csrc->mc_db;
5411 cdst->mc_dbx = csrc->mc_dbx;
5412 cdst->mc_snum = csrc->mc_snum;
5413 cdst->mc_top = csrc->mc_top;
5414 cdst->mc_flags = csrc->mc_flags;
5416 for (i=0; i<csrc->mc_snum; i++) {
5417 cdst->mc_pg[i] = csrc->mc_pg[i];
5418 cdst->mc_ki[i] = csrc->mc_ki[i];
5422 /** Rebalance the tree after a delete operation.
5423 * @param[in] mc Cursor pointing to the page where rebalancing
5425 * @return 0 on success, non-zero on failure.
5428 mdb_rebalance(MDB_cursor *mc)
5438 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5439 DPRINTF("rebalancing %s page %zu (has %u keys, %.1f%% full)",
5440 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
5441 pgno, NUMKEYS(mc->mc_pg[mc->mc_top]), (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10);
5445 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD) {
5448 COPY_PGNO(pgno, mc->mc_pg[mc->mc_top]->mp_pgno);
5449 DPRINTF("no need to rebalance page %zu, above fill threshold",
5455 if (mc->mc_snum < 2) {
5456 MDB_page *mp = mc->mc_pg[0];
5457 if (NUMKEYS(mp) == 0) {
5458 DPUTS("tree is completely empty");
5459 mc->mc_db->md_root = P_INVALID;
5460 mc->mc_db->md_depth = 0;
5461 mc->mc_db->md_leaf_pages = 0;
5462 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5466 /* Adjust other cursors pointing to mp */
5467 MDB_cursor *m2, *m3;
5468 MDB_dbi dbi = mc->mc_dbi;
5470 if (mc->mc_flags & C_SUB)
5473 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5474 if (m2 == mc) continue;
5475 if (mc->mc_flags & C_SUB)
5476 m3 = &m2->mc_xcursor->mx_cursor;
5479 if (m3->mc_snum < mc->mc_snum) continue;
5480 if (m3->mc_pg[0] == mp) {
5486 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
5487 DPUTS("collapsing root page!");
5488 mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
5489 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
5490 if ((rc = mdb_page_get(mc->mc_txn, mc->mc_db->md_root,
5493 mc->mc_db->md_depth--;
5494 mc->mc_db->md_branch_pages--;
5496 /* Adjust other cursors pointing to mp */
5497 MDB_cursor *m2, *m3;
5498 MDB_dbi dbi = mc->mc_dbi;
5500 if (mc->mc_flags & C_SUB)
5503 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5504 if (m2 == mc) continue;
5505 if (mc->mc_flags & C_SUB)
5506 m3 = &m2->mc_xcursor->mx_cursor;
5509 if (m3->mc_snum < mc->mc_snum) continue;
5510 if (m3->mc_pg[0] == mp) {
5511 m3->mc_pg[0] = mc->mc_pg[0];
5516 DPUTS("root page doesn't need rebalancing");
5520 /* The parent (branch page) must have at least 2 pointers,
5521 * otherwise the tree is invalid.
5523 ptop = mc->mc_top-1;
5524 assert(NUMKEYS(mc->mc_pg[ptop]) > 1);
5526 /* Leaf page fill factor is below the threshold.
5527 * Try to move keys from left or right neighbor, or
5528 * merge with a neighbor page.
5533 mdb_cursor_copy(mc, &mn);
5534 mn.mc_xcursor = NULL;
5536 if (mc->mc_ki[ptop] == 0) {
5537 /* We're the leftmost leaf in our parent.
5539 DPUTS("reading right neighbor");
5541 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5542 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5544 mn.mc_ki[mn.mc_top] = 0;
5545 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
5547 /* There is at least one neighbor to the left.
5549 DPUTS("reading left neighbor");
5551 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
5552 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mn.mc_pg[mn.mc_top])))
5554 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
5555 mc->mc_ki[mc->mc_top] = 0;
5558 DPRINTF("found neighbor page %zu (%u keys, %.1f%% full)",
5559 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);
5561 /* If the neighbor page is above threshold and has at least two
5562 * keys, move one key from it.
5564 * Otherwise we should try to merge them.
