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)
93 #if defined(__sun) || defined(ANDROID)
94 /* Most platforms have posix_memalign, older may only have memalign */
95 #define HAVE_MEMALIGN 1
99 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
100 #include <netinet/in.h>
101 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
104 #if defined(__APPLE__) || defined (BSD)
105 # define MDB_USE_POSIX_SEM 1
106 # define MDB_FDATASYNC fsync
107 #elif defined(ANDROID)
108 # define MDB_FDATASYNC fsync
113 #ifdef MDB_USE_POSIX_SEM
114 # define MDB_USE_HASH 1
115 #include <semaphore.h>
120 #include <valgrind/memcheck.h>
121 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
122 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
123 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
124 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
125 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
127 #define VGMEMP_CREATE(h,r,z)
128 #define VGMEMP_ALLOC(h,a,s)
129 #define VGMEMP_FREE(h,a)
130 #define VGMEMP_DESTROY(h)
131 #define VGMEMP_DEFINED(a,s)
135 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
136 /* Solaris just defines one or the other */
137 # define LITTLE_ENDIAN 1234
138 # define BIG_ENDIAN 4321
139 # ifdef _LITTLE_ENDIAN
140 # define BYTE_ORDER LITTLE_ENDIAN
142 # define BYTE_ORDER BIG_ENDIAN
145 # define BYTE_ORDER __BYTE_ORDER
149 #ifndef LITTLE_ENDIAN
150 #define LITTLE_ENDIAN __LITTLE_ENDIAN
153 #define BIG_ENDIAN __BIG_ENDIAN
156 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
157 #define MISALIGNED_OK 1
163 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
164 # error "Unknown or unsupported endianness (BYTE_ORDER)"
165 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
166 # error "Two's complement, reasonably sized integer types, please"
170 /** Put infrequently used env functions in separate section */
172 # define ESECT __attribute__ ((section("__TEXT,text_env")))
174 # define ESECT __attribute__ ((section("text_env")))
180 /** @defgroup internal LMDB Internals
183 /** @defgroup compat Compatibility Macros
184 * A bunch of macros to minimize the amount of platform-specific ifdefs
185 * needed throughout the rest of the code. When the features this library
186 * needs are similar enough to POSIX to be hidden in a one-or-two line
187 * replacement, this macro approach is used.
191 /** Features under development */
196 /** Wrapper around __func__, which is a C99 feature */
197 #if __STDC_VERSION__ >= 199901L
198 # define mdb_func_ __func__
199 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
200 # define mdb_func_ __FUNCTION__
202 /* If a debug message says <mdb_unknown>(), update the #if statements above */
203 # define mdb_func_ "<mdb_unknown>"
207 #define MDB_USE_HASH 1
208 #define MDB_PIDLOCK 0
209 #define THREAD_RET DWORD
210 #define pthread_t HANDLE
211 #define pthread_mutex_t HANDLE
212 #define pthread_cond_t HANDLE
213 #define pthread_key_t DWORD
214 #define pthread_self() GetCurrentThreadId()
215 #define pthread_key_create(x,y) \
216 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
217 #define pthread_key_delete(x) TlsFree(x)
218 #define pthread_getspecific(x) TlsGetValue(x)
219 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
220 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
221 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
222 #define pthread_cond_signal(x) SetEvent(*x)
223 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
224 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
225 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
226 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
227 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
228 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
229 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
230 #define getpid() GetCurrentProcessId()
231 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
232 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
233 #define ErrCode() GetLastError()
234 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
235 #define close(fd) (CloseHandle(fd) ? 0 : -1)
236 #define munmap(ptr,len) UnmapViewOfFile(ptr)
237 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
238 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
240 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
244 #define THREAD_RET void *
245 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
246 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
247 #define Z "z" /**< printf format modifier for size_t */
249 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
250 #define MDB_PIDLOCK 1
252 #ifdef MDB_USE_POSIX_SEM
254 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
255 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
256 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
257 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
260 mdb_sem_wait(sem_t *sem)
263 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
268 /** Lock the reader mutex.
270 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
271 /** Unlock the reader mutex.
273 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
275 /** Lock the writer mutex.
276 * Only a single write transaction is allowed at a time. Other writers
277 * will block waiting for this mutex.
279 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
280 /** Unlock the writer mutex.
282 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
283 #endif /* MDB_USE_POSIX_SEM */
285 /** Get the error code for the last failed system function.
287 #define ErrCode() errno
289 /** An abstraction for a file handle.
290 * On POSIX systems file handles are small integers. On Windows
291 * they're opaque pointers.
295 /** A value for an invalid file handle.
296 * Mainly used to initialize file variables and signify that they are
299 #define INVALID_HANDLE_VALUE (-1)
301 /** Get the size of a memory page for the system.
302 * This is the basic size that the platform's memory manager uses, and is
303 * fundamental to the use of memory-mapped files.
305 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
308 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
311 #define MNAME_LEN (sizeof(pthread_mutex_t))
317 /** A flag for opening a file and requesting synchronous data writes.
318 * This is only used when writing a meta page. It's not strictly needed;
319 * we could just do a normal write and then immediately perform a flush.
320 * But if this flag is available it saves us an extra system call.
322 * @note If O_DSYNC is undefined but exists in /usr/include,
323 * preferably set some compiler flag to get the definition.
324 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
327 # define MDB_DSYNC O_DSYNC
331 /** Function for flushing the data of a file. Define this to fsync
332 * if fdatasync() is not supported.
334 #ifndef MDB_FDATASYNC
335 # define MDB_FDATASYNC fdatasync
339 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
350 /** A page number in the database.
351 * Note that 64 bit page numbers are overkill, since pages themselves
352 * already represent 12-13 bits of addressable memory, and the OS will
353 * always limit applications to a maximum of 63 bits of address space.
355 * @note In the #MDB_node structure, we only store 48 bits of this value,
356 * which thus limits us to only 60 bits of addressable data.
358 typedef MDB_ID pgno_t;
360 /** A transaction ID.
361 * See struct MDB_txn.mt_txnid for details.
363 typedef MDB_ID txnid_t;
365 /** @defgroup debug Debug Macros
369 /** Enable debug output. Needs variable argument macros (a C99 feature).
370 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
371 * read from and written to the database (used for free space management).
377 static int mdb_debug;
378 static txnid_t mdb_debug_start;
380 /** Print a debug message with printf formatting.
381 * Requires double parenthesis around 2 or more args.
383 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
384 # define DPRINTF0(fmt, ...) \
385 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
387 # define DPRINTF(args) ((void) 0)
389 /** Print a debug string.
390 * The string is printed literally, with no format processing.
392 #define DPUTS(arg) DPRINTF(("%s", arg))
393 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
395 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
398 /** @brief The maximum size of a database page.
400 * It is 32k or 64k, since value-PAGEBASE must fit in
401 * #MDB_page.%mp_upper.
403 * LMDB will use database pages < OS pages if needed.
404 * That causes more I/O in write transactions: The OS must
405 * know (read) the whole page before writing a partial page.
407 * Note that we don't currently support Huge pages. On Linux,
408 * regular data files cannot use Huge pages, and in general
409 * Huge pages aren't actually pageable. We rely on the OS
410 * demand-pager to read our data and page it out when memory
411 * pressure from other processes is high. So until OSs have
412 * actual paging support for Huge pages, they're not viable.
414 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
416 /** The minimum number of keys required in a database page.
417 * Setting this to a larger value will place a smaller bound on the
418 * maximum size of a data item. Data items larger than this size will
419 * be pushed into overflow pages instead of being stored directly in
420 * the B-tree node. This value used to default to 4. With a page size
421 * of 4096 bytes that meant that any item larger than 1024 bytes would
422 * go into an overflow page. That also meant that on average 2-3KB of
423 * each overflow page was wasted space. The value cannot be lower than
424 * 2 because then there would no longer be a tree structure. With this
425 * value, items larger than 2KB will go into overflow pages, and on
426 * average only 1KB will be wasted.
428 #define MDB_MINKEYS 2
430 /** A stamp that identifies a file as an LMDB file.
431 * There's nothing special about this value other than that it is easily
432 * recognizable, and it will reflect any byte order mismatches.
434 #define MDB_MAGIC 0xBEEFC0DE
436 /** The version number for a database's datafile format. */
437 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
438 /** The version number for a database's lockfile format. */
439 #define MDB_LOCK_VERSION 1
441 /** @brief The max size of a key we can write, or 0 for dynamic max.
443 * Define this as 0 to compute the max from the page size. 511
444 * is default for backwards compat: liblmdb <= 0.9.10 can break
445 * when modifying a DB with keys/dupsort data bigger than its max.
446 * #MDB_DEVEL sets the default to 0.
448 * Data items in an #MDB_DUPSORT database are also limited to
449 * this size, since they're actually keys of a sub-DB. Keys and
450 * #MDB_DUPSORT data items must fit on a node in a regular page.
452 #ifndef MDB_MAXKEYSIZE
453 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
456 /** The maximum size of a key we can write to the environment. */
458 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
460 #define ENV_MAXKEY(env) ((env)->me_maxkey)
463 /** @brief The maximum size of a data item.
465 * We only store a 32 bit value for node sizes.
467 #define MAXDATASIZE 0xffffffffUL
470 /** Key size which fits in a #DKBUF.
473 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
476 * This is used for printing a hex dump of a key's contents.
478 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
479 /** Display a key in hex.
481 * Invoke a function to display a key in hex.
483 #define DKEY(x) mdb_dkey(x, kbuf)
489 /** An invalid page number.
490 * Mainly used to denote an empty tree.
492 #define P_INVALID (~(pgno_t)0)
494 /** Test if the flags \b f are set in a flag word \b w. */
495 #define F_ISSET(w, f) (((w) & (f)) == (f))
497 /** Round \b n up to an even number. */
498 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
500 /** Used for offsets within a single page.
501 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
504 typedef uint16_t indx_t;
506 /** Default size of memory map.
507 * This is certainly too small for any actual applications. Apps should always set
508 * the size explicitly using #mdb_env_set_mapsize().
510 #define DEFAULT_MAPSIZE 1048576
512 /** @defgroup readers Reader Lock Table
513 * Readers don't acquire any locks for their data access. Instead, they
514 * simply record their transaction ID in the reader table. The reader
515 * mutex is needed just to find an empty slot in the reader table. The
516 * slot's address is saved in thread-specific data so that subsequent read
517 * transactions started by the same thread need no further locking to proceed.
519 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
521 * No reader table is used if the database is on a read-only filesystem, or
522 * if #MDB_NOLOCK is set.
524 * Since the database uses multi-version concurrency control, readers don't
525 * actually need any locking. This table is used to keep track of which
526 * readers are using data from which old transactions, so that we'll know
527 * when a particular old transaction is no longer in use. Old transactions
528 * that have discarded any data pages can then have those pages reclaimed
529 * for use by a later write transaction.
531 * The lock table is constructed such that reader slots are aligned with the
532 * processor's cache line size. Any slot is only ever used by one thread.
533 * This alignment guarantees that there will be no contention or cache
534 * thrashing as threads update their own slot info, and also eliminates
535 * any need for locking when accessing a slot.
537 * A writer thread will scan every slot in the table to determine the oldest
538 * outstanding reader transaction. Any freed pages older than this will be
539 * reclaimed by the writer. The writer doesn't use any locks when scanning
540 * this table. This means that there's no guarantee that the writer will
541 * see the most up-to-date reader info, but that's not required for correct
542 * operation - all we need is to know the upper bound on the oldest reader,
543 * we don't care at all about the newest reader. So the only consequence of
544 * reading stale information here is that old pages might hang around a
545 * while longer before being reclaimed. That's actually good anyway, because
546 * the longer we delay reclaiming old pages, the more likely it is that a
547 * string of contiguous pages can be found after coalescing old pages from
548 * many old transactions together.
551 /** Number of slots in the reader table.
552 * This value was chosen somewhat arbitrarily. 126 readers plus a
553 * couple mutexes fit exactly into 8KB on my development machine.
554 * Applications should set the table size using #mdb_env_set_maxreaders().
556 #define DEFAULT_READERS 126
558 /** The size of a CPU cache line in bytes. We want our lock structures
559 * aligned to this size to avoid false cache line sharing in the
561 * This value works for most CPUs. For Itanium this should be 128.
567 /** The information we store in a single slot of the reader table.
568 * In addition to a transaction ID, we also record the process and
569 * thread ID that owns a slot, so that we can detect stale information,
570 * e.g. threads or processes that went away without cleaning up.
571 * @note We currently don't check for stale records. We simply re-init
572 * the table when we know that we're the only process opening the
575 typedef struct MDB_rxbody {
576 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
577 * Multiple readers that start at the same time will probably have the
578 * same ID here. Again, it's not important to exclude them from
579 * anything; all we need to know is which version of the DB they
580 * started from so we can avoid overwriting any data used in that
581 * particular version.
584 /** The process ID of the process owning this reader txn. */
586 /** The thread ID of the thread owning this txn. */
590 /** The actual reader record, with cacheline padding. */
591 typedef struct MDB_reader {
594 /** shorthand for mrb_txnid */
595 #define mr_txnid mru.mrx.mrb_txnid
596 #define mr_pid mru.mrx.mrb_pid
597 #define mr_tid mru.mrx.mrb_tid
598 /** cache line alignment */
599 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
603 /** The header for the reader table.
604 * The table resides in a memory-mapped file. (This is a different file
605 * than is used for the main database.)
607 * For POSIX the actual mutexes reside in the shared memory of this
608 * mapped file. On Windows, mutexes are named objects allocated by the
609 * kernel; we store the mutex names in this mapped file so that other
610 * processes can grab them. This same approach is also used on
611 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
612 * process-shared POSIX mutexes. For these cases where a named object
613 * is used, the object name is derived from a 64 bit FNV hash of the
614 * environment pathname. As such, naming collisions are extremely
615 * unlikely. If a collision occurs, the results are unpredictable.
617 typedef struct MDB_txbody {
618 /** Stamp identifying this as an LMDB file. It must be set
621 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
623 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
624 char mtb_rmname[MNAME_LEN];
626 /** Mutex protecting access to this table.
627 * This is the reader lock that #LOCK_MUTEX_R acquires.
629 pthread_mutex_t mtb_mutex;
631 /** The ID of the last transaction committed to the database.
632 * This is recorded here only for convenience; the value can always
633 * be determined by reading the main database meta pages.
636 /** The number of slots that have been used in the reader table.
637 * This always records the maximum count, it is not decremented
638 * when readers release their slots.
640 unsigned mtb_numreaders;
643 /** The actual reader table definition. */
644 typedef struct MDB_txninfo {
647 #define mti_magic mt1.mtb.mtb_magic
648 #define mti_format mt1.mtb.mtb_format
649 #define mti_mutex mt1.mtb.mtb_mutex
650 #define mti_rmname mt1.mtb.mtb_rmname
651 #define mti_txnid mt1.mtb.mtb_txnid
652 #define mti_numreaders mt1.mtb.mtb_numreaders
653 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
656 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
657 char mt2_wmname[MNAME_LEN];
658 #define mti_wmname mt2.mt2_wmname
660 pthread_mutex_t mt2_wmutex;
661 #define mti_wmutex mt2.mt2_wmutex
663 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
665 MDB_reader mti_readers[1];
668 /** Lockfile format signature: version, features and field layout */
669 #define MDB_LOCK_FORMAT \
671 ((MDB_LOCK_VERSION) \
672 /* Flags which describe functionality */ \
673 + (((MDB_PIDLOCK) != 0) << 16)))
676 /** Common header for all page types.
677 * Overflow records occupy a number of contiguous pages with no
678 * headers on any page after the first.
680 typedef struct MDB_page {
681 #define mp_pgno mp_p.p_pgno
682 #define mp_next mp_p.p_next
684 pgno_t p_pgno; /**< page number */
685 struct MDB_page *p_next; /**< for in-memory list of freed pages */
688 /** @defgroup mdb_page Page Flags
690 * Flags for the page headers.
693 #define P_BRANCH 0x01 /**< branch page */
694 #define P_LEAF 0x02 /**< leaf page */
695 #define P_OVERFLOW 0x04 /**< overflow page */
696 #define P_META 0x08 /**< meta page */
697 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
698 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
699 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
700 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
701 #define P_KEEP 0x8000 /**< leave this page alone during spill */
703 uint16_t mp_flags; /**< @ref mdb_page */
704 #define mp_lower mp_pb.pb.pb_lower
705 #define mp_upper mp_pb.pb.pb_upper
706 #define mp_pages mp_pb.pb_pages
709 indx_t pb_lower; /**< lower bound of free space */
710 indx_t pb_upper; /**< upper bound of free space */
712 uint32_t pb_pages; /**< number of overflow pages */
714 indx_t mp_ptrs[1]; /**< dynamic size */
717 /** Size of the page header, excluding dynamic data at the end */
718 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
720 /** Address of first usable data byte in a page, after the header */
721 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
723 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
724 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
726 /** Number of nodes on a page */
727 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
729 /** The amount of space remaining in the page */
730 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
732 /** The percentage of space used in the page, in tenths of a percent. */
733 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
734 ((env)->me_psize - PAGEHDRSZ))
735 /** The minimum page fill factor, in tenths of a percent.
736 * Pages emptier than this are candidates for merging.
738 #define FILL_THRESHOLD 250
740 /** Test if a page is a leaf page */
741 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
742 /** Test if a page is a LEAF2 page */
743 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
744 /** Test if a page is a branch page */
745 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
746 /** Test if a page is an overflow page */
747 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
748 /** Test if a page is a sub page */
749 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
751 /** The number of overflow pages needed to store the given size. */
752 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
754 /** Link in #MDB_txn.%mt_loose_pgs list */
755 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
757 /** Header for a single key/data pair within a page.
758 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
759 * We guarantee 2-byte alignment for 'MDB_node's.
761 typedef struct MDB_node {
762 /** lo and hi are used for data size on leaf nodes and for
763 * child pgno on branch nodes. On 64 bit platforms, flags
764 * is also used for pgno. (Branch nodes have no flags).
765 * They are in host byte order in case that lets some
766 * accesses be optimized into a 32-bit word access.
768 #if BYTE_ORDER == LITTLE_ENDIAN
769 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
771 unsigned short mn_hi, mn_lo;
773 /** @defgroup mdb_node Node Flags
775 * Flags for node headers.
778 #define F_BIGDATA 0x01 /**< data put on overflow page */
779 #define F_SUBDATA 0x02 /**< data is a sub-database */
780 #define F_DUPDATA 0x04 /**< data has duplicates */
782 /** valid flags for #mdb_node_add() */
783 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
786 unsigned short mn_flags; /**< @ref mdb_node */
787 unsigned short mn_ksize; /**< key size */
788 char mn_data[1]; /**< key and data are appended here */
791 /** Size of the node header, excluding dynamic data at the end */
792 #define NODESIZE offsetof(MDB_node, mn_data)
794 /** Bit position of top word in page number, for shifting mn_flags */
795 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
797 /** Size of a node in a branch page with a given key.
798 * This is just the node header plus the key, there is no data.
800 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
802 /** Size of a node in a leaf page with a given key and data.
803 * This is node header plus key plus data size.
805 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
807 /** Address of node \b i in page \b p */
808 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
810 /** Address of the key for the node */
811 #define NODEKEY(node) (void *)((node)->mn_data)
813 /** Address of the data for a node */
814 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
816 /** Get the page number pointed to by a branch node */
817 #define NODEPGNO(node) \
818 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
819 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
820 /** Set the page number in a branch node */
821 #define SETPGNO(node,pgno) do { \
822 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
823 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
825 /** Get the size of the data in a leaf node */
826 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
827 /** Set the size of the data for a leaf node */
828 #define SETDSZ(node,size) do { \
829 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
830 /** The size of a key in a node */
831 #define NODEKSZ(node) ((node)->mn_ksize)
833 /** Copy a page number from src to dst */
835 #define COPY_PGNO(dst,src) dst = src
837 #if SIZE_MAX > 4294967295UL
838 #define COPY_PGNO(dst,src) do { \
839 unsigned short *s, *d; \
840 s = (unsigned short *)&(src); \
841 d = (unsigned short *)&(dst); \
848 #define COPY_PGNO(dst,src) do { \
849 unsigned short *s, *d; \
850 s = (unsigned short *)&(src); \
851 d = (unsigned short *)&(dst); \
857 /** The address of a key in a LEAF2 page.
858 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
859 * There are no node headers, keys are stored contiguously.
861 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
863 /** Set the \b node's key into \b keyptr, if requested. */
864 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
865 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
867 /** Set the \b node's key into \b key. */
868 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
870 /** Information about a single database in the environment. */
871 typedef struct MDB_db {
872 uint32_t md_pad; /**< also ksize for LEAF2 pages */
873 uint16_t md_flags; /**< @ref mdb_dbi_open */
874 uint16_t md_depth; /**< depth of this tree */
875 pgno_t md_branch_pages; /**< number of internal pages */
876 pgno_t md_leaf_pages; /**< number of leaf pages */
877 pgno_t md_overflow_pages; /**< number of overflow pages */
878 size_t md_entries; /**< number of data items */
879 pgno_t md_root; /**< the root page of this tree */
882 /** mdb_dbi_open flags */
883 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
884 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
885 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
886 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
888 /** Handle for the DB used to track free pages. */
890 /** Handle for the default DB. */
893 /** Meta page content.
894 * A meta page is the start point for accessing a database snapshot.
895 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
897 typedef struct MDB_meta {
898 /** Stamp identifying this as an LMDB file. It must be set
901 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
903 void *mm_address; /**< address for fixed mapping */
904 size_t mm_mapsize; /**< size of mmap region */
905 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
906 /** The size of pages used in this DB */
907 #define mm_psize mm_dbs[0].md_pad
908 /** Any persistent environment flags. @ref mdb_env */
909 #define mm_flags mm_dbs[0].md_flags
910 pgno_t mm_last_pg; /**< last used page in file */
911 txnid_t mm_txnid; /**< txnid that committed this page */
914 /** Buffer for a stack-allocated meta page.
915 * The members define size and alignment, and silence type
916 * aliasing warnings. They are not used directly; that could
917 * mean incorrectly using several union members in parallel.
919 typedef union MDB_metabuf {
922 char mm_pad[PAGEHDRSZ];
927 /** Auxiliary DB info.
928 * The information here is mostly static/read-only. There is
929 * only a single copy of this record in the environment.
931 typedef struct MDB_dbx {
932 MDB_val md_name; /**< name of the database */
933 MDB_cmp_func *md_cmp; /**< function for comparing keys */
934 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
935 MDB_rel_func *md_rel; /**< user relocate function */
936 void *md_relctx; /**< user-provided context for md_rel */
939 /** A database transaction.
940 * Every operation requires a transaction handle.
943 MDB_txn *mt_parent; /**< parent of a nested txn */
944 MDB_txn *mt_child; /**< nested txn under this txn */
945 pgno_t mt_next_pgno; /**< next unallocated page */
946 /** The ID of this transaction. IDs are integers incrementing from 1.
947 * Only committed write transactions increment the ID. If a transaction
948 * aborts, the ID may be re-used by the next writer.
951 MDB_env *mt_env; /**< the DB environment */
952 /** The list of pages that became unused during this transaction.
955 /** The list of loose pages that became unused and may be reused
956 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
958 MDB_page *mt_loose_pgs;
959 /* #Number of loose pages (#mt_loose_pgs) */
961 /** The sorted list of dirty pages we temporarily wrote to disk
962 * because the dirty list was full. page numbers in here are
963 * shifted left by 1, deleted slots have the LSB set.
965 MDB_IDL mt_spill_pgs;
967 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
969 /** For read txns: This thread/txn's reader table slot, or NULL. */
972 /** Array of records for each DB known in the environment. */
974 /** Array of MDB_db records for each known DB */
976 /** Array of sequence numbers for each DB handle */
977 unsigned int *mt_dbiseqs;
978 /** @defgroup mt_dbflag Transaction DB Flags
982 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
983 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
984 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
985 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
987 /** In write txns, array of cursors for each DB */
988 MDB_cursor **mt_cursors;
989 /** Array of flags for each DB */
990 unsigned char *mt_dbflags;
991 /** Number of DB records in use. This number only ever increments;
992 * we don't decrement it when individual DB handles are closed.
996 /** @defgroup mdb_txn Transaction Flags
1000 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
1001 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1002 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1003 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1005 unsigned int mt_flags; /**< @ref mdb_txn */
1006 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1007 * Includes ancestor txns' dirty pages not hidden by other txns'
1008 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1009 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1011 unsigned int mt_dirty_room;
1014 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1015 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1016 * raise this on a 64 bit machine.
1018 #define CURSOR_STACK 32
1022 /** Cursors are used for all DB operations.
1023 * A cursor holds a path of (page pointer, key index) from the DB
1024 * root to a position in the DB, plus other state. #MDB_DUPSORT
1025 * cursors include an xcursor to the current data item. Write txns
1026 * track their cursors and keep them up to date when data moves.
1027 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1028 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1031 /** Next cursor on this DB in this txn */
1032 MDB_cursor *mc_next;
1033 /** Backup of the original cursor if this cursor is a shadow */
1034 MDB_cursor *mc_backup;
1035 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1036 struct MDB_xcursor *mc_xcursor;
1037 /** The transaction that owns this cursor */
1039 /** The database handle this cursor operates on */
1041 /** The database record for this cursor */
1043 /** The database auxiliary record for this cursor */
1045 /** The @ref mt_dbflag for this database */
1046 unsigned char *mc_dbflag;
1047 unsigned short mc_snum; /**< number of pushed pages */
1048 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1049 /** @defgroup mdb_cursor Cursor Flags
1051 * Cursor state flags.
