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 lock 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 */
1146 void *me_userctx; /**< User-settable context */
1147 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1150 /** Nested transaction */
1151 typedef struct MDB_ntxn {
1152 MDB_txn mnt_txn; /**< the transaction */
1153 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1156 /** max number of pages to commit in one writev() call */
1157 #define MDB_COMMIT_PAGES 64
1158 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1159 #undef MDB_COMMIT_PAGES
1160 #define MDB_COMMIT_PAGES IOV_MAX
1163 /** max bytes to write in one call */
1164 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1166 /** Check \b txn and \b dbi arguments to a function */
1167 #define TXN_DBI_EXIST(txn, dbi) \
1168 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1170 /** Check for misused \b dbi handles */
1171 #define TXN_DBI_CHANGED(txn, dbi) \
1172 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1174 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1175 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1176 static int mdb_page_touch(MDB_cursor *mc);
1178 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1179 static int mdb_page_search_root(MDB_cursor *mc,
1180 MDB_val *key, int modify);
1181 #define MDB_PS_MODIFY 1
1182 #define MDB_PS_ROOTONLY 2
1183 #define MDB_PS_FIRST 4
1184 #define MDB_PS_LAST 8
1185 static int mdb_page_search(MDB_cursor *mc,
1186 MDB_val *key, int flags);
1187 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1189 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1190 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1191 pgno_t newpgno, unsigned int nflags);
1193 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1194 static int mdb_env_pick_meta(const MDB_env *env);
1195 static int mdb_env_write_meta(MDB_txn *txn);
1196 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1197 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1199 static void mdb_env_close0(MDB_env *env, int excl);
1201 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1202 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1203 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1204 static void mdb_node_del(MDB_cursor *mc, int ksize);
1205 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1206 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1207 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1208 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1209 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1211 static int mdb_rebalance(MDB_cursor *mc);
1212 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1214 static void mdb_cursor_pop(MDB_cursor *mc);
1215 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1217 static int mdb_cursor_del0(MDB_cursor *mc);
1218 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1219 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1220 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1221 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1222 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1224 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1225 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1227 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1228 static void mdb_xcursor_init0(MDB_cursor *mc);
1229 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1231 static int mdb_drop0(MDB_cursor *mc, int subs);
1232 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1235 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1239 static SECURITY_DESCRIPTOR mdb_null_sd;
1240 static SECURITY_ATTRIBUTES mdb_all_sa;
1241 static int mdb_sec_inited;
1244 /** Return the library version info. */
1246 mdb_version(int *major, int *minor, int *patch)
1248 if (major) *major = MDB_VERSION_MAJOR;
1249 if (minor) *minor = MDB_VERSION_MINOR;
1250 if (patch) *patch = MDB_VERSION_PATCH;
1251 return MDB_VERSION_STRING;
1254 /** Table of descriptions for LMDB @ref errors */
1255 static char *const mdb_errstr[] = {
1256 "MDB_KEYEXIST: Key/data pair already exists",
1257 "MDB_NOTFOUND: No matching key/data pair found",
1258 "MDB_PAGE_NOTFOUND: Requested page not found",
1259 "MDB_CORRUPTED: Located page was wrong type",
1260 "MDB_PANIC: Update of meta page failed",
1261 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1262 "MDB_INVALID: File is not an LMDB file",
1263 "MDB_MAP_FULL: Environment mapsize limit reached",
1264 "MDB_DBS_FULL: Environment maxdbs limit reached",
1265 "MDB_READERS_FULL: Environment maxreaders limit reached",
1266 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1267 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1268 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1269 "MDB_PAGE_FULL: Internal error - page has no more space",
1270 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1271 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1272 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1273 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1274 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1275 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1279 mdb_strerror(int err)
1282 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1283 * This works as long as no function between the call to mdb_strerror
1284 * and the actual use of the message uses more than 4K of stack.
1287 char buf[1024], *ptr = buf;
1291 return ("Successful return: 0");
1293 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1294 i = err - MDB_KEYEXIST;
1295 return mdb_errstr[i];
1299 /* These are the C-runtime error codes we use. The comment indicates
1300 * their numeric value, and the Win32 error they would correspond to
1301 * if the error actually came from a Win32 API. A major mess, we should
1302 * have used LMDB-specific error codes for everything.
1305 case ENOENT: /* 2, FILE_NOT_FOUND */
1306 case EIO: /* 5, ACCESS_DENIED */
1307 case ENOMEM: /* 12, INVALID_ACCESS */
1308 case EACCES: /* 13, INVALID_DATA */
1309 case EBUSY: /* 16, CURRENT_DIRECTORY */
1310 case EINVAL: /* 22, BAD_COMMAND */
1311 case ENOSPC: /* 28, OUT_OF_PAPER */
1312 return strerror(err);
1317 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1318 FORMAT_MESSAGE_IGNORE_INSERTS,
1319 NULL, err, 0, ptr, sizeof(buf), pad);
1322 return strerror(err);
1326 /** assert(3) variant in cursor context */
1327 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1328 /** assert(3) variant in transaction context */
1329 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1330 /** assert(3) variant in environment context */
1331 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1334 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1335 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1338 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1339 const char *func, const char *file, int line)
1342 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1343 file, line, expr_txt, func);
1344 if (env->me_assert_func)
1345 env->me_assert_func(env, buf);
1346 fprintf(stderr, "%s\n", buf);
1350 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1354 /** Return the page number of \b mp which may be sub-page, for debug output */
1356 mdb_dbg_pgno(MDB_page *mp)
1359 COPY_PGNO(ret, mp->mp_pgno);
1363 /** Display a key in hexadecimal and return the address of the result.
1364 * @param[in] key the key to display
1365 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1366 * @return The key in hexadecimal form.
1369 mdb_dkey(MDB_val *key, char *buf)
1372 unsigned char *c = key->mv_data;
1378 if (key->mv_size > DKBUF_MAXKEYSIZE)
1379 return "MDB_MAXKEYSIZE";
1380 /* may want to make this a dynamic check: if the key is mostly
1381 * printable characters, print it as-is instead of converting to hex.
1385 for (i=0; i<key->mv_size; i++)
1386 ptr += sprintf(ptr, "%02x", *c++);
1388 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1394 mdb_leafnode_type(MDB_node *n)
1396 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1397 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1398 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1401 /** Display all the keys in the page. */
1403 mdb_page_list(MDB_page *mp)
1405 pgno_t pgno = mdb_dbg_pgno(mp);
1406 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1408 unsigned int i, nkeys, nsize, total = 0;
1412 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1413 case P_BRANCH: type = "Branch page"; break;
1414 case P_LEAF: type = "Leaf page"; break;
1415 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1416 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1417 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1419 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1420 pgno, mp->mp_pages, state);
1423 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1424 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1427 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1431 nkeys = NUMKEYS(mp);
1432 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1434 for (i=0; i<nkeys; i++) {
1435 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1436 key.mv_size = nsize = mp->mp_pad;
1437 key.mv_data = LEAF2KEY(mp, i, nsize);
1439 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1442 node = NODEPTR(mp, i);
1443 key.mv_size = node->mn_ksize;
1444 key.mv_data = node->mn_data;
1445 nsize = NODESIZE + key.mv_size;
1446 if (IS_BRANCH(mp)) {
1447 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1451 if (F_ISSET(node->mn_flags, F_BIGDATA))
1452 nsize += sizeof(pgno_t);
1454 nsize += NODEDSZ(node);
1456 nsize += sizeof(indx_t);
1457 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1458 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1460 total = EVEN(total);
1462 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1463 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1467 mdb_cursor_chk(MDB_cursor *mc)
1473 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1474 for (i=0; i<mc->mc_top; i++) {
1476 node = NODEPTR(mp, mc->mc_ki[i]);
1477 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1480 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1486 /** Count all the pages in each DB and in the freelist
1487 * and make sure it matches the actual number of pages
1489 * All named DBs must be open for a correct count.
1491 static void mdb_audit(MDB_txn *txn)
1495 MDB_ID freecount, count;
1500 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1501 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1502 freecount += *(MDB_ID *)data.mv_data;
1503 mdb_tassert(txn, rc == MDB_NOTFOUND);
1506 for (i = 0; i<txn->mt_numdbs; i++) {
1508 if (!(txn->mt_dbflags[i] & DB_VALID))
1510 mdb_cursor_init(&mc, txn, i, &mx);
1511 if (txn->mt_dbs[i].md_root == P_INVALID)
1513 count += txn->mt_dbs[i].md_branch_pages +
1514 txn->mt_dbs[i].md_leaf_pages +
1515 txn->mt_dbs[i].md_overflow_pages;
1516 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1517 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1518 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1521 mp = mc.mc_pg[mc.mc_top];
1522 for (j=0; j<NUMKEYS(mp); j++) {
1523 MDB_node *leaf = NODEPTR(mp, j);
1524 if (leaf->mn_flags & F_SUBDATA) {
1526 memcpy(&db, NODEDATA(leaf), sizeof(db));
1527 count += db.md_branch_pages + db.md_leaf_pages +
1528 db.md_overflow_pages;
1532 mdb_tassert(txn, rc == MDB_NOTFOUND);
1535 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1536 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1537 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1543 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1545 return txn->mt_dbxs[dbi].md_cmp(a, b);
1549 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1551 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1554 /** Allocate memory for a page.
1555 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1558 mdb_page_malloc(MDB_txn *txn, unsigned num)
1560 MDB_env *env = txn->mt_env;
1561 MDB_page *ret = env->me_dpages;
1562 size_t psize = env->me_psize, sz = psize, off;
1563 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1564 * For a single page alloc, we init everything after the page header.
1565 * For multi-page, we init the final page; if the caller needed that
1566 * many pages they will be filling in at least up to the last page.
1570 VGMEMP_ALLOC(env, ret, sz);
1571 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1572 env->me_dpages = ret->mp_next;
1575 psize -= off = PAGEHDRSZ;
1580 if ((ret = malloc(sz)) != NULL) {
1581 VGMEMP_ALLOC(env, ret, sz);
1582 if (!(env->me_flags & MDB_NOMEMINIT)) {
1583 memset((char *)ret + off, 0, psize);
1587 txn->mt_flags |= MDB_TXN_ERROR;
1591 /** Free a single page.
1592 * Saves single pages to a list, for future reuse.
1593 * (This is not used for multi-page overflow pages.)
1596 mdb_page_free(MDB_env *env, MDB_page *mp)
1598 mp->mp_next = env->me_dpages;
1599 VGMEMP_FREE(env, mp);
1600 env->me_dpages = mp;
1603 /** Free a dirty page */
1605 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1607 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1608 mdb_page_free(env, dp);
1610 /* large pages just get freed directly */
1611 VGMEMP_FREE(env, dp);
1616 /** Return all dirty pages to dpage list */
1618 mdb_dlist_free(MDB_txn *txn)
1620 MDB_env *env = txn->mt_env;
1621 MDB_ID2L dl = txn->mt_u.dirty_list;
1622 unsigned i, n = dl[0].mid;
1624 for (i = 1; i <= n; i++) {
1625 mdb_dpage_free(env, dl[i].mptr);
1630 /** Loosen or free a single page.
1631 * Saves single pages to a list for future reuse
1632 * in this same txn. It has been pulled from the freeDB
1633 * and already resides on the dirty list, but has been
1634 * deleted. Use these pages first before pulling again
1637 * If the page wasn't dirtied in this txn, just add it
1638 * to this txn's free list.
1641 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1644 pgno_t pgno = mp->mp_pgno;
1645 MDB_txn *txn = mc->mc_txn;
1647 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1648 if (txn->mt_parent) {
1649 MDB_ID2 *dl = txn->mt_u.dirty_list;
1650 /* If txn has a parent, make sure the page is in our
1654 unsigned x = mdb_mid2l_search(dl, pgno);
1655 if (x <= dl[0].mid && dl[x].mid == pgno) {
1656 if (mp != dl[x].mptr) { /* bad cursor? */
1657 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1658 txn->mt_flags |= MDB_TXN_ERROR;
1659 return MDB_CORRUPTED;
1666 /* no parent txn, so it's just ours */
1671 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1673 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1674 txn->mt_loose_pgs = mp;
1675 txn->mt_loose_count++;
1676 mp->mp_flags |= P_LOOSE;
1678 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1686 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1687 * @param[in] mc A cursor handle for the current operation.
1688 * @param[in] pflags Flags of the pages to update:
1689 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1690 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1691 * @return 0 on success, non-zero on failure.
1694 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1696 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1697 MDB_txn *txn = mc->mc_txn;
1703 int rc = MDB_SUCCESS, level;
1705 /* Mark pages seen by cursors */
1706 if (mc->mc_flags & C_UNTRACK)
1707 mc = NULL; /* will find mc in mt_cursors */
1708 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1709 for (; mc; mc=mc->mc_next) {
1710 if (!(mc->mc_flags & C_INITIALIZED))
1712 for (m3 = mc;; m3 = &mx->mx_cursor) {
1714 for (j=0; j<m3->mc_snum; j++) {
1716 if ((mp->mp_flags & Mask) == pflags)
1717 mp->mp_flags ^= P_KEEP;
1719 mx = m3->mc_xcursor;
1720 /* Proceed to mx if it is at a sub-database */
1721 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1723 if (! (mp && (mp->mp_flags & P_LEAF)))
1725 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1726 if (!(leaf->mn_flags & F_SUBDATA))
1735 /* Mark dirty root pages */
1736 for (i=0; i<txn->mt_numdbs; i++) {
1737 if (txn->mt_dbflags[i] & DB_DIRTY) {
1738 pgno_t pgno = txn->mt_dbs[i].md_root;
1739 if (pgno == P_INVALID)
1741 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1743 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1744 dp->mp_flags ^= P_KEEP;
1752 static int mdb_page_flush(MDB_txn *txn, int keep);
1754 /** Spill pages from the dirty list back to disk.
1755 * This is intended to prevent running into #MDB_TXN_FULL situations,
1756 * but note that they may still occur in a few cases:
1757 * 1) our estimate of the txn size could be too small. Currently this
1758 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1759 * 2) child txns may run out of space if their parents dirtied a
1760 * lot of pages and never spilled them. TODO: we probably should do
1761 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1762 * the parent's dirty_room is below a given threshold.
1764 * Otherwise, if not using nested txns, it is expected that apps will
1765 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1766 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1767 * If the txn never references them again, they can be left alone.
1768 * If the txn only reads them, they can be used without any fuss.
1769 * If the txn writes them again, they can be dirtied immediately without
1770 * going thru all of the work of #mdb_page_touch(). Such references are
1771 * handled by #mdb_page_unspill().
1773 * Also note, we never spill DB root pages, nor pages of active cursors,
1774 * because we'll need these back again soon anyway. And in nested txns,
1775 * we can't spill a page in a child txn if it was already spilled in a
1776 * parent txn. That would alter the parent txns' data even though
1777 * the child hasn't committed yet, and we'd have no way to undo it if
1778 * the child aborted.
1780 * @param[in] m0 cursor A cursor handle identifying the transaction and
1781 * database for which we are checking space.
1782 * @param[in] key For a put operation, the key being stored.
1783 * @param[in] data For a put operation, the data being stored.
1784 * @return 0 on success, non-zero on failure.
1787 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1789 MDB_txn *txn = m0->mc_txn;
1791 MDB_ID2L dl = txn->mt_u.dirty_list;
1792 unsigned int i, j, need;
1795 if (m0->mc_flags & C_SUB)
1798 /* Estimate how much space this op will take */
1799 i = m0->mc_db->md_depth;
1800 /* Named DBs also dirty the main DB */
1801 if (m0->mc_dbi > MAIN_DBI)
1802 i += txn->mt_dbs[MAIN_DBI].md_depth;
1803 /* For puts, roughly factor in the key+data size */
1805 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1806 i += i; /* double it for good measure */
1809 if (txn->mt_dirty_room > i)
1812 if (!txn->mt_spill_pgs) {
1813 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1814 if (!txn->mt_spill_pgs)
1817 /* purge deleted slots */
1818 MDB_IDL sl = txn->mt_spill_pgs;
1819 unsigned int num = sl[0];
1821 for (i=1; i<=num; i++) {
1828 /* Preserve pages which may soon be dirtied again */
1829 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1832 /* Less aggressive spill - we originally spilled the entire dirty list,
1833 * with a few exceptions for cursor pages and DB root pages. But this
1834 * turns out to be a lot of wasted effort because in a large txn many
1835 * of those pages will need to be used again. So now we spill only 1/8th
1836 * of the dirty pages. Testing revealed this to be a good tradeoff,
1837 * better than 1/2, 1/4, or 1/10.
1839 if (need < MDB_IDL_UM_MAX / 8)
1840 need = MDB_IDL_UM_MAX / 8;
1842 /* Save the page IDs of all the pages we're flushing */
1843 /* flush from the tail forward, this saves a lot of shifting later on. */
1844 for (i=dl[0].mid; i && need; i--) {
1845 MDB_ID pn = dl[i].mid << 1;
1847 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1849 /* Can't spill twice, make sure it's not already in a parent's
1852 if (txn->mt_parent) {
1854 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1855 if (tx2->mt_spill_pgs) {
1856 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1857 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1858 dp->mp_flags |= P_KEEP;
1866 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1870 mdb_midl_sort(txn->mt_spill_pgs);
1872 /* Flush the spilled part of dirty list */
1873 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1876 /* Reset any dirty pages we kept that page_flush didn't see */
1877 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1880 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1884 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1886 mdb_find_oldest(MDB_txn *txn)
1889 txnid_t mr, oldest = txn->mt_txnid - 1;
1890 if (txn->mt_env->me_txns) {
1891 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1892 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1903 /** Add a page to the txn's dirty list */
1905 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1908 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1910 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1911 insert = mdb_mid2l_append;
1913 insert = mdb_mid2l_insert;
1915 mid.mid = mp->mp_pgno;
1917 rc = insert(txn->mt_u.dirty_list, &mid);
1918 mdb_tassert(txn, rc == 0);
1919 txn->mt_dirty_room--;
1922 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1923 * me_pghead and mt_next_pgno.
1925 * If there are free pages available from older transactions, they
1926 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1927 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1928 * and move me_pglast to say which records were consumed. Only this
1929 * function can create me_pghead and move me_pglast/mt_next_pgno.
1930 * @param[in] mc cursor A cursor handle identifying the transaction and
1931 * database for which we are allocating.
1932 * @param[in] num the number of pages to allocate.
1933 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1934 * will always be satisfied by a single contiguous chunk of memory.
1935 * @return 0 on success, non-zero on failure.
1938 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1940 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1941 /* Get at most <Max_retries> more freeDB records once me_pghead
1942 * has enough pages. If not enough, use new pages from the map.
1943 * If <Paranoid> and mc is updating the freeDB, only get new
1944 * records if me_pghead is empty. Then the freelist cannot play
1945 * catch-up with itself by growing while trying to save it.
1947 enum { Paranoid = 1, Max_retries = 500 };
1949 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1951 int rc, retry = num * 60;
1952 MDB_txn *txn = mc->mc_txn;
1953 MDB_env *env = txn->mt_env;
1954 pgno_t pgno, *mop = env->me_pghead;
1955 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1957 txnid_t oldest = 0, last;
1962 /* If there are any loose pages, just use them */
1963 if (num == 1 && txn->mt_loose_pgs) {
1964 np = txn->mt_loose_pgs;
1965 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1966 txn->mt_loose_count--;
1967 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1975 /* If our dirty list is already full, we can't do anything */
1976 if (txn->mt_dirty_room == 0) {
1981 for (op = MDB_FIRST;; op = MDB_NEXT) {
1986 /* Seek a big enough contiguous page range. Prefer
1987 * pages at the tail, just truncating the list.
1993 if (mop[i-n2] == pgno+n2)
2000 if (op == MDB_FIRST) { /* 1st iteration */
2001 /* Prepare to fetch more and coalesce */
2002 last = env->me_pglast;
2003 oldest = env->me_pgoldest;
2004 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2007 key.mv_data = &last; /* will look up last+1 */
2008 key.mv_size = sizeof(last);
2010 if (Paranoid && mc->mc_dbi == FREE_DBI)
2013 if (Paranoid && retry < 0 && mop_len)
2017 /* Do not fetch more if the record will be too recent */
2018 if (oldest <= last) {
2020 oldest = mdb_find_oldest(txn);
2021 env->me_pgoldest = oldest;
2027 rc = mdb_cursor_get(&m2, &key, NULL, op);
2029 if (rc == MDB_NOTFOUND)
2033 last = *(txnid_t*)key.mv_data;
2034 if (oldest <= last) {
2036 oldest = mdb_find_oldest(txn);
2037 env->me_pgoldest = oldest;
2043 np = m2.mc_pg[m2.mc_top];
2044 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2045 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2048 idl = (MDB_ID *) data.mv_data;
2051 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2056 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2058 mop = env->me_pghead;
2060 env->me_pglast = last;
2062 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2063 last, txn->mt_dbs[FREE_DBI].md_root, i));
2065 DPRINTF(("IDL %"Z"u", idl[j]));
2067 /* Merge in descending sorted order */
2068 mdb_midl_xmerge(mop, idl);
2072 /* Use new pages from the map when nothing suitable in the freeDB */
2074 pgno = txn->mt_next_pgno;
2075 if (pgno + num >= env->me_maxpg) {
2076 DPUTS("DB size maxed out");
2082 if (env->me_flags & MDB_WRITEMAP) {
2083 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2085 if (!(np = mdb_page_malloc(txn, num))) {
2091 mop[0] = mop_len -= num;
2092 /* Move any stragglers down */
2093 for (j = i-num; j < mop_len; )
2094 mop[++j] = mop[++i];
2096 txn->mt_next_pgno = pgno + num;
2099 mdb_page_dirty(txn, np);
2105 txn->mt_flags |= MDB_TXN_ERROR;
2109 /** Copy the used portions of a non-overflow page.