5566 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) >= 2)
5567 return mdb_node_move(&mn, mc);
5568 else { /* FIXME: if (has_enough_room()) */
5569 mc->mc_flags &= ~C_INITIALIZED;
5570 if (mc->mc_ki[ptop] == 0)
5571 return mdb_page_merge(&mn, mc);
5573 return mdb_page_merge(mc, &mn);
5577 /** Complete a delete operation started by #mdb_cursor_del(). */
5579 mdb_cursor_del0(MDB_cursor *mc, MDB_node *leaf)
5583 /* add overflow pages to free list */
5584 if (!IS_LEAF2(mc->mc_pg[mc->mc_top]) && F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5588 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
5589 ovpages = OVPAGES(NODEDSZ(leaf), mc->mc_txn->mt_env->me_psize);
5590 mc->mc_db->md_overflow_pages -= ovpages;
5591 for (i=0; i<ovpages; i++) {
5592 DPRINTF("freed ov page %zu", pg);
5593 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
5597 mdb_node_del(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], mc->mc_db->md_pad);
5598 mc->mc_db->md_entries--;
5599 rc = mdb_rebalance(mc);
5600 if (rc != MDB_SUCCESS)
5601 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5607 mdb_del(MDB_txn *txn, MDB_dbi dbi,
5608 MDB_val *key, MDB_val *data)
5613 MDB_val rdata, *xdata;
5617 assert(key != NULL);
5619 DPRINTF("====> delete db %u key [%s]", dbi, DKEY(key));
5621 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
5624 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
5628 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
5632 mdb_cursor_init(&mc, txn, dbi, &mx);
5643 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
5645 rc = mdb_cursor_del(&mc, data ? 0 : MDB_NODUPDATA);
5649 /** Split a page and insert a new node.
5650 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
5651 * The cursor will be updated to point to the actual page and index where
5652 * the node got inserted after the split.
5653 * @param[in] newkey The key for the newly inserted node.
5654 * @param[in] newdata The data for the newly inserted node.
5655 * @param[in] newpgno The page number, if the new node is a branch node.
5656 * @return 0 on success, non-zero on failure.
5659 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
5660 unsigned int nflags)
5663 int rc = MDB_SUCCESS, ins_new = 0, new_root = 0, newpos = 1;
5666 unsigned int i, j, split_indx, nkeys, pmax;
5668 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
5670 MDB_page *mp, *rp, *pp;
5675 mp = mc->mc_pg[mc->mc_top];
5676 newindx = mc->mc_ki[mc->mc_top];
5678 DPRINTF("-----> splitting %s page %zu and adding [%s] at index %i",
5679 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
5680 DKEY(newkey), mc->mc_ki[mc->mc_top]);
5682 /* Create a right sibling. */
5683 if ((rp = mdb_page_new(mc, mp->mp_flags, 1)) == NULL)
5685 DPRINTF("new right sibling: page %zu", rp->mp_pgno);
5687 if (mc->mc_snum < 2) {
5688 if ((pp = mdb_page_new(mc, P_BRANCH, 1)) == NULL)
5690 /* shift current top to make room for new parent */
5691 mc->mc_pg[1] = mc->mc_pg[0];
5692 mc->mc_ki[1] = mc->mc_ki[0];
5695 mc->mc_db->md_root = pp->mp_pgno;
5696 DPRINTF("root split! new root = %zu", pp->mp_pgno);
5697 mc->mc_db->md_depth++;
5700 /* Add left (implicit) pointer. */
5701 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
5702 /* undo the pre-push */
5703 mc->mc_pg[0] = mc->mc_pg[1];
5704 mc->mc_ki[0] = mc->mc_ki[1];
5705 mc->mc_db->md_root = mp->mp_pgno;
5706 mc->mc_db->md_depth--;
5713 ptop = mc->mc_top-1;
5714 DPRINTF("parent branch page is %zu", mc->mc_pg[ptop]->mp_pgno);