1054 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1055 #define C_EOF 0x02 /**< No more data */
1056 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1057 #define C_DEL 0x08 /**< last op was a cursor_del */
1058 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1059 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1061 unsigned int mc_flags; /**< @ref mdb_cursor */
1062 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1063 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1066 /** Context for sorted-dup records.
1067 * We could have gone to a fully recursive design, with arbitrarily
1068 * deep nesting of sub-databases. But for now we only handle these
1069 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1071 typedef struct MDB_xcursor {
1072 /** A sub-cursor for traversing the Dup DB */
1073 MDB_cursor mx_cursor;
1074 /** The database record for this Dup DB */
1076 /** The auxiliary DB record for this Dup DB */
1078 /** The @ref mt_dbflag for this Dup DB */
1079 unsigned char mx_dbflag;
1082 /** State of FreeDB old pages, stored in the MDB_env */
1083 typedef struct MDB_pgstate {
1084 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1085 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1088 /** The database environment. */
1090 HANDLE me_fd; /**< The main data file */
1091 HANDLE me_lfd; /**< The lock file */
1092 HANDLE me_mfd; /**< just for writing the meta pages */
1093 /** Failed to update the meta page. Probably an I/O error. */
1094 #define MDB_FATAL_ERROR 0x80000000U
1095 /** Some fields are initialized. */
1096 #define MDB_ENV_ACTIVE 0x20000000U
1097 /** me_txkey is set */
1098 #define MDB_ENV_TXKEY 0x10000000U
1099 uint32_t me_flags; /**< @ref mdb_env */
1100 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1101 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1102 unsigned int me_maxreaders; /**< size of the reader table */
1103 unsigned int me_numreaders; /**< max numreaders set by this env */
1104 MDB_dbi me_numdbs; /**< number of DBs opened */
1105 MDB_dbi me_maxdbs; /**< size of the DB table */
1106 MDB_PID_T me_pid; /**< process ID of this env */
1107 char *me_path; /**< path to the DB files */
1108 char *me_map; /**< the memory map of the data file */
1109 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1110 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1111 void *me_pbuf; /**< scratch area for DUPSORT put() */
1112 MDB_txn *me_txn; /**< current write transaction */
1113 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1114 size_t me_mapsize; /**< size of the data memory map */
1115 off_t me_size; /**< current file size */
1116 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1117 MDB_dbx *me_dbxs; /**< array of static DB info */
1118 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1119 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1120 pthread_key_t me_txkey; /**< thread-key for readers */
1121 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1122 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1123 # define me_pglast me_pgstate.mf_pglast
1124 # define me_pghead me_pgstate.mf_pghead
1125 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1126 /** IDL of pages that became unused in a write txn */
1127 MDB_IDL me_free_pgs;
1128 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1129 MDB_ID2L me_dirty_list;
1130 /** Max number of freelist items that can fit in a single overflow page */
1132 /** Max size of a node on a page */
1133 unsigned int me_nodemax;
1134 #if !(MDB_MAXKEYSIZE)
1135 unsigned int me_maxkey; /**< max size of a key */
1137 int me_live_reader; /**< have liveness lock in reader table */
1139 int me_pidquery; /**< Used in OpenProcess */
1140 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1142 #elif defined(MDB_USE_POSIX_SEM)
1143 sem_t *me_rmutex; /* Shared mutexes are not supported */
1147 int me_fsynconly; /**< fdatasync is unreliable */
1149 void *me_userctx; /**< User-settable context */
1150 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1153 /** Nested transaction */
1154 typedef struct MDB_ntxn {
1155 MDB_txn mnt_txn; /**< the transaction */
1156 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1159 /** max number of pages to commit in one writev() call */
1160 #define MDB_COMMIT_PAGES 64
1161 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1162 #undef MDB_COMMIT_PAGES
1163 #define MDB_COMMIT_PAGES IOV_MAX
1166 /** max bytes to write in one call */
1167 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1169 /** Check \b txn and \b dbi arguments to a function */
1170 #define TXN_DBI_EXIST(txn, dbi) \
1171 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1173 /** Check for misused \b dbi handles */
1174 #define TXN_DBI_CHANGED(txn, dbi) \
1175 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1177 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1178 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1179 static int mdb_page_touch(MDB_cursor *mc);
1181 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1182 static int mdb_page_search_root(MDB_cursor *mc,
1183 MDB_val *key, int modify);
1184 #define MDB_PS_MODIFY 1
1185 #define MDB_PS_ROOTONLY 2
1186 #define MDB_PS_FIRST 4
1187 #define MDB_PS_LAST 8
1188 static int mdb_page_search(MDB_cursor *mc,
1189 MDB_val *key, int flags);
1190 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1192 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1193 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1194 pgno_t newpgno, unsigned int nflags);
1196 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1197 static int mdb_env_pick_meta(const MDB_env *env);
1198 static int mdb_env_write_meta(MDB_txn *txn);
1199 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1200 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1202 static void mdb_env_close0(MDB_env *env, int excl);
1204 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1205 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1206 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1207 static void mdb_node_del(MDB_cursor *mc, int ksize);
1208 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1209 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1210 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1211 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1212 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1214 static int mdb_rebalance(MDB_cursor *mc);
1215 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1217 static void mdb_cursor_pop(MDB_cursor *mc);
1218 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1220 static int mdb_cursor_del0(MDB_cursor *mc);
1221 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1222 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1223 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1224 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1225 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1227 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1228 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1230 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1231 static void mdb_xcursor_init0(MDB_cursor *mc);
1232 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1234 static int mdb_drop0(MDB_cursor *mc, int subs);
1235 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1238 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1242 static SECURITY_DESCRIPTOR mdb_null_sd;
1243 static SECURITY_ATTRIBUTES mdb_all_sa;
1244 static int mdb_sec_inited;
1247 /** Return the library version info. */
1249 mdb_version(int *major, int *minor, int *patch)
1251 if (major) *major = MDB_VERSION_MAJOR;
1252 if (minor) *minor = MDB_VERSION_MINOR;
1253 if (patch) *patch = MDB_VERSION_PATCH;
1254 return MDB_VERSION_STRING;
1257 /** Table of descriptions for LMDB @ref errors */
1258 static char *const mdb_errstr[] = {
1259 "MDB_KEYEXIST: Key/data pair already exists",
1260 "MDB_NOTFOUND: No matching key/data pair found",
1261 "MDB_PAGE_NOTFOUND: Requested page not found",
1262 "MDB_CORRUPTED: Located page was wrong type",
1263 "MDB_PANIC: Update of meta page failed",
1264 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1265 "MDB_INVALID: File is not an LMDB file",
1266 "MDB_MAP_FULL: Environment mapsize limit reached",
1267 "MDB_DBS_FULL: Environment maxdbs limit reached",
1268 "MDB_READERS_FULL: Environment maxreaders limit reached",
1269 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1270 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1271 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1272 "MDB_PAGE_FULL: Internal error - page has no more space",
1273 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1274 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1275 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1276 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1277 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1278 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1282 mdb_strerror(int err)
1285 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1286 * This works as long as no function between the call to mdb_strerror
1287 * and the actual use of the message uses more than 4K of stack.
1290 char buf[1024], *ptr = buf;
1294 return ("Successful return: 0");
1296 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1297 i = err - MDB_KEYEXIST;
1298 return mdb_errstr[i];
1302 /* These are the C-runtime error codes we use. The comment indicates
1303 * their numeric value, and the Win32 error they would correspond to
1304 * if the error actually came from a Win32 API. A major mess, we should
1305 * have used LMDB-specific error codes for everything.
1308 case ENOENT: /* 2, FILE_NOT_FOUND */
1309 case EIO: /* 5, ACCESS_DENIED */
1310 case ENOMEM: /* 12, INVALID_ACCESS */
1311 case EACCES: /* 13, INVALID_DATA */
1312 case EBUSY: /* 16, CURRENT_DIRECTORY */
1313 case EINVAL: /* 22, BAD_COMMAND */
1314 case ENOSPC: /* 28, OUT_OF_PAPER */
1315 return strerror(err);
1320 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1321 FORMAT_MESSAGE_IGNORE_INSERTS,
1322 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1325 return strerror(err);
1329 /** assert(3) variant in cursor context */
1330 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1331 /** assert(3) variant in transaction context */
1332 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1333 /** assert(3) variant in environment context */
1334 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1337 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1338 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1341 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1342 const char *func, const char *file, int line)
1345 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1346 file, line, expr_txt, func);
1347 if (env->me_assert_func)
1348 env->me_assert_func(env, buf);
1349 fprintf(stderr, "%s\n", buf);
1353 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1357 /** Return the page number of \b mp which may be sub-page, for debug output */
1359 mdb_dbg_pgno(MDB_page *mp)
1362 COPY_PGNO(ret, mp->mp_pgno);
1366 /** Display a key in hexadecimal and return the address of the result.
1367 * @param[in] key the key to display
1368 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1369 * @return The key in hexadecimal form.
1372 mdb_dkey(MDB_val *key, char *buf)
1375 unsigned char *c = key->mv_data;
1381 if (key->mv_size > DKBUF_MAXKEYSIZE)
1382 return "MDB_MAXKEYSIZE";
1383 /* may want to make this a dynamic check: if the key is mostly
1384 * printable characters, print it as-is instead of converting to hex.
1388 for (i=0; i<key->mv_size; i++)
1389 ptr += sprintf(ptr, "%02x", *c++);
1391 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1397 mdb_leafnode_type(MDB_node *n)
1399 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1400 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1401 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1404 /** Display all the keys in the page. */
1406 mdb_page_list(MDB_page *mp)
1408 pgno_t pgno = mdb_dbg_pgno(mp);
1409 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1411 unsigned int i, nkeys, nsize, total = 0;
1415 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1416 case P_BRANCH: type = "Branch page"; break;
1417 case P_LEAF: type = "Leaf page"; break;
1418 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1419 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1420 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1422 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1423 pgno, mp->mp_pages, state);
1426 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1427 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1430 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1434 nkeys = NUMKEYS(mp);
1435 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1437 for (i=0; i<nkeys; i++) {
1438 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1439 key.mv_size = nsize = mp->mp_pad;
1440 key.mv_data = LEAF2KEY(mp, i, nsize);
1442 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1445 node = NODEPTR(mp, i);
1446 key.mv_size = node->mn_ksize;
1447 key.mv_data = node->mn_data;
1448 nsize = NODESIZE + key.mv_size;
1449 if (IS_BRANCH(mp)) {
1450 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1454 if (F_ISSET(node->mn_flags, F_BIGDATA))
1455 nsize += sizeof(pgno_t);
1457 nsize += NODEDSZ(node);
1459 nsize += sizeof(indx_t);
1460 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1461 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1463 total = EVEN(total);
1465 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1466 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1470 mdb_cursor_chk(MDB_cursor *mc)
1476 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1477 for (i=0; i<mc->mc_top; i++) {
1479 node = NODEPTR(mp, mc->mc_ki[i]);
1480 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1483 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1489 /** Count all the pages in each DB and in the freelist
1490 * and make sure it matches the actual number of pages
1492 * All named DBs must be open for a correct count.
1494 static void mdb_audit(MDB_txn *txn)
1498 MDB_ID freecount, count;
1503 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1504 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1505 freecount += *(MDB_ID *)data.mv_data;
1506 mdb_tassert(txn, rc == MDB_NOTFOUND);
1509 for (i = 0; i<txn->mt_numdbs; i++) {
1511 if (!(txn->mt_dbflags[i] & DB_VALID))
1513 mdb_cursor_init(&mc, txn, i, &mx);
1514 if (txn->mt_dbs[i].md_root == P_INVALID)
1516 count += txn->mt_dbs[i].md_branch_pages +
1517 txn->mt_dbs[i].md_leaf_pages +
1518 txn->mt_dbs[i].md_overflow_pages;
1519 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1520 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1521 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1524 mp = mc.mc_pg[mc.mc_top];
1525 for (j=0; j<NUMKEYS(mp); j++) {
1526 MDB_node *leaf = NODEPTR(mp, j);
1527 if (leaf->mn_flags & F_SUBDATA) {
1529 memcpy(&db, NODEDATA(leaf), sizeof(db));
1530 count += db.md_branch_pages + db.md_leaf_pages +
1531 db.md_overflow_pages;
1535 mdb_tassert(txn, rc == MDB_NOTFOUND);
1538 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1539 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1540 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1546 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1548 return txn->mt_dbxs[dbi].md_cmp(a, b);
1552 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1554 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1557 /** Allocate memory for a page.
1558 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1561 mdb_page_malloc(MDB_txn *txn, unsigned num)
1563 MDB_env *env = txn->mt_env;
1564 MDB_page *ret = env->me_dpages;
1565 size_t psize = env->me_psize, sz = psize, off;
1566 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1567 * For a single page alloc, we init everything after the page header.
1568 * For multi-page, we init the final page; if the caller needed that
1569 * many pages they will be filling in at least up to the last page.
1573 VGMEMP_ALLOC(env, ret, sz);
1574 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1575 env->me_dpages = ret->mp_next;
1578 psize -= off = PAGEHDRSZ;
1583 if ((ret = malloc(sz)) != NULL) {
1584 VGMEMP_ALLOC(env, ret, sz);
1585 if (!(env->me_flags & MDB_NOMEMINIT)) {
1586 memset((char *)ret + off, 0, psize);
1590 txn->mt_flags |= MDB_TXN_ERROR;
1594 /** Free a single page.
1595 * Saves single pages to a list, for future reuse.
1596 * (This is not used for multi-page overflow pages.)
1599 mdb_page_free(MDB_env *env, MDB_page *mp)
1601 mp->mp_next = env->me_dpages;
1602 VGMEMP_FREE(env, mp);
1603 env->me_dpages = mp;
1606 /** Free a dirty page */
1608 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1610 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1611 mdb_page_free(env, dp);
1613 /* large pages just get freed directly */
1614 VGMEMP_FREE(env, dp);
1619 /** Return all dirty pages to dpage list */
1621 mdb_dlist_free(MDB_txn *txn)
1623 MDB_env *env = txn->mt_env;
1624 MDB_ID2L dl = txn->mt_u.dirty_list;
1625 unsigned i, n = dl[0].mid;
1627 for (i = 1; i <= n; i++) {
1628 mdb_dpage_free(env, dl[i].mptr);
1633 /** Loosen or free a single page.
1634 * Saves single pages to a list for future reuse
1635 * in this same txn. It has been pulled from the freeDB
1636 * and already resides on the dirty list, but has been
1637 * deleted. Use these pages first before pulling again
1640 * If the page wasn't dirtied in this txn, just add it
1641 * to this txn's free list.
1644 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1647 pgno_t pgno = mp->mp_pgno;
1648 MDB_txn *txn = mc->mc_txn;
1650 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1651 if (txn->mt_parent) {
1652 MDB_ID2 *dl = txn->mt_u.dirty_list;
1653 /* If txn has a parent, make sure the page is in our
1657 unsigned x = mdb_mid2l_search(dl, pgno);
1658 if (x <= dl[0].mid && dl[x].mid == pgno) {
1659 if (mp != dl[x].mptr) { /* bad cursor? */
1660 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1661 txn->mt_flags |= MDB_TXN_ERROR;
1662 return MDB_CORRUPTED;
1669 /* no parent txn, so it's just ours */
1674 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1676 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1677 txn->mt_loose_pgs = mp;
1678 txn->mt_loose_count++;
1679 mp->mp_flags |= P_LOOSE;
1681 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1689 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1690 * @param[in] mc A cursor handle for the current operation.
1691 * @param[in] pflags Flags of the pages to update:
1692 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1693 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1694 * @return 0 on success, non-zero on failure.
1697 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1699 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1700 MDB_txn *txn = mc->mc_txn;
1706 int rc = MDB_SUCCESS, level;
1708 /* Mark pages seen by cursors */
1709 if (mc->mc_flags & C_UNTRACK)
1710 mc = NULL; /* will find mc in mt_cursors */
1711 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1712 for (; mc; mc=mc->mc_next) {
1713 if (!(mc->mc_flags & C_INITIALIZED))
1715 for (m3 = mc;; m3 = &mx->mx_cursor) {
1717 for (j=0; j<m3->mc_snum; j++) {
1719 if ((mp->mp_flags & Mask) == pflags)
1720 mp->mp_flags ^= P_KEEP;
1722 mx = m3->mc_xcursor;
1723 /* Proceed to mx if it is at a sub-database */
1724 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1726 if (! (mp && (mp->mp_flags & P_LEAF)))
1728 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1729 if (!(leaf->mn_flags & F_SUBDATA))
1738 /* Mark dirty root pages */
1739 for (i=0; i<txn->mt_numdbs; i++) {
1740 if (txn->mt_dbflags[i] & DB_DIRTY) {
1741 pgno_t pgno = txn->mt_dbs[i].md_root;
1742 if (pgno == P_INVALID)
1744 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1746 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1747 dp->mp_flags ^= P_KEEP;
1755 static int mdb_page_flush(MDB_txn *txn, int keep);
1757 /** Spill pages from the dirty list back to disk.
1758 * This is intended to prevent running into #MDB_TXN_FULL situations,
1759 * but note that they may still occur in a few cases:
1760 * 1) our estimate of the txn size could be too small. Currently this
1761 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1762 * 2) child txns may run out of space if their parents dirtied a
1763 * lot of pages and never spilled them. TODO: we probably should do
1764 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1765 * the parent's dirty_room is below a given threshold.
1767 * Otherwise, if not using nested txns, it is expected that apps will
1768 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1769 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1770 * If the txn never references them again, they can be left alone.
1771 * If the txn only reads them, they can be used without any fuss.
1772 * If the txn writes them again, they can be dirtied immediately without
1773 * going thru all of the work of #mdb_page_touch(). Such references are
1774 * handled by #mdb_page_unspill().
1776 * Also note, we never spill DB root pages, nor pages of active cursors,
1777 * because we'll need these back again soon anyway. And in nested txns,
1778 * we can't spill a page in a child txn if it was already spilled in a
1779 * parent txn. That would alter the parent txns' data even though
1780 * the child hasn't committed yet, and we'd have no way to undo it if
1781 * the child aborted.
1783 * @param[in] m0 cursor A cursor handle identifying the transaction and
1784 * database for which we are checking space.
1785 * @param[in] key For a put operation, the key being stored.
1786 * @param[in] data For a put operation, the data being stored.
1787 * @return 0 on success, non-zero on failure.
1790 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1792 MDB_txn *txn = m0->mc_txn;
1794 MDB_ID2L dl = txn->mt_u.dirty_list;
1795 unsigned int i, j, need;
1798 if (m0->mc_flags & C_SUB)
1801 /* Estimate how much space this op will take */
1802 i = m0->mc_db->md_depth;
1803 /* Named DBs also dirty the main DB */
1804 if (m0->mc_dbi > MAIN_DBI)
1805 i += txn->mt_dbs[MAIN_DBI].md_depth;
1806 /* For puts, roughly factor in the key+data size */
1808 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1809 i += i; /* double it for good measure */
1812 if (txn->mt_dirty_room > i)
1815 if (!txn->mt_spill_pgs) {
1816 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1817 if (!txn->mt_spill_pgs)
1820 /* purge deleted slots */
1821 MDB_IDL sl = txn->mt_spill_pgs;
1822 unsigned int num = sl[0];
1824 for (i=1; i<=num; i++) {
1831 /* Preserve pages which may soon be dirtied again */
1832 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1835 /* Less aggressive spill - we originally spilled the entire dirty list,
1836 * with a few exceptions for cursor pages and DB root pages. But this
1837 * turns out to be a lot of wasted effort because in a large txn many
1838 * of those pages will need to be used again. So now we spill only 1/8th
1839 * of the dirty pages. Testing revealed this to be a good tradeoff,
1840 * better than 1/2, 1/4, or 1/10.
1842 if (need < MDB_IDL_UM_MAX / 8)
1843 need = MDB_IDL_UM_MAX / 8;
1845 /* Save the page IDs of all the pages we're flushing */
1846 /* flush from the tail forward, this saves a lot of shifting later on. */
1847 for (i=dl[0].mid; i && need; i--) {
1848 MDB_ID pn = dl[i].mid << 1;
1850 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1852 /* Can't spill twice, make sure it's not already in a parent's
1855 if (txn->mt_parent) {
1857 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1858 if (tx2->mt_spill_pgs) {
1859 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1860 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1861 dp->mp_flags |= P_KEEP;
1869 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1873 mdb_midl_sort(txn->mt_spill_pgs);
1875 /* Flush the spilled part of dirty list */
1876 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1879 /* Reset any dirty pages we kept that page_flush didn't see */
1880 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1883 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1887 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1889 mdb_find_oldest(MDB_txn *txn)
1892 txnid_t mr, oldest = txn->mt_txnid - 1;
1893 if (txn->mt_env->me_txns) {
1894 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1895 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1906 /** Add a page to the txn's dirty list */
1908 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1911 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1913 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1914 insert = mdb_mid2l_append;
1916 insert = mdb_mid2l_insert;
1918 mid.mid = mp->mp_pgno;
1920 rc = insert(txn->mt_u.dirty_list, &mid);
1921 mdb_tassert(txn, rc == 0);
1922 txn->mt_dirty_room--;
1925 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1926 * me_pghead and mt_next_pgno.
1928 * If there are free pages available from older transactions, they
1929 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1930 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1931 * and move me_pglast to say which records were consumed. Only this
1932 * function can create me_pghead and move me_pglast/mt_next_pgno.
1933 * @param[in] mc cursor A cursor handle identifying the transaction and
1934 * database for which we are allocating.
1935 * @param[in] num the number of pages to allocate.
1936 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1937 * will always be satisfied by a single contiguous chunk of memory.
1938 * @return 0 on success, non-zero on failure.
1941 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1943 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1944 /* Get at most <Max_retries> more freeDB records once me_pghead
1945 * has enough pages. If not enough, use new pages from the map.
1946 * If <Paranoid> and mc is updating the freeDB, only get new
1947 * records if me_pghead is empty. Then the freelist cannot play
1948 * catch-up with itself by growing while trying to save it.
1950 enum { Paranoid = 1, Max_retries = 500 };
1952 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1954 int rc, retry = num * 60;
1955 MDB_txn *txn = mc->mc_txn;
1956 MDB_env *env = txn->mt_env;
1957 pgno_t pgno, *mop = env->me_pghead;
1958 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1960 txnid_t oldest = 0, last;
1965 /* If there are any loose pages, just use them */
1966 if (num == 1 && txn->mt_loose_pgs) {
1967 np = txn->mt_loose_pgs;
1968 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1969 txn->mt_loose_count--;
1970 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1978 /* If our dirty list is already full, we can't do anything */
1979 if (txn->mt_dirty_room == 0) {
1984 for (op = MDB_FIRST;; op = MDB_NEXT) {
1989 /* Seek a big enough contiguous page range. Prefer
1990 * pages at the tail, just truncating the list.
1996 if (mop[i-n2] == pgno+n2)
2003 if (op == MDB_FIRST) { /* 1st iteration */
2004 /* Prepare to fetch more and coalesce */
2005 last = env->me_pglast;
2006 oldest = env->me_pgoldest;
2007 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2010 key.mv_data = &last; /* will look up last+1 */
2011 key.mv_size = sizeof(last);
2013 if (Paranoid && mc->mc_dbi == FREE_DBI)
2016 if (Paranoid && retry < 0 && mop_len)
2020 /* Do not fetch more if the record will be too recent */
2021 if (oldest <= last) {
2023 oldest = mdb_find_oldest(txn);
2024 env->me_pgoldest = oldest;
2030 rc = mdb_cursor_get(&m2, &key, NULL, op);
2032 if (rc == MDB_NOTFOUND)
2036 last = *(txnid_t*)key.mv_data;
2037 if (oldest <= last) {
2039 oldest = mdb_find_oldest(txn);
2040 env->me_pgoldest = oldest;
2046 np = m2.mc_pg[m2.mc_top];
2047 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2048 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2051 idl = (MDB_ID *) data.mv_data;
2054 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2059 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2061 mop = env->me_pghead;
2063 env->me_pglast = last;
2065 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2066 last, txn->mt_dbs[FREE_DBI].md_root, i));
2068 DPRINTF(("IDL %"Z"u", idl[j]));
2070 /* Merge in descending sorted order */
2071 mdb_midl_xmerge(mop, idl);
2075 /* Use new pages from the map when nothing suitable in the freeDB */
2077 pgno = txn->mt_next_pgno;
2078 if (pgno + num >= env->me_maxpg) {
2079 DPUTS("DB size maxed out");
2085 if (env->me_flags & MDB_WRITEMAP) {
2086 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2088 if (!(np = mdb_page_malloc(txn, num))) {
2094 mop[0] = mop_len -= num;
2095 /* Move any stragglers down */
2096 for (j = i-num; j < mop_len; )
2097 mop[++j] = mop[++i];
2099 txn->mt_next_pgno = pgno + num;
2102 mdb_page_dirty(txn, np);
2108 txn->mt_flags |= MDB_TXN_ERROR;
2112 /** Copy the used portions of a non-overflow page.