2110 * @param[in] dst page to copy into
2111 * @param[in] src page to copy from
2112 * @param[in] psize size of a page
2115 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2117 enum { Align = sizeof(pgno_t) };
2118 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2120 /* If page isn't full, just copy the used portion. Adjust
2121 * alignment so memcpy may copy words instead of bytes.
2123 if ((unused &= -Align) && !IS_LEAF2(src)) {
2124 upper = (upper + PAGEBASE) & -Align;
2125 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2126 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2129 memcpy(dst, src, psize - unused);
2133 /** Pull a page off the txn's spill list, if present.
2134 * If a page being referenced was spilled to disk in this txn, bring
2135 * it back and make it dirty/writable again.
2136 * @param[in] txn the transaction handle.
2137 * @param[in] mp the page being referenced. It must not be dirty.
2138 * @param[out] ret the writable page, if any. ret is unchanged if
2139 * mp wasn't spilled.
2142 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2144 MDB_env *env = txn->mt_env;
2147 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2149 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2150 if (!tx2->mt_spill_pgs)
2152 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2153 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2156 if (txn->mt_dirty_room == 0)
2157 return MDB_TXN_FULL;
2158 if (IS_OVERFLOW(mp))
2162 if (env->me_flags & MDB_WRITEMAP) {
2165 np = mdb_page_malloc(txn, num);
2169 memcpy(np, mp, num * env->me_psize);
2171 mdb_page_copy(np, mp, env->me_psize);
2174 /* If in current txn, this page is no longer spilled.
2175 * If it happens to be the last page, truncate the spill list.
2176 * Otherwise mark it as deleted by setting the LSB.
2178 if (x == txn->mt_spill_pgs[0])
2179 txn->mt_spill_pgs[0]--;
2181 txn->mt_spill_pgs[x] |= 1;
2182 } /* otherwise, if belonging to a parent txn, the
2183 * page remains spilled until child commits
2186 mdb_page_dirty(txn, np);
2187 np->mp_flags |= P_DIRTY;
2195 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2196 * @param[in] mc cursor pointing to the page to be touched
2197 * @return 0 on success, non-zero on failure.
2200 mdb_page_touch(MDB_cursor *mc)
2202 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2203 MDB_txn *txn = mc->mc_txn;
2204 MDB_cursor *m2, *m3;
2208 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2209 if (txn->mt_flags & MDB_TXN_SPILLS) {
2211 rc = mdb_page_unspill(txn, mp, &np);
2217 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2218 (rc = mdb_page_alloc(mc, 1, &np)))
2221 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2222 mp->mp_pgno, pgno));
2223 mdb_cassert(mc, mp->mp_pgno != pgno);
2224 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2225 /* Update the parent page, if any, to point to the new page */
2227 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2228 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2229 SETPGNO(node, pgno);
2231 mc->mc_db->md_root = pgno;
2233 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2234 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2236 /* If txn has a parent, make sure the page is in our
2240 unsigned x = mdb_mid2l_search(dl, pgno);
2241 if (x <= dl[0].mid && dl[x].mid == pgno) {
2242 if (mp != dl[x].mptr) { /* bad cursor? */
2243 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2244 txn->mt_flags |= MDB_TXN_ERROR;
2245 return MDB_CORRUPTED;
2250 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2252 np = mdb_page_malloc(txn, 1);
2257 rc = mdb_mid2l_insert(dl, &mid);
2258 mdb_cassert(mc, rc == 0);
2263 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2265 np->mp_flags |= P_DIRTY;
2268 /* Adjust cursors pointing to mp */
2269 mc->mc_pg[mc->mc_top] = np;
2270 m2 = txn->mt_cursors[mc->mc_dbi];
2271 if (mc->mc_flags & C_SUB) {
2272 for (; m2; m2=m2->mc_next) {
2273 m3 = &m2->mc_xcursor->mx_cursor;
2274 if (m3->mc_snum < mc->mc_snum) continue;
2275 if (m3->mc_pg[mc->mc_top] == mp)
2276 m3->mc_pg[mc->mc_top] = np;
2279 for (; m2; m2=m2->mc_next) {
2280 if (m2->mc_snum < mc->mc_snum) continue;
2281 if (m2->mc_pg[mc->mc_top] == mp) {
2282 m2->mc_pg[mc->mc_top] = np;
2283 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2285 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2287 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2288 if (!(leaf->mn_flags & F_SUBDATA))
2289 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2297 txn->mt_flags |= MDB_TXN_ERROR;
2302 mdb_env_sync(MDB_env *env, int force)
2305 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2306 if (env->me_flags & MDB_WRITEMAP) {
2307 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2308 ? MS_ASYNC : MS_SYNC;
2309 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2312 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2316 if (MDB_FDATASYNC(env->me_fd))
2323 /** Back up parent txn's cursors, then grab the originals for tracking */
2325 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2327 MDB_cursor *mc, *bk;
2332 for (i = src->mt_numdbs; --i >= 0; ) {
2333 if ((mc = src->mt_cursors[i]) != NULL) {
2334 size = sizeof(MDB_cursor);
2336 size += sizeof(MDB_xcursor);
2337 for (; mc; mc = bk->mc_next) {
2343 mc->mc_db = &dst->mt_dbs[i];
2344 /* Kill pointers into src - and dst to reduce abuse: The
2345 * user may not use mc until dst ends. Otherwise we'd...
2347 mc->mc_txn = NULL; /* ...set this to dst */
2348 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2349 if ((mx = mc->mc_xcursor) != NULL) {
2350 *(MDB_xcursor *)(bk+1) = *mx;
2351 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2353 mc->mc_next = dst->mt_cursors[i];
2354 dst->mt_cursors[i] = mc;
2361 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2362 * @param[in] txn the transaction handle.
2363 * @param[in] merge true to keep changes to parent cursors, false to revert.
2364 * @return 0 on success, non-zero on failure.
2367 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2369 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2373 for (i = txn->mt_numdbs; --i >= 0; ) {
2374 for (mc = cursors[i]; mc; mc = next) {
2376 if ((bk = mc->mc_backup) != NULL) {
2378 /* Commit changes to parent txn */
2379 mc->mc_next = bk->mc_next;
2380 mc->mc_backup = bk->mc_backup;
2381 mc->mc_txn = bk->mc_txn;
2382 mc->mc_db = bk->mc_db;
2383 mc->mc_dbflag = bk->mc_dbflag;
2384 if ((mx = mc->mc_xcursor) != NULL)
2385 mx->mx_cursor.mc_txn = bk->mc_txn;
2387 /* Abort nested txn */
2389 if ((mx = mc->mc_xcursor) != NULL)
2390 *mx = *(MDB_xcursor *)(bk+1);
2394 /* Only malloced cursors are permanently tracked. */
2402 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2405 mdb_txn_reset0(MDB_txn *txn, const char *act);
2407 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2413 Pidset = F_SETLK, Pidcheck = F_GETLK
2417 /** Set or check a pid lock. Set returns 0 on success.
2418 * Check returns 0 if the process is certainly dead, nonzero if it may
2419 * be alive (the lock exists or an error happened so we do not know).
2421 * On Windows Pidset is a no-op, we merely check for the existence
2422 * of the process with the given pid. On POSIX we use a single byte
2423 * lock on the lockfile, set at an offset equal to the pid.
2426 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2428 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2431 if (op == Pidcheck) {
2432 h = OpenProcess(env->me_pidquery, FALSE, pid);
2433 /* No documented "no such process" code, but other program use this: */
2435 return ErrCode() != ERROR_INVALID_PARAMETER;
2436 /* A process exists until all handles to it close. Has it exited? */
2437 ret = WaitForSingleObject(h, 0) != 0;
2444 struct flock lock_info;
2445 memset(&lock_info, 0, sizeof(lock_info));
2446 lock_info.l_type = F_WRLCK;
2447 lock_info.l_whence = SEEK_SET;
2448 lock_info.l_start = pid;
2449 lock_info.l_len = 1;
2450 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2451 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2453 } else if ((rc = ErrCode()) == EINTR) {
2461 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2462 * @param[in] txn the transaction handle to initialize
2463 * @return 0 on success, non-zero on failure.
2466 mdb_txn_renew0(MDB_txn *txn)
2468 MDB_env *env = txn->mt_env;
2469 MDB_txninfo *ti = env->me_txns;
2473 int rc, new_notls = 0;
2475 if (txn->mt_flags & MDB_TXN_RDONLY) {
2477 txn->mt_numdbs = env->me_numdbs;
2478 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2480 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2481 txn->mt_txnid = meta->mm_txnid;
2482 txn->mt_u.reader = NULL;
2484 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2485 pthread_getspecific(env->me_txkey);
2487 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2488 return MDB_BAD_RSLOT;
2490 MDB_PID_T pid = env->me_pid;
2491 MDB_THR_T tid = pthread_self();
2493 if (!env->me_live_reader) {
2494 rc = mdb_reader_pid(env, Pidset, pid);
2497 env->me_live_reader = 1;
2501 nr = ti->mti_numreaders;
2502 for (i=0; i<nr; i++)
2503 if (ti->mti_readers[i].mr_pid == 0)
2505 if (i == env->me_maxreaders) {
2506 UNLOCK_MUTEX_R(env);
2507 return MDB_READERS_FULL;
2509 ti->mti_readers[i].mr_pid = pid;
2510 ti->mti_readers[i].mr_tid = tid;
2512 ti->mti_numreaders = ++nr;
2513 /* Save numreaders for un-mutexed mdb_env_close() */
2514 env->me_numreaders = nr;
2515 UNLOCK_MUTEX_R(env);
2517 r = &ti->mti_readers[i];
2518 new_notls = (env->me_flags & MDB_NOTLS);
2519 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2524 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2525 txn->mt_u.reader = r;
2526 meta = env->me_metas[txn->mt_txnid & 1];
2532 txn->mt_txnid = ti->mti_txnid;
2533 meta = env->me_metas[txn->mt_txnid & 1];
2535 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2536 txn->mt_txnid = meta->mm_txnid;
2539 txn->mt_numdbs = env->me_numdbs;
2542 if (txn->mt_txnid == mdb_debug_start)
2546 txn->mt_child = NULL;
2547 txn->mt_loose_pgs = NULL;
2548 txn->mt_loose_count = 0;
2549 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2550 txn->mt_u.dirty_list = env->me_dirty_list;
2551 txn->mt_u.dirty_list[0].mid = 0;
2552 txn->mt_free_pgs = env->me_free_pgs;
2553 txn->mt_free_pgs[0] = 0;
2554 txn->mt_spill_pgs = NULL;
2556 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2559 /* Copy the DB info and flags */
2560 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2562 /* Moved to here to avoid a data race in read TXNs */
2563 txn->mt_next_pgno = meta->mm_last_pg+1;
2565 for (i=2; i<txn->mt_numdbs; i++) {
2566 x = env->me_dbflags[i];
2567 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2568 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2570 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2572 if (env->me_maxpg < txn->mt_next_pgno) {
2573 mdb_txn_reset0(txn, "renew0-mapfail");
2575 txn->mt_u.reader->mr_pid = 0;
2576 txn->mt_u.reader = NULL;
2578 return MDB_MAP_RESIZED;
2585 mdb_txn_renew(MDB_txn *txn)
2589 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2592 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2593 DPUTS("environment had fatal error, must shutdown!");
2597 rc = mdb_txn_renew0(txn);
2598 if (rc == MDB_SUCCESS) {
2599 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2600 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2601 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2607 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2611 int rc, size, tsize = sizeof(MDB_txn);
2613 if (env->me_flags & MDB_FATAL_ERROR) {
2614 DPUTS("environment had fatal error, must shutdown!");
2617 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2620 /* Nested transactions: Max 1 child, write txns only, no writemap */
2621 if (parent->mt_child ||
2622 (flags & MDB_RDONLY) ||
2623 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2624 (env->me_flags & MDB_WRITEMAP))
2626 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2628 tsize = sizeof(MDB_ntxn);
2631 if (!(flags & MDB_RDONLY)) {
2633 txn = env->me_txn0; /* just reuse preallocated write txn */
2636 /* child txns use own copy of cursors */
2637 size += env->me_maxdbs * sizeof(MDB_cursor *);
2639 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2641 if ((txn = calloc(1, size)) == NULL) {
2642 DPRINTF(("calloc: %s", strerror(errno)));
2645 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2646 if (flags & MDB_RDONLY) {
2647 txn->mt_flags |= MDB_TXN_RDONLY;
2648 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2649 txn->mt_dbiseqs = env->me_dbiseqs;
2651 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2653 txn->mt_dbiseqs = parent->mt_dbiseqs;
2654 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2656 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2657 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2665 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2666 if (!txn->mt_u.dirty_list ||
2667 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2669 free(txn->mt_u.dirty_list);
2673 txn->mt_txnid = parent->mt_txnid;
2674 txn->mt_dirty_room = parent->mt_dirty_room;
2675 txn->mt_u.dirty_list[0].mid = 0;
2676 txn->mt_spill_pgs = NULL;
2677 txn->mt_next_pgno = parent->mt_next_pgno;
2678 parent->mt_child = txn;
2679 txn->mt_parent = parent;
2680 txn->mt_numdbs = parent->mt_numdbs;
2681 txn->mt_flags = parent->mt_flags;
2682 txn->mt_dbxs = parent->mt_dbxs;
2683 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2684 /* Copy parent's mt_dbflags, but clear DB_NEW */
2685 for (i=0; i<txn->mt_numdbs; i++)
2686 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2688 ntxn = (MDB_ntxn *)txn;
2689 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2690 if (env->me_pghead) {
2691 size = MDB_IDL_SIZEOF(env->me_pghead);
2692 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2694 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2699 rc = mdb_cursor_shadow(parent, txn);
2701 mdb_txn_reset0(txn, "beginchild-fail");
2703 rc = mdb_txn_renew0(txn);
2706 if (txn != env->me_txn0)
2710 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2711 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2712 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2719 mdb_txn_env(MDB_txn *txn)
2721 if(!txn) return NULL;
2725 /** Export or close DBI handles opened in this txn. */
2727 mdb_dbis_update(MDB_txn *txn, int keep)
2730 MDB_dbi n = txn->mt_numdbs;
2731 MDB_env *env = txn->mt_env;
2732 unsigned char *tdbflags = txn->mt_dbflags;
2734 for (i = n; --i >= 2;) {
2735 if (tdbflags[i] & DB_NEW) {
2737 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2739 char *ptr = env->me_dbxs[i].md_name.mv_data;
2741 env->me_dbxs[i].md_name.mv_data = NULL;
2742 env->me_dbxs[i].md_name.mv_size = 0;
2743 env->me_dbflags[i] = 0;
2744 env->me_dbiseqs[i]++;
2750 if (keep && env->me_numdbs < n)
2754 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2755 * May be called twice for readonly txns: First reset it, then abort.
2756 * @param[in] txn the transaction handle to reset
2757 * @param[in] act why the transaction is being reset
2760 mdb_txn_reset0(MDB_txn *txn, const char *act)
2762 MDB_env *env = txn->mt_env;
2764 /* Close any DBI handles opened in this txn */
2765 mdb_dbis_update(txn, 0);
2767 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2768 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2769 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2771 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2772 if (txn->mt_u.reader) {
2773 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2774 if (!(env->me_flags & MDB_NOTLS))
2775 txn->mt_u.reader = NULL; /* txn does not own reader */
2777 txn->mt_numdbs = 0; /* close nothing if called again */
2778 txn->mt_dbxs = NULL; /* mark txn as reset */
2780 pgno_t *pghead = env->me_pghead;
2781 env->me_pghead = NULL;
2784 mdb_cursors_close(txn, 0);
2786 if (!(env->me_flags & MDB_WRITEMAP)) {
2787 mdb_dlist_free(txn);
2790 if (!txn->mt_parent) {
2791 if (mdb_midl_shrink(&txn->mt_free_pgs))
2792 env->me_free_pgs = txn->mt_free_pgs;
2795 /* The writer mutex was locked in mdb_txn_begin. */
2797 UNLOCK_MUTEX_W(env);
2799 txn->mt_parent->mt_child = NULL;
2800 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2801 mdb_midl_free(txn->mt_free_pgs);
2802 mdb_midl_free(txn->mt_spill_pgs);
2803 free(txn->mt_u.dirty_list);
2806 mdb_midl_free(pghead);
2811 mdb_txn_reset(MDB_txn *txn)
2816 /* This call is only valid for read-only txns */
2817 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2820 mdb_txn_reset0(txn, "reset");
2824 mdb_txn_abort(MDB_txn *txn)
2830 mdb_txn_abort(txn->mt_child);
2832 mdb_txn_reset0(txn, "abort");
2833 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2834 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2835 txn->mt_u.reader->mr_pid = 0;
2837 if (txn != txn->mt_env->me_txn0)
2841 /** Save the freelist as of this transaction to the freeDB.
2842 * This changes the freelist. Keep trying until it stabilizes.
2845 mdb_freelist_save(MDB_txn *txn)
2847 /* env->me_pghead[] can grow and shrink during this call.
2848 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2849 * Page numbers cannot disappear from txn->mt_free_pgs[].
2852 MDB_env *env = txn->mt_env;
2853 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2854 txnid_t pglast = 0, head_id = 0;
2855 pgno_t freecnt = 0, *free_pgs, *mop;
2856 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2858 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2860 if (env->me_pghead) {
2861 /* Make sure first page of freeDB is touched and on freelist */
2862 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2863 if (rc && rc != MDB_NOTFOUND)
2867 if (!env->me_pghead && txn->mt_loose_pgs) {
2868 /* Put loose page numbers in mt_free_pgs, since
2869 * we may be unable to return them to me_pghead.
2871 MDB_page *mp = txn->mt_loose_pgs;
2872 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2874 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2875 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2876 txn->mt_loose_pgs = NULL;
2877 txn->mt_loose_count = 0;
2880 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2881 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2882 ? SSIZE_MAX : maxfree_1pg;
2885 /* Come back here after each Put() in case freelist changed */
2890 /* If using records from freeDB which we have not yet
2891 * deleted, delete them and any we reserved for me_pghead.
2893 while (pglast < env->me_pglast) {
2894 rc = mdb_cursor_first(&mc, &key, NULL);
2897 pglast = head_id = *(txnid_t *)key.mv_data;
2898 total_room = head_room = 0;
2899 mdb_tassert(txn, pglast <= env->me_pglast);
2900 rc = mdb_cursor_del(&mc, 0);
2905 /* Save the IDL of pages freed by this txn, to a single record */
2906 if (freecnt < txn->mt_free_pgs[0]) {
2908 /* Make sure last page of freeDB is touched and on freelist */
2909 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2910 if (rc && rc != MDB_NOTFOUND)
2913 free_pgs = txn->mt_free_pgs;
2914 /* Write to last page of freeDB */
2915 key.mv_size = sizeof(txn->mt_txnid);
2916 key.mv_data = &txn->mt_txnid;
2918 freecnt = free_pgs[0];
2919 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2920 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2923 /* Retry if mt_free_pgs[] grew during the Put() */
2924 free_pgs = txn->mt_free_pgs;
2925 } while (freecnt < free_pgs[0]);
2926 mdb_midl_sort(free_pgs);
2927 memcpy(data.mv_data, free_pgs, data.mv_size);
2930 unsigned int i = free_pgs[0];
2931 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2932 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2934 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2940 mop = env->me_pghead;
2941 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2943 /* Reserve records for me_pghead[]. Split it if multi-page,
2944 * to avoid searching freeDB for a page range. Use keys in
2945 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2947 if (total_room >= mop_len) {
2948 if (total_room == mop_len || --more < 0)
2950 } else if (head_room >= maxfree_1pg && head_id > 1) {
2951 /* Keep current record (overflow page), add a new one */
2955 /* (Re)write {key = head_id, IDL length = head_room} */
2956 total_room -= head_room;
2957 head_room = mop_len - total_room;
2958 if (head_room > maxfree_1pg && head_id > 1) {
2959 /* Overflow multi-page for part of me_pghead */
2960 head_room /= head_id; /* amortize page sizes */
2961 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2962 } else if (head_room < 0) {
2963 /* Rare case, not bothering to delete this record */
2966 key.mv_size = sizeof(head_id);
2967 key.mv_data = &head_id;
2968 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2969 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2972 /* IDL is initially empty, zero out at least the length */
2973 pgs = (pgno_t *)data.mv_data;
2974 j = head_room > clean_limit ? head_room : 0;
2978 total_room += head_room;
2981 /* Return loose page numbers to me_pghead, though usually none are
2982 * left at this point. The pages themselves remain in dirty_list.
2984 if (txn->mt_loose_pgs) {
2985 MDB_page *mp = txn->mt_loose_pgs;
2986 unsigned count = txn->mt_loose_count;
2988 /* Room for loose pages + temp IDL with same */
2989 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
2991 mop = env->me_pghead;
2992 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
2993 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
2994 loose[ ++count ] = mp->mp_pgno;
2996 mdb_midl_sort(loose);
2997 mdb_midl_xmerge(mop, loose);
2998 txn->mt_loose_pgs = NULL;
2999 txn->mt_loose_count = 0;
3003 /* Fill in the reserved me_pghead records */
3009 rc = mdb_cursor_first(&mc, &key, &data);
3010 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3011 txnid_t id = *(txnid_t *)key.mv_data;
3012 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3015 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3017 if (len > mop_len) {
3019 data.mv_size = (len + 1) * sizeof(MDB_ID);
3021 data.mv_data = mop -= len;
3024 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3026 if (rc || !(mop_len -= len))
3033 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3034 * @param[in] txn the transaction that's being committed
3035 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3036 * @return 0 on success, non-zero on failure.