5717 mdb_cursor_copy(mc, &mn);
5718 mn.mc_pg[mn.mc_top] = rp;
5719 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
5721 if (nflags & MDB_APPEND) {
5722 mn.mc_ki[mn.mc_top] = 0;
5729 nkeys = NUMKEYS(mp);
5730 split_indx = (nkeys + 1) / 2;
5735 unsigned int lsize, rsize, ksize;
5736 /* Move half of the keys to the right sibling */
5738 x = mc->mc_ki[mc->mc_top] - split_indx;
5739 ksize = mc->mc_db->md_pad;
5740 split = LEAF2KEY(mp, split_indx, ksize);
5741 rsize = (nkeys - split_indx) * ksize;
5742 lsize = (nkeys - split_indx) * sizeof(indx_t);
5743 mp->mp_lower -= lsize;
5744 rp->mp_lower += lsize;
5745 mp->mp_upper += rsize - lsize;
5746 rp->mp_upper -= rsize - lsize;
5747 sepkey.mv_size = ksize;
5748 if (newindx == split_indx) {
5749 sepkey.mv_data = newkey->mv_data;
5751 sepkey.mv_data = split;
5754 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
5755 memcpy(rp->mp_ptrs, split, rsize);
5756 sepkey.mv_data = rp->mp_ptrs;
5757 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
5758 memcpy(ins, newkey->mv_data, ksize);
5759 mp->mp_lower += sizeof(indx_t);
5760 mp->mp_upper -= ksize - sizeof(indx_t);
5763 memcpy(rp->mp_ptrs, split, x * ksize);
5764 ins = LEAF2KEY(rp, x, ksize);
5765 memcpy(ins, newkey->mv_data, ksize);
5766 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
5767 rp->mp_lower += sizeof(indx_t);
5768 rp->mp_upper -= ksize - sizeof(indx_t);
5769 mc->mc_ki[mc->mc_top] = x;
5770 mc->mc_pg[mc->mc_top] = rp;
5775 /* For leaf pages, check the split point based on what
5776 * fits where, since otherwise mdb_node_add can fail.
5778 * This check is only needed when the data items are
5779 * relatively large, such that being off by one will
5780 * make the difference between success or failure.
5781 * When the size of the data items is much smaller than
5782 * one-half of a page, this check is irrelevant.
5785 unsigned int psize, nsize;
5786 /* Maximum free space in an empty page */
5787 pmax = mc->mc_txn->mt_env->me_psize - PAGEHDRSZ;
5788 nsize = mdb_leaf_size(mc->mc_txn->mt_env, newkey, newdata);
5789 if ((nkeys < 20) || (nsize > pmax/4)) {
5790 if (newindx <= split_indx) {
5793 for (i=0; i<split_indx; i++) {
5794 node = NODEPTR(mp, i);
5795 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5796 if (F_ISSET(node->mn_flags, F_BIGDATA))
5797 psize += sizeof(pgno_t);
5799 psize += NODEDSZ(node);
5802 if (i == split_indx - 1 && newindx == split_indx)
5811 for (i=nkeys-1; i>=split_indx; i--) {
5812 node = NODEPTR(mp, i);
5813 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
5814 if (F_ISSET(node->mn_flags, F_BIGDATA))
5815 psize += sizeof(pgno_t);
5817 psize += NODEDSZ(node);
5828 /* First find the separating key between the split pages.
5829 * The case where newindx == split_indx is ambiguous; the
5830 * new item could go to the new page or stay on the original
5831 * page. If newpos == 1 it goes to the new page.
5833 if (newindx == split_indx && newpos) {
5834 sepkey.mv_size = newkey->mv_size;
5835 sepkey.mv_data = newkey->mv_data;
5837 node = NODEPTR(mp, split_indx);
5838 sepkey.mv_size = node->mn_ksize;
5839 sepkey.mv_data = NODEKEY(node);
5843 DPRINTF("separator is [%s]", DKEY(&sepkey));
5845 /* Copy separator key to the parent.
5847 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(mc->mc_txn->mt_env, &sepkey)) {
5850 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
5852 /* Right page might now have changed parent.
5853 * Check if left page also changed parent.
