2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2014 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.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
73 #if defined(__mips) && defined(__linux)
74 /* MIPS has cache coherency issues, requires explicit cache control */
75 #include <asm/cachectl.h>
76 extern int cacheflush(char *addr, int nbytes, int cache);
77 #define CACHEFLUSH(addr, bytes, cache) cacheflush(addr, bytes, cache)
79 #define CACHEFLUSH(addr, bytes, cache)
82 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
83 /** fdatasync is broken on ext3/ext4fs on older kernels, see
84 * description in #mdb_env_open2 comments
86 #define BROKEN_FDATASYNC
99 #if defined(__sun) || defined(ANDROID)
100 /* Most platforms have posix_memalign, older may only have memalign */
101 #define HAVE_MEMALIGN 1
105 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
106 #include <netinet/in.h>
107 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
110 #if defined(__APPLE__) || defined (BSD)
111 # define MDB_USE_POSIX_SEM 1
112 # define MDB_FDATASYNC fsync
113 #elif defined(ANDROID)
114 # define MDB_FDATASYNC fsync
119 #ifdef MDB_USE_POSIX_SEM
120 # define MDB_USE_HASH 1
121 #include <semaphore.h>
126 #include <valgrind/memcheck.h>
127 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
128 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
129 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
130 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
131 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
133 #define VGMEMP_CREATE(h,r,z)
134 #define VGMEMP_ALLOC(h,a,s)
135 #define VGMEMP_FREE(h,a)
136 #define VGMEMP_DESTROY(h)
137 #define VGMEMP_DEFINED(a,s)
141 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
142 /* Solaris just defines one or the other */
143 # define LITTLE_ENDIAN 1234
144 # define BIG_ENDIAN 4321
145 # ifdef _LITTLE_ENDIAN
146 # define BYTE_ORDER LITTLE_ENDIAN
148 # define BYTE_ORDER BIG_ENDIAN
151 # define BYTE_ORDER __BYTE_ORDER
155 #ifndef LITTLE_ENDIAN
156 #define LITTLE_ENDIAN __LITTLE_ENDIAN
159 #define BIG_ENDIAN __BIG_ENDIAN
162 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
163 #define MISALIGNED_OK 1
169 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
170 # error "Unknown or unsupported endianness (BYTE_ORDER)"
171 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
172 # error "Two's complement, reasonably sized integer types, please"
176 /** Put infrequently used env functions in separate section */
178 # define ESECT __attribute__ ((section("__TEXT,text_env")))
180 # define ESECT __attribute__ ((section("text_env")))
186 /** @defgroup internal LMDB Internals
189 /** @defgroup compat Compatibility Macros
190 * A bunch of macros to minimize the amount of platform-specific ifdefs
191 * needed throughout the rest of the code. When the features this library
192 * needs are similar enough to POSIX to be hidden in a one-or-two line
193 * replacement, this macro approach is used.
197 /** Features under development */
202 /** Wrapper around __func__, which is a C99 feature */
203 #if __STDC_VERSION__ >= 199901L
204 # define mdb_func_ __func__
205 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
206 # define mdb_func_ __FUNCTION__
208 /* If a debug message says <mdb_unknown>(), update the #if statements above */
209 # define mdb_func_ "<mdb_unknown>"
213 #define MDB_USE_HASH 1
214 #define MDB_PIDLOCK 0
215 #define THREAD_RET DWORD
216 #define pthread_t HANDLE
217 #define pthread_mutex_t HANDLE
218 #define pthread_cond_t HANDLE
219 #define pthread_key_t DWORD
220 #define pthread_self() GetCurrentThreadId()
221 #define pthread_key_create(x,y) \
222 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
223 #define pthread_key_delete(x) TlsFree(x)
224 #define pthread_getspecific(x) TlsGetValue(x)
225 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
226 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
227 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
228 #define pthread_cond_signal(x) SetEvent(*x)
229 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
230 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
231 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
232 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
233 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
234 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
235 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
236 #define getpid() GetCurrentProcessId()
237 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
238 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
239 #define ErrCode() GetLastError()
240 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
241 #define close(fd) (CloseHandle(fd) ? 0 : -1)
242 #define munmap(ptr,len) UnmapViewOfFile(ptr)
243 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
244 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
246 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
250 #define THREAD_RET void *
251 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
252 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
253 #define Z "z" /**< printf format modifier for size_t */
255 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
256 #define MDB_PIDLOCK 1
258 #ifdef MDB_USE_POSIX_SEM
260 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
261 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
262 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
263 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
266 mdb_sem_wait(sem_t *sem)
269 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
274 /** Lock the reader mutex.
276 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
277 /** Unlock the reader mutex.
279 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
281 /** Lock the writer mutex.
282 * Only a single write transaction is allowed at a time. Other writers
283 * will block waiting for this mutex.
285 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
286 /** Unlock the writer mutex.
288 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
289 #endif /* MDB_USE_POSIX_SEM */
291 /** Get the error code for the last failed system function.
293 #define ErrCode() errno
295 /** An abstraction for a file handle.
296 * On POSIX systems file handles are small integers. On Windows
297 * they're opaque pointers.
301 /** A value for an invalid file handle.
302 * Mainly used to initialize file variables and signify that they are
305 #define INVALID_HANDLE_VALUE (-1)
307 /** Get the size of a memory page for the system.
308 * This is the basic size that the platform's memory manager uses, and is
309 * fundamental to the use of memory-mapped files.
311 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
314 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
317 #define MNAME_LEN (sizeof(pthread_mutex_t))
323 /** A flag for opening a file and requesting synchronous data writes.
324 * This is only used when writing a meta page. It's not strictly needed;
325 * we could just do a normal write and then immediately perform a flush.
326 * But if this flag is available it saves us an extra system call.
328 * @note If O_DSYNC is undefined but exists in /usr/include,
329 * preferably set some compiler flag to get the definition.
330 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
333 # define MDB_DSYNC O_DSYNC
337 /** Function for flushing the data of a file. Define this to fsync
338 * if fdatasync() is not supported.
340 #ifndef MDB_FDATASYNC
341 # define MDB_FDATASYNC fdatasync
345 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
356 /** A page number in the database.
357 * Note that 64 bit page numbers are overkill, since pages themselves
358 * already represent 12-13 bits of addressable memory, and the OS will
359 * always limit applications to a maximum of 63 bits of address space.
361 * @note In the #MDB_node structure, we only store 48 bits of this value,
362 * which thus limits us to only 60 bits of addressable data.
364 typedef MDB_ID pgno_t;
366 /** A transaction ID.
367 * See struct MDB_txn.mt_txnid for details.
369 typedef MDB_ID txnid_t;
371 /** @defgroup debug Debug Macros
375 /** Enable debug output. Needs variable argument macros (a C99 feature).
376 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
377 * read from and written to the database (used for free space management).
383 static int mdb_debug;
384 static txnid_t mdb_debug_start;
386 /** Print a debug message with printf formatting.
387 * Requires double parenthesis around 2 or more args.
389 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
390 # define DPRINTF0(fmt, ...) \
391 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
393 # define DPRINTF(args) ((void) 0)
395 /** Print a debug string.
396 * The string is printed literally, with no format processing.
398 #define DPUTS(arg) DPRINTF(("%s", arg))
399 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
401 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
404 /** @brief The maximum size of a database page.
406 * It is 32k or 64k, since value-PAGEBASE must fit in
407 * #MDB_page.%mp_upper.
409 * LMDB will use database pages < OS pages if needed.
410 * That causes more I/O in write transactions: The OS must
411 * know (read) the whole page before writing a partial page.
413 * Note that we don't currently support Huge pages. On Linux,
414 * regular data files cannot use Huge pages, and in general
415 * Huge pages aren't actually pageable. We rely on the OS
416 * demand-pager to read our data and page it out when memory
417 * pressure from other processes is high. So until OSs have
418 * actual paging support for Huge pages, they're not viable.
420 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
422 /** The minimum number of keys required in a database page.
423 * Setting this to a larger value will place a smaller bound on the
424 * maximum size of a data item. Data items larger than this size will
425 * be pushed into overflow pages instead of being stored directly in
426 * the B-tree node. This value used to default to 4. With a page size
427 * of 4096 bytes that meant that any item larger than 1024 bytes would
428 * go into an overflow page. That also meant that on average 2-3KB of
429 * each overflow page was wasted space. The value cannot be lower than
430 * 2 because then there would no longer be a tree structure. With this
431 * value, items larger than 2KB will go into overflow pages, and on
432 * average only 1KB will be wasted.
434 #define MDB_MINKEYS 2
436 /** A stamp that identifies a file as an LMDB file.
437 * There's nothing special about this value other than that it is easily
438 * recognizable, and it will reflect any byte order mismatches.
440 #define MDB_MAGIC 0xBEEFC0DE
442 /** The version number for a database's datafile format. */
443 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
444 /** The version number for a database's lockfile format. */
445 #define MDB_LOCK_VERSION 1
447 /** @brief The max size of a key we can write, or 0 for dynamic max.
449 * Define this as 0 to compute the max from the page size. 511
450 * is default for backwards compat: liblmdb <= 0.9.10 can break
451 * when modifying a DB with keys/dupsort data bigger than its max.
452 * #MDB_DEVEL sets the default to 0.
454 * Data items in an #MDB_DUPSORT database are also limited to
455 * this size, since they're actually keys of a sub-DB. Keys and
456 * #MDB_DUPSORT data items must fit on a node in a regular page.
458 #ifndef MDB_MAXKEYSIZE
459 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
462 /** The maximum size of a key we can write to the environment. */
464 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
466 #define ENV_MAXKEY(env) ((env)->me_maxkey)
469 /** @brief The maximum size of a data item.
471 * We only store a 32 bit value for node sizes.
473 #define MAXDATASIZE 0xffffffffUL
476 /** Key size which fits in a #DKBUF.
479 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
482 * This is used for printing a hex dump of a key's contents.
484 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
485 /** Display a key in hex.
487 * Invoke a function to display a key in hex.
489 #define DKEY(x) mdb_dkey(x, kbuf)
495 /** An invalid page number.
496 * Mainly used to denote an empty tree.
498 #define P_INVALID (~(pgno_t)0)
500 /** Test if the flags \b f are set in a flag word \b w. */
501 #define F_ISSET(w, f) (((w) & (f)) == (f))
503 /** Round \b n up to an even number. */
504 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
506 /** Used for offsets within a single page.
507 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
510 typedef uint16_t indx_t;
512 /** Default size of memory map.
513 * This is certainly too small for any actual applications. Apps should always set
514 * the size explicitly using #mdb_env_set_mapsize().
516 #define DEFAULT_MAPSIZE 1048576
518 /** @defgroup readers Reader Lock Table
519 * Readers don't acquire any locks for their data access. Instead, they
520 * simply record their transaction ID in the reader table. The reader
521 * mutex is needed just to find an empty slot in the reader table. The
522 * slot's address is saved in thread-specific data so that subsequent read
523 * transactions started by the same thread need no further locking to proceed.
525 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
527 * No reader table is used if the database is on a read-only filesystem, or
528 * if #MDB_NOLOCK is set.
530 * Since the database uses multi-version concurrency control, readers don't
531 * actually need any locking. This table is used to keep track of which
532 * readers are using data from which old transactions, so that we'll know
533 * when a particular old transaction is no longer in use. Old transactions
534 * that have discarded any data pages can then have those pages reclaimed
535 * for use by a later write transaction.
537 * The lock table is constructed such that reader slots are aligned with the
538 * processor's cache line size. Any slot is only ever used by one thread.
539 * This alignment guarantees that there will be no contention or cache
540 * thrashing as threads update their own slot info, and also eliminates
541 * any need for locking when accessing a slot.
543 * A writer thread will scan every slot in the table to determine the oldest
544 * outstanding reader transaction. Any freed pages older than this will be
545 * reclaimed by the writer. The writer doesn't use any locks when scanning
546 * this table. This means that there's no guarantee that the writer will
547 * see the most up-to-date reader info, but that's not required for correct
548 * operation - all we need is to know the upper bound on the oldest reader,
549 * we don't care at all about the newest reader. So the only consequence of
550 * reading stale information here is that old pages might hang around a
551 * while longer before being reclaimed. That's actually good anyway, because
552 * the longer we delay reclaiming old pages, the more likely it is that a
553 * string of contiguous pages can be found after coalescing old pages from
554 * many old transactions together.
557 /** Number of slots in the reader table.
558 * This value was chosen somewhat arbitrarily. 126 readers plus a
559 * couple mutexes fit exactly into 8KB on my development machine.
560 * Applications should set the table size using #mdb_env_set_maxreaders().
562 #define DEFAULT_READERS 126
564 /** The size of a CPU cache line in bytes. We want our lock structures
565 * aligned to this size to avoid false cache line sharing in the
567 * This value works for most CPUs. For Itanium this should be 128.
573 /** The information we store in a single slot of the reader table.
574 * In addition to a transaction ID, we also record the process and
575 * thread ID that owns a slot, so that we can detect stale information,
576 * e.g. threads or processes that went away without cleaning up.
577 * @note We currently don't check for stale records. We simply re-init
578 * the table when we know that we're the only process opening the
581 typedef struct MDB_rxbody {
582 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
583 * Multiple readers that start at the same time will probably have the
584 * same ID here. Again, it's not important to exclude them from
585 * anything; all we need to know is which version of the DB they
586 * started from so we can avoid overwriting any data used in that
587 * particular version.
590 /** The process ID of the process owning this reader txn. */
592 /** The thread ID of the thread owning this txn. */
596 /** The actual reader record, with cacheline padding. */
597 typedef struct MDB_reader {
600 /** shorthand for mrb_txnid */
601 #define mr_txnid mru.mrx.mrb_txnid
602 #define mr_pid mru.mrx.mrb_pid
603 #define mr_tid mru.mrx.mrb_tid
604 /** cache line alignment */
605 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
609 /** The header for the reader table.
610 * The table resides in a memory-mapped file. (This is a different file
611 * than is used for the main database.)
613 * For POSIX the actual mutexes reside in the shared memory of this
614 * mapped file. On Windows, mutexes are named objects allocated by the
615 * kernel; we store the mutex names in this mapped file so that other
616 * processes can grab them. This same approach is also used on
617 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
618 * process-shared POSIX mutexes. For these cases where a named object
619 * is used, the object name is derived from a 64 bit FNV hash of the
620 * environment pathname. As such, naming collisions are extremely
621 * unlikely. If a collision occurs, the results are unpredictable.
623 typedef struct MDB_txbody {
624 /** Stamp identifying this as an LMDB file. It must be set
627 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
629 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
630 char mtb_rmname[MNAME_LEN];
632 /** Mutex protecting access to this table.
633 * This is the reader lock that #LOCK_MUTEX_R acquires.
635 pthread_mutex_t mtb_mutex;
637 /** The ID of the last transaction committed to the database.
638 * This is recorded here only for convenience; the value can always
639 * be determined by reading the main database meta pages.
642 /** The number of slots that have been used in the reader table.
643 * This always records the maximum count, it is not decremented
644 * when readers release their slots.
646 unsigned mtb_numreaders;
649 /** The actual reader table definition. */
650 typedef struct MDB_txninfo {
653 #define mti_magic mt1.mtb.mtb_magic
654 #define mti_format mt1.mtb.mtb_format
655 #define mti_mutex mt1.mtb.mtb_mutex
656 #define mti_rmname mt1.mtb.mtb_rmname
657 #define mti_txnid mt1.mtb.mtb_txnid
658 #define mti_numreaders mt1.mtb.mtb_numreaders
659 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
662 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
663 char mt2_wmname[MNAME_LEN];
664 #define mti_wmname mt2.mt2_wmname
666 pthread_mutex_t mt2_wmutex;
667 #define mti_wmutex mt2.mt2_wmutex
669 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
671 MDB_reader mti_readers[1];
674 /** Lockfile format signature: version, features and field layout */
675 #define MDB_LOCK_FORMAT \
677 ((MDB_LOCK_VERSION) \
678 /* Flags which describe functionality */ \
679 + (((MDB_PIDLOCK) != 0) << 16)))
682 /** Common header for all page types.
683 * Overflow records occupy a number of contiguous pages with no
684 * headers on any page after the first.
686 typedef struct MDB_page {
687 #define mp_pgno mp_p.p_pgno
688 #define mp_next mp_p.p_next
690 pgno_t p_pgno; /**< page number */
691 struct MDB_page *p_next; /**< for in-memory list of freed pages */
694 /** @defgroup mdb_page Page Flags
696 * Flags for the page headers.
699 #define P_BRANCH 0x01 /**< branch page */
700 #define P_LEAF 0x02 /**< leaf page */
701 #define P_OVERFLOW 0x04 /**< overflow page */
702 #define P_META 0x08 /**< meta page */
703 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
704 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
705 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
706 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
707 #define P_KEEP 0x8000 /**< leave this page alone during spill */
709 uint16_t mp_flags; /**< @ref mdb_page */
710 #define mp_lower mp_pb.pb.pb_lower
711 #define mp_upper mp_pb.pb.pb_upper
712 #define mp_pages mp_pb.pb_pages
715 indx_t pb_lower; /**< lower bound of free space */
716 indx_t pb_upper; /**< upper bound of free space */
718 uint32_t pb_pages; /**< number of overflow pages */
720 indx_t mp_ptrs[1]; /**< dynamic size */
723 /** Size of the page header, excluding dynamic data at the end */
724 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
726 /** Address of first usable data byte in a page, after the header */
727 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
729 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
730 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
732 /** Number of nodes on a page */
733 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
735 /** The amount of space remaining in the page */
736 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
738 /** The percentage of space used in the page, in tenths of a percent. */
739 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
740 ((env)->me_psize - PAGEHDRSZ))
741 /** The minimum page fill factor, in tenths of a percent.
742 * Pages emptier than this are candidates for merging.
744 #define FILL_THRESHOLD 250
746 /** Test if a page is a leaf page */
747 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
748 /** Test if a page is a LEAF2 page */
749 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
750 /** Test if a page is a branch page */
751 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
752 /** Test if a page is an overflow page */
753 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
754 /** Test if a page is a sub page */
755 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
757 /** The number of overflow pages needed to store the given size. */
758 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
760 /** Link in #MDB_txn.%mt_loose_pgs list */
761 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
763 /** Header for a single key/data pair within a page.
764 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
765 * We guarantee 2-byte alignment for 'MDB_node's.
767 typedef struct MDB_node {
768 /** lo and hi are used for data size on leaf nodes and for
769 * child pgno on branch nodes. On 64 bit platforms, flags
770 * is also used for pgno. (Branch nodes have no flags).
771 * They are in host byte order in case that lets some
772 * accesses be optimized into a 32-bit word access.
774 #if BYTE_ORDER == LITTLE_ENDIAN
775 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
777 unsigned short mn_hi, mn_lo;
779 /** @defgroup mdb_node Node Flags
781 * Flags for node headers.
784 #define F_BIGDATA 0x01 /**< data put on overflow page */
785 #define F_SUBDATA 0x02 /**< data is a sub-database */
786 #define F_DUPDATA 0x04 /**< data has duplicates */
788 /** valid flags for #mdb_node_add() */
789 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
792 unsigned short mn_flags; /**< @ref mdb_node */
793 unsigned short mn_ksize; /**< key size */
794 char mn_data[1]; /**< key and data are appended here */
797 /** Size of the node header, excluding dynamic data at the end */
798 #define NODESIZE offsetof(MDB_node, mn_data)
800 /** Bit position of top word in page number, for shifting mn_flags */
801 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
803 /** Size of a node in a branch page with a given key.
804 * This is just the node header plus the key, there is no data.
806 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
808 /** Size of a node in a leaf page with a given key and data.
809 * This is node header plus key plus data size.
811 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
813 /** Address of node \b i in page \b p */
814 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
816 /** Address of the key for the node */
817 #define NODEKEY(node) (void *)((node)->mn_data)
819 /** Address of the data for a node */
820 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
822 /** Get the page number pointed to by a branch node */
823 #define NODEPGNO(node) \
824 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
825 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
826 /** Set the page number in a branch node */
827 #define SETPGNO(node,pgno) do { \
828 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
829 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
831 /** Get the size of the data in a leaf node */
832 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
833 /** Set the size of the data for a leaf node */
834 #define SETDSZ(node,size) do { \
835 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
836 /** The size of a key in a node */
837 #define NODEKSZ(node) ((node)->mn_ksize)
839 /** Copy a page number from src to dst */
841 #define COPY_PGNO(dst,src) dst = src
843 #if SIZE_MAX > 4294967295UL
844 #define COPY_PGNO(dst,src) do { \
845 unsigned short *s, *d; \
846 s = (unsigned short *)&(src); \
847 d = (unsigned short *)&(dst); \
854 #define COPY_PGNO(dst,src) do { \
855 unsigned short *s, *d; \
856 s = (unsigned short *)&(src); \
857 d = (unsigned short *)&(dst); \
863 /** The address of a key in a LEAF2 page.
864 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
865 * There are no node headers, keys are stored contiguously.
867 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
869 /** Set the \b node's key into \b keyptr, if requested. */
870 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
871 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
873 /** Set the \b node's key into \b key. */
874 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
876 /** Information about a single database in the environment. */
877 typedef struct MDB_db {
878 uint32_t md_pad; /**< also ksize for LEAF2 pages */
879 uint16_t md_flags; /**< @ref mdb_dbi_open */
880 uint16_t md_depth; /**< depth of this tree */
881 pgno_t md_branch_pages; /**< number of internal pages */
882 pgno_t md_leaf_pages; /**< number of leaf pages */
883 pgno_t md_overflow_pages; /**< number of overflow pages */
884 size_t md_entries; /**< number of data items */
885 pgno_t md_root; /**< the root page of this tree */
888 /** mdb_dbi_open flags */
889 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
890 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
891 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
892 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
894 /** Handle for the DB used to track free pages. */
896 /** Handle for the default DB. */
899 /** Meta page content.
900 * A meta page is the start point for accessing a database snapshot.
901 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
903 typedef struct MDB_meta {
904 /** Stamp identifying this as an LMDB file. It must be set
907 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
909 void *mm_address; /**< address for fixed mapping */
910 size_t mm_mapsize; /**< size of mmap region */
911 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
912 /** The size of pages used in this DB */
913 #define mm_psize mm_dbs[0].md_pad
914 /** Any persistent environment flags. @ref mdb_env */
915 #define mm_flags mm_dbs[0].md_flags
916 pgno_t mm_last_pg; /**< last used page in file */
917 txnid_t mm_txnid; /**< txnid that committed this page */
920 /** Buffer for a stack-allocated meta page.
921 * The members define size and alignment, and silence type
922 * aliasing warnings. They are not used directly; that could
923 * mean incorrectly using several union members in parallel.
925 typedef union MDB_metabuf {
928 char mm_pad[PAGEHDRSZ];
933 /** Auxiliary DB info.
934 * The information here is mostly static/read-only. There is
935 * only a single copy of this record in the environment.
937 typedef struct MDB_dbx {
938 MDB_val md_name; /**< name of the database */
939 MDB_cmp_func *md_cmp; /**< function for comparing keys */
940 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
941 MDB_rel_func *md_rel; /**< user relocate function */
942 void *md_relctx; /**< user-provided context for md_rel */
945 /** A database transaction.
946 * Every operation requires a transaction handle.
949 MDB_txn *mt_parent; /**< parent of a nested txn */
950 MDB_txn *mt_child; /**< nested txn under this txn */
951 pgno_t mt_next_pgno; /**< next unallocated page */
952 /** The ID of this transaction. IDs are integers incrementing from 1.
953 * Only committed write transactions increment the ID. If a transaction
954 * aborts, the ID may be re-used by the next writer.
957 MDB_env *mt_env; /**< the DB environment */
958 /** The list of pages that became unused during this transaction.
961 /** The list of loose pages that became unused and may be reused
962 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
964 MDB_page *mt_loose_pgs;
965 /* #Number of loose pages (#mt_loose_pgs) */
967 /** The sorted list of dirty pages we temporarily wrote to disk
968 * because the dirty list was full. page numbers in here are
969 * shifted left by 1, deleted slots have the LSB set.
971 MDB_IDL mt_spill_pgs;
973 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
975 /** For read txns: This thread/txn's reader table slot, or NULL. */
978 /** Array of records for each DB known in the environment. */
980 /** Array of MDB_db records for each known DB */
982 /** Array of sequence numbers for each DB handle */
983 unsigned int *mt_dbiseqs;
984 /** @defgroup mt_dbflag Transaction DB Flags
988 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
989 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
990 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
991 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
993 /** In write txns, array of cursors for each DB */
994 MDB_cursor **mt_cursors;
995 /** Array of flags for each DB */
996 unsigned char *mt_dbflags;
997 /** Number of DB records in use. This number only ever increments;
998 * we don't decrement it when individual DB handles are closed.
1002 /** @defgroup mdb_txn Transaction Flags
1006 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
1007 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1008 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1009 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1011 unsigned int mt_flags; /**< @ref mdb_txn */
1012 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1013 * Includes ancestor txns' dirty pages not hidden by other txns'
1014 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1015 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1017 unsigned int mt_dirty_room;
1020 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1021 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1022 * raise this on a 64 bit machine.
1024 #define CURSOR_STACK 32
1028 /** Cursors are used for all DB operations.
1029 * A cursor holds a path of (page pointer, key index) from the DB
1030 * root to a position in the DB, plus other state. #MDB_DUPSORT
1031 * cursors include an xcursor to the current data item. Write txns
1032 * track their cursors and keep them up to date when data moves.