3039 mdb_page_flush(MDB_txn *txn, int keep)
3041 MDB_env *env = txn->mt_env;
3042 MDB_ID2L dl = txn->mt_u.dirty_list;
3043 unsigned psize = env->me_psize, j;
3044 int i, pagecount = dl[0].mid, rc;
3045 size_t size = 0, pos = 0;
3047 MDB_page *dp = NULL;
3051 struct iovec iov[MDB_COMMIT_PAGES];
3052 ssize_t wpos = 0, wsize = 0, wres;
3053 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3059 if (env->me_flags & MDB_WRITEMAP) {
3060 /* Clear dirty flags */
3061 while (++i <= pagecount) {
3063 /* Don't flush this page yet */
3064 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3065 dp->mp_flags &= ~P_KEEP;
3069 dp->mp_flags &= ~P_DIRTY;
3074 /* Write the pages */
3076 if (++i <= pagecount) {
3078 /* Don't flush this page yet */
3079 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3080 dp->mp_flags &= ~P_KEEP;
3085 /* clear dirty flag */
3086 dp->mp_flags &= ~P_DIRTY;
3089 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3094 /* Windows actually supports scatter/gather I/O, but only on
3095 * unbuffered file handles. Since we're relying on the OS page
3096 * cache for all our data, that's self-defeating. So we just
3097 * write pages one at a time. We use the ov structure to set
3098 * the write offset, to at least save the overhead of a Seek
3101 DPRINTF(("committing page %"Z"u", pgno));
3102 memset(&ov, 0, sizeof(ov));
3103 ov.Offset = pos & 0xffffffff;
3104 ov.OffsetHigh = pos >> 16 >> 16;
3105 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3107 DPRINTF(("WriteFile: %d", rc));
3111 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3112 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3114 /* Write previous page(s) */
3115 #ifdef MDB_USE_PWRITEV
3116 wres = pwritev(env->me_fd, iov, n, wpos);
3119 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3121 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3123 DPRINTF(("lseek: %s", strerror(rc)));
3126 wres = writev(env->me_fd, iov, n);
3129 if (wres != wsize) {
3132 DPRINTF(("Write error: %s", strerror(rc)));
3134 rc = EIO; /* TODO: Use which error code? */
3135 DPUTS("short write, filesystem full?");
3146 DPRINTF(("committing page %"Z"u", pgno));
3147 next_pos = pos + size;
3148 iov[n].iov_len = size;
3149 iov[n].iov_base = (char *)dp;
3155 /* MIPS has cache coherency issues, this is a no-op everywhere else
3156 * Note: for any size >= on-chip cache size, entire on-chip cache is
3159 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3161 for (i = keep; ++i <= pagecount; ) {
3163 /* This is a page we skipped above */
3166 dl[j].mid = dp->mp_pgno;
3169 mdb_dpage_free(env, dp);
3174 txn->mt_dirty_room += i - j;
3180 mdb_txn_commit(MDB_txn *txn)
3186 if (txn == NULL || txn->mt_env == NULL)
3189 if (txn->mt_child) {
3190 rc = mdb_txn_commit(txn->mt_child);
3191 txn->mt_child = NULL;
3198 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3199 mdb_dbis_update(txn, 1);
3200 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3205 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3206 DPUTS("error flag is set, can't commit");
3208 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3213 if (txn->mt_parent) {
3214 MDB_txn *parent = txn->mt_parent;
3218 unsigned x, y, len, ps_len;
3220 /* Append our free list to parent's */
3221 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3224 mdb_midl_free(txn->mt_free_pgs);
3225 /* Failures after this must either undo the changes
3226 * to the parent or set MDB_TXN_ERROR in the parent.
3229 parent->mt_next_pgno = txn->mt_next_pgno;
3230 parent->mt_flags = txn->mt_flags;
3232 /* Merge our cursors into parent's and close them */
3233 mdb_cursors_close(txn, 1);
3235 /* Update parent's DB table. */
3236 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3237 parent->mt_numdbs = txn->mt_numdbs;
3238 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3239 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3240 for (i=2; i<txn->mt_numdbs; i++) {
3241 /* preserve parent's DB_NEW status */
3242 x = parent->mt_dbflags[i] & DB_NEW;
3243 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3246 dst = parent->mt_u.dirty_list;
3247 src = txn->mt_u.dirty_list;
3248 /* Remove anything in our dirty list from parent's spill list */
3249 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3251 pspill[0] = (pgno_t)-1;
3252 /* Mark our dirty pages as deleted in parent spill list */
3253 for (i=0, len=src[0].mid; ++i <= len; ) {
3254 MDB_ID pn = src[i].mid << 1;
3255 while (pn > pspill[x])
3257 if (pn == pspill[x]) {
3262 /* Squash deleted pagenums if we deleted any */
3263 for (x=y; ++x <= ps_len; )
3264 if (!(pspill[x] & 1))
3265 pspill[++y] = pspill[x];
3269 /* Find len = length of merging our dirty list with parent's */
3271 dst[0].mid = 0; /* simplify loops */
3272 if (parent->mt_parent) {
3273 len = x + src[0].mid;
3274 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3275 for (i = x; y && i; y--) {
3276 pgno_t yp = src[y].mid;
3277 while (yp < dst[i].mid)
3279 if (yp == dst[i].mid) {
3284 } else { /* Simplify the above for single-ancestor case */
3285 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3287 /* Merge our dirty list with parent's */
3289 for (i = len; y; dst[i--] = src[y--]) {
3290 pgno_t yp = src[y].mid;
3291 while (yp < dst[x].mid)
3292 dst[i--] = dst[x--];
3293 if (yp == dst[x].mid)
3294 free(dst[x--].mptr);
3296 mdb_tassert(txn, i == x);
3298 free(txn->mt_u.dirty_list);
3299 parent->mt_dirty_room = txn->mt_dirty_room;
3300 if (txn->mt_spill_pgs) {
3301 if (parent->mt_spill_pgs) {
3302 /* TODO: Prevent failure here, so parent does not fail */
3303 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3305 parent->mt_flags |= MDB_TXN_ERROR;
3306 mdb_midl_free(txn->mt_spill_pgs);
3307 mdb_midl_sort(parent->mt_spill_pgs);
3309 parent->mt_spill_pgs = txn->mt_spill_pgs;
3313 /* Append our loose page list to parent's */
3314 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3316 *lp = txn->mt_loose_pgs;
3317 parent->mt_loose_count += txn->mt_loose_count;
3319 parent->mt_child = NULL;
3320 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3325 if (txn != env->me_txn) {
3326 DPUTS("attempt to commit unknown transaction");
3331 mdb_cursors_close(txn, 0);
3333 if (!txn->mt_u.dirty_list[0].mid &&
3334 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3337 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3338 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3340 /* Update DB root pointers */
3341 if (txn->mt_numdbs > 2) {
3345 data.mv_size = sizeof(MDB_db);
3347 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3348 for (i = 2; i < txn->mt_numdbs; i++) {
3349 if (txn->mt_dbflags[i] & DB_DIRTY) {
3350 if (TXN_DBI_CHANGED(txn, i)) {
3354 data.mv_data = &txn->mt_dbs[i];
3355 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3362 rc = mdb_freelist_save(txn);
3366 mdb_midl_free(env->me_pghead);
3367 env->me_pghead = NULL;
3368 if (mdb_midl_shrink(&txn->mt_free_pgs))
3369 env->me_free_pgs = txn->mt_free_pgs;
3375 if ((rc = mdb_page_flush(txn, 0)) ||
3376 (rc = mdb_env_sync(env, 0)) ||
3377 (rc = mdb_env_write_meta(txn)))
3380 /* Free P_LOOSE pages left behind in dirty_list */
3381 if (!(env->me_flags & MDB_WRITEMAP))
3382 mdb_dlist_free(txn);
3387 mdb_dbis_update(txn, 1);
3390 UNLOCK_MUTEX_W(env);
3391 if (txn != env->me_txn0)
3401 /** Read the environment parameters of a DB environment before
3402 * mapping it into memory.
3403 * @param[in] env the environment handle
3404 * @param[out] meta address of where to store the meta information
3405 * @return 0 on success, non-zero on failure.
3408 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3414 enum { Size = sizeof(pbuf) };
3416 /* We don't know the page size yet, so use a minimum value.
3417 * Read both meta pages so we can use the latest one.
3420 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3424 memset(&ov, 0, sizeof(ov));
3426 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3427 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3430 rc = pread(env->me_fd, &pbuf, Size, off);
3433 if (rc == 0 && off == 0)
3435 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3436 DPRINTF(("read: %s", mdb_strerror(rc)));
3440 p = (MDB_page *)&pbuf;
3442 if (!F_ISSET(p->mp_flags, P_META)) {
3443 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3448 if (m->mm_magic != MDB_MAGIC) {
3449 DPUTS("meta has invalid magic");
3453 if (m->mm_version != MDB_DATA_VERSION) {
3454 DPRINTF(("database is version %u, expected version %u",
3455 m->mm_version, MDB_DATA_VERSION));
3456 return MDB_VERSION_MISMATCH;
3459 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3466 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3468 meta->mm_magic = MDB_MAGIC;
3469 meta->mm_version = MDB_DATA_VERSION;
3470 meta->mm_mapsize = env->me_mapsize;
3471 meta->mm_psize = env->me_psize;
3472 meta->mm_last_pg = 1;
3473 meta->mm_flags = env->me_flags & 0xffff;
3474 meta->mm_flags |= MDB_INTEGERKEY;
3475 meta->mm_dbs[0].md_root = P_INVALID;
3476 meta->mm_dbs[1].md_root = P_INVALID;
3479 /** Write the environment parameters of a freshly created DB environment.
3480 * @param[in] env the environment handle
3481 * @param[out] meta address of where to store the meta information
3482 * @return 0 on success, non-zero on failure.
3485 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3493 memset(&ov, 0, sizeof(ov));
3494 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3496 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3499 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3500 len = pwrite(fd, ptr, size, pos); \
3501 rc = (len >= 0); } while(0)
3504 DPUTS("writing new meta page");
3506 psize = env->me_psize;
3508 mdb_env_init_meta0(env, meta);
3510 p = calloc(2, psize);
3512 p->mp_flags = P_META;
3513 *(MDB_meta *)METADATA(p) = *meta;
3515 q = (MDB_page *)((char *)p + psize);
3517 q->mp_flags = P_META;
3518 *(MDB_meta *)METADATA(q) = *meta;
3520 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3523 else if ((unsigned) len == psize * 2)
3531 /** Update the environment info to commit a transaction.
3532 * @param[in] txn the transaction that's being committed
3533 * @return 0 on success, non-zero on failure.
3536 mdb_env_write_meta(MDB_txn *txn)
3539 MDB_meta meta, metab, *mp;
3542 int rc, len, toggle;
3551 toggle = txn->mt_txnid & 1;
3552 DPRINTF(("writing meta page %d for root page %"Z"u",
3553 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3556 mp = env->me_metas[toggle];
3557 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3558 /* Persist any increases of mapsize config */
3559 if (mapsize < env->me_mapsize)
3560 mapsize = env->me_mapsize;
3562 if (env->me_flags & MDB_WRITEMAP) {
3563 mp->mm_mapsize = mapsize;
3564 mp->mm_dbs[0] = txn->mt_dbs[0];
3565 mp->mm_dbs[1] = txn->mt_dbs[1];
3566 mp->mm_last_pg = txn->mt_next_pgno - 1;
3567 mp->mm_txnid = txn->mt_txnid;
3568 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3569 unsigned meta_size = env->me_psize;
3570 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3573 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3574 if (meta_size < env->me_os_psize)
3575 meta_size += meta_size;
3580 if (MDB_MSYNC(ptr, meta_size, rc)) {
3587 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3588 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3590 meta.mm_mapsize = mapsize;
3591 meta.mm_dbs[0] = txn->mt_dbs[0];
3592 meta.mm_dbs[1] = txn->mt_dbs[1];
3593 meta.mm_last_pg = txn->mt_next_pgno - 1;
3594 meta.mm_txnid = txn->mt_txnid;
3596 off = offsetof(MDB_meta, mm_mapsize);
3597 ptr = (char *)&meta + off;
3598 len = sizeof(MDB_meta) - off;
3600 off += env->me_psize;
3603 /* Write to the SYNC fd */
3604 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3605 env->me_fd : env->me_mfd;
3608 memset(&ov, 0, sizeof(ov));
3610 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3614 rc = pwrite(mfd, ptr, len, off);
3617 rc = rc < 0 ? ErrCode() : EIO;
3618 DPUTS("write failed, disk error?");
3619 /* On a failure, the pagecache still contains the new data.
3620 * Write some old data back, to prevent it from being used.
3621 * Use the non-SYNC fd; we know it will fail anyway.
3623 meta.mm_last_pg = metab.mm_last_pg;
3624 meta.mm_txnid = metab.mm_txnid;
3626 memset(&ov, 0, sizeof(ov));
3628 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3630 r2 = pwrite(env->me_fd, ptr, len, off);
3631 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3634 env->me_flags |= MDB_FATAL_ERROR;
3637 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3638 CACHEFLUSH(env->me_map + off, len, DCACHE);
3640 /* Memory ordering issues are irrelevant; since the entire writer
3641 * is wrapped by wmutex, all of these changes will become visible
3642 * after the wmutex is unlocked. Since the DB is multi-version,
3643 * readers will get consistent data regardless of how fresh or
3644 * how stale their view of these values is.
3647 env->me_txns->mti_txnid = txn->mt_txnid;
3652 /** Check both meta pages to see which one is newer.
3653 * @param[in] env the environment handle
3654 * @return meta toggle (0 or 1).
3657 mdb_env_pick_meta(const MDB_env *env)
3659 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3663 mdb_env_create(MDB_env **env)
3667 e = calloc(1, sizeof(MDB_env));
3671 e->me_maxreaders = DEFAULT_READERS;
3672 e->me_maxdbs = e->me_numdbs = 2;
3673 e->me_fd = INVALID_HANDLE_VALUE;
3674 e->me_lfd = INVALID_HANDLE_VALUE;
3675 e->me_mfd = INVALID_HANDLE_VALUE;
3676 #ifdef MDB_USE_POSIX_SEM
3677 e->me_rmutex = SEM_FAILED;
3678 e->me_wmutex = SEM_FAILED;
3680 e->me_pid = getpid();
3681 GET_PAGESIZE(e->me_os_psize);
3682 VGMEMP_CREATE(e,0,0);
3688 mdb_env_map(MDB_env *env, void *addr)
3691 unsigned int flags = env->me_flags;
3695 LONG sizelo, sizehi;
3698 if (flags & MDB_RDONLY) {
3699 /* Don't set explicit map size, use whatever exists */
3704 msize = env->me_mapsize;
3705 sizelo = msize & 0xffffffff;
3706 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3708 /* Windows won't create mappings for zero length files.
3709 * and won't map more than the file size.
3710 * Just set the maxsize right now.
3712 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3713 || !SetEndOfFile(env->me_fd)
3714 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3718 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3719 PAGE_READWRITE : PAGE_READONLY,
3720 sizehi, sizelo, NULL);
3723 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3724 FILE_MAP_WRITE : FILE_MAP_READ,
3726 rc = env->me_map ? 0 : ErrCode();
3731 int prot = PROT_READ;
3732 if (flags & MDB_WRITEMAP) {
3734 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3737 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3739 if (env->me_map == MAP_FAILED) {
3744 if (flags & MDB_NORDAHEAD) {
3745 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3747 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3749 #ifdef POSIX_MADV_RANDOM
3750 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3751 #endif /* POSIX_MADV_RANDOM */
3752 #endif /* MADV_RANDOM */
3756 /* Can happen because the address argument to mmap() is just a
3757 * hint. mmap() can pick another, e.g. if the range is in use.
3758 * The MAP_FIXED flag would prevent that, but then mmap could
3759 * instead unmap existing pages to make room for the new map.
3761 if (addr && env->me_map != addr)
3762 return EBUSY; /* TODO: Make a new MDB_* error code? */
3764 p = (MDB_page *)env->me_map;
3765 env->me_metas[0] = METADATA(p);
3766 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3772 mdb_env_set_mapsize(MDB_env *env, size_t size)
3774 /* If env is already open, caller is responsible for making
3775 * sure there are no active txns.
3783 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3784 else if (size < env->me_mapsize) {
3785 /* If the configured size is smaller, make sure it's
3786 * still big enough. Silently round up to minimum if not.
3788 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3792 munmap(env->me_map, env->me_mapsize);
3793 env->me_mapsize = size;
3794 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3795 rc = mdb_env_map(env, old);
3799 env->me_mapsize = size;
3801 env->me_maxpg = env->me_mapsize / env->me_psize;
3806 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3810 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3815 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3817 if (env->me_map || readers < 1)
3819 env->me_maxreaders = readers;
3824 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3826 if (!env || !readers)
3828 *readers = env->me_maxreaders;
3833 mdb_fsize(HANDLE fd, size_t *size)
3836 LARGE_INTEGER fsize;
3838 if (!GetFileSizeEx(fd, &fsize))
3841 *size = fsize.QuadPart;
3853 /** Further setup required for opening an LMDB environment
3856 mdb_env_open2(MDB_env *env)
3858 unsigned int flags = env->me_flags;
3859 int i, newenv = 0, rc;
3863 /* See if we should use QueryLimited */
3865 if ((rc & 0xff) > 5)
3866 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3868 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3871 memset(&meta, 0, sizeof(meta));
3873 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3876 DPUTS("new mdbenv");
3878 env->me_psize = env->me_os_psize;
3879 if (env->me_psize > MAX_PAGESIZE)
3880 env->me_psize = MAX_PAGESIZE;
3882 env->me_psize = meta.mm_psize;
3885 /* Was a mapsize configured? */
3886 if (!env->me_mapsize) {
3887 /* If this is a new environment, take the default,
3888 * else use the size recorded in the existing env.
3890 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3891 } else if (env->me_mapsize < meta.mm_mapsize) {
3892 /* If the configured size is smaller, make sure it's
3893 * still big enough. Silently round up to minimum if not.
3895 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3896 if (env->me_mapsize < minsize)
3897 env->me_mapsize = minsize;
3900 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3905 if (flags & MDB_FIXEDMAP)
3906 meta.mm_address = env->me_map;
3907 i = mdb_env_init_meta(env, &meta);
3908 if (i != MDB_SUCCESS) {
3913 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3914 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3916 #if !(MDB_MAXKEYSIZE)
3917 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3919 env->me_maxpg = env->me_mapsize / env->me_psize;
3923 int toggle = mdb_env_pick_meta(env);
3924 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3926 DPRINTF(("opened database version %u, pagesize %u",
3927 env->me_metas[0]->mm_version, env->me_psize));
3928 DPRINTF(("using meta page %d", toggle));
3929 DPRINTF(("depth: %u", db->md_depth));
3930 DPRINTF(("entries: %"Z"u", db->md_entries));
3931 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3932 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3933 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3934 DPRINTF(("root: %"Z"u", db->md_root));
3942 /** Release a reader thread's slot in the reader lock table.
3943 * This function is called automatically when a thread exits.
3944 * @param[in] ptr This points to the slot in the reader lock table.
3947 mdb_env_reader_dest(void *ptr)
3949 MDB_reader *reader = ptr;
3955 /** Junk for arranging thread-specific callbacks on Windows. This is
3956 * necessarily platform and compiler-specific. Windows supports up
3957 * to 1088 keys. Let's assume nobody opens more than 64 environments
3958 * in a single process, for now. They can override this if needed.
3960 #ifndef MAX_TLS_KEYS
3961 #define MAX_TLS_KEYS 64
3963 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3964 static int mdb_tls_nkeys;
3966 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3970 case DLL_PROCESS_ATTACH: break;
3971 case DLL_THREAD_ATTACH: break;
3972 case DLL_THREAD_DETACH:
3973 for (i=0; i<mdb_tls_nkeys; i++) {
3974 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3976 mdb_env_reader_dest(r);
3980 case DLL_PROCESS_DETACH: break;
3985 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3987 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3991 /* Force some symbol references.
3992 * _tls_used forces the linker to create the TLS directory if not already done
3993 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3995 #pragma comment(linker, "/INCLUDE:_tls_used")
3996 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3997 #pragma const_seg(".CRT$XLB")
3998 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3999 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4002 #pragma comment(linker, "/INCLUDE:__tls_used")
4003 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4004 #pragma data_seg(".CRT$XLB")
4005 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4007 #endif /* WIN 32/64 */
4008 #endif /* !__GNUC__ */
4011 /** Downgrade the exclusive lock on the region back to shared */
4013 mdb_env_share_locks(MDB_env *env, int *excl)
4015 int rc = 0, toggle = mdb_env_pick_meta(env);
4017 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4022 /* First acquire a shared lock. The Unlock will
4023 * then release the existing exclusive lock.
4025 memset(&ov, 0, sizeof(ov));
4026 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4029 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4035 struct flock lock_info;
4036 /* The shared lock replaces the existing lock */
4037 memset((void *)&lock_info, 0, sizeof(lock_info));
4038 lock_info.l_type = F_RDLCK;
4039 lock_info.l_whence = SEEK_SET;
4040 lock_info.l_start = 0;
4041 lock_info.l_len = 1;
4042 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4043 (rc = ErrCode()) == EINTR) ;
4044 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4051 /** Try to get exlusive lock, otherwise shared.