2113 * @param[in] dst page to copy into
2114 * @param[in] src page to copy from
2115 * @param[in] psize size of a page
2118 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2120 enum { Align = sizeof(pgno_t) };
2121 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2123 /* If page isn't full, just copy the used portion. Adjust
2124 * alignment so memcpy may copy words instead of bytes.
2126 if ((unused &= -Align) && !IS_LEAF2(src)) {
2127 upper = (upper + PAGEBASE) & -Align;
2128 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2129 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2132 memcpy(dst, src, psize - unused);
2136 /** Pull a page off the txn's spill list, if present.
2137 * If a page being referenced was spilled to disk in this txn, bring
2138 * it back and make it dirty/writable again.
2139 * @param[in] txn the transaction handle.
2140 * @param[in] mp the page being referenced. It must not be dirty.
2141 * @param[out] ret the writable page, if any. ret is unchanged if
2142 * mp wasn't spilled.
2145 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2147 MDB_env *env = txn->mt_env;
2150 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2152 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2153 if (!tx2->mt_spill_pgs)
2155 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2156 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2159 if (txn->mt_dirty_room == 0)
2160 return MDB_TXN_FULL;
2161 if (IS_OVERFLOW(mp))
2165 if (env->me_flags & MDB_WRITEMAP) {
2168 np = mdb_page_malloc(txn, num);
2172 memcpy(np, mp, num * env->me_psize);
2174 mdb_page_copy(np, mp, env->me_psize);
2177 /* If in current txn, this page is no longer spilled.
2178 * If it happens to be the last page, truncate the spill list.
2179 * Otherwise mark it as deleted by setting the LSB.
2181 if (x == txn->mt_spill_pgs[0])
2182 txn->mt_spill_pgs[0]--;
2184 txn->mt_spill_pgs[x] |= 1;
2185 } /* otherwise, if belonging to a parent txn, the
2186 * page remains spilled until child commits
2189 mdb_page_dirty(txn, np);
2190 np->mp_flags |= P_DIRTY;
2198 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2199 * @param[in] mc cursor pointing to the page to be touched
2200 * @return 0 on success, non-zero on failure.
2203 mdb_page_touch(MDB_cursor *mc)
2205 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2206 MDB_txn *txn = mc->mc_txn;
2207 MDB_cursor *m2, *m3;
2211 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2212 if (txn->mt_flags & MDB_TXN_SPILLS) {
2214 rc = mdb_page_unspill(txn, mp, &np);
2220 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2221 (rc = mdb_page_alloc(mc, 1, &np)))
2224 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2225 mp->mp_pgno, pgno));
2226 mdb_cassert(mc, mp->mp_pgno != pgno);
2227 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2228 /* Update the parent page, if any, to point to the new page */
2230 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2231 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2232 SETPGNO(node, pgno);
2234 mc->mc_db->md_root = pgno;
2236 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2237 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2239 /* If txn has a parent, make sure the page is in our
2243 unsigned x = mdb_mid2l_search(dl, pgno);
2244 if (x <= dl[0].mid && dl[x].mid == pgno) {
2245 if (mp != dl[x].mptr) { /* bad cursor? */
2246 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2247 txn->mt_flags |= MDB_TXN_ERROR;
2248 return MDB_CORRUPTED;
2253 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2255 np = mdb_page_malloc(txn, 1);
2260 rc = mdb_mid2l_insert(dl, &mid);
2261 mdb_cassert(mc, rc == 0);
2266 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2268 np->mp_flags |= P_DIRTY;
2271 /* Adjust cursors pointing to mp */
2272 mc->mc_pg[mc->mc_top] = np;
2273 m2 = txn->mt_cursors[mc->mc_dbi];
2274 if (mc->mc_flags & C_SUB) {
2275 for (; m2; m2=m2->mc_next) {
2276 m3 = &m2->mc_xcursor->mx_cursor;
2277 if (m3->mc_snum < mc->mc_snum) continue;
2278 if (m3->mc_pg[mc->mc_top] == mp)
2279 m3->mc_pg[mc->mc_top] = np;
2282 for (; m2; m2=m2->mc_next) {
2283 if (m2->mc_snum < mc->mc_snum) continue;
2284 if (m2->mc_pg[mc->mc_top] == mp) {
2285 m2->mc_pg[mc->mc_top] = np;
2286 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2288 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2290 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2291 if (!(leaf->mn_flags & F_SUBDATA))
2292 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2300 txn->mt_flags |= MDB_TXN_ERROR;
2305 mdb_env_sync(MDB_env *env, int force)
2308 if (env->me_flags & MDB_RDONLY)
2310 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2311 if (env->me_flags & MDB_WRITEMAP) {
2312 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2313 ? MS_ASYNC : MS_SYNC;
2314 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2317 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2322 if (env->me_fsynconly) {
2323 if (fsync(env->me_fd))
2327 if (MDB_FDATASYNC(env->me_fd))
2334 /** Back up parent txn's cursors, then grab the originals for tracking */
2336 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2338 MDB_cursor *mc, *bk;
2343 for (i = src->mt_numdbs; --i >= 0; ) {
2344 if ((mc = src->mt_cursors[i]) != NULL) {
2345 size = sizeof(MDB_cursor);
2347 size += sizeof(MDB_xcursor);
2348 for (; mc; mc = bk->mc_next) {
2354 mc->mc_db = &dst->mt_dbs[i];
2355 /* Kill pointers into src - and dst to reduce abuse: The
2356 * user may not use mc until dst ends. Otherwise we'd...
2358 mc->mc_txn = NULL; /* ...set this to dst */
2359 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2360 if ((mx = mc->mc_xcursor) != NULL) {
2361 *(MDB_xcursor *)(bk+1) = *mx;
2362 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2364 mc->mc_next = dst->mt_cursors[i];
2365 dst->mt_cursors[i] = mc;
2372 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2373 * @param[in] txn the transaction handle.
2374 * @param[in] merge true to keep changes to parent cursors, false to revert.
2375 * @return 0 on success, non-zero on failure.
2378 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2380 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2384 for (i = txn->mt_numdbs; --i >= 0; ) {
2385 for (mc = cursors[i]; mc; mc = next) {
2387 if ((bk = mc->mc_backup) != NULL) {
2389 /* Commit changes to parent txn */
2390 mc->mc_next = bk->mc_next;
2391 mc->mc_backup = bk->mc_backup;
2392 mc->mc_txn = bk->mc_txn;
2393 mc->mc_db = bk->mc_db;
2394 mc->mc_dbflag = bk->mc_dbflag;
2395 if ((mx = mc->mc_xcursor) != NULL)
2396 mx->mx_cursor.mc_txn = bk->mc_txn;
2398 /* Abort nested txn */
2400 if ((mx = mc->mc_xcursor) != NULL)
2401 *mx = *(MDB_xcursor *)(bk+1);
2405 /* Only malloced cursors are permanently tracked. */
2413 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2416 mdb_txn_reset0(MDB_txn *txn, const char *act);
2418 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2424 Pidset = F_SETLK, Pidcheck = F_GETLK
2428 /** Set or check a pid lock. Set returns 0 on success.
2429 * Check returns 0 if the process is certainly dead, nonzero if it may
2430 * be alive (the lock exists or an error happened so we do not know).
2432 * On Windows Pidset is a no-op, we merely check for the existence
2433 * of the process with the given pid. On POSIX we use a single byte
2434 * lock on the lockfile, set at an offset equal to the pid.
2437 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2439 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2442 if (op == Pidcheck) {
2443 h = OpenProcess(env->me_pidquery, FALSE, pid);
2444 /* No documented "no such process" code, but other program use this: */
2446 return ErrCode() != ERROR_INVALID_PARAMETER;
2447 /* A process exists until all handles to it close. Has it exited? */
2448 ret = WaitForSingleObject(h, 0) != 0;
2455 struct flock lock_info;
2456 memset(&lock_info, 0, sizeof(lock_info));
2457 lock_info.l_type = F_WRLCK;
2458 lock_info.l_whence = SEEK_SET;
2459 lock_info.l_start = pid;
2460 lock_info.l_len = 1;
2461 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2462 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2464 } else if ((rc = ErrCode()) == EINTR) {
2472 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2473 * @param[in] txn the transaction handle to initialize
2474 * @return 0 on success, non-zero on failure.
2477 mdb_txn_renew0(MDB_txn *txn)
2479 MDB_env *env = txn->mt_env;
2480 MDB_txninfo *ti = env->me_txns;
2484 int rc, new_notls = 0;
2486 if (txn->mt_flags & MDB_TXN_RDONLY) {
2488 txn->mt_numdbs = env->me_numdbs;
2489 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2491 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2492 txn->mt_txnid = meta->mm_txnid;
2493 txn->mt_u.reader = NULL;
2495 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2496 pthread_getspecific(env->me_txkey);
2498 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2499 return MDB_BAD_RSLOT;
2501 MDB_PID_T pid = env->me_pid;
2502 MDB_THR_T tid = pthread_self();
2504 if (!env->me_live_reader) {
2505 rc = mdb_reader_pid(env, Pidset, pid);
2508 env->me_live_reader = 1;
2512 nr = ti->mti_numreaders;
2513 for (i=0; i<nr; i++)
2514 if (ti->mti_readers[i].mr_pid == 0)
2516 if (i == env->me_maxreaders) {
2517 UNLOCK_MUTEX_R(env);
2518 return MDB_READERS_FULL;
2520 ti->mti_readers[i].mr_pid = pid;
2521 ti->mti_readers[i].mr_tid = tid;
2523 ti->mti_numreaders = ++nr;
2524 /* Save numreaders for un-mutexed mdb_env_close() */
2525 env->me_numreaders = nr;
2526 UNLOCK_MUTEX_R(env);
2528 r = &ti->mti_readers[i];
2529 new_notls = (env->me_flags & MDB_NOTLS);
2530 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2535 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2536 txn->mt_u.reader = r;
2537 meta = env->me_metas[txn->mt_txnid & 1];
2543 txn->mt_txnid = ti->mti_txnid;
2544 meta = env->me_metas[txn->mt_txnid & 1];
2546 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2547 txn->mt_txnid = meta->mm_txnid;
2550 txn->mt_numdbs = env->me_numdbs;
2553 if (txn->mt_txnid == mdb_debug_start)
2557 txn->mt_child = NULL;
2558 txn->mt_loose_pgs = NULL;
2559 txn->mt_loose_count = 0;
2560 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2561 txn->mt_u.dirty_list = env->me_dirty_list;
2562 txn->mt_u.dirty_list[0].mid = 0;
2563 txn->mt_free_pgs = env->me_free_pgs;
2564 txn->mt_free_pgs[0] = 0;
2565 txn->mt_spill_pgs = NULL;
2567 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2570 /* Copy the DB info and flags */
2571 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2573 /* Moved to here to avoid a data race in read TXNs */
2574 txn->mt_next_pgno = meta->mm_last_pg+1;
2576 for (i=2; i<txn->mt_numdbs; i++) {
2577 x = env->me_dbflags[i];
2578 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2579 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2581 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2583 if (env->me_maxpg < txn->mt_next_pgno) {
2584 mdb_txn_reset0(txn, "renew0-mapfail");
2586 txn->mt_u.reader->mr_pid = 0;
2587 txn->mt_u.reader = NULL;
2589 return MDB_MAP_RESIZED;
2596 mdb_txn_renew(MDB_txn *txn)
2600 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2603 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2604 DPUTS("environment had fatal error, must shutdown!");
2608 rc = mdb_txn_renew0(txn);
2609 if (rc == MDB_SUCCESS) {
2610 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2611 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2612 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2618 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2622 int rc, size, tsize = sizeof(MDB_txn);
2624 if (env->me_flags & MDB_FATAL_ERROR) {
2625 DPUTS("environment had fatal error, must shutdown!");
2628 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2631 /* Nested transactions: Max 1 child, write txns only, no writemap */
2632 if (parent->mt_child ||
2633 (flags & MDB_RDONLY) ||
2634 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2635 (env->me_flags & MDB_WRITEMAP))
2637 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2639 tsize = sizeof(MDB_ntxn);
2642 if (!(flags & MDB_RDONLY)) {
2644 txn = env->me_txn0; /* just reuse preallocated write txn */
2647 /* child txns use own copy of cursors */
2648 size += env->me_maxdbs * sizeof(MDB_cursor *);
2650 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2652 if ((txn = calloc(1, size)) == NULL) {
2653 DPRINTF(("calloc: %s", strerror(errno)));
2656 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2657 if (flags & MDB_RDONLY) {
2658 txn->mt_flags |= MDB_TXN_RDONLY;
2659 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2660 txn->mt_dbiseqs = env->me_dbiseqs;
2662 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2664 txn->mt_dbiseqs = parent->mt_dbiseqs;
2665 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2667 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2668 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2676 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2677 if (!txn->mt_u.dirty_list ||
2678 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2680 free(txn->mt_u.dirty_list);
2684 txn->mt_txnid = parent->mt_txnid;
2685 txn->mt_dirty_room = parent->mt_dirty_room;
2686 txn->mt_u.dirty_list[0].mid = 0;
2687 txn->mt_spill_pgs = NULL;
2688 txn->mt_next_pgno = parent->mt_next_pgno;
2689 parent->mt_child = txn;
2690 txn->mt_parent = parent;
2691 txn->mt_numdbs = parent->mt_numdbs;
2692 txn->mt_flags = parent->mt_flags;
2693 txn->mt_dbxs = parent->mt_dbxs;
2694 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2695 /* Copy parent's mt_dbflags, but clear DB_NEW */
2696 for (i=0; i<txn->mt_numdbs; i++)
2697 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2699 ntxn = (MDB_ntxn *)txn;
2700 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2701 if (env->me_pghead) {
2702 size = MDB_IDL_SIZEOF(env->me_pghead);
2703 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2705 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2710 rc = mdb_cursor_shadow(parent, txn);
2712 mdb_txn_reset0(txn, "beginchild-fail");
2714 rc = mdb_txn_renew0(txn);
2717 if (txn != env->me_txn0)
2721 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2722 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2723 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2730 mdb_txn_env(MDB_txn *txn)
2732 if(!txn) return NULL;
2736 /** Export or close DBI handles opened in this txn. */
2738 mdb_dbis_update(MDB_txn *txn, int keep)
2741 MDB_dbi n = txn->mt_numdbs;
2742 MDB_env *env = txn->mt_env;
2743 unsigned char *tdbflags = txn->mt_dbflags;
2745 for (i = n; --i >= 2;) {
2746 if (tdbflags[i] & DB_NEW) {
2748 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2750 char *ptr = env->me_dbxs[i].md_name.mv_data;
2752 env->me_dbxs[i].md_name.mv_data = NULL;
2753 env->me_dbxs[i].md_name.mv_size = 0;
2754 env->me_dbflags[i] = 0;
2755 env->me_dbiseqs[i]++;
2761 if (keep && env->me_numdbs < n)
2765 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2766 * May be called twice for readonly txns: First reset it, then abort.
2767 * @param[in] txn the transaction handle to reset
2768 * @param[in] act why the transaction is being reset
2771 mdb_txn_reset0(MDB_txn *txn, const char *act)
2773 MDB_env *env = txn->mt_env;
2775 /* Close any DBI handles opened in this txn */
2776 mdb_dbis_update(txn, 0);
2778 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2779 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2780 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2782 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2783 if (txn->mt_u.reader) {
2784 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2785 if (!(env->me_flags & MDB_NOTLS))
2786 txn->mt_u.reader = NULL; /* txn does not own reader */
2788 txn->mt_numdbs = 0; /* close nothing if called again */
2789 txn->mt_dbxs = NULL; /* mark txn as reset */
2791 pgno_t *pghead = env->me_pghead;
2793 mdb_cursors_close(txn, 0);
2794 if (!(env->me_flags & MDB_WRITEMAP)) {
2795 mdb_dlist_free(txn);
2798 if (!txn->mt_parent) {
2799 if (mdb_midl_shrink(&txn->mt_free_pgs))
2800 env->me_free_pgs = txn->mt_free_pgs;
2802 env->me_pghead = NULL;
2806 /* The writer mutex was locked in mdb_txn_begin. */
2808 UNLOCK_MUTEX_W(env);
2810 txn->mt_parent->mt_child = NULL;
2811 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2812 mdb_midl_free(txn->mt_free_pgs);
2813 mdb_midl_free(txn->mt_spill_pgs);
2814 free(txn->mt_u.dirty_list);
2817 mdb_midl_free(pghead);
2822 mdb_txn_reset(MDB_txn *txn)
2827 /* This call is only valid for read-only txns */
2828 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2831 mdb_txn_reset0(txn, "reset");
2835 mdb_txn_abort(MDB_txn *txn)
2841 mdb_txn_abort(txn->mt_child);
2843 mdb_txn_reset0(txn, "abort");
2844 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2845 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2846 txn->mt_u.reader->mr_pid = 0;
2848 if (txn != txn->mt_env->me_txn0)
2852 /** Save the freelist as of this transaction to the freeDB.
2853 * This changes the freelist. Keep trying until it stabilizes.
2856 mdb_freelist_save(MDB_txn *txn)
2858 /* env->me_pghead[] can grow and shrink during this call.
2859 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2860 * Page numbers cannot disappear from txn->mt_free_pgs[].
2863 MDB_env *env = txn->mt_env;
2864 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2865 txnid_t pglast = 0, head_id = 0;
2866 pgno_t freecnt = 0, *free_pgs, *mop;
2867 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2869 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2871 if (env->me_pghead) {
2872 /* Make sure first page of freeDB is touched and on freelist */
2873 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2874 if (rc && rc != MDB_NOTFOUND)
2878 if (!env->me_pghead && txn->mt_loose_pgs) {
2879 /* Put loose page numbers in mt_free_pgs, since
2880 * we may be unable to return them to me_pghead.
2882 MDB_page *mp = txn->mt_loose_pgs;
2883 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2885 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2886 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2887 txn->mt_loose_pgs = NULL;
2888 txn->mt_loose_count = 0;
2891 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2892 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2893 ? SSIZE_MAX : maxfree_1pg;
2896 /* Come back here after each Put() in case freelist changed */
2901 /* If using records from freeDB which we have not yet
2902 * deleted, delete them and any we reserved for me_pghead.
2904 while (pglast < env->me_pglast) {
2905 rc = mdb_cursor_first(&mc, &key, NULL);
2908 pglast = head_id = *(txnid_t *)key.mv_data;
2909 total_room = head_room = 0;
2910 mdb_tassert(txn, pglast <= env->me_pglast);
2911 rc = mdb_cursor_del(&mc, 0);
2916 /* Save the IDL of pages freed by this txn, to a single record */
2917 if (freecnt < txn->mt_free_pgs[0]) {
2919 /* Make sure last page of freeDB is touched and on freelist */
2920 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2921 if (rc && rc != MDB_NOTFOUND)
2924 free_pgs = txn->mt_free_pgs;
2925 /* Write to last page of freeDB */
2926 key.mv_size = sizeof(txn->mt_txnid);
2927 key.mv_data = &txn->mt_txnid;
2929 freecnt = free_pgs[0];
2930 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2931 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2934 /* Retry if mt_free_pgs[] grew during the Put() */
2935 free_pgs = txn->mt_free_pgs;
2936 } while (freecnt < free_pgs[0]);
2937 mdb_midl_sort(free_pgs);
2938 memcpy(data.mv_data, free_pgs, data.mv_size);
2941 unsigned int i = free_pgs[0];
2942 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2943 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2945 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2951 mop = env->me_pghead;
2952 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2954 /* Reserve records for me_pghead[]. Split it if multi-page,
2955 * to avoid searching freeDB for a page range. Use keys in
2956 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2958 if (total_room >= mop_len) {
2959 if (total_room == mop_len || --more < 0)
2961 } else if (head_room >= maxfree_1pg && head_id > 1) {
2962 /* Keep current record (overflow page), add a new one */
2966 /* (Re)write {key = head_id, IDL length = head_room} */
2967 total_room -= head_room;
2968 head_room = mop_len - total_room;
2969 if (head_room > maxfree_1pg && head_id > 1) {
2970 /* Overflow multi-page for part of me_pghead */
2971 head_room /= head_id; /* amortize page sizes */
2972 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2973 } else if (head_room < 0) {
2974 /* Rare case, not bothering to delete this record */
2977 key.mv_size = sizeof(head_id);
2978 key.mv_data = &head_id;
2979 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2980 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2983 /* IDL is initially empty, zero out at least the length */
2984 pgs = (pgno_t *)data.mv_data;
2985 j = head_room > clean_limit ? head_room : 0;
2989 total_room += head_room;
2992 /* Return loose page numbers to me_pghead, though usually none are
2993 * left at this point. The pages themselves remain in dirty_list.
2995 if (txn->mt_loose_pgs) {
2996 MDB_page *mp = txn->mt_loose_pgs;
2997 unsigned count = txn->mt_loose_count;
2999 /* Room for loose pages + temp IDL with same */
3000 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3002 mop = env->me_pghead;
3003 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3004 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3005 loose[ ++count ] = mp->mp_pgno;
3007 mdb_midl_sort(loose);
3008 mdb_midl_xmerge(mop, loose);
3009 txn->mt_loose_pgs = NULL;
3010 txn->mt_loose_count = 0;
3014 /* Fill in the reserved me_pghead records */
3020 rc = mdb_cursor_first(&mc, &key, &data);
3021 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3022 txnid_t id = *(txnid_t *)key.mv_data;
3023 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3026 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3028 if (len > mop_len) {
3030 data.mv_size = (len + 1) * sizeof(MDB_ID);
3032 data.mv_data = mop -= len;
3035 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3037 if (rc || !(mop_len -= len))
3044 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3045 * @param[in] txn the transaction that's being committed
3046 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3047 * @return 0 on success, non-zero on failure.
3050 mdb_page_flush(MDB_txn *txn, int keep)
3052 MDB_env *env = txn->mt_env;
3053 MDB_ID2L dl = txn->mt_u.dirty_list;
3054 unsigned psize = env->me_psize, j;
3055 int i, pagecount = dl[0].mid, rc;
3056 size_t size = 0, pos = 0;
3058 MDB_page *dp = NULL;
3062 struct iovec iov[MDB_COMMIT_PAGES];
3063 ssize_t wpos = 0, wsize = 0, wres;
3064 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3070 if (env->me_flags & MDB_WRITEMAP) {
3071 /* Clear dirty flags */
3072 while (++i <= pagecount) {
3074 /* Don't flush this page yet */
3075 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3076 dp->mp_flags &= ~P_KEEP;
3080 dp->mp_flags &= ~P_DIRTY;
3085 /* Write the pages */
3087 if (++i <= pagecount) {
3089 /* Don't flush this page yet */
3090 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3091 dp->mp_flags &= ~P_KEEP;
3096 /* clear dirty flag */
3097 dp->mp_flags &= ~P_DIRTY;
3100 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3105 /* Windows actually supports scatter/gather I/O, but only on
3106 * unbuffered file handles. Since we're relying on the OS page
3107 * cache for all our data, that's self-defeating. So we just
3108 * write pages one at a time. We use the ov structure to set
3109 * the write offset, to at least save the overhead of a Seek
3112 DPRINTF(("committing page %"Z"u", pgno));
3113 memset(&ov, 0, sizeof(ov));
3114 ov.Offset = pos & 0xffffffff;
3115 ov.OffsetHigh = pos >> 16 >> 16;
3116 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3118 DPRINTF(("WriteFile: %d", rc));
3122 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3123 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3125 /* Write previous page(s) */
3126 #ifdef MDB_USE_PWRITEV
3127 wres = pwritev(env->me_fd, iov, n, wpos);
3130 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3132 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3134 DPRINTF(("lseek: %s", strerror(rc)));
3137 wres = writev(env->me_fd, iov, n);
3140 if (wres != wsize) {
3143 DPRINTF(("Write error: %s", strerror(rc)));
3145 rc = EIO; /* TODO: Use which error code? */
3146 DPUTS("short write, filesystem full?");
3157 DPRINTF(("committing page %"Z"u", pgno));
3158 next_pos = pos + size;
3159 iov[n].iov_len = size;
3160 iov[n].iov_base = (char *)dp;
3166 /* MIPS has cache coherency issues, this is a no-op everywhere else
3167 * Note: for any size >= on-chip cache size, entire on-chip cache is
3170 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3172 for (i = keep; ++i <= pagecount; ) {
3174 /* This is a page we skipped above */
3177 dl[j].mid = dp->mp_pgno;
3180 mdb_dpage_free(env, dp);
3185 txn->mt_dirty_room += i - j;
3191 mdb_txn_commit(MDB_txn *txn)
3197 if (txn == NULL || txn->mt_env == NULL)
3200 if (txn->mt_child) {
3201 rc = mdb_txn_commit(txn->mt_child);
3202 txn->mt_child = NULL;
3209 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3210 mdb_dbis_update(txn, 1);
3211 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3216 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3217 DPUTS("error flag is set, can't commit");
3219 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3224 if (txn->mt_parent) {
3225 MDB_txn *parent = txn->mt_parent;
3229 unsigned x, y, len, ps_len;
3231 /* Append our free list to parent's */
3232 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3235 mdb_midl_free(txn->mt_free_pgs);
3236 /* Failures after this must either undo the changes
3237 * to the parent or set MDB_TXN_ERROR in the parent.
