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)
2545 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2546 txn->mt_u.dirty_list = env->me_dirty_list;
2547 txn->mt_u.dirty_list[0].mid = 0;
2548 txn->mt_free_pgs = env->me_free_pgs;
2549 txn->mt_free_pgs[0] = 0;
2550 txn->mt_spill_pgs = NULL;
2552 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2555 /* Copy the DB info and flags */
2556 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2558 /* Moved to here to avoid a data race in read TXNs */
2559 txn->mt_next_pgno = meta->mm_last_pg+1;
2561 for (i=2; i<txn->mt_numdbs; i++) {
2562 x = env->me_dbflags[i];
2563 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2564 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2566 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2568 if (env->me_maxpg < txn->mt_next_pgno) {
2569 mdb_txn_reset0(txn, "renew0-mapfail");
2571 txn->mt_u.reader->mr_pid = 0;
2572 txn->mt_u.reader = NULL;
2574 return MDB_MAP_RESIZED;
2581 mdb_txn_renew(MDB_txn *txn)
2585 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2588 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2589 DPUTS("environment had fatal error, must shutdown!");
2593 rc = mdb_txn_renew0(txn);
2594 if (rc == MDB_SUCCESS) {
2595 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2596 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2597 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2603 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2607 int rc, size, tsize = sizeof(MDB_txn);
2609 if (env->me_flags & MDB_FATAL_ERROR) {
2610 DPUTS("environment had fatal error, must shutdown!");
2613 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2616 /* Nested transactions: Max 1 child, write txns only, no writemap */
2617 if (parent->mt_child ||
2618 (flags & MDB_RDONLY) ||
2619 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2620 (env->me_flags & MDB_WRITEMAP))
2622 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2624 tsize = sizeof(MDB_ntxn);
2627 if (!(flags & MDB_RDONLY)) {
2629 txn = env->me_txn0; /* just reuse preallocated write txn */
2633 /* child txns use own copy of cursors */
2634 size += env->me_maxdbs * sizeof(MDB_cursor *);
2636 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2638 if ((txn = calloc(1, size)) == NULL) {
2639 DPRINTF(("calloc: %s", strerror(errno)));
2642 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2643 if (flags & MDB_RDONLY) {
2644 txn->mt_flags |= MDB_TXN_RDONLY;
2645 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2646 txn->mt_dbiseqs = env->me_dbiseqs;
2648 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2650 txn->mt_dbiseqs = parent->mt_dbiseqs;
2651 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2653 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2654 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2662 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2663 if (!txn->mt_u.dirty_list ||
2664 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2666 free(txn->mt_u.dirty_list);
2670 txn->mt_txnid = parent->mt_txnid;
2671 txn->mt_dirty_room = parent->mt_dirty_room;
2672 txn->mt_u.dirty_list[0].mid = 0;
2673 txn->mt_spill_pgs = NULL;
2674 txn->mt_next_pgno = parent->mt_next_pgno;
2675 parent->mt_child = txn;
2676 txn->mt_parent = parent;
2677 txn->mt_numdbs = parent->mt_numdbs;
2678 txn->mt_flags = parent->mt_flags;
2679 txn->mt_dbxs = parent->mt_dbxs;
2680 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2681 /* Copy parent's mt_dbflags, but clear DB_NEW */
2682 for (i=0; i<txn->mt_numdbs; i++)
2683 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2685 ntxn = (MDB_ntxn *)txn;
2686 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2687 if (env->me_pghead) {
2688 size = MDB_IDL_SIZEOF(env->me_pghead);
2689 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2691 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2696 rc = mdb_cursor_shadow(parent, txn);
2698 mdb_txn_reset0(txn, "beginchild-fail");
2700 rc = mdb_txn_renew0(txn);
2703 if (txn != env->me_txn0)
2707 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2708 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2709 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2716 mdb_txn_env(MDB_txn *txn)
2718 if(!txn) return NULL;
2722 /** Export or close DBI handles opened in this txn. */
2724 mdb_dbis_update(MDB_txn *txn, int keep)
2727 MDB_dbi n = txn->mt_numdbs;
2728 MDB_env *env = txn->mt_env;
2729 unsigned char *tdbflags = txn->mt_dbflags;
2731 for (i = n; --i >= 2;) {
2732 if (tdbflags[i] & DB_NEW) {
2734 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2736 char *ptr = env->me_dbxs[i].md_name.mv_data;
2738 env->me_dbxs[i].md_name.mv_data = NULL;
2739 env->me_dbxs[i].md_name.mv_size = 0;
2740 env->me_dbflags[i] = 0;
2741 env->me_dbiseqs[i]++;
2747 if (keep && env->me_numdbs < n)
2751 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2752 * May be called twice for readonly txns: First reset it, then abort.
2753 * @param[in] txn the transaction handle to reset
2754 * @param[in] act why the transaction is being reset
2757 mdb_txn_reset0(MDB_txn *txn, const char *act)
2759 MDB_env *env = txn->mt_env;
2761 /* Close any DBI handles opened in this txn */
2762 mdb_dbis_update(txn, 0);
2764 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2765 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2766 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2768 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2769 if (txn->mt_u.reader) {
2770 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2771 if (!(env->me_flags & MDB_NOTLS))
2772 txn->mt_u.reader = NULL; /* txn does not own reader */
2774 txn->mt_numdbs = 0; /* close nothing if called again */
2775 txn->mt_dbxs = NULL; /* mark txn as reset */
2777 pgno_t *pghead = env->me_pghead;
2778 env->me_pghead = NULL;
2781 if (!(env->me_flags & MDB_WRITEMAP)) {
2782 mdb_dlist_free(txn);
2785 if (!txn->mt_parent) {
2786 if (mdb_midl_shrink(&txn->mt_free_pgs))
2787 env->me_free_pgs = txn->mt_free_pgs;
2790 /* The writer mutex was locked in mdb_txn_begin. */
2792 UNLOCK_MUTEX_W(env);
2795 mdb_cursors_close(txn, 0);
2797 mdb_midl_free(pghead);
2799 if (txn->mt_parent) {
2800 txn->mt_parent->mt_child = NULL;
2801 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2802 mdb_midl_free(txn->mt_free_pgs);
2803 mdb_midl_free(txn->mt_spill_pgs);
2804 free(txn->mt_u.dirty_list);
2810 mdb_txn_reset(MDB_txn *txn)
2815 /* This call is only valid for read-only txns */
2816 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2819 mdb_txn_reset0(txn, "reset");
2823 mdb_txn_abort(MDB_txn *txn)
2829 mdb_txn_abort(txn->mt_child);
2831 mdb_txn_reset0(txn, "abort");
2832 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2833 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2834 txn->mt_u.reader->mr_pid = 0;
2836 if (txn != txn->mt_env->me_txn0)
2840 /** Save the freelist as of this transaction to the freeDB.
2841 * This changes the freelist. Keep trying until it stabilizes.
2844 mdb_freelist_save(MDB_txn *txn)
2846 /* env->me_pghead[] can grow and shrink during this call.
2847 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2848 * Page numbers cannot disappear from txn->mt_free_pgs[].
2851 MDB_env *env = txn->mt_env;
2852 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2853 txnid_t pglast = 0, head_id = 0;
2854 pgno_t freecnt = 0, *free_pgs, *mop;
2855 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2857 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2859 if (env->me_pghead) {
2860 /* Make sure first page of freeDB is touched and on freelist */
2861 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2862 if (rc && rc != MDB_NOTFOUND)
2866 if (!env->me_pghead && txn->mt_loose_pgs) {
2867 /* Put loose page numbers in mt_free_pgs, since
2868 * we may be unable to return them to me_pghead.
2870 MDB_page *mp = txn->mt_loose_pgs;
2871 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2873 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2874 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2875 txn->mt_loose_pgs = NULL;
2876 txn->mt_loose_count = 0;
2879 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2880 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2881 ? SSIZE_MAX : maxfree_1pg;
2884 /* Come back here after each Put() in case freelist changed */
2889 /* If using records from freeDB which we have not yet
2890 * deleted, delete them and any we reserved for me_pghead.
2892 while (pglast < env->me_pglast) {
2893 rc = mdb_cursor_first(&mc, &key, NULL);
2896 pglast = head_id = *(txnid_t *)key.mv_data;
2897 total_room = head_room = 0;
2898 mdb_tassert(txn, pglast <= env->me_pglast);
2899 rc = mdb_cursor_del(&mc, 0);
2904 /* Save the IDL of pages freed by this txn, to a single record */
2905 if (freecnt < txn->mt_free_pgs[0]) {
2907 /* Make sure last page of freeDB is touched and on freelist */
2908 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2909 if (rc && rc != MDB_NOTFOUND)
2912 free_pgs = txn->mt_free_pgs;
2913 /* Write to last page of freeDB */
2914 key.mv_size = sizeof(txn->mt_txnid);
2915 key.mv_data = &txn->mt_txnid;
2917 freecnt = free_pgs[0];
2918 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2919 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2922 /* Retry if mt_free_pgs[] grew during the Put() */
2923 free_pgs = txn->mt_free_pgs;
2924 } while (freecnt < free_pgs[0]);
2925 mdb_midl_sort(free_pgs);
2926 memcpy(data.mv_data, free_pgs, data.mv_size);
2929 unsigned int i = free_pgs[0];
2930 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2931 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2933 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2939 mop = env->me_pghead;
2940 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2942 /* Reserve records for me_pghead[]. Split it if multi-page,
2943 * to avoid searching freeDB for a page range. Use keys in
2944 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2946 if (total_room >= mop_len) {
2947 if (total_room == mop_len || --more < 0)
2949 } else if (head_room >= maxfree_1pg && head_id > 1) {
2950 /* Keep current record (overflow page), add a new one */
2954 /* (Re)write {key = head_id, IDL length = head_room} */
2955 total_room -= head_room;
2956 head_room = mop_len - total_room;
2957 if (head_room > maxfree_1pg && head_id > 1) {
2958 /* Overflow multi-page for part of me_pghead */
2959 head_room /= head_id; /* amortize page sizes */
2960 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2961 } else if (head_room < 0) {
2962 /* Rare case, not bothering to delete this record */
2965 key.mv_size = sizeof(head_id);
2966 key.mv_data = &head_id;
2967 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2968 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2971 /* IDL is initially empty, zero out at least the length */
2972 pgs = (pgno_t *)data.mv_data;
2973 j = head_room > clean_limit ? head_room : 0;
2977 total_room += head_room;
2980 /* Return loose page numbers to me_pghead, though usually none are
2981 * left at this point. The pages themselves remain in dirty_list.
2983 if (txn->mt_loose_pgs) {
2984 MDB_page *mp = txn->mt_loose_pgs;
2985 unsigned count = txn->mt_loose_count;
2987 /* Room for loose pages + temp IDL with same */
2988 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
2990 mop = env->me_pghead;
2991 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
2992 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
2993 loose[ ++count ] = mp->mp_pgno;
2995 mdb_midl_sort(loose);
2996 mdb_midl_xmerge(mop, loose);
2997 txn->mt_loose_pgs = NULL;
2998 txn->mt_loose_count = 0;
3002 /* Fill in the reserved me_pghead records */
3008 rc = mdb_cursor_first(&mc, &key, &data);
3009 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3010 txnid_t id = *(txnid_t *)key.mv_data;
3011 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3014 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3016 if (len > mop_len) {
3018 data.mv_size = (len + 1) * sizeof(MDB_ID);
3020 data.mv_data = mop -= len;
3023 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3025 if (rc || !(mop_len -= len))
3032 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3033 * @param[in] txn the transaction that's being committed
3034 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3035 * @return 0 on success, non-zero on failure.
3038 mdb_page_flush(MDB_txn *txn, int keep)
3040 MDB_env *env = txn->mt_env;
3041 MDB_ID2L dl = txn->mt_u.dirty_list;
3042 unsigned psize = env->me_psize, j;
3043 int i, pagecount = dl[0].mid, rc;
3044 size_t size = 0, pos = 0;
3046 MDB_page *dp = NULL;
3050 struct iovec iov[MDB_COMMIT_PAGES];
3051 ssize_t wpos = 0, wsize = 0, wres;
3052 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3058 if (env->me_flags & MDB_WRITEMAP) {
3059 /* Clear dirty flags */
3060 while (++i <= pagecount) {
3062 /* Don't flush this page yet */
3063 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3064 dp->mp_flags &= ~P_KEEP;
3068 dp->mp_flags &= ~P_DIRTY;
3073 /* Write the pages */
3075 if (++i <= pagecount) {
3077 /* Don't flush this page yet */
3078 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3079 dp->mp_flags &= ~P_KEEP;
3084 /* clear dirty flag */
3085 dp->mp_flags &= ~P_DIRTY;
3088 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3093 /* Windows actually supports scatter/gather I/O, but only on
3094 * unbuffered file handles. Since we're relying on the OS page
3095 * cache for all our data, that's self-defeating. So we just
3096 * write pages one at a time. We use the ov structure to set
3097 * the write offset, to at least save the overhead of a Seek
3100 DPRINTF(("committing page %"Z"u", pgno));
3101 memset(&ov, 0, sizeof(ov));
3102 ov.Offset = pos & 0xffffffff;
3103 ov.OffsetHigh = pos >> 16 >> 16;
3104 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3106 DPRINTF(("WriteFile: %d", rc));
3110 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3111 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3113 /* Write previous page(s) */
3114 #ifdef MDB_USE_PWRITEV
3115 wres = pwritev(env->me_fd, iov, n, wpos);
3118 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3120 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3122 DPRINTF(("lseek: %s", strerror(rc)));
3125 wres = writev(env->me_fd, iov, n);
3128 if (wres != wsize) {
3131 DPRINTF(("Write error: %s", strerror(rc)));
3133 rc = EIO; /* TODO: Use which error code? */
3134 DPUTS("short write, filesystem full?");
3145 DPRINTF(("committing page %"Z"u", pgno));
3146 next_pos = pos + size;
3147 iov[n].iov_len = size;
3148 iov[n].iov_base = (char *)dp;
3154 /* MIPS has cache coherency issues, this is a no-op everywhere else
3155 * Note: for any size >= on-chip cache size, entire on-chip cache is
3158 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3160 for (i = keep; ++i <= pagecount; ) {
3162 /* This is a page we skipped above */
3165 dl[j].mid = dp->mp_pgno;
3168 mdb_dpage_free(env, dp);
3173 txn->mt_dirty_room += i - j;
3179 mdb_txn_commit(MDB_txn *txn)
3185 if (txn == NULL || txn->mt_env == NULL)
3188 if (txn->mt_child) {
3189 rc = mdb_txn_commit(txn->mt_child);
3190 txn->mt_child = NULL;
3197 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3198 mdb_dbis_update(txn, 1);
3199 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3204 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3205 DPUTS("error flag is set, can't commit");
3207 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3212 if (txn->mt_parent) {
3213 MDB_txn *parent = txn->mt_parent;
3217 unsigned x, y, len, ps_len;
3219 /* Append our free list to parent's */
3220 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3223 mdb_midl_free(txn->mt_free_pgs);
3224 /* Failures after this must either undo the changes
3225 * to the parent or set MDB_TXN_ERROR in the parent.