1033 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1034 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1037 /** Next cursor on this DB in this txn */
1038 MDB_cursor *mc_next;
1039 /** Backup of the original cursor if this cursor is a shadow */
1040 MDB_cursor *mc_backup;
1041 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1042 struct MDB_xcursor *mc_xcursor;
1043 /** The transaction that owns this cursor */
1045 /** The database handle this cursor operates on */
1047 /** The database record for this cursor */
1049 /** The database auxiliary record for this cursor */
1051 /** The @ref mt_dbflag for this database */
1052 unsigned char *mc_dbflag;
1053 unsigned short mc_snum; /**< number of pushed pages */
1054 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1055 /** @defgroup mdb_cursor Cursor Flags
1057 * Cursor state flags.
1060 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1061 #define C_EOF 0x02 /**< No more data */
1062 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1063 #define C_DEL 0x08 /**< last op was a cursor_del */
1064 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1065 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1067 unsigned int mc_flags; /**< @ref mdb_cursor */
1068 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1069 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1072 /** Context for sorted-dup records.
1073 * We could have gone to a fully recursive design, with arbitrarily
1074 * deep nesting of sub-databases. But for now we only handle these
1075 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1077 typedef struct MDB_xcursor {
1078 /** A sub-cursor for traversing the Dup DB */
1079 MDB_cursor mx_cursor;
1080 /** The database record for this Dup DB */
1082 /** The auxiliary DB record for this Dup DB */
1084 /** The @ref mt_dbflag for this Dup DB */
1085 unsigned char mx_dbflag;
1088 /** State of FreeDB old pages, stored in the MDB_env */
1089 typedef struct MDB_pgstate {
1090 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1091 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1094 /** The database environment. */
1096 HANDLE me_fd; /**< The main data file */
1097 HANDLE me_lfd; /**< The lock file */
1098 HANDLE me_mfd; /**< just for writing the meta pages */
1099 /** Failed to update the meta page. Probably an I/O error. */
1100 #define MDB_FATAL_ERROR 0x80000000U
1101 /** Some fields are initialized. */
1102 #define MDB_ENV_ACTIVE 0x20000000U
1103 /** me_txkey is set */
1104 #define MDB_ENV_TXKEY 0x10000000U
1105 uint32_t me_flags; /**< @ref mdb_env */
1106 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1107 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1108 unsigned int me_maxreaders; /**< size of the reader table */
1109 unsigned int me_numreaders; /**< max numreaders set by this env */
1110 MDB_dbi me_numdbs; /**< number of DBs opened */
1111 MDB_dbi me_maxdbs; /**< size of the DB table */
1112 MDB_PID_T me_pid; /**< process ID of this env */
1113 char *me_path; /**< path to the DB files */
1114 char *me_map; /**< the memory map of the data file */
1115 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1116 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1117 void *me_pbuf; /**< scratch area for DUPSORT put() */
1118 MDB_txn *me_txn; /**< current write transaction */
1119 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1120 size_t me_mapsize; /**< size of the data memory map */
1121 off_t me_size; /**< current file size */
1122 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1123 MDB_dbx *me_dbxs; /**< array of static DB info */
1124 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1125 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1126 pthread_key_t me_txkey; /**< thread-key for readers */
1127 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1128 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1129 # define me_pglast me_pgstate.mf_pglast
1130 # define me_pghead me_pgstate.mf_pghead
1131 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1132 /** IDL of pages that became unused in a write txn */
1133 MDB_IDL me_free_pgs;
1134 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1135 MDB_ID2L me_dirty_list;
1136 /** Max number of freelist items that can fit in a single overflow page */
1138 /** Max size of a node on a page */
1139 unsigned int me_nodemax;
1140 #if !(MDB_MAXKEYSIZE)
1141 unsigned int me_maxkey; /**< max size of a key */
1143 int me_live_reader; /**< have liveness lock in reader table */
1145 int me_pidquery; /**< Used in OpenProcess */
1146 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1148 #elif defined(MDB_USE_POSIX_SEM)
1149 sem_t *me_rmutex; /* Shared mutexes are not supported */
1152 #ifdef BROKEN_FDATASYMC
1153 int me_fsynconly; /**< fdatasync is unreliable */
1155 void *me_userctx; /**< User-settable context */
1156 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1159 /** Nested transaction */
1160 typedef struct MDB_ntxn {
1161 MDB_txn mnt_txn; /**< the transaction */
1162 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1165 /** max number of pages to commit in one writev() call */
1166 #define MDB_COMMIT_PAGES 64
1167 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1168 #undef MDB_COMMIT_PAGES
1169 #define MDB_COMMIT_PAGES IOV_MAX
1172 /** max bytes to write in one call */
1173 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1175 /** Check \b txn and \b dbi arguments to a function */
1176 #define TXN_DBI_EXIST(txn, dbi) \
1177 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1179 /** Check for misused \b dbi handles */
1180 #define TXN_DBI_CHANGED(txn, dbi) \
1181 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1183 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1184 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1185 static int mdb_page_touch(MDB_cursor *mc);
1187 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1188 static int mdb_page_search_root(MDB_cursor *mc,
1189 MDB_val *key, int modify);
1190 #define MDB_PS_MODIFY 1
1191 #define MDB_PS_ROOTONLY 2
1192 #define MDB_PS_FIRST 4
1193 #define MDB_PS_LAST 8
1194 static int mdb_page_search(MDB_cursor *mc,
1195 MDB_val *key, int flags);
1196 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1198 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1199 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1200 pgno_t newpgno, unsigned int nflags);
1202 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1203 static int mdb_env_pick_meta(const MDB_env *env);
1204 static int mdb_env_write_meta(MDB_txn *txn);
1205 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1206 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1208 static void mdb_env_close0(MDB_env *env, int excl);
1210 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1211 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1212 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1213 static void mdb_node_del(MDB_cursor *mc, int ksize);
1214 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1215 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1216 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1217 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1218 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1220 static int mdb_rebalance(MDB_cursor *mc);
1221 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1223 static void mdb_cursor_pop(MDB_cursor *mc);
1224 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1226 static int mdb_cursor_del0(MDB_cursor *mc);
1227 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1228 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1229 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1230 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1231 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1233 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1234 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1236 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1237 static void mdb_xcursor_init0(MDB_cursor *mc);
1238 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1240 static int mdb_drop0(MDB_cursor *mc, int subs);
1241 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1244 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1248 static SECURITY_DESCRIPTOR mdb_null_sd;
1249 static SECURITY_ATTRIBUTES mdb_all_sa;
1250 static int mdb_sec_inited;
1253 /** Return the library version info. */
1255 mdb_version(int *major, int *minor, int *patch)
1257 if (major) *major = MDB_VERSION_MAJOR;
1258 if (minor) *minor = MDB_VERSION_MINOR;
1259 if (patch) *patch = MDB_VERSION_PATCH;
1260 return MDB_VERSION_STRING;
1263 /** Table of descriptions for LMDB @ref errors */
1264 static char *const mdb_errstr[] = {
1265 "MDB_KEYEXIST: Key/data pair already exists",
1266 "MDB_NOTFOUND: No matching key/data pair found",
1267 "MDB_PAGE_NOTFOUND: Requested page not found",
1268 "MDB_CORRUPTED: Located page was wrong type",
1269 "MDB_PANIC: Update of meta page failed",
1270 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1271 "MDB_INVALID: File is not an LMDB file",
1272 "MDB_MAP_FULL: Environment mapsize limit reached",
1273 "MDB_DBS_FULL: Environment maxdbs limit reached",
1274 "MDB_READERS_FULL: Environment maxreaders limit reached",
1275 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1276 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1277 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1278 "MDB_PAGE_FULL: Internal error - page has no more space",
1279 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1280 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1281 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1282 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1283 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1284 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1288 mdb_strerror(int err)
1291 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1292 * This works as long as no function between the call to mdb_strerror
1293 * and the actual use of the message uses more than 4K of stack.
1296 char buf[1024], *ptr = buf;
1300 return ("Successful return: 0");
1302 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1303 i = err - MDB_KEYEXIST;
1304 return mdb_errstr[i];
1308 /* These are the C-runtime error codes we use. The comment indicates
1309 * their numeric value, and the Win32 error they would correspond to
1310 * if the error actually came from a Win32 API. A major mess, we should
1311 * have used LMDB-specific error codes for everything.
1314 case ENOENT: /* 2, FILE_NOT_FOUND */
1315 case EIO: /* 5, ACCESS_DENIED */
1316 case ENOMEM: /* 12, INVALID_ACCESS */
1317 case EACCES: /* 13, INVALID_DATA */
1318 case EBUSY: /* 16, CURRENT_DIRECTORY */
1319 case EINVAL: /* 22, BAD_COMMAND */
1320 case ENOSPC: /* 28, OUT_OF_PAPER */
1321 return strerror(err);
1326 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1327 FORMAT_MESSAGE_IGNORE_INSERTS,
1328 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1331 return strerror(err);
1335 /** assert(3) variant in cursor context */
1336 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1337 /** assert(3) variant in transaction context */
1338 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1339 /** assert(3) variant in environment context */
1340 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1343 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1344 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1347 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1348 const char *func, const char *file, int line)
1351 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1352 file, line, expr_txt, func);
1353 if (env->me_assert_func)
1354 env->me_assert_func(env, buf);
1355 fprintf(stderr, "%s\n", buf);
1359 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1363 /** Return the page number of \b mp which may be sub-page, for debug output */
1365 mdb_dbg_pgno(MDB_page *mp)
1368 COPY_PGNO(ret, mp->mp_pgno);
1372 /** Display a key in hexadecimal and return the address of the result.
1373 * @param[in] key the key to display
1374 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1375 * @return The key in hexadecimal form.
1378 mdb_dkey(MDB_val *key, char *buf)
1381 unsigned char *c = key->mv_data;
1387 if (key->mv_size > DKBUF_MAXKEYSIZE)
1388 return "MDB_MAXKEYSIZE";
1389 /* may want to make this a dynamic check: if the key is mostly
1390 * printable characters, print it as-is instead of converting to hex.
1394 for (i=0; i<key->mv_size; i++)
1395 ptr += sprintf(ptr, "%02x", *c++);
1397 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1403 mdb_leafnode_type(MDB_node *n)
1405 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1406 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1407 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1410 /** Display all the keys in the page. */
1412 mdb_page_list(MDB_page *mp)
1414 pgno_t pgno = mdb_dbg_pgno(mp);
1415 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1417 unsigned int i, nkeys, nsize, total = 0;
1421 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1422 case P_BRANCH: type = "Branch page"; break;
1423 case P_LEAF: type = "Leaf page"; break;
1424 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1425 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1426 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1428 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1429 pgno, mp->mp_pages, state);
1432 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1433 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1436 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1440 nkeys = NUMKEYS(mp);
1441 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1443 for (i=0; i<nkeys; i++) {
1444 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1445 key.mv_size = nsize = mp->mp_pad;
1446 key.mv_data = LEAF2KEY(mp, i, nsize);
1448 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1451 node = NODEPTR(mp, i);
1452 key.mv_size = node->mn_ksize;
1453 key.mv_data = node->mn_data;
1454 nsize = NODESIZE + key.mv_size;
1455 if (IS_BRANCH(mp)) {
1456 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1460 if (F_ISSET(node->mn_flags, F_BIGDATA))
1461 nsize += sizeof(pgno_t);
1463 nsize += NODEDSZ(node);
1465 nsize += sizeof(indx_t);
1466 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1467 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1469 total = EVEN(total);
1471 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1472 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1476 mdb_cursor_chk(MDB_cursor *mc)
1482 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1483 for (i=0; i<mc->mc_top; i++) {
1485 node = NODEPTR(mp, mc->mc_ki[i]);
1486 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1489 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1495 /** Count all the pages in each DB and in the freelist
1496 * and make sure it matches the actual number of pages
1498 * All named DBs must be open for a correct count.
1500 static void mdb_audit(MDB_txn *txn)
1504 MDB_ID freecount, count;
1509 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1510 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1511 freecount += *(MDB_ID *)data.mv_data;
1512 mdb_tassert(txn, rc == MDB_NOTFOUND);
1515 for (i = 0; i<txn->mt_numdbs; i++) {
1517 if (!(txn->mt_dbflags[i] & DB_VALID))
1519 mdb_cursor_init(&mc, txn, i, &mx);
1520 if (txn->mt_dbs[i].md_root == P_INVALID)
1522 count += txn->mt_dbs[i].md_branch_pages +
1523 txn->mt_dbs[i].md_leaf_pages +
1524 txn->mt_dbs[i].md_overflow_pages;
1525 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1526 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1527 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1530 mp = mc.mc_pg[mc.mc_top];
1531 for (j=0; j<NUMKEYS(mp); j++) {
1532 MDB_node *leaf = NODEPTR(mp, j);
1533 if (leaf->mn_flags & F_SUBDATA) {
1535 memcpy(&db, NODEDATA(leaf), sizeof(db));
1536 count += db.md_branch_pages + db.md_leaf_pages +
1537 db.md_overflow_pages;
1541 mdb_tassert(txn, rc == MDB_NOTFOUND);
1544 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1545 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1546 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1552 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1554 return txn->mt_dbxs[dbi].md_cmp(a, b);
1558 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1560 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1563 /** Allocate memory for a page.
1564 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1567 mdb_page_malloc(MDB_txn *txn, unsigned num)
1569 MDB_env *env = txn->mt_env;
1570 MDB_page *ret = env->me_dpages;
1571 size_t psize = env->me_psize, sz = psize, off;
1572 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1573 * For a single page alloc, we init everything after the page header.
1574 * For multi-page, we init the final page; if the caller needed that
1575 * many pages they will be filling in at least up to the last page.
1579 VGMEMP_ALLOC(env, ret, sz);
1580 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1581 env->me_dpages = ret->mp_next;
1584 psize -= off = PAGEHDRSZ;
1589 if ((ret = malloc(sz)) != NULL) {
1590 VGMEMP_ALLOC(env, ret, sz);
1591 if (!(env->me_flags & MDB_NOMEMINIT)) {
1592 memset((char *)ret + off, 0, psize);
1596 txn->mt_flags |= MDB_TXN_ERROR;
1600 /** Free a single page.
1601 * Saves single pages to a list, for future reuse.
1602 * (This is not used for multi-page overflow pages.)
1605 mdb_page_free(MDB_env *env, MDB_page *mp)
1607 mp->mp_next = env->me_dpages;
1608 VGMEMP_FREE(env, mp);
1609 env->me_dpages = mp;
1612 /** Free a dirty page */
1614 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1616 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1617 mdb_page_free(env, dp);
1619 /* large pages just get freed directly */
1620 VGMEMP_FREE(env, dp);
1625 /** Return all dirty pages to dpage list */
1627 mdb_dlist_free(MDB_txn *txn)
1629 MDB_env *env = txn->mt_env;
1630 MDB_ID2L dl = txn->mt_u.dirty_list;
1631 unsigned i, n = dl[0].mid;
1633 for (i = 1; i <= n; i++) {
1634 mdb_dpage_free(env, dl[i].mptr);
1639 /** Loosen or free a single page.
1640 * Saves single pages to a list for future reuse
1641 * in this same txn. It has been pulled from the freeDB
1642 * and already resides on the dirty list, but has been
1643 * deleted. Use these pages first before pulling again
1646 * If the page wasn't dirtied in this txn, just add it
1647 * to this txn's free list.
1650 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1653 pgno_t pgno = mp->mp_pgno;
1654 MDB_txn *txn = mc->mc_txn;
1656 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1657 if (txn->mt_parent) {
1658 MDB_ID2 *dl = txn->mt_u.dirty_list;
1659 /* If txn has a parent, make sure the page is in our
1663 unsigned x = mdb_mid2l_search(dl, pgno);
1664 if (x <= dl[0].mid && dl[x].mid == pgno) {
1665 if (mp != dl[x].mptr) { /* bad cursor? */
1666 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1667 txn->mt_flags |= MDB_TXN_ERROR;
1668 return MDB_CORRUPTED;
1675 /* no parent txn, so it's just ours */
1680 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1682 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1683 txn->mt_loose_pgs = mp;
1684 txn->mt_loose_count++;
1685 mp->mp_flags |= P_LOOSE;
1687 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1695 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1696 * @param[in] mc A cursor handle for the current operation.
1697 * @param[in] pflags Flags of the pages to update:
1698 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1699 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1700 * @return 0 on success, non-zero on failure.
1703 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1705 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1706 MDB_txn *txn = mc->mc_txn;
1712 int rc = MDB_SUCCESS, level;
1714 /* Mark pages seen by cursors */
1715 if (mc->mc_flags & C_UNTRACK)
1716 mc = NULL; /* will find mc in mt_cursors */
1717 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1718 for (; mc; mc=mc->mc_next) {
1719 if (!(mc->mc_flags & C_INITIALIZED))
1721 for (m3 = mc;; m3 = &mx->mx_cursor) {
1723 for (j=0; j<m3->mc_snum; j++) {
1725 if ((mp->mp_flags & Mask) == pflags)
1726 mp->mp_flags ^= P_KEEP;
1728 mx = m3->mc_xcursor;
1729 /* Proceed to mx if it is at a sub-database */
1730 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1732 if (! (mp && (mp->mp_flags & P_LEAF)))
1734 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1735 if (!(leaf->mn_flags & F_SUBDATA))
1744 /* Mark dirty root pages */
1745 for (i=0; i<txn->mt_numdbs; i++) {
1746 if (txn->mt_dbflags[i] & DB_DIRTY) {
1747 pgno_t pgno = txn->mt_dbs[i].md_root;
1748 if (pgno == P_INVALID)
1750 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1752 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1753 dp->mp_flags ^= P_KEEP;
1761 static int mdb_page_flush(MDB_txn *txn, int keep);
1763 /** Spill pages from the dirty list back to disk.
1764 * This is intended to prevent running into #MDB_TXN_FULL situations,
1765 * but note that they may still occur in a few cases:
1766 * 1) our estimate of the txn size could be too small. Currently this
1767 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1768 * 2) child txns may run out of space if their parents dirtied a
1769 * lot of pages and never spilled them. TODO: we probably should do
1770 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1771 * the parent's dirty_room is below a given threshold.
1773 * Otherwise, if not using nested txns, it is expected that apps will
1774 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1775 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1776 * If the txn never references them again, they can be left alone.
1777 * If the txn only reads them, they can be used without any fuss.
1778 * If the txn writes them again, they can be dirtied immediately without
1779 * going thru all of the work of #mdb_page_touch(). Such references are
1780 * handled by #mdb_page_unspill().
1782 * Also note, we never spill DB root pages, nor pages of active cursors,
1783 * because we'll need these back again soon anyway. And in nested txns,
1784 * we can't spill a page in a child txn if it was already spilled in a
1785 * parent txn. That would alter the parent txns' data even though
1786 * the child hasn't committed yet, and we'd have no way to undo it if
1787 * the child aborted.
1789 * @param[in] m0 cursor A cursor handle identifying the transaction and
1790 * database for which we are checking space.
1791 * @param[in] key For a put operation, the key being stored.
1792 * @param[in] data For a put operation, the data being stored.
1793 * @return 0 on success, non-zero on failure.
1796 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1798 MDB_txn *txn = m0->mc_txn;
1800 MDB_ID2L dl = txn->mt_u.dirty_list;
1801 unsigned int i, j, need;
1804 if (m0->mc_flags & C_SUB)
1807 /* Estimate how much space this op will take */
1808 i = m0->mc_db->md_depth;
1809 /* Named DBs also dirty the main DB */
1810 if (m0->mc_dbi > MAIN_DBI)
1811 i += txn->mt_dbs[MAIN_DBI].md_depth;
1812 /* For puts, roughly factor in the key+data size */
1814 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1815 i += i; /* double it for good measure */
1818 if (txn->mt_dirty_room > i)
1821 if (!txn->mt_spill_pgs) {
1822 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1823 if (!txn->mt_spill_pgs)
1826 /* purge deleted slots */
1827 MDB_IDL sl = txn->mt_spill_pgs;
1828 unsigned int num = sl[0];
1830 for (i=1; i<=num; i++) {
1837 /* Preserve pages which may soon be dirtied again */
1838 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1841 /* Less aggressive spill - we originally spilled the entire dirty list,
1842 * with a few exceptions for cursor pages and DB root pages. But this
1843 * turns out to be a lot of wasted effort because in a large txn many
1844 * of those pages will need to be used again. So now we spill only 1/8th
1845 * of the dirty pages. Testing revealed this to be a good tradeoff,
1846 * better than 1/2, 1/4, or 1/10.
1848 if (need < MDB_IDL_UM_MAX / 8)
1849 need = MDB_IDL_UM_MAX / 8;
1851 /* Save the page IDs of all the pages we're flushing */
1852 /* flush from the tail forward, this saves a lot of shifting later on. */
1853 for (i=dl[0].mid; i && need; i--) {
1854 MDB_ID pn = dl[i].mid << 1;
1856 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1858 /* Can't spill twice, make sure it's not already in a parent's
1861 if (txn->mt_parent) {
1863 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1864 if (tx2->mt_spill_pgs) {
1865 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1866 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1867 dp->mp_flags |= P_KEEP;
1875 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1879 mdb_midl_sort(txn->mt_spill_pgs);
1881 /* Flush the spilled part of dirty list */
1882 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1885 /* Reset any dirty pages we kept that page_flush didn't see */
1886 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1889 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1893 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1895 mdb_find_oldest(MDB_txn *txn)
1898 txnid_t mr, oldest = txn->mt_txnid - 1;
1899 if (txn->mt_env->me_txns) {
1900 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1901 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1912 /** Add a page to the txn's dirty list */
1914 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1917 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1919 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1920 insert = mdb_mid2l_append;
1922 insert = mdb_mid2l_insert;
1924 mid.mid = mp->mp_pgno;
1926 rc = insert(txn->mt_u.dirty_list, &mid);
1927 mdb_tassert(txn, rc == 0);
1928 txn->mt_dirty_room--;
1931 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1932 * me_pghead and mt_next_pgno.
1934 * If there are free pages available from older transactions, they
1935 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1936 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1937 * and move me_pglast to say which records were consumed. Only this
1938 * function can create me_pghead and move me_pglast/mt_next_pgno.
1939 * @param[in] mc cursor A cursor handle identifying the transaction and
1940 * database for which we are allocating.
1941 * @param[in] num the number of pages to allocate.
1942 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1943 * will always be satisfied by a single contiguous chunk of memory.
1944 * @return 0 on success, non-zero on failure.
1947 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1949 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1950 /* Get at most <Max_retries> more freeDB records once me_pghead
1951 * has enough pages. If not enough, use new pages from the map.
1952 * If <Paranoid> and mc is updating the freeDB, only get new
1953 * records if me_pghead is empty. Then the freelist cannot play
1954 * catch-up with itself by growing while trying to save it.
1956 enum { Paranoid = 1, Max_retries = 500 };
1958 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1960 int rc, retry = num * 60;
1961 MDB_txn *txn = mc->mc_txn;
1962 MDB_env *env = txn->mt_env;
1963 pgno_t pgno, *mop = env->me_pghead;
1964 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1966 txnid_t oldest = 0, last;
1971 /* If there are any loose pages, just use them */
1972 if (num == 1 && txn->mt_loose_pgs) {
1973 np = txn->mt_loose_pgs;
1974 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1975 txn->mt_loose_count--;
1976 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1984 /* If our dirty list is already full, we can't do anything */
1985 if (txn->mt_dirty_room == 0) {
1990 for (op = MDB_FIRST;; op = MDB_NEXT) {
1995 /* Seek a big enough contiguous page range. Prefer
1996 * pages at the tail, just truncating the list.
2002 if (mop[i-n2] == pgno+n2)
2009 if (op == MDB_FIRST) { /* 1st iteration */
2010 /* Prepare to fetch more and coalesce */
2011 last = env->me_pglast;
2012 oldest = env->me_pgoldest;
2013 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2016 key.mv_data = &last; /* will look up last+1 */
2017 key.mv_size = sizeof(last);
2019 if (Paranoid && mc->mc_dbi == FREE_DBI)
2022 if (Paranoid && retry < 0 && mop_len)
2026 /* Do not fetch more if the record will be too recent */
2027 if (oldest <= last) {
2029 oldest = mdb_find_oldest(txn);
2030 env->me_pgoldest = oldest;
2036 rc = mdb_cursor_get(&m2, &key, NULL, op);
2038 if (rc == MDB_NOTFOUND)
2042 last = *(txnid_t*)key.mv_data;
2043 if (oldest <= last) {
2045 oldest = mdb_find_oldest(txn);
2046 env->me_pgoldest = oldest;
2052 np = m2.mc_pg[m2.mc_top];
2053 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2054 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2057 idl = (MDB_ID *) data.mv_data;
2060 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2065 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2067 mop = env->me_pghead;
2069 env->me_pglast = last;
2071 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2072 last, txn->mt_dbs[FREE_DBI].md_root, i));
2074 DPRINTF(("IDL %"Z"u", idl[j]));
2076 /* Merge in descending sorted order */
2077 mdb_midl_xmerge(mop, idl);
2081 /* Use new pages from the map when nothing suitable in the freeDB */
2083 pgno = txn->mt_next_pgno;
2084 if (pgno + num >= env->me_maxpg) {
2085 DPUTS("DB size maxed out");
2091 if (env->me_flags & MDB_WRITEMAP) {
2092 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2094 if (!(np = mdb_page_malloc(txn, num))) {
2100 mop[0] = mop_len -= num;
2101 /* Move any stragglers down */
2102 for (j = i-num; j < mop_len; )
2103 mop[++j] = mop[++i];
2105 txn->mt_next_pgno = pgno + num;
2108 mdb_page_dirty(txn, np);
2114 txn->mt_flags |= MDB_TXN_ERROR;
2118 /** Copy the used portions of a non-overflow page.