3240 parent->mt_next_pgno = txn->mt_next_pgno;
3241 parent->mt_flags = txn->mt_flags;
3243 /* Merge our cursors into parent's and close them */
3244 mdb_cursors_close(txn, 1);
3246 /* Update parent's DB table. */
3247 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3248 parent->mt_numdbs = txn->mt_numdbs;
3249 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3250 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3251 for (i=2; i<txn->mt_numdbs; i++) {
3252 /* preserve parent's DB_NEW status */
3253 x = parent->mt_dbflags[i] & DB_NEW;
3254 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3257 dst = parent->mt_u.dirty_list;
3258 src = txn->mt_u.dirty_list;
3259 /* Remove anything in our dirty list from parent's spill list */
3260 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3262 pspill[0] = (pgno_t)-1;
3263 /* Mark our dirty pages as deleted in parent spill list */
3264 for (i=0, len=src[0].mid; ++i <= len; ) {
3265 MDB_ID pn = src[i].mid << 1;
3266 while (pn > pspill[x])
3268 if (pn == pspill[x]) {
3273 /* Squash deleted pagenums if we deleted any */
3274 for (x=y; ++x <= ps_len; )
3275 if (!(pspill[x] & 1))
3276 pspill[++y] = pspill[x];
3280 /* Find len = length of merging our dirty list with parent's */
3282 dst[0].mid = 0; /* simplify loops */
3283 if (parent->mt_parent) {
3284 len = x + src[0].mid;
3285 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3286 for (i = x; y && i; y--) {
3287 pgno_t yp = src[y].mid;
3288 while (yp < dst[i].mid)
3290 if (yp == dst[i].mid) {
3295 } else { /* Simplify the above for single-ancestor case */
3296 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3298 /* Merge our dirty list with parent's */
3300 for (i = len; y; dst[i--] = src[y--]) {
3301 pgno_t yp = src[y].mid;
3302 while (yp < dst[x].mid)
3303 dst[i--] = dst[x--];
3304 if (yp == dst[x].mid)
3305 free(dst[x--].mptr);
3307 mdb_tassert(txn, i == x);
3309 free(txn->mt_u.dirty_list);
3310 parent->mt_dirty_room = txn->mt_dirty_room;
3311 if (txn->mt_spill_pgs) {
3312 if (parent->mt_spill_pgs) {
3313 /* TODO: Prevent failure here, so parent does not fail */
3314 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3316 parent->mt_flags |= MDB_TXN_ERROR;
3317 mdb_midl_free(txn->mt_spill_pgs);
3318 mdb_midl_sort(parent->mt_spill_pgs);
3320 parent->mt_spill_pgs = txn->mt_spill_pgs;
3324 /* Append our loose page list to parent's */
3325 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3327 *lp = txn->mt_loose_pgs;
3328 parent->mt_loose_count += txn->mt_loose_count;
3330 parent->mt_child = NULL;
3331 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3336 if (txn != env->me_txn) {
3337 DPUTS("attempt to commit unknown transaction");
3342 mdb_cursors_close(txn, 0);
3344 if (!txn->mt_u.dirty_list[0].mid &&
3345 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3348 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3349 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3351 /* Update DB root pointers */
3352 if (txn->mt_numdbs > 2) {
3356 data.mv_size = sizeof(MDB_db);
3358 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3359 for (i = 2; i < txn->mt_numdbs; i++) {
3360 if (txn->mt_dbflags[i] & DB_DIRTY) {
3361 if (TXN_DBI_CHANGED(txn, i)) {
3365 data.mv_data = &txn->mt_dbs[i];
3366 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3373 rc = mdb_freelist_save(txn);
3377 mdb_midl_free(env->me_pghead);
3378 env->me_pghead = NULL;
3379 if (mdb_midl_shrink(&txn->mt_free_pgs))
3380 env->me_free_pgs = txn->mt_free_pgs;
3386 if ((rc = mdb_page_flush(txn, 0)) ||
3387 (rc = mdb_env_sync(env, 0)) ||
3388 (rc = mdb_env_write_meta(txn)))
3391 /* Free P_LOOSE pages left behind in dirty_list */
3392 if (!(env->me_flags & MDB_WRITEMAP))
3393 mdb_dlist_free(txn);
3398 mdb_dbis_update(txn, 1);
3401 UNLOCK_MUTEX_W(env);
3402 if (txn != env->me_txn0)
3412 /** Read the environment parameters of a DB environment before
3413 * mapping it into memory.
3414 * @param[in] env the environment handle
3415 * @param[out] meta address of where to store the meta information
3416 * @return 0 on success, non-zero on failure.
3419 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3425 enum { Size = sizeof(pbuf) };
3427 /* We don't know the page size yet, so use a minimum value.
3428 * Read both meta pages so we can use the latest one.
3431 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3435 memset(&ov, 0, sizeof(ov));
3437 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3438 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3441 rc = pread(env->me_fd, &pbuf, Size, off);
3444 if (rc == 0 && off == 0)
3446 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3447 DPRINTF(("read: %s", mdb_strerror(rc)));
3451 p = (MDB_page *)&pbuf;
3453 if (!F_ISSET(p->mp_flags, P_META)) {
3454 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3459 if (m->mm_magic != MDB_MAGIC) {
3460 DPUTS("meta has invalid magic");
3464 if (m->mm_version != MDB_DATA_VERSION) {
3465 DPRINTF(("database is version %u, expected version %u",
3466 m->mm_version, MDB_DATA_VERSION));
3467 return MDB_VERSION_MISMATCH;
3470 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3477 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3479 meta->mm_magic = MDB_MAGIC;
3480 meta->mm_version = MDB_DATA_VERSION;
3481 meta->mm_mapsize = env->me_mapsize;
3482 meta->mm_psize = env->me_psize;
3483 meta->mm_last_pg = 1;
3484 meta->mm_flags = env->me_flags & 0xffff;
3485 meta->mm_flags |= MDB_INTEGERKEY;
3486 meta->mm_dbs[0].md_root = P_INVALID;
3487 meta->mm_dbs[1].md_root = P_INVALID;
3490 /** Write the environment parameters of a freshly created DB environment.
3491 * @param[in] env the environment handle
3492 * @param[out] meta address of where to store the meta information
3493 * @return 0 on success, non-zero on failure.
3496 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3504 memset(&ov, 0, sizeof(ov));
3505 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3507 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3510 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3511 len = pwrite(fd, ptr, size, pos); \
3512 rc = (len >= 0); } while(0)
3515 DPUTS("writing new meta page");
3517 psize = env->me_psize;
3519 mdb_env_init_meta0(env, meta);
3521 p = calloc(2, psize);
3523 p->mp_flags = P_META;
3524 *(MDB_meta *)METADATA(p) = *meta;
3526 q = (MDB_page *)((char *)p + psize);
3528 q->mp_flags = P_META;
3529 *(MDB_meta *)METADATA(q) = *meta;
3531 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3534 else if ((unsigned) len == psize * 2)
3542 /** Update the environment info to commit a transaction.
3543 * @param[in] txn the transaction that's being committed
3544 * @return 0 on success, non-zero on failure.
3547 mdb_env_write_meta(MDB_txn *txn)
3550 MDB_meta meta, metab, *mp;
3553 int rc, len, toggle;
3562 toggle = txn->mt_txnid & 1;
3563 DPRINTF(("writing meta page %d for root page %"Z"u",
3564 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3567 mp = env->me_metas[toggle];
3568 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3569 /* Persist any increases of mapsize config */
3570 if (mapsize < env->me_mapsize)
3571 mapsize = env->me_mapsize;
3573 if (env->me_flags & MDB_WRITEMAP) {
3574 mp->mm_mapsize = mapsize;
3575 mp->mm_dbs[0] = txn->mt_dbs[0];
3576 mp->mm_dbs[1] = txn->mt_dbs[1];
3577 mp->mm_last_pg = txn->mt_next_pgno - 1;
3578 mp->mm_txnid = txn->mt_txnid;
3579 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3580 unsigned meta_size = env->me_psize;
3581 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3584 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3585 if (meta_size < env->me_os_psize)
3586 meta_size += meta_size;
3591 if (MDB_MSYNC(ptr, meta_size, rc)) {
3598 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3599 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3601 meta.mm_mapsize = mapsize;
3602 meta.mm_dbs[0] = txn->mt_dbs[0];
3603 meta.mm_dbs[1] = txn->mt_dbs[1];
3604 meta.mm_last_pg = txn->mt_next_pgno - 1;
3605 meta.mm_txnid = txn->mt_txnid;
3607 off = offsetof(MDB_meta, mm_mapsize);
3608 ptr = (char *)&meta + off;
3609 len = sizeof(MDB_meta) - off;
3611 off += env->me_psize;
3614 /* Write to the SYNC fd */
3615 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3616 env->me_fd : env->me_mfd;
3619 memset(&ov, 0, sizeof(ov));
3621 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3625 rc = pwrite(mfd, ptr, len, off);
3628 rc = rc < 0 ? ErrCode() : EIO;
3629 DPUTS("write failed, disk error?");
3630 /* On a failure, the pagecache still contains the new data.
3631 * Write some old data back, to prevent it from being used.
3632 * Use the non-SYNC fd; we know it will fail anyway.
3634 meta.mm_last_pg = metab.mm_last_pg;
3635 meta.mm_txnid = metab.mm_txnid;
3637 memset(&ov, 0, sizeof(ov));
3639 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3641 r2 = pwrite(env->me_fd, ptr, len, off);
3642 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3645 env->me_flags |= MDB_FATAL_ERROR;
3648 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3649 CACHEFLUSH(env->me_map + off, len, DCACHE);
3651 /* Memory ordering issues are irrelevant; since the entire writer
3652 * is wrapped by wmutex, all of these changes will become visible
3653 * after the wmutex is unlocked. Since the DB is multi-version,
3654 * readers will get consistent data regardless of how fresh or
3655 * how stale their view of these values is.
3658 env->me_txns->mti_txnid = txn->mt_txnid;
3663 /** Check both meta pages to see which one is newer.
3664 * @param[in] env the environment handle
3665 * @return meta toggle (0 or 1).
3668 mdb_env_pick_meta(const MDB_env *env)
3670 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3674 mdb_env_create(MDB_env **env)
3678 e = calloc(1, sizeof(MDB_env));
3682 e->me_maxreaders = DEFAULT_READERS;
3683 e->me_maxdbs = e->me_numdbs = 2;
3684 e->me_fd = INVALID_HANDLE_VALUE;
3685 e->me_lfd = INVALID_HANDLE_VALUE;
3686 e->me_mfd = INVALID_HANDLE_VALUE;
3687 #ifdef MDB_USE_POSIX_SEM
3688 e->me_rmutex = SEM_FAILED;
3689 e->me_wmutex = SEM_FAILED;
3691 e->me_pid = getpid();
3692 GET_PAGESIZE(e->me_os_psize);
3693 VGMEMP_CREATE(e,0,0);
3699 mdb_env_map(MDB_env *env, void *addr)
3702 unsigned int flags = env->me_flags;
3706 LONG sizelo, sizehi;
3709 if (flags & MDB_RDONLY) {
3710 /* Don't set explicit map size, use whatever exists */
3715 msize = env->me_mapsize;
3716 sizelo = msize & 0xffffffff;
3717 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3719 /* Windows won't create mappings for zero length files.
3720 * and won't map more than the file size.
3721 * Just set the maxsize right now.
3723 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3724 || !SetEndOfFile(env->me_fd)
3725 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3729 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3730 PAGE_READWRITE : PAGE_READONLY,
3731 sizehi, sizelo, NULL);
3734 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3735 FILE_MAP_WRITE : FILE_MAP_READ,
3737 rc = env->me_map ? 0 : ErrCode();
3742 int prot = PROT_READ;
3743 if (flags & MDB_WRITEMAP) {
3745 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3748 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3750 if (env->me_map == MAP_FAILED) {
3755 if (flags & MDB_NORDAHEAD) {
3756 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3758 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3760 #ifdef POSIX_MADV_RANDOM
3761 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3762 #endif /* POSIX_MADV_RANDOM */
3763 #endif /* MADV_RANDOM */
3767 /* Can happen because the address argument to mmap() is just a
3768 * hint. mmap() can pick another, e.g. if the range is in use.
3769 * The MAP_FIXED flag would prevent that, but then mmap could
3770 * instead unmap existing pages to make room for the new map.
3772 if (addr && env->me_map != addr)
3773 return EBUSY; /* TODO: Make a new MDB_* error code? */
3775 p = (MDB_page *)env->me_map;
3776 env->me_metas[0] = METADATA(p);
3777 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3783 mdb_env_set_mapsize(MDB_env *env, size_t size)
3785 /* If env is already open, caller is responsible for making
3786 * sure there are no active txns.
3794 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3795 else if (size < env->me_mapsize) {
3796 /* If the configured size is smaller, make sure it's
3797 * still big enough. Silently round up to minimum if not.
3799 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3803 munmap(env->me_map, env->me_mapsize);
3804 env->me_mapsize = size;
3805 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3806 rc = mdb_env_map(env, old);
3810 env->me_mapsize = size;
3812 env->me_maxpg = env->me_mapsize / env->me_psize;
3817 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3821 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3826 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3828 if (env->me_map || readers < 1)
3830 env->me_maxreaders = readers;
3835 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3837 if (!env || !readers)
3839 *readers = env->me_maxreaders;
3844 mdb_fsize(HANDLE fd, size_t *size)
3847 LARGE_INTEGER fsize;
3849 if (!GetFileSizeEx(fd, &fsize))
3852 *size = fsize.QuadPart;
3865 #include <sys/utsname.h>
3866 #include <sys/vfs.h>
3869 /** Further setup required for opening an LMDB environment
3872 mdb_env_open2(MDB_env *env)
3874 unsigned int flags = env->me_flags;
3875 int i, newenv = 0, rc;
3879 /* See if we should use QueryLimited */
3881 if ((rc & 0xff) > 5)
3882 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3884 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3887 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
3888 * https://lkml.org/lkml/2012/9/3/83
3889 * Kernels after 3.6-rc6 are known good.
3890 * https://lkml.org/lkml/2012/9/10/556
3891 * See if the DB is on ext3/ext4, then check for new enough kernel
3892 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
3897 fstatfs(env->me_fd, &st);
3898 while (st.f_type == 0xEF53) {
3902 if (uts.release[0] < '3') {
3903 if (!strncmp(uts.release, "2.6.32.", 7)) {
3904 i = atoi(uts.release+7);
3906 break; /* 2.6.32.60 and newer is OK */
3907 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
3908 i = atoi(uts.release+7);
3910 break; /* 2.6.34.15 and newer is OK */
3912 } else if (uts.release[0] == '3') {
3913 i = atoi(uts.release+2);
3915 break; /* 3.6 and newer is OK */
3917 i = atoi(uts.release+4);
3919 break; /* 3.5.4 and newer is OK */
3920 } else if (i == 2) {
3921 i = atoi(uts.release+4);
3923 break; /* 3.2.30 and newer is OK */
3925 } else { /* 4.x and newer is OK */
3928 env->me_fsynconly = 1;
3934 memset(&meta, 0, sizeof(meta));
3936 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3939 DPUTS("new mdbenv");
3941 env->me_psize = env->me_os_psize;
3942 if (env->me_psize > MAX_PAGESIZE)
3943 env->me_psize = MAX_PAGESIZE;
3945 env->me_psize = meta.mm_psize;
3948 /* Was a mapsize configured? */
3949 if (!env->me_mapsize) {
3950 /* If this is a new environment, take the default,
3951 * else use the size recorded in the existing env.
3953 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3954 } else if (env->me_mapsize < meta.mm_mapsize) {
3955 /* If the configured size is smaller, make sure it's
3956 * still big enough. Silently round up to minimum if not.
3958 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3959 if (env->me_mapsize < minsize)
3960 env->me_mapsize = minsize;
3963 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3968 if (flags & MDB_FIXEDMAP)
3969 meta.mm_address = env->me_map;
3970 i = mdb_env_init_meta(env, &meta);
3971 if (i != MDB_SUCCESS) {
3976 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3977 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3979 #if !(MDB_MAXKEYSIZE)
3980 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3982 env->me_maxpg = env->me_mapsize / env->me_psize;
3986 int toggle = mdb_env_pick_meta(env);
3987 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3989 DPRINTF(("opened database version %u, pagesize %u",
3990 env->me_metas[0]->mm_version, env->me_psize));
3991 DPRINTF(("using meta page %d", toggle));
3992 DPRINTF(("depth: %u", db->md_depth));
3993 DPRINTF(("entries: %"Z"u", db->md_entries));
3994 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3995 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3996 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3997 DPRINTF(("root: %"Z"u", db->md_root));
4005 /** Release a reader thread's slot in the reader lock table.
4006 * This function is called automatically when a thread exits.
4007 * @param[in] ptr This points to the slot in the reader lock table.
4010 mdb_env_reader_dest(void *ptr)
4012 MDB_reader *reader = ptr;
4018 /** Junk for arranging thread-specific callbacks on Windows. This is
4019 * necessarily platform and compiler-specific. Windows supports up
4020 * to 1088 keys. Let's assume nobody opens more than 64 environments
4021 * in a single process, for now. They can override this if needed.
4023 #ifndef MAX_TLS_KEYS
4024 #define MAX_TLS_KEYS 64
4026 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4027 static int mdb_tls_nkeys;
4029 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4033 case DLL_PROCESS_ATTACH: break;
4034 case DLL_THREAD_ATTACH: break;
4035 case DLL_THREAD_DETACH:
4036 for (i=0; i<mdb_tls_nkeys; i++) {
4037 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4039 mdb_env_reader_dest(r);
4043 case DLL_PROCESS_DETACH: break;
4048 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4050 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4054 /* Force some symbol references.
4055 * _tls_used forces the linker to create the TLS directory if not already done
4056 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4058 #pragma comment(linker, "/INCLUDE:_tls_used")
4059 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4060 #pragma const_seg(".CRT$XLB")
4061 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4062 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4065 #pragma comment(linker, "/INCLUDE:__tls_used")
4066 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4067 #pragma data_seg(".CRT$XLB")
4068 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4070 #endif /* WIN 32/64 */
4071 #endif /* !__GNUC__ */
4074 /** Downgrade the exclusive lock on the region back to shared */
4076 mdb_env_share_locks(MDB_env *env, int *excl)
4078 int rc = 0, toggle = mdb_env_pick_meta(env);
4080 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4085 /* First acquire a shared lock. The Unlock will
4086 * then release the existing exclusive lock.
4088 memset(&ov, 0, sizeof(ov));
4089 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4092 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4098 struct flock lock_info;
4099 /* The shared lock replaces the existing lock */
4100 memset((void *)&lock_info, 0, sizeof(lock_info));
4101 lock_info.l_type = F_RDLCK;
4102 lock_info.l_whence = SEEK_SET;
4103 lock_info.l_start = 0;
4104 lock_info.l_len = 1;
4105 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4106 (rc = ErrCode()) == EINTR) ;
4107 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4114 /** Try to get exclusive lock, otherwise shared.
4115 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4118 mdb_env_excl_lock(MDB_env *env, int *excl)
4122 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4126 memset(&ov, 0, sizeof(ov));
4127 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4134 struct flock lock_info;
4135 memset((void *)&lock_info, 0, sizeof(lock_info));
4136 lock_info.l_type = F_WRLCK;
4137 lock_info.l_whence = SEEK_SET;
4138 lock_info.l_start = 0;
4139 lock_info.l_len = 1;
4140 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4141 (rc = ErrCode()) == EINTR) ;
4145 # ifdef MDB_USE_POSIX_SEM
4146 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4149 lock_info.l_type = F_RDLCK;
4150 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4151 (rc = ErrCode()) == EINTR) ;
4161 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4163 * @(#) $Revision: 5.1 $
4164 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4165 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4167 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4171 * Please do not copyright this code. This code is in the public domain.
4173 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4174 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4175 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4176 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4177 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4178 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4179 * PERFORMANCE OF THIS SOFTWARE.
4182 * chongo <Landon Curt Noll> /\oo/\
4183 * http://www.isthe.com/chongo/
4185 * Share and Enjoy! :-)
4188 typedef unsigned long long mdb_hash_t;
4189 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4191 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4192 * @param[in] val value to hash
4193 * @param[in] hval initial value for hash
4194 * @return 64 bit hash
4196 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4197 * hval arg on the first call.
4200 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4202 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4203 unsigned char *end = s + val->mv_size;
4205 * FNV-1a hash each octet of the string
4208 /* xor the bottom with the current octet */
4209 hval ^= (mdb_hash_t)*s++;
4211 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4212 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4213 (hval << 7) + (hval << 8) + (hval << 40);
4215 /* return our new hash value */
4219 /** Hash the string and output the encoded hash.
4220 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4221 * very short name limits. We don't care about the encoding being reversible,
4222 * we just want to preserve as many bits of the input as possible in a
4223 * small printable string.
4224 * @param[in] str string to hash
4225 * @param[out] encbuf an array of 11 chars to hold the hash
4227 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4230 mdb_pack85(unsigned long l, char *out)
4234 for (i=0; i<5; i++) {
4235 *out++ = mdb_a85[l % 85];
4241 mdb_hash_enc(MDB_val *val, char *encbuf)
4243 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4245 mdb_pack85(h, encbuf);
4246 mdb_pack85(h>>32, encbuf+5);
4251 /** Open and/or initialize the lock region for the environment.
4252 * @param[in] env The LMDB environment.
4253 * @param[in] lpath The pathname of the file used for the lock region.
4254 * @param[in] mode The Unix permissions for the file, if we create it.
4255 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4256 * @return 0 on success, non-zero on failure.
4259 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4262 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4264 # define MDB_ERRCODE_ROFS EROFS
4265 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4266 # define MDB_CLOEXEC O_CLOEXEC
4269 # define MDB_CLOEXEC 0
4276 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4277 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4278 FILE_ATTRIBUTE_NORMAL, NULL);
4280 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4282 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4284 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4289 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4290 /* Lose record locks when exec*() */
4291 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4292 fcntl(env->me_lfd, F_SETFD, fdflags);
4295 if (!(env->me_flags & MDB_NOTLS)) {
4296 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4299 env->me_flags |= MDB_ENV_TXKEY;
4301 /* Windows TLS callbacks need help finding their TLS info. */
4302 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4306 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4310 /* Try to get exclusive lock. If we succeed, then
4311 * nobody is using the lock region and we should initialize it.
4313 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4316 size = GetFileSize(env->me_lfd, NULL);
4318 size = lseek(env->me_lfd, 0, SEEK_END);
4319 if (size == -1) goto fail_errno;
4321 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4322 if (size < rsize && *excl > 0) {
4324 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4325 || !SetEndOfFile(env->me_lfd))
4328 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4332 size = rsize - sizeof(MDB_txninfo);
4333 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4338 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4340 if (!mh) goto fail_errno;
4341 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4343 if (!env->me_txns) goto fail_errno;
4345 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4347 if (m == MAP_FAILED) goto fail_errno;
4353 BY_HANDLE_FILE_INFORMATION stbuf;
4362 if (!mdb_sec_inited) {
4363 InitializeSecurityDescriptor(&mdb_null_sd,
4364 SECURITY_DESCRIPTOR_REVISION);
4365 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4366 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4367 mdb_all_sa.bInheritHandle = FALSE;
4368 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4371 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4372 idbuf.volume = stbuf.dwVolumeSerialNumber;
4373 idbuf.nhigh = stbuf.nFileIndexHigh;
4374 idbuf.nlow = stbuf.nFileIndexLow;
4375 val.mv_data = &idbuf;
4376 val.mv_size = sizeof(idbuf);
4377 mdb_hash_enc(&val, encbuf);
4378 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4379 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4380 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4381 if (!env->me_rmutex) goto fail_errno;
4382 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4383 if (!env->me_wmutex) goto fail_errno;
4384 #elif defined(MDB_USE_POSIX_SEM)
4393 #if defined(__NetBSD__)
4394 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4396 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4397 idbuf.dev = stbuf.st_dev;
4398 idbuf.ino = stbuf.st_ino;
4399 val.mv_data = &idbuf;
4400 val.mv_size = sizeof(idbuf);
4401 mdb_hash_enc(&val, encbuf);
4402 #ifdef MDB_SHORT_SEMNAMES
4403 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4405 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4406 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4407 /* Clean up after a previous run, if needed: Try to
4408 * remove both semaphores before doing anything else.
4410 sem_unlink(env->me_txns->mti_rmname);
4411 sem_unlink(env->me_txns->mti_wmname);
4412 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4413 O_CREAT|O_EXCL, mode, 1);
4414 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4415 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4416 O_CREAT|O_EXCL, mode, 1);
4417 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4418 #else /* MDB_USE_POSIX_SEM */
4419 pthread_mutexattr_t mattr;
4421 if ((rc = pthread_mutexattr_init(&mattr))
4422 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4423 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4424 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4426 pthread_mutexattr_destroy(&mattr);
4427 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4429 env->me_txns->mti_magic = MDB_MAGIC;
4430 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4431 env->me_txns->mti_txnid = 0;
4432 env->me_txns->mti_numreaders = 0;
4435 if (env->me_txns->mti_magic != MDB_MAGIC) {
4436 DPUTS("lock region has invalid magic");
4440 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4441 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4442 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4443 rc = MDB_VERSION_MISMATCH;
4447 if (rc && rc != EACCES && rc != EAGAIN) {
4451 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4452 if (!env->me_rmutex) goto fail_errno;
4453 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4454 if (!env->me_wmutex) goto fail_errno;
4455 #elif defined(MDB_USE_POSIX_SEM)
4456 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4457 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4458 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4459 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4470 /** The name of the lock file in the DB environment */
4471 #define LOCKNAME "/lock.mdb"
4472 /** The name of the data file in the DB environment */
4473 #define DATANAME "/data.mdb"
4474 /** The suffix of the lock file when no subdir is used */
4475 #define LOCKSUFF "-lock"
4476 /** Only a subset of the @ref mdb_env flags can be changed
4477 * at runtime. Changing other flags requires closing the
4478 * environment and re-opening it with the new flags.