3228 parent->mt_next_pgno = txn->mt_next_pgno;
3229 parent->mt_flags = txn->mt_flags;
3231 /* Merge our cursors into parent's and close them */
3232 mdb_cursors_close(txn, 1);
3234 /* Update parent's DB table. */
3235 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3236 parent->mt_numdbs = txn->mt_numdbs;
3237 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3238 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3239 for (i=2; i<txn->mt_numdbs; i++) {
3240 /* preserve parent's DB_NEW status */
3241 x = parent->mt_dbflags[i] & DB_NEW;
3242 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3245 dst = parent->mt_u.dirty_list;
3246 src = txn->mt_u.dirty_list;
3247 /* Remove anything in our dirty list from parent's spill list */
3248 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3250 pspill[0] = (pgno_t)-1;
3251 /* Mark our dirty pages as deleted in parent spill list */
3252 for (i=0, len=src[0].mid; ++i <= len; ) {
3253 MDB_ID pn = src[i].mid << 1;
3254 while (pn > pspill[x])
3256 if (pn == pspill[x]) {
3261 /* Squash deleted pagenums if we deleted any */
3262 for (x=y; ++x <= ps_len; )
3263 if (!(pspill[x] & 1))
3264 pspill[++y] = pspill[x];
3268 /* Find len = length of merging our dirty list with parent's */
3270 dst[0].mid = 0; /* simplify loops */
3271 if (parent->mt_parent) {
3272 len = x + src[0].mid;
3273 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3274 for (i = x; y && i; y--) {
3275 pgno_t yp = src[y].mid;
3276 while (yp < dst[i].mid)
3278 if (yp == dst[i].mid) {
3283 } else { /* Simplify the above for single-ancestor case */
3284 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3286 /* Merge our dirty list with parent's */
3288 for (i = len; y; dst[i--] = src[y--]) {
3289 pgno_t yp = src[y].mid;
3290 while (yp < dst[x].mid)
3291 dst[i--] = dst[x--];
3292 if (yp == dst[x].mid)
3293 free(dst[x--].mptr);
3295 mdb_tassert(txn, i == x);
3297 free(txn->mt_u.dirty_list);
3298 parent->mt_dirty_room = txn->mt_dirty_room;
3299 if (txn->mt_spill_pgs) {
3300 if (parent->mt_spill_pgs) {
3301 /* TODO: Prevent failure here, so parent does not fail */
3302 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3304 parent->mt_flags |= MDB_TXN_ERROR;
3305 mdb_midl_free(txn->mt_spill_pgs);
3306 mdb_midl_sort(parent->mt_spill_pgs);
3308 parent->mt_spill_pgs = txn->mt_spill_pgs;
3312 /* Append our loose page list to parent's */
3313 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3315 *lp = txn->mt_loose_pgs;
3316 parent->mt_loose_count += txn->mt_loose_count;
3318 parent->mt_child = NULL;
3319 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3324 if (txn != env->me_txn) {
3325 DPUTS("attempt to commit unknown transaction");
3330 mdb_cursors_close(txn, 0);
3332 if (!txn->mt_u.dirty_list[0].mid &&
3333 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3336 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3337 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3339 /* Update DB root pointers */
3340 if (txn->mt_numdbs > 2) {
3344 data.mv_size = sizeof(MDB_db);
3346 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3347 for (i = 2; i < txn->mt_numdbs; i++) {
3348 if (txn->mt_dbflags[i] & DB_DIRTY) {
3349 if (TXN_DBI_CHANGED(txn, i)) {
3353 data.mv_data = &txn->mt_dbs[i];
3354 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3361 rc = mdb_freelist_save(txn);
3365 mdb_midl_free(env->me_pghead);
3366 env->me_pghead = NULL;
3367 if (mdb_midl_shrink(&txn->mt_free_pgs))
3368 env->me_free_pgs = txn->mt_free_pgs;
3374 if ((rc = mdb_page_flush(txn, 0)) ||
3375 (rc = mdb_env_sync(env, 0)) ||
3376 (rc = mdb_env_write_meta(txn)))
3379 /* Free P_LOOSE pages left behind in dirty_list */
3380 if (!(env->me_flags & MDB_WRITEMAP))
3381 mdb_dlist_free(txn);
3386 mdb_dbis_update(txn, 1);
3389 UNLOCK_MUTEX_W(env);
3390 if (txn != env->me_txn0)
3400 /** Read the environment parameters of a DB environment before
3401 * mapping it into memory.
3402 * @param[in] env the environment handle
3403 * @param[out] meta address of where to store the meta information
3404 * @return 0 on success, non-zero on failure.
3407 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3413 enum { Size = sizeof(pbuf) };
3415 /* We don't know the page size yet, so use a minimum value.
3416 * Read both meta pages so we can use the latest one.
3419 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3423 memset(&ov, 0, sizeof(ov));
3425 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3426 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3429 rc = pread(env->me_fd, &pbuf, Size, off);
3432 if (rc == 0 && off == 0)
3434 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3435 DPRINTF(("read: %s", mdb_strerror(rc)));
3439 p = (MDB_page *)&pbuf;
3441 if (!F_ISSET(p->mp_flags, P_META)) {
3442 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3447 if (m->mm_magic != MDB_MAGIC) {
3448 DPUTS("meta has invalid magic");
3452 if (m->mm_version != MDB_DATA_VERSION) {
3453 DPRINTF(("database is version %u, expected version %u",
3454 m->mm_version, MDB_DATA_VERSION));
3455 return MDB_VERSION_MISMATCH;
3458 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3465 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3467 meta->mm_magic = MDB_MAGIC;
3468 meta->mm_version = MDB_DATA_VERSION;
3469 meta->mm_mapsize = env->me_mapsize;
3470 meta->mm_psize = env->me_psize;
3471 meta->mm_last_pg = 1;
3472 meta->mm_flags = env->me_flags & 0xffff;
3473 meta->mm_flags |= MDB_INTEGERKEY;
3474 meta->mm_dbs[0].md_root = P_INVALID;
3475 meta->mm_dbs[1].md_root = P_INVALID;
3478 /** Write the environment parameters of a freshly created DB environment.
3479 * @param[in] env the environment handle
3480 * @param[out] meta address of where to store the meta information
3481 * @return 0 on success, non-zero on failure.
3484 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3492 memset(&ov, 0, sizeof(ov));
3493 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3495 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3498 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3499 len = pwrite(fd, ptr, size, pos); \
3500 rc = (len >= 0); } while(0)
3503 DPUTS("writing new meta page");
3505 psize = env->me_psize;
3507 mdb_env_init_meta0(env, meta);
3509 p = calloc(2, psize);
3511 p->mp_flags = P_META;
3512 *(MDB_meta *)METADATA(p) = *meta;
3514 q = (MDB_page *)((char *)p + psize);
3516 q->mp_flags = P_META;
3517 *(MDB_meta *)METADATA(q) = *meta;
3519 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3522 else if ((unsigned) len == psize * 2)
3530 /** Update the environment info to commit a transaction.
3531 * @param[in] txn the transaction that's being committed
3532 * @return 0 on success, non-zero on failure.
3535 mdb_env_write_meta(MDB_txn *txn)
3538 MDB_meta meta, metab, *mp;
3541 int rc, len, toggle;
3550 toggle = txn->mt_txnid & 1;
3551 DPRINTF(("writing meta page %d for root page %"Z"u",
3552 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3555 mp = env->me_metas[toggle];
3556 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3557 /* Persist any increases of mapsize config */
3558 if (mapsize < env->me_mapsize)
3559 mapsize = env->me_mapsize;
3561 if (env->me_flags & MDB_WRITEMAP) {
3562 mp->mm_mapsize = mapsize;
3563 mp->mm_dbs[0] = txn->mt_dbs[0];
3564 mp->mm_dbs[1] = txn->mt_dbs[1];
3565 mp->mm_last_pg = txn->mt_next_pgno - 1;
3566 mp->mm_txnid = txn->mt_txnid;
3567 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3568 unsigned meta_size = env->me_psize;
3569 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3572 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3573 if (meta_size < env->me_os_psize)
3574 meta_size += meta_size;
3579 if (MDB_MSYNC(ptr, meta_size, rc)) {
3586 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3587 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3589 meta.mm_mapsize = mapsize;
3590 meta.mm_dbs[0] = txn->mt_dbs[0];
3591 meta.mm_dbs[1] = txn->mt_dbs[1];
3592 meta.mm_last_pg = txn->mt_next_pgno - 1;
3593 meta.mm_txnid = txn->mt_txnid;
3595 off = offsetof(MDB_meta, mm_mapsize);
3596 ptr = (char *)&meta + off;
3597 len = sizeof(MDB_meta) - off;
3599 off += env->me_psize;
3602 /* Write to the SYNC fd */
3603 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3604 env->me_fd : env->me_mfd;
3607 memset(&ov, 0, sizeof(ov));
3609 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3613 rc = pwrite(mfd, ptr, len, off);
3616 rc = rc < 0 ? ErrCode() : EIO;
3617 DPUTS("write failed, disk error?");
3618 /* On a failure, the pagecache still contains the new data.
3619 * Write some old data back, to prevent it from being used.
3620 * Use the non-SYNC fd; we know it will fail anyway.
3622 meta.mm_last_pg = metab.mm_last_pg;
3623 meta.mm_txnid = metab.mm_txnid;
3625 memset(&ov, 0, sizeof(ov));
3627 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3629 r2 = pwrite(env->me_fd, ptr, len, off);
3630 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3633 env->me_flags |= MDB_FATAL_ERROR;
3636 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3637 CACHEFLUSH(env->me_map + off, len, DCACHE);
3639 /* Memory ordering issues are irrelevant; since the entire writer
3640 * is wrapped by wmutex, all of these changes will become visible
3641 * after the wmutex is unlocked. Since the DB is multi-version,
3642 * readers will get consistent data regardless of how fresh or
3643 * how stale their view of these values is.
3646 env->me_txns->mti_txnid = txn->mt_txnid;
3651 /** Check both meta pages to see which one is newer.
3652 * @param[in] env the environment handle
3653 * @return meta toggle (0 or 1).
3656 mdb_env_pick_meta(const MDB_env *env)
3658 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3662 mdb_env_create(MDB_env **env)
3666 e = calloc(1, sizeof(MDB_env));
3670 e->me_maxreaders = DEFAULT_READERS;
3671 e->me_maxdbs = e->me_numdbs = 2;
3672 e->me_fd = INVALID_HANDLE_VALUE;
3673 e->me_lfd = INVALID_HANDLE_VALUE;
3674 e->me_mfd = INVALID_HANDLE_VALUE;
3675 #ifdef MDB_USE_POSIX_SEM
3676 e->me_rmutex = SEM_FAILED;
3677 e->me_wmutex = SEM_FAILED;
3679 e->me_pid = getpid();
3680 GET_PAGESIZE(e->me_os_psize);
3681 VGMEMP_CREATE(e,0,0);
3687 mdb_env_map(MDB_env *env, void *addr)
3690 unsigned int flags = env->me_flags;
3694 LONG sizelo, sizehi;
3697 if (flags & MDB_RDONLY) {
3698 /* Don't set explicit map size, use whatever exists */
3703 msize = env->me_mapsize;
3704 sizelo = msize & 0xffffffff;
3705 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3707 /* Windows won't create mappings for zero length files.
3708 * and won't map more than the file size.
3709 * Just set the maxsize right now.
3711 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3712 || !SetEndOfFile(env->me_fd)
3713 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3717 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3718 PAGE_READWRITE : PAGE_READONLY,
3719 sizehi, sizelo, NULL);
3722 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3723 FILE_MAP_WRITE : FILE_MAP_READ,
3725 rc = env->me_map ? 0 : ErrCode();
3730 int prot = PROT_READ;
3731 if (flags & MDB_WRITEMAP) {
3733 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3736 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3738 if (env->me_map == MAP_FAILED) {
3743 if (flags & MDB_NORDAHEAD) {
3744 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3746 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3748 #ifdef POSIX_MADV_RANDOM
3749 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3750 #endif /* POSIX_MADV_RANDOM */
3751 #endif /* MADV_RANDOM */
3755 /* Can happen because the address argument to mmap() is just a
3756 * hint. mmap() can pick another, e.g. if the range is in use.
3757 * The MAP_FIXED flag would prevent that, but then mmap could
3758 * instead unmap existing pages to make room for the new map.
3760 if (addr && env->me_map != addr)
3761 return EBUSY; /* TODO: Make a new MDB_* error code? */
3763 p = (MDB_page *)env->me_map;
3764 env->me_metas[0] = METADATA(p);
3765 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3771 mdb_env_set_mapsize(MDB_env *env, size_t size)
3773 /* If env is already open, caller is responsible for making
3774 * sure there are no active txns.
3782 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3783 else if (size < env->me_mapsize) {
3784 /* If the configured size is smaller, make sure it's
3785 * still big enough. Silently round up to minimum if not.
3787 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3791 munmap(env->me_map, env->me_mapsize);
3792 env->me_mapsize = size;
3793 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3794 rc = mdb_env_map(env, old);
3798 env->me_mapsize = size;
3800 env->me_maxpg = env->me_mapsize / env->me_psize;
3805 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3809 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3814 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3816 if (env->me_map || readers < 1)
3818 env->me_maxreaders = readers;
3823 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3825 if (!env || !readers)
3827 *readers = env->me_maxreaders;
3831 /** Further setup required for opening an LMDB environment
3834 mdb_env_open2(MDB_env *env)
3836 unsigned int flags = env->me_flags;
3837 int i, newenv = 0, rc;
3841 /* See if we should use QueryLimited */
3843 if ((rc & 0xff) > 5)
3844 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3846 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3849 memset(&meta, 0, sizeof(meta));
3851 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3854 DPUTS("new mdbenv");
3856 env->me_psize = env->me_os_psize;
3857 if (env->me_psize > MAX_PAGESIZE)
3858 env->me_psize = MAX_PAGESIZE;
3860 env->me_psize = meta.mm_psize;
3863 /* Was a mapsize configured? */
3864 if (!env->me_mapsize) {
3865 /* If this is a new environment, take the default,
3866 * else use the size recorded in the existing env.
3868 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3869 } else if (env->me_mapsize < meta.mm_mapsize) {
3870 /* If the configured size is smaller, make sure it's
3871 * still big enough. Silently round up to minimum if not.
3873 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3874 if (env->me_mapsize < minsize)
3875 env->me_mapsize = minsize;
3878 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3883 if (flags & MDB_FIXEDMAP)
3884 meta.mm_address = env->me_map;
3885 i = mdb_env_init_meta(env, &meta);
3886 if (i != MDB_SUCCESS) {
3891 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3892 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3894 #if !(MDB_MAXKEYSIZE)
3895 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3897 env->me_maxpg = env->me_mapsize / env->me_psize;
3901 int toggle = mdb_env_pick_meta(env);
3902 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3904 DPRINTF(("opened database version %u, pagesize %u",
3905 env->me_metas[0]->mm_version, env->me_psize));
3906 DPRINTF(("using meta page %d", toggle));
3907 DPRINTF(("depth: %u", db->md_depth));
3908 DPRINTF(("entries: %"Z"u", db->md_entries));
3909 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3910 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3911 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3912 DPRINTF(("root: %"Z"u", db->md_root));
3920 /** Release a reader thread's slot in the reader lock table.
3921 * This function is called automatically when a thread exits.
3922 * @param[in] ptr This points to the slot in the reader lock table.
3925 mdb_env_reader_dest(void *ptr)
3927 MDB_reader *reader = ptr;
3933 /** Junk for arranging thread-specific callbacks on Windows. This is
3934 * necessarily platform and compiler-specific. Windows supports up
3935 * to 1088 keys. Let's assume nobody opens more than 64 environments
3936 * in a single process, for now. They can override this if needed.
3938 #ifndef MAX_TLS_KEYS
3939 #define MAX_TLS_KEYS 64
3941 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3942 static int mdb_tls_nkeys;
3944 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3948 case DLL_PROCESS_ATTACH: break;
3949 case DLL_THREAD_ATTACH: break;
3950 case DLL_THREAD_DETACH:
3951 for (i=0; i<mdb_tls_nkeys; i++) {
3952 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3954 mdb_env_reader_dest(r);
3958 case DLL_PROCESS_DETACH: break;
3963 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3965 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3969 /* Force some symbol references.
3970 * _tls_used forces the linker to create the TLS directory if not already done
3971 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3973 #pragma comment(linker, "/INCLUDE:_tls_used")
3974 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3975 #pragma const_seg(".CRT$XLB")
3976 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3977 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3980 #pragma comment(linker, "/INCLUDE:__tls_used")
3981 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3982 #pragma data_seg(".CRT$XLB")
3983 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3985 #endif /* WIN 32/64 */
3986 #endif /* !__GNUC__ */
3989 /** Downgrade the exclusive lock on the region back to shared */
3991 mdb_env_share_locks(MDB_env *env, int *excl)
3993 int rc = 0, toggle = mdb_env_pick_meta(env);
3995 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4000 /* First acquire a shared lock. The Unlock will
4001 * then release the existing exclusive lock.