5855 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
5856 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
5857 mc->mc_pg[ptop] = mn.mc_pg[ptop];
5858 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
5862 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
5865 if (rc != MDB_SUCCESS) {
5868 if (nflags & MDB_APPEND) {
5869 mc->mc_pg[mc->mc_top] = rp;
5870 mc->mc_ki[mc->mc_top] = 0;
5871 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
5880 /* Move half of the keys to the right sibling. */
5882 /* grab a page to hold a temporary copy */
5883 copy = mdb_page_malloc(mc);
5887 copy->mp_pgno = mp->mp_pgno;
5888 copy->mp_flags = mp->mp_flags;
5889 copy->mp_lower = PAGEHDRSZ;
5890 copy->mp_upper = mc->mc_txn->mt_env->me_psize;
5891 mc->mc_pg[mc->mc_top] = copy;
5892 for (i = j = 0; i <= nkeys; j++) {
5893 if (i == split_indx) {
5894 /* Insert in right sibling. */
5895 /* Reset insert index for right sibling. */
5896 if (i != newindx || (newpos ^ ins_new)) {
5898 mc->mc_pg[mc->mc_top] = rp;
5902 if (i == newindx && !ins_new) {
5903 /* Insert the original entry that caused the split. */
5904 rkey.mv_data = newkey->mv_data;
5905 rkey.mv_size = newkey->mv_size;
5914 /* Update index for the new key. */
5915 mc->mc_ki[mc->mc_top] = j;
5916 } else if (i == nkeys) {
5919 node = NODEPTR(mp, i);
5920 rkey.mv_data = NODEKEY(node);
5921 rkey.mv_size = node->mn_ksize;
5923 xdata.mv_data = NODEDATA(node);
5924 xdata.mv_size = NODEDSZ(node);
5927 pgno = NODEPGNO(node);
5928 flags = node->mn_flags;
5933 if (!IS_LEAF(mp) && j == 0) {
5934 /* First branch index doesn't need key data. */
5938 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
5942 nkeys = NUMKEYS(copy);
5943 for (i=0; i<nkeys; i++)
5944 mp->mp_ptrs[i] = copy->mp_ptrs[i];
5945 mp->mp_lower = copy->mp_lower;
5946 mp->mp_upper = copy->mp_upper;
5947 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
5948 mc->mc_txn->mt_env->me_psize - copy->mp_upper);
5950 /* reset back to original page */
5951 if (newindx < split_indx || (!newpos && newindx == split_indx)) {
5952 mc->mc_pg[mc->mc_top] = mp;
5953 if (nflags & MDB_RESERVE) {
5954 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5955 if (!(node->mn_flags & F_BIGDATA))
5956 newdata->mv_data = NODEDATA(node);
5960 /* return tmp page to freelist */
5961 copy->mp_next = mc->mc_txn->mt_env->me_dpages;
5962 VGMEMP_FREE(mc->mc_txn->mt_env, copy);
5963 mc->mc_txn->mt_env->me_dpages = copy;
5966 /* Adjust other cursors pointing to mp */
5967 MDB_cursor *m2, *m3;
5968 MDB_dbi dbi = mc->mc_dbi;
5970 if (mc->mc_flags & C_SUB)
5973 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
5974 if (m2 == mc) continue;
5975 if (mc->mc_flags & C_SUB)
5976 m3 = &m2->mc_xcursor->mx_cursor;
5979 if (!(m3->mc_flags & C_INITIALIZED))
5984 for (k=m3->mc_top; k>=0; k--) {
5985 m3->mc_ki[k+1] = m3->mc_ki[k];
5986 m3->mc_pg[k+1] = m3->mc_pg[k];
5988 m3->mc_ki[0] = mc->mc_ki[0];
5989 m3->mc_pg[0] = mc->mc_pg[0];
5993 if (m3->mc_pg[mc->mc_top] == mp) {
5994 if (m3->mc_ki[m3->mc_top] >= split_indx) {
5995 m3->mc_pg[m3->mc_top] = rp;
5996 m3->mc_ki[m3->mc_top] -= split_indx;
6005 mdb_put(MDB_txn *txn, MDB_dbi dbi,
6006 MDB_val *key, MDB_val *data, unsigned int flags)
6011 assert(key != NULL);
6012 assert(data != NULL);
6014 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6017 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
6021 if (key->mv_size == 0 || key->mv_size > MAXKEYSIZE) {
6025 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND)) != flags)
6028 mdb_cursor_init(&mc, txn, dbi, &mx);
6029 return mdb_cursor_put(&mc, key, data, flags);
6032 /** Only a subset of the @ref mdb_env flags can be changed
6033 * at runtime. Changing other flags requires closing the environment
6034 * and re-opening it with the new flags.