4480 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4481 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4482 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4484 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4485 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4489 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4491 int oflags, rc, len, excl = -1;
4492 char *lpath, *dpath;
4494 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4498 if (flags & MDB_NOSUBDIR) {
4499 rc = len + sizeof(LOCKSUFF) + len + 1;
4501 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4506 if (flags & MDB_NOSUBDIR) {
4507 dpath = lpath + len + sizeof(LOCKSUFF);
4508 sprintf(lpath, "%s" LOCKSUFF, path);
4509 strcpy(dpath, path);
4511 dpath = lpath + len + sizeof(LOCKNAME);
4512 sprintf(lpath, "%s" LOCKNAME, path);
4513 sprintf(dpath, "%s" DATANAME, path);
4517 flags |= env->me_flags;
4518 if (flags & MDB_RDONLY) {
4519 /* silently ignore WRITEMAP when we're only getting read access */
4520 flags &= ~MDB_WRITEMAP;
4522 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4523 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4526 env->me_flags = flags |= MDB_ENV_ACTIVE;
4530 env->me_path = strdup(path);
4531 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4532 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4533 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4534 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4539 /* For RDONLY, get lockfile after we know datafile exists */
4540 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4541 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4547 if (F_ISSET(flags, MDB_RDONLY)) {
4548 oflags = GENERIC_READ;
4549 len = OPEN_EXISTING;
4551 oflags = GENERIC_READ|GENERIC_WRITE;
4554 mode = FILE_ATTRIBUTE_NORMAL;
4555 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4556 NULL, len, mode, NULL);
4558 if (F_ISSET(flags, MDB_RDONLY))
4561 oflags = O_RDWR | O_CREAT;
4563 env->me_fd = open(dpath, oflags, mode);
4565 if (env->me_fd == INVALID_HANDLE_VALUE) {
4570 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4571 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4576 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4577 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4578 env->me_mfd = env->me_fd;
4580 /* Synchronous fd for meta writes. Needed even with
4581 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4584 len = OPEN_EXISTING;
4585 env->me_mfd = CreateFile(dpath, oflags,
4586 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4587 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4590 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4592 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4597 DPRINTF(("opened dbenv %p", (void *) env));
4599 rc = mdb_env_share_locks(env, &excl);
4603 if (!((flags & MDB_RDONLY) ||
4604 (env->me_pbuf = calloc(1, env->me_psize))))
4606 if (!(flags & MDB_RDONLY)) {
4608 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4609 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4610 txn = calloc(1, size);
4612 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4613 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4614 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4615 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4617 txn->mt_dbxs = env->me_dbxs;
4627 mdb_env_close0(env, excl);
4633 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4635 mdb_env_close0(MDB_env *env, int excl)
4639 if (!(env->me_flags & MDB_ENV_ACTIVE))
4642 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4643 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4644 free(env->me_dbxs[i].md_name.mv_data);
4647 free(env->me_dbiseqs);
4648 free(env->me_dbflags);
4651 free(env->me_dirty_list);
4653 mdb_midl_free(env->me_free_pgs);
4655 if (env->me_flags & MDB_ENV_TXKEY) {
4656 pthread_key_delete(env->me_txkey);
4658 /* Delete our key from the global list */
4659 for (i=0; i<mdb_tls_nkeys; i++)
4660 if (mdb_tls_keys[i] == env->me_txkey) {
4661 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4669 munmap(env->me_map, env->me_mapsize);
4671 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4672 (void) close(env->me_mfd);
4673 if (env->me_fd != INVALID_HANDLE_VALUE)
4674 (void) close(env->me_fd);
4676 MDB_PID_T pid = env->me_pid;
4677 /* Clearing readers is done in this function because
4678 * me_txkey with its destructor must be disabled first.
4680 for (i = env->me_numreaders; --i >= 0; )
4681 if (env->me_txns->mti_readers[i].mr_pid == pid)
4682 env->me_txns->mti_readers[i].mr_pid = 0;
4684 if (env->me_rmutex) {
4685 CloseHandle(env->me_rmutex);
4686 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4688 /* Windows automatically destroys the mutexes when
4689 * the last handle closes.
4691 #elif defined(MDB_USE_POSIX_SEM)
4692 if (env->me_rmutex != SEM_FAILED) {
4693 sem_close(env->me_rmutex);
4694 if (env->me_wmutex != SEM_FAILED)
4695 sem_close(env->me_wmutex);
4696 /* If we have the filelock: If we are the
4697 * only remaining user, clean up semaphores.
4700 mdb_env_excl_lock(env, &excl);
4702 sem_unlink(env->me_txns->mti_rmname);
4703 sem_unlink(env->me_txns->mti_wmname);
4707 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4709 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4712 /* Unlock the lockfile. Windows would have unlocked it
4713 * after closing anyway, but not necessarily at once.
4715 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4718 (void) close(env->me_lfd);
4721 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4726 mdb_env_close(MDB_env *env)
4733 VGMEMP_DESTROY(env);
4734 while ((dp = env->me_dpages) != NULL) {
4735 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4736 env->me_dpages = dp->mp_next;
4740 mdb_env_close0(env, 0);
4744 /** Compare two items pointing at aligned size_t's */
4746 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4748 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4749 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4752 /** Compare two items pointing at aligned unsigned int's */
4754 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4756 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4757 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4760 /** Compare two items pointing at unsigned ints of unknown alignment.
4761 * Nodes and keys are guaranteed to be 2-byte aligned.
4764 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4766 #if BYTE_ORDER == LITTLE_ENDIAN
4767 unsigned short *u, *c;
4770 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4771 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4774 } while(!x && u > (unsigned short *)a->mv_data);
4777 unsigned short *u, *c, *end;
4780 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4781 u = (unsigned short *)a->mv_data;
4782 c = (unsigned short *)b->mv_data;
4785 } while(!x && u < end);
4790 /** Compare two items pointing at size_t's of unknown alignment. */
4791 #ifdef MISALIGNED_OK
4792 # define mdb_cmp_clong mdb_cmp_long
4794 # define mdb_cmp_clong mdb_cmp_cint
4797 /** Compare two items lexically */
4799 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4806 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4812 diff = memcmp(a->mv_data, b->mv_data, len);
4813 return diff ? diff : len_diff<0 ? -1 : len_diff;
4816 /** Compare two items in reverse byte order */
4818 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4820 const unsigned char *p1, *p2, *p1_lim;
4824 p1_lim = (const unsigned char *)a->mv_data;
4825 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4826 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4828 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4834 while (p1 > p1_lim) {
4835 diff = *--p1 - *--p2;
4839 return len_diff<0 ? -1 : len_diff;
4842 /** Search for key within a page, using binary search.
4843 * Returns the smallest entry larger or equal to the key.
4844 * If exactp is non-null, stores whether the found entry was an exact match
4845 * in *exactp (1 or 0).
4846 * Updates the cursor index with the index of the found entry.
4847 * If no entry larger or equal to the key is found, returns NULL.
4850 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4852 unsigned int i = 0, nkeys;
4855 MDB_page *mp = mc->mc_pg[mc->mc_top];
4856 MDB_node *node = NULL;
4861 nkeys = NUMKEYS(mp);
4863 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4864 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4867 low = IS_LEAF(mp) ? 0 : 1;
4869 cmp = mc->mc_dbx->md_cmp;
4871 /* Branch pages have no data, so if using integer keys,
4872 * alignment is guaranteed. Use faster mdb_cmp_int.
4874 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4875 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4882 nodekey.mv_size = mc->mc_db->md_pad;
4883 node = NODEPTR(mp, 0); /* fake */
4884 while (low <= high) {
4885 i = (low + high) >> 1;
4886 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4887 rc = cmp(key, &nodekey);
4888 DPRINTF(("found leaf index %u [%s], rc = %i",
4889 i, DKEY(&nodekey), rc));
4898 while (low <= high) {
4899 i = (low + high) >> 1;
4901 node = NODEPTR(mp, i);
4902 nodekey.mv_size = NODEKSZ(node);
4903 nodekey.mv_data = NODEKEY(node);
4905 rc = cmp(key, &nodekey);
4908 DPRINTF(("found leaf index %u [%s], rc = %i",
4909 i, DKEY(&nodekey), rc));
4911 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4912 i, DKEY(&nodekey), NODEPGNO(node), rc));
4923 if (rc > 0) { /* Found entry is less than the key. */
4924 i++; /* Skip to get the smallest entry larger than key. */
4926 node = NODEPTR(mp, i);
4929 *exactp = (rc == 0 && nkeys > 0);
4930 /* store the key index */
4931 mc->mc_ki[mc->mc_top] = i;
4933 /* There is no entry larger or equal to the key. */
4936 /* nodeptr is fake for LEAF2 */
4942 mdb_cursor_adjust(MDB_cursor *mc, func)
4946 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4947 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4954 /** Pop a page off the top of the cursor's stack. */
4956 mdb_cursor_pop(MDB_cursor *mc)
4960 MDB_page *top = mc->mc_pg[mc->mc_top];
4966 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4967 DDBI(mc), (void *) mc));
4971 /** Push a page onto the top of the cursor's stack. */
4973 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4975 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4976 DDBI(mc), (void *) mc));
4978 if (mc->mc_snum >= CURSOR_STACK) {
4979 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4980 return MDB_CURSOR_FULL;
4983 mc->mc_top = mc->mc_snum++;
4984 mc->mc_pg[mc->mc_top] = mp;
4985 mc->mc_ki[mc->mc_top] = 0;
4990 /** Find the address of the page corresponding to a given page number.
4991 * @param[in] txn the transaction for this access.
4992 * @param[in] pgno the page number for the page to retrieve.
4993 * @param[out] ret address of a pointer where the page's address will be stored.
4994 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4995 * @return 0 on success, non-zero on failure.
4998 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5000 MDB_env *env = txn->mt_env;
5004 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
5008 MDB_ID2L dl = tx2->mt_u.dirty_list;
5010 /* Spilled pages were dirtied in this txn and flushed
5011 * because the dirty list got full. Bring this page
5012 * back in from the map (but don't unspill it here,
5013 * leave that unless page_touch happens again).
5015 if (tx2->mt_spill_pgs) {
5016 MDB_ID pn = pgno << 1;
5017 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5018 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5019 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5024 unsigned x = mdb_mid2l_search(dl, pgno);
5025 if (x <= dl[0].mid && dl[x].mid == pgno) {
5031 } while ((tx2 = tx2->mt_parent) != NULL);
5034 if (pgno < txn->mt_next_pgno) {
5036 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5038 DPRINTF(("page %"Z"u not found", pgno));
5039 txn->mt_flags |= MDB_TXN_ERROR;
5040 return MDB_PAGE_NOTFOUND;
5050 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5051 * The cursor is at the root page, set up the rest of it.
5054 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5056 MDB_page *mp = mc->mc_pg[mc->mc_top];
5060 while (IS_BRANCH(mp)) {
5064 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5065 mdb_cassert(mc, NUMKEYS(mp) > 1);
5066 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5068 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5070 if (flags & MDB_PS_LAST)
5071 i = NUMKEYS(mp) - 1;
5074 node = mdb_node_search(mc, key, &exact);
5076 i = NUMKEYS(mp) - 1;
5078 i = mc->mc_ki[mc->mc_top];
5080 mdb_cassert(mc, i > 0);
5084 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5087 mdb_cassert(mc, i < NUMKEYS(mp));
5088 node = NODEPTR(mp, i);
5090 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5093 mc->mc_ki[mc->mc_top] = i;
5094 if ((rc = mdb_cursor_push(mc, mp)))
5097 if (flags & MDB_PS_MODIFY) {
5098 if ((rc = mdb_page_touch(mc)) != 0)
5100 mp = mc->mc_pg[mc->mc_top];
5105 DPRINTF(("internal error, index points to a %02X page!?",
5107 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5108 return MDB_CORRUPTED;
5111 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5112 key ? DKEY(key) : "null"));
5113 mc->mc_flags |= C_INITIALIZED;
5114 mc->mc_flags &= ~C_EOF;
5119 /** Search for the lowest key under the current branch page.
5120 * This just bypasses a NUMKEYS check in the current page
5121 * before calling mdb_page_search_root(), because the callers
5122 * are all in situations where the current page is known to
5126 mdb_page_search_lowest(MDB_cursor *mc)
5128 MDB_page *mp = mc->mc_pg[mc->mc_top];
5129 MDB_node *node = NODEPTR(mp, 0);
5132 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5135 mc->mc_ki[mc->mc_top] = 0;
5136 if ((rc = mdb_cursor_push(mc, mp)))
5138 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5141 /** Search for the page a given key should be in.
5142 * Push it and its parent pages on the cursor stack.
5143 * @param[in,out] mc the cursor for this operation.
5144 * @param[in] key the key to search for, or NULL for first/last page.
5145 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5146 * are touched (updated with new page numbers).
5147 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5148 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5149 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5150 * @return 0 on success, non-zero on failure.
5153 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5158 /* Make sure the txn is still viable, then find the root from
5159 * the txn's db table and set it as the root of the cursor's stack.
5161 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5162 DPUTS("transaction has failed, must abort");
5165 /* Make sure we're using an up-to-date root */
5166 if (*mc->mc_dbflag & DB_STALE) {
5168 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5170 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5171 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5178 MDB_node *leaf = mdb_node_search(&mc2,
5179 &mc->mc_dbx->md_name, &exact);
5181 return MDB_NOTFOUND;
5182 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5185 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5187 /* The txn may not know this DBI, or another process may
5188 * have dropped and recreated the DB with other flags.
5190 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5191 return MDB_INCOMPATIBLE;
5192 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5194 *mc->mc_dbflag &= ~DB_STALE;
5196 root = mc->mc_db->md_root;
5198 if (root == P_INVALID) { /* Tree is empty. */
5199 DPUTS("tree is empty");
5200 return MDB_NOTFOUND;
5204 mdb_cassert(mc, root > 1);
5205 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5206 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5212 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5213 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5215 if (flags & MDB_PS_MODIFY) {
5216 if ((rc = mdb_page_touch(mc)))
5220 if (flags & MDB_PS_ROOTONLY)
5223 return mdb_page_search_root(mc, key, flags);
5227 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5229 MDB_txn *txn = mc->mc_txn;
5230 pgno_t pg = mp->mp_pgno;
5231 unsigned x = 0, ovpages = mp->mp_pages;
5232 MDB_env *env = txn->mt_env;
5233 MDB_IDL sl = txn->mt_spill_pgs;
5234 MDB_ID pn = pg << 1;
5237 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5238 /* If the page is dirty or on the spill list we just acquired it,
5239 * so we should give it back to our current free list, if any.
5240 * Otherwise put it onto the list of pages we freed in this txn.
5242 * Won't create me_pghead: me_pglast must be inited along with it.
5243 * Unsupported in nested txns: They would need to hide the page
5244 * range in ancestor txns' dirty and spilled lists.
5246 if (env->me_pghead &&
5248 ((mp->mp_flags & P_DIRTY) ||
5249 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5253 MDB_ID2 *dl, ix, iy;
5254 rc = mdb_midl_need(&env->me_pghead, ovpages);
5257 if (!(mp->mp_flags & P_DIRTY)) {
5258 /* This page is no longer spilled */
5265 /* Remove from dirty list */
5266 dl = txn->mt_u.dirty_list;
5268 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5274 mdb_cassert(mc, x > 1);
5276 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5277 txn->mt_flags |= MDB_TXN_ERROR;
5278 return MDB_CORRUPTED;
5281 if (!(env->me_flags & MDB_WRITEMAP))
5282 mdb_dpage_free(env, mp);
5284 /* Insert in me_pghead */
5285 mop = env->me_pghead;
5286 j = mop[0] + ovpages;
5287 for (i = mop[0]; i && mop[i] < pg; i--)
5293 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5297 mc->mc_db->md_overflow_pages -= ovpages;
5301 /** Return the data associated with a given node.
5302 * @param[in] txn The transaction for this operation.
5303 * @param[in] leaf The node being read.
5304 * @param[out] data Updated to point to the node's data.
5305 * @return 0 on success, non-zero on failure.
5308 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5310 MDB_page *omp; /* overflow page */
5314 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5315 data->mv_size = NODEDSZ(leaf);
5316 data->mv_data = NODEDATA(leaf);
5320 /* Read overflow data.
5322 data->mv_size = NODEDSZ(leaf);
5323 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5324 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5325 DPRINTF(("read overflow page %"Z"u failed", pgno));
5328 data->mv_data = METADATA(omp);
5334 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5335 MDB_val *key, MDB_val *data)
5342 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5344 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5347 if (txn->mt_flags & MDB_TXN_ERROR)
5350 mdb_cursor_init(&mc, txn, dbi, &mx);
5351 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5354 /** Find a sibling for a page.
5355 * Replaces the page at the top of the cursor's stack with the
5356 * specified sibling, if one exists.
5357 * @param[in] mc The cursor for this operation.
5358 * @param[in] move_right Non-zero if the right sibling is requested,
5359 * otherwise the left sibling.
5360 * @return 0 on success, non-zero on failure.
5363 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5369 if (mc->mc_snum < 2) {
5370 return MDB_NOTFOUND; /* root has no siblings */
5374 DPRINTF(("parent page is page %"Z"u, index %u",
5375 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5377 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5378 : (mc->mc_ki[mc->mc_top] == 0)) {
5379 DPRINTF(("no more keys left, moving to %s sibling",
5380 move_right ? "right" : "left"));
5381 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5382 /* undo cursor_pop before returning */
5389 mc->mc_ki[mc->mc_top]++;
5391 mc->mc_ki[mc->mc_top]--;
5392 DPRINTF(("just moving to %s index key %u",
5393 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5395 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5397 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5398 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5399 /* mc will be inconsistent if caller does mc_snum++ as above */
5400 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5404 mdb_cursor_push(mc, mp);
5406 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5411 /** Move the cursor to the next data item. */
5413 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5419 if (mc->mc_flags & C_EOF) {
5420 return MDB_NOTFOUND;
5423 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5425 mp = mc->mc_pg[mc->mc_top];
5427 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5428 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5429 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5430 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5431 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5432 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5433 if (rc == MDB_SUCCESS)
5434 MDB_GET_KEY(leaf, key);
5439 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5440 if (op == MDB_NEXT_DUP)
5441 return MDB_NOTFOUND;
5445 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5446 mdb_dbg_pgno(mp), (void *) mc));
5447 if (mc->mc_flags & C_DEL)
5450 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5451 DPUTS("=====> move to next sibling page");
5452 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5453 mc->mc_flags |= C_EOF;
5456 mp = mc->mc_pg[mc->mc_top];
5457 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5459 mc->mc_ki[mc->mc_top]++;
5462 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5463 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5466 key->mv_size = mc->mc_db->md_pad;
5467 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5471 mdb_cassert(mc, IS_LEAF(mp));
5472 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5474 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5475 mdb_xcursor_init1(mc, leaf);
5478 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5481 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5482 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5483 if (rc != MDB_SUCCESS)
5488 MDB_GET_KEY(leaf, key);
5492 /** Move the cursor to the previous data item. */
5494 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5500 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5502 mp = mc->mc_pg[mc->mc_top];
5504 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5505 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5506 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5507 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5508 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5509 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5510 if (rc == MDB_SUCCESS) {
5511 MDB_GET_KEY(leaf, key);
5512 mc->mc_flags &= ~C_EOF;
5518 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5519 if (op == MDB_PREV_DUP)
5520 return MDB_NOTFOUND;
5524 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5525 mdb_dbg_pgno(mp), (void *) mc));
5527 if (mc->mc_ki[mc->mc_top] == 0) {
5528 DPUTS("=====> move to prev sibling page");
5529 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5532 mp = mc->mc_pg[mc->mc_top];
5533 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5534 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5536 mc->mc_ki[mc->mc_top]--;
5538 mc->mc_flags &= ~C_EOF;
5540 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5541 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5544 key->mv_size = mc->mc_db->md_pad;
5545 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5549 mdb_cassert(mc, IS_LEAF(mp));
5550 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5552 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5553 mdb_xcursor_init1(mc, leaf);
5556 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5559 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5560 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5561 if (rc != MDB_SUCCESS)
5566 MDB_GET_KEY(leaf, key);
5570 /** Set the cursor on a specific data item. */
5572 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5573 MDB_cursor_op op, int *exactp)
5577 MDB_node *leaf = NULL;
5580 if (key->mv_size == 0)
5581 return MDB_BAD_VALSIZE;
5584 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5586 /* See if we're already on the right page */
5587 if (mc->mc_flags & C_INITIALIZED) {
5590 mp = mc->mc_pg[mc->mc_top];
5592 mc->mc_ki[mc->mc_top] = 0;
5593 return MDB_NOTFOUND;
5595 if (mp->mp_flags & P_LEAF2) {
5596 nodekey.mv_size = mc->mc_db->md_pad;
5597 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5599 leaf = NODEPTR(mp, 0);
5600 MDB_GET_KEY2(leaf, nodekey);
5602 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5604 /* Probably happens rarely, but first node on the page
5605 * was the one we wanted.
5607 mc->mc_ki[mc->mc_top] = 0;
5614 unsigned int nkeys = NUMKEYS(mp);
5616 if (mp->mp_flags & P_LEAF2) {
5617 nodekey.mv_data = LEAF2KEY(mp,
5618 nkeys-1, nodekey.mv_size);
5620 leaf = NODEPTR(mp, nkeys-1);
5621 MDB_GET_KEY2(leaf, nodekey);
5623 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5625 /* last node was the one we wanted */
5626 mc->mc_ki[mc->mc_top] = nkeys-1;
5632 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5633 /* This is definitely the right page, skip search_page */
5634 if (mp->mp_flags & P_LEAF2) {
5635 nodekey.mv_data = LEAF2KEY(mp,
5636 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5638 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5639 MDB_GET_KEY2(leaf, nodekey);
5641 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5643 /* current node was the one we wanted */
5653 /* If any parents have right-sibs, search.
5654 * Otherwise, there's nothing further.
5656 for (i=0; i<mc->mc_top; i++)
5658 NUMKEYS(mc->mc_pg[i])-1)
5660 if (i == mc->mc_top) {
5661 /* There are no other pages */
5662 mc->mc_ki[mc->mc_top] = nkeys;
5663 return MDB_NOTFOUND;
5667 /* There are no other pages */
5668 mc->mc_ki[mc->mc_top] = 0;
5669 if (op == MDB_SET_RANGE && !exactp) {
5673 return MDB_NOTFOUND;
5677 rc = mdb_page_search(mc, key, 0);
5678 if (rc != MDB_SUCCESS)
5681 mp = mc->mc_pg[mc->mc_top];
5682 mdb_cassert(mc, IS_LEAF(mp));
5685 leaf = mdb_node_search(mc, key, exactp);
5686 if (exactp != NULL && !*exactp) {
5687 /* MDB_SET specified and not an exact match. */
5688 return MDB_NOTFOUND;
5692 DPUTS("===> inexact leaf not found, goto sibling");
5693 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5694 return rc; /* no entries matched */
5695 mp = mc->mc_pg[mc->mc_top];
5696 mdb_cassert(mc, IS_LEAF(mp));
5697 leaf = NODEPTR(mp, 0);
5701 mc->mc_flags |= C_INITIALIZED;
5702 mc->mc_flags &= ~C_EOF;
5705 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5706 key->mv_size = mc->mc_db->md_pad;
5707 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5712 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5713 mdb_xcursor_init1(mc, leaf);
5716 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5717 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5718 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5721 if (op == MDB_GET_BOTH) {
5727 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5728 if (rc != MDB_SUCCESS)
5731 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5733 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5735 rc = mc->mc_dbx->md_dcmp(data, &d2);
5737 if (op == MDB_GET_BOTH || rc > 0)
5738 return MDB_NOTFOUND;
5745 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5746 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5751 /* The key already matches in all other cases */
5752 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5753 MDB_GET_KEY(leaf, key);
5754 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5759 /** Move the cursor to the first item in the database. */
5761 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5767 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5769 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5770 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5771 if (rc != MDB_SUCCESS)
5774 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5776 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5777 mc->mc_flags |= C_INITIALIZED;
5778 mc->mc_flags &= ~C_EOF;
5780 mc->mc_ki[mc->mc_top] = 0;
5782 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5783 key->mv_size = mc->mc_db->md_pad;
5784 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5789 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5790 mdb_xcursor_init1(mc, leaf);
5791 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5795 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5799 MDB_GET_KEY(leaf, key);
5803 /** Move the cursor to the last item in the database. */
5805 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5811 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5813 if (!(mc->mc_flags & C_EOF)) {
5815 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5816 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5817 if (rc != MDB_SUCCESS)
5820 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5823 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5824 mc->mc_flags |= C_INITIALIZED|C_EOF;
5825 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5827 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5828 key->mv_size = mc->mc_db->md_pad;
5829 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5834 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5835 mdb_xcursor_init1(mc, leaf);
5836 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5840 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5845 MDB_GET_KEY(leaf, key);
5850 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5855 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5860 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5864 case MDB_GET_CURRENT:
5865 if (!(mc->mc_flags & C_INITIALIZED)) {
5868 MDB_page *mp = mc->mc_pg[mc->mc_top];
5869 int nkeys = NUMKEYS(mp);
5870 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5871 mc->mc_ki[mc->mc_top] = nkeys;
5877 key->mv_size = mc->mc_db->md_pad;
5878 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5880 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5881 MDB_GET_KEY(leaf, key);
5883 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5884 if (mc->mc_flags & C_DEL)
5885 mdb_xcursor_init1(mc, leaf);
5886 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5888 rc = mdb_node_read(mc->mc_txn, leaf, data);
5895 case MDB_GET_BOTH_RANGE:
5900 if (mc->mc_xcursor == NULL) {
5901 rc = MDB_INCOMPATIBLE;
5911 rc = mdb_cursor_set(mc, key, data, op,
5912 op == MDB_SET_RANGE ? NULL : &exact);
5915 case MDB_GET_MULTIPLE:
5916 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5920 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5921 rc = MDB_INCOMPATIBLE;
5925 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5926 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5929 case MDB_NEXT_MULTIPLE:
5934 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5935 rc = MDB_INCOMPATIBLE;
5938 if (!(mc->mc_flags & C_INITIALIZED))
5939 rc = mdb_cursor_first(mc, key, data);
5941 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5942 if (rc == MDB_SUCCESS) {
5943 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5946 mx = &mc->mc_xcursor->mx_cursor;
5947 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5949 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5950 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5958 case MDB_NEXT_NODUP:
5959 if (!(mc->mc_flags & C_INITIALIZED))
5960 rc = mdb_cursor_first(mc, key, data);
5962 rc = mdb_cursor_next(mc, key, data, op);
5966 case MDB_PREV_NODUP:
5967 if (!(mc->mc_flags & C_INITIALIZED)) {
5968 rc = mdb_cursor_last(mc, key, data);
5971 mc->mc_flags |= C_INITIALIZED;
5972 mc->mc_ki[mc->mc_top]++;
5974 rc = mdb_cursor_prev(mc, key, data, op);
5977 rc = mdb_cursor_first(mc, key, data);
5980 mfunc = mdb_cursor_first;
5982 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5986 if (mc->mc_xcursor == NULL) {
5987 rc = MDB_INCOMPATIBLE;
5991 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5992 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5993 MDB_GET_KEY(leaf, key);
5994 rc = mdb_node_read(mc->mc_txn, leaf, data);
5998 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6002 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6005 rc = mdb_cursor_last(mc, key, data);
6008 mfunc = mdb_cursor_last;
6011 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6016 if (mc->mc_flags & C_DEL)
6017 mc->mc_flags ^= C_DEL;
6022 /** Touch all the pages in the cursor stack. Set mc_top.