4003 memset(&ov, 0, sizeof(ov));
4004 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4007 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4013 struct flock lock_info;
4014 /* The shared lock replaces the existing lock */
4015 memset((void *)&lock_info, 0, sizeof(lock_info));
4016 lock_info.l_type = F_RDLCK;
4017 lock_info.l_whence = SEEK_SET;
4018 lock_info.l_start = 0;
4019 lock_info.l_len = 1;
4020 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4021 (rc = ErrCode()) == EINTR) ;
4022 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4029 /** Try to get exlusive lock, otherwise shared.
4030 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4033 mdb_env_excl_lock(MDB_env *env, int *excl)
4037 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4041 memset(&ov, 0, sizeof(ov));
4042 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4049 struct flock lock_info;
4050 memset((void *)&lock_info, 0, sizeof(lock_info));
4051 lock_info.l_type = F_WRLCK;
4052 lock_info.l_whence = SEEK_SET;
4053 lock_info.l_start = 0;
4054 lock_info.l_len = 1;
4055 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4056 (rc = ErrCode()) == EINTR) ;
4060 # ifdef MDB_USE_POSIX_SEM
4061 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4064 lock_info.l_type = F_RDLCK;
4065 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4066 (rc = ErrCode()) == EINTR) ;
4076 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4078 * @(#) $Revision: 5.1 $
4079 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4080 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4082 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4086 * Please do not copyright this code. This code is in the public domain.
4088 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4089 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4090 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4091 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4092 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4093 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4094 * PERFORMANCE OF THIS SOFTWARE.
4097 * chongo <Landon Curt Noll> /\oo/\
4098 * http://www.isthe.com/chongo/
4100 * Share and Enjoy! :-)
4103 typedef unsigned long long mdb_hash_t;
4104 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4106 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4107 * @param[in] val value to hash
4108 * @param[in] hval initial value for hash
4109 * @return 64 bit hash
4111 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4112 * hval arg on the first call.
4115 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4117 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4118 unsigned char *end = s + val->mv_size;
4120 * FNV-1a hash each octet of the string
4123 /* xor the bottom with the current octet */
4124 hval ^= (mdb_hash_t)*s++;
4126 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4127 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4128 (hval << 7) + (hval << 8) + (hval << 40);
4130 /* return our new hash value */
4134 /** Hash the string and output the encoded hash.
4135 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4136 * very short name limits. We don't care about the encoding being reversible,
4137 * we just want to preserve as many bits of the input as possible in a
4138 * small printable string.
4139 * @param[in] str string to hash
4140 * @param[out] encbuf an array of 11 chars to hold the hash
4142 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4145 mdb_pack85(unsigned long l, char *out)
4149 for (i=0; i<5; i++) {
4150 *out++ = mdb_a85[l % 85];
4156 mdb_hash_enc(MDB_val *val, char *encbuf)
4158 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4160 mdb_pack85(h, encbuf);
4161 mdb_pack85(h>>32, encbuf+5);
4166 /** Open and/or initialize the lock region for the environment.
4167 * @param[in] env The LMDB environment.
4168 * @param[in] lpath The pathname of the file used for the lock region.
4169 * @param[in] mode The Unix permissions for the file, if we create it.
4170 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4171 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4172 * @return 0 on success, non-zero on failure.
4175 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4178 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4180 # define MDB_ERRCODE_ROFS EROFS
4181 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4182 # define MDB_CLOEXEC O_CLOEXEC
4185 # define MDB_CLOEXEC 0
4192 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4193 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4194 FILE_ATTRIBUTE_NORMAL, NULL);
4196 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4198 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4200 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4205 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4206 /* Lose record locks when exec*() */
4207 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4208 fcntl(env->me_lfd, F_SETFD, fdflags);
4211 if (!(env->me_flags & MDB_NOTLS)) {
4212 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4215 env->me_flags |= MDB_ENV_TXKEY;
4217 /* Windows TLS callbacks need help finding their TLS info. */
4218 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4222 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4226 /* Try to get exclusive lock. If we succeed, then
4227 * nobody is using the lock region and we should initialize it.
4229 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4232 size = GetFileSize(env->me_lfd, NULL);
4234 size = lseek(env->me_lfd, 0, SEEK_END);
4235 if (size == -1) goto fail_errno;
4237 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4238 if (size < rsize && *excl > 0) {
4240 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4241 || !SetEndOfFile(env->me_lfd))
4244 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4248 size = rsize - sizeof(MDB_txninfo);
4249 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4254 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4256 if (!mh) goto fail_errno;
4257 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4259 if (!env->me_txns) goto fail_errno;
4261 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4263 if (m == MAP_FAILED) goto fail_errno;
4269 BY_HANDLE_FILE_INFORMATION stbuf;
4278 if (!mdb_sec_inited) {
4279 InitializeSecurityDescriptor(&mdb_null_sd,
4280 SECURITY_DESCRIPTOR_REVISION);
4281 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4282 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4283 mdb_all_sa.bInheritHandle = FALSE;
4284 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4287 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4288 idbuf.volume = stbuf.dwVolumeSerialNumber;
4289 idbuf.nhigh = stbuf.nFileIndexHigh;
4290 idbuf.nlow = stbuf.nFileIndexLow;
4291 val.mv_data = &idbuf;
4292 val.mv_size = sizeof(idbuf);
4293 mdb_hash_enc(&val, encbuf);
4294 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4295 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4296 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4297 if (!env->me_rmutex) goto fail_errno;
4298 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4299 if (!env->me_wmutex) goto fail_errno;
4300 #elif defined(MDB_USE_POSIX_SEM)
4309 #if defined(__NetBSD__)
4310 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4312 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4313 idbuf.dev = stbuf.st_dev;
4314 idbuf.ino = stbuf.st_ino;
4315 val.mv_data = &idbuf;
4316 val.mv_size = sizeof(idbuf);
4317 mdb_hash_enc(&val, encbuf);
4318 #ifdef MDB_SHORT_SEMNAMES
4319 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4321 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4322 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4323 /* Clean up after a previous run, if needed: Try to
4324 * remove both semaphores before doing anything else.
4326 sem_unlink(env->me_txns->mti_rmname);
4327 sem_unlink(env->me_txns->mti_wmname);
4328 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4329 O_CREAT|O_EXCL, mode, 1);
4330 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4331 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4332 O_CREAT|O_EXCL, mode, 1);
4333 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4334 #else /* MDB_USE_POSIX_SEM */
4335 pthread_mutexattr_t mattr;
4337 if ((rc = pthread_mutexattr_init(&mattr))
4338 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4339 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4340 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4342 pthread_mutexattr_destroy(&mattr);
4343 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4345 env->me_txns->mti_magic = MDB_MAGIC;
4346 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4347 env->me_txns->mti_txnid = 0;
4348 env->me_txns->mti_numreaders = 0;
4351 if (env->me_txns->mti_magic != MDB_MAGIC) {
4352 DPUTS("lock region has invalid magic");
4356 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4357 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4358 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4359 rc = MDB_VERSION_MISMATCH;
4363 if (rc && rc != EACCES && rc != EAGAIN) {
4367 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4368 if (!env->me_rmutex) goto fail_errno;
4369 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4370 if (!env->me_wmutex) goto fail_errno;
4371 #elif defined(MDB_USE_POSIX_SEM)
4372 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4373 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4374 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4375 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4386 /** The name of the lock file in the DB environment */
4387 #define LOCKNAME "/lock.mdb"
4388 /** The name of the data file in the DB environment */
4389 #define DATANAME "/data.mdb"
4390 /** The suffix of the lock file when no subdir is used */
4391 #define LOCKSUFF "-lock"
4392 /** Only a subset of the @ref mdb_env flags can be changed
4393 * at runtime. Changing other flags requires closing the
4394 * environment and re-opening it with the new flags.
4396 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4397 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4398 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4400 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4401 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4405 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4407 int oflags, rc, len, excl = -1;
4408 char *lpath, *dpath;
4410 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4414 if (flags & MDB_NOSUBDIR) {
4415 rc = len + sizeof(LOCKSUFF) + len + 1;
4417 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4422 if (flags & MDB_NOSUBDIR) {
4423 dpath = lpath + len + sizeof(LOCKSUFF);
4424 sprintf(lpath, "%s" LOCKSUFF, path);
4425 strcpy(dpath, path);
4427 dpath = lpath + len + sizeof(LOCKNAME);
4428 sprintf(lpath, "%s" LOCKNAME, path);
4429 sprintf(dpath, "%s" DATANAME, path);
4433 flags |= env->me_flags;
4434 if (flags & MDB_RDONLY) {
4435 /* silently ignore WRITEMAP when we're only getting read access */
4436 flags &= ~MDB_WRITEMAP;
4438 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4439 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4442 env->me_flags = flags |= MDB_ENV_ACTIVE;
4446 env->me_path = strdup(path);
4447 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4448 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4449 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4450 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4455 /* For RDONLY, get lockfile after we know datafile exists */
4456 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4457 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4463 if (F_ISSET(flags, MDB_RDONLY)) {
4464 oflags = GENERIC_READ;
4465 len = OPEN_EXISTING;
4467 oflags = GENERIC_READ|GENERIC_WRITE;
4470 mode = FILE_ATTRIBUTE_NORMAL;
4471 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4472 NULL, len, mode, NULL);
4474 if (F_ISSET(flags, MDB_RDONLY))
4477 oflags = O_RDWR | O_CREAT;
4479 env->me_fd = open(dpath, oflags, mode);
4481 if (env->me_fd == INVALID_HANDLE_VALUE) {
4486 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4487 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4492 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4493 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4494 env->me_mfd = env->me_fd;
4496 /* Synchronous fd for meta writes. Needed even with
4497 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4500 len = OPEN_EXISTING;
4501 env->me_mfd = CreateFile(dpath, oflags,
4502 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4503 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4506 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4508 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4513 DPRINTF(("opened dbenv %p", (void *) env));
4515 rc = mdb_env_share_locks(env, &excl);
4519 if (!((flags & MDB_RDONLY) ||
4520 (env->me_pbuf = calloc(1, env->me_psize))))
4522 if (!(flags & MDB_RDONLY)) {
4524 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4525 (sizeof(MDB_db)+sizeof(MDB_cursor)+sizeof(unsigned int)+1);
4526 txn = calloc(1, size);
4528 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4529 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4530 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4531 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4533 txn->mt_dbxs = env->me_dbxs;
4543 mdb_env_close0(env, excl);
4549 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4551 mdb_env_close0(MDB_env *env, int excl)
4555 if (!(env->me_flags & MDB_ENV_ACTIVE))
4558 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4559 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4560 free(env->me_dbxs[i].md_name.mv_data);
4563 free(env->me_dbiseqs);
4564 free(env->me_dbflags);
4567 free(env->me_dirty_list);
4569 mdb_midl_free(env->me_free_pgs);
4571 if (env->me_flags & MDB_ENV_TXKEY) {
4572 pthread_key_delete(env->me_txkey);
4574 /* Delete our key from the global list */
4575 for (i=0; i<mdb_tls_nkeys; i++)
4576 if (mdb_tls_keys[i] == env->me_txkey) {
4577 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4585 munmap(env->me_map, env->me_mapsize);
4587 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4588 (void) close(env->me_mfd);
4589 if (env->me_fd != INVALID_HANDLE_VALUE)
4590 (void) close(env->me_fd);
4592 MDB_PID_T pid = env->me_pid;
4593 /* Clearing readers is done in this function because
4594 * me_txkey with its destructor must be disabled first.
4596 for (i = env->me_numreaders; --i >= 0; )
4597 if (env->me_txns->mti_readers[i].mr_pid == pid)
4598 env->me_txns->mti_readers[i].mr_pid = 0;
4600 if (env->me_rmutex) {
4601 CloseHandle(env->me_rmutex);
4602 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4604 /* Windows automatically destroys the mutexes when
4605 * the last handle closes.
4607 #elif defined(MDB_USE_POSIX_SEM)
4608 if (env->me_rmutex != SEM_FAILED) {
4609 sem_close(env->me_rmutex);
4610 if (env->me_wmutex != SEM_FAILED)
4611 sem_close(env->me_wmutex);
4612 /* If we have the filelock: If we are the
4613 * only remaining user, clean up semaphores.
4616 mdb_env_excl_lock(env, &excl);
4618 sem_unlink(env->me_txns->mti_rmname);
4619 sem_unlink(env->me_txns->mti_wmname);
4623 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4625 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4628 /* Unlock the lockfile. Windows would have unlocked it
4629 * after closing anyway, but not necessarily at once.
4631 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4634 (void) close(env->me_lfd);
4637 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4642 mdb_env_close(MDB_env *env)
4649 VGMEMP_DESTROY(env);
4650 while ((dp = env->me_dpages) != NULL) {
4651 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4652 env->me_dpages = dp->mp_next;
4656 mdb_env_close0(env, 0);
4660 /** Compare two items pointing at aligned size_t's */
4662 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4664 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4665 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4668 /** Compare two items pointing at aligned unsigned int's */
4670 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4672 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4673 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4676 /** Compare two items pointing at unsigned ints of unknown alignment.
4677 * Nodes and keys are guaranteed to be 2-byte aligned.
4680 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4682 #if BYTE_ORDER == LITTLE_ENDIAN
4683 unsigned short *u, *c;
4686 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4687 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4690 } while(!x && u > (unsigned short *)a->mv_data);
4693 unsigned short *u, *c, *end;
4696 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4697 u = (unsigned short *)a->mv_data;
4698 c = (unsigned short *)b->mv_data;
4701 } while(!x && u < end);
4706 /** Compare two items pointing at size_t's of unknown alignment. */
4707 #ifdef MISALIGNED_OK
4708 # define mdb_cmp_clong mdb_cmp_long
4710 # define mdb_cmp_clong mdb_cmp_cint
4713 /** Compare two items lexically */
4715 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4722 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4728 diff = memcmp(a->mv_data, b->mv_data, len);
4729 return diff ? diff : len_diff<0 ? -1 : len_diff;
4732 /** Compare two items in reverse byte order */
4734 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4736 const unsigned char *p1, *p2, *p1_lim;
4740 p1_lim = (const unsigned char *)a->mv_data;
4741 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4742 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4744 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4750 while (p1 > p1_lim) {
4751 diff = *--p1 - *--p2;
4755 return len_diff<0 ? -1 : len_diff;
4758 /** Search for key within a page, using binary search.
4759 * Returns the smallest entry larger or equal to the key.
4760 * If exactp is non-null, stores whether the found entry was an exact match
4761 * in *exactp (1 or 0).
4762 * Updates the cursor index with the index of the found entry.
4763 * If no entry larger or equal to the key is found, returns NULL.
4766 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4768 unsigned int i = 0, nkeys;
4771 MDB_page *mp = mc->mc_pg[mc->mc_top];
4772 MDB_node *node = NULL;
4777 nkeys = NUMKEYS(mp);
4779 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4780 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4783 low = IS_LEAF(mp) ? 0 : 1;
4785 cmp = mc->mc_dbx->md_cmp;
4787 /* Branch pages have no data, so if using integer keys,
4788 * alignment is guaranteed. Use faster mdb_cmp_int.