2119 * @param[in] dst page to copy into
2120 * @param[in] src page to copy from
2121 * @param[in] psize size of a page
2124 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2126 enum { Align = sizeof(pgno_t) };
2127 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2129 /* If page isn't full, just copy the used portion. Adjust
2130 * alignment so memcpy may copy words instead of bytes.
2132 if ((unused &= -Align) && !IS_LEAF2(src)) {
2133 upper = (upper + PAGEBASE) & -Align;
2134 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2135 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2138 memcpy(dst, src, psize - unused);
2142 /** Pull a page off the txn's spill list, if present.
2143 * If a page being referenced was spilled to disk in this txn, bring
2144 * it back and make it dirty/writable again.
2145 * @param[in] txn the transaction handle.
2146 * @param[in] mp the page being referenced. It must not be dirty.
2147 * @param[out] ret the writable page, if any. ret is unchanged if
2148 * mp wasn't spilled.
2151 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2153 MDB_env *env = txn->mt_env;
2156 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2158 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2159 if (!tx2->mt_spill_pgs)
2161 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2162 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2165 if (txn->mt_dirty_room == 0)
2166 return MDB_TXN_FULL;
2167 if (IS_OVERFLOW(mp))
2171 if (env->me_flags & MDB_WRITEMAP) {
2174 np = mdb_page_malloc(txn, num);
2178 memcpy(np, mp, num * env->me_psize);
2180 mdb_page_copy(np, mp, env->me_psize);
2183 /* If in current txn, this page is no longer spilled.
2184 * If it happens to be the last page, truncate the spill list.
2185 * Otherwise mark it as deleted by setting the LSB.
2187 if (x == txn->mt_spill_pgs[0])
2188 txn->mt_spill_pgs[0]--;
2190 txn->mt_spill_pgs[x] |= 1;
2191 } /* otherwise, if belonging to a parent txn, the
2192 * page remains spilled until child commits
2195 mdb_page_dirty(txn, np);
2196 np->mp_flags |= P_DIRTY;
2204 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2205 * @param[in] mc cursor pointing to the page to be touched
2206 * @return 0 on success, non-zero on failure.
2209 mdb_page_touch(MDB_cursor *mc)
2211 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2212 MDB_txn *txn = mc->mc_txn;
2213 MDB_cursor *m2, *m3;
2217 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2218 if (txn->mt_flags & MDB_TXN_SPILLS) {
2220 rc = mdb_page_unspill(txn, mp, &np);
2226 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2227 (rc = mdb_page_alloc(mc, 1, &np)))
2230 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2231 mp->mp_pgno, pgno));
2232 mdb_cassert(mc, mp->mp_pgno != pgno);
2233 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2234 /* Update the parent page, if any, to point to the new page */
2236 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2237 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2238 SETPGNO(node, pgno);
2240 mc->mc_db->md_root = pgno;
2242 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2243 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2245 /* If txn has a parent, make sure the page is in our
2249 unsigned x = mdb_mid2l_search(dl, pgno);
2250 if (x <= dl[0].mid && dl[x].mid == pgno) {
2251 if (mp != dl[x].mptr) { /* bad cursor? */
2252 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2253 txn->mt_flags |= MDB_TXN_ERROR;
2254 return MDB_CORRUPTED;
2259 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2261 np = mdb_page_malloc(txn, 1);
2266 rc = mdb_mid2l_insert(dl, &mid);
2267 mdb_cassert(mc, rc == 0);
2272 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2274 np->mp_flags |= P_DIRTY;
2277 /* Adjust cursors pointing to mp */
2278 mc->mc_pg[mc->mc_top] = np;
2279 m2 = txn->mt_cursors[mc->mc_dbi];
2280 if (mc->mc_flags & C_SUB) {
2281 for (; m2; m2=m2->mc_next) {
2282 m3 = &m2->mc_xcursor->mx_cursor;
2283 if (m3->mc_snum < mc->mc_snum) continue;
2284 if (m3->mc_pg[mc->mc_top] == mp)
2285 m3->mc_pg[mc->mc_top] = np;
2288 for (; m2; m2=m2->mc_next) {
2289 if (m2->mc_snum < mc->mc_snum) continue;
2290 if (m2->mc_pg[mc->mc_top] == mp) {
2291 m2->mc_pg[mc->mc_top] = np;
2292 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2294 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2296 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2297 if (!(leaf->mn_flags & F_SUBDATA))
2298 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2306 txn->mt_flags |= MDB_TXN_ERROR;
2311 mdb_env_sync(MDB_env *env, int force)
2314 if (env->me_flags & MDB_RDONLY)
2316 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2317 if (env->me_flags & MDB_WRITEMAP) {
2318 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2319 ? MS_ASYNC : MS_SYNC;
2320 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2323 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2327 #ifdef BROKEN_FDATASYNC
2328 if (env->me_fsynconly) {
2329 if (fsync(env->me_fd))
2333 if (MDB_FDATASYNC(env->me_fd))
2340 /** Back up parent txn's cursors, then grab the originals for tracking */
2342 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2344 MDB_cursor *mc, *bk;
2349 for (i = src->mt_numdbs; --i >= 0; ) {
2350 if ((mc = src->mt_cursors[i]) != NULL) {
2351 size = sizeof(MDB_cursor);
2353 size += sizeof(MDB_xcursor);
2354 for (; mc; mc = bk->mc_next) {
2360 mc->mc_db = &dst->mt_dbs[i];
2361 /* Kill pointers into src - and dst to reduce abuse: The
2362 * user may not use mc until dst ends. Otherwise we'd...
2364 mc->mc_txn = NULL; /* ...set this to dst */
2365 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2366 if ((mx = mc->mc_xcursor) != NULL) {
2367 *(MDB_xcursor *)(bk+1) = *mx;
2368 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2370 mc->mc_next = dst->mt_cursors[i];
2371 dst->mt_cursors[i] = mc;
2378 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2379 * @param[in] txn the transaction handle.
2380 * @param[in] merge true to keep changes to parent cursors, false to revert.
2381 * @return 0 on success, non-zero on failure.
2384 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2386 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2390 for (i = txn->mt_numdbs; --i >= 0; ) {
2391 for (mc = cursors[i]; mc; mc = next) {
2393 if ((bk = mc->mc_backup) != NULL) {
2395 /* Commit changes to parent txn */
2396 mc->mc_next = bk->mc_next;
2397 mc->mc_backup = bk->mc_backup;
2398 mc->mc_txn = bk->mc_txn;
2399 mc->mc_db = bk->mc_db;
2400 mc->mc_dbflag = bk->mc_dbflag;
2401 if ((mx = mc->mc_xcursor) != NULL)
2402 mx->mx_cursor.mc_txn = bk->mc_txn;
2404 /* Abort nested txn */
2406 if ((mx = mc->mc_xcursor) != NULL)
2407 *mx = *(MDB_xcursor *)(bk+1);
2411 /* Only malloced cursors are permanently tracked. */
2419 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2422 mdb_txn_reset0(MDB_txn *txn, const char *act);
2424 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2430 Pidset = F_SETLK, Pidcheck = F_GETLK
2434 /** Set or check a pid lock. Set returns 0 on success.
2435 * Check returns 0 if the process is certainly dead, nonzero if it may
2436 * be alive (the lock exists or an error happened so we do not know).
2438 * On Windows Pidset is a no-op, we merely check for the existence
2439 * of the process with the given pid. On POSIX we use a single byte
2440 * lock on the lockfile, set at an offset equal to the pid.
2443 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2445 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2448 if (op == Pidcheck) {
2449 h = OpenProcess(env->me_pidquery, FALSE, pid);
2450 /* No documented "no such process" code, but other program use this: */
2452 return ErrCode() != ERROR_INVALID_PARAMETER;
2453 /* A process exists until all handles to it close. Has it exited? */
2454 ret = WaitForSingleObject(h, 0) != 0;
2461 struct flock lock_info;
2462 memset(&lock_info, 0, sizeof(lock_info));
2463 lock_info.l_type = F_WRLCK;
2464 lock_info.l_whence = SEEK_SET;
2465 lock_info.l_start = pid;
2466 lock_info.l_len = 1;
2467 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2468 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2470 } else if ((rc = ErrCode()) == EINTR) {
2478 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2479 * @param[in] txn the transaction handle to initialize
2480 * @return 0 on success, non-zero on failure.
2483 mdb_txn_renew0(MDB_txn *txn)
2485 MDB_env *env = txn->mt_env;
2486 MDB_txninfo *ti = env->me_txns;
2490 int rc, new_notls = 0;
2492 if (txn->mt_flags & MDB_TXN_RDONLY) {
2494 txn->mt_numdbs = env->me_numdbs;
2495 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2497 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2498 txn->mt_txnid = meta->mm_txnid;
2499 txn->mt_u.reader = NULL;
2501 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2502 pthread_getspecific(env->me_txkey);
2504 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2505 return MDB_BAD_RSLOT;
2507 MDB_PID_T pid = env->me_pid;
2508 MDB_THR_T tid = pthread_self();
2510 if (!env->me_live_reader) {
2511 rc = mdb_reader_pid(env, Pidset, pid);
2514 env->me_live_reader = 1;
2518 nr = ti->mti_numreaders;
2519 for (i=0; i<nr; i++)
2520 if (ti->mti_readers[i].mr_pid == 0)
2522 if (i == env->me_maxreaders) {
2523 UNLOCK_MUTEX_R(env);
2524 return MDB_READERS_FULL;
2526 ti->mti_readers[i].mr_pid = pid;
2527 ti->mti_readers[i].mr_tid = tid;
2529 ti->mti_numreaders = ++nr;
2530 /* Save numreaders for un-mutexed mdb_env_close() */
2531 env->me_numreaders = nr;
2532 UNLOCK_MUTEX_R(env);
2534 r = &ti->mti_readers[i];
2535 new_notls = (env->me_flags & MDB_NOTLS);
2536 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2541 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2542 txn->mt_u.reader = r;
2543 meta = env->me_metas[txn->mt_txnid & 1];
2549 txn->mt_txnid = ti->mti_txnid;
2550 meta = env->me_metas[txn->mt_txnid & 1];
2552 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2553 txn->mt_txnid = meta->mm_txnid;
2556 txn->mt_numdbs = env->me_numdbs;
2559 if (txn->mt_txnid == mdb_debug_start)
2563 txn->mt_child = NULL;
2564 txn->mt_loose_pgs = NULL;
2565 txn->mt_loose_count = 0;
2566 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2567 txn->mt_u.dirty_list = env->me_dirty_list;
2568 txn->mt_u.dirty_list[0].mid = 0;
2569 txn->mt_free_pgs = env->me_free_pgs;
2570 txn->mt_free_pgs[0] = 0;
2571 txn->mt_spill_pgs = NULL;
2573 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2576 /* Copy the DB info and flags */
2577 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2579 /* Moved to here to avoid a data race in read TXNs */
2580 txn->mt_next_pgno = meta->mm_last_pg+1;
2582 for (i=2; i<txn->mt_numdbs; i++) {
2583 x = env->me_dbflags[i];
2584 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2585 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2587 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2589 if (env->me_maxpg < txn->mt_next_pgno) {
2590 mdb_txn_reset0(txn, "renew0-mapfail");
2592 txn->mt_u.reader->mr_pid = 0;
2593 txn->mt_u.reader = NULL;
2595 return MDB_MAP_RESIZED;
2602 mdb_txn_renew(MDB_txn *txn)
2606 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2609 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2610 DPUTS("environment had fatal error, must shutdown!");
2614 rc = mdb_txn_renew0(txn);
2615 if (rc == MDB_SUCCESS) {
2616 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2617 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2618 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2624 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2628 int rc, size, tsize = sizeof(MDB_txn);
2630 if (env->me_flags & MDB_FATAL_ERROR) {
2631 DPUTS("environment had fatal error, must shutdown!");
2634 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2637 /* Nested transactions: Max 1 child, write txns only, no writemap */
2638 if (parent->mt_child ||
2639 (flags & MDB_RDONLY) ||
2640 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2641 (env->me_flags & MDB_WRITEMAP))
2643 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2645 tsize = sizeof(MDB_ntxn);
2648 if (!(flags & MDB_RDONLY)) {
2650 txn = env->me_txn0; /* just reuse preallocated write txn */
2653 /* child txns use own copy of cursors */
2654 size += env->me_maxdbs * sizeof(MDB_cursor *);
2656 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2658 if ((txn = calloc(1, size)) == NULL) {
2659 DPRINTF(("calloc: %s", strerror(errno)));
2662 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2663 if (flags & MDB_RDONLY) {
2664 txn->mt_flags |= MDB_TXN_RDONLY;
2665 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2666 txn->mt_dbiseqs = env->me_dbiseqs;
2668 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2670 txn->mt_dbiseqs = parent->mt_dbiseqs;
2671 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2673 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2674 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2682 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2683 if (!txn->mt_u.dirty_list ||
2684 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2686 free(txn->mt_u.dirty_list);
2690 txn->mt_txnid = parent->mt_txnid;
2691 txn->mt_dirty_room = parent->mt_dirty_room;
2692 txn->mt_u.dirty_list[0].mid = 0;
2693 txn->mt_spill_pgs = NULL;
2694 txn->mt_next_pgno = parent->mt_next_pgno;
2695 parent->mt_child = txn;
2696 txn->mt_parent = parent;
2697 txn->mt_numdbs = parent->mt_numdbs;
2698 txn->mt_flags = parent->mt_flags;
2699 txn->mt_dbxs = parent->mt_dbxs;
2700 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2701 /* Copy parent's mt_dbflags, but clear DB_NEW */
2702 for (i=0; i<txn->mt_numdbs; i++)
2703 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2705 ntxn = (MDB_ntxn *)txn;
2706 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2707 if (env->me_pghead) {
2708 size = MDB_IDL_SIZEOF(env->me_pghead);
2709 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2711 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2716 rc = mdb_cursor_shadow(parent, txn);
2718 mdb_txn_reset0(txn, "beginchild-fail");
2720 rc = mdb_txn_renew0(txn);
2723 if (txn != env->me_txn0)
2727 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2728 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2729 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2736 mdb_txn_env(MDB_txn *txn)
2738 if(!txn) return NULL;
2742 /** Export or close DBI handles opened in this txn. */
2744 mdb_dbis_update(MDB_txn *txn, int keep)
2747 MDB_dbi n = txn->mt_numdbs;
2748 MDB_env *env = txn->mt_env;
2749 unsigned char *tdbflags = txn->mt_dbflags;
2751 for (i = n; --i >= 2;) {
2752 if (tdbflags[i] & DB_NEW) {
2754 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2756 char *ptr = env->me_dbxs[i].md_name.mv_data;
2758 env->me_dbxs[i].md_name.mv_data = NULL;
2759 env->me_dbxs[i].md_name.mv_size = 0;
2760 env->me_dbflags[i] = 0;
2761 env->me_dbiseqs[i]++;
2767 if (keep && env->me_numdbs < n)
2771 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2772 * May be called twice for readonly txns: First reset it, then abort.
2773 * @param[in] txn the transaction handle to reset
2774 * @param[in] act why the transaction is being reset
2777 mdb_txn_reset0(MDB_txn *txn, const char *act)
2779 MDB_env *env = txn->mt_env;
2781 /* Close any DBI handles opened in this txn */
2782 mdb_dbis_update(txn, 0);
2784 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2785 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2786 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2788 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2789 if (txn->mt_u.reader) {
2790 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2791 if (!(env->me_flags & MDB_NOTLS))
2792 txn->mt_u.reader = NULL; /* txn does not own reader */
2794 txn->mt_numdbs = 0; /* close nothing if called again */
2795 txn->mt_dbxs = NULL; /* mark txn as reset */
2797 pgno_t *pghead = env->me_pghead;
2799 mdb_cursors_close(txn, 0);
2800 if (!(env->me_flags & MDB_WRITEMAP)) {
2801 mdb_dlist_free(txn);
2804 if (!txn->mt_parent) {
2805 if (mdb_midl_shrink(&txn->mt_free_pgs))
2806 env->me_free_pgs = txn->mt_free_pgs;
2808 env->me_pghead = NULL;
2812 /* The writer mutex was locked in mdb_txn_begin. */
2814 UNLOCK_MUTEX_W(env);
2816 txn->mt_parent->mt_child = NULL;
2817 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2818 mdb_midl_free(txn->mt_free_pgs);
2819 mdb_midl_free(txn->mt_spill_pgs);
2820 free(txn->mt_u.dirty_list);
2823 mdb_midl_free(pghead);
2828 mdb_txn_reset(MDB_txn *txn)
2833 /* This call is only valid for read-only txns */
2834 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2837 mdb_txn_reset0(txn, "reset");
2841 mdb_txn_abort(MDB_txn *txn)
2847 mdb_txn_abort(txn->mt_child);
2849 mdb_txn_reset0(txn, "abort");
2850 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2851 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2852 txn->mt_u.reader->mr_pid = 0;
2854 if (txn != txn->mt_env->me_txn0)
2858 /** Save the freelist as of this transaction to the freeDB.
2859 * This changes the freelist. Keep trying until it stabilizes.
2862 mdb_freelist_save(MDB_txn *txn)
2864 /* env->me_pghead[] can grow and shrink during this call.
2865 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2866 * Page numbers cannot disappear from txn->mt_free_pgs[].
2869 MDB_env *env = txn->mt_env;
2870 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2871 txnid_t pglast = 0, head_id = 0;
2872 pgno_t freecnt = 0, *free_pgs, *mop;
2873 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2875 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2877 if (env->me_pghead) {
2878 /* Make sure first page of freeDB is touched and on freelist */
2879 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2880 if (rc && rc != MDB_NOTFOUND)
2884 if (!env->me_pghead && txn->mt_loose_pgs) {
2885 /* Put loose page numbers in mt_free_pgs, since
2886 * we may be unable to return them to me_pghead.
2888 MDB_page *mp = txn->mt_loose_pgs;
2889 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2891 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2892 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2893 txn->mt_loose_pgs = NULL;
2894 txn->mt_loose_count = 0;
2897 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2898 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2899 ? SSIZE_MAX : maxfree_1pg;
2902 /* Come back here after each Put() in case freelist changed */
2907 /* If using records from freeDB which we have not yet
2908 * deleted, delete them and any we reserved for me_pghead.
2910 while (pglast < env->me_pglast) {
2911 rc = mdb_cursor_first(&mc, &key, NULL);
2914 pglast = head_id = *(txnid_t *)key.mv_data;
2915 total_room = head_room = 0;
2916 mdb_tassert(txn, pglast <= env->me_pglast);
2917 rc = mdb_cursor_del(&mc, 0);
2922 /* Save the IDL of pages freed by this txn, to a single record */
2923 if (freecnt < txn->mt_free_pgs[0]) {
2925 /* Make sure last page of freeDB is touched and on freelist */
2926 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2927 if (rc && rc != MDB_NOTFOUND)
2930 free_pgs = txn->mt_free_pgs;
2931 /* Write to last page of freeDB */
2932 key.mv_size = sizeof(txn->mt_txnid);
2933 key.mv_data = &txn->mt_txnid;
2935 freecnt = free_pgs[0];
2936 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2937 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2940 /* Retry if mt_free_pgs[] grew during the Put() */
2941 free_pgs = txn->mt_free_pgs;
2942 } while (freecnt < free_pgs[0]);
2943 mdb_midl_sort(free_pgs);
2944 memcpy(data.mv_data, free_pgs, data.mv_size);
2947 unsigned int i = free_pgs[0];
2948 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2949 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2951 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2957 mop = env->me_pghead;
2958 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2960 /* Reserve records for me_pghead[]. Split it if multi-page,
2961 * to avoid searching freeDB for a page range. Use keys in
2962 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2964 if (total_room >= mop_len) {
2965 if (total_room == mop_len || --more < 0)
2967 } else if (head_room >= maxfree_1pg && head_id > 1) {
2968 /* Keep current record (overflow page), add a new one */
2972 /* (Re)write {key = head_id, IDL length = head_room} */
2973 total_room -= head_room;
2974 head_room = mop_len - total_room;
2975 if (head_room > maxfree_1pg && head_id > 1) {
2976 /* Overflow multi-page for part of me_pghead */
2977 head_room /= head_id; /* amortize page sizes */
2978 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2979 } else if (head_room < 0) {
2980 /* Rare case, not bothering to delete this record */
2983 key.mv_size = sizeof(head_id);
2984 key.mv_data = &head_id;
2985 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2986 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2989 /* IDL is initially empty, zero out at least the length */
2990 pgs = (pgno_t *)data.mv_data;
2991 j = head_room > clean_limit ? head_room : 0;
2995 total_room += head_room;
2998 /* Return loose page numbers to me_pghead, though usually none are
2999 * left at this point. The pages themselves remain in dirty_list.
3001 if (txn->mt_loose_pgs) {
3002 MDB_page *mp = txn->mt_loose_pgs;
3003 unsigned count = txn->mt_loose_count;
3005 /* Room for loose pages + temp IDL with same */
3006 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3008 mop = env->me_pghead;
3009 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3010 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3011 loose[ ++count ] = mp->mp_pgno;
3013 mdb_midl_sort(loose);
3014 mdb_midl_xmerge(mop, loose);
3015 txn->mt_loose_pgs = NULL;
3016 txn->mt_loose_count = 0;
3020 /* Fill in the reserved me_pghead records */
3026 rc = mdb_cursor_first(&mc, &key, &data);
3027 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3028 txnid_t id = *(txnid_t *)key.mv_data;
3029 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3032 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3034 if (len > mop_len) {
3036 data.mv_size = (len + 1) * sizeof(MDB_ID);
3038 data.mv_data = mop -= len;
3041 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3043 if (rc || !(mop_len -= len))
3050 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3051 * @param[in] txn the transaction that's being committed
3052 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3053 * @return 0 on success, non-zero on failure.
3056 mdb_page_flush(MDB_txn *txn, int keep)
3058 MDB_env *env = txn->mt_env;
3059 MDB_ID2L dl = txn->mt_u.dirty_list;
3060 unsigned psize = env->me_psize, j;
3061 int i, pagecount = dl[0].mid, rc;
3062 size_t size = 0, pos = 0;
3064 MDB_page *dp = NULL;
3068 struct iovec iov[MDB_COMMIT_PAGES];
3069 ssize_t wpos = 0, wsize = 0, wres;
3070 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3076 if (env->me_flags & MDB_WRITEMAP) {
3077 /* Clear dirty flags */
3078 while (++i <= pagecount) {
3080 /* Don't flush this page yet */
3081 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3082 dp->mp_flags &= ~P_KEEP;
3086 dp->mp_flags &= ~P_DIRTY;
3091 /* Write the pages */
3093 if (++i <= pagecount) {
3095 /* Don't flush this page yet */
3096 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3097 dp->mp_flags &= ~P_KEEP;
3102 /* clear dirty flag */
3103 dp->mp_flags &= ~P_DIRTY;
3106 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3111 /* Windows actually supports scatter/gather I/O, but only on
3112 * unbuffered file handles. Since we're relying on the OS page
3113 * cache for all our data, that's self-defeating. So we just
3114 * write pages one at a time. We use the ov structure to set
3115 * the write offset, to at least save the overhead of a Seek
3118 DPRINTF(("committing page %"Z"u", pgno));
3119 memset(&ov, 0, sizeof(ov));
3120 ov.Offset = pos & 0xffffffff;
3121 ov.OffsetHigh = pos >> 16 >> 16;
3122 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3124 DPRINTF(("WriteFile: %d", rc));
3128 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3129 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3131 /* Write previous page(s) */
3132 #ifdef MDB_USE_PWRITEV
3133 wres = pwritev(env->me_fd, iov, n, wpos);
3136 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3138 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3140 DPRINTF(("lseek: %s", strerror(rc)));
3143 wres = writev(env->me_fd, iov, n);
3146 if (wres != wsize) {
3149 DPRINTF(("Write error: %s", strerror(rc)));
3151 rc = EIO; /* TODO: Use which error code? */
3152 DPUTS("short write, filesystem full?");
3163 DPRINTF(("committing page %"Z"u", pgno));
3164 next_pos = pos + size;
3165 iov[n].iov_len = size;
3166 iov[n].iov_base = (char *)dp;
3172 /* MIPS has cache coherency issues, this is a no-op everywhere else
3173 * Note: for any size >= on-chip cache size, entire on-chip cache is
3176 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3178 for (i = keep; ++i <= pagecount; ) {
3180 /* This is a page we skipped above */
3183 dl[j].mid = dp->mp_pgno;
3186 mdb_dpage_free(env, dp);
3191 txn->mt_dirty_room += i - j;
3197 mdb_txn_commit(MDB_txn *txn)
3203 if (txn == NULL || txn->mt_env == NULL)
3206 if (txn->mt_child) {
3207 rc = mdb_txn_commit(txn->mt_child);
3208 txn->mt_child = NULL;
3215 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3216 mdb_dbis_update(txn, 1);
3217 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3222 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3223 DPUTS("error flag is set, can't commit");
3225 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3230 if (txn->mt_parent) {
3231 MDB_txn *parent = txn->mt_parent;
3235 unsigned x, y, len, ps_len;
3237 /* Append our free list to parent's */
3238 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3241 mdb_midl_free(txn->mt_free_pgs);
3242 /* Failures after this must either undo the changes
3243 * to the parent or set MDB_TXN_ERROR in the parent.