4052 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4055 mdb_env_excl_lock(MDB_env *env, int *excl)
4059 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4063 memset(&ov, 0, sizeof(ov));
4064 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4071 struct flock lock_info;
4072 memset((void *)&lock_info, 0, sizeof(lock_info));
4073 lock_info.l_type = F_WRLCK;
4074 lock_info.l_whence = SEEK_SET;
4075 lock_info.l_start = 0;
4076 lock_info.l_len = 1;
4077 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4078 (rc = ErrCode()) == EINTR) ;
4082 # ifdef MDB_USE_POSIX_SEM
4083 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4086 lock_info.l_type = F_RDLCK;
4087 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4088 (rc = ErrCode()) == EINTR) ;
4098 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4100 * @(#) $Revision: 5.1 $
4101 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4102 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4104 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4108 * Please do not copyright this code. This code is in the public domain.
4110 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4111 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4112 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4113 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4114 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4115 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4116 * PERFORMANCE OF THIS SOFTWARE.
4119 * chongo <Landon Curt Noll> /\oo/\
4120 * http://www.isthe.com/chongo/
4122 * Share and Enjoy! :-)
4125 typedef unsigned long long mdb_hash_t;
4126 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4128 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4129 * @param[in] val value to hash
4130 * @param[in] hval initial value for hash
4131 * @return 64 bit hash
4133 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4134 * hval arg on the first call.
4137 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4139 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4140 unsigned char *end = s + val->mv_size;
4142 * FNV-1a hash each octet of the string
4145 /* xor the bottom with the current octet */
4146 hval ^= (mdb_hash_t)*s++;
4148 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4149 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4150 (hval << 7) + (hval << 8) + (hval << 40);
4152 /* return our new hash value */
4156 /** Hash the string and output the encoded hash.
4157 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4158 * very short name limits. We don't care about the encoding being reversible,
4159 * we just want to preserve as many bits of the input as possible in a
4160 * small printable string.
4161 * @param[in] str string to hash
4162 * @param[out] encbuf an array of 11 chars to hold the hash
4164 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4167 mdb_pack85(unsigned long l, char *out)
4171 for (i=0; i<5; i++) {
4172 *out++ = mdb_a85[l % 85];
4178 mdb_hash_enc(MDB_val *val, char *encbuf)
4180 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4182 mdb_pack85(h, encbuf);
4183 mdb_pack85(h>>32, encbuf+5);
4188 /** Open and/or initialize the lock region for the environment.
4189 * @param[in] env The LMDB environment.
4190 * @param[in] lpath The pathname of the file used for the lock region.
4191 * @param[in] mode The Unix permissions for the file, if we create it.
4192 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4193 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4194 * @return 0 on success, non-zero on failure.
4197 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4200 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4202 # define MDB_ERRCODE_ROFS EROFS
4203 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4204 # define MDB_CLOEXEC O_CLOEXEC
4207 # define MDB_CLOEXEC 0
4214 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4215 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4216 FILE_ATTRIBUTE_NORMAL, NULL);
4218 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4220 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4222 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4227 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4228 /* Lose record locks when exec*() */
4229 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4230 fcntl(env->me_lfd, F_SETFD, fdflags);
4233 if (!(env->me_flags & MDB_NOTLS)) {
4234 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4237 env->me_flags |= MDB_ENV_TXKEY;
4239 /* Windows TLS callbacks need help finding their TLS info. */
4240 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4244 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4248 /* Try to get exclusive lock. If we succeed, then
4249 * nobody is using the lock region and we should initialize it.
4251 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4254 size = GetFileSize(env->me_lfd, NULL);
4256 size = lseek(env->me_lfd, 0, SEEK_END);
4257 if (size == -1) goto fail_errno;
4259 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4260 if (size < rsize && *excl > 0) {
4262 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4263 || !SetEndOfFile(env->me_lfd))
4266 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4270 size = rsize - sizeof(MDB_txninfo);
4271 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4276 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4278 if (!mh) goto fail_errno;
4279 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4281 if (!env->me_txns) goto fail_errno;
4283 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4285 if (m == MAP_FAILED) goto fail_errno;
4291 BY_HANDLE_FILE_INFORMATION stbuf;
4300 if (!mdb_sec_inited) {
4301 InitializeSecurityDescriptor(&mdb_null_sd,
4302 SECURITY_DESCRIPTOR_REVISION);
4303 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4304 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4305 mdb_all_sa.bInheritHandle = FALSE;
4306 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4309 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4310 idbuf.volume = stbuf.dwVolumeSerialNumber;
4311 idbuf.nhigh = stbuf.nFileIndexHigh;
4312 idbuf.nlow = stbuf.nFileIndexLow;
4313 val.mv_data = &idbuf;
4314 val.mv_size = sizeof(idbuf);
4315 mdb_hash_enc(&val, encbuf);
4316 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4317 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4318 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4319 if (!env->me_rmutex) goto fail_errno;
4320 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4321 if (!env->me_wmutex) goto fail_errno;
4322 #elif defined(MDB_USE_POSIX_SEM)
4331 #if defined(__NetBSD__)
4332 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4334 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4335 idbuf.dev = stbuf.st_dev;
4336 idbuf.ino = stbuf.st_ino;
4337 val.mv_data = &idbuf;
4338 val.mv_size = sizeof(idbuf);
4339 mdb_hash_enc(&val, encbuf);
4340 #ifdef MDB_SHORT_SEMNAMES
4341 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4343 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4344 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4345 /* Clean up after a previous run, if needed: Try to
4346 * remove both semaphores before doing anything else.
4348 sem_unlink(env->me_txns->mti_rmname);
4349 sem_unlink(env->me_txns->mti_wmname);
4350 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4351 O_CREAT|O_EXCL, mode, 1);
4352 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4353 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4354 O_CREAT|O_EXCL, mode, 1);
4355 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4356 #else /* MDB_USE_POSIX_SEM */
4357 pthread_mutexattr_t mattr;
4359 if ((rc = pthread_mutexattr_init(&mattr))
4360 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4361 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4362 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4364 pthread_mutexattr_destroy(&mattr);
4365 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4367 env->me_txns->mti_magic = MDB_MAGIC;
4368 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4369 env->me_txns->mti_txnid = 0;
4370 env->me_txns->mti_numreaders = 0;
4373 if (env->me_txns->mti_magic != MDB_MAGIC) {
4374 DPUTS("lock region has invalid magic");
4378 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4379 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4380 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4381 rc = MDB_VERSION_MISMATCH;
4385 if (rc && rc != EACCES && rc != EAGAIN) {
4389 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4390 if (!env->me_rmutex) goto fail_errno;
4391 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4392 if (!env->me_wmutex) goto fail_errno;
4393 #elif defined(MDB_USE_POSIX_SEM)
4394 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4395 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4396 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4397 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4408 /** The name of the lock file in the DB environment */
4409 #define LOCKNAME "/lock.mdb"
4410 /** The name of the data file in the DB environment */
4411 #define DATANAME "/data.mdb"
4412 /** The suffix of the lock file when no subdir is used */
4413 #define LOCKSUFF "-lock"
4414 /** Only a subset of the @ref mdb_env flags can be changed
4415 * at runtime. Changing other flags requires closing the
4416 * environment and re-opening it with the new flags.
4418 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4419 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4420 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4422 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4423 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4427 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4429 int oflags, rc, len, excl = -1;
4430 char *lpath, *dpath;
4432 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4436 if (flags & MDB_NOSUBDIR) {
4437 rc = len + sizeof(LOCKSUFF) + len + 1;
4439 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4444 if (flags & MDB_NOSUBDIR) {
4445 dpath = lpath + len + sizeof(LOCKSUFF);
4446 sprintf(lpath, "%s" LOCKSUFF, path);
4447 strcpy(dpath, path);
4449 dpath = lpath + len + sizeof(LOCKNAME);
4450 sprintf(lpath, "%s" LOCKNAME, path);
4451 sprintf(dpath, "%s" DATANAME, path);
4455 flags |= env->me_flags;
4456 if (flags & MDB_RDONLY) {
4457 /* silently ignore WRITEMAP when we're only getting read access */
4458 flags &= ~MDB_WRITEMAP;
4460 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4461 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4464 env->me_flags = flags |= MDB_ENV_ACTIVE;
4468 env->me_path = strdup(path);
4469 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4470 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4471 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4472 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4477 /* For RDONLY, get lockfile after we know datafile exists */
4478 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4479 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4485 if (F_ISSET(flags, MDB_RDONLY)) {
4486 oflags = GENERIC_READ;
4487 len = OPEN_EXISTING;
4489 oflags = GENERIC_READ|GENERIC_WRITE;
4492 mode = FILE_ATTRIBUTE_NORMAL;
4493 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4494 NULL, len, mode, NULL);
4496 if (F_ISSET(flags, MDB_RDONLY))
4499 oflags = O_RDWR | O_CREAT;
4501 env->me_fd = open(dpath, oflags, mode);
4503 if (env->me_fd == INVALID_HANDLE_VALUE) {
4508 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4509 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4514 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4515 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4516 env->me_mfd = env->me_fd;
4518 /* Synchronous fd for meta writes. Needed even with
4519 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4522 len = OPEN_EXISTING;
4523 env->me_mfd = CreateFile(dpath, oflags,
4524 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4525 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4528 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4530 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4535 DPRINTF(("opened dbenv %p", (void *) env));
4537 rc = mdb_env_share_locks(env, &excl);
4541 if (!((flags & MDB_RDONLY) ||
4542 (env->me_pbuf = calloc(1, env->me_psize))))
4544 if (!(flags & MDB_RDONLY)) {
4546 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4547 (sizeof(MDB_db)+sizeof(MDB_cursor)+sizeof(unsigned int)+1);
4548 txn = calloc(1, size);
4550 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4551 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4552 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4553 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4555 txn->mt_dbxs = env->me_dbxs;
4565 mdb_env_close0(env, excl);
4571 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4573 mdb_env_close0(MDB_env *env, int excl)
4577 if (!(env->me_flags & MDB_ENV_ACTIVE))
4580 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4581 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4582 free(env->me_dbxs[i].md_name.mv_data);
4585 free(env->me_dbiseqs);
4586 free(env->me_dbflags);
4589 free(env->me_dirty_list);
4591 mdb_midl_free(env->me_free_pgs);
4593 if (env->me_flags & MDB_ENV_TXKEY) {
4594 pthread_key_delete(env->me_txkey);
4596 /* Delete our key from the global list */
4597 for (i=0; i<mdb_tls_nkeys; i++)
4598 if (mdb_tls_keys[i] == env->me_txkey) {
4599 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4607 munmap(env->me_map, env->me_mapsize);
4609 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4610 (void) close(env->me_mfd);
4611 if (env->me_fd != INVALID_HANDLE_VALUE)
4612 (void) close(env->me_fd);
4614 MDB_PID_T pid = env->me_pid;
4615 /* Clearing readers is done in this function because
4616 * me_txkey with its destructor must be disabled first.
4618 for (i = env->me_numreaders; --i >= 0; )
4619 if (env->me_txns->mti_readers[i].mr_pid == pid)
4620 env->me_txns->mti_readers[i].mr_pid = 0;
4622 if (env->me_rmutex) {
4623 CloseHandle(env->me_rmutex);
4624 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4626 /* Windows automatically destroys the mutexes when
4627 * the last handle closes.
4629 #elif defined(MDB_USE_POSIX_SEM)
4630 if (env->me_rmutex != SEM_FAILED) {
4631 sem_close(env->me_rmutex);
4632 if (env->me_wmutex != SEM_FAILED)
4633 sem_close(env->me_wmutex);
4634 /* If we have the filelock: If we are the
4635 * only remaining user, clean up semaphores.
4638 mdb_env_excl_lock(env, &excl);
4640 sem_unlink(env->me_txns->mti_rmname);
4641 sem_unlink(env->me_txns->mti_wmname);
4645 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4647 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4650 /* Unlock the lockfile. Windows would have unlocked it
4651 * after closing anyway, but not necessarily at once.
4653 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4656 (void) close(env->me_lfd);
4659 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4664 mdb_env_close(MDB_env *env)
4671 VGMEMP_DESTROY(env);
4672 while ((dp = env->me_dpages) != NULL) {
4673 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4674 env->me_dpages = dp->mp_next;
4678 mdb_env_close0(env, 0);
4682 /** Compare two items pointing at aligned size_t's */
4684 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4686 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4687 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4690 /** Compare two items pointing at aligned unsigned int's */
4692 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4694 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4695 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4698 /** Compare two items pointing at unsigned ints of unknown alignment.
4699 * Nodes and keys are guaranteed to be 2-byte aligned.
4702 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4704 #if BYTE_ORDER == LITTLE_ENDIAN
4705 unsigned short *u, *c;
4708 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4709 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4712 } while(!x && u > (unsigned short *)a->mv_data);
4715 unsigned short *u, *c, *end;
4718 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4719 u = (unsigned short *)a->mv_data;
4720 c = (unsigned short *)b->mv_data;
4723 } while(!x && u < end);
4728 /** Compare two items pointing at size_t's of unknown alignment. */
4729 #ifdef MISALIGNED_OK
4730 # define mdb_cmp_clong mdb_cmp_long
4732 # define mdb_cmp_clong mdb_cmp_cint
4735 /** Compare two items lexically */
4737 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4744 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4750 diff = memcmp(a->mv_data, b->mv_data, len);
4751 return diff ? diff : len_diff<0 ? -1 : len_diff;
4754 /** Compare two items in reverse byte order */
4756 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4758 const unsigned char *p1, *p2, *p1_lim;
4762 p1_lim = (const unsigned char *)a->mv_data;
4763 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4764 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4766 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4772 while (p1 > p1_lim) {
4773 diff = *--p1 - *--p2;
4777 return len_diff<0 ? -1 : len_diff;
4780 /** Search for key within a page, using binary search.
4781 * Returns the smallest entry larger or equal to the key.
4782 * If exactp is non-null, stores whether the found entry was an exact match
4783 * in *exactp (1 or 0).
4784 * Updates the cursor index with the index of the found entry.
4785 * If no entry larger or equal to the key is found, returns NULL.
4788 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4790 unsigned int i = 0, nkeys;
4793 MDB_page *mp = mc->mc_pg[mc->mc_top];
4794 MDB_node *node = NULL;
4799 nkeys = NUMKEYS(mp);
4801 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4802 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4805 low = IS_LEAF(mp) ? 0 : 1;
4807 cmp = mc->mc_dbx->md_cmp;
4809 /* Branch pages have no data, so if using integer keys,
4810 * alignment is guaranteed. Use faster mdb_cmp_int.
4812 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4813 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4820 nodekey.mv_size = mc->mc_db->md_pad;
4821 node = NODEPTR(mp, 0); /* fake */
4822 while (low <= high) {
4823 i = (low + high) >> 1;
4824 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4825 rc = cmp(key, &nodekey);
4826 DPRINTF(("found leaf index %u [%s], rc = %i",
4827 i, DKEY(&nodekey), rc));
4836 while (low <= high) {
4837 i = (low + high) >> 1;
4839 node = NODEPTR(mp, i);
4840 nodekey.mv_size = NODEKSZ(node);
4841 nodekey.mv_data = NODEKEY(node);
4843 rc = cmp(key, &nodekey);
4846 DPRINTF(("found leaf index %u [%s], rc = %i",
4847 i, DKEY(&nodekey), rc));
4849 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4850 i, DKEY(&nodekey), NODEPGNO(node), rc));
4861 if (rc > 0) { /* Found entry is less than the key. */
4862 i++; /* Skip to get the smallest entry larger than key. */
4864 node = NODEPTR(mp, i);
4867 *exactp = (rc == 0 && nkeys > 0);
4868 /* store the key index */
4869 mc->mc_ki[mc->mc_top] = i;
4871 /* There is no entry larger or equal to the key. */
4874 /* nodeptr is fake for LEAF2 */
4880 mdb_cursor_adjust(MDB_cursor *mc, func)
4884 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4885 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4892 /** Pop a page off the top of the cursor's stack. */
4894 mdb_cursor_pop(MDB_cursor *mc)
4898 MDB_page *top = mc->mc_pg[mc->mc_top];
4904 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4905 DDBI(mc), (void *) mc));
4909 /** Push a page onto the top of the cursor's stack. */
4911 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4913 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4914 DDBI(mc), (void *) mc));
4916 if (mc->mc_snum >= CURSOR_STACK) {
4917 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4918 return MDB_CURSOR_FULL;
4921 mc->mc_top = mc->mc_snum++;
4922 mc->mc_pg[mc->mc_top] = mp;
4923 mc->mc_ki[mc->mc_top] = 0;
4928 /** Find the address of the page corresponding to a given page number.
4929 * @param[in] txn the transaction for this access.
4930 * @param[in] pgno the page number for the page to retrieve.
4931 * @param[out] ret address of a pointer where the page's address will be stored.
4932 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4933 * @return 0 on success, non-zero on failure.
4936 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4938 MDB_env *env = txn->mt_env;
4942 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4946 MDB_ID2L dl = tx2->mt_u.dirty_list;
4948 /* Spilled pages were dirtied in this txn and flushed
4949 * because the dirty list got full. Bring this page
4950 * back in from the map (but don't unspill it here,
4951 * leave that unless page_touch happens again).
4953 if (tx2->mt_spill_pgs) {
4954 MDB_ID pn = pgno << 1;
4955 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4956 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4957 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4962 unsigned x = mdb_mid2l_search(dl, pgno);
4963 if (x <= dl[0].mid && dl[x].mid == pgno) {
4969 } while ((tx2 = tx2->mt_parent) != NULL);
4972 if (pgno < txn->mt_next_pgno) {
4974 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4976 DPRINTF(("page %"Z"u not found", pgno));
4977 txn->mt_flags |= MDB_TXN_ERROR;
4978 return MDB_PAGE_NOTFOUND;
4988 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4989 * The cursor is at the root page, set up the rest of it.
4992 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4994 MDB_page *mp = mc->mc_pg[mc->mc_top];
4998 while (IS_BRANCH(mp)) {
5002 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5003 mdb_cassert(mc, NUMKEYS(mp) > 1);
5004 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5006 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5008 if (flags & MDB_PS_LAST)
5009 i = NUMKEYS(mp) - 1;
5012 node = mdb_node_search(mc, key, &exact);
5014 i = NUMKEYS(mp) - 1;
5016 i = mc->mc_ki[mc->mc_top];
5018 mdb_cassert(mc, i > 0);
5022 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5025 mdb_cassert(mc, i < NUMKEYS(mp));
5026 node = NODEPTR(mp, i);
5028 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5031 mc->mc_ki[mc->mc_top] = i;
5032 if ((rc = mdb_cursor_push(mc, mp)))
5035 if (flags & MDB_PS_MODIFY) {
5036 if ((rc = mdb_page_touch(mc)) != 0)
5038 mp = mc->mc_pg[mc->mc_top];
5043 DPRINTF(("internal error, index points to a %02X page!?",
5045 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5046 return MDB_CORRUPTED;
5049 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5050 key ? DKEY(key) : "null"));
5051 mc->mc_flags |= C_INITIALIZED;
5052 mc->mc_flags &= ~C_EOF;
5057 /** Search for the lowest key under the current branch page.
5058 * This just bypasses a NUMKEYS check in the current page
5059 * before calling mdb_page_search_root(), because the callers
5060 * are all in situations where the current page is known to
5064 mdb_page_search_lowest(MDB_cursor *mc)
5066 MDB_page *mp = mc->mc_pg[mc->mc_top];
5067 MDB_node *node = NODEPTR(mp, 0);
5070 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5073 mc->mc_ki[mc->mc_top] = 0;
5074 if ((rc = mdb_cursor_push(mc, mp)))
5076 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5079 /** Search for the page a given key should be in.
5080 * Push it and its parent pages on the cursor stack.
5081 * @param[in,out] mc the cursor for this operation.
5082 * @param[in] key the key to search for, or NULL for first/last page.
5083 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5084 * are touched (updated with new page numbers).
5085 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5086 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5087 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5088 * @return 0 on success, non-zero on failure.
5091 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5096 /* Make sure the txn is still viable, then find the root from
5097 * the txn's db table and set it as the root of the cursor's stack.
5099 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5100 DPUTS("transaction has failed, must abort");
5103 /* Make sure we're using an up-to-date root */
5104 if (*mc->mc_dbflag & DB_STALE) {
5106 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5108 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5109 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5116 MDB_node *leaf = mdb_node_search(&mc2,
5117 &mc->mc_dbx->md_name, &exact);
5119 return MDB_NOTFOUND;
5120 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5123 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5125 /* The txn may not know this DBI, or another process may
5126 * have dropped and recreated the DB with other flags.
5128 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5129 return MDB_INCOMPATIBLE;
5130 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5132 *mc->mc_dbflag &= ~DB_STALE;
5134 root = mc->mc_db->md_root;
5136 if (root == P_INVALID) { /* Tree is empty. */
5137 DPUTS("tree is empty");
5138 return MDB_NOTFOUND;
5142 mdb_cassert(mc, root > 1);
5143 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5144 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5150 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5151 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5153 if (flags & MDB_PS_MODIFY) {
5154 if ((rc = mdb_page_touch(mc)))
5158 if (flags & MDB_PS_ROOTONLY)
5161 return mdb_page_search_root(mc, key, flags);
5165 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5167 MDB_txn *txn = mc->mc_txn;
5168 pgno_t pg = mp->mp_pgno;
5169 unsigned x = 0, ovpages = mp->mp_pages;
5170 MDB_env *env = txn->mt_env;
5171 MDB_IDL sl = txn->mt_spill_pgs;
5172 MDB_ID pn = pg << 1;
5175 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5176 /* If the page is dirty or on the spill list we just acquired it,
5177 * so we should give it back to our current free list, if any.
5178 * Otherwise put it onto the list of pages we freed in this txn.
5180 * Won't create me_pghead: me_pglast must be inited along with it.
5181 * Unsupported in nested txns: They would need to hide the page
5182 * range in ancestor txns' dirty and spilled lists.