4790 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4791 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4798 nodekey.mv_size = mc->mc_db->md_pad;
4799 node = NODEPTR(mp, 0); /* fake */
4800 while (low <= high) {
4801 i = (low + high) >> 1;
4802 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4803 rc = cmp(key, &nodekey);
4804 DPRINTF(("found leaf index %u [%s], rc = %i",
4805 i, DKEY(&nodekey), rc));
4814 while (low <= high) {
4815 i = (low + high) >> 1;
4817 node = NODEPTR(mp, i);
4818 nodekey.mv_size = NODEKSZ(node);
4819 nodekey.mv_data = NODEKEY(node);
4821 rc = cmp(key, &nodekey);
4824 DPRINTF(("found leaf index %u [%s], rc = %i",
4825 i, DKEY(&nodekey), rc));
4827 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4828 i, DKEY(&nodekey), NODEPGNO(node), rc));
4839 if (rc > 0) { /* Found entry is less than the key. */
4840 i++; /* Skip to get the smallest entry larger than key. */
4842 node = NODEPTR(mp, i);
4845 *exactp = (rc == 0 && nkeys > 0);
4846 /* store the key index */
4847 mc->mc_ki[mc->mc_top] = i;
4849 /* There is no entry larger or equal to the key. */
4852 /* nodeptr is fake for LEAF2 */
4858 mdb_cursor_adjust(MDB_cursor *mc, func)
4862 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4863 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4870 /** Pop a page off the top of the cursor's stack. */
4872 mdb_cursor_pop(MDB_cursor *mc)
4876 MDB_page *top = mc->mc_pg[mc->mc_top];
4882 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4883 DDBI(mc), (void *) mc));
4887 /** Push a page onto the top of the cursor's stack. */
4889 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4891 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4892 DDBI(mc), (void *) mc));
4894 if (mc->mc_snum >= CURSOR_STACK) {
4895 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4896 return MDB_CURSOR_FULL;
4899 mc->mc_top = mc->mc_snum++;
4900 mc->mc_pg[mc->mc_top] = mp;
4901 mc->mc_ki[mc->mc_top] = 0;
4906 /** Find the address of the page corresponding to a given page number.
4907 * @param[in] txn the transaction for this access.
4908 * @param[in] pgno the page number for the page to retrieve.
4909 * @param[out] ret address of a pointer where the page's address will be stored.
4910 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4911 * @return 0 on success, non-zero on failure.
4914 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4916 MDB_env *env = txn->mt_env;
4920 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4924 MDB_ID2L dl = tx2->mt_u.dirty_list;
4926 /* Spilled pages were dirtied in this txn and flushed
4927 * because the dirty list got full. Bring this page
4928 * back in from the map (but don't unspill it here,
4929 * leave that unless page_touch happens again).
4931 if (tx2->mt_spill_pgs) {
4932 MDB_ID pn = pgno << 1;
4933 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4934 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4935 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4940 unsigned x = mdb_mid2l_search(dl, pgno);
4941 if (x <= dl[0].mid && dl[x].mid == pgno) {
4947 } while ((tx2 = tx2->mt_parent) != NULL);
4950 if (pgno < txn->mt_next_pgno) {
4952 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4954 DPRINTF(("page %"Z"u not found", pgno));
4955 txn->mt_flags |= MDB_TXN_ERROR;
4956 return MDB_PAGE_NOTFOUND;
4966 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4967 * The cursor is at the root page, set up the rest of it.
4970 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4972 MDB_page *mp = mc->mc_pg[mc->mc_top];
4976 while (IS_BRANCH(mp)) {
4980 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4981 mdb_cassert(mc, NUMKEYS(mp) > 1);
4982 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4984 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4986 if (flags & MDB_PS_LAST)
4987 i = NUMKEYS(mp) - 1;
4990 node = mdb_node_search(mc, key, &exact);
4992 i = NUMKEYS(mp) - 1;
4994 i = mc->mc_ki[mc->mc_top];
4996 mdb_cassert(mc, i > 0);
5000 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5003 mdb_cassert(mc, i < NUMKEYS(mp));
5004 node = NODEPTR(mp, i);
5006 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5009 mc->mc_ki[mc->mc_top] = i;
5010 if ((rc = mdb_cursor_push(mc, mp)))
5013 if (flags & MDB_PS_MODIFY) {
5014 if ((rc = mdb_page_touch(mc)) != 0)
5016 mp = mc->mc_pg[mc->mc_top];
5021 DPRINTF(("internal error, index points to a %02X page!?",
5023 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5024 return MDB_CORRUPTED;
5027 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5028 key ? DKEY(key) : "null"));
5029 mc->mc_flags |= C_INITIALIZED;
5030 mc->mc_flags &= ~C_EOF;
5035 /** Search for the lowest key under the current branch page.
5036 * This just bypasses a NUMKEYS check in the current page
5037 * before calling mdb_page_search_root(), because the callers
5038 * are all in situations where the current page is known to
5042 mdb_page_search_lowest(MDB_cursor *mc)
5044 MDB_page *mp = mc->mc_pg[mc->mc_top];
5045 MDB_node *node = NODEPTR(mp, 0);
5048 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5051 mc->mc_ki[mc->mc_top] = 0;
5052 if ((rc = mdb_cursor_push(mc, mp)))
5054 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5057 /** Search for the page a given key should be in.
5058 * Push it and its parent pages on the cursor stack.
5059 * @param[in,out] mc the cursor for this operation.
5060 * @param[in] key the key to search for, or NULL for first/last page.
5061 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5062 * are touched (updated with new page numbers).
5063 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5064 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5065 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5066 * @return 0 on success, non-zero on failure.
5069 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5074 /* Make sure the txn is still viable, then find the root from
5075 * the txn's db table and set it as the root of the cursor's stack.
5077 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5078 DPUTS("transaction has failed, must abort");
5081 /* Make sure we're using an up-to-date root */
5082 if (*mc->mc_dbflag & DB_STALE) {
5084 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5086 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5087 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5094 MDB_node *leaf = mdb_node_search(&mc2,
5095 &mc->mc_dbx->md_name, &exact);
5097 return MDB_NOTFOUND;
5098 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5101 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5103 /* The txn may not know this DBI, or another process may
5104 * have dropped and recreated the DB with other flags.
5106 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5107 return MDB_INCOMPATIBLE;
5108 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5110 *mc->mc_dbflag &= ~DB_STALE;
5112 root = mc->mc_db->md_root;
5114 if (root == P_INVALID) { /* Tree is empty. */
5115 DPUTS("tree is empty");
5116 return MDB_NOTFOUND;
5120 mdb_cassert(mc, root > 1);
5121 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5122 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5128 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5129 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5131 if (flags & MDB_PS_MODIFY) {
5132 if ((rc = mdb_page_touch(mc)))
5136 if (flags & MDB_PS_ROOTONLY)
5139 return mdb_page_search_root(mc, key, flags);
5143 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5145 MDB_txn *txn = mc->mc_txn;
5146 pgno_t pg = mp->mp_pgno;
5147 unsigned x = 0, ovpages = mp->mp_pages;
5148 MDB_env *env = txn->mt_env;
5149 MDB_IDL sl = txn->mt_spill_pgs;
5150 MDB_ID pn = pg << 1;
5153 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5154 /* If the page is dirty or on the spill list we just acquired it,
5155 * so we should give it back to our current free list, if any.
5156 * Otherwise put it onto the list of pages we freed in this txn.
5158 * Won't create me_pghead: me_pglast must be inited along with it.
5159 * Unsupported in nested txns: They would need to hide the page
5160 * range in ancestor txns' dirty and spilled lists.
5162 if (env->me_pghead &&
5164 ((mp->mp_flags & P_DIRTY) ||
5165 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5169 MDB_ID2 *dl, ix, iy;
5170 rc = mdb_midl_need(&env->me_pghead, ovpages);
5173 if (!(mp->mp_flags & P_DIRTY)) {
5174 /* This page is no longer spilled */
5181 /* Remove from dirty list */
5182 dl = txn->mt_u.dirty_list;
5184 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5190 mdb_cassert(mc, x > 1);
5192 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5193 txn->mt_flags |= MDB_TXN_ERROR;
5194 return MDB_CORRUPTED;
5197 if (!(env->me_flags & MDB_WRITEMAP))
5198 mdb_dpage_free(env, mp);
5200 /* Insert in me_pghead */
5201 mop = env->me_pghead;
5202 j = mop[0] + ovpages;
5203 for (i = mop[0]; i && mop[i] < pg; i--)
5209 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5213 mc->mc_db->md_overflow_pages -= ovpages;
5217 /** Return the data associated with a given node.
5218 * @param[in] txn The transaction for this operation.
5219 * @param[in] leaf The node being read.
5220 * @param[out] data Updated to point to the node's data.
5221 * @return 0 on success, non-zero on failure.
5224 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5226 MDB_page *omp; /* overflow page */
5230 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5231 data->mv_size = NODEDSZ(leaf);
5232 data->mv_data = NODEDATA(leaf);
5236 /* Read overflow data.
5238 data->mv_size = NODEDSZ(leaf);
5239 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5240 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5241 DPRINTF(("read overflow page %"Z"u failed", pgno));
5244 data->mv_data = METADATA(omp);
5250 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5251 MDB_val *key, MDB_val *data)
5258 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5260 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5263 if (txn->mt_flags & MDB_TXN_ERROR)
5266 mdb_cursor_init(&mc, txn, dbi, &mx);
5267 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5270 /** Find a sibling for a page.
5271 * Replaces the page at the top of the cursor's stack with the
5272 * specified sibling, if one exists.
5273 * @param[in] mc The cursor for this operation.
5274 * @param[in] move_right Non-zero if the right sibling is requested,
5275 * otherwise the left sibling.
5276 * @return 0 on success, non-zero on failure.
5279 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5285 if (mc->mc_snum < 2) {
5286 return MDB_NOTFOUND; /* root has no siblings */
5290 DPRINTF(("parent page is page %"Z"u, index %u",
5291 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5293 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5294 : (mc->mc_ki[mc->mc_top] == 0)) {
5295 DPRINTF(("no more keys left, moving to %s sibling",
5296 move_right ? "right" : "left"));
5297 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5298 /* undo cursor_pop before returning */
5305 mc->mc_ki[mc->mc_top]++;
5307 mc->mc_ki[mc->mc_top]--;
5308 DPRINTF(("just moving to %s index key %u",
5309 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5311 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5313 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5314 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5315 /* mc will be inconsistent if caller does mc_snum++ as above */
5316 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5320 mdb_cursor_push(mc, mp);
5322 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5327 /** Move the cursor to the next data item. */
5329 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5335 if (mc->mc_flags & C_EOF) {
5336 return MDB_NOTFOUND;
5339 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5341 mp = mc->mc_pg[mc->mc_top];
5343 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5344 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5345 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5346 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5347 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5348 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5349 if (rc == MDB_SUCCESS)
5350 MDB_GET_KEY(leaf, key);
5355 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5356 if (op == MDB_NEXT_DUP)
5357 return MDB_NOTFOUND;
5361 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5362 mdb_dbg_pgno(mp), (void *) mc));
5363 if (mc->mc_flags & C_DEL)
5366 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5367 DPUTS("=====> move to next sibling page");
5368 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5369 mc->mc_flags |= C_EOF;
5372 mp = mc->mc_pg[mc->mc_top];
5373 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5375 mc->mc_ki[mc->mc_top]++;
5378 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5379 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5382 key->mv_size = mc->mc_db->md_pad;
5383 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5387 mdb_cassert(mc, IS_LEAF(mp));
5388 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5390 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5391 mdb_xcursor_init1(mc, leaf);
5394 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5397 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5398 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5399 if (rc != MDB_SUCCESS)
5404 MDB_GET_KEY(leaf, key);
5408 /** Move the cursor to the previous data item. */
5410 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5416 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5418 mp = mc->mc_pg[mc->mc_top];
5420 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5421 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5422 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5423 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5424 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5425 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5426 if (rc == MDB_SUCCESS) {
5427 MDB_GET_KEY(leaf, key);
5428 mc->mc_flags &= ~C_EOF;
5434 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5435 if (op == MDB_PREV_DUP)
5436 return MDB_NOTFOUND;
5440 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5441 mdb_dbg_pgno(mp), (void *) mc));
5443 if (mc->mc_ki[mc->mc_top] == 0) {
5444 DPUTS("=====> move to prev sibling page");
5445 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5448 mp = mc->mc_pg[mc->mc_top];
5449 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5450 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5452 mc->mc_ki[mc->mc_top]--;
5454 mc->mc_flags &= ~C_EOF;
5456 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5457 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5460 key->mv_size = mc->mc_db->md_pad;
5461 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5465 mdb_cassert(mc, IS_LEAF(mp));
5466 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5468 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5469 mdb_xcursor_init1(mc, leaf);
5472 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5475 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5476 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5477 if (rc != MDB_SUCCESS)
5482 MDB_GET_KEY(leaf, key);
5486 /** Set the cursor on a specific data item. */
5488 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5489 MDB_cursor_op op, int *exactp)
5493 MDB_node *leaf = NULL;
5496 if (key->mv_size == 0)
5497 return MDB_BAD_VALSIZE;
5500 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5502 /* See if we're already on the right page */
5503 if (mc->mc_flags & C_INITIALIZED) {
5506 mp = mc->mc_pg[mc->mc_top];
5508 mc->mc_ki[mc->mc_top] = 0;
5509 return MDB_NOTFOUND;
5511 if (mp->mp_flags & P_LEAF2) {
5512 nodekey.mv_size = mc->mc_db->md_pad;
5513 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5515 leaf = NODEPTR(mp, 0);
5516 MDB_GET_KEY2(leaf, nodekey);
5518 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5520 /* Probably happens rarely, but first node on the page
5521 * was the one we wanted.
5523 mc->mc_ki[mc->mc_top] = 0;
5530 unsigned int nkeys = NUMKEYS(mp);
5532 if (mp->mp_flags & P_LEAF2) {
5533 nodekey.mv_data = LEAF2KEY(mp,
5534 nkeys-1, nodekey.mv_size);
5536 leaf = NODEPTR(mp, nkeys-1);
5537 MDB_GET_KEY2(leaf, nodekey);
5539 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5541 /* last node was the one we wanted */
5542 mc->mc_ki[mc->mc_top] = nkeys-1;
5548 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5549 /* This is definitely the right page, skip search_page */
5550 if (mp->mp_flags & P_LEAF2) {
5551 nodekey.mv_data = LEAF2KEY(mp,
5552 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5554 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5555 MDB_GET_KEY2(leaf, nodekey);
5557 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5559 /* current node was the one we wanted */
5569 /* If any parents have right-sibs, search.
5570 * Otherwise, there's nothing further.