3246 parent->mt_next_pgno = txn->mt_next_pgno;
3247 parent->mt_flags = txn->mt_flags;
3249 /* Merge our cursors into parent's and close them */
3250 mdb_cursors_close(txn, 1);
3252 /* Update parent's DB table. */
3253 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3254 parent->mt_numdbs = txn->mt_numdbs;
3255 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3256 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3257 for (i=2; i<txn->mt_numdbs; i++) {
3258 /* preserve parent's DB_NEW status */
3259 x = parent->mt_dbflags[i] & DB_NEW;
3260 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3263 dst = parent->mt_u.dirty_list;
3264 src = txn->mt_u.dirty_list;
3265 /* Remove anything in our dirty list from parent's spill list */
3266 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3268 pspill[0] = (pgno_t)-1;
3269 /* Mark our dirty pages as deleted in parent spill list */
3270 for (i=0, len=src[0].mid; ++i <= len; ) {
3271 MDB_ID pn = src[i].mid << 1;
3272 while (pn > pspill[x])
3274 if (pn == pspill[x]) {
3279 /* Squash deleted pagenums if we deleted any */
3280 for (x=y; ++x <= ps_len; )
3281 if (!(pspill[x] & 1))
3282 pspill[++y] = pspill[x];
3286 /* Find len = length of merging our dirty list with parent's */
3288 dst[0].mid = 0; /* simplify loops */
3289 if (parent->mt_parent) {
3290 len = x + src[0].mid;
3291 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3292 for (i = x; y && i; y--) {
3293 pgno_t yp = src[y].mid;
3294 while (yp < dst[i].mid)
3296 if (yp == dst[i].mid) {
3301 } else { /* Simplify the above for single-ancestor case */
3302 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3304 /* Merge our dirty list with parent's */
3306 for (i = len; y; dst[i--] = src[y--]) {
3307 pgno_t yp = src[y].mid;
3308 while (yp < dst[x].mid)
3309 dst[i--] = dst[x--];
3310 if (yp == dst[x].mid)
3311 free(dst[x--].mptr);
3313 mdb_tassert(txn, i == x);
3315 free(txn->mt_u.dirty_list);
3316 parent->mt_dirty_room = txn->mt_dirty_room;
3317 if (txn->mt_spill_pgs) {
3318 if (parent->mt_spill_pgs) {
3319 /* TODO: Prevent failure here, so parent does not fail */
3320 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3322 parent->mt_flags |= MDB_TXN_ERROR;
3323 mdb_midl_free(txn->mt_spill_pgs);
3324 mdb_midl_sort(parent->mt_spill_pgs);
3326 parent->mt_spill_pgs = txn->mt_spill_pgs;
3330 /* Append our loose page list to parent's */
3331 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3333 *lp = txn->mt_loose_pgs;
3334 parent->mt_loose_count += txn->mt_loose_count;
3336 parent->mt_child = NULL;
3337 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3342 if (txn != env->me_txn) {
3343 DPUTS("attempt to commit unknown transaction");
3348 mdb_cursors_close(txn, 0);
3350 if (!txn->mt_u.dirty_list[0].mid &&
3351 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3354 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3355 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3357 /* Update DB root pointers */
3358 if (txn->mt_numdbs > 2) {
3362 data.mv_size = sizeof(MDB_db);
3364 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3365 for (i = 2; i < txn->mt_numdbs; i++) {
3366 if (txn->mt_dbflags[i] & DB_DIRTY) {
3367 if (TXN_DBI_CHANGED(txn, i)) {
3371 data.mv_data = &txn->mt_dbs[i];
3372 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3379 rc = mdb_freelist_save(txn);
3383 mdb_midl_free(env->me_pghead);
3384 env->me_pghead = NULL;
3385 if (mdb_midl_shrink(&txn->mt_free_pgs))
3386 env->me_free_pgs = txn->mt_free_pgs;
3392 if ((rc = mdb_page_flush(txn, 0)) ||
3393 (rc = mdb_env_sync(env, 0)) ||
3394 (rc = mdb_env_write_meta(txn)))
3397 /* Free P_LOOSE pages left behind in dirty_list */
3398 if (!(env->me_flags & MDB_WRITEMAP))
3399 mdb_dlist_free(txn);
3404 mdb_dbis_update(txn, 1);
3407 UNLOCK_MUTEX_W(env);
3408 if (txn != env->me_txn0)
3418 /** Read the environment parameters of a DB environment before
3419 * mapping it into memory.
3420 * @param[in] env the environment handle
3421 * @param[out] meta address of where to store the meta information
3422 * @return 0 on success, non-zero on failure.
3425 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3431 enum { Size = sizeof(pbuf) };
3433 /* We don't know the page size yet, so use a minimum value.
3434 * Read both meta pages so we can use the latest one.
3437 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3441 memset(&ov, 0, sizeof(ov));
3443 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3444 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3447 rc = pread(env->me_fd, &pbuf, Size, off);
3450 if (rc == 0 && off == 0)
3452 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3453 DPRINTF(("read: %s", mdb_strerror(rc)));
3457 p = (MDB_page *)&pbuf;
3459 if (!F_ISSET(p->mp_flags, P_META)) {
3460 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3465 if (m->mm_magic != MDB_MAGIC) {
3466 DPUTS("meta has invalid magic");
3470 if (m->mm_version != MDB_DATA_VERSION) {
3471 DPRINTF(("database is version %u, expected version %u",
3472 m->mm_version, MDB_DATA_VERSION));
3473 return MDB_VERSION_MISMATCH;
3476 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3483 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3485 meta->mm_magic = MDB_MAGIC;
3486 meta->mm_version = MDB_DATA_VERSION;
3487 meta->mm_mapsize = env->me_mapsize;
3488 meta->mm_psize = env->me_psize;
3489 meta->mm_last_pg = 1;
3490 meta->mm_flags = env->me_flags & 0xffff;
3491 meta->mm_flags |= MDB_INTEGERKEY;
3492 meta->mm_dbs[0].md_root = P_INVALID;
3493 meta->mm_dbs[1].md_root = P_INVALID;
3496 /** Write the environment parameters of a freshly created DB environment.
3497 * @param[in] env the environment handle
3498 * @param[out] meta address of where to store the meta information
3499 * @return 0 on success, non-zero on failure.
3502 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3510 memset(&ov, 0, sizeof(ov));
3511 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3513 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3516 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3517 len = pwrite(fd, ptr, size, pos); \
3518 rc = (len >= 0); } while(0)
3521 DPUTS("writing new meta page");
3523 psize = env->me_psize;
3525 mdb_env_init_meta0(env, meta);
3527 p = calloc(2, psize);
3529 p->mp_flags = P_META;
3530 *(MDB_meta *)METADATA(p) = *meta;
3532 q = (MDB_page *)((char *)p + psize);
3534 q->mp_flags = P_META;
3535 *(MDB_meta *)METADATA(q) = *meta;
3537 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3540 else if ((unsigned) len == psize * 2)
3548 /** Update the environment info to commit a transaction.
3549 * @param[in] txn the transaction that's being committed
3550 * @return 0 on success, non-zero on failure.
3553 mdb_env_write_meta(MDB_txn *txn)
3556 MDB_meta meta, metab, *mp;
3559 int rc, len, toggle;
3568 toggle = txn->mt_txnid & 1;
3569 DPRINTF(("writing meta page %d for root page %"Z"u",
3570 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3573 mp = env->me_metas[toggle];
3574 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3575 /* Persist any increases of mapsize config */
3576 if (mapsize < env->me_mapsize)
3577 mapsize = env->me_mapsize;
3579 if (env->me_flags & MDB_WRITEMAP) {
3580 mp->mm_mapsize = mapsize;
3581 mp->mm_dbs[0] = txn->mt_dbs[0];
3582 mp->mm_dbs[1] = txn->mt_dbs[1];
3583 mp->mm_last_pg = txn->mt_next_pgno - 1;
3584 mp->mm_txnid = txn->mt_txnid;
3585 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3586 unsigned meta_size = env->me_psize;
3587 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3590 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3591 if (meta_size < env->me_os_psize)
3592 meta_size += meta_size;
3597 if (MDB_MSYNC(ptr, meta_size, rc)) {
3604 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3605 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3607 meta.mm_mapsize = mapsize;
3608 meta.mm_dbs[0] = txn->mt_dbs[0];
3609 meta.mm_dbs[1] = txn->mt_dbs[1];
3610 meta.mm_last_pg = txn->mt_next_pgno - 1;
3611 meta.mm_txnid = txn->mt_txnid;
3613 off = offsetof(MDB_meta, mm_mapsize);
3614 ptr = (char *)&meta + off;
3615 len = sizeof(MDB_meta) - off;
3617 off += env->me_psize;
3620 /* Write to the SYNC fd */
3621 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3622 env->me_fd : env->me_mfd;
3625 memset(&ov, 0, sizeof(ov));
3627 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3631 rc = pwrite(mfd, ptr, len, off);
3634 rc = rc < 0 ? ErrCode() : EIO;
3635 DPUTS("write failed, disk error?");
3636 /* On a failure, the pagecache still contains the new data.
3637 * Write some old data back, to prevent it from being used.
3638 * Use the non-SYNC fd; we know it will fail anyway.
3640 meta.mm_last_pg = metab.mm_last_pg;
3641 meta.mm_txnid = metab.mm_txnid;
3643 memset(&ov, 0, sizeof(ov));
3645 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3647 r2 = pwrite(env->me_fd, ptr, len, off);
3648 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3651 env->me_flags |= MDB_FATAL_ERROR;
3654 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3655 CACHEFLUSH(env->me_map + off, len, DCACHE);
3657 /* Memory ordering issues are irrelevant; since the entire writer
3658 * is wrapped by wmutex, all of these changes will become visible
3659 * after the wmutex is unlocked. Since the DB is multi-version,
3660 * readers will get consistent data regardless of how fresh or
3661 * how stale their view of these values is.
3664 env->me_txns->mti_txnid = txn->mt_txnid;
3669 /** Check both meta pages to see which one is newer.
3670 * @param[in] env the environment handle
3671 * @return meta toggle (0 or 1).
3674 mdb_env_pick_meta(const MDB_env *env)
3676 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3680 mdb_env_create(MDB_env **env)
3684 e = calloc(1, sizeof(MDB_env));
3688 e->me_maxreaders = DEFAULT_READERS;
3689 e->me_maxdbs = e->me_numdbs = 2;
3690 e->me_fd = INVALID_HANDLE_VALUE;
3691 e->me_lfd = INVALID_HANDLE_VALUE;
3692 e->me_mfd = INVALID_HANDLE_VALUE;
3693 #ifdef MDB_USE_POSIX_SEM
3694 e->me_rmutex = SEM_FAILED;
3695 e->me_wmutex = SEM_FAILED;
3697 e->me_pid = getpid();
3698 GET_PAGESIZE(e->me_os_psize);
3699 VGMEMP_CREATE(e,0,0);
3705 mdb_env_map(MDB_env *env, void *addr)
3708 unsigned int flags = env->me_flags;
3712 LONG sizelo, sizehi;
3715 if (flags & MDB_RDONLY) {
3716 /* Don't set explicit map size, use whatever exists */
3721 msize = env->me_mapsize;
3722 sizelo = msize & 0xffffffff;
3723 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3725 /* Windows won't create mappings for zero length files.
3726 * and won't map more than the file size.
3727 * Just set the maxsize right now.
3729 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3730 || !SetEndOfFile(env->me_fd)
3731 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3735 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3736 PAGE_READWRITE : PAGE_READONLY,
3737 sizehi, sizelo, NULL);
3740 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3741 FILE_MAP_WRITE : FILE_MAP_READ,
3743 rc = env->me_map ? 0 : ErrCode();
3748 int prot = PROT_READ;
3749 if (flags & MDB_WRITEMAP) {
3751 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3754 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3756 if (env->me_map == MAP_FAILED) {
3761 if (flags & MDB_NORDAHEAD) {
3762 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3764 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3766 #ifdef POSIX_MADV_RANDOM
3767 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3768 #endif /* POSIX_MADV_RANDOM */
3769 #endif /* MADV_RANDOM */
3773 /* Can happen because the address argument to mmap() is just a
3774 * hint. mmap() can pick another, e.g. if the range is in use.
3775 * The MAP_FIXED flag would prevent that, but then mmap could
3776 * instead unmap existing pages to make room for the new map.
3778 if (addr && env->me_map != addr)
3779 return EBUSY; /* TODO: Make a new MDB_* error code? */
3781 p = (MDB_page *)env->me_map;
3782 env->me_metas[0] = METADATA(p);
3783 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3789 mdb_env_set_mapsize(MDB_env *env, size_t size)
3791 /* If env is already open, caller is responsible for making
3792 * sure there are no active txns.
3800 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3801 else if (size < env->me_mapsize) {
3802 /* If the configured size is smaller, make sure it's
3803 * still big enough. Silently round up to minimum if not.
3805 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3809 munmap(env->me_map, env->me_mapsize);
3810 env->me_mapsize = size;
3811 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3812 rc = mdb_env_map(env, old);
3816 env->me_mapsize = size;
3818 env->me_maxpg = env->me_mapsize / env->me_psize;
3823 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3827 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3832 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3834 if (env->me_map || readers < 1)
3836 env->me_maxreaders = readers;
3841 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3843 if (!env || !readers)
3845 *readers = env->me_maxreaders;
3850 mdb_fsize(HANDLE fd, size_t *size)
3853 LARGE_INTEGER fsize;
3855 if (!GetFileSizeEx(fd, &fsize))
3858 *size = fsize.QuadPart;
3870 #ifdef BROKEN_FDATASYNC
3871 #include <sys/utsname.h>
3872 #include <sys/vfs.h>
3875 /** Further setup required for opening an LMDB environment
3878 mdb_env_open2(MDB_env *env)
3880 unsigned int flags = env->me_flags;
3881 int i, newenv = 0, rc;
3885 /* See if we should use QueryLimited */
3887 if ((rc & 0xff) > 5)
3888 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3890 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3892 #ifdef BROKEN_FDATASYNC
3893 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
3894 * https://lkml.org/lkml/2012/9/3/83
3895 * Kernels after 3.6-rc6 are known good.
3896 * https://lkml.org/lkml/2012/9/10/556
3897 * See if the DB is on ext3/ext4, then check for new enough kernel
3898 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
3903 fstatfs(env->me_fd, &st);
3904 while (st.f_type == 0xEF53) {
3908 if (uts.release[0] < '3') {
3909 if (!strncmp(uts.release, "2.6.32.", 7)) {
3910 i = atoi(uts.release+7);
3912 break; /* 2.6.32.60 and newer is OK */
3913 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
3914 i = atoi(uts.release+7);
3916 break; /* 2.6.34.15 and newer is OK */
3918 } else if (uts.release[0] == '3') {
3919 i = atoi(uts.release+2);
3921 break; /* 3.6 and newer is OK */
3923 i = atoi(uts.release+4);
3925 break; /* 3.5.4 and newer is OK */
3926 } else if (i == 2) {
3927 i = atoi(uts.release+4);
3929 break; /* 3.2.30 and newer is OK */
3931 } else { /* 4.x and newer is OK */
3934 env->me_fsynconly = 1;
3940 memset(&meta, 0, sizeof(meta));
3942 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3945 DPUTS("new mdbenv");
3947 env->me_psize = env->me_os_psize;
3948 if (env->me_psize > MAX_PAGESIZE)
3949 env->me_psize = MAX_PAGESIZE;
3951 env->me_psize = meta.mm_psize;
3954 /* Was a mapsize configured? */
3955 if (!env->me_mapsize) {
3956 /* If this is a new environment, take the default,
3957 * else use the size recorded in the existing env.
3959 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3960 } else if (env->me_mapsize < meta.mm_mapsize) {
3961 /* If the configured size is smaller, make sure it's
3962 * still big enough. Silently round up to minimum if not.
3964 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3965 if (env->me_mapsize < minsize)
3966 env->me_mapsize = minsize;
3969 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3974 if (flags & MDB_FIXEDMAP)
3975 meta.mm_address = env->me_map;
3976 i = mdb_env_init_meta(env, &meta);
3977 if (i != MDB_SUCCESS) {
3982 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3983 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3985 #if !(MDB_MAXKEYSIZE)
3986 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3988 env->me_maxpg = env->me_mapsize / env->me_psize;
3992 int toggle = mdb_env_pick_meta(env);
3993 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3995 DPRINTF(("opened database version %u, pagesize %u",
3996 env->me_metas[0]->mm_version, env->me_psize));
3997 DPRINTF(("using meta page %d", toggle));
3998 DPRINTF(("depth: %u", db->md_depth));
3999 DPRINTF(("entries: %"Z"u", db->md_entries));
4000 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4001 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4002 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4003 DPRINTF(("root: %"Z"u", db->md_root));
4011 /** Release a reader thread's slot in the reader lock table.
4012 * This function is called automatically when a thread exits.
4013 * @param[in] ptr This points to the slot in the reader lock table.
4016 mdb_env_reader_dest(void *ptr)
4018 MDB_reader *reader = ptr;
4024 /** Junk for arranging thread-specific callbacks on Windows. This is
4025 * necessarily platform and compiler-specific. Windows supports up
4026 * to 1088 keys. Let's assume nobody opens more than 64 environments
4027 * in a single process, for now. They can override this if needed.
4029 #ifndef MAX_TLS_KEYS
4030 #define MAX_TLS_KEYS 64
4032 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4033 static int mdb_tls_nkeys;
4035 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4039 case DLL_PROCESS_ATTACH: break;
4040 case DLL_THREAD_ATTACH: break;
4041 case DLL_THREAD_DETACH:
4042 for (i=0; i<mdb_tls_nkeys; i++) {
4043 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4045 mdb_env_reader_dest(r);
4049 case DLL_PROCESS_DETACH: break;
4054 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4056 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4060 /* Force some symbol references.
4061 * _tls_used forces the linker to create the TLS directory if not already done
4062 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4064 #pragma comment(linker, "/INCLUDE:_tls_used")
4065 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4066 #pragma const_seg(".CRT$XLB")
4067 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4068 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4071 #pragma comment(linker, "/INCLUDE:__tls_used")
4072 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4073 #pragma data_seg(".CRT$XLB")
4074 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4076 #endif /* WIN 32/64 */
4077 #endif /* !__GNUC__ */
4080 /** Downgrade the exclusive lock on the region back to shared */
4082 mdb_env_share_locks(MDB_env *env, int *excl)
4084 int rc = 0, toggle = mdb_env_pick_meta(env);
4086 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4091 /* First acquire a shared lock. The Unlock will
4092 * then release the existing exclusive lock.
4094 memset(&ov, 0, sizeof(ov));
4095 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4098 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4104 struct flock lock_info;
4105 /* The shared lock replaces the existing lock */
4106 memset((void *)&lock_info, 0, sizeof(lock_info));
4107 lock_info.l_type = F_RDLCK;
4108 lock_info.l_whence = SEEK_SET;
4109 lock_info.l_start = 0;
4110 lock_info.l_len = 1;
4111 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4112 (rc = ErrCode()) == EINTR) ;
4113 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4120 /** Try to get exclusive lock, otherwise shared.
4121 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4124 mdb_env_excl_lock(MDB_env *env, int *excl)
4128 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4132 memset(&ov, 0, sizeof(ov));
4133 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4140 struct flock lock_info;
4141 memset((void *)&lock_info, 0, sizeof(lock_info));
4142 lock_info.l_type = F_WRLCK;
4143 lock_info.l_whence = SEEK_SET;
4144 lock_info.l_start = 0;
4145 lock_info.l_len = 1;
4146 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4147 (rc = ErrCode()) == EINTR) ;
4151 # ifdef MDB_USE_POSIX_SEM
4152 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4155 lock_info.l_type = F_RDLCK;
4156 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4157 (rc = ErrCode()) == EINTR) ;
4167 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4169 * @(#) $Revision: 5.1 $
4170 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4171 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4173 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4177 * Please do not copyright this code. This code is in the public domain.
4179 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4180 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4181 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4182 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4183 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4184 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4185 * PERFORMANCE OF THIS SOFTWARE.
4188 * chongo <Landon Curt Noll> /\oo/\
4189 * http://www.isthe.com/chongo/
4191 * Share and Enjoy! :-)
4194 typedef unsigned long long mdb_hash_t;
4195 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4197 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4198 * @param[in] val value to hash
4199 * @param[in] hval initial value for hash
4200 * @return 64 bit hash
4202 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4203 * hval arg on the first call.
4206 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4208 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4209 unsigned char *end = s + val->mv_size;
4211 * FNV-1a hash each octet of the string
4214 /* xor the bottom with the current octet */
4215 hval ^= (mdb_hash_t)*s++;
4217 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4218 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4219 (hval << 7) + (hval << 8) + (hval << 40);
4221 /* return our new hash value */
4225 /** Hash the string and output the encoded hash.
4226 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4227 * very short name limits. We don't care about the encoding being reversible,
4228 * we just want to preserve as many bits of the input as possible in a
4229 * small printable string.
4230 * @param[in] str string to hash
4231 * @param[out] encbuf an array of 11 chars to hold the hash
4233 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4236 mdb_pack85(unsigned long l, char *out)
4240 for (i=0; i<5; i++) {
4241 *out++ = mdb_a85[l % 85];
4247 mdb_hash_enc(MDB_val *val, char *encbuf)
4249 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4251 mdb_pack85(h, encbuf);
4252 mdb_pack85(h>>32, encbuf+5);
4257 /** Open and/or initialize the lock region for the environment.
4258 * @param[in] env The LMDB environment.
4259 * @param[in] lpath The pathname of the file used for the lock region.
4260 * @param[in] mode The Unix permissions for the file, if we create it.
4261 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4262 * @return 0 on success, non-zero on failure.
4265 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4268 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4270 # define MDB_ERRCODE_ROFS EROFS
4271 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4272 # define MDB_CLOEXEC O_CLOEXEC
4275 # define MDB_CLOEXEC 0
4282 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4283 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4284 FILE_ATTRIBUTE_NORMAL, NULL);
4286 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4288 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4290 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4295 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4296 /* Lose record locks when exec*() */
4297 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4298 fcntl(env->me_lfd, F_SETFD, fdflags);
4301 if (!(env->me_flags & MDB_NOTLS)) {
4302 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4305 env->me_flags |= MDB_ENV_TXKEY;
4307 /* Windows TLS callbacks need help finding their TLS info. */
4308 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4312 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4316 /* Try to get exclusive lock. If we succeed, then
4317 * nobody is using the lock region and we should initialize it.
4319 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4322 size = GetFileSize(env->me_lfd, NULL);
4324 size = lseek(env->me_lfd, 0, SEEK_END);
4325 if (size == -1) goto fail_errno;
4327 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4328 if (size < rsize && *excl > 0) {
4330 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4331 || !SetEndOfFile(env->me_lfd))
4334 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4338 size = rsize - sizeof(MDB_txninfo);
4339 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4344 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4346 if (!mh) goto fail_errno;
4347 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4349 if (!env->me_txns) goto fail_errno;
4351 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4353 if (m == MAP_FAILED) goto fail_errno;
4359 BY_HANDLE_FILE_INFORMATION stbuf;
4368 if (!mdb_sec_inited) {
4369 InitializeSecurityDescriptor(&mdb_null_sd,
4370 SECURITY_DESCRIPTOR_REVISION);
4371 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4372 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4373 mdb_all_sa.bInheritHandle = FALSE;
4374 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4377 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4378 idbuf.volume = stbuf.dwVolumeSerialNumber;
4379 idbuf.nhigh = stbuf.nFileIndexHigh;
4380 idbuf.nlow = stbuf.nFileIndexLow;
4381 val.mv_data = &idbuf;
4382 val.mv_size = sizeof(idbuf);
4383 mdb_hash_enc(&val, encbuf);
4384 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4385 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4386 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4387 if (!env->me_rmutex) goto fail_errno;
4388 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4389 if (!env->me_wmutex) goto fail_errno;
4390 #elif defined(MDB_USE_POSIX_SEM)
4399 #if defined(__NetBSD__)
4400 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4402 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4403 idbuf.dev = stbuf.st_dev;
4404 idbuf.ino = stbuf.st_ino;
4405 val.mv_data = &idbuf;
4406 val.mv_size = sizeof(idbuf);
4407 mdb_hash_enc(&val, encbuf);
4408 #ifdef MDB_SHORT_SEMNAMES
4409 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4411 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4412 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4413 /* Clean up after a previous run, if needed: Try to
4414 * remove both semaphores before doing anything else.
4416 sem_unlink(env->me_txns->mti_rmname);
4417 sem_unlink(env->me_txns->mti_wmname);
4418 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4419 O_CREAT|O_EXCL, mode, 1);
4420 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4421 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4422 O_CREAT|O_EXCL, mode, 1);
4423 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4424 #else /* MDB_USE_POSIX_SEM */
4425 pthread_mutexattr_t mattr;
4427 if ((rc = pthread_mutexattr_init(&mattr))
4428 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4429 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4430 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4432 pthread_mutexattr_destroy(&mattr);
4433 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4435 env->me_txns->mti_magic = MDB_MAGIC;
4436 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4437 env->me_txns->mti_txnid = 0;
4438 env->me_txns->mti_numreaders = 0;
4441 if (env->me_txns->mti_magic != MDB_MAGIC) {
4442 DPUTS("lock region has invalid magic");
4446 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4447 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4448 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4449 rc = MDB_VERSION_MISMATCH;
4453 if (rc && rc != EACCES && rc != EAGAIN) {
4457 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4458 if (!env->me_rmutex) goto fail_errno;
4459 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4460 if (!env->me_wmutex) goto fail_errno;
4461 #elif defined(MDB_USE_POSIX_SEM)
4462 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4463 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4464 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4465 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4476 /** The name of the lock file in the DB environment */
4477 #define LOCKNAME "/lock.mdb"
4478 /** The name of the data file in the DB environment */
4479 #define DATANAME "/data.mdb"
4480 /** The suffix of the lock file when no subdir is used */
4481 #define LOCKSUFF "-lock"
4482 /** Only a subset of the @ref mdb_env flags can be changed
4483 * at runtime. Changing other flags requires closing the
4484 * environment and re-opening it with the new flags.