5184 if (env->me_pghead &&
5186 ((mp->mp_flags & P_DIRTY) ||
5187 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5191 MDB_ID2 *dl, ix, iy;
5192 rc = mdb_midl_need(&env->me_pghead, ovpages);
5195 if (!(mp->mp_flags & P_DIRTY)) {
5196 /* This page is no longer spilled */
5203 /* Remove from dirty list */
5204 dl = txn->mt_u.dirty_list;
5206 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5212 mdb_cassert(mc, x > 1);
5214 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5215 txn->mt_flags |= MDB_TXN_ERROR;
5216 return MDB_CORRUPTED;
5219 if (!(env->me_flags & MDB_WRITEMAP))
5220 mdb_dpage_free(env, mp);
5222 /* Insert in me_pghead */
5223 mop = env->me_pghead;
5224 j = mop[0] + ovpages;
5225 for (i = mop[0]; i && mop[i] < pg; i--)
5231 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5235 mc->mc_db->md_overflow_pages -= ovpages;
5239 /** Return the data associated with a given node.
5240 * @param[in] txn The transaction for this operation.
5241 * @param[in] leaf The node being read.
5242 * @param[out] data Updated to point to the node's data.
5243 * @return 0 on success, non-zero on failure.
5246 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5248 MDB_page *omp; /* overflow page */
5252 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5253 data->mv_size = NODEDSZ(leaf);
5254 data->mv_data = NODEDATA(leaf);
5258 /* Read overflow data.
5260 data->mv_size = NODEDSZ(leaf);
5261 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5262 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5263 DPRINTF(("read overflow page %"Z"u failed", pgno));
5266 data->mv_data = METADATA(omp);
5272 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5273 MDB_val *key, MDB_val *data)
5280 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5282 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5285 if (txn->mt_flags & MDB_TXN_ERROR)
5288 mdb_cursor_init(&mc, txn, dbi, &mx);
5289 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5292 /** Find a sibling for a page.
5293 * Replaces the page at the top of the cursor's stack with the
5294 * specified sibling, if one exists.
5295 * @param[in] mc The cursor for this operation.
5296 * @param[in] move_right Non-zero if the right sibling is requested,
5297 * otherwise the left sibling.
5298 * @return 0 on success, non-zero on failure.
5301 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5307 if (mc->mc_snum < 2) {
5308 return MDB_NOTFOUND; /* root has no siblings */
5312 DPRINTF(("parent page is page %"Z"u, index %u",
5313 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5315 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5316 : (mc->mc_ki[mc->mc_top] == 0)) {
5317 DPRINTF(("no more keys left, moving to %s sibling",
5318 move_right ? "right" : "left"));
5319 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5320 /* undo cursor_pop before returning */
5327 mc->mc_ki[mc->mc_top]++;
5329 mc->mc_ki[mc->mc_top]--;
5330 DPRINTF(("just moving to %s index key %u",
5331 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5333 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5335 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5336 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5337 /* mc will be inconsistent if caller does mc_snum++ as above */
5338 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5342 mdb_cursor_push(mc, mp);
5344 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5349 /** Move the cursor to the next data item. */
5351 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5357 if (mc->mc_flags & C_EOF) {
5358 return MDB_NOTFOUND;
5361 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5363 mp = mc->mc_pg[mc->mc_top];
5365 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5366 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5367 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5368 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5369 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5370 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5371 if (rc == MDB_SUCCESS)
5372 MDB_GET_KEY(leaf, key);
5377 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5378 if (op == MDB_NEXT_DUP)
5379 return MDB_NOTFOUND;
5383 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5384 mdb_dbg_pgno(mp), (void *) mc));
5385 if (mc->mc_flags & C_DEL)
5388 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5389 DPUTS("=====> move to next sibling page");
5390 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5391 mc->mc_flags |= C_EOF;
5394 mp = mc->mc_pg[mc->mc_top];
5395 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5397 mc->mc_ki[mc->mc_top]++;
5400 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5401 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5404 key->mv_size = mc->mc_db->md_pad;
5405 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5409 mdb_cassert(mc, IS_LEAF(mp));
5410 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5412 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5413 mdb_xcursor_init1(mc, leaf);
5416 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5419 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5420 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5421 if (rc != MDB_SUCCESS)
5426 MDB_GET_KEY(leaf, key);
5430 /** Move the cursor to the previous data item. */
5432 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5438 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5440 mp = mc->mc_pg[mc->mc_top];
5442 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5443 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5444 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5445 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5446 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5447 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5448 if (rc == MDB_SUCCESS) {
5449 MDB_GET_KEY(leaf, key);
5450 mc->mc_flags &= ~C_EOF;
5456 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5457 if (op == MDB_PREV_DUP)
5458 return MDB_NOTFOUND;
5462 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5463 mdb_dbg_pgno(mp), (void *) mc));
5465 if (mc->mc_ki[mc->mc_top] == 0) {
5466 DPUTS("=====> move to prev sibling page");
5467 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5470 mp = mc->mc_pg[mc->mc_top];
5471 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5472 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5474 mc->mc_ki[mc->mc_top]--;
5476 mc->mc_flags &= ~C_EOF;
5478 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5479 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5482 key->mv_size = mc->mc_db->md_pad;
5483 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5487 mdb_cassert(mc, IS_LEAF(mp));
5488 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5490 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5491 mdb_xcursor_init1(mc, leaf);
5494 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5497 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5498 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5499 if (rc != MDB_SUCCESS)
5504 MDB_GET_KEY(leaf, key);
5508 /** Set the cursor on a specific data item. */
5510 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5511 MDB_cursor_op op, int *exactp)
5515 MDB_node *leaf = NULL;
5518 if (key->mv_size == 0)
5519 return MDB_BAD_VALSIZE;
5522 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5524 /* See if we're already on the right page */
5525 if (mc->mc_flags & C_INITIALIZED) {
5528 mp = mc->mc_pg[mc->mc_top];
5530 mc->mc_ki[mc->mc_top] = 0;
5531 return MDB_NOTFOUND;
5533 if (mp->mp_flags & P_LEAF2) {
5534 nodekey.mv_size = mc->mc_db->md_pad;
5535 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5537 leaf = NODEPTR(mp, 0);
5538 MDB_GET_KEY2(leaf, nodekey);
5540 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5542 /* Probably happens rarely, but first node on the page
5543 * was the one we wanted.
5545 mc->mc_ki[mc->mc_top] = 0;
5552 unsigned int nkeys = NUMKEYS(mp);
5554 if (mp->mp_flags & P_LEAF2) {
5555 nodekey.mv_data = LEAF2KEY(mp,
5556 nkeys-1, nodekey.mv_size);
5558 leaf = NODEPTR(mp, nkeys-1);
5559 MDB_GET_KEY2(leaf, nodekey);
5561 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5563 /* last node was the one we wanted */
5564 mc->mc_ki[mc->mc_top] = nkeys-1;
5570 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5571 /* This is definitely the right page, skip search_page */
5572 if (mp->mp_flags & P_LEAF2) {
5573 nodekey.mv_data = LEAF2KEY(mp,
5574 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5576 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5577 MDB_GET_KEY2(leaf, nodekey);
5579 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5581 /* current node was the one we wanted */
5591 /* If any parents have right-sibs, search.
5592 * Otherwise, there's nothing further.
5594 for (i=0; i<mc->mc_top; i++)
5596 NUMKEYS(mc->mc_pg[i])-1)
5598 if (i == mc->mc_top) {
5599 /* There are no other pages */
5600 mc->mc_ki[mc->mc_top] = nkeys;
5601 return MDB_NOTFOUND;
5605 /* There are no other pages */
5606 mc->mc_ki[mc->mc_top] = 0;
5607 if (op == MDB_SET_RANGE && !exactp) {
5611 return MDB_NOTFOUND;
5615 rc = mdb_page_search(mc, key, 0);
5616 if (rc != MDB_SUCCESS)
5619 mp = mc->mc_pg[mc->mc_top];
5620 mdb_cassert(mc, IS_LEAF(mp));
5623 leaf = mdb_node_search(mc, key, exactp);
5624 if (exactp != NULL && !*exactp) {
5625 /* MDB_SET specified and not an exact match. */
5626 return MDB_NOTFOUND;
5630 DPUTS("===> inexact leaf not found, goto sibling");
5631 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5632 return rc; /* no entries matched */
5633 mp = mc->mc_pg[mc->mc_top];
5634 mdb_cassert(mc, IS_LEAF(mp));
5635 leaf = NODEPTR(mp, 0);
5639 mc->mc_flags |= C_INITIALIZED;
5640 mc->mc_flags &= ~C_EOF;
5643 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5644 key->mv_size = mc->mc_db->md_pad;
5645 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5650 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5651 mdb_xcursor_init1(mc, leaf);
5654 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5655 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5656 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5659 if (op == MDB_GET_BOTH) {
5665 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5666 if (rc != MDB_SUCCESS)
5669 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5671 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5673 rc = mc->mc_dbx->md_dcmp(data, &d2);
5675 if (op == MDB_GET_BOTH || rc > 0)
5676 return MDB_NOTFOUND;
5683 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5684 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5689 /* The key already matches in all other cases */
5690 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5691 MDB_GET_KEY(leaf, key);
5692 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5697 /** Move the cursor to the first item in the database. */
5699 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5705 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5707 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5708 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5709 if (rc != MDB_SUCCESS)
5712 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5714 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5715 mc->mc_flags |= C_INITIALIZED;
5716 mc->mc_flags &= ~C_EOF;
5718 mc->mc_ki[mc->mc_top] = 0;
5720 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5721 key->mv_size = mc->mc_db->md_pad;
5722 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5727 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5728 mdb_xcursor_init1(mc, leaf);
5729 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5733 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5737 MDB_GET_KEY(leaf, key);
5741 /** Move the cursor to the last item in the database. */
5743 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5749 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5751 if (!(mc->mc_flags & C_EOF)) {
5753 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5754 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5755 if (rc != MDB_SUCCESS)
5758 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5761 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5762 mc->mc_flags |= C_INITIALIZED|C_EOF;
5763 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5765 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5766 key->mv_size = mc->mc_db->md_pad;
5767 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5772 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5773 mdb_xcursor_init1(mc, leaf);
5774 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5778 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5783 MDB_GET_KEY(leaf, key);
5788 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5793 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5798 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5802 case MDB_GET_CURRENT:
5803 if (!(mc->mc_flags & C_INITIALIZED)) {
5806 MDB_page *mp = mc->mc_pg[mc->mc_top];
5807 int nkeys = NUMKEYS(mp);
5808 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5809 mc->mc_ki[mc->mc_top] = nkeys;
5815 key->mv_size = mc->mc_db->md_pad;
5816 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5818 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5819 MDB_GET_KEY(leaf, key);
5821 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5822 if (mc->mc_flags & C_DEL)
5823 mdb_xcursor_init1(mc, leaf);
5824 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5826 rc = mdb_node_read(mc->mc_txn, leaf, data);
5833 case MDB_GET_BOTH_RANGE:
5838 if (mc->mc_xcursor == NULL) {
5839 rc = MDB_INCOMPATIBLE;
5849 rc = mdb_cursor_set(mc, key, data, op,
5850 op == MDB_SET_RANGE ? NULL : &exact);
5853 case MDB_GET_MULTIPLE:
5854 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5858 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5859 rc = MDB_INCOMPATIBLE;
5863 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5864 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5867 case MDB_NEXT_MULTIPLE:
5872 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5873 rc = MDB_INCOMPATIBLE;
5876 if (!(mc->mc_flags & C_INITIALIZED))
5877 rc = mdb_cursor_first(mc, key, data);
5879 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5880 if (rc == MDB_SUCCESS) {
5881 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5884 mx = &mc->mc_xcursor->mx_cursor;
5885 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5887 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5888 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5896 case MDB_NEXT_NODUP:
5897 if (!(mc->mc_flags & C_INITIALIZED))
5898 rc = mdb_cursor_first(mc, key, data);
5900 rc = mdb_cursor_next(mc, key, data, op);
5904 case MDB_PREV_NODUP:
5905 if (!(mc->mc_flags & C_INITIALIZED)) {
5906 rc = mdb_cursor_last(mc, key, data);
5909 mc->mc_flags |= C_INITIALIZED;
5910 mc->mc_ki[mc->mc_top]++;
5912 rc = mdb_cursor_prev(mc, key, data, op);
5915 rc = mdb_cursor_first(mc, key, data);
5918 mfunc = mdb_cursor_first;
5920 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5924 if (mc->mc_xcursor == NULL) {
5925 rc = MDB_INCOMPATIBLE;
5929 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5930 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5931 MDB_GET_KEY(leaf, key);
5932 rc = mdb_node_read(mc->mc_txn, leaf, data);
5936 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5940 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5943 rc = mdb_cursor_last(mc, key, data);
5946 mfunc = mdb_cursor_last;
5949 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5954 if (mc->mc_flags & C_DEL)
5955 mc->mc_flags ^= C_DEL;
5960 /** Touch all the pages in the cursor stack. Set mc_top.
5961 * Makes sure all the pages are writable, before attempting a write operation.
5962 * @param[in] mc The cursor to operate on.
5965 mdb_cursor_touch(MDB_cursor *mc)
5967 int rc = MDB_SUCCESS;
5969 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5972 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5974 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5975 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5978 *mc->mc_dbflag |= DB_DIRTY;
5983 rc = mdb_page_touch(mc);
5984 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5985 mc->mc_top = mc->mc_snum-1;
5990 /** Do not spill pages to disk if txn is getting full, may fail instead */
5991 #define MDB_NOSPILL 0x8000
5994 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5997 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5999 MDB_node *leaf = NULL;
6002 MDB_val xdata, *rdata, dkey, olddata;
6004 int do_sub = 0, insert_key, insert_data;
6005 unsigned int mcount = 0, dcount = 0, nospill;
6008 unsigned int nflags;
6011 if (mc == NULL || key == NULL)
6014 env = mc->mc_txn->mt_env;
6016 /* Check this first so counter will always be zero on any
6019 if (flags & MDB_MULTIPLE) {
6020 dcount = data[1].mv_size;
6021 data[1].mv_size = 0;
6022 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6023 return MDB_INCOMPATIBLE;
6026 nospill = flags & MDB_NOSPILL;
6027 flags &= ~MDB_NOSPILL;
6029 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6030 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6032 if (key->mv_size-1 >= ENV_MAXKEY(env))
6033 return MDB_BAD_VALSIZE;
6035 #if SIZE_MAX > MAXDATASIZE
6036 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6037 return MDB_BAD_VALSIZE;
6039 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6040 return MDB_BAD_VALSIZE;
6043 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6044 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6048 if (flags == MDB_CURRENT) {
6049 if (!(mc->mc_flags & C_INITIALIZED))
6052 } else if (mc->mc_db->md_root == P_INVALID) {
6053 /* new database, cursor has nothing to point to */
6056 mc->mc_flags &= ~C_INITIALIZED;
6061 if (flags & MDB_APPEND) {
6063 rc = mdb_cursor_last(mc, &k2, &d2);
6065 rc = mc->mc_dbx->md_cmp(key, &k2);
6068 mc->mc_ki[mc->mc_top]++;
6070 /* new key is <= last key */
6075 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6077 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6078 DPRINTF(("duplicate key [%s]", DKEY(key)));
6080 return MDB_KEYEXIST;
6082 if (rc && rc != MDB_NOTFOUND)
6086 if (mc->mc_flags & C_DEL)
6087 mc->mc_flags ^= C_DEL;
6089 /* Cursor is positioned, check for room in the dirty list */
6091 if (flags & MDB_MULTIPLE) {
6093 xdata.mv_size = data->mv_size * dcount;
6097 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6101 if (rc == MDB_NO_ROOT) {
6103 /* new database, write a root leaf page */
6104 DPUTS("allocating new root leaf page");
6105 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6108 mdb_cursor_push(mc, np);
6109 mc->mc_db->md_root = np->mp_pgno;
6110 mc->mc_db->md_depth++;
6111 *mc->mc_dbflag |= DB_DIRTY;
6112 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6114 np->mp_flags |= P_LEAF2;
6115 mc->mc_flags |= C_INITIALIZED;
6117 /* make sure all cursor pages are writable */
6118 rc2 = mdb_cursor_touch(mc);
6123 insert_key = insert_data = rc;
6125 /* The key does not exist */
6126 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6127 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6128 LEAFSIZE(key, data) > env->me_nodemax)
6130 /* Too big for a node, insert in sub-DB. Set up an empty
6131 * "old sub-page" for prep_subDB to expand to a full page.
6133 fp_flags = P_LEAF|P_DIRTY;
6135 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6136 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6137 olddata.mv_size = PAGEHDRSZ;
6141 /* there's only a key anyway, so this is a no-op */
6142 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6144 unsigned int ksize = mc->mc_db->md_pad;
6145 if (key->mv_size != ksize)
6146 return MDB_BAD_VALSIZE;
6147 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6148 memcpy(ptr, key->mv_data, ksize);
6150 /* if overwriting slot 0 of leaf, need to
6151 * update branch key if there is a parent page
6153 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6154 unsigned short top = mc->mc_top;
6156 /* slot 0 is always an empty key, find real slot */
6157 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6159 if (mc->mc_ki[mc->mc_top])
6160 rc2 = mdb_update_key(mc, key);
6171 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6172 olddata.mv_size = NODEDSZ(leaf);
6173 olddata.mv_data = NODEDATA(leaf);
6176 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6177 /* Prepare (sub-)page/sub-DB to accept the new item,
6178 * if needed. fp: old sub-page or a header faking
6179 * it. mp: new (sub-)page. offset: growth in page
6180 * size. xdata: node data with new page or DB.
6182 unsigned i, offset = 0;
6183 mp = fp = xdata.mv_data = env->me_pbuf;
6184 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6186 /* Was a single item before, must convert now */
6187 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6188 /* Just overwrite the current item */
6189 if (flags == MDB_CURRENT)
6192 #if UINT_MAX < SIZE_MAX
6193 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6194 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6196 /* does data match? */
6197 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6198 if (flags & MDB_NODUPDATA)
6199 return MDB_KEYEXIST;
6204 /* Back up original data item */
6205 dkey.mv_size = olddata.mv_size;
6206 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6208 /* Make sub-page header for the dup items, with dummy body */
6209 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6210 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6211 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6212 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6213 fp->mp_flags |= P_LEAF2;
6214 fp->mp_pad = data->mv_size;
6215 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6217 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6218 (dkey.mv_size & 1) + (data->mv_size & 1);
6220 fp->mp_upper = xdata.mv_size - PAGEBASE;
6221 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6222 } else if (leaf->mn_flags & F_SUBDATA) {
6223 /* Data is on sub-DB, just store it */
6224 flags |= F_DUPDATA|F_SUBDATA;
6227 /* Data is on sub-page */
6228 fp = olddata.mv_data;
6231 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6232 offset = EVEN(NODESIZE + sizeof(indx_t) +
6236 offset = fp->mp_pad;
6237 if (SIZELEFT(fp) < offset) {
6238 offset *= 4; /* space for 4 more */
6241 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6243 fp->mp_flags |= P_DIRTY;
6244 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6245 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6249 xdata.mv_size = olddata.mv_size + offset;
6252 fp_flags = fp->mp_flags;
6253 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6254 /* Too big for a sub-page, convert to sub-DB */
6255 fp_flags &= ~P_SUBP;
6257 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6258 fp_flags |= P_LEAF2;
6259 dummy.md_pad = fp->mp_pad;
6260 dummy.md_flags = MDB_DUPFIXED;
6261 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6262 dummy.md_flags |= MDB_INTEGERKEY;
6268 dummy.md_branch_pages = 0;
6269 dummy.md_leaf_pages = 1;
6270 dummy.md_overflow_pages = 0;
6271 dummy.md_entries = NUMKEYS(fp);
6272 xdata.mv_size = sizeof(MDB_db);
6273 xdata.mv_data = &dummy;
6274 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6276 offset = env->me_psize - olddata.mv_size;
6277 flags |= F_DUPDATA|F_SUBDATA;
6278 dummy.md_root = mp->mp_pgno;
6281 mp->mp_flags = fp_flags | P_DIRTY;
6282 mp->mp_pad = fp->mp_pad;
6283 mp->mp_lower = fp->mp_lower;
6284 mp->mp_upper = fp->mp_upper + offset;
6285 if (fp_flags & P_LEAF2) {
6286 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6288 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6289 olddata.mv_size - fp->mp_upper - PAGEBASE);
6290 for (i=0; i<NUMKEYS(fp); i++)
6291 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6299 mdb_node_del(mc, 0);
6303 /* overflow page overwrites need special handling */
6304 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6307 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6309 memcpy(&pg, olddata.mv_data, sizeof(pg));
6310 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6312 ovpages = omp->mp_pages;
6314 /* Is the ov page large enough? */
6315 if (ovpages >= dpages) {
6316 if (!(omp->mp_flags & P_DIRTY) &&
6317 (level || (env->me_flags & MDB_WRITEMAP)))
6319 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6322 level = 0; /* dirty in this txn or clean */
6325 if (omp->mp_flags & P_DIRTY) {
6326 /* yes, overwrite it. Note in this case we don't
6327 * bother to try shrinking the page if the new data
6328 * is smaller than the overflow threshold.
6331 /* It is writable only in a parent txn */
6332 size_t sz = (size_t) env->me_psize * ovpages, off;
6333 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6339 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6340 mdb_cassert(mc, rc2 == 0);
6341 if (!(flags & MDB_RESERVE)) {
6342 /* Copy end of page, adjusting alignment so
6343 * compiler may copy words instead of bytes.