5572 for (i=0; i<mc->mc_top; i++)
5574 NUMKEYS(mc->mc_pg[i])-1)
5576 if (i == mc->mc_top) {
5577 /* There are no other pages */
5578 mc->mc_ki[mc->mc_top] = nkeys;
5579 return MDB_NOTFOUND;
5583 /* There are no other pages */
5584 mc->mc_ki[mc->mc_top] = 0;
5585 if (op == MDB_SET_RANGE && !exactp) {
5589 return MDB_NOTFOUND;
5593 rc = mdb_page_search(mc, key, 0);
5594 if (rc != MDB_SUCCESS)
5597 mp = mc->mc_pg[mc->mc_top];
5598 mdb_cassert(mc, IS_LEAF(mp));
5601 leaf = mdb_node_search(mc, key, exactp);
5602 if (exactp != NULL && !*exactp) {
5603 /* MDB_SET specified and not an exact match. */
5604 return MDB_NOTFOUND;
5608 DPUTS("===> inexact leaf not found, goto sibling");
5609 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5610 return rc; /* no entries matched */
5611 mp = mc->mc_pg[mc->mc_top];
5612 mdb_cassert(mc, IS_LEAF(mp));
5613 leaf = NODEPTR(mp, 0);
5617 mc->mc_flags |= C_INITIALIZED;
5618 mc->mc_flags &= ~C_EOF;
5621 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5622 key->mv_size = mc->mc_db->md_pad;
5623 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5628 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5629 mdb_xcursor_init1(mc, leaf);
5632 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5633 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5634 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5637 if (op == MDB_GET_BOTH) {
5643 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5644 if (rc != MDB_SUCCESS)
5647 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5649 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5651 rc = mc->mc_dbx->md_dcmp(data, &d2);
5653 if (op == MDB_GET_BOTH || rc > 0)
5654 return MDB_NOTFOUND;
5661 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5662 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5667 /* The key already matches in all other cases */
5668 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5669 MDB_GET_KEY(leaf, key);
5670 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5675 /** Move the cursor to the first item in the database. */
5677 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5683 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5685 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5686 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5687 if (rc != MDB_SUCCESS)
5690 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5692 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5693 mc->mc_flags |= C_INITIALIZED;
5694 mc->mc_flags &= ~C_EOF;
5696 mc->mc_ki[mc->mc_top] = 0;
5698 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5699 key->mv_size = mc->mc_db->md_pad;
5700 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5705 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5706 mdb_xcursor_init1(mc, leaf);
5707 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5711 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5715 MDB_GET_KEY(leaf, key);
5719 /** Move the cursor to the last item in the database. */
5721 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5727 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5729 if (!(mc->mc_flags & C_EOF)) {
5731 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5732 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5733 if (rc != MDB_SUCCESS)
5736 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5739 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5740 mc->mc_flags |= C_INITIALIZED|C_EOF;
5741 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5743 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5744 key->mv_size = mc->mc_db->md_pad;
5745 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5750 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5751 mdb_xcursor_init1(mc, leaf);
5752 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5756 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5761 MDB_GET_KEY(leaf, key);
5766 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5771 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5776 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5780 case MDB_GET_CURRENT:
5781 if (!(mc->mc_flags & C_INITIALIZED)) {
5784 MDB_page *mp = mc->mc_pg[mc->mc_top];
5785 int nkeys = NUMKEYS(mp);
5786 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5787 mc->mc_ki[mc->mc_top] = nkeys;
5793 key->mv_size = mc->mc_db->md_pad;
5794 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5796 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5797 MDB_GET_KEY(leaf, key);
5799 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5800 if (mc->mc_flags & C_DEL)
5801 mdb_xcursor_init1(mc, leaf);
5802 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5804 rc = mdb_node_read(mc->mc_txn, leaf, data);
5811 case MDB_GET_BOTH_RANGE:
5816 if (mc->mc_xcursor == NULL) {
5817 rc = MDB_INCOMPATIBLE;
5827 rc = mdb_cursor_set(mc, key, data, op,
5828 op == MDB_SET_RANGE ? NULL : &exact);
5831 case MDB_GET_MULTIPLE:
5832 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5836 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5837 rc = MDB_INCOMPATIBLE;
5841 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5842 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5845 case MDB_NEXT_MULTIPLE:
5850 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5851 rc = MDB_INCOMPATIBLE;
5854 if (!(mc->mc_flags & C_INITIALIZED))
5855 rc = mdb_cursor_first(mc, key, data);
5857 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5858 if (rc == MDB_SUCCESS) {
5859 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5862 mx = &mc->mc_xcursor->mx_cursor;
5863 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5865 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5866 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5874 case MDB_NEXT_NODUP:
5875 if (!(mc->mc_flags & C_INITIALIZED))
5876 rc = mdb_cursor_first(mc, key, data);
5878 rc = mdb_cursor_next(mc, key, data, op);
5882 case MDB_PREV_NODUP:
5883 if (!(mc->mc_flags & C_INITIALIZED)) {
5884 rc = mdb_cursor_last(mc, key, data);
5887 mc->mc_flags |= C_INITIALIZED;
5888 mc->mc_ki[mc->mc_top]++;
5890 rc = mdb_cursor_prev(mc, key, data, op);
5893 rc = mdb_cursor_first(mc, key, data);
5896 mfunc = mdb_cursor_first;
5898 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5902 if (mc->mc_xcursor == NULL) {
5903 rc = MDB_INCOMPATIBLE;
5907 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5908 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5909 MDB_GET_KEY(leaf, key);
5910 rc = mdb_node_read(mc->mc_txn, leaf, data);
5914 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5918 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5921 rc = mdb_cursor_last(mc, key, data);
5924 mfunc = mdb_cursor_last;
5927 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5932 if (mc->mc_flags & C_DEL)
5933 mc->mc_flags ^= C_DEL;
5938 /** Touch all the pages in the cursor stack. Set mc_top.
5939 * Makes sure all the pages are writable, before attempting a write operation.
5940 * @param[in] mc The cursor to operate on.
5943 mdb_cursor_touch(MDB_cursor *mc)
5945 int rc = MDB_SUCCESS;
5947 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5950 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5952 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5953 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5956 *mc->mc_dbflag |= DB_DIRTY;
5961 rc = mdb_page_touch(mc);
5962 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5963 mc->mc_top = mc->mc_snum-1;
5968 /** Do not spill pages to disk if txn is getting full, may fail instead */
5969 #define MDB_NOSPILL 0x8000
5972 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5975 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5977 MDB_node *leaf = NULL;
5980 MDB_val xdata, *rdata, dkey, olddata;
5982 int do_sub = 0, insert_key, insert_data;
5983 unsigned int mcount = 0, dcount = 0, nospill;
5986 unsigned int nflags;
5989 if (mc == NULL || key == NULL)
5992 env = mc->mc_txn->mt_env;
5994 /* Check this first so counter will always be zero on any
5997 if (flags & MDB_MULTIPLE) {
5998 dcount = data[1].mv_size;
5999 data[1].mv_size = 0;
6000 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6001 return MDB_INCOMPATIBLE;
6004 nospill = flags & MDB_NOSPILL;
6005 flags &= ~MDB_NOSPILL;
6007 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6008 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6010 if (key->mv_size-1 >= ENV_MAXKEY(env))
6011 return MDB_BAD_VALSIZE;
6013 #if SIZE_MAX > MAXDATASIZE
6014 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6015 return MDB_BAD_VALSIZE;
6017 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6018 return MDB_BAD_VALSIZE;
6021 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6022 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6026 if (flags == MDB_CURRENT) {
6027 if (!(mc->mc_flags & C_INITIALIZED))
6030 } else if (mc->mc_db->md_root == P_INVALID) {
6031 /* new database, cursor has nothing to point to */
6034 mc->mc_flags &= ~C_INITIALIZED;
6039 if (flags & MDB_APPEND) {
6041 rc = mdb_cursor_last(mc, &k2, &d2);
6043 rc = mc->mc_dbx->md_cmp(key, &k2);
6046 mc->mc_ki[mc->mc_top]++;
6048 /* new key is <= last key */
6053 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6055 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6056 DPRINTF(("duplicate key [%s]", DKEY(key)));
6058 return MDB_KEYEXIST;
6060 if (rc && rc != MDB_NOTFOUND)
6064 if (mc->mc_flags & C_DEL)
6065 mc->mc_flags ^= C_DEL;
6067 /* Cursor is positioned, check for room in the dirty list */
6069 if (flags & MDB_MULTIPLE) {
6071 xdata.mv_size = data->mv_size * dcount;
6075 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6079 if (rc == MDB_NO_ROOT) {
6081 /* new database, write a root leaf page */
6082 DPUTS("allocating new root leaf page");
6083 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6086 mdb_cursor_push(mc, np);
6087 mc->mc_db->md_root = np->mp_pgno;
6088 mc->mc_db->md_depth++;
6089 *mc->mc_dbflag |= DB_DIRTY;
6090 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6092 np->mp_flags |= P_LEAF2;
6093 mc->mc_flags |= C_INITIALIZED;
6095 /* make sure all cursor pages are writable */
6096 rc2 = mdb_cursor_touch(mc);
6101 insert_key = insert_data = rc;
6103 /* The key does not exist */
6104 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6105 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6106 LEAFSIZE(key, data) > env->me_nodemax)
6108 /* Too big for a node, insert in sub-DB. Set up an empty
6109 * "old sub-page" for prep_subDB to expand to a full page.
6111 fp_flags = P_LEAF|P_DIRTY;
6113 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6114 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6115 olddata.mv_size = PAGEHDRSZ;
6119 /* there's only a key anyway, so this is a no-op */
6120 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6122 unsigned int ksize = mc->mc_db->md_pad;
6123 if (key->mv_size != ksize)
6124 return MDB_BAD_VALSIZE;
6125 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6126 memcpy(ptr, key->mv_data, ksize);
6128 /* if overwriting slot 0 of leaf, need to
6129 * update branch key if there is a parent page
6131 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6132 unsigned short top = mc->mc_top;
6134 /* slot 0 is always an empty key, find real slot */
6135 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6137 if (mc->mc_ki[mc->mc_top])
6138 rc2 = mdb_update_key(mc, key);
6149 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6150 olddata.mv_size = NODEDSZ(leaf);
6151 olddata.mv_data = NODEDATA(leaf);
6154 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6155 /* Prepare (sub-)page/sub-DB to accept the new item,
6156 * if needed. fp: old sub-page or a header faking
6157 * it. mp: new (sub-)page. offset: growth in page
6158 * size. xdata: node data with new page or DB.
6160 unsigned i, offset = 0;
6161 mp = fp = xdata.mv_data = env->me_pbuf;
6162 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6164 /* Was a single item before, must convert now */
6165 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6166 /* Just overwrite the current item */
6167 if (flags == MDB_CURRENT)
6170 #if UINT_MAX < SIZE_MAX
6171 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6172 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6174 /* does data match? */
6175 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6176 if (flags & MDB_NODUPDATA)
6177 return MDB_KEYEXIST;
6182 /* Back up original data item */
6183 dkey.mv_size = olddata.mv_size;
6184 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6186 /* Make sub-page header for the dup items, with dummy body */
6187 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6188 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6189 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6190 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6191 fp->mp_flags |= P_LEAF2;
6192 fp->mp_pad = data->mv_size;
6193 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6195 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6196 (dkey.mv_size & 1) + (data->mv_size & 1);
6198 fp->mp_upper = xdata.mv_size - PAGEBASE;
6199 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6200 } else if (leaf->mn_flags & F_SUBDATA) {
6201 /* Data is on sub-DB, just store it */
6202 flags |= F_DUPDATA|F_SUBDATA;
6205 /* Data is on sub-page */
6206 fp = olddata.mv_data;
6209 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6210 offset = EVEN(NODESIZE + sizeof(indx_t) +
6214 offset = fp->mp_pad;
6215 if (SIZELEFT(fp) < offset) {
6216 offset *= 4; /* space for 4 more */
6219 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6221 fp->mp_flags |= P_DIRTY;
6222 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6223 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6227 xdata.mv_size = olddata.mv_size + offset;
6230 fp_flags = fp->mp_flags;
6231 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6232 /* Too big for a sub-page, convert to sub-DB */
6233 fp_flags &= ~P_SUBP;
6235 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6236 fp_flags |= P_LEAF2;
6237 dummy.md_pad = fp->mp_pad;
6238 dummy.md_flags = MDB_DUPFIXED;
6239 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6240 dummy.md_flags |= MDB_INTEGERKEY;
6246 dummy.md_branch_pages = 0;
6247 dummy.md_leaf_pages = 1;
6248 dummy.md_overflow_pages = 0;
6249 dummy.md_entries = NUMKEYS(fp);
6250 xdata.mv_size = sizeof(MDB_db);
6251 xdata.mv_data = &dummy;
6252 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6254 offset = env->me_psize - olddata.mv_size;
6255 flags |= F_DUPDATA|F_SUBDATA;
6256 dummy.md_root = mp->mp_pgno;
6259 mp->mp_flags = fp_flags | P_DIRTY;
6260 mp->mp_pad = fp->mp_pad;
6261 mp->mp_lower = fp->mp_lower;
6262 mp->mp_upper = fp->mp_upper + offset;
6263 if (fp_flags & P_LEAF2) {
6264 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6266 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6267 olddata.mv_size - fp->mp_upper - PAGEBASE);
6268 for (i=0; i<NUMKEYS(fp); i++)
6269 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6277 mdb_node_del(mc, 0);
6281 /* overflow page overwrites need special handling */
6282 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6285 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6287 memcpy(&pg, olddata.mv_data, sizeof(pg));
6288 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6290 ovpages = omp->mp_pages;
6292 /* Is the ov page large enough? */
6293 if (ovpages >= dpages) {
6294 if (!(omp->mp_flags & P_DIRTY) &&
6295 (level || (env->me_flags & MDB_WRITEMAP)))
6297 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6300 level = 0; /* dirty in this txn or clean */
6303 if (omp->mp_flags & P_DIRTY) {
6304 /* yes, overwrite it. Note in this case we don't
6305 * bother to try shrinking the page if the new data
6306 * is smaller than the overflow threshold.
6309 /* It is writable only in a parent txn */
6310 size_t sz = (size_t) env->me_psize * ovpages, off;
6311 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6317 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6318 mdb_cassert(mc, rc2 == 0);
6319 if (!(flags & MDB_RESERVE)) {
6320 /* Copy end of page, adjusting alignment so
6321 * compiler may copy words instead of bytes.
6323 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6324 memcpy((size_t *)((char *)np + off),
6325 (size_t *)((char *)omp + off), sz - off);
6328 memcpy(np, omp, sz); /* Copy beginning of page */
6331 SETDSZ(leaf, data->mv_size);
6332 if (F_ISSET(flags, MDB_RESERVE))
6333 data->mv_data = METADATA(omp);
6335 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6339 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6341 } else if (data->mv_size == olddata.mv_size) {
6342 /* same size, just replace it. Note that we could
6343 * also reuse this node if the new data is smaller,
6344 * but instead we opt to shrink the node in that case.
6346 if (F_ISSET(flags, MDB_RESERVE))
6347 data->mv_data = olddata.mv_data;
6348 else if (!(mc->mc_flags & C_SUB))
6349 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6351 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6356 mdb_node_del(mc, 0);
6362 nflags = flags & NODE_ADD_FLAGS;
6363 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6364 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6365 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6366 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6368 nflags |= MDB_SPLIT_REPLACE;
6369 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6371 /* There is room already in this leaf page. */
6372 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6373 if (rc == 0 && insert_key) {
6374 /* Adjust other cursors pointing to mp */
6375 MDB_cursor *m2, *m3;
6376 MDB_dbi dbi = mc->mc_dbi;
6377 unsigned i = mc->mc_top;
6378 MDB_page *mp = mc->mc_pg[i];
6380 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6381 if (mc->mc_flags & C_SUB)
6382 m3 = &m2->mc_xcursor->mx_cursor;
6385 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6386 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6393 if (rc == MDB_SUCCESS) {
6394 /* Now store the actual data in the child DB. Note that we're
6395 * storing the user data in the keys field, so there are strict
6396 * size limits on dupdata. The actual data fields of the child
6397 * DB are all zero size.
6405 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6406 if (flags & MDB_CURRENT) {
6407 xflags = MDB_CURRENT|MDB_NOSPILL;
6409 mdb_xcursor_init1(mc, leaf);
6410 xflags = (flags & MDB_NODUPDATA) ?
6411 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6413 /* converted, write the original data first */
6415 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6419 /* Adjust other cursors pointing to mp */
6421 unsigned i = mc->mc_top;
6422 MDB_page *mp = mc->mc_pg[i];
6424 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6425 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6426 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6427 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6428 mdb_xcursor_init1(m2, leaf);
6432 /* we've done our job */
6435 ecount = mc->mc_xcursor->mx_db.md_entries;
6436 if (flags & MDB_APPENDDUP)
6437 xflags |= MDB_APPEND;
6438 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6439 if (flags & F_SUBDATA) {
6440 void *db = NODEDATA(leaf);
6441 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6443 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6445 /* Increment count unless we just replaced an existing item. */
6447 mc->mc_db->md_entries++;
6449 /* Invalidate txn if we created an empty sub-DB */
6452 /* If we succeeded and the key didn't exist before,
6453 * make sure the cursor is marked valid.
6455 mc->mc_flags |= C_INITIALIZED;
6457 if (flags & MDB_MULTIPLE) {
6460 /* let caller know how many succeeded, if any */
6461 data[1].mv_size = mcount;
6462 if (mcount < dcount) {
6463 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6464 insert_key = insert_data = 0;
6471 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6474 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6479 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6485 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6486 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6488 if (!(mc->mc_flags & C_INITIALIZED))
6491 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6492 return MDB_NOTFOUND;
6494 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6497 rc = mdb_cursor_touch(mc);
6501 mp = mc->mc_pg[mc->mc_top];
6504 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6506 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6507 if (flags & MDB_NODUPDATA) {
6508 /* mdb_cursor_del0() will subtract the final entry */
6509 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6511 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6512 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6514 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6517 /* If sub-DB still has entries, we're done */
6518 if (mc->mc_xcursor->mx_db.md_entries) {
6519 if (leaf->mn_flags & F_SUBDATA) {
6520 /* update subDB info */
6521 void *db = NODEDATA(leaf);
6522 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6525 /* shrink fake page */
6526 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6527 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6528 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6529 /* fix other sub-DB cursors pointed at this fake page */
6530 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6531 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6532 if (m2->mc_pg[mc->mc_top] == mp &&
6533 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6534 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6537 mc->mc_db->md_entries--;
6538 mc->mc_flags |= C_DEL;
6541 /* otherwise fall thru and delete the sub-DB */
6544 if (leaf->mn_flags & F_SUBDATA) {
6545 /* add all the child DB's pages to the free list */
6546 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6552 /* add overflow pages to free list */
6553 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6557 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6558 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6559 (rc = mdb_ovpage_free(mc, omp)))
6564 return mdb_cursor_del0(mc);
6567 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6571 /** Allocate and initialize new pages for a database.