4486 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4487 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4488 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4490 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4491 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4495 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4497 int oflags, rc, len, excl = -1;
4498 char *lpath, *dpath;
4500 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4504 if (flags & MDB_NOSUBDIR) {
4505 rc = len + sizeof(LOCKSUFF) + len + 1;
4507 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4512 if (flags & MDB_NOSUBDIR) {
4513 dpath = lpath + len + sizeof(LOCKSUFF);
4514 sprintf(lpath, "%s" LOCKSUFF, path);
4515 strcpy(dpath, path);
4517 dpath = lpath + len + sizeof(LOCKNAME);
4518 sprintf(lpath, "%s" LOCKNAME, path);
4519 sprintf(dpath, "%s" DATANAME, path);
4523 flags |= env->me_flags;
4524 if (flags & MDB_RDONLY) {
4525 /* silently ignore WRITEMAP when we're only getting read access */
4526 flags &= ~MDB_WRITEMAP;
4528 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4529 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4532 env->me_flags = flags |= MDB_ENV_ACTIVE;
4536 env->me_path = strdup(path);
4537 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4538 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4539 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4540 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4545 /* For RDONLY, get lockfile after we know datafile exists */
4546 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4547 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4553 if (F_ISSET(flags, MDB_RDONLY)) {
4554 oflags = GENERIC_READ;
4555 len = OPEN_EXISTING;
4557 oflags = GENERIC_READ|GENERIC_WRITE;
4560 mode = FILE_ATTRIBUTE_NORMAL;
4561 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4562 NULL, len, mode, NULL);
4564 if (F_ISSET(flags, MDB_RDONLY))
4567 oflags = O_RDWR | O_CREAT;
4569 env->me_fd = open(dpath, oflags, mode);
4571 if (env->me_fd == INVALID_HANDLE_VALUE) {
4576 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4577 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4582 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4583 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4584 env->me_mfd = env->me_fd;
4586 /* Synchronous fd for meta writes. Needed even with
4587 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4590 len = OPEN_EXISTING;
4591 env->me_mfd = CreateFile(dpath, oflags,
4592 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4593 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4596 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4598 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4603 DPRINTF(("opened dbenv %p", (void *) env));
4605 rc = mdb_env_share_locks(env, &excl);
4609 if (!((flags & MDB_RDONLY) ||
4610 (env->me_pbuf = calloc(1, env->me_psize))))
4612 if (!(flags & MDB_RDONLY)) {
4614 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4615 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4616 txn = calloc(1, size);
4618 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4619 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4620 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4621 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4623 txn->mt_dbxs = env->me_dbxs;
4633 mdb_env_close0(env, excl);
4639 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4641 mdb_env_close0(MDB_env *env, int excl)
4645 if (!(env->me_flags & MDB_ENV_ACTIVE))
4648 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4649 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4650 free(env->me_dbxs[i].md_name.mv_data);
4653 free(env->me_dbiseqs);
4654 free(env->me_dbflags);
4657 free(env->me_dirty_list);
4659 mdb_midl_free(env->me_free_pgs);
4661 if (env->me_flags & MDB_ENV_TXKEY) {
4662 pthread_key_delete(env->me_txkey);
4664 /* Delete our key from the global list */
4665 for (i=0; i<mdb_tls_nkeys; i++)
4666 if (mdb_tls_keys[i] == env->me_txkey) {
4667 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4675 munmap(env->me_map, env->me_mapsize);
4677 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4678 (void) close(env->me_mfd);
4679 if (env->me_fd != INVALID_HANDLE_VALUE)
4680 (void) close(env->me_fd);
4682 MDB_PID_T pid = env->me_pid;
4683 /* Clearing readers is done in this function because
4684 * me_txkey with its destructor must be disabled first.
4686 for (i = env->me_numreaders; --i >= 0; )
4687 if (env->me_txns->mti_readers[i].mr_pid == pid)
4688 env->me_txns->mti_readers[i].mr_pid = 0;
4690 if (env->me_rmutex) {
4691 CloseHandle(env->me_rmutex);
4692 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4694 /* Windows automatically destroys the mutexes when
4695 * the last handle closes.
4697 #elif defined(MDB_USE_POSIX_SEM)
4698 if (env->me_rmutex != SEM_FAILED) {
4699 sem_close(env->me_rmutex);
4700 if (env->me_wmutex != SEM_FAILED)
4701 sem_close(env->me_wmutex);
4702 /* If we have the filelock: If we are the
4703 * only remaining user, clean up semaphores.
4706 mdb_env_excl_lock(env, &excl);
4708 sem_unlink(env->me_txns->mti_rmname);
4709 sem_unlink(env->me_txns->mti_wmname);
4713 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4715 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4718 /* Unlock the lockfile. Windows would have unlocked it
4719 * after closing anyway, but not necessarily at once.
4721 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4724 (void) close(env->me_lfd);
4727 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4732 mdb_env_close(MDB_env *env)
4739 VGMEMP_DESTROY(env);
4740 while ((dp = env->me_dpages) != NULL) {
4741 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4742 env->me_dpages = dp->mp_next;
4746 mdb_env_close0(env, 0);
4750 /** Compare two items pointing at aligned size_t's */
4752 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4754 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4755 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4758 /** Compare two items pointing at aligned unsigned int's */
4760 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4762 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4763 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4766 /** Compare two items pointing at unsigned ints of unknown alignment.
4767 * Nodes and keys are guaranteed to be 2-byte aligned.
4770 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4772 #if BYTE_ORDER == LITTLE_ENDIAN
4773 unsigned short *u, *c;
4776 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4777 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4780 } while(!x && u > (unsigned short *)a->mv_data);
4783 unsigned short *u, *c, *end;
4786 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4787 u = (unsigned short *)a->mv_data;
4788 c = (unsigned short *)b->mv_data;
4791 } while(!x && u < end);
4796 /** Compare two items pointing at size_t's of unknown alignment. */
4797 #ifdef MISALIGNED_OK
4798 # define mdb_cmp_clong mdb_cmp_long
4800 # define mdb_cmp_clong mdb_cmp_cint
4803 /** Compare two items lexically */
4805 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4812 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4818 diff = memcmp(a->mv_data, b->mv_data, len);
4819 return diff ? diff : len_diff<0 ? -1 : len_diff;
4822 /** Compare two items in reverse byte order */
4824 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4826 const unsigned char *p1, *p2, *p1_lim;
4830 p1_lim = (const unsigned char *)a->mv_data;
4831 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4832 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4834 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4840 while (p1 > p1_lim) {
4841 diff = *--p1 - *--p2;
4845 return len_diff<0 ? -1 : len_diff;
4848 /** Search for key within a page, using binary search.
4849 * Returns the smallest entry larger or equal to the key.
4850 * If exactp is non-null, stores whether the found entry was an exact match
4851 * in *exactp (1 or 0).
4852 * Updates the cursor index with the index of the found entry.
4853 * If no entry larger or equal to the key is found, returns NULL.
4856 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4858 unsigned int i = 0, nkeys;
4861 MDB_page *mp = mc->mc_pg[mc->mc_top];
4862 MDB_node *node = NULL;
4867 nkeys = NUMKEYS(mp);
4869 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4870 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4873 low = IS_LEAF(mp) ? 0 : 1;
4875 cmp = mc->mc_dbx->md_cmp;
4877 /* Branch pages have no data, so if using integer keys,
4878 * alignment is guaranteed. Use faster mdb_cmp_int.
4880 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4881 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4888 nodekey.mv_size = mc->mc_db->md_pad;
4889 node = NODEPTR(mp, 0); /* fake */
4890 while (low <= high) {
4891 i = (low + high) >> 1;
4892 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4893 rc = cmp(key, &nodekey);
4894 DPRINTF(("found leaf index %u [%s], rc = %i",
4895 i, DKEY(&nodekey), rc));
4904 while (low <= high) {
4905 i = (low + high) >> 1;
4907 node = NODEPTR(mp, i);
4908 nodekey.mv_size = NODEKSZ(node);
4909 nodekey.mv_data = NODEKEY(node);
4911 rc = cmp(key, &nodekey);
4914 DPRINTF(("found leaf index %u [%s], rc = %i",
4915 i, DKEY(&nodekey), rc));
4917 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4918 i, DKEY(&nodekey), NODEPGNO(node), rc));
4929 if (rc > 0) { /* Found entry is less than the key. */
4930 i++; /* Skip to get the smallest entry larger than key. */
4932 node = NODEPTR(mp, i);
4935 *exactp = (rc == 0 && nkeys > 0);
4936 /* store the key index */
4937 mc->mc_ki[mc->mc_top] = i;
4939 /* There is no entry larger or equal to the key. */
4942 /* nodeptr is fake for LEAF2 */
4948 mdb_cursor_adjust(MDB_cursor *mc, func)
4952 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4953 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4960 /** Pop a page off the top of the cursor's stack. */
4962 mdb_cursor_pop(MDB_cursor *mc)
4966 MDB_page *top = mc->mc_pg[mc->mc_top];
4972 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4973 DDBI(mc), (void *) mc));
4977 /** Push a page onto the top of the cursor's stack. */
4979 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4981 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4982 DDBI(mc), (void *) mc));
4984 if (mc->mc_snum >= CURSOR_STACK) {
4985 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4986 return MDB_CURSOR_FULL;
4989 mc->mc_top = mc->mc_snum++;
4990 mc->mc_pg[mc->mc_top] = mp;
4991 mc->mc_ki[mc->mc_top] = 0;
4996 /** Find the address of the page corresponding to a given page number.
4997 * @param[in] txn the transaction for this access.
4998 * @param[in] pgno the page number for the page to retrieve.
4999 * @param[out] ret address of a pointer where the page's address will be stored.
5000 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5001 * @return 0 on success, non-zero on failure.
5004 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5006 MDB_env *env = txn->mt_env;
5010 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
5014 MDB_ID2L dl = tx2->mt_u.dirty_list;
5016 /* Spilled pages were dirtied in this txn and flushed
5017 * because the dirty list got full. Bring this page
5018 * back in from the map (but don't unspill it here,
5019 * leave that unless page_touch happens again).
5021 if (tx2->mt_spill_pgs) {
5022 MDB_ID pn = pgno << 1;
5023 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5024 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5025 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5030 unsigned x = mdb_mid2l_search(dl, pgno);
5031 if (x <= dl[0].mid && dl[x].mid == pgno) {
5037 } while ((tx2 = tx2->mt_parent) != NULL);
5040 if (pgno < txn->mt_next_pgno) {
5042 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5044 DPRINTF(("page %"Z"u not found", pgno));
5045 txn->mt_flags |= MDB_TXN_ERROR;
5046 return MDB_PAGE_NOTFOUND;
5056 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5057 * The cursor is at the root page, set up the rest of it.
5060 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5062 MDB_page *mp = mc->mc_pg[mc->mc_top];
5066 while (IS_BRANCH(mp)) {
5070 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5071 mdb_cassert(mc, NUMKEYS(mp) > 1);
5072 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5074 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5076 if (flags & MDB_PS_LAST)
5077 i = NUMKEYS(mp) - 1;
5080 node = mdb_node_search(mc, key, &exact);
5082 i = NUMKEYS(mp) - 1;
5084 i = mc->mc_ki[mc->mc_top];
5086 mdb_cassert(mc, i > 0);
5090 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5093 mdb_cassert(mc, i < NUMKEYS(mp));
5094 node = NODEPTR(mp, i);
5096 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5099 mc->mc_ki[mc->mc_top] = i;
5100 if ((rc = mdb_cursor_push(mc, mp)))
5103 if (flags & MDB_PS_MODIFY) {
5104 if ((rc = mdb_page_touch(mc)) != 0)
5106 mp = mc->mc_pg[mc->mc_top];
5111 DPRINTF(("internal error, index points to a %02X page!?",
5113 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5114 return MDB_CORRUPTED;
5117 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5118 key ? DKEY(key) : "null"));
5119 mc->mc_flags |= C_INITIALIZED;
5120 mc->mc_flags &= ~C_EOF;
5125 /** Search for the lowest key under the current branch page.
5126 * This just bypasses a NUMKEYS check in the current page
5127 * before calling mdb_page_search_root(), because the callers
5128 * are all in situations where the current page is known to
5132 mdb_page_search_lowest(MDB_cursor *mc)
5134 MDB_page *mp = mc->mc_pg[mc->mc_top];
5135 MDB_node *node = NODEPTR(mp, 0);
5138 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5141 mc->mc_ki[mc->mc_top] = 0;
5142 if ((rc = mdb_cursor_push(mc, mp)))
5144 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5147 /** Search for the page a given key should be in.
5148 * Push it and its parent pages on the cursor stack.
5149 * @param[in,out] mc the cursor for this operation.
5150 * @param[in] key the key to search for, or NULL for first/last page.
5151 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5152 * are touched (updated with new page numbers).
5153 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5154 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5155 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5156 * @return 0 on success, non-zero on failure.
5159 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5164 /* Make sure the txn is still viable, then find the root from
5165 * the txn's db table and set it as the root of the cursor's stack.
5167 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5168 DPUTS("transaction has failed, must abort");
5171 /* Make sure we're using an up-to-date root */
5172 if (*mc->mc_dbflag & DB_STALE) {
5174 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5176 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5177 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5184 MDB_node *leaf = mdb_node_search(&mc2,
5185 &mc->mc_dbx->md_name, &exact);
5187 return MDB_NOTFOUND;
5188 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5191 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5193 /* The txn may not know this DBI, or another process may
5194 * have dropped and recreated the DB with other flags.
5196 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5197 return MDB_INCOMPATIBLE;
5198 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5200 *mc->mc_dbflag &= ~DB_STALE;
5202 root = mc->mc_db->md_root;
5204 if (root == P_INVALID) { /* Tree is empty. */
5205 DPUTS("tree is empty");
5206 return MDB_NOTFOUND;
5210 mdb_cassert(mc, root > 1);
5211 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5212 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5218 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5219 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5221 if (flags & MDB_PS_MODIFY) {
5222 if ((rc = mdb_page_touch(mc)))
5226 if (flags & MDB_PS_ROOTONLY)
5229 return mdb_page_search_root(mc, key, flags);
5233 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5235 MDB_txn *txn = mc->mc_txn;
5236 pgno_t pg = mp->mp_pgno;
5237 unsigned x = 0, ovpages = mp->mp_pages;
5238 MDB_env *env = txn->mt_env;
5239 MDB_IDL sl = txn->mt_spill_pgs;
5240 MDB_ID pn = pg << 1;
5243 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5244 /* If the page is dirty or on the spill list we just acquired it,
5245 * so we should give it back to our current free list, if any.
5246 * Otherwise put it onto the list of pages we freed in this txn.
5248 * Won't create me_pghead: me_pglast must be inited along with it.
5249 * Unsupported in nested txns: They would need to hide the page
5250 * range in ancestor txns' dirty and spilled lists.
5252 if (env->me_pghead &&
5254 ((mp->mp_flags & P_DIRTY) ||
5255 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5259 MDB_ID2 *dl, ix, iy;
5260 rc = mdb_midl_need(&env->me_pghead, ovpages);
5263 if (!(mp->mp_flags & P_DIRTY)) {
5264 /* This page is no longer spilled */
5271 /* Remove from dirty list */
5272 dl = txn->mt_u.dirty_list;
5274 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5280 mdb_cassert(mc, x > 1);
5282 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5283 txn->mt_flags |= MDB_TXN_ERROR;
5284 return MDB_CORRUPTED;
5287 if (!(env->me_flags & MDB_WRITEMAP))
5288 mdb_dpage_free(env, mp);
5290 /* Insert in me_pghead */
5291 mop = env->me_pghead;
5292 j = mop[0] + ovpages;
5293 for (i = mop[0]; i && mop[i] < pg; i--)
5299 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5303 mc->mc_db->md_overflow_pages -= ovpages;
5307 /** Return the data associated with a given node.
5308 * @param[in] txn The transaction for this operation.
5309 * @param[in] leaf The node being read.
5310 * @param[out] data Updated to point to the node's data.
5311 * @return 0 on success, non-zero on failure.
5314 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5316 MDB_page *omp; /* overflow page */
5320 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5321 data->mv_size = NODEDSZ(leaf);
5322 data->mv_data = NODEDATA(leaf);
5326 /* Read overflow data.
5328 data->mv_size = NODEDSZ(leaf);
5329 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5330 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5331 DPRINTF(("read overflow page %"Z"u failed", pgno));
5334 data->mv_data = METADATA(omp);
5340 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5341 MDB_val *key, MDB_val *data)
5348 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5350 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5353 if (txn->mt_flags & MDB_TXN_ERROR)
5356 mdb_cursor_init(&mc, txn, dbi, &mx);
5357 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5360 /** Find a sibling for a page.
5361 * Replaces the page at the top of the cursor's stack with the
5362 * specified sibling, if one exists.
5363 * @param[in] mc The cursor for this operation.
5364 * @param[in] move_right Non-zero if the right sibling is requested,
5365 * otherwise the left sibling.
5366 * @return 0 on success, non-zero on failure.
5369 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5375 if (mc->mc_snum < 2) {
5376 return MDB_NOTFOUND; /* root has no siblings */
5380 DPRINTF(("parent page is page %"Z"u, index %u",
5381 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5383 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5384 : (mc->mc_ki[mc->mc_top] == 0)) {
5385 DPRINTF(("no more keys left, moving to %s sibling",
5386 move_right ? "right" : "left"));
5387 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5388 /* undo cursor_pop before returning */
5395 mc->mc_ki[mc->mc_top]++;
5397 mc->mc_ki[mc->mc_top]--;
5398 DPRINTF(("just moving to %s index key %u",
5399 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5401 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5403 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5404 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5405 /* mc will be inconsistent if caller does mc_snum++ as above */
5406 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5410 mdb_cursor_push(mc, mp);
5412 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5417 /** Move the cursor to the next data item. */
5419 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5425 if (mc->mc_flags & C_EOF) {
5426 return MDB_NOTFOUND;
5429 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5431 mp = mc->mc_pg[mc->mc_top];
5433 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5434 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5435 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5436 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5437 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5438 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5439 if (rc == MDB_SUCCESS)
5440 MDB_GET_KEY(leaf, key);
5445 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5446 if (op == MDB_NEXT_DUP)
5447 return MDB_NOTFOUND;
5451 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5452 mdb_dbg_pgno(mp), (void *) mc));
5453 if (mc->mc_flags & C_DEL)
5456 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5457 DPUTS("=====> move to next sibling page");
5458 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5459 mc->mc_flags |= C_EOF;
5462 mp = mc->mc_pg[mc->mc_top];
5463 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5465 mc->mc_ki[mc->mc_top]++;
5468 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5469 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5472 key->mv_size = mc->mc_db->md_pad;
5473 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5477 mdb_cassert(mc, IS_LEAF(mp));
5478 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5480 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5481 mdb_xcursor_init1(mc, leaf);
5484 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5487 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5488 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5489 if (rc != MDB_SUCCESS)
5494 MDB_GET_KEY(leaf, key);
5498 /** Move the cursor to the previous data item. */
5500 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5506 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5508 mp = mc->mc_pg[mc->mc_top];
5510 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5511 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5512 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5513 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5514 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5515 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5516 if (rc == MDB_SUCCESS) {
5517 MDB_GET_KEY(leaf, key);
5518 mc->mc_flags &= ~C_EOF;
5524 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5525 if (op == MDB_PREV_DUP)
5526 return MDB_NOTFOUND;
5530 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5531 mdb_dbg_pgno(mp), (void *) mc));
5533 if (mc->mc_ki[mc->mc_top] == 0) {
5534 DPUTS("=====> move to prev sibling page");
5535 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5538 mp = mc->mc_pg[mc->mc_top];
5539 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5540 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5542 mc->mc_ki[mc->mc_top]--;
5544 mc->mc_flags &= ~C_EOF;
5546 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5547 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5550 key->mv_size = mc->mc_db->md_pad;
5551 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5555 mdb_cassert(mc, IS_LEAF(mp));
5556 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5558 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5559 mdb_xcursor_init1(mc, leaf);
5562 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5565 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5566 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5567 if (rc != MDB_SUCCESS)
5572 MDB_GET_KEY(leaf, key);
5576 /** Set the cursor on a specific data item. */
5578 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5579 MDB_cursor_op op, int *exactp)
5583 MDB_node *leaf = NULL;
5586 if (key->mv_size == 0)
5587 return MDB_BAD_VALSIZE;
5590 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5592 /* See if we're already on the right page */
5593 if (mc->mc_flags & C_INITIALIZED) {
5596 mp = mc->mc_pg[mc->mc_top];
5598 mc->mc_ki[mc->mc_top] = 0;
5599 return MDB_NOTFOUND;
5601 if (mp->mp_flags & P_LEAF2) {
5602 nodekey.mv_size = mc->mc_db->md_pad;
5603 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5605 leaf = NODEPTR(mp, 0);
5606 MDB_GET_KEY2(leaf, nodekey);
5608 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5610 /* Probably happens rarely, but first node on the page
5611 * was the one we wanted.
5613 mc->mc_ki[mc->mc_top] = 0;
5620 unsigned int nkeys = NUMKEYS(mp);
5622 if (mp->mp_flags & P_LEAF2) {
5623 nodekey.mv_data = LEAF2KEY(mp,
5624 nkeys-1, nodekey.mv_size);
5626 leaf = NODEPTR(mp, nkeys-1);
5627 MDB_GET_KEY2(leaf, nodekey);
5629 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5631 /* last node was the one we wanted */
5632 mc->mc_ki[mc->mc_top] = nkeys-1;
5638 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5639 /* This is definitely the right page, skip search_page */
5640 if (mp->mp_flags & P_LEAF2) {
5641 nodekey.mv_data = LEAF2KEY(mp,
5642 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5644 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5645 MDB_GET_KEY2(leaf, nodekey);
5647 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5649 /* current node was the one we wanted */
5659 /* If any parents have right-sibs, search.
5660 * Otherwise, there's nothing further.