6023 * Makes sure all the pages are writable, before attempting a write operation.
6024 * @param[in] mc The cursor to operate on.
6027 mdb_cursor_touch(MDB_cursor *mc)
6029 int rc = MDB_SUCCESS;
6031 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6034 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6036 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6037 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6040 *mc->mc_dbflag |= DB_DIRTY;
6045 rc = mdb_page_touch(mc);
6046 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6047 mc->mc_top = mc->mc_snum-1;
6052 /** Do not spill pages to disk if txn is getting full, may fail instead */
6053 #define MDB_NOSPILL 0x8000
6056 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6059 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
6061 MDB_node *leaf = NULL;
6064 MDB_val xdata, *rdata, dkey, olddata;
6066 int do_sub = 0, insert_key, insert_data;
6067 unsigned int mcount = 0, dcount = 0, nospill;
6070 unsigned int nflags;
6073 if (mc == NULL || key == NULL)
6076 env = mc->mc_txn->mt_env;
6078 /* Check this first so counter will always be zero on any
6081 if (flags & MDB_MULTIPLE) {
6082 dcount = data[1].mv_size;
6083 data[1].mv_size = 0;
6084 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6085 return MDB_INCOMPATIBLE;
6088 nospill = flags & MDB_NOSPILL;
6089 flags &= ~MDB_NOSPILL;
6091 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6092 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6094 if (key->mv_size-1 >= ENV_MAXKEY(env))
6095 return MDB_BAD_VALSIZE;
6097 #if SIZE_MAX > MAXDATASIZE
6098 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6099 return MDB_BAD_VALSIZE;
6101 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6102 return MDB_BAD_VALSIZE;
6105 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6106 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6110 if (flags == MDB_CURRENT) {
6111 if (!(mc->mc_flags & C_INITIALIZED))
6114 } else if (mc->mc_db->md_root == P_INVALID) {
6115 /* new database, cursor has nothing to point to */
6118 mc->mc_flags &= ~C_INITIALIZED;
6123 if (flags & MDB_APPEND) {
6125 rc = mdb_cursor_last(mc, &k2, &d2);
6127 rc = mc->mc_dbx->md_cmp(key, &k2);
6130 mc->mc_ki[mc->mc_top]++;
6132 /* new key is <= last key */
6137 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6139 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6140 DPRINTF(("duplicate key [%s]", DKEY(key)));
6142 return MDB_KEYEXIST;
6144 if (rc && rc != MDB_NOTFOUND)
6148 if (mc->mc_flags & C_DEL)
6149 mc->mc_flags ^= C_DEL;
6151 /* Cursor is positioned, check for room in the dirty list */
6153 if (flags & MDB_MULTIPLE) {
6155 xdata.mv_size = data->mv_size * dcount;
6159 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6163 if (rc == MDB_NO_ROOT) {
6165 /* new database, write a root leaf page */
6166 DPUTS("allocating new root leaf page");
6167 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6170 mdb_cursor_push(mc, np);
6171 mc->mc_db->md_root = np->mp_pgno;
6172 mc->mc_db->md_depth++;
6173 *mc->mc_dbflag |= DB_DIRTY;
6174 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6176 np->mp_flags |= P_LEAF2;
6177 mc->mc_flags |= C_INITIALIZED;
6179 /* make sure all cursor pages are writable */
6180 rc2 = mdb_cursor_touch(mc);
6185 insert_key = insert_data = rc;
6187 /* The key does not exist */
6188 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6189 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6190 LEAFSIZE(key, data) > env->me_nodemax)
6192 /* Too big for a node, insert in sub-DB. Set up an empty
6193 * "old sub-page" for prep_subDB to expand to a full page.
6195 fp_flags = P_LEAF|P_DIRTY;
6197 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6198 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6199 olddata.mv_size = PAGEHDRSZ;
6203 /* there's only a key anyway, so this is a no-op */
6204 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6206 unsigned int ksize = mc->mc_db->md_pad;
6207 if (key->mv_size != ksize)
6208 return MDB_BAD_VALSIZE;
6209 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6210 memcpy(ptr, key->mv_data, ksize);
6212 /* if overwriting slot 0 of leaf, need to
6213 * update branch key if there is a parent page
6215 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6216 unsigned short top = mc->mc_top;
6218 /* slot 0 is always an empty key, find real slot */
6219 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6221 if (mc->mc_ki[mc->mc_top])
6222 rc2 = mdb_update_key(mc, key);
6233 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6234 olddata.mv_size = NODEDSZ(leaf);
6235 olddata.mv_data = NODEDATA(leaf);
6238 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6239 /* Prepare (sub-)page/sub-DB to accept the new item,
6240 * if needed. fp: old sub-page or a header faking
6241 * it. mp: new (sub-)page. offset: growth in page
6242 * size. xdata: node data with new page or DB.
6244 unsigned i, offset = 0;
6245 mp = fp = xdata.mv_data = env->me_pbuf;
6246 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6248 /* Was a single item before, must convert now */
6249 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6250 /* Just overwrite the current item */
6251 if (flags == MDB_CURRENT)
6254 #if UINT_MAX < SIZE_MAX
6255 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6256 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6258 /* does data match? */
6259 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6260 if (flags & MDB_NODUPDATA)
6261 return MDB_KEYEXIST;
6266 /* Back up original data item */
6267 dkey.mv_size = olddata.mv_size;
6268 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6270 /* Make sub-page header for the dup items, with dummy body */
6271 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6272 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6273 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6274 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6275 fp->mp_flags |= P_LEAF2;
6276 fp->mp_pad = data->mv_size;
6277 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6279 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6280 (dkey.mv_size & 1) + (data->mv_size & 1);
6282 fp->mp_upper = xdata.mv_size - PAGEBASE;
6283 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6284 } else if (leaf->mn_flags & F_SUBDATA) {
6285 /* Data is on sub-DB, just store it */
6286 flags |= F_DUPDATA|F_SUBDATA;
6289 /* Data is on sub-page */
6290 fp = olddata.mv_data;
6293 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6294 offset = EVEN(NODESIZE + sizeof(indx_t) +
6298 offset = fp->mp_pad;
6299 if (SIZELEFT(fp) < offset) {
6300 offset *= 4; /* space for 4 more */
6303 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6305 fp->mp_flags |= P_DIRTY;
6306 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6307 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6311 xdata.mv_size = olddata.mv_size + offset;
6314 fp_flags = fp->mp_flags;
6315 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6316 /* Too big for a sub-page, convert to sub-DB */
6317 fp_flags &= ~P_SUBP;
6319 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6320 fp_flags |= P_LEAF2;
6321 dummy.md_pad = fp->mp_pad;
6322 dummy.md_flags = MDB_DUPFIXED;
6323 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6324 dummy.md_flags |= MDB_INTEGERKEY;
6330 dummy.md_branch_pages = 0;
6331 dummy.md_leaf_pages = 1;
6332 dummy.md_overflow_pages = 0;
6333 dummy.md_entries = NUMKEYS(fp);
6334 xdata.mv_size = sizeof(MDB_db);
6335 xdata.mv_data = &dummy;
6336 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6338 offset = env->me_psize - olddata.mv_size;
6339 flags |= F_DUPDATA|F_SUBDATA;
6340 dummy.md_root = mp->mp_pgno;
6343 mp->mp_flags = fp_flags | P_DIRTY;
6344 mp->mp_pad = fp->mp_pad;
6345 mp->mp_lower = fp->mp_lower;
6346 mp->mp_upper = fp->mp_upper + offset;
6347 if (fp_flags & P_LEAF2) {
6348 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6350 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6351 olddata.mv_size - fp->mp_upper - PAGEBASE);
6352 for (i=0; i<NUMKEYS(fp); i++)
6353 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6361 mdb_node_del(mc, 0);
6365 /* overflow page overwrites need special handling */
6366 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6369 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6371 memcpy(&pg, olddata.mv_data, sizeof(pg));
6372 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6374 ovpages = omp->mp_pages;
6376 /* Is the ov page large enough? */
6377 if (ovpages >= dpages) {
6378 if (!(omp->mp_flags & P_DIRTY) &&
6379 (level || (env->me_flags & MDB_WRITEMAP)))
6381 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6384 level = 0; /* dirty in this txn or clean */
6387 if (omp->mp_flags & P_DIRTY) {
6388 /* yes, overwrite it. Note in this case we don't
6389 * bother to try shrinking the page if the new data
6390 * is smaller than the overflow threshold.
6393 /* It is writable only in a parent txn */
6394 size_t sz = (size_t) env->me_psize * ovpages, off;
6395 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6401 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6402 mdb_cassert(mc, rc2 == 0);
6403 if (!(flags & MDB_RESERVE)) {
6404 /* Copy end of page, adjusting alignment so
6405 * compiler may copy words instead of bytes.
6407 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6408 memcpy((size_t *)((char *)np + off),
6409 (size_t *)((char *)omp + off), sz - off);
6412 memcpy(np, omp, sz); /* Copy beginning of page */
6415 SETDSZ(leaf, data->mv_size);
6416 if (F_ISSET(flags, MDB_RESERVE))
6417 data->mv_data = METADATA(omp);
6419 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6423 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6425 } else if (data->mv_size == olddata.mv_size) {
6426 /* same size, just replace it. Note that we could
6427 * also reuse this node if the new data is smaller,
6428 * but instead we opt to shrink the node in that case.
6430 if (F_ISSET(flags, MDB_RESERVE))
6431 data->mv_data = olddata.mv_data;
6432 else if (!(mc->mc_flags & C_SUB))
6433 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6435 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6440 mdb_node_del(mc, 0);
6446 nflags = flags & NODE_ADD_FLAGS;
6447 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6448 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6449 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6450 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6452 nflags |= MDB_SPLIT_REPLACE;
6453 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6455 /* There is room already in this leaf page. */
6456 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6457 if (rc == 0 && insert_key) {
6458 /* Adjust other cursors pointing to mp */
6459 MDB_cursor *m2, *m3;
6460 MDB_dbi dbi = mc->mc_dbi;
6461 unsigned i = mc->mc_top;
6462 MDB_page *mp = mc->mc_pg[i];
6464 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6465 if (mc->mc_flags & C_SUB)
6466 m3 = &m2->mc_xcursor->mx_cursor;
6469 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6470 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6477 if (rc == MDB_SUCCESS) {
6478 /* Now store the actual data in the child DB. Note that we're
6479 * storing the user data in the keys field, so there are strict
6480 * size limits on dupdata. The actual data fields of the child
6481 * DB are all zero size.
6489 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6490 if (flags & MDB_CURRENT) {
6491 xflags = MDB_CURRENT|MDB_NOSPILL;
6493 mdb_xcursor_init1(mc, leaf);
6494 xflags = (flags & MDB_NODUPDATA) ?
6495 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6497 /* converted, write the original data first */
6499 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6503 /* Adjust other cursors pointing to mp */
6505 unsigned i = mc->mc_top;
6506 MDB_page *mp = mc->mc_pg[i];
6508 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6509 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6510 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6511 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6512 mdb_xcursor_init1(m2, leaf);
6516 /* we've done our job */
6519 ecount = mc->mc_xcursor->mx_db.md_entries;
6520 if (flags & MDB_APPENDDUP)
6521 xflags |= MDB_APPEND;
6522 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6523 if (flags & F_SUBDATA) {
6524 void *db = NODEDATA(leaf);
6525 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6527 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6529 /* Increment count unless we just replaced an existing item. */
6531 mc->mc_db->md_entries++;
6533 /* Invalidate txn if we created an empty sub-DB */
6536 /* If we succeeded and the key didn't exist before,
6537 * make sure the cursor is marked valid.
6539 mc->mc_flags |= C_INITIALIZED;
6541 if (flags & MDB_MULTIPLE) {
6544 /* let caller know how many succeeded, if any */
6545 data[1].mv_size = mcount;
6546 if (mcount < dcount) {
6547 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6548 insert_key = insert_data = 0;
6555 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6558 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6563 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6569 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6570 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6572 if (!(mc->mc_flags & C_INITIALIZED))
6575 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6576 return MDB_NOTFOUND;
6578 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6581 rc = mdb_cursor_touch(mc);
6585 mp = mc->mc_pg[mc->mc_top];
6588 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6590 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6591 if (flags & MDB_NODUPDATA) {
6592 /* mdb_cursor_del0() will subtract the final entry */
6593 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6595 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6596 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6598 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6601 /* If sub-DB still has entries, we're done */
6602 if (mc->mc_xcursor->mx_db.md_entries) {
6603 if (leaf->mn_flags & F_SUBDATA) {
6604 /* update subDB info */
6605 void *db = NODEDATA(leaf);
6606 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6609 /* shrink fake page */
6610 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6611 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6612 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6613 /* fix other sub-DB cursors pointed at this fake page */
6614 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6615 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6616 if (m2->mc_pg[mc->mc_top] == mp &&
6617 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6618 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6621 mc->mc_db->md_entries--;
6622 mc->mc_flags |= C_DEL;
6625 /* otherwise fall thru and delete the sub-DB */
6628 if (leaf->mn_flags & F_SUBDATA) {
6629 /* add all the child DB's pages to the free list */
6630 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6636 /* add overflow pages to free list */
6637 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6641 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6642 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6643 (rc = mdb_ovpage_free(mc, omp)))
6648 return mdb_cursor_del0(mc);
6651 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6655 /** Allocate and initialize new pages for a database.
6656 * @param[in] mc a cursor on the database being added to.
6657 * @param[in] flags flags defining what type of page is being allocated.
6658 * @param[in] num the number of pages to allocate. This is usually 1,
6659 * unless allocating overflow pages for a large record.
6660 * @param[out] mp Address of a page, or NULL on failure.
6661 * @return 0 on success, non-zero on failure.
6664 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6669 if ((rc = mdb_page_alloc(mc, num, &np)))
6671 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6672 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6673 np->mp_flags = flags | P_DIRTY;
6674 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6675 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6678 mc->mc_db->md_branch_pages++;
6679 else if (IS_LEAF(np))
6680 mc->mc_db->md_leaf_pages++;
6681 else if (IS_OVERFLOW(np)) {
6682 mc->mc_db->md_overflow_pages += num;
6690 /** Calculate the size of a leaf node.
6691 * The size depends on the environment's page size; if a data item
6692 * is too large it will be put onto an overflow page and the node
6693 * size will only include the key and not the data. Sizes are always
6694 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6695 * of the #MDB_node headers.
6696 * @param[in] env The environment handle.
6697 * @param[in] key The key for the node.
6698 * @param[in] data The data for the node.
6699 * @return The number of bytes needed to store the node.
6702 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6706 sz = LEAFSIZE(key, data);
6707 if (sz > env->me_nodemax) {
6708 /* put on overflow page */
6709 sz -= data->mv_size - sizeof(pgno_t);
6712 return EVEN(sz + sizeof(indx_t));
6715 /** Calculate the size of a branch node.
6716 * The size should depend on the environment's page size but since
6717 * we currently don't support spilling large keys onto overflow
6718 * pages, it's simply the size of the #MDB_node header plus the
6719 * size of the key. Sizes are always rounded up to an even number
6720 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6721 * @param[in] env The environment handle.
6722 * @param[in] key The key for the node.
6723 * @return The number of bytes needed to store the node.
6726 mdb_branch_size(MDB_env *env, MDB_val *key)
6731 if (sz > env->me_nodemax) {
6732 /* put on overflow page */
6733 /* not implemented */
6734 /* sz -= key->size - sizeof(pgno_t); */
6737 return sz + sizeof(indx_t);
6740 /** Add a node to the page pointed to by the cursor.
6741 * @param[in] mc The cursor for this operation.
6742 * @param[in] indx The index on the page where the new node should be added.
6743 * @param[in] key The key for the new node.
6744 * @param[in] data The data for the new node, if any.
6745 * @param[in] pgno The page number, if adding a branch node.
6746 * @param[in] flags Flags for the node.
6747 * @return 0 on success, non-zero on failure. Possible errors are:
6749 * <li>ENOMEM - failed to allocate overflow pages for the node.
6750 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6751 * should never happen since all callers already calculate the
6752 * page's free space before calling this function.
6756 mdb_node_add(MDB_cursor *mc, indx_t indx,
6757 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6760 size_t node_size = NODESIZE;
6764 MDB_page *mp = mc->mc_pg[mc->mc_top];
6765 MDB_page *ofp = NULL; /* overflow page */
6768 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6770 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6771 IS_LEAF(mp) ? "leaf" : "branch",
6772 IS_SUBP(mp) ? "sub-" : "",
6773 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6774 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6777 /* Move higher keys up one slot. */
6778 int ksize = mc->mc_db->md_pad, dif;
6779 char *ptr = LEAF2KEY(mp, indx, ksize);
6780 dif = NUMKEYS(mp) - indx;
6782 memmove(ptr+ksize, ptr, dif*ksize);
6783 /* insert new key */
6784 memcpy(ptr, key->mv_data, ksize);
6786 /* Just using these for counting */
6787 mp->mp_lower += sizeof(indx_t);
6788 mp->mp_upper -= ksize - sizeof(indx_t);
6792 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6794 node_size += key->mv_size;
6796 mdb_cassert(mc, data);
6797 if (F_ISSET(flags, F_BIGDATA)) {
6798 /* Data already on overflow page. */
6799 node_size += sizeof(pgno_t);
6800 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6801 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6803 /* Put data on overflow page. */
6804 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6805 data->mv_size, node_size+data->mv_size));
6806 node_size = EVEN(node_size + sizeof(pgno_t));
6807 if ((ssize_t)node_size > room)
6809 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6811 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6815 node_size += data->mv_size;
6818 node_size = EVEN(node_size);
6819 if ((ssize_t)node_size > room)
6823 /* Move higher pointers up one slot. */
6824 for (i = NUMKEYS(mp); i > indx; i--)
6825 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6827 /* Adjust free space offsets. */
6828 ofs = mp->mp_upper - node_size;
6829 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6830 mp->mp_ptrs[indx] = ofs;
6832 mp->mp_lower += sizeof(indx_t);
6834 /* Write the node data. */
6835 node = NODEPTR(mp, indx);
6836 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6837 node->mn_flags = flags;
6839 SETDSZ(node,data->mv_size);
6844 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6847 mdb_cassert(mc, key);
6849 if (F_ISSET(flags, F_BIGDATA))
6850 memcpy(node->mn_data + key->mv_size, data->mv_data,
6852 else if (F_ISSET(flags, MDB_RESERVE))
6853 data->mv_data = node->mn_data + key->mv_size;
6855 memcpy(node->mn_data + key->mv_size, data->mv_data,
6858 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6860 if (F_ISSET(flags, MDB_RESERVE))
6861 data->mv_data = METADATA(ofp);
6863 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6870 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6871 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6872 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6873 DPRINTF(("node size = %"Z"u", node_size));
6874 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6875 return MDB_PAGE_FULL;
6878 /** Delete the specified node from a page.
6879 * @param[in] mc Cursor pointing to the node to delete.
6880 * @param[in] ksize The size of a node. Only used if the page is
6881 * part of a #MDB_DUPFIXED database.
6884 mdb_node_del(MDB_cursor *mc, int ksize)
6886 MDB_page *mp = mc->mc_pg[mc->mc_top];
6887 indx_t indx = mc->mc_ki[mc->mc_top];
6889 indx_t i, j, numkeys, ptr;
6893 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6894 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6895 numkeys = NUMKEYS(mp);
6896 mdb_cassert(mc, indx < numkeys);
6899 int x = numkeys - 1 - indx;
6900 base = LEAF2KEY(mp, indx, ksize);
6902 memmove(base, base + ksize, x * ksize);
6903 mp->mp_lower -= sizeof(indx_t);
6904 mp->mp_upper += ksize - sizeof(indx_t);
6908 node = NODEPTR(mp, indx);
6909 sz = NODESIZE + node->mn_ksize;
6911 if (F_ISSET(node->mn_flags, F_BIGDATA))
6912 sz += sizeof(pgno_t);
6914 sz += NODEDSZ(node);
6918 ptr = mp->mp_ptrs[indx];
6919 for (i = j = 0; i < numkeys; i++) {
6921 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6922 if (mp->mp_ptrs[i] < ptr)
6923 mp->mp_ptrs[j] += sz;
6928 base = (char *)mp + mp->mp_upper + PAGEBASE;
6929 memmove(base + sz, base, ptr - mp->mp_upper);
6931 mp->mp_lower -= sizeof(indx_t);
6935 /** Compact the main page after deleting a node on a subpage.
6936 * @param[in] mp The main page to operate on.
6937 * @param[in] indx The index of the subpage on the main page.
6940 mdb_node_shrink(MDB_page *mp, indx_t indx)
6946 indx_t i, numkeys, ptr;
6948 node = NODEPTR(mp, indx);
6949 sp = (MDB_page *)NODEDATA(node);
6950 delta = SIZELEFT(sp);
6951 xp = (MDB_page *)((char *)sp + delta);
6953 /* shift subpage upward */
6955 nsize = NUMKEYS(sp) * sp->mp_pad;
6957 return; /* do not make the node uneven-sized */
6958 memmove(METADATA(xp), METADATA(sp), nsize);
6961 numkeys = NUMKEYS(sp);
6962 for (i=numkeys-1; i>=0; i--)
6963 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6965 xp->mp_upper = sp->mp_lower;
6966 xp->mp_lower = sp->mp_lower;
6967 xp->mp_flags = sp->mp_flags;
6968 xp->mp_pad = sp->mp_pad;
6969 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6971 nsize = NODEDSZ(node) - delta;
6972 SETDSZ(node, nsize);
6974 /* shift lower nodes upward */
6975 ptr = mp->mp_ptrs[indx];
6976 numkeys = NUMKEYS(mp);
6977 for (i = 0; i < numkeys; i++) {
6978 if (mp->mp_ptrs[i] <= ptr)
6979 mp->mp_ptrs[i] += delta;
6982 base = (char *)mp + mp->mp_upper + PAGEBASE;
6983 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6984 mp->mp_upper += delta;
6987 /** Initial setup of a sorted-dups cursor.
6988 * Sorted duplicates are implemented as a sub-database for the given key.
6989 * The duplicate data items are actually keys of the sub-database.
6990 * Operations on the duplicate data items are performed using a sub-cursor
6991 * initialized when the sub-database is first accessed. This function does
6992 * the preliminary setup of the sub-cursor, filling in the fields that
6993 * depend only on the parent DB.
6994 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6997 mdb_xcursor_init0(MDB_cursor *mc)
6999 MDB_xcursor *mx = mc->mc_xcursor;
7001 mx->mx_cursor.mc_xcursor = NULL;
7002 mx->mx_cursor.mc_txn = mc->mc_txn;
7003 mx->mx_cursor.mc_db = &mx->mx_db;
7004 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7005 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7006 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7007 mx->mx_cursor.mc_snum = 0;
7008 mx->mx_cursor.mc_top = 0;
7009 mx->mx_cursor.mc_flags = C_SUB;
7010 mx->mx_dbx.md_name.mv_size = 0;
7011 mx->mx_dbx.md_name.mv_data = NULL;
7012 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7013 mx->mx_dbx.md_dcmp = NULL;
7014 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7017 /** Final setup of a sorted-dups cursor.