6572 * @param[in] mc a cursor on the database being added to.
6573 * @param[in] flags flags defining what type of page is being allocated.
6574 * @param[in] num the number of pages to allocate. This is usually 1,
6575 * unless allocating overflow pages for a large record.
6576 * @param[out] mp Address of a page, or NULL on failure.
6577 * @return 0 on success, non-zero on failure.
6580 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6585 if ((rc = mdb_page_alloc(mc, num, &np)))
6587 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6588 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6589 np->mp_flags = flags | P_DIRTY;
6590 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6591 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6594 mc->mc_db->md_branch_pages++;
6595 else if (IS_LEAF(np))
6596 mc->mc_db->md_leaf_pages++;
6597 else if (IS_OVERFLOW(np)) {
6598 mc->mc_db->md_overflow_pages += num;
6606 /** Calculate the size of a leaf node.
6607 * The size depends on the environment's page size; if a data item
6608 * is too large it will be put onto an overflow page and the node
6609 * size will only include the key and not the data. Sizes are always
6610 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6611 * of the #MDB_node headers.
6612 * @param[in] env The environment handle.
6613 * @param[in] key The key for the node.
6614 * @param[in] data The data for the node.
6615 * @return The number of bytes needed to store the node.
6618 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6622 sz = LEAFSIZE(key, data);
6623 if (sz > env->me_nodemax) {
6624 /* put on overflow page */
6625 sz -= data->mv_size - sizeof(pgno_t);
6628 return EVEN(sz + sizeof(indx_t));
6631 /** Calculate the size of a branch node.
6632 * The size should depend on the environment's page size but since
6633 * we currently don't support spilling large keys onto overflow
6634 * pages, it's simply the size of the #MDB_node header plus the
6635 * size of the key. Sizes are always rounded up to an even number
6636 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6637 * @param[in] env The environment handle.
6638 * @param[in] key The key for the node.
6639 * @return The number of bytes needed to store the node.
6642 mdb_branch_size(MDB_env *env, MDB_val *key)
6647 if (sz > env->me_nodemax) {
6648 /* put on overflow page */
6649 /* not implemented */
6650 /* sz -= key->size - sizeof(pgno_t); */
6653 return sz + sizeof(indx_t);
6656 /** Add a node to the page pointed to by the cursor.
6657 * @param[in] mc The cursor for this operation.
6658 * @param[in] indx The index on the page where the new node should be added.
6659 * @param[in] key The key for the new node.
6660 * @param[in] data The data for the new node, if any.
6661 * @param[in] pgno The page number, if adding a branch node.
6662 * @param[in] flags Flags for the node.
6663 * @return 0 on success, non-zero on failure. Possible errors are:
6665 * <li>ENOMEM - failed to allocate overflow pages for the node.
6666 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6667 * should never happen since all callers already calculate the
6668 * page's free space before calling this function.
6672 mdb_node_add(MDB_cursor *mc, indx_t indx,
6673 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6676 size_t node_size = NODESIZE;
6680 MDB_page *mp = mc->mc_pg[mc->mc_top];
6681 MDB_page *ofp = NULL; /* overflow page */
6684 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6686 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6687 IS_LEAF(mp) ? "leaf" : "branch",
6688 IS_SUBP(mp) ? "sub-" : "",
6689 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6690 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6693 /* Move higher keys up one slot. */
6694 int ksize = mc->mc_db->md_pad, dif;
6695 char *ptr = LEAF2KEY(mp, indx, ksize);
6696 dif = NUMKEYS(mp) - indx;
6698 memmove(ptr+ksize, ptr, dif*ksize);
6699 /* insert new key */
6700 memcpy(ptr, key->mv_data, ksize);
6702 /* Just using these for counting */
6703 mp->mp_lower += sizeof(indx_t);
6704 mp->mp_upper -= ksize - sizeof(indx_t);
6708 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6710 node_size += key->mv_size;
6712 mdb_cassert(mc, data);
6713 if (F_ISSET(flags, F_BIGDATA)) {
6714 /* Data already on overflow page. */
6715 node_size += sizeof(pgno_t);
6716 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6717 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6719 /* Put data on overflow page. */
6720 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6721 data->mv_size, node_size+data->mv_size));
6722 node_size = EVEN(node_size + sizeof(pgno_t));
6723 if ((ssize_t)node_size > room)
6725 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6727 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6731 node_size += data->mv_size;
6734 node_size = EVEN(node_size);
6735 if ((ssize_t)node_size > room)
6739 /* Move higher pointers up one slot. */
6740 for (i = NUMKEYS(mp); i > indx; i--)
6741 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6743 /* Adjust free space offsets. */
6744 ofs = mp->mp_upper - node_size;
6745 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6746 mp->mp_ptrs[indx] = ofs;
6748 mp->mp_lower += sizeof(indx_t);
6750 /* Write the node data. */
6751 node = NODEPTR(mp, indx);
6752 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6753 node->mn_flags = flags;
6755 SETDSZ(node,data->mv_size);
6760 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6763 mdb_cassert(mc, key);
6765 if (F_ISSET(flags, F_BIGDATA))
6766 memcpy(node->mn_data + key->mv_size, data->mv_data,
6768 else if (F_ISSET(flags, MDB_RESERVE))
6769 data->mv_data = node->mn_data + key->mv_size;
6771 memcpy(node->mn_data + key->mv_size, data->mv_data,
6774 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6776 if (F_ISSET(flags, MDB_RESERVE))
6777 data->mv_data = METADATA(ofp);
6779 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6786 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6787 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6788 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6789 DPRINTF(("node size = %"Z"u", node_size));
6790 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6791 return MDB_PAGE_FULL;
6794 /** Delete the specified node from a page.
6795 * @param[in] mc Cursor pointing to the node to delete.
6796 * @param[in] ksize The size of a node. Only used if the page is
6797 * part of a #MDB_DUPFIXED database.
6800 mdb_node_del(MDB_cursor *mc, int ksize)
6802 MDB_page *mp = mc->mc_pg[mc->mc_top];
6803 indx_t indx = mc->mc_ki[mc->mc_top];
6805 indx_t i, j, numkeys, ptr;
6809 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6810 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6811 numkeys = NUMKEYS(mp);
6812 mdb_cassert(mc, indx < numkeys);
6815 int x = numkeys - 1 - indx;
6816 base = LEAF2KEY(mp, indx, ksize);
6818 memmove(base, base + ksize, x * ksize);
6819 mp->mp_lower -= sizeof(indx_t);
6820 mp->mp_upper += ksize - sizeof(indx_t);
6824 node = NODEPTR(mp, indx);
6825 sz = NODESIZE + node->mn_ksize;
6827 if (F_ISSET(node->mn_flags, F_BIGDATA))
6828 sz += sizeof(pgno_t);
6830 sz += NODEDSZ(node);
6834 ptr = mp->mp_ptrs[indx];
6835 for (i = j = 0; i < numkeys; i++) {
6837 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6838 if (mp->mp_ptrs[i] < ptr)
6839 mp->mp_ptrs[j] += sz;
6844 base = (char *)mp + mp->mp_upper + PAGEBASE;
6845 memmove(base + sz, base, ptr - mp->mp_upper);
6847 mp->mp_lower -= sizeof(indx_t);
6851 /** Compact the main page after deleting a node on a subpage.
6852 * @param[in] mp The main page to operate on.
6853 * @param[in] indx The index of the subpage on the main page.
6856 mdb_node_shrink(MDB_page *mp, indx_t indx)
6862 indx_t i, numkeys, ptr;
6864 node = NODEPTR(mp, indx);
6865 sp = (MDB_page *)NODEDATA(node);
6866 delta = SIZELEFT(sp);
6867 xp = (MDB_page *)((char *)sp + delta);
6869 /* shift subpage upward */
6871 nsize = NUMKEYS(sp) * sp->mp_pad;
6873 return; /* do not make the node uneven-sized */
6874 memmove(METADATA(xp), METADATA(sp), nsize);
6877 numkeys = NUMKEYS(sp);
6878 for (i=numkeys-1; i>=0; i--)
6879 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6881 xp->mp_upper = sp->mp_lower;
6882 xp->mp_lower = sp->mp_lower;
6883 xp->mp_flags = sp->mp_flags;
6884 xp->mp_pad = sp->mp_pad;
6885 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6887 nsize = NODEDSZ(node) - delta;
6888 SETDSZ(node, nsize);
6890 /* shift lower nodes upward */
6891 ptr = mp->mp_ptrs[indx];
6892 numkeys = NUMKEYS(mp);
6893 for (i = 0; i < numkeys; i++) {
6894 if (mp->mp_ptrs[i] <= ptr)
6895 mp->mp_ptrs[i] += delta;
6898 base = (char *)mp + mp->mp_upper + PAGEBASE;
6899 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6900 mp->mp_upper += delta;
6903 /** Initial setup of a sorted-dups cursor.
6904 * Sorted duplicates are implemented as a sub-database for the given key.
6905 * The duplicate data items are actually keys of the sub-database.
6906 * Operations on the duplicate data items are performed using a sub-cursor
6907 * initialized when the sub-database is first accessed. This function does
6908 * the preliminary setup of the sub-cursor, filling in the fields that
6909 * depend only on the parent DB.
6910 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6913 mdb_xcursor_init0(MDB_cursor *mc)
6915 MDB_xcursor *mx = mc->mc_xcursor;
6917 mx->mx_cursor.mc_xcursor = NULL;
6918 mx->mx_cursor.mc_txn = mc->mc_txn;
6919 mx->mx_cursor.mc_db = &mx->mx_db;
6920 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6921 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6922 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6923 mx->mx_cursor.mc_snum = 0;
6924 mx->mx_cursor.mc_top = 0;
6925 mx->mx_cursor.mc_flags = C_SUB;
6926 mx->mx_dbx.md_name.mv_size = 0;
6927 mx->mx_dbx.md_name.mv_data = NULL;
6928 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6929 mx->mx_dbx.md_dcmp = NULL;
6930 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6933 /** Final setup of a sorted-dups cursor.
6934 * Sets up the fields that depend on the data from the main cursor.
6935 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6936 * @param[in] node The data containing the #MDB_db record for the
6937 * sorted-dup database.
6940 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6942 MDB_xcursor *mx = mc->mc_xcursor;
6944 if (node->mn_flags & F_SUBDATA) {
6945 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6946 mx->mx_cursor.mc_pg[0] = 0;
6947 mx->mx_cursor.mc_snum = 0;
6948 mx->mx_cursor.mc_top = 0;
6949 mx->mx_cursor.mc_flags = C_SUB;
6951 MDB_page *fp = NODEDATA(node);
6952 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6953 mx->mx_db.md_flags = 0;
6954 mx->mx_db.md_depth = 1;
6955 mx->mx_db.md_branch_pages = 0;
6956 mx->mx_db.md_leaf_pages = 1;
6957 mx->mx_db.md_overflow_pages = 0;
6958 mx->mx_db.md_entries = NUMKEYS(fp);
6959 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6960 mx->mx_cursor.mc_snum = 1;
6961 mx->mx_cursor.mc_top = 0;
6962 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6963 mx->mx_cursor.mc_pg[0] = fp;
6964 mx->mx_cursor.mc_ki[0] = 0;
6965 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6966 mx->mx_db.md_flags = MDB_DUPFIXED;
6967 mx->mx_db.md_pad = fp->mp_pad;
6968 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6969 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6972 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6973 mx->mx_db.md_root));
6974 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6975 #if UINT_MAX < SIZE_MAX
6976 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6977 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6981 /** Initialize a cursor for a given transaction and database. */
6983 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6986 mc->mc_backup = NULL;
6989 mc->mc_db = &txn->mt_dbs[dbi];
6990 mc->mc_dbx = &txn->mt_dbxs[dbi];
6991 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6996 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6997 mdb_tassert(txn, mx != NULL);
6998 mc->mc_xcursor = mx;
6999 mdb_xcursor_init0(mc);
7001 mc->mc_xcursor = NULL;
7003 if (*mc->mc_dbflag & DB_STALE) {
7004 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7009 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7012 size_t size = sizeof(MDB_cursor);
7014 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7017 if (txn->mt_flags & MDB_TXN_ERROR)
7020 /* Allow read access to the freelist */
7021 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7024 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7025 size += sizeof(MDB_xcursor);
7027 if ((mc = malloc(size)) != NULL) {
7028 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7029 if (txn->mt_cursors) {
7030 mc->mc_next = txn->mt_cursors[dbi];
7031 txn->mt_cursors[dbi] = mc;
7032 mc->mc_flags |= C_UNTRACK;
7044 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7046 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7049 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7052 if (txn->mt_flags & MDB_TXN_ERROR)
7055 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7059 /* Return the count of duplicate data items for the current key */
7061 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7065 if (mc == NULL || countp == NULL)
7068 if (mc->mc_xcursor == NULL)
7069 return MDB_INCOMPATIBLE;
7071 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7074 if (!(mc->mc_flags & C_INITIALIZED))
7077 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7078 return MDB_NOTFOUND;
7080 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7081 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7084 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7087 *countp = mc->mc_xcursor->mx_db.md_entries;
7093 mdb_cursor_close(MDB_cursor *mc)
7095 if (mc && !mc->mc_backup) {
7096 /* remove from txn, if tracked */
7097 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7098 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7099 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7101 *prev = mc->mc_next;
7108 mdb_cursor_txn(MDB_cursor *mc)
7110 if (!mc) return NULL;
7115 mdb_cursor_dbi(MDB_cursor *mc)
7120 /** Replace the key for a branch node with a new key.
7121 * @param[in] mc Cursor pointing to the node to operate on.
7122 * @param[in] key The new key to use.
7123 * @return 0 on success, non-zero on failure.
7126 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7132 int delta, ksize, oksize;
7133 indx_t ptr, i, numkeys, indx;
7136 indx = mc->mc_ki[mc->mc_top];
7137 mp = mc->mc_pg[mc->mc_top];
7138 node = NODEPTR(mp, indx);
7139 ptr = mp->mp_ptrs[indx];
7143 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7144 k2.mv_data = NODEKEY(node);
7145 k2.mv_size = node->mn_ksize;
7146 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7148 mdb_dkey(&k2, kbuf2),
7154 /* Sizes must be 2-byte aligned. */
7155 ksize = EVEN(key->mv_size);
7156 oksize = EVEN(node->mn_ksize);
7157 delta = ksize - oksize;
7159 /* Shift node contents if EVEN(key length) changed. */
7161 if (delta > 0 && SIZELEFT(mp) < delta) {
7163 /* not enough space left, do a delete and split */
7164 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7165 pgno = NODEPGNO(node);
7166 mdb_node_del(mc, 0);
7167 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7170 numkeys = NUMKEYS(mp);
7171 for (i = 0; i < numkeys; i++) {
7172 if (mp->mp_ptrs[i] <= ptr)
7173 mp->mp_ptrs[i] -= delta;
7176 base = (char *)mp + mp->mp_upper + PAGEBASE;
7177 len = ptr - mp->mp_upper + NODESIZE;
7178 memmove(base - delta, base, len);
7179 mp->mp_upper -= delta;
7181 node = NODEPTR(mp, indx);
7184 /* But even if no shift was needed, update ksize */
7185 if (node->mn_ksize != key->mv_size)
7186 node->mn_ksize = key->mv_size;
7189 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7195 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7197 /** Move a node from csrc to cdst.