6036 #define CHANGEABLE (MDB_NOSYNC)
6038 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
6040 if ((flag & CHANGEABLE) != flag)
6043 env->me_flags |= flag;
6045 env->me_flags &= ~flag;
6050 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
6055 *arg = env->me_flags;
6060 mdb_env_get_path(MDB_env *env, const char **arg)
6065 *arg = env->me_path;
6069 /** Common code for #mdb_stat() and #mdb_env_stat().
6070 * @param[in] env the environment to operate in.
6071 * @param[in] db the #MDB_db record containing the stats to return.
6072 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
6073 * @return 0, this function always succeeds.
6076 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
6078 arg->ms_psize = env->me_psize;
6079 arg->ms_depth = db->md_depth;
6080 arg->ms_branch_pages = db->md_branch_pages;
6081 arg->ms_leaf_pages = db->md_leaf_pages;
6082 arg->ms_overflow_pages = db->md_overflow_pages;
6083 arg->ms_entries = db->md_entries;
6088 mdb_env_stat(MDB_env *env, MDB_stat *arg)
6092 if (env == NULL || arg == NULL)
6095 mdb_env_read_meta(env, &toggle);
6097 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
6100 /** Set the default comparison functions for a database.
6101 * Called immediately after a database is opened to set the defaults.
6102 * The user can then override them with #mdb_set_compare() or
6103 * #mdb_set_dupsort().
6104 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
6105 * @param[in] dbi A database handle returned by #mdb_open()
6108 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
6110 if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEKEY)
6111 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memnr;
6112 else if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERKEY)
6113 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_cint;
6115 txn->mt_dbxs[dbi].md_cmp = mdb_cmp_memn;
6117 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6118 if (txn->mt_dbs[dbi].md_flags & MDB_INTEGERDUP) {
6119 if (txn->mt_dbs[dbi].md_flags & MDB_DUPFIXED)
6120 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_int;
6122 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_cint;
6123 } else if (txn->mt_dbs[dbi].md_flags & MDB_REVERSEDUP) {
6124 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memnr;
6126 txn->mt_dbxs[dbi].md_dcmp = mdb_cmp_memn;
6129 txn->mt_dbxs[dbi].md_dcmp = NULL;
6133 int mdb_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
6138 int rc, dbflag, exact;
6141 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
6142 mdb_default_cmp(txn, FREE_DBI);
6148 if (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY))
6149 txn->mt_dbs[MAIN_DBI].md_flags |= (flags & (MDB_DUPSORT|MDB_REVERSEKEY|MDB_INTEGERKEY));
6150 mdb_default_cmp(txn, MAIN_DBI);
6154 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
6155 mdb_default_cmp(txn, MAIN_DBI);
6158 /* Is the DB already open? */
6160 for (i=2; i<txn->mt_numdbs; i++) {
6161 if (len == txn->mt_dbxs[i].md_name.mv_size &&
6162 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
6168 if (txn->mt_numdbs >= txn->mt_env->me_maxdbs - 1)
6171 /* Find the DB info */
6175 key.mv_data = (void *)name;
6176 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
6177 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
6178 if (rc == MDB_SUCCESS) {
6179 /* make sure this is actually a DB */
6180 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
6181 if (!(node->mn_flags & F_SUBDATA))
6183 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
6184 /* Create if requested */
6186 data.mv_size = sizeof(MDB_db);
6187 data.mv_data = &dummy;
6188 memset(&dummy, 0, sizeof(dummy));
6189 dummy.md_root = P_INVALID;
6190 dummy.md_flags = flags & 0xffff;
6191 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
6195 /* OK, got info, add to table */
6196 if (rc == MDB_SUCCESS) {
6197 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_data = strdup(name);
6198 txn->mt_dbxs[txn->mt_numdbs].md_name.mv_size = len;
6199 txn->mt_dbxs[txn->mt_numdbs].md_rel = NULL;
6200 txn->mt_dbflags[txn->mt_numdbs] = dbflag;
6201 memcpy(&txn->mt_dbs[txn->mt_numdbs], data.mv_data, sizeof(MDB_db));
6202 *dbi = txn->mt_numdbs;
6203 txn->mt_env->me_dbs[0][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
6204 txn->mt_env->me_dbs[1][txn->mt_numdbs] = txn->mt_dbs[txn->mt_numdbs];
6205 mdb_default_cmp(txn, txn->mt_numdbs);
6212 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
6214 if (txn == NULL || arg == NULL || dbi >= txn->mt_numdbs)
6217 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
6220 void mdb_close(MDB_env *env, MDB_dbi dbi)
6223 if (dbi <= MAIN_DBI || dbi >= env->me_numdbs)
6225 ptr = env->me_dbxs[dbi].md_name.mv_data;
6226 env->me_dbxs[dbi].md_name.mv_data = NULL;
6227 env->me_dbxs[dbi].md_name.mv_size = 0;
6231 /** Add all the DB's pages to the free list.