6345 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6346 memcpy((size_t *)((char *)np + off),
6347 (size_t *)((char *)omp + off), sz - off);
6350 memcpy(np, omp, sz); /* Copy beginning of page */
6353 SETDSZ(leaf, data->mv_size);
6354 if (F_ISSET(flags, MDB_RESERVE))
6355 data->mv_data = METADATA(omp);
6357 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6361 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6363 } else if (data->mv_size == olddata.mv_size) {
6364 /* same size, just replace it. Note that we could
6365 * also reuse this node if the new data is smaller,
6366 * but instead we opt to shrink the node in that case.
6368 if (F_ISSET(flags, MDB_RESERVE))
6369 data->mv_data = olddata.mv_data;
6370 else if (!(mc->mc_flags & C_SUB))
6371 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6373 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6378 mdb_node_del(mc, 0);
6384 nflags = flags & NODE_ADD_FLAGS;
6385 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6386 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6387 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6388 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6390 nflags |= MDB_SPLIT_REPLACE;
6391 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6393 /* There is room already in this leaf page. */
6394 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6395 if (rc == 0 && insert_key) {
6396 /* Adjust other cursors pointing to mp */
6397 MDB_cursor *m2, *m3;
6398 MDB_dbi dbi = mc->mc_dbi;
6399 unsigned i = mc->mc_top;
6400 MDB_page *mp = mc->mc_pg[i];
6402 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6403 if (mc->mc_flags & C_SUB)
6404 m3 = &m2->mc_xcursor->mx_cursor;
6407 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6408 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6415 if (rc == MDB_SUCCESS) {
6416 /* Now store the actual data in the child DB. Note that we're
6417 * storing the user data in the keys field, so there are strict
6418 * size limits on dupdata. The actual data fields of the child
6419 * DB are all zero size.
6427 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6428 if (flags & MDB_CURRENT) {
6429 xflags = MDB_CURRENT|MDB_NOSPILL;
6431 mdb_xcursor_init1(mc, leaf);
6432 xflags = (flags & MDB_NODUPDATA) ?
6433 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6435 /* converted, write the original data first */
6437 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6441 /* Adjust other cursors pointing to mp */
6443 unsigned i = mc->mc_top;
6444 MDB_page *mp = mc->mc_pg[i];
6446 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6447 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6448 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6449 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6450 mdb_xcursor_init1(m2, leaf);
6454 /* we've done our job */
6457 ecount = mc->mc_xcursor->mx_db.md_entries;
6458 if (flags & MDB_APPENDDUP)
6459 xflags |= MDB_APPEND;
6460 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6461 if (flags & F_SUBDATA) {
6462 void *db = NODEDATA(leaf);
6463 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6465 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6467 /* Increment count unless we just replaced an existing item. */
6469 mc->mc_db->md_entries++;
6471 /* Invalidate txn if we created an empty sub-DB */
6474 /* If we succeeded and the key didn't exist before,
6475 * make sure the cursor is marked valid.
6477 mc->mc_flags |= C_INITIALIZED;
6479 if (flags & MDB_MULTIPLE) {
6482 /* let caller know how many succeeded, if any */
6483 data[1].mv_size = mcount;
6484 if (mcount < dcount) {
6485 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6486 insert_key = insert_data = 0;
6493 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6496 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6501 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6507 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6508 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6510 if (!(mc->mc_flags & C_INITIALIZED))
6513 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6514 return MDB_NOTFOUND;
6516 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6519 rc = mdb_cursor_touch(mc);
6523 mp = mc->mc_pg[mc->mc_top];
6526 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6528 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6529 if (flags & MDB_NODUPDATA) {
6530 /* mdb_cursor_del0() will subtract the final entry */
6531 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6533 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6534 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6536 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6539 /* If sub-DB still has entries, we're done */
6540 if (mc->mc_xcursor->mx_db.md_entries) {
6541 if (leaf->mn_flags & F_SUBDATA) {
6542 /* update subDB info */
6543 void *db = NODEDATA(leaf);
6544 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6547 /* shrink fake page */
6548 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6549 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6550 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6551 /* fix other sub-DB cursors pointed at this fake page */
6552 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6553 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6554 if (m2->mc_pg[mc->mc_top] == mp &&
6555 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6556 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6559 mc->mc_db->md_entries--;
6560 mc->mc_flags |= C_DEL;
6563 /* otherwise fall thru and delete the sub-DB */
6566 if (leaf->mn_flags & F_SUBDATA) {
6567 /* add all the child DB's pages to the free list */
6568 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6574 /* add overflow pages to free list */
6575 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6579 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6580 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6581 (rc = mdb_ovpage_free(mc, omp)))
6586 return mdb_cursor_del0(mc);
6589 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6593 /** Allocate and initialize new pages for a database.
6594 * @param[in] mc a cursor on the database being added to.
6595 * @param[in] flags flags defining what type of page is being allocated.
6596 * @param[in] num the number of pages to allocate. This is usually 1,
6597 * unless allocating overflow pages for a large record.
6598 * @param[out] mp Address of a page, or NULL on failure.
6599 * @return 0 on success, non-zero on failure.
6602 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6607 if ((rc = mdb_page_alloc(mc, num, &np)))
6609 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6610 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6611 np->mp_flags = flags | P_DIRTY;
6612 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6613 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6616 mc->mc_db->md_branch_pages++;
6617 else if (IS_LEAF(np))
6618 mc->mc_db->md_leaf_pages++;
6619 else if (IS_OVERFLOW(np)) {
6620 mc->mc_db->md_overflow_pages += num;
6628 /** Calculate the size of a leaf node.
6629 * The size depends on the environment's page size; if a data item
6630 * is too large it will be put onto an overflow page and the node
6631 * size will only include the key and not the data. Sizes are always
6632 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6633 * of the #MDB_node headers.
6634 * @param[in] env The environment handle.
6635 * @param[in] key The key for the node.
6636 * @param[in] data The data for the node.
6637 * @return The number of bytes needed to store the node.
6640 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6644 sz = LEAFSIZE(key, data);
6645 if (sz > env->me_nodemax) {
6646 /* put on overflow page */
6647 sz -= data->mv_size - sizeof(pgno_t);
6650 return EVEN(sz + sizeof(indx_t));
6653 /** Calculate the size of a branch node.
6654 * The size should depend on the environment's page size but since
6655 * we currently don't support spilling large keys onto overflow
6656 * pages, it's simply the size of the #MDB_node header plus the
6657 * size of the key. Sizes are always rounded up to an even number
6658 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6659 * @param[in] env The environment handle.
6660 * @param[in] key The key for the node.
6661 * @return The number of bytes needed to store the node.
6664 mdb_branch_size(MDB_env *env, MDB_val *key)
6669 if (sz > env->me_nodemax) {
6670 /* put on overflow page */
6671 /* not implemented */
6672 /* sz -= key->size - sizeof(pgno_t); */
6675 return sz + sizeof(indx_t);
6678 /** Add a node to the page pointed to by the cursor.
6679 * @param[in] mc The cursor for this operation.
6680 * @param[in] indx The index on the page where the new node should be added.
6681 * @param[in] key The key for the new node.
6682 * @param[in] data The data for the new node, if any.
6683 * @param[in] pgno The page number, if adding a branch node.
6684 * @param[in] flags Flags for the node.
6685 * @return 0 on success, non-zero on failure. Possible errors are:
6687 * <li>ENOMEM - failed to allocate overflow pages for the node.
6688 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6689 * should never happen since all callers already calculate the
6690 * page's free space before calling this function.
6694 mdb_node_add(MDB_cursor *mc, indx_t indx,
6695 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6698 size_t node_size = NODESIZE;
6702 MDB_page *mp = mc->mc_pg[mc->mc_top];
6703 MDB_page *ofp = NULL; /* overflow page */
6706 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6708 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6709 IS_LEAF(mp) ? "leaf" : "branch",
6710 IS_SUBP(mp) ? "sub-" : "",
6711 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6712 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6715 /* Move higher keys up one slot. */
6716 int ksize = mc->mc_db->md_pad, dif;
6717 char *ptr = LEAF2KEY(mp, indx, ksize);
6718 dif = NUMKEYS(mp) - indx;
6720 memmove(ptr+ksize, ptr, dif*ksize);
6721 /* insert new key */
6722 memcpy(ptr, key->mv_data, ksize);
6724 /* Just using these for counting */
6725 mp->mp_lower += sizeof(indx_t);
6726 mp->mp_upper -= ksize - sizeof(indx_t);
6730 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6732 node_size += key->mv_size;
6734 mdb_cassert(mc, data);
6735 if (F_ISSET(flags, F_BIGDATA)) {
6736 /* Data already on overflow page. */
6737 node_size += sizeof(pgno_t);
6738 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6739 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6741 /* Put data on overflow page. */
6742 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6743 data->mv_size, node_size+data->mv_size));
6744 node_size = EVEN(node_size + sizeof(pgno_t));
6745 if ((ssize_t)node_size > room)
6747 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6749 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6753 node_size += data->mv_size;
6756 node_size = EVEN(node_size);
6757 if ((ssize_t)node_size > room)
6761 /* Move higher pointers up one slot. */
6762 for (i = NUMKEYS(mp); i > indx; i--)
6763 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6765 /* Adjust free space offsets. */
6766 ofs = mp->mp_upper - node_size;
6767 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6768 mp->mp_ptrs[indx] = ofs;
6770 mp->mp_lower += sizeof(indx_t);
6772 /* Write the node data. */
6773 node = NODEPTR(mp, indx);
6774 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6775 node->mn_flags = flags;
6777 SETDSZ(node,data->mv_size);
6782 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6785 mdb_cassert(mc, key);
6787 if (F_ISSET(flags, F_BIGDATA))
6788 memcpy(node->mn_data + key->mv_size, data->mv_data,
6790 else if (F_ISSET(flags, MDB_RESERVE))
6791 data->mv_data = node->mn_data + key->mv_size;
6793 memcpy(node->mn_data + key->mv_size, data->mv_data,
6796 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6798 if (F_ISSET(flags, MDB_RESERVE))
6799 data->mv_data = METADATA(ofp);
6801 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6808 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6809 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6810 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6811 DPRINTF(("node size = %"Z"u", node_size));
6812 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6813 return MDB_PAGE_FULL;
6816 /** Delete the specified node from a page.
6817 * @param[in] mc Cursor pointing to the node to delete.
6818 * @param[in] ksize The size of a node. Only used if the page is
6819 * part of a #MDB_DUPFIXED database.
6822 mdb_node_del(MDB_cursor *mc, int ksize)
6824 MDB_page *mp = mc->mc_pg[mc->mc_top];
6825 indx_t indx = mc->mc_ki[mc->mc_top];
6827 indx_t i, j, numkeys, ptr;
6831 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6832 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6833 numkeys = NUMKEYS(mp);
6834 mdb_cassert(mc, indx < numkeys);
6837 int x = numkeys - 1 - indx;
6838 base = LEAF2KEY(mp, indx, ksize);
6840 memmove(base, base + ksize, x * ksize);
6841 mp->mp_lower -= sizeof(indx_t);
6842 mp->mp_upper += ksize - sizeof(indx_t);
6846 node = NODEPTR(mp, indx);
6847 sz = NODESIZE + node->mn_ksize;
6849 if (F_ISSET(node->mn_flags, F_BIGDATA))
6850 sz += sizeof(pgno_t);
6852 sz += NODEDSZ(node);
6856 ptr = mp->mp_ptrs[indx];
6857 for (i = j = 0; i < numkeys; i++) {
6859 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6860 if (mp->mp_ptrs[i] < ptr)
6861 mp->mp_ptrs[j] += sz;
6866 base = (char *)mp + mp->mp_upper + PAGEBASE;
6867 memmove(base + sz, base, ptr - mp->mp_upper);
6869 mp->mp_lower -= sizeof(indx_t);
6873 /** Compact the main page after deleting a node on a subpage.
6874 * @param[in] mp The main page to operate on.
6875 * @param[in] indx The index of the subpage on the main page.
6878 mdb_node_shrink(MDB_page *mp, indx_t indx)
6884 indx_t i, numkeys, ptr;
6886 node = NODEPTR(mp, indx);
6887 sp = (MDB_page *)NODEDATA(node);
6888 delta = SIZELEFT(sp);
6889 xp = (MDB_page *)((char *)sp + delta);
6891 /* shift subpage upward */
6893 nsize = NUMKEYS(sp) * sp->mp_pad;
6895 return; /* do not make the node uneven-sized */
6896 memmove(METADATA(xp), METADATA(sp), nsize);
6899 numkeys = NUMKEYS(sp);
6900 for (i=numkeys-1; i>=0; i--)
6901 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6903 xp->mp_upper = sp->mp_lower;
6904 xp->mp_lower = sp->mp_lower;
6905 xp->mp_flags = sp->mp_flags;
6906 xp->mp_pad = sp->mp_pad;
6907 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6909 nsize = NODEDSZ(node) - delta;
6910 SETDSZ(node, nsize);
6912 /* shift lower nodes upward */
6913 ptr = mp->mp_ptrs[indx];
6914 numkeys = NUMKEYS(mp);
6915 for (i = 0; i < numkeys; i++) {
6916 if (mp->mp_ptrs[i] <= ptr)
6917 mp->mp_ptrs[i] += delta;
6920 base = (char *)mp + mp->mp_upper + PAGEBASE;
6921 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6922 mp->mp_upper += delta;
6925 /** Initial setup of a sorted-dups cursor.
6926 * Sorted duplicates are implemented as a sub-database for the given key.
6927 * The duplicate data items are actually keys of the sub-database.
6928 * Operations on the duplicate data items are performed using a sub-cursor
6929 * initialized when the sub-database is first accessed. This function does
6930 * the preliminary setup of the sub-cursor, filling in the fields that
6931 * depend only on the parent DB.
6932 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6935 mdb_xcursor_init0(MDB_cursor *mc)
6937 MDB_xcursor *mx = mc->mc_xcursor;
6939 mx->mx_cursor.mc_xcursor = NULL;
6940 mx->mx_cursor.mc_txn = mc->mc_txn;
6941 mx->mx_cursor.mc_db = &mx->mx_db;
6942 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6943 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6944 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6945 mx->mx_cursor.mc_snum = 0;
6946 mx->mx_cursor.mc_top = 0;
6947 mx->mx_cursor.mc_flags = C_SUB;
6948 mx->mx_dbx.md_name.mv_size = 0;
6949 mx->mx_dbx.md_name.mv_data = NULL;
6950 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6951 mx->mx_dbx.md_dcmp = NULL;
6952 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6955 /** Final setup of a sorted-dups cursor.
6956 * Sets up the fields that depend on the data from the main cursor.
6957 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6958 * @param[in] node The data containing the #MDB_db record for the
6959 * sorted-dup database.
6962 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6964 MDB_xcursor *mx = mc->mc_xcursor;
6966 if (node->mn_flags & F_SUBDATA) {
6967 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6968 mx->mx_cursor.mc_pg[0] = 0;
6969 mx->mx_cursor.mc_snum = 0;
6970 mx->mx_cursor.mc_top = 0;
6971 mx->mx_cursor.mc_flags = C_SUB;
6973 MDB_page *fp = NODEDATA(node);
6974 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6975 mx->mx_db.md_flags = 0;
6976 mx->mx_db.md_depth = 1;
6977 mx->mx_db.md_branch_pages = 0;
6978 mx->mx_db.md_leaf_pages = 1;
6979 mx->mx_db.md_overflow_pages = 0;
6980 mx->mx_db.md_entries = NUMKEYS(fp);
6981 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6982 mx->mx_cursor.mc_snum = 1;
6983 mx->mx_cursor.mc_top = 0;
6984 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6985 mx->mx_cursor.mc_pg[0] = fp;
6986 mx->mx_cursor.mc_ki[0] = 0;
6987 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6988 mx->mx_db.md_flags = MDB_DUPFIXED;
6989 mx->mx_db.md_pad = fp->mp_pad;
6990 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6991 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6994 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6995 mx->mx_db.md_root));
6996 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6997 #if UINT_MAX < SIZE_MAX
6998 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6999 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7003 /** Initialize a cursor for a given transaction and database. */
7005 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7008 mc->mc_backup = NULL;
7011 mc->mc_db = &txn->mt_dbs[dbi];
7012 mc->mc_dbx = &txn->mt_dbxs[dbi];
7013 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7018 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7019 mdb_tassert(txn, mx != NULL);
7020 mc->mc_xcursor = mx;
7021 mdb_xcursor_init0(mc);
7023 mc->mc_xcursor = NULL;
7025 if (*mc->mc_dbflag & DB_STALE) {
7026 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7031 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7034 size_t size = sizeof(MDB_cursor);
7036 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7039 if (txn->mt_flags & MDB_TXN_ERROR)
7042 /* Allow read access to the freelist */
7043 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7046 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7047 size += sizeof(MDB_xcursor);
7049 if ((mc = malloc(size)) != NULL) {
7050 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7051 if (txn->mt_cursors) {
7052 mc->mc_next = txn->mt_cursors[dbi];
7053 txn->mt_cursors[dbi] = mc;
7054 mc->mc_flags |= C_UNTRACK;
7066 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7068 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7071 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7074 if (txn->mt_flags & MDB_TXN_ERROR)
7077 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7081 /* Return the count of duplicate data items for the current key */
7083 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7087 if (mc == NULL || countp == NULL)
7090 if (mc->mc_xcursor == NULL)
7091 return MDB_INCOMPATIBLE;
7093 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7096 if (!(mc->mc_flags & C_INITIALIZED))
7099 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7100 return MDB_NOTFOUND;
7102 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7103 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7106 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7109 *countp = mc->mc_xcursor->mx_db.md_entries;
7115 mdb_cursor_close(MDB_cursor *mc)
7117 if (mc && !mc->mc_backup) {
7118 /* remove from txn, if tracked */
7119 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7120 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7121 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7123 *prev = mc->mc_next;
7130 mdb_cursor_txn(MDB_cursor *mc)
7132 if (!mc) return NULL;
7137 mdb_cursor_dbi(MDB_cursor *mc)
7142 /** Replace the key for a branch node with a new key.
7143 * @param[in] mc Cursor pointing to the node to operate on.
7144 * @param[in] key The new key to use.
7145 * @return 0 on success, non-zero on failure.
7148 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7154 int delta, ksize, oksize;
7155 indx_t ptr, i, numkeys, indx;
7158 indx = mc->mc_ki[mc->mc_top];
7159 mp = mc->mc_pg[mc->mc_top];
7160 node = NODEPTR(mp, indx);
7161 ptr = mp->mp_ptrs[indx];
7165 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7166 k2.mv_data = NODEKEY(node);
7167 k2.mv_size = node->mn_ksize;
7168 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7170 mdb_dkey(&k2, kbuf2),
7176 /* Sizes must be 2-byte aligned. */
7177 ksize = EVEN(key->mv_size);
7178 oksize = EVEN(node->mn_ksize);
7179 delta = ksize - oksize;
7181 /* Shift node contents if EVEN(key length) changed. */
7183 if (delta > 0 && SIZELEFT(mp) < delta) {
7185 /* not enough space left, do a delete and split */
7186 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7187 pgno = NODEPGNO(node);
7188 mdb_node_del(mc, 0);
7189 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7192 numkeys = NUMKEYS(mp);
7193 for (i = 0; i < numkeys; i++) {
7194 if (mp->mp_ptrs[i] <= ptr)
7195 mp->mp_ptrs[i] -= delta;
7198 base = (char *)mp + mp->mp_upper + PAGEBASE;
7199 len = ptr - mp->mp_upper + NODESIZE;
7200 memmove(base - delta, base, len);
7201 mp->mp_upper -= delta;
7203 node = NODEPTR(mp, indx);
7206 /* But even if no shift was needed, update ksize */
7207 if (node->mn_ksize != key->mv_size)
7208 node->mn_ksize = key->mv_size;
7211 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7217 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7219 /** Move a node from csrc to cdst.
7222 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7229 unsigned short flags;
7233 /* Mark src and dst as dirty. */
7234 if ((rc = mdb_page_touch(csrc)) ||
7235 (rc = mdb_page_touch(cdst)))
7238 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7239 key.mv_size = csrc->mc_db->md_pad;
7240 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7242 data.mv_data = NULL;
7246 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7247 mdb_cassert(csrc, !((size_t)srcnode & 1));
7248 srcpg = NODEPGNO(srcnode);
7249 flags = srcnode->mn_flags;
7250 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7251 unsigned int snum = csrc->mc_snum;
7253 /* must find the lowest key below src */
7254 rc = mdb_page_search_lowest(csrc);
7257 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7258 key.mv_size = csrc->mc_db->md_pad;
7259 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7261 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7262 key.mv_size = NODEKSZ(s2);
7263 key.mv_data = NODEKEY(s2);
7265 csrc->mc_snum = snum--;
7266 csrc->mc_top = snum;
7268 key.mv_size = NODEKSZ(srcnode);
7269 key.mv_data = NODEKEY(srcnode);
7271 data.mv_size = NODEDSZ(srcnode);
7272 data.mv_data = NODEDATA(srcnode);
7274 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7275 unsigned int snum = cdst->mc_snum;
7278 /* must find the lowest key below dst */
7279 mdb_cursor_copy(cdst, &mn);
7280 rc = mdb_page_search_lowest(&mn);
7283 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7284 bkey.mv_size = mn.mc_db->md_pad;
7285 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7287 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7288 bkey.mv_size = NODEKSZ(s2);
7289 bkey.mv_data = NODEKEY(s2);
7291 mn.mc_snum = snum--;
7294 rc = mdb_update_key(&mn, &bkey);
7299 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7300 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7301 csrc->mc_ki[csrc->mc_top],
7303 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7304 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7306 /* Add the node to the destination page.