7200 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7207 unsigned short flags;
7211 /* Mark src and dst as dirty. */
7212 if ((rc = mdb_page_touch(csrc)) ||
7213 (rc = mdb_page_touch(cdst)))
7216 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7217 key.mv_size = csrc->mc_db->md_pad;
7218 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7220 data.mv_data = NULL;
7224 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7225 mdb_cassert(csrc, !((size_t)srcnode & 1));
7226 srcpg = NODEPGNO(srcnode);
7227 flags = srcnode->mn_flags;
7228 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7229 unsigned int snum = csrc->mc_snum;
7231 /* must find the lowest key below src */
7232 rc = mdb_page_search_lowest(csrc);
7235 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7236 key.mv_size = csrc->mc_db->md_pad;
7237 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7239 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7240 key.mv_size = NODEKSZ(s2);
7241 key.mv_data = NODEKEY(s2);
7243 csrc->mc_snum = snum--;
7244 csrc->mc_top = snum;
7246 key.mv_size = NODEKSZ(srcnode);
7247 key.mv_data = NODEKEY(srcnode);
7249 data.mv_size = NODEDSZ(srcnode);
7250 data.mv_data = NODEDATA(srcnode);
7252 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7253 unsigned int snum = cdst->mc_snum;
7256 /* must find the lowest key below dst */
7257 mdb_cursor_copy(cdst, &mn);
7258 rc = mdb_page_search_lowest(&mn);
7261 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7262 bkey.mv_size = mn.mc_db->md_pad;
7263 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7265 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7266 bkey.mv_size = NODEKSZ(s2);
7267 bkey.mv_data = NODEKEY(s2);
7269 mn.mc_snum = snum--;
7272 rc = mdb_update_key(&mn, &bkey);
7277 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7278 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7279 csrc->mc_ki[csrc->mc_top],
7281 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7282 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7284 /* Add the node to the destination page.
7286 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7287 if (rc != MDB_SUCCESS)
7290 /* Delete the node from the source page.
7292 mdb_node_del(csrc, key.mv_size);
7295 /* Adjust other cursors pointing to mp */
7296 MDB_cursor *m2, *m3;
7297 MDB_dbi dbi = csrc->mc_dbi;
7298 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7300 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7301 if (csrc->mc_flags & C_SUB)
7302 m3 = &m2->mc_xcursor->mx_cursor;
7305 if (m3 == csrc) continue;
7306 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7307 csrc->mc_ki[csrc->mc_top]) {
7308 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7309 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7314 /* Update the parent separators.
7316 if (csrc->mc_ki[csrc->mc_top] == 0) {
7317 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7318 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7319 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7321 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7322 key.mv_size = NODEKSZ(srcnode);
7323 key.mv_data = NODEKEY(srcnode);
7325 DPRINTF(("update separator for source page %"Z"u to [%s]",
7326 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7327 mdb_cursor_copy(csrc, &mn);
7330 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7333 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7335 indx_t ix = csrc->mc_ki[csrc->mc_top];
7336 nullkey.mv_size = 0;
7337 csrc->mc_ki[csrc->mc_top] = 0;
7338 rc = mdb_update_key(csrc, &nullkey);
7339 csrc->mc_ki[csrc->mc_top] = ix;
7340 mdb_cassert(csrc, rc == MDB_SUCCESS);
7344 if (cdst->mc_ki[cdst->mc_top] == 0) {
7345 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7346 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7347 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7349 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7350 key.mv_size = NODEKSZ(srcnode);
7351 key.mv_data = NODEKEY(srcnode);
7353 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7354 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7355 mdb_cursor_copy(cdst, &mn);
7358 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7361 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7363 indx_t ix = cdst->mc_ki[cdst->mc_top];
7364 nullkey.mv_size = 0;
7365 cdst->mc_ki[cdst->mc_top] = 0;
7366 rc = mdb_update_key(cdst, &nullkey);
7367 cdst->mc_ki[cdst->mc_top] = ix;
7368 mdb_cassert(csrc, rc == MDB_SUCCESS);
7375 /** Merge one page into another.
7376 * The nodes from the page pointed to by \b csrc will
7377 * be copied to the page pointed to by \b cdst and then
7378 * the \b csrc page will be freed.
7379 * @param[in] csrc Cursor pointing to the source page.
7380 * @param[in] cdst Cursor pointing to the destination page.
7381 * @return 0 on success, non-zero on failure.
7384 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7386 MDB_page *psrc, *pdst;
7393 psrc = csrc->mc_pg[csrc->mc_top];
7394 pdst = cdst->mc_pg[cdst->mc_top];
7396 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7398 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7399 mdb_cassert(csrc, cdst->mc_snum > 1);
7401 /* Mark dst as dirty. */
7402 if ((rc = mdb_page_touch(cdst)))
7405 /* Move all nodes from src to dst.
7407 j = nkeys = NUMKEYS(pdst);
7408 if (IS_LEAF2(psrc)) {
7409 key.mv_size = csrc->mc_db->md_pad;
7410 key.mv_data = METADATA(psrc);
7411 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7412 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7413 if (rc != MDB_SUCCESS)
7415 key.mv_data = (char *)key.mv_data + key.mv_size;
7418 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7419 srcnode = NODEPTR(psrc, i);
7420 if (i == 0 && IS_BRANCH(psrc)) {
7423 mdb_cursor_copy(csrc, &mn);
7424 /* must find the lowest key below src */
7425 rc = mdb_page_search_lowest(&mn);
7428 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7429 key.mv_size = mn.mc_db->md_pad;
7430 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7432 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7433 key.mv_size = NODEKSZ(s2);
7434 key.mv_data = NODEKEY(s2);
7437 key.mv_size = srcnode->mn_ksize;
7438 key.mv_data = NODEKEY(srcnode);
7441 data.mv_size = NODEDSZ(srcnode);
7442 data.mv_data = NODEDATA(srcnode);
7443 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7444 if (rc != MDB_SUCCESS)
7449 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7450 pdst->mp_pgno, NUMKEYS(pdst),
7451 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7453 /* Unlink the src page from parent and add to free list.
7456 mdb_node_del(csrc, 0);
7457 if (csrc->mc_ki[csrc->mc_top] == 0) {
7459 rc = mdb_update_key(csrc, &key);
7467 psrc = csrc->mc_pg[csrc->mc_top];
7468 /* If not operating on FreeDB, allow this page to be reused
7469 * in this txn. Otherwise just add to free list.
7471 rc = mdb_page_loose(csrc, psrc);
7475 csrc->mc_db->md_leaf_pages--;
7477 csrc->mc_db->md_branch_pages--;
7479 /* Adjust other cursors pointing to mp */
7480 MDB_cursor *m2, *m3;
7481 MDB_dbi dbi = csrc->mc_dbi;
7483 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7484 if (csrc->mc_flags & C_SUB)
7485 m3 = &m2->mc_xcursor->mx_cursor;
7488 if (m3 == csrc) continue;
7489 if (m3->mc_snum < csrc->mc_snum) continue;
7490 if (m3->mc_pg[csrc->mc_top] == psrc) {
7491 m3->mc_pg[csrc->mc_top] = pdst;
7492 m3->mc_ki[csrc->mc_top] += nkeys;
7497 unsigned int snum = cdst->mc_snum;
7498 uint16_t depth = cdst->mc_db->md_depth;
7499 mdb_cursor_pop(cdst);
7500 rc = mdb_rebalance(cdst);
7501 /* Did the tree shrink? */
7502 if (depth > cdst->mc_db->md_depth)
7504 cdst->mc_snum = snum;
7505 cdst->mc_top = snum-1;
7510 /** Copy the contents of a cursor.
7511 * @param[in] csrc The cursor to copy from.
7512 * @param[out] cdst The cursor to copy to.
7515 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7519 cdst->mc_txn = csrc->mc_txn;
7520 cdst->mc_dbi = csrc->mc_dbi;
7521 cdst->mc_db = csrc->mc_db;
7522 cdst->mc_dbx = csrc->mc_dbx;
7523 cdst->mc_snum = csrc->mc_snum;
7524 cdst->mc_top = csrc->mc_top;
7525 cdst->mc_flags = csrc->mc_flags;
7527 for (i=0; i<csrc->mc_snum; i++) {
7528 cdst->mc_pg[i] = csrc->mc_pg[i];
7529 cdst->mc_ki[i] = csrc->mc_ki[i];
7533 /** Rebalance the tree after a delete operation.
7534 * @param[in] mc Cursor pointing to the page where rebalancing
7536 * @return 0 on success, non-zero on failure.
7539 mdb_rebalance(MDB_cursor *mc)
7543 unsigned int ptop, minkeys;
7547 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7548 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7549 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7550 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7551 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7553 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7554 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7555 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7556 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7560 if (mc->mc_snum < 2) {
7561 MDB_page *mp = mc->mc_pg[0];
7563 DPUTS("Can't rebalance a subpage, ignoring");
7566 if (NUMKEYS(mp) == 0) {
7567 DPUTS("tree is completely empty");
7568 mc->mc_db->md_root = P_INVALID;
7569 mc->mc_db->md_depth = 0;
7570 mc->mc_db->md_leaf_pages = 0;
7571 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7574 /* Adjust cursors pointing to mp */
7577 mc->mc_flags &= ~C_INITIALIZED;
7579 MDB_cursor *m2, *m3;
7580 MDB_dbi dbi = mc->mc_dbi;
7582 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7583 if (mc->mc_flags & C_SUB)
7584 m3 = &m2->mc_xcursor->mx_cursor;
7587 if (m3->mc_snum < mc->mc_snum) continue;
7588 if (m3->mc_pg[0] == mp) {
7591 m3->mc_flags &= ~C_INITIALIZED;
7595 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7597 DPUTS("collapsing root page!");
7598 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7601 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7602 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7605 mc->mc_db->md_depth--;
7606 mc->mc_db->md_branch_pages--;
7607 mc->mc_ki[0] = mc->mc_ki[1];
7608 for (i = 1; i<mc->mc_db->md_depth; i++) {
7609 mc->mc_pg[i] = mc->mc_pg[i+1];
7610 mc->mc_ki[i] = mc->mc_ki[i+1];
7613 /* Adjust other cursors pointing to mp */
7614 MDB_cursor *m2, *m3;
7615 MDB_dbi dbi = mc->mc_dbi;
7617 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7618 if (mc->mc_flags & C_SUB)
7619 m3 = &m2->mc_xcursor->mx_cursor;
7622 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7623 if (m3->mc_pg[0] == mp) {
7626 for (i=0; i<m3->mc_snum; i++) {
7627 m3->mc_pg[i] = m3->mc_pg[i+1];
7628 m3->mc_ki[i] = m3->mc_ki[i+1];
7634 DPUTS("root page doesn't need rebalancing");
7638 /* The parent (branch page) must have at least 2 pointers,
7639 * otherwise the tree is invalid.
7641 ptop = mc->mc_top-1;
7642 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7644 /* Leaf page fill factor is below the threshold.
7645 * Try to move keys from left or right neighbor, or
7646 * merge with a neighbor page.
7651 mdb_cursor_copy(mc, &mn);
7652 mn.mc_xcursor = NULL;
7654 oldki = mc->mc_ki[mc->mc_top];
7655 if (mc->mc_ki[ptop] == 0) {
7656 /* We're the leftmost leaf in our parent.
7658 DPUTS("reading right neighbor");
7660 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7661 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7664 mn.mc_ki[mn.mc_top] = 0;
7665 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7667 /* There is at least one neighbor to the left.
7669 DPUTS("reading left neighbor");
7671 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7672 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7675 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7676 mc->mc_ki[mc->mc_top] = 0;
7679 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7680 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7681 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7683 /* If the neighbor page is above threshold and has enough keys,
7684 * move one key from it. Otherwise we should try to merge them.
7685 * (A branch page must never have less than 2 keys.)
7687 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7688 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7689 rc = mdb_node_move(&mn, mc);
7690 if (mc->mc_ki[ptop]) {
7694 if (mc->mc_ki[ptop] == 0) {
7695 rc = mdb_page_merge(&mn, mc);
7697 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7698 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7699 rc = mdb_page_merge(mc, &mn);
7700 mdb_cursor_copy(&mn, mc);
7702 mc->mc_flags &= ~C_EOF;
7704 mc->mc_ki[mc->mc_top] = oldki;
7708 /** Complete a delete operation started by #mdb_cursor_del(). */
7710 mdb_cursor_del0(MDB_cursor *mc)
7717 ki = mc->mc_ki[mc->mc_top];
7718 mdb_node_del(mc, mc->mc_db->md_pad);
7719 mc->mc_db->md_entries--;
7720 rc = mdb_rebalance(mc);
7722 if (rc == MDB_SUCCESS) {
7723 MDB_cursor *m2, *m3;
7724 MDB_dbi dbi = mc->mc_dbi;
7726 mp = mc->mc_pg[mc->mc_top];
7727 nkeys = NUMKEYS(mp);
7729 /* if mc points past last node in page, find next sibling */
7730 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7731 rc = mdb_cursor_sibling(mc, 1);
7732 if (rc == MDB_NOTFOUND) {
7733 mc->mc_flags |= C_EOF;
7738 /* Adjust other cursors pointing to mp */
7739 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7740 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7741 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7743 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7745 if (m3->mc_pg[mc->mc_top] == mp) {
7746 if (m3->mc_ki[mc->mc_top] >= ki) {
7747 m3->mc_flags |= C_DEL;
7748 if (m3->mc_ki[mc->mc_top] > ki)
7749 m3->mc_ki[mc->mc_top]--;
7750 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7751 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7753 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7754 rc = mdb_cursor_sibling(m3, 1);
7755 if (rc == MDB_NOTFOUND) {
7756 m3->mc_flags |= C_EOF;
7762 mc->mc_flags |= C_DEL;
7766 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7771 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7772 MDB_val *key, MDB_val *data)
7774 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7777 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7778 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7780 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7781 /* must ignore any data */
7785 return mdb_del0(txn, dbi, key, data, 0);
7789 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7790 MDB_val *key, MDB_val *data, unsigned flags)
7795 MDB_val rdata, *xdata;
7799 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7801 mdb_cursor_init(&mc, txn, dbi, &mx);
7810 flags |= MDB_NODUPDATA;
7812 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7814 /* let mdb_page_split know about this cursor if needed:
7815 * delete will trigger a rebalance; if it needs to move
7816 * a node from one page to another, it will have to
7817 * update the parent's separator key(s). If the new sepkey
7818 * is larger than the current one, the parent page may
7819 * run out of space, triggering a split. We need this
7820 * cursor to be consistent until the end of the rebalance.
7822 mc.mc_flags |= C_UNTRACK;
7823 mc.mc_next = txn->mt_cursors[dbi];
7824 txn->mt_cursors[dbi] = &mc;
7825 rc = mdb_cursor_del(&mc, flags);
7826 txn->mt_cursors[dbi] = mc.mc_next;
7831 /** Split a page and insert a new node.
7832 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7833 * The cursor will be updated to point to the actual page and index where
7834 * the node got inserted after the split.
7835 * @param[in] newkey The key for the newly inserted node.
7836 * @param[in] newdata The data for the newly inserted node.
7837 * @param[in] newpgno The page number, if the new node is a branch node.
7838 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7839 * @return 0 on success, non-zero on failure.