5662 for (i=0; i<mc->mc_top; i++)
5664 NUMKEYS(mc->mc_pg[i])-1)
5666 if (i == mc->mc_top) {
5667 /* There are no other pages */
5668 mc->mc_ki[mc->mc_top] = nkeys;
5669 return MDB_NOTFOUND;
5673 /* There are no other pages */
5674 mc->mc_ki[mc->mc_top] = 0;
5675 if (op == MDB_SET_RANGE && !exactp) {
5679 return MDB_NOTFOUND;
5683 rc = mdb_page_search(mc, key, 0);
5684 if (rc != MDB_SUCCESS)
5687 mp = mc->mc_pg[mc->mc_top];
5688 mdb_cassert(mc, IS_LEAF(mp));
5691 leaf = mdb_node_search(mc, key, exactp);
5692 if (exactp != NULL && !*exactp) {
5693 /* MDB_SET specified and not an exact match. */
5694 return MDB_NOTFOUND;
5698 DPUTS("===> inexact leaf not found, goto sibling");
5699 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5700 return rc; /* no entries matched */
5701 mp = mc->mc_pg[mc->mc_top];
5702 mdb_cassert(mc, IS_LEAF(mp));
5703 leaf = NODEPTR(mp, 0);
5707 mc->mc_flags |= C_INITIALIZED;
5708 mc->mc_flags &= ~C_EOF;
5711 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5712 key->mv_size = mc->mc_db->md_pad;
5713 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5718 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5719 mdb_xcursor_init1(mc, leaf);
5722 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5723 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5724 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5727 if (op == MDB_GET_BOTH) {
5733 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5734 if (rc != MDB_SUCCESS)
5737 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5739 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5741 rc = mc->mc_dbx->md_dcmp(data, &d2);
5743 if (op == MDB_GET_BOTH || rc > 0)
5744 return MDB_NOTFOUND;
5751 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5752 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5757 /* The key already matches in all other cases */
5758 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5759 MDB_GET_KEY(leaf, key);
5760 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5765 /** Move the cursor to the first item in the database. */
5767 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5773 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5775 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5776 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5777 if (rc != MDB_SUCCESS)
5780 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5782 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5783 mc->mc_flags |= C_INITIALIZED;
5784 mc->mc_flags &= ~C_EOF;
5786 mc->mc_ki[mc->mc_top] = 0;
5788 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5789 key->mv_size = mc->mc_db->md_pad;
5790 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5795 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5796 mdb_xcursor_init1(mc, leaf);
5797 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5801 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5805 MDB_GET_KEY(leaf, key);
5809 /** Move the cursor to the last item in the database. */
5811 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5817 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5819 if (!(mc->mc_flags & C_EOF)) {
5821 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5822 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5823 if (rc != MDB_SUCCESS)
5826 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5829 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5830 mc->mc_flags |= C_INITIALIZED|C_EOF;
5831 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5833 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5834 key->mv_size = mc->mc_db->md_pad;
5835 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5840 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5841 mdb_xcursor_init1(mc, leaf);
5842 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5846 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5851 MDB_GET_KEY(leaf, key);
5856 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5861 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5866 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5870 case MDB_GET_CURRENT:
5871 if (!(mc->mc_flags & C_INITIALIZED)) {
5874 MDB_page *mp = mc->mc_pg[mc->mc_top];
5875 int nkeys = NUMKEYS(mp);
5876 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5877 mc->mc_ki[mc->mc_top] = nkeys;
5883 key->mv_size = mc->mc_db->md_pad;
5884 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5886 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5887 MDB_GET_KEY(leaf, key);
5889 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5890 if (mc->mc_flags & C_DEL)
5891 mdb_xcursor_init1(mc, leaf);
5892 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5894 rc = mdb_node_read(mc->mc_txn, leaf, data);
5901 case MDB_GET_BOTH_RANGE:
5906 if (mc->mc_xcursor == NULL) {
5907 rc = MDB_INCOMPATIBLE;
5917 rc = mdb_cursor_set(mc, key, data, op,
5918 op == MDB_SET_RANGE ? NULL : &exact);
5921 case MDB_GET_MULTIPLE:
5922 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5926 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5927 rc = MDB_INCOMPATIBLE;
5931 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5932 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5935 case MDB_NEXT_MULTIPLE:
5940 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5941 rc = MDB_INCOMPATIBLE;
5944 if (!(mc->mc_flags & C_INITIALIZED))
5945 rc = mdb_cursor_first(mc, key, data);
5947 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5948 if (rc == MDB_SUCCESS) {
5949 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5952 mx = &mc->mc_xcursor->mx_cursor;
5953 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5955 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5956 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5964 case MDB_NEXT_NODUP:
5965 if (!(mc->mc_flags & C_INITIALIZED))
5966 rc = mdb_cursor_first(mc, key, data);
5968 rc = mdb_cursor_next(mc, key, data, op);
5972 case MDB_PREV_NODUP:
5973 if (!(mc->mc_flags & C_INITIALIZED)) {
5974 rc = mdb_cursor_last(mc, key, data);
5977 mc->mc_flags |= C_INITIALIZED;
5978 mc->mc_ki[mc->mc_top]++;
5980 rc = mdb_cursor_prev(mc, key, data, op);
5983 rc = mdb_cursor_first(mc, key, data);
5986 mfunc = mdb_cursor_first;
5988 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5992 if (mc->mc_xcursor == NULL) {
5993 rc = MDB_INCOMPATIBLE;
5997 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5998 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5999 MDB_GET_KEY(leaf, key);
6000 rc = mdb_node_read(mc->mc_txn, leaf, data);
6004 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6008 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6011 rc = mdb_cursor_last(mc, key, data);
6014 mfunc = mdb_cursor_last;
6017 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6022 if (mc->mc_flags & C_DEL)
6023 mc->mc_flags ^= C_DEL;
6028 /** Touch all the pages in the cursor stack. Set mc_top.
6029 * Makes sure all the pages are writable, before attempting a write operation.
6030 * @param[in] mc The cursor to operate on.
6033 mdb_cursor_touch(MDB_cursor *mc)
6035 int rc = MDB_SUCCESS;
6037 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6040 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6042 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6043 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6046 *mc->mc_dbflag |= DB_DIRTY;
6051 rc = mdb_page_touch(mc);
6052 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6053 mc->mc_top = mc->mc_snum-1;
6058 /** Do not spill pages to disk if txn is getting full, may fail instead */
6059 #define MDB_NOSPILL 0x8000
6062 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6065 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
6067 MDB_node *leaf = NULL;
6070 MDB_val xdata, *rdata, dkey, olddata;
6072 int do_sub = 0, insert_key, insert_data;
6073 unsigned int mcount = 0, dcount = 0, nospill;
6076 unsigned int nflags;
6079 if (mc == NULL || key == NULL)
6082 env = mc->mc_txn->mt_env;
6084 /* Check this first so counter will always be zero on any
6087 if (flags & MDB_MULTIPLE) {
6088 dcount = data[1].mv_size;
6089 data[1].mv_size = 0;
6090 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6091 return MDB_INCOMPATIBLE;
6094 nospill = flags & MDB_NOSPILL;
6095 flags &= ~MDB_NOSPILL;
6097 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6098 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6100 if (key->mv_size-1 >= ENV_MAXKEY(env))
6101 return MDB_BAD_VALSIZE;
6103 #if SIZE_MAX > MAXDATASIZE
6104 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6105 return MDB_BAD_VALSIZE;
6107 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6108 return MDB_BAD_VALSIZE;
6111 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6112 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6116 if (flags == MDB_CURRENT) {
6117 if (!(mc->mc_flags & C_INITIALIZED))
6120 } else if (mc->mc_db->md_root == P_INVALID) {
6121 /* new database, cursor has nothing to point to */
6124 mc->mc_flags &= ~C_INITIALIZED;
6129 if (flags & MDB_APPEND) {
6131 rc = mdb_cursor_last(mc, &k2, &d2);
6133 rc = mc->mc_dbx->md_cmp(key, &k2);
6136 mc->mc_ki[mc->mc_top]++;
6138 /* new key is <= last key */
6143 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6145 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6146 DPRINTF(("duplicate key [%s]", DKEY(key)));
6148 return MDB_KEYEXIST;
6150 if (rc && rc != MDB_NOTFOUND)
6154 if (mc->mc_flags & C_DEL)
6155 mc->mc_flags ^= C_DEL;
6157 /* Cursor is positioned, check for room in the dirty list */
6159 if (flags & MDB_MULTIPLE) {
6161 xdata.mv_size = data->mv_size * dcount;
6165 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6169 if (rc == MDB_NO_ROOT) {
6171 /* new database, write a root leaf page */
6172 DPUTS("allocating new root leaf page");
6173 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6176 mdb_cursor_push(mc, np);
6177 mc->mc_db->md_root = np->mp_pgno;
6178 mc->mc_db->md_depth++;
6179 *mc->mc_dbflag |= DB_DIRTY;
6180 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6182 np->mp_flags |= P_LEAF2;
6183 mc->mc_flags |= C_INITIALIZED;
6185 /* make sure all cursor pages are writable */
6186 rc2 = mdb_cursor_touch(mc);
6191 insert_key = insert_data = rc;
6193 /* The key does not exist */
6194 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6195 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6196 LEAFSIZE(key, data) > env->me_nodemax)
6198 /* Too big for a node, insert in sub-DB. Set up an empty
6199 * "old sub-page" for prep_subDB to expand to a full page.
6201 fp_flags = P_LEAF|P_DIRTY;
6203 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6204 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6205 olddata.mv_size = PAGEHDRSZ;
6209 /* there's only a key anyway, so this is a no-op */
6210 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6212 unsigned int ksize = mc->mc_db->md_pad;
6213 if (key->mv_size != ksize)
6214 return MDB_BAD_VALSIZE;
6215 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6216 memcpy(ptr, key->mv_data, ksize);
6218 /* if overwriting slot 0 of leaf, need to
6219 * update branch key if there is a parent page
6221 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6222 unsigned short top = mc->mc_top;
6224 /* slot 0 is always an empty key, find real slot */
6225 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6227 if (mc->mc_ki[mc->mc_top])
6228 rc2 = mdb_update_key(mc, key);
6239 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6240 olddata.mv_size = NODEDSZ(leaf);
6241 olddata.mv_data = NODEDATA(leaf);
6244 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6245 /* Prepare (sub-)page/sub-DB to accept the new item,
6246 * if needed. fp: old sub-page or a header faking
6247 * it. mp: new (sub-)page. offset: growth in page
6248 * size. xdata: node data with new page or DB.
6250 unsigned i, offset = 0;
6251 mp = fp = xdata.mv_data = env->me_pbuf;
6252 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6254 /* Was a single item before, must convert now */
6255 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6256 /* Just overwrite the current item */
6257 if (flags == MDB_CURRENT)
6260 #if UINT_MAX < SIZE_MAX
6261 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6262 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6264 /* does data match? */
6265 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6266 if (flags & MDB_NODUPDATA)
6267 return MDB_KEYEXIST;
6272 /* Back up original data item */
6273 dkey.mv_size = olddata.mv_size;
6274 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6276 /* Make sub-page header for the dup items, with dummy body */
6277 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6278 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6279 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6280 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6281 fp->mp_flags |= P_LEAF2;
6282 fp->mp_pad = data->mv_size;
6283 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6285 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6286 (dkey.mv_size & 1) + (data->mv_size & 1);
6288 fp->mp_upper = xdata.mv_size - PAGEBASE;
6289 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6290 } else if (leaf->mn_flags & F_SUBDATA) {
6291 /* Data is on sub-DB, just store it */
6292 flags |= F_DUPDATA|F_SUBDATA;
6295 /* Data is on sub-page */
6296 fp = olddata.mv_data;
6299 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6300 offset = EVEN(NODESIZE + sizeof(indx_t) +
6304 offset = fp->mp_pad;
6305 if (SIZELEFT(fp) < offset) {
6306 offset *= 4; /* space for 4 more */
6309 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6311 fp->mp_flags |= P_DIRTY;
6312 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6313 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6317 xdata.mv_size = olddata.mv_size + offset;
6320 fp_flags = fp->mp_flags;
6321 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6322 /* Too big for a sub-page, convert to sub-DB */
6323 fp_flags &= ~P_SUBP;
6325 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6326 fp_flags |= P_LEAF2;
6327 dummy.md_pad = fp->mp_pad;
6328 dummy.md_flags = MDB_DUPFIXED;
6329 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6330 dummy.md_flags |= MDB_INTEGERKEY;
6336 dummy.md_branch_pages = 0;
6337 dummy.md_leaf_pages = 1;
6338 dummy.md_overflow_pages = 0;
6339 dummy.md_entries = NUMKEYS(fp);
6340 xdata.mv_size = sizeof(MDB_db);
6341 xdata.mv_data = &dummy;
6342 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6344 offset = env->me_psize - olddata.mv_size;
6345 flags |= F_DUPDATA|F_SUBDATA;
6346 dummy.md_root = mp->mp_pgno;
6349 mp->mp_flags = fp_flags | P_DIRTY;
6350 mp->mp_pad = fp->mp_pad;
6351 mp->mp_lower = fp->mp_lower;
6352 mp->mp_upper = fp->mp_upper + offset;
6353 if (fp_flags & P_LEAF2) {
6354 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6356 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6357 olddata.mv_size - fp->mp_upper - PAGEBASE);
6358 for (i=0; i<NUMKEYS(fp); i++)
6359 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6367 mdb_node_del(mc, 0);
6371 /* overflow page overwrites need special handling */
6372 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6375 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6377 memcpy(&pg, olddata.mv_data, sizeof(pg));
6378 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6380 ovpages = omp->mp_pages;
6382 /* Is the ov page large enough? */
6383 if (ovpages >= dpages) {
6384 if (!(omp->mp_flags & P_DIRTY) &&
6385 (level || (env->me_flags & MDB_WRITEMAP)))
6387 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6390 level = 0; /* dirty in this txn or clean */
6393 if (omp->mp_flags & P_DIRTY) {
6394 /* yes, overwrite it. Note in this case we don't
6395 * bother to try shrinking the page if the new data
6396 * is smaller than the overflow threshold.
6399 /* It is writable only in a parent txn */
6400 size_t sz = (size_t) env->me_psize * ovpages, off;
6401 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6407 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6408 mdb_cassert(mc, rc2 == 0);
6409 if (!(flags & MDB_RESERVE)) {
6410 /* Copy end of page, adjusting alignment so
6411 * compiler may copy words instead of bytes.
6413 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6414 memcpy((size_t *)((char *)np + off),
6415 (size_t *)((char *)omp + off), sz - off);
6418 memcpy(np, omp, sz); /* Copy beginning of page */
6421 SETDSZ(leaf, data->mv_size);
6422 if (F_ISSET(flags, MDB_RESERVE))
6423 data->mv_data = METADATA(omp);
6425 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6429 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6431 } else if (data->mv_size == olddata.mv_size) {
6432 /* same size, just replace it. Note that we could
6433 * also reuse this node if the new data is smaller,
6434 * but instead we opt to shrink the node in that case.
6436 if (F_ISSET(flags, MDB_RESERVE))
6437 data->mv_data = olddata.mv_data;
6438 else if (!(mc->mc_flags & C_SUB))
6439 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6441 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6446 mdb_node_del(mc, 0);
6452 nflags = flags & NODE_ADD_FLAGS;
6453 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6454 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6455 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6456 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6458 nflags |= MDB_SPLIT_REPLACE;
6459 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6461 /* There is room already in this leaf page. */
6462 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6463 if (rc == 0 && insert_key) {
6464 /* Adjust other cursors pointing to mp */
6465 MDB_cursor *m2, *m3;
6466 MDB_dbi dbi = mc->mc_dbi;
6467 unsigned i = mc->mc_top;
6468 MDB_page *mp = mc->mc_pg[i];
6470 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6471 if (mc->mc_flags & C_SUB)
6472 m3 = &m2->mc_xcursor->mx_cursor;
6475 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6476 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6483 if (rc == MDB_SUCCESS) {
6484 /* Now store the actual data in the child DB. Note that we're
6485 * storing the user data in the keys field, so there are strict
6486 * size limits on dupdata. The actual data fields of the child
6487 * DB are all zero size.
6495 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6496 if (flags & MDB_CURRENT) {
6497 xflags = MDB_CURRENT|MDB_NOSPILL;
6499 mdb_xcursor_init1(mc, leaf);
6500 xflags = (flags & MDB_NODUPDATA) ?
6501 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6503 /* converted, write the original data first */
6505 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6509 /* Adjust other cursors pointing to mp */
6511 unsigned i = mc->mc_top;
6512 MDB_page *mp = mc->mc_pg[i];
6514 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6515 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6516 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6517 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6518 mdb_xcursor_init1(m2, leaf);
6522 /* we've done our job */
6525 ecount = mc->mc_xcursor->mx_db.md_entries;
6526 if (flags & MDB_APPENDDUP)
6527 xflags |= MDB_APPEND;
6528 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6529 if (flags & F_SUBDATA) {
6530 void *db = NODEDATA(leaf);
6531 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6533 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6535 /* Increment count unless we just replaced an existing item. */
6537 mc->mc_db->md_entries++;
6539 /* Invalidate txn if we created an empty sub-DB */
6542 /* If we succeeded and the key didn't exist before,
6543 * make sure the cursor is marked valid.
6545 mc->mc_flags |= C_INITIALIZED;
6547 if (flags & MDB_MULTIPLE) {
6550 /* let caller know how many succeeded, if any */
6551 data[1].mv_size = mcount;
6552 if (mcount < dcount) {
6553 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6554 insert_key = insert_data = 0;
6561 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6564 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6569 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6575 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6576 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6578 if (!(mc->mc_flags & C_INITIALIZED))
6581 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6582 return MDB_NOTFOUND;
6584 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6587 rc = mdb_cursor_touch(mc);
6591 mp = mc->mc_pg[mc->mc_top];
6594 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6596 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6597 if (flags & MDB_NODUPDATA) {
6598 /* mdb_cursor_del0() will subtract the final entry */
6599 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6601 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6602 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6604 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6607 /* If sub-DB still has entries, we're done */
6608 if (mc->mc_xcursor->mx_db.md_entries) {
6609 if (leaf->mn_flags & F_SUBDATA) {
6610 /* update subDB info */
6611 void *db = NODEDATA(leaf);
6612 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6615 /* shrink fake page */
6616 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6617 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6618 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6619 /* fix other sub-DB cursors pointed at this fake page */
6620 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6621 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6622 if (m2->mc_pg[mc->mc_top] == mp &&
6623 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6624 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6627 mc->mc_db->md_entries--;
6628 mc->mc_flags |= C_DEL;
6631 /* otherwise fall thru and delete the sub-DB */
6634 if (leaf->mn_flags & F_SUBDATA) {
6635 /* add all the child DB's pages to the free list */
6636 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6642 /* add overflow pages to free list */
6643 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6647 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6648 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6649 (rc = mdb_ovpage_free(mc, omp)))
6654 return mdb_cursor_del0(mc);
6657 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6661 /** Allocate and initialize new pages for a database.
6662 * @param[in] mc a cursor on the database being added to.
6663 * @param[in] flags flags defining what type of page is being allocated.
6664 * @param[in] num the number of pages to allocate. This is usually 1,
6665 * unless allocating overflow pages for a large record.
6666 * @param[out] mp Address of a page, or NULL on failure.
6667 * @return 0 on success, non-zero on failure.
6670 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6675 if ((rc = mdb_page_alloc(mc, num, &np)))
6677 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6678 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6679 np->mp_flags = flags | P_DIRTY;
6680 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6681 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6684 mc->mc_db->md_branch_pages++;
6685 else if (IS_LEAF(np))
6686 mc->mc_db->md_leaf_pages++;
6687 else if (IS_OVERFLOW(np)) {
6688 mc->mc_db->md_overflow_pages += num;
6696 /** Calculate the size of a leaf node.
6697 * The size depends on the environment's page size; if a data item
6698 * is too large it will be put onto an overflow page and the node
6699 * size will only include the key and not the data. Sizes are always
6700 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6701 * of the #MDB_node headers.
6702 * @param[in] env The environment handle.
6703 * @param[in] key The key for the node.
6704 * @param[in] data The data for the node.
6705 * @return The number of bytes needed to store the node.
6708 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6712 sz = LEAFSIZE(key, data);
6713 if (sz > env->me_nodemax) {
6714 /* put on overflow page */
6715 sz -= data->mv_size - sizeof(pgno_t);
6718 return EVEN(sz + sizeof(indx_t));
6721 /** Calculate the size of a branch node.
6722 * The size should depend on the environment's page size but since
6723 * we currently don't support spilling large keys onto overflow
6724 * pages, it's simply the size of the #MDB_node header plus the
6725 * size of the key. Sizes are always rounded up to an even number
6726 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6727 * @param[in] env The environment handle.
6728 * @param[in] key The key for the node.
6729 * @return The number of bytes needed to store the node.
6732 mdb_branch_size(MDB_env *env, MDB_val *key)
6737 if (sz > env->me_nodemax) {
6738 /* put on overflow page */
6739 /* not implemented */
6740 /* sz -= key->size - sizeof(pgno_t); */
6743 return sz + sizeof(indx_t);
6746 /** Add a node to the page pointed to by the cursor.
6747 * @param[in] mc The cursor for this operation.
6748 * @param[in] indx The index on the page where the new node should be added.
6749 * @param[in] key The key for the new node.
6750 * @param[in] data The data for the new node, if any.
6751 * @param[in] pgno The page number, if adding a branch node.
6752 * @param[in] flags Flags for the node.
6753 * @return 0 on success, non-zero on failure. Possible errors are:
6755 * <li>ENOMEM - failed to allocate overflow pages for the node.
6756 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6757 * should never happen since all callers already calculate the
6758 * page's free space before calling this function.
6762 mdb_node_add(MDB_cursor *mc, indx_t indx,
6763 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6766 size_t node_size = NODESIZE;
6770 MDB_page *mp = mc->mc_pg[mc->mc_top];
6771 MDB_page *ofp = NULL; /* overflow page */
6774 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6776 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6777 IS_LEAF(mp) ? "leaf" : "branch",
6778 IS_SUBP(mp) ? "sub-" : "",
6779 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6780 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6783 /* Move higher keys up one slot. */
6784 int ksize = mc->mc_db->md_pad, dif;
6785 char *ptr = LEAF2KEY(mp, indx, ksize);
6786 dif = NUMKEYS(mp) - indx;
6788 memmove(ptr+ksize, ptr, dif*ksize);
6789 /* insert new key */
6790 memcpy(ptr, key->mv_data, ksize);
6792 /* Just using these for counting */
6793 mp->mp_lower += sizeof(indx_t);
6794 mp->mp_upper -= ksize - sizeof(indx_t);
6798 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6800 node_size += key->mv_size;
6802 mdb_cassert(mc, data);
6803 if (F_ISSET(flags, F_BIGDATA)) {
6804 /* Data already on overflow page. */
6805 node_size += sizeof(pgno_t);
6806 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6807 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6809 /* Put data on overflow page. */
6810 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6811 data->mv_size, node_size+data->mv_size));
6812 node_size = EVEN(node_size + sizeof(pgno_t));
6813 if ((ssize_t)node_size > room)
6815 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6817 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6821 node_size += data->mv_size;
6824 node_size = EVEN(node_size);
6825 if ((ssize_t)node_size > room)
6829 /* Move higher pointers up one slot. */
6830 for (i = NUMKEYS(mp); i > indx; i--)
6831 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6833 /* Adjust free space offsets. */
6834 ofs = mp->mp_upper - node_size;
6835 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6836 mp->mp_ptrs[indx] = ofs;
6838 mp->mp_lower += sizeof(indx_t);
6840 /* Write the node data. */
6841 node = NODEPTR(mp, indx);
6842 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6843 node->mn_flags = flags;
6845 SETDSZ(node,data->mv_size);
6850 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6853 mdb_cassert(mc, key);
6855 if (F_ISSET(flags, F_BIGDATA))
6856 memcpy(node->mn_data + key->mv_size, data->mv_data,
6858 else if (F_ISSET(flags, MDB_RESERVE))
6859 data->mv_data = node->mn_data + key->mv_size;
6861 memcpy(node->mn_data + key->mv_size, data->mv_data,
6864 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6866 if (F_ISSET(flags, MDB_RESERVE))
6867 data->mv_data = METADATA(ofp);
6869 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6876 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6877 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6878 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6879 DPRINTF(("node size = %"Z"u", node_size));
6880 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6881 return MDB_PAGE_FULL;
6884 /** Delete the specified node from a page.
6885 * @param[in] mc Cursor pointing to the node to delete.
6886 * @param[in] ksize The size of a node. Only used if the page is
6887 * part of a #MDB_DUPFIXED database.
6890 mdb_node_del(MDB_cursor *mc, int ksize)
6892 MDB_page *mp = mc->mc_pg[mc->mc_top];
6893 indx_t indx = mc->mc_ki[mc->mc_top];
6895 indx_t i, j, numkeys, ptr;
6899 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6900 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6901 numkeys = NUMKEYS(mp);
6902 mdb_cassert(mc, indx < numkeys);
6905 int x = numkeys - 1 - indx;
6906 base = LEAF2KEY(mp, indx, ksize);
6908 memmove(base, base + ksize, x * ksize);
6909 mp->mp_lower -= sizeof(indx_t);
6910 mp->mp_upper += ksize - sizeof(indx_t);
6914 node = NODEPTR(mp, indx);
6915 sz = NODESIZE + node->mn_ksize;
6917 if (F_ISSET(node->mn_flags, F_BIGDATA))
6918 sz += sizeof(pgno_t);
6920 sz += NODEDSZ(node);
6924 ptr = mp->mp_ptrs[indx];
6925 for (i = j = 0; i < numkeys; i++) {
6927 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6928 if (mp->mp_ptrs[i] < ptr)
6929 mp->mp_ptrs[j] += sz;
6934 base = (char *)mp + mp->mp_upper + PAGEBASE;
6935 memmove(base + sz, base, ptr - mp->mp_upper);
6937 mp->mp_lower -= sizeof(indx_t);
6941 /** Compact the main page after deleting a node on a subpage.
6942 * @param[in] mp The main page to operate on.
6943 * @param[in] indx The index of the subpage on the main page.
6946 mdb_node_shrink(MDB_page *mp, indx_t indx)
6952 indx_t i, numkeys, ptr;
6954 node = NODEPTR(mp, indx);
6955 sp = (MDB_page *)NODEDATA(node);
6956 delta = SIZELEFT(sp);
6957 xp = (MDB_page *)((char *)sp + delta);
6959 /* shift subpage upward */
6961 nsize = NUMKEYS(sp) * sp->mp_pad;
6963 return; /* do not make the node uneven-sized */
6964 memmove(METADATA(xp), METADATA(sp), nsize);
6967 numkeys = NUMKEYS(sp);
6968 for (i=numkeys-1; i>=0; i--)
6969 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6971 xp->mp_upper = sp->mp_lower;
6972 xp->mp_lower = sp->mp_lower;
6973 xp->mp_flags = sp->mp_flags;
6974 xp->mp_pad = sp->mp_pad;
6975 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6977 nsize = NODEDSZ(node) - delta;
6978 SETDSZ(node, nsize);
6980 /* shift lower nodes upward */
6981 ptr = mp->mp_ptrs[indx];
6982 numkeys = NUMKEYS(mp);
6983 for (i = 0; i < numkeys; i++) {
6984 if (mp->mp_ptrs[i] <= ptr)
6985 mp->mp_ptrs[i] += delta;
6988 base = (char *)mp + mp->mp_upper + PAGEBASE;
6989 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6990 mp->mp_upper += delta;
6993 /** Initial setup of a sorted-dups cursor.
6994 * Sorted duplicates are implemented as a sub-database for the given key.
6995 * The duplicate data items are actually keys of the sub-database.
6996 * Operations on the duplicate data items are performed using a sub-cursor
6997 * initialized when the sub-database is first accessed. This function does
6998 * the preliminary setup of the sub-cursor, filling in the fields that
6999 * depend only on the parent DB.