7308 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7309 if (rc != MDB_SUCCESS)
7312 /* Delete the node from the source page.
7314 mdb_node_del(csrc, key.mv_size);
7317 /* Adjust other cursors pointing to mp */
7318 MDB_cursor *m2, *m3;
7319 MDB_dbi dbi = csrc->mc_dbi;
7320 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7322 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7323 if (csrc->mc_flags & C_SUB)
7324 m3 = &m2->mc_xcursor->mx_cursor;
7327 if (m3 == csrc) continue;
7328 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7329 csrc->mc_ki[csrc->mc_top]) {
7330 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7331 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7336 /* Update the parent separators.
7338 if (csrc->mc_ki[csrc->mc_top] == 0) {
7339 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7340 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7341 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7343 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7344 key.mv_size = NODEKSZ(srcnode);
7345 key.mv_data = NODEKEY(srcnode);
7347 DPRINTF(("update separator for source page %"Z"u to [%s]",
7348 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7349 mdb_cursor_copy(csrc, &mn);
7352 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7355 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7357 indx_t ix = csrc->mc_ki[csrc->mc_top];
7358 nullkey.mv_size = 0;
7359 csrc->mc_ki[csrc->mc_top] = 0;
7360 rc = mdb_update_key(csrc, &nullkey);
7361 csrc->mc_ki[csrc->mc_top] = ix;
7362 mdb_cassert(csrc, rc == MDB_SUCCESS);
7366 if (cdst->mc_ki[cdst->mc_top] == 0) {
7367 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7368 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7369 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7371 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7372 key.mv_size = NODEKSZ(srcnode);
7373 key.mv_data = NODEKEY(srcnode);
7375 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7376 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7377 mdb_cursor_copy(cdst, &mn);
7380 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7383 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7385 indx_t ix = cdst->mc_ki[cdst->mc_top];
7386 nullkey.mv_size = 0;
7387 cdst->mc_ki[cdst->mc_top] = 0;
7388 rc = mdb_update_key(cdst, &nullkey);
7389 cdst->mc_ki[cdst->mc_top] = ix;
7390 mdb_cassert(csrc, rc == MDB_SUCCESS);
7397 /** Merge one page into another.
7398 * The nodes from the page pointed to by \b csrc will
7399 * be copied to the page pointed to by \b cdst and then
7400 * the \b csrc page will be freed.
7401 * @param[in] csrc Cursor pointing to the source page.
7402 * @param[in] cdst Cursor pointing to the destination page.
7403 * @return 0 on success, non-zero on failure.
7406 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7408 MDB_page *psrc, *pdst;
7415 psrc = csrc->mc_pg[csrc->mc_top];
7416 pdst = cdst->mc_pg[cdst->mc_top];
7418 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7420 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7421 mdb_cassert(csrc, cdst->mc_snum > 1);
7423 /* Mark dst as dirty. */
7424 if ((rc = mdb_page_touch(cdst)))
7427 /* Move all nodes from src to dst.
7429 j = nkeys = NUMKEYS(pdst);
7430 if (IS_LEAF2(psrc)) {
7431 key.mv_size = csrc->mc_db->md_pad;
7432 key.mv_data = METADATA(psrc);
7433 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7434 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7435 if (rc != MDB_SUCCESS)
7437 key.mv_data = (char *)key.mv_data + key.mv_size;
7440 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7441 srcnode = NODEPTR(psrc, i);
7442 if (i == 0 && IS_BRANCH(psrc)) {
7445 mdb_cursor_copy(csrc, &mn);
7446 /* must find the lowest key below src */
7447 rc = mdb_page_search_lowest(&mn);
7450 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7451 key.mv_size = mn.mc_db->md_pad;
7452 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7454 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7455 key.mv_size = NODEKSZ(s2);
7456 key.mv_data = NODEKEY(s2);
7459 key.mv_size = srcnode->mn_ksize;
7460 key.mv_data = NODEKEY(srcnode);
7463 data.mv_size = NODEDSZ(srcnode);
7464 data.mv_data = NODEDATA(srcnode);
7465 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7466 if (rc != MDB_SUCCESS)
7471 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7472 pdst->mp_pgno, NUMKEYS(pdst),
7473 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7475 /* Unlink the src page from parent and add to free list.
7478 mdb_node_del(csrc, 0);
7479 if (csrc->mc_ki[csrc->mc_top] == 0) {
7481 rc = mdb_update_key(csrc, &key);
7489 psrc = csrc->mc_pg[csrc->mc_top];
7490 /* If not operating on FreeDB, allow this page to be reused
7491 * in this txn. Otherwise just add to free list.
7493 rc = mdb_page_loose(csrc, psrc);
7497 csrc->mc_db->md_leaf_pages--;
7499 csrc->mc_db->md_branch_pages--;
7501 /* Adjust other cursors pointing to mp */
7502 MDB_cursor *m2, *m3;
7503 MDB_dbi dbi = csrc->mc_dbi;
7505 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7506 if (csrc->mc_flags & C_SUB)
7507 m3 = &m2->mc_xcursor->mx_cursor;
7510 if (m3 == csrc) continue;
7511 if (m3->mc_snum < csrc->mc_snum) continue;
7512 if (m3->mc_pg[csrc->mc_top] == psrc) {
7513 m3->mc_pg[csrc->mc_top] = pdst;
7514 m3->mc_ki[csrc->mc_top] += nkeys;
7519 unsigned int snum = cdst->mc_snum;
7520 uint16_t depth = cdst->mc_db->md_depth;
7521 mdb_cursor_pop(cdst);
7522 rc = mdb_rebalance(cdst);
7523 /* Did the tree shrink? */
7524 if (depth > cdst->mc_db->md_depth)
7526 cdst->mc_snum = snum;
7527 cdst->mc_top = snum-1;
7532 /** Copy the contents of a cursor.
7533 * @param[in] csrc The cursor to copy from.
7534 * @param[out] cdst The cursor to copy to.
7537 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7541 cdst->mc_txn = csrc->mc_txn;
7542 cdst->mc_dbi = csrc->mc_dbi;
7543 cdst->mc_db = csrc->mc_db;
7544 cdst->mc_dbx = csrc->mc_dbx;
7545 cdst->mc_snum = csrc->mc_snum;
7546 cdst->mc_top = csrc->mc_top;
7547 cdst->mc_flags = csrc->mc_flags;
7549 for (i=0; i<csrc->mc_snum; i++) {
7550 cdst->mc_pg[i] = csrc->mc_pg[i];
7551 cdst->mc_ki[i] = csrc->mc_ki[i];
7555 /** Rebalance the tree after a delete operation.
7556 * @param[in] mc Cursor pointing to the page where rebalancing
7558 * @return 0 on success, non-zero on failure.
7561 mdb_rebalance(MDB_cursor *mc)
7565 unsigned int ptop, minkeys;
7569 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7570 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7571 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7572 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7573 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7575 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7576 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7577 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7578 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7582 if (mc->mc_snum < 2) {
7583 MDB_page *mp = mc->mc_pg[0];
7585 DPUTS("Can't rebalance a subpage, ignoring");
7588 if (NUMKEYS(mp) == 0) {
7589 DPUTS("tree is completely empty");
7590 mc->mc_db->md_root = P_INVALID;
7591 mc->mc_db->md_depth = 0;
7592 mc->mc_db->md_leaf_pages = 0;
7593 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7596 /* Adjust cursors pointing to mp */
7599 mc->mc_flags &= ~C_INITIALIZED;
7601 MDB_cursor *m2, *m3;
7602 MDB_dbi dbi = mc->mc_dbi;
7604 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7605 if (mc->mc_flags & C_SUB)
7606 m3 = &m2->mc_xcursor->mx_cursor;
7609 if (m3->mc_snum < mc->mc_snum) continue;
7610 if (m3->mc_pg[0] == mp) {
7613 m3->mc_flags &= ~C_INITIALIZED;
7617 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7619 DPUTS("collapsing root page!");
7620 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7623 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7624 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7627 mc->mc_db->md_depth--;
7628 mc->mc_db->md_branch_pages--;
7629 mc->mc_ki[0] = mc->mc_ki[1];
7630 for (i = 1; i<mc->mc_db->md_depth; i++) {
7631 mc->mc_pg[i] = mc->mc_pg[i+1];
7632 mc->mc_ki[i] = mc->mc_ki[i+1];
7635 /* Adjust other cursors pointing to mp */
7636 MDB_cursor *m2, *m3;
7637 MDB_dbi dbi = mc->mc_dbi;
7639 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7640 if (mc->mc_flags & C_SUB)
7641 m3 = &m2->mc_xcursor->mx_cursor;
7644 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7645 if (m3->mc_pg[0] == mp) {
7648 for (i=0; i<m3->mc_snum; i++) {
7649 m3->mc_pg[i] = m3->mc_pg[i+1];
7650 m3->mc_ki[i] = m3->mc_ki[i+1];
7656 DPUTS("root page doesn't need rebalancing");
7660 /* The parent (branch page) must have at least 2 pointers,
7661 * otherwise the tree is invalid.
7663 ptop = mc->mc_top-1;
7664 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7666 /* Leaf page fill factor is below the threshold.
7667 * Try to move keys from left or right neighbor, or
7668 * merge with a neighbor page.
7673 mdb_cursor_copy(mc, &mn);
7674 mn.mc_xcursor = NULL;
7676 oldki = mc->mc_ki[mc->mc_top];
7677 if (mc->mc_ki[ptop] == 0) {
7678 /* We're the leftmost leaf in our parent.
7680 DPUTS("reading right neighbor");
7682 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7683 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7686 mn.mc_ki[mn.mc_top] = 0;
7687 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7689 /* There is at least one neighbor to the left.
7691 DPUTS("reading left neighbor");
7693 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7694 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7697 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7698 mc->mc_ki[mc->mc_top] = 0;
7701 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7702 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7703 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7705 /* If the neighbor page is above threshold and has enough keys,
7706 * move one key from it. Otherwise we should try to merge them.
7707 * (A branch page must never have less than 2 keys.)
7709 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7710 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7711 rc = mdb_node_move(&mn, mc);
7712 if (mc->mc_ki[ptop]) {
7716 if (mc->mc_ki[ptop] == 0) {
7717 rc = mdb_page_merge(&mn, mc);
7719 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7720 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7721 rc = mdb_page_merge(mc, &mn);
7722 mdb_cursor_copy(&mn, mc);
7724 mc->mc_flags &= ~C_EOF;
7726 mc->mc_ki[mc->mc_top] = oldki;
7730 /** Complete a delete operation started by #mdb_cursor_del(). */
7732 mdb_cursor_del0(MDB_cursor *mc)
7739 ki = mc->mc_ki[mc->mc_top];
7740 mdb_node_del(mc, mc->mc_db->md_pad);
7741 mc->mc_db->md_entries--;
7742 rc = mdb_rebalance(mc);
7744 if (rc == MDB_SUCCESS) {
7745 MDB_cursor *m2, *m3;
7746 MDB_dbi dbi = mc->mc_dbi;
7748 mp = mc->mc_pg[mc->mc_top];
7749 nkeys = NUMKEYS(mp);
7751 /* if mc points past last node in page, find next sibling */
7752 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7753 rc = mdb_cursor_sibling(mc, 1);
7754 if (rc == MDB_NOTFOUND) {
7755 mc->mc_flags |= C_EOF;
7760 /* Adjust other cursors pointing to mp */
7761 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7762 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7763 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7765 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7767 if (m3->mc_pg[mc->mc_top] == mp) {
7768 if (m3->mc_ki[mc->mc_top] >= ki) {
7769 m3->mc_flags |= C_DEL;
7770 if (m3->mc_ki[mc->mc_top] > ki)
7771 m3->mc_ki[mc->mc_top]--;
7772 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7773 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7775 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7776 rc = mdb_cursor_sibling(m3, 1);
7777 if (rc == MDB_NOTFOUND) {
7778 m3->mc_flags |= C_EOF;
7784 mc->mc_flags |= C_DEL;
7788 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7793 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7794 MDB_val *key, MDB_val *data)
7796 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7799 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7800 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7802 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7803 /* must ignore any data */
7807 return mdb_del0(txn, dbi, key, data, 0);
7811 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7812 MDB_val *key, MDB_val *data, unsigned flags)
7817 MDB_val rdata, *xdata;
7821 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7823 mdb_cursor_init(&mc, txn, dbi, &mx);
7832 flags |= MDB_NODUPDATA;
7834 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7836 /* let mdb_page_split know about this cursor if needed:
7837 * delete will trigger a rebalance; if it needs to move
7838 * a node from one page to another, it will have to
7839 * update the parent's separator key(s). If the new sepkey
7840 * is larger than the current one, the parent page may
7841 * run out of space, triggering a split. We need this
7842 * cursor to be consistent until the end of the rebalance.
7844 mc.mc_flags |= C_UNTRACK;
7845 mc.mc_next = txn->mt_cursors[dbi];
7846 txn->mt_cursors[dbi] = &mc;
7847 rc = mdb_cursor_del(&mc, flags);
7848 txn->mt_cursors[dbi] = mc.mc_next;
7853 /** Split a page and insert a new node.
7854 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7855 * The cursor will be updated to point to the actual page and index where
7856 * the node got inserted after the split.
7857 * @param[in] newkey The key for the newly inserted node.
7858 * @param[in] newdata The data for the newly inserted node.
7859 * @param[in] newpgno The page number, if the new node is a branch node.
7860 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7861 * @return 0 on success, non-zero on failure.
7864 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7865 unsigned int nflags)
7868 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7871 int i, j, split_indx, nkeys, pmax;
7872 MDB_env *env = mc->mc_txn->mt_env;
7874 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7875 MDB_page *copy = NULL;
7876 MDB_page *mp, *rp, *pp;
7881 mp = mc->mc_pg[mc->mc_top];
7882 newindx = mc->mc_ki[mc->mc_top];
7883 nkeys = NUMKEYS(mp);
7885 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7886 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7887 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7889 /* Create a right sibling. */
7890 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7892 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7894 if (mc->mc_snum < 2) {
7895 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7897 /* shift current top to make room for new parent */
7898 mc->mc_pg[1] = mc->mc_pg[0];
7899 mc->mc_ki[1] = mc->mc_ki[0];
7902 mc->mc_db->md_root = pp->mp_pgno;
7903 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7904 mc->mc_db->md_depth++;
7907 /* Add left (implicit) pointer. */
7908 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7909 /* undo the pre-push */
7910 mc->mc_pg[0] = mc->mc_pg[1];
7911 mc->mc_ki[0] = mc->mc_ki[1];
7912 mc->mc_db->md_root = mp->mp_pgno;
7913 mc->mc_db->md_depth--;
7920 ptop = mc->mc_top-1;
7921 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7924 mc->mc_flags |= C_SPLITTING;
7925 mdb_cursor_copy(mc, &mn);
7926 mn.mc_pg[mn.mc_top] = rp;
7927 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7929 if (nflags & MDB_APPEND) {
7930 mn.mc_ki[mn.mc_top] = 0;
7932 split_indx = newindx;
7936 split_indx = (nkeys+1) / 2;
7941 unsigned int lsize, rsize, ksize;
7942 /* Move half of the keys to the right sibling */
7943 x = mc->mc_ki[mc->mc_top] - split_indx;
7944 ksize = mc->mc_db->md_pad;
7945 split = LEAF2KEY(mp, split_indx, ksize);
7946 rsize = (nkeys - split_indx) * ksize;
7947 lsize = (nkeys - split_indx) * sizeof(indx_t);
7948 mp->mp_lower -= lsize;
7949 rp->mp_lower += lsize;
7950 mp->mp_upper += rsize - lsize;
7951 rp->mp_upper -= rsize - lsize;
7952 sepkey.mv_size = ksize;
7953 if (newindx == split_indx) {
7954 sepkey.mv_data = newkey->mv_data;
7956 sepkey.mv_data = split;
7959 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7960 memcpy(rp->mp_ptrs, split, rsize);
7961 sepkey.mv_data = rp->mp_ptrs;
7962 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7963 memcpy(ins, newkey->mv_data, ksize);
7964 mp->mp_lower += sizeof(indx_t);
7965 mp->mp_upper -= ksize - sizeof(indx_t);
7968 memcpy(rp->mp_ptrs, split, x * ksize);
7969 ins = LEAF2KEY(rp, x, ksize);
7970 memcpy(ins, newkey->mv_data, ksize);
7971 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7972 rp->mp_lower += sizeof(indx_t);
7973 rp->mp_upper -= ksize - sizeof(indx_t);
7974 mc->mc_ki[mc->mc_top] = x;
7975 mc->mc_pg[mc->mc_top] = rp;
7978 int psize, nsize, k;
7979 /* Maximum free space in an empty page */
7980 pmax = env->me_psize - PAGEHDRSZ;
7982 nsize = mdb_leaf_size(env, newkey, newdata);
7984 nsize = mdb_branch_size(env, newkey);
7985 nsize = EVEN(nsize);
7987 /* grab a page to hold a temporary copy */
7988 copy = mdb_page_malloc(mc->mc_txn, 1);
7993 copy->mp_pgno = mp->mp_pgno;
7994 copy->mp_flags = mp->mp_flags;
7995 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7996 copy->mp_upper = env->me_psize - PAGEBASE;
7998 /* prepare to insert */
7999 for (i=0, j=0; i<nkeys; i++) {
8001 copy->mp_ptrs[j++] = 0;
8003 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8006 /* When items are relatively large the split point needs
8007 * to be checked, because being off-by-one will make the
8008 * difference between success or failure in mdb_node_add.
8010 * It's also relevant if a page happens to be laid out
8011 * such that one half of its nodes are all "small" and
8012 * the other half of its nodes are "large." If the new
8013 * item is also "large" and falls on the half with
8014 * "large" nodes, it also may not fit.
8016 * As a final tweak, if the new item goes on the last
8017 * spot on the page (and thus, onto the new page), bias
8018 * the split so the new page is emptier than the old page.
8019 * This yields better packing during sequential inserts.
8021 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8022 /* Find split point */
8024 if (newindx <= split_indx || newindx >= nkeys) {
8026 k = newindx >= nkeys ? nkeys : split_indx+2;
8031 for (; i!=k; i+=j) {
8036 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8037 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8039 if (F_ISSET(node->mn_flags, F_BIGDATA))
8040 psize += sizeof(pgno_t);
8042 psize += NODEDSZ(node);
8044 psize = EVEN(psize);
8046 if (psize > pmax || i == k-j) {
8047 split_indx = i + (j<0);
8052 if (split_indx == newindx) {
8053 sepkey.mv_size = newkey->mv_size;
8054 sepkey.mv_data = newkey->mv_data;
8056 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8057 sepkey.mv_size = node->mn_ksize;
8058 sepkey.mv_data = NODEKEY(node);
8063 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8065 /* Copy separator key to the parent.
8067 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8071 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8076 if (mn.mc_snum == mc->mc_snum) {
8077 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8078 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8079 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8080 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8085 /* Right page might now have changed parent.
8086 * Check if left page also changed parent.
8088 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8089 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8090 for (i=0; i<ptop; i++) {
8091 mc->mc_pg[i] = mn.mc_pg[i];
8092 mc->mc_ki[i] = mn.mc_ki[i];
8094 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8095 if (mn.mc_ki[ptop]) {
8096 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8098 /* find right page's left sibling */
8099 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8100 mdb_cursor_sibling(mc, 0);
8105 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8108 mc->mc_flags ^= C_SPLITTING;
8109 if (rc != MDB_SUCCESS) {
8112 if (nflags & MDB_APPEND) {
8113 mc->mc_pg[mc->mc_top] = rp;
8114 mc->mc_ki[mc->mc_top] = 0;
8115 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8118 for (i=0; i<mc->mc_top; i++)
8119 mc->mc_ki[i] = mn.mc_ki[i];
8120 } else if (!IS_LEAF2(mp)) {
8122 mc->mc_pg[mc->mc_top] = rp;
8127 rkey.mv_data = newkey->mv_data;
8128 rkey.mv_size = newkey->mv_size;
8134 /* Update index for the new key. */
8135 mc->mc_ki[mc->mc_top] = j;
8137 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8138 rkey.mv_data = NODEKEY(node);
8139 rkey.mv_size = node->mn_ksize;
8141 xdata.mv_data = NODEDATA(node);
8142 xdata.mv_size = NODEDSZ(node);
8145 pgno = NODEPGNO(node);
8146 flags = node->mn_flags;
8149 if (!IS_LEAF(mp) && j == 0) {
8150 /* First branch index doesn't need key data. */
8154 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8160 mc->mc_pg[mc->mc_top] = copy;
8165 } while (i != split_indx);
8167 nkeys = NUMKEYS(copy);
8168 for (i=0; i<nkeys; i++)
8169 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8170 mp->mp_lower = copy->mp_lower;
8171 mp->mp_upper = copy->mp_upper;
8172 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8173 env->me_psize - copy->mp_upper - PAGEBASE);
8175 /* reset back to original page */
8176 if (newindx < split_indx) {
8177 mc->mc_pg[mc->mc_top] = mp;
8178 if (nflags & MDB_RESERVE) {
8179 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8180 if (!(node->mn_flags & F_BIGDATA))
8181 newdata->mv_data = NODEDATA(node);
8184 mc->mc_pg[mc->mc_top] = rp;
8186 /* Make sure mc_ki is still valid.