7018 * Sets up the fields that depend on the data from the main cursor.
7019 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7020 * @param[in] node The data containing the #MDB_db record for the
7021 * sorted-dup database.
7024 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7026 MDB_xcursor *mx = mc->mc_xcursor;
7028 if (node->mn_flags & F_SUBDATA) {
7029 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7030 mx->mx_cursor.mc_pg[0] = 0;
7031 mx->mx_cursor.mc_snum = 0;
7032 mx->mx_cursor.mc_top = 0;
7033 mx->mx_cursor.mc_flags = C_SUB;
7035 MDB_page *fp = NODEDATA(node);
7036 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7037 mx->mx_db.md_flags = 0;
7038 mx->mx_db.md_depth = 1;
7039 mx->mx_db.md_branch_pages = 0;
7040 mx->mx_db.md_leaf_pages = 1;
7041 mx->mx_db.md_overflow_pages = 0;
7042 mx->mx_db.md_entries = NUMKEYS(fp);
7043 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7044 mx->mx_cursor.mc_snum = 1;
7045 mx->mx_cursor.mc_top = 0;
7046 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7047 mx->mx_cursor.mc_pg[0] = fp;
7048 mx->mx_cursor.mc_ki[0] = 0;
7049 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7050 mx->mx_db.md_flags = MDB_DUPFIXED;
7051 mx->mx_db.md_pad = fp->mp_pad;
7052 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7053 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7056 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7057 mx->mx_db.md_root));
7058 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7059 #if UINT_MAX < SIZE_MAX
7060 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7061 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7065 /** Initialize a cursor for a given transaction and database. */
7067 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7070 mc->mc_backup = NULL;
7073 mc->mc_db = &txn->mt_dbs[dbi];
7074 mc->mc_dbx = &txn->mt_dbxs[dbi];
7075 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7080 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7081 mdb_tassert(txn, mx != NULL);
7082 mc->mc_xcursor = mx;
7083 mdb_xcursor_init0(mc);
7085 mc->mc_xcursor = NULL;
7087 if (*mc->mc_dbflag & DB_STALE) {
7088 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7093 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7096 size_t size = sizeof(MDB_cursor);
7098 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7101 if (txn->mt_flags & MDB_TXN_ERROR)
7104 /* Allow read access to the freelist */
7105 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7108 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7109 size += sizeof(MDB_xcursor);
7111 if ((mc = malloc(size)) != NULL) {
7112 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7113 if (txn->mt_cursors) {
7114 mc->mc_next = txn->mt_cursors[dbi];
7115 txn->mt_cursors[dbi] = mc;
7116 mc->mc_flags |= C_UNTRACK;
7128 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7130 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7133 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7136 if (txn->mt_flags & MDB_TXN_ERROR)
7139 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7143 /* Return the count of duplicate data items for the current key */
7145 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7149 if (mc == NULL || countp == NULL)
7152 if (mc->mc_xcursor == NULL)
7153 return MDB_INCOMPATIBLE;
7155 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7158 if (!(mc->mc_flags & C_INITIALIZED))
7161 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7162 return MDB_NOTFOUND;
7164 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7165 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7168 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7171 *countp = mc->mc_xcursor->mx_db.md_entries;
7177 mdb_cursor_close(MDB_cursor *mc)
7179 if (mc && !mc->mc_backup) {
7180 /* remove from txn, if tracked */
7181 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7182 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7183 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7185 *prev = mc->mc_next;
7192 mdb_cursor_txn(MDB_cursor *mc)
7194 if (!mc) return NULL;
7199 mdb_cursor_dbi(MDB_cursor *mc)
7204 /** Replace the key for a branch node with a new key.
7205 * @param[in] mc Cursor pointing to the node to operate on.
7206 * @param[in] key The new key to use.
7207 * @return 0 on success, non-zero on failure.
7210 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7216 int delta, ksize, oksize;
7217 indx_t ptr, i, numkeys, indx;
7220 indx = mc->mc_ki[mc->mc_top];
7221 mp = mc->mc_pg[mc->mc_top];
7222 node = NODEPTR(mp, indx);
7223 ptr = mp->mp_ptrs[indx];
7227 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7228 k2.mv_data = NODEKEY(node);
7229 k2.mv_size = node->mn_ksize;
7230 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7232 mdb_dkey(&k2, kbuf2),
7238 /* Sizes must be 2-byte aligned. */
7239 ksize = EVEN(key->mv_size);
7240 oksize = EVEN(node->mn_ksize);
7241 delta = ksize - oksize;
7243 /* Shift node contents if EVEN(key length) changed. */
7245 if (delta > 0 && SIZELEFT(mp) < delta) {
7247 /* not enough space left, do a delete and split */
7248 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7249 pgno = NODEPGNO(node);
7250 mdb_node_del(mc, 0);
7251 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7254 numkeys = NUMKEYS(mp);
7255 for (i = 0; i < numkeys; i++) {
7256 if (mp->mp_ptrs[i] <= ptr)
7257 mp->mp_ptrs[i] -= delta;
7260 base = (char *)mp + mp->mp_upper + PAGEBASE;
7261 len = ptr - mp->mp_upper + NODESIZE;
7262 memmove(base - delta, base, len);
7263 mp->mp_upper -= delta;
7265 node = NODEPTR(mp, indx);
7268 /* But even if no shift was needed, update ksize */
7269 if (node->mn_ksize != key->mv_size)
7270 node->mn_ksize = key->mv_size;
7273 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7279 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7281 /** Move a node from csrc to cdst.
7284 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7291 unsigned short flags;
7295 /* Mark src and dst as dirty. */
7296 if ((rc = mdb_page_touch(csrc)) ||
7297 (rc = mdb_page_touch(cdst)))
7300 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7301 key.mv_size = csrc->mc_db->md_pad;
7302 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7304 data.mv_data = NULL;
7308 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7309 mdb_cassert(csrc, !((size_t)srcnode & 1));
7310 srcpg = NODEPGNO(srcnode);
7311 flags = srcnode->mn_flags;
7312 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7313 unsigned int snum = csrc->mc_snum;
7315 /* must find the lowest key below src */
7316 rc = mdb_page_search_lowest(csrc);
7319 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7320 key.mv_size = csrc->mc_db->md_pad;
7321 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7323 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7324 key.mv_size = NODEKSZ(s2);
7325 key.mv_data = NODEKEY(s2);
7327 csrc->mc_snum = snum--;
7328 csrc->mc_top = snum;
7330 key.mv_size = NODEKSZ(srcnode);
7331 key.mv_data = NODEKEY(srcnode);
7333 data.mv_size = NODEDSZ(srcnode);
7334 data.mv_data = NODEDATA(srcnode);
7336 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7337 unsigned int snum = cdst->mc_snum;
7340 /* must find the lowest key below dst */
7341 mdb_cursor_copy(cdst, &mn);
7342 rc = mdb_page_search_lowest(&mn);
7345 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7346 bkey.mv_size = mn.mc_db->md_pad;
7347 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7349 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7350 bkey.mv_size = NODEKSZ(s2);
7351 bkey.mv_data = NODEKEY(s2);
7353 mn.mc_snum = snum--;
7356 rc = mdb_update_key(&mn, &bkey);
7361 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7362 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7363 csrc->mc_ki[csrc->mc_top],
7365 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7366 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7368 /* Add the node to the destination page.
7370 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7371 if (rc != MDB_SUCCESS)
7374 /* Delete the node from the source page.
7376 mdb_node_del(csrc, key.mv_size);
7379 /* Adjust other cursors pointing to mp */
7380 MDB_cursor *m2, *m3;
7381 MDB_dbi dbi = csrc->mc_dbi;
7382 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7384 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7385 if (csrc->mc_flags & C_SUB)
7386 m3 = &m2->mc_xcursor->mx_cursor;
7389 if (m3 == csrc) continue;
7390 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7391 csrc->mc_ki[csrc->mc_top]) {
7392 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7393 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7398 /* Update the parent separators.
7400 if (csrc->mc_ki[csrc->mc_top] == 0) {
7401 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7402 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7403 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7405 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7406 key.mv_size = NODEKSZ(srcnode);
7407 key.mv_data = NODEKEY(srcnode);
7409 DPRINTF(("update separator for source page %"Z"u to [%s]",
7410 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7411 mdb_cursor_copy(csrc, &mn);
7414 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7417 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7419 indx_t ix = csrc->mc_ki[csrc->mc_top];
7420 nullkey.mv_size = 0;
7421 csrc->mc_ki[csrc->mc_top] = 0;
7422 rc = mdb_update_key(csrc, &nullkey);
7423 csrc->mc_ki[csrc->mc_top] = ix;
7424 mdb_cassert(csrc, rc == MDB_SUCCESS);
7428 if (cdst->mc_ki[cdst->mc_top] == 0) {
7429 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7430 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7431 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7433 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7434 key.mv_size = NODEKSZ(srcnode);
7435 key.mv_data = NODEKEY(srcnode);
7437 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7438 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7439 mdb_cursor_copy(cdst, &mn);
7442 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7445 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7447 indx_t ix = cdst->mc_ki[cdst->mc_top];
7448 nullkey.mv_size = 0;
7449 cdst->mc_ki[cdst->mc_top] = 0;
7450 rc = mdb_update_key(cdst, &nullkey);
7451 cdst->mc_ki[cdst->mc_top] = ix;
7452 mdb_cassert(csrc, rc == MDB_SUCCESS);
7459 /** Merge one page into another.
7460 * The nodes from the page pointed to by \b csrc will
7461 * be copied to the page pointed to by \b cdst and then
7462 * the \b csrc page will be freed.
7463 * @param[in] csrc Cursor pointing to the source page.
7464 * @param[in] cdst Cursor pointing to the destination page.
7465 * @return 0 on success, non-zero on failure.
7468 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7470 MDB_page *psrc, *pdst;
7477 psrc = csrc->mc_pg[csrc->mc_top];
7478 pdst = cdst->mc_pg[cdst->mc_top];
7480 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7482 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7483 mdb_cassert(csrc, cdst->mc_snum > 1);
7485 /* Mark dst as dirty. */
7486 if ((rc = mdb_page_touch(cdst)))
7489 /* Move all nodes from src to dst.
7491 j = nkeys = NUMKEYS(pdst);
7492 if (IS_LEAF2(psrc)) {
7493 key.mv_size = csrc->mc_db->md_pad;
7494 key.mv_data = METADATA(psrc);
7495 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7496 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7497 if (rc != MDB_SUCCESS)
7499 key.mv_data = (char *)key.mv_data + key.mv_size;
7502 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7503 srcnode = NODEPTR(psrc, i);
7504 if (i == 0 && IS_BRANCH(psrc)) {
7507 mdb_cursor_copy(csrc, &mn);
7508 /* must find the lowest key below src */
7509 rc = mdb_page_search_lowest(&mn);
7512 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7513 key.mv_size = mn.mc_db->md_pad;
7514 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7516 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7517 key.mv_size = NODEKSZ(s2);
7518 key.mv_data = NODEKEY(s2);
7521 key.mv_size = srcnode->mn_ksize;
7522 key.mv_data = NODEKEY(srcnode);
7525 data.mv_size = NODEDSZ(srcnode);
7526 data.mv_data = NODEDATA(srcnode);
7527 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7528 if (rc != MDB_SUCCESS)
7533 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7534 pdst->mp_pgno, NUMKEYS(pdst),
7535 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7537 /* Unlink the src page from parent and add to free list.
7540 mdb_node_del(csrc, 0);
7541 if (csrc->mc_ki[csrc->mc_top] == 0) {
7543 rc = mdb_update_key(csrc, &key);
7551 psrc = csrc->mc_pg[csrc->mc_top];
7552 /* If not operating on FreeDB, allow this page to be reused
7553 * in this txn. Otherwise just add to free list.
7555 rc = mdb_page_loose(csrc, psrc);
7559 csrc->mc_db->md_leaf_pages--;
7561 csrc->mc_db->md_branch_pages--;
7563 /* Adjust other cursors pointing to mp */
7564 MDB_cursor *m2, *m3;
7565 MDB_dbi dbi = csrc->mc_dbi;
7567 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7568 if (csrc->mc_flags & C_SUB)
7569 m3 = &m2->mc_xcursor->mx_cursor;
7572 if (m3 == csrc) continue;
7573 if (m3->mc_snum < csrc->mc_snum) continue;
7574 if (m3->mc_pg[csrc->mc_top] == psrc) {
7575 m3->mc_pg[csrc->mc_top] = pdst;
7576 m3->mc_ki[csrc->mc_top] += nkeys;
7581 unsigned int snum = cdst->mc_snum;
7582 uint16_t depth = cdst->mc_db->md_depth;
7583 mdb_cursor_pop(cdst);
7584 rc = mdb_rebalance(cdst);
7585 /* Did the tree shrink? */
7586 if (depth > cdst->mc_db->md_depth)
7588 cdst->mc_snum = snum;
7589 cdst->mc_top = snum-1;
7594 /** Copy the contents of a cursor.
7595 * @param[in] csrc The cursor to copy from.
7596 * @param[out] cdst The cursor to copy to.
7599 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7603 cdst->mc_txn = csrc->mc_txn;
7604 cdst->mc_dbi = csrc->mc_dbi;
7605 cdst->mc_db = csrc->mc_db;
7606 cdst->mc_dbx = csrc->mc_dbx;
7607 cdst->mc_snum = csrc->mc_snum;
7608 cdst->mc_top = csrc->mc_top;
7609 cdst->mc_flags = csrc->mc_flags;
7611 for (i=0; i<csrc->mc_snum; i++) {
7612 cdst->mc_pg[i] = csrc->mc_pg[i];
7613 cdst->mc_ki[i] = csrc->mc_ki[i];
7617 /** Rebalance the tree after a delete operation.
7618 * @param[in] mc Cursor pointing to the page where rebalancing
7620 * @return 0 on success, non-zero on failure.
7623 mdb_rebalance(MDB_cursor *mc)
7627 unsigned int ptop, minkeys;
7631 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7632 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7633 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7634 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7635 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7637 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7638 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7639 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7640 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7644 if (mc->mc_snum < 2) {
7645 MDB_page *mp = mc->mc_pg[0];
7647 DPUTS("Can't rebalance a subpage, ignoring");
7650 if (NUMKEYS(mp) == 0) {
7651 DPUTS("tree is completely empty");
7652 mc->mc_db->md_root = P_INVALID;
7653 mc->mc_db->md_depth = 0;
7654 mc->mc_db->md_leaf_pages = 0;
7655 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7658 /* Adjust cursors pointing to mp */
7661 mc->mc_flags &= ~C_INITIALIZED;
7663 MDB_cursor *m2, *m3;
7664 MDB_dbi dbi = mc->mc_dbi;
7666 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7667 if (mc->mc_flags & C_SUB)
7668 m3 = &m2->mc_xcursor->mx_cursor;
7671 if (m3->mc_snum < mc->mc_snum) continue;
7672 if (m3->mc_pg[0] == mp) {
7675 m3->mc_flags &= ~C_INITIALIZED;
7679 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7681 DPUTS("collapsing root page!");
7682 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7685 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7686 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7689 mc->mc_db->md_depth--;
7690 mc->mc_db->md_branch_pages--;
7691 mc->mc_ki[0] = mc->mc_ki[1];
7692 for (i = 1; i<mc->mc_db->md_depth; i++) {
7693 mc->mc_pg[i] = mc->mc_pg[i+1];
7694 mc->mc_ki[i] = mc->mc_ki[i+1];
7697 /* Adjust other cursors pointing to mp */
7698 MDB_cursor *m2, *m3;
7699 MDB_dbi dbi = mc->mc_dbi;
7701 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7702 if (mc->mc_flags & C_SUB)
7703 m3 = &m2->mc_xcursor->mx_cursor;
7706 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7707 if (m3->mc_pg[0] == mp) {
7710 for (i=0; i<m3->mc_snum; i++) {
7711 m3->mc_pg[i] = m3->mc_pg[i+1];
7712 m3->mc_ki[i] = m3->mc_ki[i+1];
7718 DPUTS("root page doesn't need rebalancing");
7722 /* The parent (branch page) must have at least 2 pointers,
7723 * otherwise the tree is invalid.
7725 ptop = mc->mc_top-1;
7726 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7728 /* Leaf page fill factor is below the threshold.
7729 * Try to move keys from left or right neighbor, or
7730 * merge with a neighbor page.
7735 mdb_cursor_copy(mc, &mn);
7736 mn.mc_xcursor = NULL;
7738 oldki = mc->mc_ki[mc->mc_top];
7739 if (mc->mc_ki[ptop] == 0) {
7740 /* We're the leftmost leaf in our parent.
7742 DPUTS("reading right neighbor");
7744 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7745 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7748 mn.mc_ki[mn.mc_top] = 0;
7749 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7751 /* There is at least one neighbor to the left.
7753 DPUTS("reading left neighbor");
7755 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7756 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7759 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7760 mc->mc_ki[mc->mc_top] = 0;
7763 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7764 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7765 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7767 /* If the neighbor page is above threshold and has enough keys,
7768 * move one key from it. Otherwise we should try to merge them.
7769 * (A branch page must never have less than 2 keys.)
7771 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7772 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7773 rc = mdb_node_move(&mn, mc);
7774 if (mc->mc_ki[ptop]) {
7778 if (mc->mc_ki[ptop] == 0) {
7779 rc = mdb_page_merge(&mn, mc);
7781 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7782 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7783 rc = mdb_page_merge(mc, &mn);
7784 mdb_cursor_copy(&mn, mc);
7786 mc->mc_flags &= ~C_EOF;
7788 mc->mc_ki[mc->mc_top] = oldki;
7792 /** Complete a delete operation started by #mdb_cursor_del(). */
7794 mdb_cursor_del0(MDB_cursor *mc)
7801 ki = mc->mc_ki[mc->mc_top];
7802 mdb_node_del(mc, mc->mc_db->md_pad);
7803 mc->mc_db->md_entries--;
7804 rc = mdb_rebalance(mc);
7806 if (rc == MDB_SUCCESS) {
7807 MDB_cursor *m2, *m3;
7808 MDB_dbi dbi = mc->mc_dbi;
7810 mp = mc->mc_pg[mc->mc_top];
7811 nkeys = NUMKEYS(mp);
7813 /* if mc points past last node in page, find next sibling */
7814 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7815 rc = mdb_cursor_sibling(mc, 1);
7816 if (rc == MDB_NOTFOUND) {
7817 mc->mc_flags |= C_EOF;
7822 /* Adjust other cursors pointing to mp */
7823 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7824 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7825 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7827 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7829 if (m3->mc_pg[mc->mc_top] == mp) {
7830 if (m3->mc_ki[mc->mc_top] >= ki) {
7831 m3->mc_flags |= C_DEL;
7832 if (m3->mc_ki[mc->mc_top] > ki)
7833 m3->mc_ki[mc->mc_top]--;
7834 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7835 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7837 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7838 rc = mdb_cursor_sibling(m3, 1);
7839 if (rc == MDB_NOTFOUND) {
7840 m3->mc_flags |= C_EOF;
7846 mc->mc_flags |= C_DEL;
7850 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7855 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7856 MDB_val *key, MDB_val *data)
7858 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7861 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7862 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7864 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7865 /* must ignore any data */
7869 return mdb_del0(txn, dbi, key, data, 0);
7873 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7874 MDB_val *key, MDB_val *data, unsigned flags)
7879 MDB_val rdata, *xdata;
7883 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7885 mdb_cursor_init(&mc, txn, dbi, &mx);
7894 flags |= MDB_NODUPDATA;
7896 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7898 /* let mdb_page_split know about this cursor if needed:
7899 * delete will trigger a rebalance; if it needs to move
7900 * a node from one page to another, it will have to
7901 * update the parent's separator key(s). If the new sepkey
7902 * is larger than the current one, the parent page may
7903 * run out of space, triggering a split. We need this
7904 * cursor to be consistent until the end of the rebalance.
7906 mc.mc_flags |= C_UNTRACK;
7907 mc.mc_next = txn->mt_cursors[dbi];
7908 txn->mt_cursors[dbi] = &mc;
7909 rc = mdb_cursor_del(&mc, flags);
7910 txn->mt_cursors[dbi] = mc.mc_next;
7915 /** Split a page and insert a new node.
7916 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7917 * The cursor will be updated to point to the actual page and index where
7918 * the node got inserted after the split.
7919 * @param[in] newkey The key for the newly inserted node.
7920 * @param[in] newdata The data for the newly inserted node.
7921 * @param[in] newpgno The page number, if the new node is a branch node.
7922 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7923 * @return 0 on success, non-zero on failure.
7926 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7927 unsigned int nflags)
7930 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7933 int i, j, split_indx, nkeys, pmax;
7934 MDB_env *env = mc->mc_txn->mt_env;
7936 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7937 MDB_page *copy = NULL;
7938 MDB_page *mp, *rp, *pp;
7943 mp = mc->mc_pg[mc->mc_top];
7944 newindx = mc->mc_ki[mc->mc_top];
7945 nkeys = NUMKEYS(mp);
7947 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7948 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7949 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7951 /* Create a right sibling. */
7952 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7954 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7956 if (mc->mc_snum < 2) {
7957 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7959 /* shift current top to make room for new parent */
7960 mc->mc_pg[1] = mc->mc_pg[0];
7961 mc->mc_ki[1] = mc->mc_ki[0];
7964 mc->mc_db->md_root = pp->mp_pgno;
7965 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7966 mc->mc_db->md_depth++;
7969 /* Add left (implicit) pointer. */
7970 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7971 /* undo the pre-push */
7972 mc->mc_pg[0] = mc->mc_pg[1];
7973 mc->mc_ki[0] = mc->mc_ki[1];
7974 mc->mc_db->md_root = mp->mp_pgno;
7975 mc->mc_db->md_depth--;
7982 ptop = mc->mc_top-1;
7983 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7986 mc->mc_flags |= C_SPLITTING;
7987 mdb_cursor_copy(mc, &mn);
7988 mn.mc_pg[mn.mc_top] = rp;
7989 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7991 if (nflags & MDB_APPEND) {
7992 mn.mc_ki[mn.mc_top] = 0;
7994 split_indx = newindx;
7998 split_indx = (nkeys+1) / 2;
8003 unsigned int lsize, rsize, ksize;
8004 /* Move half of the keys to the right sibling */
8005 x = mc->mc_ki[mc->mc_top] - split_indx;
8006 ksize = mc->mc_db->md_pad;
8007 split = LEAF2KEY(mp, split_indx, ksize);
8008 rsize = (nkeys - split_indx) * ksize;
8009 lsize = (nkeys - split_indx) * sizeof(indx_t);
8010 mp->mp_lower -= lsize;
8011 rp->mp_lower += lsize;
8012 mp->mp_upper += rsize - lsize;
8013 rp->mp_upper -= rsize - lsize;
8014 sepkey.mv_size = ksize;
8015 if (newindx == split_indx) {
8016 sepkey.mv_data = newkey->mv_data;
8018 sepkey.mv_data = split;
8021 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8022 memcpy(rp->mp_ptrs, split, rsize);
8023 sepkey.mv_data = rp->mp_ptrs;
8024 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8025 memcpy(ins, newkey->mv_data, ksize);
8026 mp->mp_lower += sizeof(indx_t);
8027 mp->mp_upper -= ksize - sizeof(indx_t);
8030 memcpy(rp->mp_ptrs, split, x * ksize);
8031 ins = LEAF2KEY(rp, x, ksize);
8032 memcpy(ins, newkey->mv_data, ksize);
8033 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8034 rp->mp_lower += sizeof(indx_t);
8035 rp->mp_upper -= ksize - sizeof(indx_t);
8036 mc->mc_ki[mc->mc_top] = x;
8037 mc->mc_pg[mc->mc_top] = rp;
8040 int psize, nsize, k;
8041 /* Maximum free space in an empty page */
8042 pmax = env->me_psize - PAGEHDRSZ;
8044 nsize = mdb_leaf_size(env, newkey, newdata);
8046 nsize = mdb_branch_size(env, newkey);
8047 nsize = EVEN(nsize);
8049 /* grab a page to hold a temporary copy */
8050 copy = mdb_page_malloc(mc->mc_txn, 1);
8055 copy->mp_pgno = mp->mp_pgno;
8056 copy->mp_flags = mp->mp_flags;
8057 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8058 copy->mp_upper = env->me_psize - PAGEBASE;
8060 /* prepare to insert */
8061 for (i=0, j=0; i<nkeys; i++) {
8063 copy->mp_ptrs[j++] = 0;
8065 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8068 /* When items are relatively large the split point needs
8069 * to be checked, because being off-by-one will make the
8070 * difference between success or failure in mdb_node_add.
8072 * It's also relevant if a page happens to be laid out
8073 * such that one half of its nodes are all "small" and
8074 * the other half of its nodes are "large." If the new
8075 * item is also "large" and falls on the half with
8076 * "large" nodes, it also may not fit.
8078 * As a final tweak, if the new item goes on the last
8079 * spot on the page (and thus, onto the new page), bias
8080 * the split so the new page is emptier than the old page.
8081 * This yields better packing during sequential inserts.