7000 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7003 mdb_xcursor_init0(MDB_cursor *mc)
7005 MDB_xcursor *mx = mc->mc_xcursor;
7007 mx->mx_cursor.mc_xcursor = NULL;
7008 mx->mx_cursor.mc_txn = mc->mc_txn;
7009 mx->mx_cursor.mc_db = &mx->mx_db;
7010 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7011 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7012 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7013 mx->mx_cursor.mc_snum = 0;
7014 mx->mx_cursor.mc_top = 0;
7015 mx->mx_cursor.mc_flags = C_SUB;
7016 mx->mx_dbx.md_name.mv_size = 0;
7017 mx->mx_dbx.md_name.mv_data = NULL;
7018 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7019 mx->mx_dbx.md_dcmp = NULL;
7020 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7023 /** Final setup of a sorted-dups cursor.
7024 * Sets up the fields that depend on the data from the main cursor.
7025 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7026 * @param[in] node The data containing the #MDB_db record for the
7027 * sorted-dup database.
7030 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7032 MDB_xcursor *mx = mc->mc_xcursor;
7034 if (node->mn_flags & F_SUBDATA) {
7035 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7036 mx->mx_cursor.mc_pg[0] = 0;
7037 mx->mx_cursor.mc_snum = 0;
7038 mx->mx_cursor.mc_top = 0;
7039 mx->mx_cursor.mc_flags = C_SUB;
7041 MDB_page *fp = NODEDATA(node);
7042 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7043 mx->mx_db.md_flags = 0;
7044 mx->mx_db.md_depth = 1;
7045 mx->mx_db.md_branch_pages = 0;
7046 mx->mx_db.md_leaf_pages = 1;
7047 mx->mx_db.md_overflow_pages = 0;
7048 mx->mx_db.md_entries = NUMKEYS(fp);
7049 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7050 mx->mx_cursor.mc_snum = 1;
7051 mx->mx_cursor.mc_top = 0;
7052 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7053 mx->mx_cursor.mc_pg[0] = fp;
7054 mx->mx_cursor.mc_ki[0] = 0;
7055 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7056 mx->mx_db.md_flags = MDB_DUPFIXED;
7057 mx->mx_db.md_pad = fp->mp_pad;
7058 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7059 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7062 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7063 mx->mx_db.md_root));
7064 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7065 #if UINT_MAX < SIZE_MAX
7066 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7067 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7071 /** Initialize a cursor for a given transaction and database. */
7073 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7076 mc->mc_backup = NULL;
7079 mc->mc_db = &txn->mt_dbs[dbi];
7080 mc->mc_dbx = &txn->mt_dbxs[dbi];
7081 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7086 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7087 mdb_tassert(txn, mx != NULL);
7088 mc->mc_xcursor = mx;
7089 mdb_xcursor_init0(mc);
7091 mc->mc_xcursor = NULL;
7093 if (*mc->mc_dbflag & DB_STALE) {
7094 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7099 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7102 size_t size = sizeof(MDB_cursor);
7104 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7107 if (txn->mt_flags & MDB_TXN_ERROR)
7110 /* Allow read access to the freelist */
7111 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7114 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7115 size += sizeof(MDB_xcursor);
7117 if ((mc = malloc(size)) != NULL) {
7118 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7119 if (txn->mt_cursors) {
7120 mc->mc_next = txn->mt_cursors[dbi];
7121 txn->mt_cursors[dbi] = mc;
7122 mc->mc_flags |= C_UNTRACK;
7134 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7136 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7139 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7142 if (txn->mt_flags & MDB_TXN_ERROR)
7145 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7149 /* Return the count of duplicate data items for the current key */
7151 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7155 if (mc == NULL || countp == NULL)
7158 if (mc->mc_xcursor == NULL)
7159 return MDB_INCOMPATIBLE;
7161 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7164 if (!(mc->mc_flags & C_INITIALIZED))
7167 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7168 return MDB_NOTFOUND;
7170 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7171 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7174 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7177 *countp = mc->mc_xcursor->mx_db.md_entries;
7183 mdb_cursor_close(MDB_cursor *mc)
7185 if (mc && !mc->mc_backup) {
7186 /* remove from txn, if tracked */
7187 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7188 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7189 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7191 *prev = mc->mc_next;
7198 mdb_cursor_txn(MDB_cursor *mc)
7200 if (!mc) return NULL;
7205 mdb_cursor_dbi(MDB_cursor *mc)
7210 /** Replace the key for a branch node with a new key.
7211 * @param[in] mc Cursor pointing to the node to operate on.
7212 * @param[in] key The new key to use.
7213 * @return 0 on success, non-zero on failure.
7216 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7222 int delta, ksize, oksize;
7223 indx_t ptr, i, numkeys, indx;
7226 indx = mc->mc_ki[mc->mc_top];
7227 mp = mc->mc_pg[mc->mc_top];
7228 node = NODEPTR(mp, indx);
7229 ptr = mp->mp_ptrs[indx];
7233 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7234 k2.mv_data = NODEKEY(node);
7235 k2.mv_size = node->mn_ksize;
7236 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7238 mdb_dkey(&k2, kbuf2),
7244 /* Sizes must be 2-byte aligned. */
7245 ksize = EVEN(key->mv_size);
7246 oksize = EVEN(node->mn_ksize);
7247 delta = ksize - oksize;
7249 /* Shift node contents if EVEN(key length) changed. */
7251 if (delta > 0 && SIZELEFT(mp) < delta) {
7253 /* not enough space left, do a delete and split */
7254 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7255 pgno = NODEPGNO(node);
7256 mdb_node_del(mc, 0);
7257 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7260 numkeys = NUMKEYS(mp);
7261 for (i = 0; i < numkeys; i++) {
7262 if (mp->mp_ptrs[i] <= ptr)
7263 mp->mp_ptrs[i] -= delta;
7266 base = (char *)mp + mp->mp_upper + PAGEBASE;
7267 len = ptr - mp->mp_upper + NODESIZE;
7268 memmove(base - delta, base, len);
7269 mp->mp_upper -= delta;
7271 node = NODEPTR(mp, indx);
7274 /* But even if no shift was needed, update ksize */
7275 if (node->mn_ksize != key->mv_size)
7276 node->mn_ksize = key->mv_size;
7279 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7285 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7287 /** Move a node from csrc to cdst.
7290 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7297 unsigned short flags;
7301 /* Mark src and dst as dirty. */
7302 if ((rc = mdb_page_touch(csrc)) ||
7303 (rc = mdb_page_touch(cdst)))
7306 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7307 key.mv_size = csrc->mc_db->md_pad;
7308 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7310 data.mv_data = NULL;
7314 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7315 mdb_cassert(csrc, !((size_t)srcnode & 1));
7316 srcpg = NODEPGNO(srcnode);
7317 flags = srcnode->mn_flags;
7318 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7319 unsigned int snum = csrc->mc_snum;
7321 /* must find the lowest key below src */
7322 rc = mdb_page_search_lowest(csrc);
7325 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7326 key.mv_size = csrc->mc_db->md_pad;
7327 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7329 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7330 key.mv_size = NODEKSZ(s2);
7331 key.mv_data = NODEKEY(s2);
7333 csrc->mc_snum = snum--;
7334 csrc->mc_top = snum;
7336 key.mv_size = NODEKSZ(srcnode);
7337 key.mv_data = NODEKEY(srcnode);
7339 data.mv_size = NODEDSZ(srcnode);
7340 data.mv_data = NODEDATA(srcnode);
7342 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7343 unsigned int snum = cdst->mc_snum;
7346 /* must find the lowest key below dst */
7347 mdb_cursor_copy(cdst, &mn);
7348 rc = mdb_page_search_lowest(&mn);
7351 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7352 bkey.mv_size = mn.mc_db->md_pad;
7353 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7355 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7356 bkey.mv_size = NODEKSZ(s2);
7357 bkey.mv_data = NODEKEY(s2);
7359 mn.mc_snum = snum--;
7362 rc = mdb_update_key(&mn, &bkey);
7367 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7368 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7369 csrc->mc_ki[csrc->mc_top],
7371 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7372 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7374 /* Add the node to the destination page.
7376 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7377 if (rc != MDB_SUCCESS)
7380 /* Delete the node from the source page.
7382 mdb_node_del(csrc, key.mv_size);
7385 /* Adjust other cursors pointing to mp */
7386 MDB_cursor *m2, *m3;
7387 MDB_dbi dbi = csrc->mc_dbi;
7388 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7390 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7391 if (csrc->mc_flags & C_SUB)
7392 m3 = &m2->mc_xcursor->mx_cursor;
7395 if (m3 == csrc) continue;
7396 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7397 csrc->mc_ki[csrc->mc_top]) {
7398 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7399 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7404 /* Update the parent separators.
7406 if (csrc->mc_ki[csrc->mc_top] == 0) {
7407 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7408 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7409 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7411 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7412 key.mv_size = NODEKSZ(srcnode);
7413 key.mv_data = NODEKEY(srcnode);
7415 DPRINTF(("update separator for source page %"Z"u to [%s]",
7416 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7417 mdb_cursor_copy(csrc, &mn);
7420 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7423 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7425 indx_t ix = csrc->mc_ki[csrc->mc_top];
7426 nullkey.mv_size = 0;
7427 csrc->mc_ki[csrc->mc_top] = 0;
7428 rc = mdb_update_key(csrc, &nullkey);
7429 csrc->mc_ki[csrc->mc_top] = ix;
7430 mdb_cassert(csrc, rc == MDB_SUCCESS);
7434 if (cdst->mc_ki[cdst->mc_top] == 0) {
7435 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7436 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7437 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7439 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7440 key.mv_size = NODEKSZ(srcnode);
7441 key.mv_data = NODEKEY(srcnode);
7443 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7444 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7445 mdb_cursor_copy(cdst, &mn);
7448 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7451 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7453 indx_t ix = cdst->mc_ki[cdst->mc_top];
7454 nullkey.mv_size = 0;
7455 cdst->mc_ki[cdst->mc_top] = 0;
7456 rc = mdb_update_key(cdst, &nullkey);
7457 cdst->mc_ki[cdst->mc_top] = ix;
7458 mdb_cassert(csrc, rc == MDB_SUCCESS);
7465 /** Merge one page into another.
7466 * The nodes from the page pointed to by \b csrc will
7467 * be copied to the page pointed to by \b cdst and then
7468 * the \b csrc page will be freed.
7469 * @param[in] csrc Cursor pointing to the source page.
7470 * @param[in] cdst Cursor pointing to the destination page.
7471 * @return 0 on success, non-zero on failure.
7474 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7476 MDB_page *psrc, *pdst;
7483 psrc = csrc->mc_pg[csrc->mc_top];
7484 pdst = cdst->mc_pg[cdst->mc_top];
7486 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7488 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7489 mdb_cassert(csrc, cdst->mc_snum > 1);
7491 /* Mark dst as dirty. */
7492 if ((rc = mdb_page_touch(cdst)))
7495 /* Move all nodes from src to dst.
7497 j = nkeys = NUMKEYS(pdst);
7498 if (IS_LEAF2(psrc)) {
7499 key.mv_size = csrc->mc_db->md_pad;
7500 key.mv_data = METADATA(psrc);
7501 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7502 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7503 if (rc != MDB_SUCCESS)
7505 key.mv_data = (char *)key.mv_data + key.mv_size;
7508 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7509 srcnode = NODEPTR(psrc, i);
7510 if (i == 0 && IS_BRANCH(psrc)) {
7513 mdb_cursor_copy(csrc, &mn);
7514 /* must find the lowest key below src */
7515 rc = mdb_page_search_lowest(&mn);
7518 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7519 key.mv_size = mn.mc_db->md_pad;
7520 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7522 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7523 key.mv_size = NODEKSZ(s2);
7524 key.mv_data = NODEKEY(s2);
7527 key.mv_size = srcnode->mn_ksize;
7528 key.mv_data = NODEKEY(srcnode);
7531 data.mv_size = NODEDSZ(srcnode);
7532 data.mv_data = NODEDATA(srcnode);
7533 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7534 if (rc != MDB_SUCCESS)
7539 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7540 pdst->mp_pgno, NUMKEYS(pdst),
7541 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7543 /* Unlink the src page from parent and add to free list.
7546 mdb_node_del(csrc, 0);
7547 if (csrc->mc_ki[csrc->mc_top] == 0) {
7549 rc = mdb_update_key(csrc, &key);
7557 psrc = csrc->mc_pg[csrc->mc_top];
7558 /* If not operating on FreeDB, allow this page to be reused
7559 * in this txn. Otherwise just add to free list.
7561 rc = mdb_page_loose(csrc, psrc);
7565 csrc->mc_db->md_leaf_pages--;
7567 csrc->mc_db->md_branch_pages--;
7569 /* Adjust other cursors pointing to mp */
7570 MDB_cursor *m2, *m3;
7571 MDB_dbi dbi = csrc->mc_dbi;
7573 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7574 if (csrc->mc_flags & C_SUB)
7575 m3 = &m2->mc_xcursor->mx_cursor;
7578 if (m3 == csrc) continue;
7579 if (m3->mc_snum < csrc->mc_snum) continue;
7580 if (m3->mc_pg[csrc->mc_top] == psrc) {
7581 m3->mc_pg[csrc->mc_top] = pdst;
7582 m3->mc_ki[csrc->mc_top] += nkeys;
7587 unsigned int snum = cdst->mc_snum;
7588 uint16_t depth = cdst->mc_db->md_depth;
7589 mdb_cursor_pop(cdst);
7590 rc = mdb_rebalance(cdst);
7591 /* Did the tree shrink? */
7592 if (depth > cdst->mc_db->md_depth)
7594 cdst->mc_snum = snum;
7595 cdst->mc_top = snum-1;
7600 /** Copy the contents of a cursor.
7601 * @param[in] csrc The cursor to copy from.
7602 * @param[out] cdst The cursor to copy to.
7605 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7609 cdst->mc_txn = csrc->mc_txn;
7610 cdst->mc_dbi = csrc->mc_dbi;
7611 cdst->mc_db = csrc->mc_db;
7612 cdst->mc_dbx = csrc->mc_dbx;
7613 cdst->mc_snum = csrc->mc_snum;
7614 cdst->mc_top = csrc->mc_top;
7615 cdst->mc_flags = csrc->mc_flags;
7617 for (i=0; i<csrc->mc_snum; i++) {
7618 cdst->mc_pg[i] = csrc->mc_pg[i];
7619 cdst->mc_ki[i] = csrc->mc_ki[i];
7623 /** Rebalance the tree after a delete operation.
7624 * @param[in] mc Cursor pointing to the page where rebalancing
7626 * @return 0 on success, non-zero on failure.
7629 mdb_rebalance(MDB_cursor *mc)
7633 unsigned int ptop, minkeys;
7637 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7638 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7639 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7640 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7641 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7643 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7644 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7645 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7646 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7650 if (mc->mc_snum < 2) {
7651 MDB_page *mp = mc->mc_pg[0];
7653 DPUTS("Can't rebalance a subpage, ignoring");
7656 if (NUMKEYS(mp) == 0) {
7657 DPUTS("tree is completely empty");
7658 mc->mc_db->md_root = P_INVALID;
7659 mc->mc_db->md_depth = 0;
7660 mc->mc_db->md_leaf_pages = 0;
7661 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7664 /* Adjust cursors pointing to mp */
7667 mc->mc_flags &= ~C_INITIALIZED;
7669 MDB_cursor *m2, *m3;
7670 MDB_dbi dbi = mc->mc_dbi;
7672 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7673 if (mc->mc_flags & C_SUB)
7674 m3 = &m2->mc_xcursor->mx_cursor;
7677 if (m3->mc_snum < mc->mc_snum) continue;
7678 if (m3->mc_pg[0] == mp) {
7681 m3->mc_flags &= ~C_INITIALIZED;
7685 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7687 DPUTS("collapsing root page!");
7688 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7691 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7692 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7695 mc->mc_db->md_depth--;
7696 mc->mc_db->md_branch_pages--;
7697 mc->mc_ki[0] = mc->mc_ki[1];
7698 for (i = 1; i<mc->mc_db->md_depth; i++) {
7699 mc->mc_pg[i] = mc->mc_pg[i+1];
7700 mc->mc_ki[i] = mc->mc_ki[i+1];
7703 /* Adjust other cursors pointing to mp */
7704 MDB_cursor *m2, *m3;
7705 MDB_dbi dbi = mc->mc_dbi;
7707 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7708 if (mc->mc_flags & C_SUB)
7709 m3 = &m2->mc_xcursor->mx_cursor;
7712 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7713 if (m3->mc_pg[0] == mp) {
7716 for (i=0; i<m3->mc_snum; i++) {
7717 m3->mc_pg[i] = m3->mc_pg[i+1];
7718 m3->mc_ki[i] = m3->mc_ki[i+1];
7724 DPUTS("root page doesn't need rebalancing");
7728 /* The parent (branch page) must have at least 2 pointers,
7729 * otherwise the tree is invalid.
7731 ptop = mc->mc_top-1;
7732 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7734 /* Leaf page fill factor is below the threshold.
7735 * Try to move keys from left or right neighbor, or
7736 * merge with a neighbor page.
7741 mdb_cursor_copy(mc, &mn);
7742 mn.mc_xcursor = NULL;
7744 oldki = mc->mc_ki[mc->mc_top];
7745 if (mc->mc_ki[ptop] == 0) {
7746 /* We're the leftmost leaf in our parent.
7748 DPUTS("reading right neighbor");
7750 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7751 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7754 mn.mc_ki[mn.mc_top] = 0;
7755 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7757 /* There is at least one neighbor to the left.
7759 DPUTS("reading left neighbor");
7761 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7762 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7765 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7766 mc->mc_ki[mc->mc_top] = 0;
7769 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7770 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7771 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7773 /* If the neighbor page is above threshold and has enough keys,
7774 * move one key from it. Otherwise we should try to merge them.
7775 * (A branch page must never have less than 2 keys.)
7777 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7778 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7779 rc = mdb_node_move(&mn, mc);
7780 if (mc->mc_ki[ptop]) {
7784 if (mc->mc_ki[ptop] == 0) {
7785 rc = mdb_page_merge(&mn, mc);
7787 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7788 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7789 rc = mdb_page_merge(mc, &mn);
7790 mdb_cursor_copy(&mn, mc);
7792 mc->mc_flags &= ~C_EOF;
7794 mc->mc_ki[mc->mc_top] = oldki;
7798 /** Complete a delete operation started by #mdb_cursor_del(). */
7800 mdb_cursor_del0(MDB_cursor *mc)
7807 ki = mc->mc_ki[mc->mc_top];
7808 mdb_node_del(mc, mc->mc_db->md_pad);
7809 mc->mc_db->md_entries--;
7810 rc = mdb_rebalance(mc);
7812 if (rc == MDB_SUCCESS) {
7813 MDB_cursor *m2, *m3;
7814 MDB_dbi dbi = mc->mc_dbi;
7816 mp = mc->mc_pg[mc->mc_top];
7817 nkeys = NUMKEYS(mp);
7819 /* if mc points past last node in page, find next sibling */
7820 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7821 rc = mdb_cursor_sibling(mc, 1);
7822 if (rc == MDB_NOTFOUND) {
7823 mc->mc_flags |= C_EOF;
7828 /* Adjust other cursors pointing to mp */
7829 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7830 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7831 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7833 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7835 if (m3->mc_pg[mc->mc_top] == mp) {
7836 if (m3->mc_ki[mc->mc_top] >= ki) {
7837 m3->mc_flags |= C_DEL;
7838 if (m3->mc_ki[mc->mc_top] > ki)
7839 m3->mc_ki[mc->mc_top]--;
7840 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7841 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7843 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7844 rc = mdb_cursor_sibling(m3, 1);
7845 if (rc == MDB_NOTFOUND) {
7846 m3->mc_flags |= C_EOF;
7852 mc->mc_flags |= C_DEL;
7856 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7861 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7862 MDB_val *key, MDB_val *data)
7864 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7867 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7868 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7870 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7871 /* must ignore any data */
7875 return mdb_del0(txn, dbi, key, data, 0);
7879 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7880 MDB_val *key, MDB_val *data, unsigned flags)
7885 MDB_val rdata, *xdata;
7889 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7891 mdb_cursor_init(&mc, txn, dbi, &mx);
7900 flags |= MDB_NODUPDATA;
7902 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7904 /* let mdb_page_split know about this cursor if needed:
7905 * delete will trigger a rebalance; if it needs to move
7906 * a node from one page to another, it will have to
7907 * update the parent's separator key(s). If the new sepkey
7908 * is larger than the current one, the parent page may
7909 * run out of space, triggering a split. We need this
7910 * cursor to be consistent until the end of the rebalance.
7912 mc.mc_flags |= C_UNTRACK;
7913 mc.mc_next = txn->mt_cursors[dbi];
7914 txn->mt_cursors[dbi] = &mc;
7915 rc = mdb_cursor_del(&mc, flags);
7916 txn->mt_cursors[dbi] = mc.mc_next;
7921 /** Split a page and insert a new node.
7922 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7923 * The cursor will be updated to point to the actual page and index where
7924 * the node got inserted after the split.
7925 * @param[in] newkey The key for the newly inserted node.
7926 * @param[in] newdata The data for the newly inserted node.
7927 * @param[in] newpgno The page number, if the new node is a branch node.
7928 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7929 * @return 0 on success, non-zero on failure.
7932 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7933 unsigned int nflags)
7936 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7939 int i, j, split_indx, nkeys, pmax;
7940 MDB_env *env = mc->mc_txn->mt_env;
7942 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7943 MDB_page *copy = NULL;
7944 MDB_page *mp, *rp, *pp;
7949 mp = mc->mc_pg[mc->mc_top];
7950 newindx = mc->mc_ki[mc->mc_top];
7951 nkeys = NUMKEYS(mp);
7953 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7954 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7955 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7957 /* Create a right sibling. */
7958 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7960 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7962 if (mc->mc_snum < 2) {
7963 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7965 /* shift current top to make room for new parent */
7966 mc->mc_pg[1] = mc->mc_pg[0];
7967 mc->mc_ki[1] = mc->mc_ki[0];
7970 mc->mc_db->md_root = pp->mp_pgno;
7971 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7972 mc->mc_db->md_depth++;
7975 /* Add left (implicit) pointer. */
7976 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7977 /* undo the pre-push */
7978 mc->mc_pg[0] = mc->mc_pg[1];
7979 mc->mc_ki[0] = mc->mc_ki[1];
7980 mc->mc_db->md_root = mp->mp_pgno;
7981 mc->mc_db->md_depth--;
7988 ptop = mc->mc_top-1;
7989 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7992 mc->mc_flags |= C_SPLITTING;
7993 mdb_cursor_copy(mc, &mn);
7994 mn.mc_pg[mn.mc_top] = rp;
7995 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7997 if (nflags & MDB_APPEND) {
7998 mn.mc_ki[mn.mc_top] = 0;
8000 split_indx = newindx;
8004 split_indx = (nkeys+1) / 2;
8009 unsigned int lsize, rsize, ksize;
8010 /* Move half of the keys to the right sibling */
8011 x = mc->mc_ki[mc->mc_top] - split_indx;
8012 ksize = mc->mc_db->md_pad;
8013 split = LEAF2KEY(mp, split_indx, ksize);
8014 rsize = (nkeys - split_indx) * ksize;
8015 lsize = (nkeys - split_indx) * sizeof(indx_t);
8016 mp->mp_lower -= lsize;
8017 rp->mp_lower += lsize;
8018 mp->mp_upper += rsize - lsize;
8019 rp->mp_upper -= rsize - lsize;
8020 sepkey.mv_size = ksize;
8021 if (newindx == split_indx) {
8022 sepkey.mv_data = newkey->mv_data;
8024 sepkey.mv_data = split;
8027 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8028 memcpy(rp->mp_ptrs, split, rsize);
8029 sepkey.mv_data = rp->mp_ptrs;
8030 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8031 memcpy(ins, newkey->mv_data, ksize);
8032 mp->mp_lower += sizeof(indx_t);
8033 mp->mp_upper -= ksize - sizeof(indx_t);
8036 memcpy(rp->mp_ptrs, split, x * ksize);
8037 ins = LEAF2KEY(rp, x, ksize);
8038 memcpy(ins, newkey->mv_data, ksize);
8039 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8040 rp->mp_lower += sizeof(indx_t);
8041 rp->mp_upper -= ksize - sizeof(indx_t);
8042 mc->mc_ki[mc->mc_top] = x;
8043 mc->mc_pg[mc->mc_top] = rp;
8046 int psize, nsize, k;
8047 /* Maximum free space in an empty page */
8048 pmax = env->me_psize - PAGEHDRSZ;
8050 nsize = mdb_leaf_size(env, newkey, newdata);
8052 nsize = mdb_branch_size(env, newkey);
8053 nsize = EVEN(nsize);
8055 /* grab a page to hold a temporary copy */
8056 copy = mdb_page_malloc(mc->mc_txn, 1);
8061 copy->mp_pgno = mp->mp_pgno;
8062 copy->mp_flags = mp->mp_flags;
8063 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8064 copy->mp_upper = env->me_psize - PAGEBASE;
8066 /* prepare to insert */
8067 for (i=0, j=0; i<nkeys; i++) {
8069 copy->mp_ptrs[j++] = 0;
8071 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8074 /* When items are relatively large the split point needs
8075 * to be checked, because being off-by-one will make the
8076 * difference between success or failure in mdb_node_add.
8078 * It's also relevant if a page happens to be laid out
8079 * such that one half of its nodes are all "small" and
8080 * the other half of its nodes are "large." If the new
8081 * item is also "large" and falls on the half with
8082 * "large" nodes, it also may not fit.
8084 * As a final tweak, if the new item goes on the last
8085 * spot on the page (and thus, onto the new page), bias
8086 * the split so the new page is emptier than the old page.