8188 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8189 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8190 for (i=0; i<=ptop; i++) {
8191 mc->mc_pg[i] = mn.mc_pg[i];
8192 mc->mc_ki[i] = mn.mc_ki[i];
8199 /* Adjust other cursors pointing to mp */
8200 MDB_cursor *m2, *m3;
8201 MDB_dbi dbi = mc->mc_dbi;
8202 int fixup = NUMKEYS(mp);
8204 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8205 if (mc->mc_flags & C_SUB)
8206 m3 = &m2->mc_xcursor->mx_cursor;
8211 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8213 if (m3->mc_flags & C_SPLITTING)
8218 for (k=m3->mc_top; k>=0; k--) {
8219 m3->mc_ki[k+1] = m3->mc_ki[k];
8220 m3->mc_pg[k+1] = m3->mc_pg[k];
8222 if (m3->mc_ki[0] >= split_indx) {
8227 m3->mc_pg[0] = mc->mc_pg[0];
8231 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8232 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8233 m3->mc_ki[mc->mc_top]++;
8234 if (m3->mc_ki[mc->mc_top] >= fixup) {
8235 m3->mc_pg[mc->mc_top] = rp;
8236 m3->mc_ki[mc->mc_top] -= fixup;
8237 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8239 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8240 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8245 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8248 if (copy) /* tmp page */
8249 mdb_page_free(env, copy);
8251 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8256 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8257 MDB_val *key, MDB_val *data, unsigned int flags)
8262 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8265 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8268 mdb_cursor_init(&mc, txn, dbi, &mx);
8269 return mdb_cursor_put(&mc, key, data, flags);
8273 #define MDB_WBUF (1024*1024)
8276 /** State needed for a compacting copy. */
8277 typedef struct mdb_copy {
8278 pthread_mutex_t mc_mutex;
8279 pthread_cond_t mc_cond;
8286 pgno_t mc_next_pgno;
8289 volatile int mc_new;
8294 /** Dedicated writer thread for compacting copy. */
8295 static THREAD_RET ESECT
8296 mdb_env_copythr(void *arg)
8300 int toggle = 0, wsize, rc;
8303 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8306 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8309 pthread_mutex_lock(&my->mc_mutex);
8311 pthread_cond_signal(&my->mc_cond);
8314 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8315 if (my->mc_new < 0) {
8320 wsize = my->mc_wlen[toggle];
8321 ptr = my->mc_wbuf[toggle];
8324 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8328 } else if (len > 0) {
8342 /* If there's an overflow page tail, write it too */
8343 if (my->mc_olen[toggle]) {
8344 wsize = my->mc_olen[toggle];
8345 ptr = my->mc_over[toggle];
8346 my->mc_olen[toggle] = 0;
8349 my->mc_wlen[toggle] = 0;
8351 pthread_cond_signal(&my->mc_cond);
8353 pthread_cond_signal(&my->mc_cond);
8354 pthread_mutex_unlock(&my->mc_mutex);
8355 return (THREAD_RET)0;
8359 /** Tell the writer thread there's a buffer ready to write */
8361 mdb_env_cthr_toggle(mdb_copy *my, int st)
8363 int toggle = my->mc_toggle ^ 1;
8364 pthread_mutex_lock(&my->mc_mutex);
8365 if (my->mc_status) {
8366 pthread_mutex_unlock(&my->mc_mutex);
8367 return my->mc_status;
8369 while (my->mc_new == 1)
8370 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8372 my->mc_toggle = toggle;
8373 pthread_cond_signal(&my->mc_cond);
8374 pthread_mutex_unlock(&my->mc_mutex);
8378 /** Depth-first tree traversal for compacting copy. */
8380 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8383 MDB_txn *txn = my->mc_txn;
8385 MDB_page *mo, *mp, *leaf;
8390 /* Empty DB, nothing to do */
8391 if (*pg == P_INVALID)
8398 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8401 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8405 /* Make cursor pages writable */
8406 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8410 for (i=0; i<mc.mc_top; i++) {
8411 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8412 mc.mc_pg[i] = (MDB_page *)ptr;
8413 ptr += my->mc_env->me_psize;
8416 /* This is writable space for a leaf page. Usually not needed. */
8417 leaf = (MDB_page *)ptr;
8419 toggle = my->mc_toggle;
8420 while (mc.mc_snum > 0) {
8422 mp = mc.mc_pg[mc.mc_top];
8426 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8427 for (i=0; i<n; i++) {
8428 ni = NODEPTR(mp, i);
8429 if (ni->mn_flags & F_BIGDATA) {
8433 /* Need writable leaf */
8435 mc.mc_pg[mc.mc_top] = leaf;
8436 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8438 ni = NODEPTR(mp, i);
8441 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8442 rc = mdb_page_get(txn, pg, &omp, NULL);
8445 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8446 rc = mdb_env_cthr_toggle(my, 1);
8449 toggle = my->mc_toggle;
8451 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8452 memcpy(mo, omp, my->mc_env->me_psize);
8453 mo->mp_pgno = my->mc_next_pgno;
8454 my->mc_next_pgno += omp->mp_pages;
8455 my->mc_wlen[toggle] += my->mc_env->me_psize;
8456 if (omp->mp_pages > 1) {
8457 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8458 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8459 rc = mdb_env_cthr_toggle(my, 1);
8462 toggle = my->mc_toggle;
8464 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8465 } else if (ni->mn_flags & F_SUBDATA) {
8468 /* Need writable leaf */
8470 mc.mc_pg[mc.mc_top] = leaf;
8471 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8473 ni = NODEPTR(mp, i);
8476 memcpy(&db, NODEDATA(ni), sizeof(db));
8477 my->mc_toggle = toggle;
8478 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8481 toggle = my->mc_toggle;
8482 memcpy(NODEDATA(ni), &db, sizeof(db));
8487 mc.mc_ki[mc.mc_top]++;
8488 if (mc.mc_ki[mc.mc_top] < n) {
8491 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8493 rc = mdb_page_get(txn, pg, &mp, NULL);
8498 mc.mc_ki[mc.mc_top] = 0;
8499 if (IS_BRANCH(mp)) {
8500 /* Whenever we advance to a sibling branch page,
8501 * we must proceed all the way down to its first leaf.
8503 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8506 mc.mc_pg[mc.mc_top] = mp;
8510 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8511 rc = mdb_env_cthr_toggle(my, 1);
8514 toggle = my->mc_toggle;
8516 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8517 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8518 mo->mp_pgno = my->mc_next_pgno++;
8519 my->mc_wlen[toggle] += my->mc_env->me_psize;
8521 /* Update parent if there is one */
8522 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8523 SETPGNO(ni, mo->mp_pgno);
8524 mdb_cursor_pop(&mc);
8526 /* Otherwise we're done */
8536 /** Copy environment with compaction. */
8538 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8543 MDB_txn *txn = NULL;
8548 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8549 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8550 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8551 if (my.mc_wbuf[0] == NULL)
8554 pthread_mutex_init(&my.mc_mutex, NULL);
8555 pthread_cond_init(&my.mc_cond, NULL);
8556 #ifdef HAVE_MEMALIGN
8557 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8558 if (my.mc_wbuf[0] == NULL)
8561 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8566 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8567 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8572 my.mc_next_pgno = 2;
8578 THREAD_CREATE(thr, mdb_env_copythr, &my);
8580 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8584 mp = (MDB_page *)my.mc_wbuf[0];
8585 memset(mp, 0, 2*env->me_psize);
8587 mp->mp_flags = P_META;
8588 mm = (MDB_meta *)METADATA(mp);
8589 mdb_env_init_meta0(env, mm);
8590 mm->mm_address = env->me_metas[0]->mm_address;
8592 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8594 mp->mp_flags = P_META;
8595 *(MDB_meta *)METADATA(mp) = *mm;
8596 mm = (MDB_meta *)METADATA(mp);
8598 /* Count the number of free pages, subtract from lastpg to find
8599 * number of active pages
8602 MDB_ID freecount = 0;
8605 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8606 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8607 freecount += *(MDB_ID *)data.mv_data;
8608 freecount += txn->mt_dbs[0].md_branch_pages +
8609 txn->mt_dbs[0].md_leaf_pages +
8610 txn->mt_dbs[0].md_overflow_pages;
8612 /* Set metapage 1 */
8613 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8614 mm->mm_dbs[1] = txn->mt_dbs[1];
8615 if (mm->mm_last_pg > 1) {
8616 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8619 mm->mm_dbs[1].md_root = P_INVALID;
8622 my.mc_wlen[0] = env->me_psize * 2;
8624 pthread_mutex_lock(&my.mc_mutex);
8626 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8627 pthread_mutex_unlock(&my.mc_mutex);
8628 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8629 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8630 rc = mdb_env_cthr_toggle(&my, 1);
8631 mdb_env_cthr_toggle(&my, -1);
8632 pthread_mutex_lock(&my.mc_mutex);
8634 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8635 pthread_mutex_unlock(&my.mc_mutex);
8640 CloseHandle(my.mc_cond);
8641 CloseHandle(my.mc_mutex);
8642 _aligned_free(my.mc_wbuf[0]);
8644 pthread_cond_destroy(&my.mc_cond);
8645 pthread_mutex_destroy(&my.mc_mutex);
8646 free(my.mc_wbuf[0]);
8651 /** Copy environment as-is. */
8653 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8655 MDB_txn *txn = NULL;
8661 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8665 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8668 /* Do the lock/unlock of the reader mutex before starting the
8669 * write txn. Otherwise other read txns could block writers.
8671 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8676 /* We must start the actual read txn after blocking writers */
8677 mdb_txn_reset0(txn, "reset-stage1");
8679 /* Temporarily block writers until we snapshot the meta pages */
8682 rc = mdb_txn_renew0(txn);
8684 UNLOCK_MUTEX_W(env);
8689 wsize = env->me_psize * 2;
8693 DO_WRITE(rc, fd, ptr, w2, len);
8697 } else if (len > 0) {
8703 /* Non-blocking or async handles are not supported */
8709 UNLOCK_MUTEX_W(env);
8714 w2 = txn->mt_next_pgno * env->me_psize;
8717 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8724 if (wsize > MAX_WRITE)
8728 DO_WRITE(rc, fd, ptr, w2, len);
8732 } else if (len > 0) {
8749 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8751 if (flags & MDB_CP_COMPACT)
8752 return mdb_env_copyfd1(env, fd);
8754 return mdb_env_copyfd0(env, fd);
8758 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8760 return mdb_env_copyfd2(env, fd, 0);
8764 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8768 HANDLE newfd = INVALID_HANDLE_VALUE;
8770 if (env->me_flags & MDB_NOSUBDIR) {
8771 lpath = (char *)path;
8774 len += sizeof(DATANAME);
8775 lpath = malloc(len);
8778 sprintf(lpath, "%s" DATANAME, path);
8781 /* The destination path must exist, but the destination file must not.
8782 * We don't want the OS to cache the writes, since the source data is
8783 * already in the OS cache.
8786 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8787 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8789 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8791 if (newfd == INVALID_HANDLE_VALUE) {
8796 if (env->me_psize >= env->me_os_psize) {
8798 /* Set O_DIRECT if the file system supports it */
8799 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8800 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8802 #ifdef F_NOCACHE /* __APPLE__ */
8803 rc = fcntl(newfd, F_NOCACHE, 1);
8811 rc = mdb_env_copyfd2(env, newfd, flags);
8814 if (!(env->me_flags & MDB_NOSUBDIR))
8816 if (newfd != INVALID_HANDLE_VALUE)
8817 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8824 mdb_env_copy(MDB_env *env, const char *path)
8826 return mdb_env_copy2(env, path, 0);
8830 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8832 if ((flag & CHANGEABLE) != flag)
8835 env->me_flags |= flag;
8837 env->me_flags &= ~flag;
8842 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8847 *arg = env->me_flags;
8852 mdb_env_set_userctx(MDB_env *env, void *ctx)
8856 env->me_userctx = ctx;
8861 mdb_env_get_userctx(MDB_env *env)
8863 return env ? env->me_userctx : NULL;
8867 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8872 env->me_assert_func = func;
8878 mdb_env_get_path(MDB_env *env, const char **arg)
8883 *arg = env->me_path;
8888 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8897 /** Common code for #mdb_stat() and #mdb_env_stat().
8898 * @param[in] env the environment to operate in.
8899 * @param[in] db the #MDB_db record containing the stats to return.
8900 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8901 * @return 0, this function always succeeds.
8904 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8906 arg->ms_psize = env->me_psize;
8907 arg->ms_depth = db->md_depth;
8908 arg->ms_branch_pages = db->md_branch_pages;
8909 arg->ms_leaf_pages = db->md_leaf_pages;
8910 arg->ms_overflow_pages = db->md_overflow_pages;
8911 arg->ms_entries = db->md_entries;
8917 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8921 if (env == NULL || arg == NULL)
8924 toggle = mdb_env_pick_meta(env);
8926 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8930 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8934 if (env == NULL || arg == NULL)
8937 toggle = mdb_env_pick_meta(env);
8938 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8939 arg->me_mapsize = env->me_mapsize;
8940 arg->me_maxreaders = env->me_maxreaders;
8942 /* me_numreaders may be zero if this process never used any readers. Use
8943 * the shared numreader count if it exists.
8945 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8947 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8948 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8952 /** Set the default comparison functions for a database.
8953 * Called immediately after a database is opened to set the defaults.
8954 * The user can then override them with #mdb_set_compare() or
8955 * #mdb_set_dupsort().
8956 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8957 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8960 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8962 uint16_t f = txn->mt_dbs[dbi].md_flags;
8964 txn->mt_dbxs[dbi].md_cmp =
8965 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8966 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8968 txn->mt_dbxs[dbi].md_dcmp =
8969 !(f & MDB_DUPSORT) ? 0 :
8970 ((f & MDB_INTEGERDUP)
8971 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8972 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8975 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8981 int rc, dbflag, exact;
8982 unsigned int unused = 0, seq;
8985 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8986 mdb_default_cmp(txn, FREE_DBI);
8989 if ((flags & VALID_FLAGS) != flags)
8991 if (txn->mt_flags & MDB_TXN_ERROR)
8997 if (flags & PERSISTENT_FLAGS) {
8998 uint16_t f2 = flags & PERSISTENT_FLAGS;
8999 /* make sure flag changes get committed */
9000 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9001 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9002 txn->mt_flags |= MDB_TXN_DIRTY;
9005 mdb_default_cmp(txn, MAIN_DBI);
9009 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9010 mdb_default_cmp(txn, MAIN_DBI);
9013 /* Is the DB already open? */
9015 for (i=2; i<txn->mt_numdbs; i++) {
9016 if (!txn->mt_dbxs[i].md_name.mv_size) {
9017 /* Remember this free slot */
9018 if (!unused) unused = i;
9021 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9022 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9028 /* If no free slot and max hit, fail */
9029 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9030 return MDB_DBS_FULL;
9032 /* Cannot mix named databases with some mainDB flags */
9033 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9034 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9036 /* Find the DB info */
9037 dbflag = DB_NEW|DB_VALID;
9040 key.mv_data = (void *)name;
9041 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9042 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9043 if (rc == MDB_SUCCESS) {
9044 /* make sure this is actually a DB */
9045 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9046 if (!(node->mn_flags & F_SUBDATA))
9047 return MDB_INCOMPATIBLE;
9048 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9049 /* Create if requested */
9050 data.mv_size = sizeof(MDB_db);
9051 data.mv_data = &dummy;
9052 memset(&dummy, 0, sizeof(dummy));
9053 dummy.md_root = P_INVALID;
9054 dummy.md_flags = flags & PERSISTENT_FLAGS;
9055 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9059 /* OK, got info, add to table */
9060 if (rc == MDB_SUCCESS) {
9061 unsigned int slot = unused ? unused : txn->mt_numdbs;
9062 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9063 txn->mt_dbxs[slot].md_name.mv_size = len;
9064 txn->mt_dbxs[slot].md_rel = NULL;
9065 txn->mt_dbflags[slot] = dbflag;
9066 /* txn-> and env-> are the same in read txns, use
9067 * tmp variable to avoid undefined assignment
9069 seq = ++txn->mt_env->me_dbiseqs[slot];
9070 txn->mt_dbiseqs[slot] = seq;
9072 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9074 mdb_default_cmp(txn, slot);
9083 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9085 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9088 if (txn->mt_flags & MDB_TXN_ERROR)
9091 if (txn->mt_dbflags[dbi] & DB_STALE) {
9094 /* Stale, must read the DB's root. cursor_init does it for us. */
9095 mdb_cursor_init(&mc, txn, dbi, &mx);
9097 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9100 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9103 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9105 ptr = env->me_dbxs[dbi].md_name.mv_data;
9106 /* If there was no name, this was already closed */
9108 env->me_dbxs[dbi].md_name.mv_data = NULL;
9109 env->me_dbxs[dbi].md_name.mv_size = 0;
9110 env->me_dbflags[dbi] = 0;
9111 env->me_dbiseqs[dbi]++;
9116 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9118 /* We could return the flags for the FREE_DBI too but what's the point? */
9119 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9121 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9125 /** Add all the DB's pages to the free list.
9126 * @param[in] mc Cursor on the DB to free.
9127 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9128 * @return 0 on success, non-zero on failure.
9131 mdb_drop0(MDB_cursor *mc, int subs)
9135 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9136 if (rc == MDB_SUCCESS) {
9137 MDB_txn *txn = mc->mc_txn;
9142 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9143 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9146 mdb_cursor_copy(mc, &mx);
9147 while (mc->mc_snum > 0) {
9148 MDB_page *mp = mc->mc_pg[mc->mc_top];
9149 unsigned n = NUMKEYS(mp);
9151 for (i=0; i<n; i++) {
9152 ni = NODEPTR(mp, i);
9153 if (ni->mn_flags & F_BIGDATA) {
9156 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9157 rc = mdb_page_get(txn, pg, &omp, NULL);
9160 mdb_cassert(mc, IS_OVERFLOW(omp));
9161 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9165 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9166 mdb_xcursor_init1(mc, ni);
9167 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9173 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9175 for (i=0; i<n; i++) {
9177 ni = NODEPTR(mp, i);
9180 mdb_midl_xappend(txn->mt_free_pgs, pg);
9185 mc->mc_ki[mc->mc_top] = i;
9186 rc = mdb_cursor_sibling(mc, 1);
9188 if (rc != MDB_NOTFOUND)
9190 /* no more siblings, go back to beginning
9191 * of previous level.
9195 for (i=1; i<mc->mc_snum; i++) {
9197 mc->mc_pg[i] = mx.mc_pg[i];
9202 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9205 txn->mt_flags |= MDB_TXN_ERROR;
9206 } else if (rc == MDB_NOTFOUND) {
9212 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9214 MDB_cursor *mc, *m2;
9217 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9220 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9223 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9226 rc = mdb_cursor_open(txn, dbi, &mc);
9230 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9231 /* Invalidate the dropped DB's cursors */
9232 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9233 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9237 /* Can't delete the main DB */
9238 if (del && dbi > MAIN_DBI) {
9239 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9241 txn->mt_dbflags[dbi] = DB_STALE;
9242 mdb_dbi_close(txn->mt_env, dbi);
9244 txn->mt_flags |= MDB_TXN_ERROR;
9247 /* reset the DB record, mark it dirty */
9248 txn->mt_dbflags[dbi] |= DB_DIRTY;
9249 txn->mt_dbs[dbi].md_depth = 0;
9250 txn->mt_dbs[dbi].md_branch_pages = 0;
9251 txn->mt_dbs[dbi].md_leaf_pages = 0;
9252 txn->mt_dbs[dbi].md_overflow_pages = 0;
9253 txn->mt_dbs[dbi].md_entries = 0;
9254 txn->mt_dbs[dbi].md_root = P_INVALID;
9256 txn->mt_flags |= MDB_TXN_DIRTY;
9259 mdb_cursor_close(mc);
9263 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9265 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9268 txn->mt_dbxs[dbi].md_cmp = cmp;
9272 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9274 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9277 txn->mt_dbxs[dbi].md_dcmp = cmp;
9281 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9283 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9286 txn->mt_dbxs[dbi].md_rel = rel;
9290 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9292 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9295 txn->mt_dbxs[dbi].md_relctx = ctx;
9300 mdb_env_get_maxkeysize(MDB_env *env)
9302 return ENV_MAXKEY(env);
9306 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9308 unsigned int i, rdrs;
9311 int rc = 0, first = 1;
9315 if (!env->me_txns) {
9316 return func("(no reader locks)\n", ctx);
9318 rdrs = env->me_txns->mti_numreaders;
9319 mr = env->me_txns->mti_readers;
9320 for (i=0; i<rdrs; i++) {
9322 txnid_t txnid = mr[i].mr_txnid;
9323 sprintf(buf, txnid == (txnid_t)-1 ?
9324 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9325 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9328 rc = func(" pid thread txnid\n", ctx);
9332 rc = func(buf, ctx);
9338 rc = func("(no active readers)\n", ctx);
9343 /** Insert pid into list if not already present.
9344 * return -1 if already present.
9347 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9349 /* binary search of pid in list */
9351 unsigned cursor = 1;
9353 unsigned n = ids[0];
9356 unsigned pivot = n >> 1;
9357 cursor = base + pivot + 1;
9358 val = pid - ids[cursor];
9363 } else if ( val > 0 ) {
9368 /* found, so it's a duplicate */
9377 for (n = ids[0]; n > cursor; n--)
9384 mdb_reader_check(MDB_env *env, int *dead)
9386 unsigned int i, j, rdrs;
9388 MDB_PID_T *pids, pid;
9397 rdrs = env->me_txns->mti_numreaders;
9398 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9402 mr = env->me_txns->mti_readers;
9403 for (i=0; i<rdrs; i++) {
9404 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9406 if (mdb_pid_insert(pids, pid) == 0) {
9407 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9409 /* Recheck, a new process may have reused pid */
9410 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9411 for (j=i; j<rdrs; j++)
9412 if (mr[j].mr_pid == pid) {
9413 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9414 (unsigned) pid, mr[j].mr_txnid));
9419 UNLOCK_MUTEX_R(env);