6232 * @param[in] mc Cursor on the DB to free.
6233 * @param[in] subs non-Zero to check for sub-DBs in this DB.
6234 * @return 0 on success, non-zero on failure.
6237 mdb_drop0(MDB_cursor *mc, int subs)
6241 rc = mdb_page_search(mc, NULL, 0);
6242 if (rc == MDB_SUCCESS) {
6247 /* LEAF2 pages have no nodes, cannot have sub-DBs */
6248 if (!subs || IS_LEAF2(mc->mc_pg[mc->mc_top]))
6251 mdb_cursor_copy(mc, &mx);
6252 while (mc->mc_snum > 0) {
6253 if (IS_LEAF(mc->mc_pg[mc->mc_top])) {
6254 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6255 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6256 if (ni->mn_flags & F_SUBDATA) {
6257 mdb_xcursor_init1(mc, ni);
6258 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6264 for (i=0; i<NUMKEYS(mc->mc_pg[mc->mc_top]); i++) {
6266 ni = NODEPTR(mc->mc_pg[mc->mc_top], i);
6269 mdb_midl_append(&mc->mc_txn->mt_free_pgs, pg);
6274 rc = mdb_cursor_sibling(mc, 1);
6276 /* no more siblings, go back to beginning
6277 * of previous level. (stack was already popped
6278 * by mdb_cursor_sibling)
6280 for (i=1; i<mc->mc_top; i++)
6281 mc->mc_pg[i] = mx.mc_pg[i];
6285 mdb_midl_append(&mc->mc_txn->mt_free_pgs,
6286 mc->mc_db->md_root);
6291 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
6296 if (!txn || !dbi || dbi >= txn->mt_numdbs)
6299 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6302 rc = mdb_cursor_open(txn, dbi, &mc);
6306 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
6310 /* Can't delete the main DB */
6311 if (del && dbi > MAIN_DBI) {
6312 rc = mdb_del(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL);
6314 mdb_close(txn->mt_env, dbi);
6316 txn->mt_dbflags[dbi] |= DB_DIRTY;
6317 txn->mt_dbs[dbi].md_depth = 0;
6318 txn->mt_dbs[dbi].md_branch_pages = 0;
6319 txn->mt_dbs[dbi].md_leaf_pages = 0;
6320 txn->mt_dbs[dbi].md_overflow_pages = 0;
6321 txn->mt_dbs[dbi].md_entries = 0;
6322 txn->mt_dbs[dbi].md_root = P_INVALID;
6325 mdb_cursor_close(mc);
6329 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6331 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6334 txn->mt_dbxs[dbi].md_cmp = cmp;
6338 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
6340 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6343 txn->mt_dbxs[dbi].md_dcmp = cmp;
6347 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
6349 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6352 txn->mt_dbxs[dbi].md_rel = rel;
6356 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
6358 if (txn == NULL || !dbi || dbi >= txn->mt_numdbs)
6361 txn->mt_dbxs[dbi].md_relctx = ctx;