8087 * This yields better packing during sequential inserts.
8089 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8090 /* Find split point */
8092 if (newindx <= split_indx || newindx >= nkeys) {
8094 k = newindx >= nkeys ? nkeys : split_indx+2;
8099 for (; i!=k; i+=j) {
8104 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8105 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8107 if (F_ISSET(node->mn_flags, F_BIGDATA))
8108 psize += sizeof(pgno_t);
8110 psize += NODEDSZ(node);
8112 psize = EVEN(psize);
8114 if (psize > pmax || i == k-j) {
8115 split_indx = i + (j<0);
8120 if (split_indx == newindx) {
8121 sepkey.mv_size = newkey->mv_size;
8122 sepkey.mv_data = newkey->mv_data;
8124 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8125 sepkey.mv_size = node->mn_ksize;
8126 sepkey.mv_data = NODEKEY(node);
8131 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8133 /* Copy separator key to the parent.
8135 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8139 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8144 if (mn.mc_snum == mc->mc_snum) {
8145 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8146 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8147 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8148 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8153 /* Right page might now have changed parent.
8154 * Check if left page also changed parent.
8156 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8157 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8158 for (i=0; i<ptop; i++) {
8159 mc->mc_pg[i] = mn.mc_pg[i];
8160 mc->mc_ki[i] = mn.mc_ki[i];
8162 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8163 if (mn.mc_ki[ptop]) {
8164 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8166 /* find right page's left sibling */
8167 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8168 mdb_cursor_sibling(mc, 0);
8173 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8176 mc->mc_flags ^= C_SPLITTING;
8177 if (rc != MDB_SUCCESS) {
8180 if (nflags & MDB_APPEND) {
8181 mc->mc_pg[mc->mc_top] = rp;
8182 mc->mc_ki[mc->mc_top] = 0;
8183 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8186 for (i=0; i<mc->mc_top; i++)
8187 mc->mc_ki[i] = mn.mc_ki[i];
8188 } else if (!IS_LEAF2(mp)) {
8190 mc->mc_pg[mc->mc_top] = rp;
8195 rkey.mv_data = newkey->mv_data;
8196 rkey.mv_size = newkey->mv_size;
8202 /* Update index for the new key. */
8203 mc->mc_ki[mc->mc_top] = j;
8205 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8206 rkey.mv_data = NODEKEY(node);
8207 rkey.mv_size = node->mn_ksize;
8209 xdata.mv_data = NODEDATA(node);
8210 xdata.mv_size = NODEDSZ(node);
8213 pgno = NODEPGNO(node);
8214 flags = node->mn_flags;
8217 if (!IS_LEAF(mp) && j == 0) {
8218 /* First branch index doesn't need key data. */
8222 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8228 mc->mc_pg[mc->mc_top] = copy;
8233 } while (i != split_indx);
8235 nkeys = NUMKEYS(copy);
8236 for (i=0; i<nkeys; i++)
8237 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8238 mp->mp_lower = copy->mp_lower;
8239 mp->mp_upper = copy->mp_upper;
8240 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8241 env->me_psize - copy->mp_upper - PAGEBASE);
8243 /* reset back to original page */
8244 if (newindx < split_indx) {
8245 mc->mc_pg[mc->mc_top] = mp;
8246 if (nflags & MDB_RESERVE) {
8247 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8248 if (!(node->mn_flags & F_BIGDATA))
8249 newdata->mv_data = NODEDATA(node);
8252 mc->mc_pg[mc->mc_top] = rp;
8254 /* Make sure mc_ki is still valid.
8256 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8257 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8258 for (i=0; i<=ptop; i++) {
8259 mc->mc_pg[i] = mn.mc_pg[i];
8260 mc->mc_ki[i] = mn.mc_ki[i];
8267 /* Adjust other cursors pointing to mp */
8268 MDB_cursor *m2, *m3;
8269 MDB_dbi dbi = mc->mc_dbi;
8270 int fixup = NUMKEYS(mp);
8272 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8273 if (mc->mc_flags & C_SUB)
8274 m3 = &m2->mc_xcursor->mx_cursor;
8279 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8281 if (m3->mc_flags & C_SPLITTING)
8286 for (k=m3->mc_top; k>=0; k--) {
8287 m3->mc_ki[k+1] = m3->mc_ki[k];
8288 m3->mc_pg[k+1] = m3->mc_pg[k];
8290 if (m3->mc_ki[0] >= split_indx) {
8295 m3->mc_pg[0] = mc->mc_pg[0];
8299 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8300 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8301 m3->mc_ki[mc->mc_top]++;
8302 if (m3->mc_ki[mc->mc_top] >= fixup) {
8303 m3->mc_pg[mc->mc_top] = rp;
8304 m3->mc_ki[mc->mc_top] -= fixup;
8305 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8307 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8308 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8313 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8316 if (copy) /* tmp page */
8317 mdb_page_free(env, copy);
8319 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8324 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8325 MDB_val *key, MDB_val *data, unsigned int flags)
8330 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8333 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8336 mdb_cursor_init(&mc, txn, dbi, &mx);
8337 return mdb_cursor_put(&mc, key, data, flags);
8341 #define MDB_WBUF (1024*1024)
8344 /** State needed for a compacting copy. */
8345 typedef struct mdb_copy {
8346 pthread_mutex_t mc_mutex;
8347 pthread_cond_t mc_cond;
8354 pgno_t mc_next_pgno;
8357 volatile int mc_new;
8362 /** Dedicated writer thread for compacting copy. */
8363 static THREAD_RET ESECT
8364 mdb_env_copythr(void *arg)
8368 int toggle = 0, wsize, rc;
8371 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8374 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8377 pthread_mutex_lock(&my->mc_mutex);
8379 pthread_cond_signal(&my->mc_cond);
8382 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8383 if (my->mc_new < 0) {
8388 wsize = my->mc_wlen[toggle];
8389 ptr = my->mc_wbuf[toggle];
8392 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8396 } else if (len > 0) {
8410 /* If there's an overflow page tail, write it too */
8411 if (my->mc_olen[toggle]) {
8412 wsize = my->mc_olen[toggle];
8413 ptr = my->mc_over[toggle];
8414 my->mc_olen[toggle] = 0;
8417 my->mc_wlen[toggle] = 0;
8419 pthread_cond_signal(&my->mc_cond);
8421 pthread_cond_signal(&my->mc_cond);
8422 pthread_mutex_unlock(&my->mc_mutex);
8423 return (THREAD_RET)0;
8427 /** Tell the writer thread there's a buffer ready to write */
8429 mdb_env_cthr_toggle(mdb_copy *my, int st)
8431 int toggle = my->mc_toggle ^ 1;
8432 pthread_mutex_lock(&my->mc_mutex);
8433 if (my->mc_status) {
8434 pthread_mutex_unlock(&my->mc_mutex);
8435 return my->mc_status;
8437 while (my->mc_new == 1)
8438 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8440 my->mc_toggle = toggle;
8441 pthread_cond_signal(&my->mc_cond);
8442 pthread_mutex_unlock(&my->mc_mutex);
8446 /** Depth-first tree traversal for compacting copy. */
8448 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8451 MDB_txn *txn = my->mc_txn;
8453 MDB_page *mo, *mp, *leaf;
8458 /* Empty DB, nothing to do */
8459 if (*pg == P_INVALID)
8466 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8469 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8473 /* Make cursor pages writable */
8474 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8478 for (i=0; i<mc.mc_top; i++) {
8479 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8480 mc.mc_pg[i] = (MDB_page *)ptr;
8481 ptr += my->mc_env->me_psize;
8484 /* This is writable space for a leaf page. Usually not needed. */
8485 leaf = (MDB_page *)ptr;
8487 toggle = my->mc_toggle;
8488 while (mc.mc_snum > 0) {
8490 mp = mc.mc_pg[mc.mc_top];
8494 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8495 for (i=0; i<n; i++) {
8496 ni = NODEPTR(mp, i);
8497 if (ni->mn_flags & F_BIGDATA) {
8501 /* Need writable leaf */
8503 mc.mc_pg[mc.mc_top] = leaf;
8504 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8506 ni = NODEPTR(mp, i);
8509 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8510 rc = mdb_page_get(txn, pg, &omp, NULL);
8513 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8514 rc = mdb_env_cthr_toggle(my, 1);
8517 toggle = my->mc_toggle;
8519 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8520 memcpy(mo, omp, my->mc_env->me_psize);
8521 mo->mp_pgno = my->mc_next_pgno;
8522 my->mc_next_pgno += omp->mp_pages;
8523 my->mc_wlen[toggle] += my->mc_env->me_psize;
8524 if (omp->mp_pages > 1) {
8525 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8526 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8527 rc = mdb_env_cthr_toggle(my, 1);
8530 toggle = my->mc_toggle;
8532 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8533 } else if (ni->mn_flags & F_SUBDATA) {
8536 /* Need writable leaf */
8538 mc.mc_pg[mc.mc_top] = leaf;
8539 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8541 ni = NODEPTR(mp, i);
8544 memcpy(&db, NODEDATA(ni), sizeof(db));
8545 my->mc_toggle = toggle;
8546 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8549 toggle = my->mc_toggle;
8550 memcpy(NODEDATA(ni), &db, sizeof(db));
8555 mc.mc_ki[mc.mc_top]++;
8556 if (mc.mc_ki[mc.mc_top] < n) {
8559 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8561 rc = mdb_page_get(txn, pg, &mp, NULL);
8566 mc.mc_ki[mc.mc_top] = 0;
8567 if (IS_BRANCH(mp)) {
8568 /* Whenever we advance to a sibling branch page,
8569 * we must proceed all the way down to its first leaf.
8571 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8574 mc.mc_pg[mc.mc_top] = mp;
8578 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8579 rc = mdb_env_cthr_toggle(my, 1);
8582 toggle = my->mc_toggle;
8584 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8585 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8586 mo->mp_pgno = my->mc_next_pgno++;
8587 my->mc_wlen[toggle] += my->mc_env->me_psize;
8589 /* Update parent if there is one */
8590 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8591 SETPGNO(ni, mo->mp_pgno);
8592 mdb_cursor_pop(&mc);
8594 /* Otherwise we're done */
8604 /** Copy environment with compaction. */
8606 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8611 MDB_txn *txn = NULL;
8616 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8617 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8618 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8619 if (my.mc_wbuf[0] == NULL)
8622 pthread_mutex_init(&my.mc_mutex, NULL);
8623 pthread_cond_init(&my.mc_cond, NULL);
8624 #ifdef HAVE_MEMALIGN
8625 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8626 if (my.mc_wbuf[0] == NULL)
8629 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8634 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8635 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8640 my.mc_next_pgno = 2;
8646 THREAD_CREATE(thr, mdb_env_copythr, &my);
8648 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8652 mp = (MDB_page *)my.mc_wbuf[0];
8653 memset(mp, 0, 2*env->me_psize);
8655 mp->mp_flags = P_META;
8656 mm = (MDB_meta *)METADATA(mp);
8657 mdb_env_init_meta0(env, mm);
8658 mm->mm_address = env->me_metas[0]->mm_address;
8660 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8662 mp->mp_flags = P_META;
8663 *(MDB_meta *)METADATA(mp) = *mm;
8664 mm = (MDB_meta *)METADATA(mp);
8666 /* Count the number of free pages, subtract from lastpg to find
8667 * number of active pages
8670 MDB_ID freecount = 0;
8673 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8674 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8675 freecount += *(MDB_ID *)data.mv_data;
8676 freecount += txn->mt_dbs[0].md_branch_pages +
8677 txn->mt_dbs[0].md_leaf_pages +
8678 txn->mt_dbs[0].md_overflow_pages;
8680 /* Set metapage 1 */
8681 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8682 mm->mm_dbs[1] = txn->mt_dbs[1];
8683 if (mm->mm_last_pg > 1) {
8684 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8687 mm->mm_dbs[1].md_root = P_INVALID;
8690 my.mc_wlen[0] = env->me_psize * 2;
8692 pthread_mutex_lock(&my.mc_mutex);
8694 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8695 pthread_mutex_unlock(&my.mc_mutex);
8696 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8697 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8698 rc = mdb_env_cthr_toggle(&my, 1);
8699 mdb_env_cthr_toggle(&my, -1);
8700 pthread_mutex_lock(&my.mc_mutex);
8702 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8703 pthread_mutex_unlock(&my.mc_mutex);
8708 CloseHandle(my.mc_cond);
8709 CloseHandle(my.mc_mutex);
8710 _aligned_free(my.mc_wbuf[0]);
8712 pthread_cond_destroy(&my.mc_cond);
8713 pthread_mutex_destroy(&my.mc_mutex);
8714 free(my.mc_wbuf[0]);
8719 /** Copy environment as-is. */
8721 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8723 MDB_txn *txn = NULL;
8729 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8733 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8736 /* Do the lock/unlock of the reader mutex before starting the
8737 * write txn. Otherwise other read txns could block writers.
8739 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8744 /* We must start the actual read txn after blocking writers */
8745 mdb_txn_reset0(txn, "reset-stage1");
8747 /* Temporarily block writers until we snapshot the meta pages */
8750 rc = mdb_txn_renew0(txn);
8752 UNLOCK_MUTEX_W(env);
8757 wsize = env->me_psize * 2;
8761 DO_WRITE(rc, fd, ptr, w2, len);
8765 } else if (len > 0) {
8771 /* Non-blocking or async handles are not supported */
8777 UNLOCK_MUTEX_W(env);
8782 w2 = txn->mt_next_pgno * env->me_psize;
8785 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8792 if (wsize > MAX_WRITE)
8796 DO_WRITE(rc, fd, ptr, w2, len);
8800 } else if (len > 0) {
8817 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8819 if (flags & MDB_CP_COMPACT)
8820 return mdb_env_copyfd1(env, fd);
8822 return mdb_env_copyfd0(env, fd);
8826 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8828 return mdb_env_copyfd2(env, fd, 0);
8832 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8836 HANDLE newfd = INVALID_HANDLE_VALUE;
8838 if (env->me_flags & MDB_NOSUBDIR) {
8839 lpath = (char *)path;
8842 len += sizeof(DATANAME);
8843 lpath = malloc(len);
8846 sprintf(lpath, "%s" DATANAME, path);
8849 /* The destination path must exist, but the destination file must not.
8850 * We don't want the OS to cache the writes, since the source data is
8851 * already in the OS cache.
8854 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8855 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8857 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8859 if (newfd == INVALID_HANDLE_VALUE) {
8864 if (env->me_psize >= env->me_os_psize) {
8866 /* Set O_DIRECT if the file system supports it */
8867 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8868 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8870 #ifdef F_NOCACHE /* __APPLE__ */
8871 rc = fcntl(newfd, F_NOCACHE, 1);
8879 rc = mdb_env_copyfd2(env, newfd, flags);
8882 if (!(env->me_flags & MDB_NOSUBDIR))
8884 if (newfd != INVALID_HANDLE_VALUE)
8885 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8892 mdb_env_copy(MDB_env *env, const char *path)
8894 return mdb_env_copy2(env, path, 0);
8898 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8900 if ((flag & CHANGEABLE) != flag)
8903 env->me_flags |= flag;
8905 env->me_flags &= ~flag;
8910 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8915 *arg = env->me_flags;
8920 mdb_env_set_userctx(MDB_env *env, void *ctx)
8924 env->me_userctx = ctx;
8929 mdb_env_get_userctx(MDB_env *env)
8931 return env ? env->me_userctx : NULL;
8935 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8940 env->me_assert_func = func;
8946 mdb_env_get_path(MDB_env *env, const char **arg)
8951 *arg = env->me_path;
8956 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8965 /** Common code for #mdb_stat() and #mdb_env_stat().
8966 * @param[in] env the environment to operate in.
8967 * @param[in] db the #MDB_db record containing the stats to return.
8968 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8969 * @return 0, this function always succeeds.
8972 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8974 arg->ms_psize = env->me_psize;
8975 arg->ms_depth = db->md_depth;
8976 arg->ms_branch_pages = db->md_branch_pages;
8977 arg->ms_leaf_pages = db->md_leaf_pages;
8978 arg->ms_overflow_pages = db->md_overflow_pages;
8979 arg->ms_entries = db->md_entries;
8985 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8989 if (env == NULL || arg == NULL)
8992 toggle = mdb_env_pick_meta(env);
8994 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8998 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9002 if (env == NULL || arg == NULL)
9005 toggle = mdb_env_pick_meta(env);
9006 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9007 arg->me_mapsize = env->me_mapsize;
9008 arg->me_maxreaders = env->me_maxreaders;
9010 /* me_numreaders may be zero if this process never used any readers. Use
9011 * the shared numreader count if it exists.
9013 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
9015 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9016 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9020 /** Set the default comparison functions for a database.
9021 * Called immediately after a database is opened to set the defaults.
9022 * The user can then override them with #mdb_set_compare() or
9023 * #mdb_set_dupsort().
9024 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9025 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9028 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9030 uint16_t f = txn->mt_dbs[dbi].md_flags;
9032 txn->mt_dbxs[dbi].md_cmp =
9033 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9034 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9036 txn->mt_dbxs[dbi].md_dcmp =
9037 !(f & MDB_DUPSORT) ? 0 :
9038 ((f & MDB_INTEGERDUP)
9039 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9040 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9043 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9049 int rc, dbflag, exact;
9050 unsigned int unused = 0, seq;
9053 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9054 mdb_default_cmp(txn, FREE_DBI);
9057 if ((flags & VALID_FLAGS) != flags)
9059 if (txn->mt_flags & MDB_TXN_ERROR)
9065 if (flags & PERSISTENT_FLAGS) {
9066 uint16_t f2 = flags & PERSISTENT_FLAGS;
9067 /* make sure flag changes get committed */
9068 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9069 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9070 txn->mt_flags |= MDB_TXN_DIRTY;
9073 mdb_default_cmp(txn, MAIN_DBI);
9077 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9078 mdb_default_cmp(txn, MAIN_DBI);
9081 /* Is the DB already open? */
9083 for (i=2; i<txn->mt_numdbs; i++) {
9084 if (!txn->mt_dbxs[i].md_name.mv_size) {
9085 /* Remember this free slot */
9086 if (!unused) unused = i;
9089 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9090 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9096 /* If no free slot and max hit, fail */
9097 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9098 return MDB_DBS_FULL;
9100 /* Cannot mix named databases with some mainDB flags */
9101 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9102 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9104 /* Find the DB info */
9105 dbflag = DB_NEW|DB_VALID;
9108 key.mv_data = (void *)name;
9109 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9110 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9111 if (rc == MDB_SUCCESS) {
9112 /* make sure this is actually a DB */
9113 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9114 if (!(node->mn_flags & F_SUBDATA))
9115 return MDB_INCOMPATIBLE;
9116 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9117 /* Create if requested */
9118 data.mv_size = sizeof(MDB_db);
9119 data.mv_data = &dummy;
9120 memset(&dummy, 0, sizeof(dummy));
9121 dummy.md_root = P_INVALID;
9122 dummy.md_flags = flags & PERSISTENT_FLAGS;
9123 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9127 /* OK, got info, add to table */
9128 if (rc == MDB_SUCCESS) {
9129 unsigned int slot = unused ? unused : txn->mt_numdbs;
9130 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9131 txn->mt_dbxs[slot].md_name.mv_size = len;
9132 txn->mt_dbxs[slot].md_rel = NULL;
9133 txn->mt_dbflags[slot] = dbflag;
9134 /* txn-> and env-> are the same in read txns, use
9135 * tmp variable to avoid undefined assignment
9137 seq = ++txn->mt_env->me_dbiseqs[slot];
9138 txn->mt_dbiseqs[slot] = seq;
9140 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9142 mdb_default_cmp(txn, slot);
9151 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9153 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9156 if (txn->mt_flags & MDB_TXN_ERROR)
9159 if (txn->mt_dbflags[dbi] & DB_STALE) {
9162 /* Stale, must read the DB's root. cursor_init does it for us. */
9163 mdb_cursor_init(&mc, txn, dbi, &mx);
9165 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9168 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9171 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9173 ptr = env->me_dbxs[dbi].md_name.mv_data;
9174 /* If there was no name, this was already closed */
9176 env->me_dbxs[dbi].md_name.mv_data = NULL;
9177 env->me_dbxs[dbi].md_name.mv_size = 0;
9178 env->me_dbflags[dbi] = 0;
9179 env->me_dbiseqs[dbi]++;
9184 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9186 /* We could return the flags for the FREE_DBI too but what's the point? */
9187 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9189 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9193 /** Add all the DB's pages to the free list.
9194 * @param[in] mc Cursor on the DB to free.
9195 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9196 * @return 0 on success, non-zero on failure.
9199 mdb_drop0(MDB_cursor *mc, int subs)
9203 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9204 if (rc == MDB_SUCCESS) {
9205 MDB_txn *txn = mc->mc_txn;
9210 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9211 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9214 mdb_cursor_copy(mc, &mx);
9215 while (mc->mc_snum > 0) {
9216 MDB_page *mp = mc->mc_pg[mc->mc_top];
9217 unsigned n = NUMKEYS(mp);
9219 for (i=0; i<n; i++) {
9220 ni = NODEPTR(mp, i);
9221 if (ni->mn_flags & F_BIGDATA) {
9224 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9225 rc = mdb_page_get(txn, pg, &omp, NULL);
9228 mdb_cassert(mc, IS_OVERFLOW(omp));
9229 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9233 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9234 mdb_xcursor_init1(mc, ni);
9235 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9241 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9243 for (i=0; i<n; i++) {
9245 ni = NODEPTR(mp, i);
9248 mdb_midl_xappend(txn->mt_free_pgs, pg);
9253 mc->mc_ki[mc->mc_top] = i;
9254 rc = mdb_cursor_sibling(mc, 1);
9256 if (rc != MDB_NOTFOUND)
9258 /* no more siblings, go back to beginning
9259 * of previous level.
9263 for (i=1; i<mc->mc_snum; i++) {
9265 mc->mc_pg[i] = mx.mc_pg[i];
9270 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9273 txn->mt_flags |= MDB_TXN_ERROR;
9274 } else if (rc == MDB_NOTFOUND) {
9280 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9282 MDB_cursor *mc, *m2;
9285 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9288 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9291 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9294 rc = mdb_cursor_open(txn, dbi, &mc);
9298 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9299 /* Invalidate the dropped DB's cursors */
9300 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9301 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9305 /* Can't delete the main DB */
9306 if (del && dbi > MAIN_DBI) {
9307 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9309 txn->mt_dbflags[dbi] = DB_STALE;
9310 mdb_dbi_close(txn->mt_env, dbi);
9312 txn->mt_flags |= MDB_TXN_ERROR;
9315 /* reset the DB record, mark it dirty */
9316 txn->mt_dbflags[dbi] |= DB_DIRTY;
9317 txn->mt_dbs[dbi].md_depth = 0;
9318 txn->mt_dbs[dbi].md_branch_pages = 0;
9319 txn->mt_dbs[dbi].md_leaf_pages = 0;
9320 txn->mt_dbs[dbi].md_overflow_pages = 0;
9321 txn->mt_dbs[dbi].md_entries = 0;
9322 txn->mt_dbs[dbi].md_root = P_INVALID;
9324 txn->mt_flags |= MDB_TXN_DIRTY;
9327 mdb_cursor_close(mc);
9331 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9333 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9336 txn->mt_dbxs[dbi].md_cmp = cmp;
9340 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9342 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9345 txn->mt_dbxs[dbi].md_dcmp = cmp;
9349 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9351 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9354 txn->mt_dbxs[dbi].md_rel = rel;
9358 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9360 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9363 txn->mt_dbxs[dbi].md_relctx = ctx;
9368 mdb_env_get_maxkeysize(MDB_env *env)
9370 return ENV_MAXKEY(env);
9374 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9376 unsigned int i, rdrs;
9379 int rc = 0, first = 1;
9383 if (!env->me_txns) {
9384 return func("(no reader locks)\n", ctx);
9386 rdrs = env->me_txns->mti_numreaders;
9387 mr = env->me_txns->mti_readers;
9388 for (i=0; i<rdrs; i++) {
9390 txnid_t txnid = mr[i].mr_txnid;
9391 sprintf(buf, txnid == (txnid_t)-1 ?
9392 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9393 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9396 rc = func(" pid thread txnid\n", ctx);
9400 rc = func(buf, ctx);
9406 rc = func("(no active readers)\n", ctx);
9411 /** Insert pid into list if not already present.
9412 * return -1 if already present.
9415 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9417 /* binary search of pid in list */
9419 unsigned cursor = 1;
9421 unsigned n = ids[0];
9424 unsigned pivot = n >> 1;
9425 cursor = base + pivot + 1;
9426 val = pid - ids[cursor];
9431 } else if ( val > 0 ) {
9436 /* found, so it's a duplicate */
9445 for (n = ids[0]; n > cursor; n--)
9452 mdb_reader_check(MDB_env *env, int *dead)
9454 unsigned int i, j, rdrs;
9456 MDB_PID_T *pids, pid;
9465 rdrs = env->me_txns->mti_numreaders;
9466 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9470 mr = env->me_txns->mti_readers;
9471 for (i=0; i<rdrs; i++) {
9472 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9474 if (mdb_pid_insert(pids, pid) == 0) {
9475 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9477 /* Recheck, a new process may have reused pid */
9478 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9479 for (j=i; j<rdrs; j++)
9480 if (mr[j].mr_pid == pid) {
9481 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9482 (unsigned) pid, mr[j].mr_txnid));
9487 UNLOCK_MUTEX_R(env);