8083 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8084 /* Find split point */
8086 if (newindx <= split_indx || newindx >= nkeys) {
8088 k = newindx >= nkeys ? nkeys : split_indx+2;
8093 for (; i!=k; i+=j) {
8098 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8099 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8101 if (F_ISSET(node->mn_flags, F_BIGDATA))
8102 psize += sizeof(pgno_t);
8104 psize += NODEDSZ(node);
8106 psize = EVEN(psize);
8108 if (psize > pmax || i == k-j) {
8109 split_indx = i + (j<0);
8114 if (split_indx == newindx) {
8115 sepkey.mv_size = newkey->mv_size;
8116 sepkey.mv_data = newkey->mv_data;
8118 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8119 sepkey.mv_size = node->mn_ksize;
8120 sepkey.mv_data = NODEKEY(node);
8125 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8127 /* Copy separator key to the parent.
8129 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8133 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8138 if (mn.mc_snum == mc->mc_snum) {
8139 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8140 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8141 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8142 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8147 /* Right page might now have changed parent.
8148 * Check if left page also changed parent.
8150 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8151 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8152 for (i=0; i<ptop; i++) {
8153 mc->mc_pg[i] = mn.mc_pg[i];
8154 mc->mc_ki[i] = mn.mc_ki[i];
8156 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8157 if (mn.mc_ki[ptop]) {
8158 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8160 /* find right page's left sibling */
8161 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8162 mdb_cursor_sibling(mc, 0);
8167 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8170 mc->mc_flags ^= C_SPLITTING;
8171 if (rc != MDB_SUCCESS) {
8174 if (nflags & MDB_APPEND) {
8175 mc->mc_pg[mc->mc_top] = rp;
8176 mc->mc_ki[mc->mc_top] = 0;
8177 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8180 for (i=0; i<mc->mc_top; i++)
8181 mc->mc_ki[i] = mn.mc_ki[i];
8182 } else if (!IS_LEAF2(mp)) {
8184 mc->mc_pg[mc->mc_top] = rp;
8189 rkey.mv_data = newkey->mv_data;
8190 rkey.mv_size = newkey->mv_size;
8196 /* Update index for the new key. */
8197 mc->mc_ki[mc->mc_top] = j;
8199 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8200 rkey.mv_data = NODEKEY(node);
8201 rkey.mv_size = node->mn_ksize;
8203 xdata.mv_data = NODEDATA(node);
8204 xdata.mv_size = NODEDSZ(node);
8207 pgno = NODEPGNO(node);
8208 flags = node->mn_flags;
8211 if (!IS_LEAF(mp) && j == 0) {
8212 /* First branch index doesn't need key data. */
8216 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8222 mc->mc_pg[mc->mc_top] = copy;
8227 } while (i != split_indx);
8229 nkeys = NUMKEYS(copy);
8230 for (i=0; i<nkeys; i++)
8231 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8232 mp->mp_lower = copy->mp_lower;
8233 mp->mp_upper = copy->mp_upper;
8234 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8235 env->me_psize - copy->mp_upper - PAGEBASE);
8237 /* reset back to original page */
8238 if (newindx < split_indx) {
8239 mc->mc_pg[mc->mc_top] = mp;
8240 if (nflags & MDB_RESERVE) {
8241 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8242 if (!(node->mn_flags & F_BIGDATA))
8243 newdata->mv_data = NODEDATA(node);
8246 mc->mc_pg[mc->mc_top] = rp;
8248 /* Make sure mc_ki is still valid.
8250 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8251 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8252 for (i=0; i<=ptop; i++) {
8253 mc->mc_pg[i] = mn.mc_pg[i];
8254 mc->mc_ki[i] = mn.mc_ki[i];
8261 /* Adjust other cursors pointing to mp */
8262 MDB_cursor *m2, *m3;
8263 MDB_dbi dbi = mc->mc_dbi;
8264 int fixup = NUMKEYS(mp);
8266 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8267 if (mc->mc_flags & C_SUB)
8268 m3 = &m2->mc_xcursor->mx_cursor;
8273 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8275 if (m3->mc_flags & C_SPLITTING)
8280 for (k=m3->mc_top; k>=0; k--) {
8281 m3->mc_ki[k+1] = m3->mc_ki[k];
8282 m3->mc_pg[k+1] = m3->mc_pg[k];
8284 if (m3->mc_ki[0] >= split_indx) {
8289 m3->mc_pg[0] = mc->mc_pg[0];
8293 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8294 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8295 m3->mc_ki[mc->mc_top]++;
8296 if (m3->mc_ki[mc->mc_top] >= fixup) {
8297 m3->mc_pg[mc->mc_top] = rp;
8298 m3->mc_ki[mc->mc_top] -= fixup;
8299 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8301 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8302 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8307 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8310 if (copy) /* tmp page */
8311 mdb_page_free(env, copy);
8313 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8318 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8319 MDB_val *key, MDB_val *data, unsigned int flags)
8324 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8327 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8330 mdb_cursor_init(&mc, txn, dbi, &mx);
8331 return mdb_cursor_put(&mc, key, data, flags);
8335 #define MDB_WBUF (1024*1024)
8338 /** State needed for a compacting copy. */
8339 typedef struct mdb_copy {
8340 pthread_mutex_t mc_mutex;
8341 pthread_cond_t mc_cond;
8348 pgno_t mc_next_pgno;
8351 volatile int mc_new;
8356 /** Dedicated writer thread for compacting copy. */
8357 static THREAD_RET ESECT
8358 mdb_env_copythr(void *arg)
8362 int toggle = 0, wsize, rc;
8365 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8368 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8371 pthread_mutex_lock(&my->mc_mutex);
8373 pthread_cond_signal(&my->mc_cond);
8376 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8377 if (my->mc_new < 0) {
8382 wsize = my->mc_wlen[toggle];
8383 ptr = my->mc_wbuf[toggle];
8386 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8390 } else if (len > 0) {
8404 /* If there's an overflow page tail, write it too */
8405 if (my->mc_olen[toggle]) {
8406 wsize = my->mc_olen[toggle];
8407 ptr = my->mc_over[toggle];
8408 my->mc_olen[toggle] = 0;
8411 my->mc_wlen[toggle] = 0;
8413 pthread_cond_signal(&my->mc_cond);
8415 pthread_cond_signal(&my->mc_cond);
8416 pthread_mutex_unlock(&my->mc_mutex);
8417 return (THREAD_RET)0;
8421 /** Tell the writer thread there's a buffer ready to write */
8423 mdb_env_cthr_toggle(mdb_copy *my, int st)
8425 int toggle = my->mc_toggle ^ 1;
8426 pthread_mutex_lock(&my->mc_mutex);
8427 if (my->mc_status) {
8428 pthread_mutex_unlock(&my->mc_mutex);
8429 return my->mc_status;
8431 while (my->mc_new == 1)
8432 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8434 my->mc_toggle = toggle;
8435 pthread_cond_signal(&my->mc_cond);
8436 pthread_mutex_unlock(&my->mc_mutex);
8440 /** Depth-first tree traversal for compacting copy. */
8442 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8445 MDB_txn *txn = my->mc_txn;
8447 MDB_page *mo, *mp, *leaf;
8452 /* Empty DB, nothing to do */
8453 if (*pg == P_INVALID)
8460 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8463 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8467 /* Make cursor pages writable */
8468 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8472 for (i=0; i<mc.mc_top; i++) {
8473 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8474 mc.mc_pg[i] = (MDB_page *)ptr;
8475 ptr += my->mc_env->me_psize;
8478 /* This is writable space for a leaf page. Usually not needed. */
8479 leaf = (MDB_page *)ptr;
8481 toggle = my->mc_toggle;
8482 while (mc.mc_snum > 0) {
8484 mp = mc.mc_pg[mc.mc_top];
8488 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8489 for (i=0; i<n; i++) {
8490 ni = NODEPTR(mp, i);
8491 if (ni->mn_flags & F_BIGDATA) {
8495 /* Need writable leaf */
8497 mc.mc_pg[mc.mc_top] = leaf;
8498 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8500 ni = NODEPTR(mp, i);
8503 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8504 rc = mdb_page_get(txn, pg, &omp, NULL);
8507 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8508 rc = mdb_env_cthr_toggle(my, 1);
8511 toggle = my->mc_toggle;
8513 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8514 memcpy(mo, omp, my->mc_env->me_psize);
8515 mo->mp_pgno = my->mc_next_pgno;
8516 my->mc_next_pgno += omp->mp_pages;
8517 my->mc_wlen[toggle] += my->mc_env->me_psize;
8518 if (omp->mp_pages > 1) {
8519 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8520 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8521 rc = mdb_env_cthr_toggle(my, 1);
8524 toggle = my->mc_toggle;
8526 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8527 } else if (ni->mn_flags & F_SUBDATA) {
8530 /* Need writable leaf */
8532 mc.mc_pg[mc.mc_top] = leaf;
8533 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8535 ni = NODEPTR(mp, i);
8538 memcpy(&db, NODEDATA(ni), sizeof(db));
8539 my->mc_toggle = toggle;
8540 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8543 toggle = my->mc_toggle;
8544 memcpy(NODEDATA(ni), &db, sizeof(db));
8549 mc.mc_ki[mc.mc_top]++;
8550 if (mc.mc_ki[mc.mc_top] < n) {
8553 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8555 rc = mdb_page_get(txn, pg, &mp, NULL);
8560 mc.mc_ki[mc.mc_top] = 0;
8561 if (IS_BRANCH(mp)) {
8562 /* Whenever we advance to a sibling branch page,
8563 * we must proceed all the way down to its first leaf.
8565 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8568 mc.mc_pg[mc.mc_top] = mp;
8572 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8573 rc = mdb_env_cthr_toggle(my, 1);
8576 toggle = my->mc_toggle;
8578 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8579 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8580 mo->mp_pgno = my->mc_next_pgno++;
8581 my->mc_wlen[toggle] += my->mc_env->me_psize;
8583 /* Update parent if there is one */
8584 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8585 SETPGNO(ni, mo->mp_pgno);
8586 mdb_cursor_pop(&mc);
8588 /* Otherwise we're done */
8598 /** Copy environment with compaction. */
8600 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8605 MDB_txn *txn = NULL;
8610 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8611 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8612 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8613 if (my.mc_wbuf[0] == NULL)
8616 pthread_mutex_init(&my.mc_mutex, NULL);
8617 pthread_cond_init(&my.mc_cond, NULL);
8618 #ifdef HAVE_MEMALIGN
8619 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8620 if (my.mc_wbuf[0] == NULL)
8623 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8628 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8629 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8634 my.mc_next_pgno = 2;
8640 THREAD_CREATE(thr, mdb_env_copythr, &my);
8642 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8646 mp = (MDB_page *)my.mc_wbuf[0];
8647 memset(mp, 0, 2*env->me_psize);
8649 mp->mp_flags = P_META;
8650 mm = (MDB_meta *)METADATA(mp);
8651 mdb_env_init_meta0(env, mm);
8652 mm->mm_address = env->me_metas[0]->mm_address;
8654 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8656 mp->mp_flags = P_META;
8657 *(MDB_meta *)METADATA(mp) = *mm;
8658 mm = (MDB_meta *)METADATA(mp);
8660 /* Count the number of free pages, subtract from lastpg to find
8661 * number of active pages
8664 MDB_ID freecount = 0;
8667 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8668 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8669 freecount += *(MDB_ID *)data.mv_data;
8670 freecount += txn->mt_dbs[0].md_branch_pages +
8671 txn->mt_dbs[0].md_leaf_pages +
8672 txn->mt_dbs[0].md_overflow_pages;
8674 /* Set metapage 1 */
8675 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8676 mm->mm_dbs[1] = txn->mt_dbs[1];
8677 if (mm->mm_last_pg > 1) {
8678 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8681 mm->mm_dbs[1].md_root = P_INVALID;
8684 my.mc_wlen[0] = env->me_psize * 2;
8686 pthread_mutex_lock(&my.mc_mutex);
8688 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8689 pthread_mutex_unlock(&my.mc_mutex);
8690 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8691 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8692 rc = mdb_env_cthr_toggle(&my, 1);
8693 mdb_env_cthr_toggle(&my, -1);
8694 pthread_mutex_lock(&my.mc_mutex);
8696 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8697 pthread_mutex_unlock(&my.mc_mutex);
8702 CloseHandle(my.mc_cond);
8703 CloseHandle(my.mc_mutex);
8704 _aligned_free(my.mc_wbuf[0]);
8706 pthread_cond_destroy(&my.mc_cond);
8707 pthread_mutex_destroy(&my.mc_mutex);
8708 free(my.mc_wbuf[0]);
8713 /** Copy environment as-is. */
8715 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8717 MDB_txn *txn = NULL;
8723 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8727 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8730 /* Do the lock/unlock of the reader mutex before starting the
8731 * write txn. Otherwise other read txns could block writers.
8733 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8738 /* We must start the actual read txn after blocking writers */
8739 mdb_txn_reset0(txn, "reset-stage1");
8741 /* Temporarily block writers until we snapshot the meta pages */
8744 rc = mdb_txn_renew0(txn);
8746 UNLOCK_MUTEX_W(env);
8751 wsize = env->me_psize * 2;
8755 DO_WRITE(rc, fd, ptr, w2, len);
8759 } else if (len > 0) {
8765 /* Non-blocking or async handles are not supported */
8771 UNLOCK_MUTEX_W(env);
8776 w2 = txn->mt_next_pgno * env->me_psize;
8779 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8786 if (wsize > MAX_WRITE)
8790 DO_WRITE(rc, fd, ptr, w2, len);
8794 } else if (len > 0) {
8811 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8813 if (flags & MDB_CP_COMPACT)
8814 return mdb_env_copyfd1(env, fd);
8816 return mdb_env_copyfd0(env, fd);
8820 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8822 return mdb_env_copyfd2(env, fd, 0);
8826 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8830 HANDLE newfd = INVALID_HANDLE_VALUE;
8832 if (env->me_flags & MDB_NOSUBDIR) {
8833 lpath = (char *)path;
8836 len += sizeof(DATANAME);
8837 lpath = malloc(len);
8840 sprintf(lpath, "%s" DATANAME, path);
8843 /* The destination path must exist, but the destination file must not.
8844 * We don't want the OS to cache the writes, since the source data is
8845 * already in the OS cache.
8848 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8849 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8851 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8853 if (newfd == INVALID_HANDLE_VALUE) {
8858 if (env->me_psize >= env->me_os_psize) {
8860 /* Set O_DIRECT if the file system supports it */
8861 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8862 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8864 #ifdef F_NOCACHE /* __APPLE__ */
8865 rc = fcntl(newfd, F_NOCACHE, 1);
8873 rc = mdb_env_copyfd2(env, newfd, flags);
8876 if (!(env->me_flags & MDB_NOSUBDIR))
8878 if (newfd != INVALID_HANDLE_VALUE)
8879 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8886 mdb_env_copy(MDB_env *env, const char *path)
8888 return mdb_env_copy2(env, path, 0);
8892 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8894 if ((flag & CHANGEABLE) != flag)
8897 env->me_flags |= flag;
8899 env->me_flags &= ~flag;
8904 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8909 *arg = env->me_flags;
8914 mdb_env_set_userctx(MDB_env *env, void *ctx)
8918 env->me_userctx = ctx;
8923 mdb_env_get_userctx(MDB_env *env)
8925 return env ? env->me_userctx : NULL;
8929 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8934 env->me_assert_func = func;
8940 mdb_env_get_path(MDB_env *env, const char **arg)
8945 *arg = env->me_path;
8950 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8959 /** Common code for #mdb_stat() and #mdb_env_stat().
8960 * @param[in] env the environment to operate in.
8961 * @param[in] db the #MDB_db record containing the stats to return.
8962 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8963 * @return 0, this function always succeeds.
8966 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8968 arg->ms_psize = env->me_psize;
8969 arg->ms_depth = db->md_depth;
8970 arg->ms_branch_pages = db->md_branch_pages;
8971 arg->ms_leaf_pages = db->md_leaf_pages;
8972 arg->ms_overflow_pages = db->md_overflow_pages;
8973 arg->ms_entries = db->md_entries;
8979 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8983 if (env == NULL || arg == NULL)
8986 toggle = mdb_env_pick_meta(env);
8988 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8992 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8996 if (env == NULL || arg == NULL)
8999 toggle = mdb_env_pick_meta(env);
9000 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9001 arg->me_mapsize = env->me_mapsize;
9002 arg->me_maxreaders = env->me_maxreaders;
9004 /* me_numreaders may be zero if this process never used any readers. Use
9005 * the shared numreader count if it exists.
9007 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
9009 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9010 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9014 /** Set the default comparison functions for a database.
9015 * Called immediately after a database is opened to set the defaults.
9016 * The user can then override them with #mdb_set_compare() or
9017 * #mdb_set_dupsort().
9018 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9019 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9022 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9024 uint16_t f = txn->mt_dbs[dbi].md_flags;
9026 txn->mt_dbxs[dbi].md_cmp =
9027 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9028 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9030 txn->mt_dbxs[dbi].md_dcmp =
9031 !(f & MDB_DUPSORT) ? 0 :
9032 ((f & MDB_INTEGERDUP)
9033 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9034 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9037 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9043 int rc, dbflag, exact;
9044 unsigned int unused = 0, seq;
9047 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9048 mdb_default_cmp(txn, FREE_DBI);
9051 if ((flags & VALID_FLAGS) != flags)
9053 if (txn->mt_flags & MDB_TXN_ERROR)
9059 if (flags & PERSISTENT_FLAGS) {
9060 uint16_t f2 = flags & PERSISTENT_FLAGS;
9061 /* make sure flag changes get committed */
9062 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9063 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9064 txn->mt_flags |= MDB_TXN_DIRTY;
9067 mdb_default_cmp(txn, MAIN_DBI);
9071 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9072 mdb_default_cmp(txn, MAIN_DBI);
9075 /* Is the DB already open? */
9077 for (i=2; i<txn->mt_numdbs; i++) {
9078 if (!txn->mt_dbxs[i].md_name.mv_size) {
9079 /* Remember this free slot */
9080 if (!unused) unused = i;
9083 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9084 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9090 /* If no free slot and max hit, fail */
9091 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9092 return MDB_DBS_FULL;
9094 /* Cannot mix named databases with some mainDB flags */
9095 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9096 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9098 /* Find the DB info */
9099 dbflag = DB_NEW|DB_VALID;
9102 key.mv_data = (void *)name;
9103 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9104 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9105 if (rc == MDB_SUCCESS) {
9106 /* make sure this is actually a DB */
9107 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9108 if (!(node->mn_flags & F_SUBDATA))
9109 return MDB_INCOMPATIBLE;
9110 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9111 /* Create if requested */
9112 data.mv_size = sizeof(MDB_db);
9113 data.mv_data = &dummy;
9114 memset(&dummy, 0, sizeof(dummy));
9115 dummy.md_root = P_INVALID;
9116 dummy.md_flags = flags & PERSISTENT_FLAGS;
9117 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9121 /* OK, got info, add to table */
9122 if (rc == MDB_SUCCESS) {
9123 unsigned int slot = unused ? unused : txn->mt_numdbs;
9124 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9125 txn->mt_dbxs[slot].md_name.mv_size = len;
9126 txn->mt_dbxs[slot].md_rel = NULL;
9127 txn->mt_dbflags[slot] = dbflag;
9128 /* txn-> and env-> are the same in read txns, use
9129 * tmp variable to avoid undefined assignment
9131 seq = ++txn->mt_env->me_dbiseqs[slot];
9132 txn->mt_dbiseqs[slot] = seq;
9134 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9136 mdb_default_cmp(txn, slot);
9145 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9147 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9150 if (txn->mt_flags & MDB_TXN_ERROR)
9153 if (txn->mt_dbflags[dbi] & DB_STALE) {
9156 /* Stale, must read the DB's root. cursor_init does it for us. */
9157 mdb_cursor_init(&mc, txn, dbi, &mx);
9159 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9162 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9165 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9167 ptr = env->me_dbxs[dbi].md_name.mv_data;
9168 /* If there was no name, this was already closed */
9170 env->me_dbxs[dbi].md_name.mv_data = NULL;
9171 env->me_dbxs[dbi].md_name.mv_size = 0;
9172 env->me_dbflags[dbi] = 0;
9173 env->me_dbiseqs[dbi]++;
9178 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9180 /* We could return the flags for the FREE_DBI too but what's the point? */
9181 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9183 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9187 /** Add all the DB's pages to the free list.
9188 * @param[in] mc Cursor on the DB to free.
9189 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9190 * @return 0 on success, non-zero on failure.
9193 mdb_drop0(MDB_cursor *mc, int subs)
9197 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9198 if (rc == MDB_SUCCESS) {
9199 MDB_txn *txn = mc->mc_txn;
9204 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9205 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9208 mdb_cursor_copy(mc, &mx);
9209 while (mc->mc_snum > 0) {
9210 MDB_page *mp = mc->mc_pg[mc->mc_top];
9211 unsigned n = NUMKEYS(mp);
9213 for (i=0; i<n; i++) {
9214 ni = NODEPTR(mp, i);
9215 if (ni->mn_flags & F_BIGDATA) {
9218 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9219 rc = mdb_page_get(txn, pg, &omp, NULL);
9222 mdb_cassert(mc, IS_OVERFLOW(omp));
9223 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9227 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9228 mdb_xcursor_init1(mc, ni);
9229 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9235 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9237 for (i=0; i<n; i++) {
9239 ni = NODEPTR(mp, i);
9242 mdb_midl_xappend(txn->mt_free_pgs, pg);
9247 mc->mc_ki[mc->mc_top] = i;
9248 rc = mdb_cursor_sibling(mc, 1);
9250 if (rc != MDB_NOTFOUND)
9252 /* no more siblings, go back to beginning
9253 * of previous level.
9257 for (i=1; i<mc->mc_snum; i++) {
9259 mc->mc_pg[i] = mx.mc_pg[i];
9264 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9267 txn->mt_flags |= MDB_TXN_ERROR;
9268 } else if (rc == MDB_NOTFOUND) {
9274 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9276 MDB_cursor *mc, *m2;
9279 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9282 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9285 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9288 rc = mdb_cursor_open(txn, dbi, &mc);
9292 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9293 /* Invalidate the dropped DB's cursors */
9294 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9295 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9299 /* Can't delete the main DB */
9300 if (del && dbi > MAIN_DBI) {
9301 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9303 txn->mt_dbflags[dbi] = DB_STALE;
9304 mdb_dbi_close(txn->mt_env, dbi);
9306 txn->mt_flags |= MDB_TXN_ERROR;
9309 /* reset the DB record, mark it dirty */
9310 txn->mt_dbflags[dbi] |= DB_DIRTY;
9311 txn->mt_dbs[dbi].md_depth = 0;
9312 txn->mt_dbs[dbi].md_branch_pages = 0;
9313 txn->mt_dbs[dbi].md_leaf_pages = 0;
9314 txn->mt_dbs[dbi].md_overflow_pages = 0;
9315 txn->mt_dbs[dbi].md_entries = 0;
9316 txn->mt_dbs[dbi].md_root = P_INVALID;
9318 txn->mt_flags |= MDB_TXN_DIRTY;
9321 mdb_cursor_close(mc);
9325 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9327 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9330 txn->mt_dbxs[dbi].md_cmp = cmp;
9334 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9336 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9339 txn->mt_dbxs[dbi].md_dcmp = cmp;
9343 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9345 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9348 txn->mt_dbxs[dbi].md_rel = rel;
9352 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9354 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9357 txn->mt_dbxs[dbi].md_relctx = ctx;
9362 mdb_env_get_maxkeysize(MDB_env *env)
9364 return ENV_MAXKEY(env);
9368 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9370 unsigned int i, rdrs;
9373 int rc = 0, first = 1;
9377 if (!env->me_txns) {
9378 return func("(no reader locks)\n", ctx);
9380 rdrs = env->me_txns->mti_numreaders;
9381 mr = env->me_txns->mti_readers;
9382 for (i=0; i<rdrs; i++) {
9384 txnid_t txnid = mr[i].mr_txnid;
9385 sprintf(buf, txnid == (txnid_t)-1 ?
9386 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9387 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9390 rc = func(" pid thread txnid\n", ctx);
9394 rc = func(buf, ctx);
9400 rc = func("(no active readers)\n", ctx);
9405 /** Insert pid into list if not already present.
9406 * return -1 if already present.
9409 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9411 /* binary search of pid in list */
9413 unsigned cursor = 1;
9415 unsigned n = ids[0];
9418 unsigned pivot = n >> 1;
9419 cursor = base + pivot + 1;
9420 val = pid - ids[cursor];
9425 } else if ( val > 0 ) {
9430 /* found, so it's a duplicate */
9439 for (n = ids[0]; n > cursor; n--)
9446 mdb_reader_check(MDB_env *env, int *dead)
9448 unsigned int i, j, rdrs;
9450 MDB_PID_T *pids, pid;
9459 rdrs = env->me_txns->mti_numreaders;
9460 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9464 mr = env->me_txns->mti_readers;
9465 for (i=0; i<rdrs; i++) {
9466 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9468 if (mdb_pid_insert(pids, pid) == 0) {
9469 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9471 /* Recheck, a new process may have reused pid */
9472 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9473 for (j=i; j<rdrs; j++)
9474 if (mr[j].mr_pid == pid) {
9475 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9476 (unsigned) pid, mr[j].mr_txnid));
9481 UNLOCK_MUTEX_R(env);