7842 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7843 unsigned int nflags)
7846 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7849 int i, j, split_indx, nkeys, pmax;
7850 MDB_env *env = mc->mc_txn->mt_env;
7852 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7853 MDB_page *copy = NULL;
7854 MDB_page *mp, *rp, *pp;
7859 mp = mc->mc_pg[mc->mc_top];
7860 newindx = mc->mc_ki[mc->mc_top];
7861 nkeys = NUMKEYS(mp);
7863 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7864 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7865 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7867 /* Create a right sibling. */
7868 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7870 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7872 if (mc->mc_snum < 2) {
7873 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7875 /* shift current top to make room for new parent */
7876 mc->mc_pg[1] = mc->mc_pg[0];
7877 mc->mc_ki[1] = mc->mc_ki[0];
7880 mc->mc_db->md_root = pp->mp_pgno;
7881 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7882 mc->mc_db->md_depth++;
7885 /* Add left (implicit) pointer. */
7886 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7887 /* undo the pre-push */
7888 mc->mc_pg[0] = mc->mc_pg[1];
7889 mc->mc_ki[0] = mc->mc_ki[1];
7890 mc->mc_db->md_root = mp->mp_pgno;
7891 mc->mc_db->md_depth--;
7898 ptop = mc->mc_top-1;
7899 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7902 mc->mc_flags |= C_SPLITTING;
7903 mdb_cursor_copy(mc, &mn);
7904 mn.mc_pg[mn.mc_top] = rp;
7905 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7907 if (nflags & MDB_APPEND) {
7908 mn.mc_ki[mn.mc_top] = 0;
7910 split_indx = newindx;
7914 split_indx = (nkeys+1) / 2;
7919 unsigned int lsize, rsize, ksize;
7920 /* Move half of the keys to the right sibling */
7921 x = mc->mc_ki[mc->mc_top] - split_indx;
7922 ksize = mc->mc_db->md_pad;
7923 split = LEAF2KEY(mp, split_indx, ksize);
7924 rsize = (nkeys - split_indx) * ksize;
7925 lsize = (nkeys - split_indx) * sizeof(indx_t);
7926 mp->mp_lower -= lsize;
7927 rp->mp_lower += lsize;
7928 mp->mp_upper += rsize - lsize;
7929 rp->mp_upper -= rsize - lsize;
7930 sepkey.mv_size = ksize;
7931 if (newindx == split_indx) {
7932 sepkey.mv_data = newkey->mv_data;
7934 sepkey.mv_data = split;
7937 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7938 memcpy(rp->mp_ptrs, split, rsize);
7939 sepkey.mv_data = rp->mp_ptrs;
7940 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7941 memcpy(ins, newkey->mv_data, ksize);
7942 mp->mp_lower += sizeof(indx_t);
7943 mp->mp_upper -= ksize - sizeof(indx_t);
7946 memcpy(rp->mp_ptrs, split, x * ksize);
7947 ins = LEAF2KEY(rp, x, ksize);
7948 memcpy(ins, newkey->mv_data, ksize);
7949 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7950 rp->mp_lower += sizeof(indx_t);
7951 rp->mp_upper -= ksize - sizeof(indx_t);
7952 mc->mc_ki[mc->mc_top] = x;
7953 mc->mc_pg[mc->mc_top] = rp;
7956 int psize, nsize, k;
7957 /* Maximum free space in an empty page */
7958 pmax = env->me_psize - PAGEHDRSZ;
7960 nsize = mdb_leaf_size(env, newkey, newdata);
7962 nsize = mdb_branch_size(env, newkey);
7963 nsize = EVEN(nsize);
7965 /* grab a page to hold a temporary copy */
7966 copy = mdb_page_malloc(mc->mc_txn, 1);
7971 copy->mp_pgno = mp->mp_pgno;
7972 copy->mp_flags = mp->mp_flags;
7973 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7974 copy->mp_upper = env->me_psize - PAGEBASE;
7976 /* prepare to insert */
7977 for (i=0, j=0; i<nkeys; i++) {
7979 copy->mp_ptrs[j++] = 0;
7981 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7984 /* When items are relatively large the split point needs
7985 * to be checked, because being off-by-one will make the
7986 * difference between success or failure in mdb_node_add.
7988 * It's also relevant if a page happens to be laid out
7989 * such that one half of its nodes are all "small" and
7990 * the other half of its nodes are "large." If the new
7991 * item is also "large" and falls on the half with
7992 * "large" nodes, it also may not fit.
7994 * As a final tweak, if the new item goes on the last
7995 * spot on the page (and thus, onto the new page), bias
7996 * the split so the new page is emptier than the old page.
7997 * This yields better packing during sequential inserts.
7999 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8000 /* Find split point */
8002 if (newindx <= split_indx || newindx >= nkeys) {
8004 k = newindx >= nkeys ? nkeys : split_indx+2;
8009 for (; i!=k; i+=j) {
8014 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8015 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8017 if (F_ISSET(node->mn_flags, F_BIGDATA))
8018 psize += sizeof(pgno_t);
8020 psize += NODEDSZ(node);
8022 psize = EVEN(psize);
8024 if (psize > pmax || i == k-j) {
8025 split_indx = i + (j<0);
8030 if (split_indx == newindx) {
8031 sepkey.mv_size = newkey->mv_size;
8032 sepkey.mv_data = newkey->mv_data;
8034 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8035 sepkey.mv_size = node->mn_ksize;
8036 sepkey.mv_data = NODEKEY(node);
8041 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8043 /* Copy separator key to the parent.
8045 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8049 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8054 if (mn.mc_snum == mc->mc_snum) {
8055 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8056 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8057 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8058 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8063 /* Right page might now have changed parent.
8064 * Check if left page also changed parent.
8066 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8067 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8068 for (i=0; i<ptop; i++) {
8069 mc->mc_pg[i] = mn.mc_pg[i];
8070 mc->mc_ki[i] = mn.mc_ki[i];
8072 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8073 if (mn.mc_ki[ptop]) {
8074 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8076 /* find right page's left sibling */
8077 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8078 mdb_cursor_sibling(mc, 0);
8083 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8086 mc->mc_flags ^= C_SPLITTING;
8087 if (rc != MDB_SUCCESS) {
8090 if (nflags & MDB_APPEND) {
8091 mc->mc_pg[mc->mc_top] = rp;
8092 mc->mc_ki[mc->mc_top] = 0;
8093 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8096 for (i=0; i<mc->mc_top; i++)
8097 mc->mc_ki[i] = mn.mc_ki[i];
8098 } else if (!IS_LEAF2(mp)) {
8100 mc->mc_pg[mc->mc_top] = rp;
8105 rkey.mv_data = newkey->mv_data;
8106 rkey.mv_size = newkey->mv_size;
8112 /* Update index for the new key. */
8113 mc->mc_ki[mc->mc_top] = j;
8115 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8116 rkey.mv_data = NODEKEY(node);
8117 rkey.mv_size = node->mn_ksize;
8119 xdata.mv_data = NODEDATA(node);
8120 xdata.mv_size = NODEDSZ(node);
8123 pgno = NODEPGNO(node);
8124 flags = node->mn_flags;
8127 if (!IS_LEAF(mp) && j == 0) {
8128 /* First branch index doesn't need key data. */
8132 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8138 mc->mc_pg[mc->mc_top] = copy;
8143 } while (i != split_indx);
8145 nkeys = NUMKEYS(copy);
8146 for (i=0; i<nkeys; i++)
8147 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8148 mp->mp_lower = copy->mp_lower;
8149 mp->mp_upper = copy->mp_upper;
8150 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8151 env->me_psize - copy->mp_upper - PAGEBASE);
8153 /* reset back to original page */
8154 if (newindx < split_indx) {
8155 mc->mc_pg[mc->mc_top] = mp;
8156 if (nflags & MDB_RESERVE) {
8157 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8158 if (!(node->mn_flags & F_BIGDATA))
8159 newdata->mv_data = NODEDATA(node);
8162 mc->mc_pg[mc->mc_top] = rp;
8164 /* Make sure mc_ki is still valid.
8166 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8167 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8168 for (i=0; i<=ptop; i++) {
8169 mc->mc_pg[i] = mn.mc_pg[i];
8170 mc->mc_ki[i] = mn.mc_ki[i];
8177 /* Adjust other cursors pointing to mp */
8178 MDB_cursor *m2, *m3;
8179 MDB_dbi dbi = mc->mc_dbi;
8180 int fixup = NUMKEYS(mp);
8182 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8183 if (mc->mc_flags & C_SUB)
8184 m3 = &m2->mc_xcursor->mx_cursor;
8189 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8191 if (m3->mc_flags & C_SPLITTING)
8196 for (k=m3->mc_top; k>=0; k--) {
8197 m3->mc_ki[k+1] = m3->mc_ki[k];
8198 m3->mc_pg[k+1] = m3->mc_pg[k];
8200 if (m3->mc_ki[0] >= split_indx) {
8205 m3->mc_pg[0] = mc->mc_pg[0];
8209 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8210 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8211 m3->mc_ki[mc->mc_top]++;
8212 if (m3->mc_ki[mc->mc_top] >= fixup) {
8213 m3->mc_pg[mc->mc_top] = rp;
8214 m3->mc_ki[mc->mc_top] -= fixup;
8215 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8217 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8218 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8223 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8226 if (copy) /* tmp page */
8227 mdb_page_free(env, copy);
8229 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8234 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8235 MDB_val *key, MDB_val *data, unsigned int flags)
8240 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8243 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8246 mdb_cursor_init(&mc, txn, dbi, &mx);
8247 return mdb_cursor_put(&mc, key, data, flags);
8251 #define MDB_WBUF (1024*1024)
8254 /** State needed for a compacting copy. */
8255 typedef struct mdb_copy {
8256 pthread_mutex_t mc_mutex;
8257 pthread_cond_t mc_cond;
8264 pgno_t mc_next_pgno;
8267 volatile int mc_new;
8272 /** Dedicated writer thread for compacting copy. */
8273 static THREAD_RET ESECT
8274 mdb_env_copythr(void *arg)
8278 int toggle = 0, wsize, rc;
8281 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8284 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8287 pthread_mutex_lock(&my->mc_mutex);
8289 pthread_cond_signal(&my->mc_cond);
8292 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8293 if (my->mc_new < 0) {
8298 wsize = my->mc_wlen[toggle];
8299 ptr = my->mc_wbuf[toggle];
8302 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8306 } else if (len > 0) {
8320 /* If there's an overflow page tail, write it too */
8321 if (my->mc_olen[toggle]) {
8322 wsize = my->mc_olen[toggle];
8323 ptr = my->mc_over[toggle];
8324 my->mc_olen[toggle] = 0;
8327 my->mc_wlen[toggle] = 0;
8329 pthread_cond_signal(&my->mc_cond);
8331 pthread_cond_signal(&my->mc_cond);
8332 pthread_mutex_unlock(&my->mc_mutex);
8333 return (THREAD_RET)0;
8337 /** Tell the writer thread there's a buffer ready to write */
8339 mdb_env_cthr_toggle(mdb_copy *my, int st)
8341 int toggle = my->mc_toggle ^ 1;
8342 pthread_mutex_lock(&my->mc_mutex);
8343 if (my->mc_status) {
8344 pthread_mutex_unlock(&my->mc_mutex);
8345 return my->mc_status;
8347 while (my->mc_new == 1)
8348 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8350 my->mc_toggle = toggle;
8351 pthread_cond_signal(&my->mc_cond);
8352 pthread_mutex_unlock(&my->mc_mutex);
8356 /** Depth-first tree traversal for compacting copy. */
8358 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8361 MDB_txn *txn = my->mc_txn;
8363 MDB_page *mo, *mp, *leaf;
8368 /* Empty DB, nothing to do */
8369 if (*pg == P_INVALID)
8376 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8379 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8383 /* Make cursor pages writable */
8384 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8388 for (i=0; i<mc.mc_top; i++) {
8389 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8390 mc.mc_pg[i] = (MDB_page *)ptr;
8391 ptr += my->mc_env->me_psize;
8394 /* This is writable space for a leaf page. Usually not needed. */
8395 leaf = (MDB_page *)ptr;
8397 toggle = my->mc_toggle;
8398 while (mc.mc_snum > 0) {
8400 mp = mc.mc_pg[mc.mc_top];
8404 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8405 for (i=0; i<n; i++) {
8406 ni = NODEPTR(mp, i);
8407 if (ni->mn_flags & F_BIGDATA) {
8411 /* Need writable leaf */
8413 mc.mc_pg[mc.mc_top] = leaf;
8414 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8416 ni = NODEPTR(mp, i);
8419 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8420 rc = mdb_page_get(txn, pg, &omp, NULL);
8423 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8424 rc = mdb_env_cthr_toggle(my, 1);
8427 toggle = my->mc_toggle;
8429 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8430 memcpy(mo, omp, my->mc_env->me_psize);
8431 mo->mp_pgno = my->mc_next_pgno;
8432 my->mc_next_pgno += omp->mp_pages;
8433 my->mc_wlen[toggle] += my->mc_env->me_psize;
8434 if (omp->mp_pages > 1) {
8435 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8436 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8437 rc = mdb_env_cthr_toggle(my, 1);
8440 toggle = my->mc_toggle;
8442 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8443 } else if (ni->mn_flags & F_SUBDATA) {
8446 /* Need writable leaf */
8448 mc.mc_pg[mc.mc_top] = leaf;
8449 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8451 ni = NODEPTR(mp, i);
8454 memcpy(&db, NODEDATA(ni), sizeof(db));
8455 my->mc_toggle = toggle;
8456 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8459 toggle = my->mc_toggle;
8460 memcpy(NODEDATA(ni), &db, sizeof(db));
8465 mc.mc_ki[mc.mc_top]++;
8466 if (mc.mc_ki[mc.mc_top] < n) {
8469 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8471 rc = mdb_page_get(txn, pg, &mp, NULL);
8476 mc.mc_ki[mc.mc_top] = 0;
8477 if (IS_BRANCH(mp)) {
8478 /* Whenever we advance to a sibling branch page,
8479 * we must proceed all the way down to its first leaf.
8481 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8484 mc.mc_pg[mc.mc_top] = mp;
8488 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8489 rc = mdb_env_cthr_toggle(my, 1);
8492 toggle = my->mc_toggle;
8494 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8495 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8496 mo->mp_pgno = my->mc_next_pgno++;
8497 my->mc_wlen[toggle] += my->mc_env->me_psize;
8499 /* Update parent if there is one */
8500 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8501 SETPGNO(ni, mo->mp_pgno);
8502 mdb_cursor_pop(&mc);
8504 /* Otherwise we're done */
8514 /** Copy environment with compaction. */
8516 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8521 MDB_txn *txn = NULL;
8526 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8527 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8528 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8529 if (my.mc_wbuf[0] == NULL)
8532 pthread_mutex_init(&my.mc_mutex, NULL);
8533 pthread_cond_init(&my.mc_cond, NULL);
8534 #ifdef HAVE_MEMALIGN
8535 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8536 if (my.mc_wbuf[0] == NULL)
8539 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8544 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8545 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8550 my.mc_next_pgno = 2;
8556 THREAD_CREATE(thr, mdb_env_copythr, &my);
8558 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8562 mp = (MDB_page *)my.mc_wbuf[0];
8563 memset(mp, 0, 2*env->me_psize);
8565 mp->mp_flags = P_META;
8566 mm = (MDB_meta *)METADATA(mp);
8567 mdb_env_init_meta0(env, mm);
8568 mm->mm_address = env->me_metas[0]->mm_address;
8570 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8572 mp->mp_flags = P_META;
8573 *(MDB_meta *)METADATA(mp) = *mm;
8574 mm = (MDB_meta *)METADATA(mp);
8576 /* Count the number of free pages, subtract from lastpg to find
8577 * number of active pages
8580 MDB_ID freecount = 0;
8583 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8584 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8585 freecount += *(MDB_ID *)data.mv_data;
8586 freecount += txn->mt_dbs[0].md_branch_pages +
8587 txn->mt_dbs[0].md_leaf_pages +
8588 txn->mt_dbs[0].md_overflow_pages;
8590 /* Set metapage 1 */
8591 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8592 mm->mm_dbs[1] = txn->mt_dbs[1];
8593 if (mm->mm_last_pg > 1) {
8594 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8597 mm->mm_dbs[1].md_root = P_INVALID;
8600 my.mc_wlen[0] = env->me_psize * 2;
8602 pthread_mutex_lock(&my.mc_mutex);
8604 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8605 pthread_mutex_unlock(&my.mc_mutex);
8606 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8607 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8608 rc = mdb_env_cthr_toggle(&my, 1);
8609 mdb_env_cthr_toggle(&my, -1);
8610 pthread_mutex_lock(&my.mc_mutex);
8612 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8613 pthread_mutex_unlock(&my.mc_mutex);
8618 CloseHandle(my.mc_cond);
8619 CloseHandle(my.mc_mutex);
8620 _aligned_free(my.mc_wbuf[0]);
8622 pthread_cond_destroy(&my.mc_cond);
8623 pthread_mutex_destroy(&my.mc_mutex);
8624 free(my.mc_wbuf[0]);
8629 /** Copy environment as-is. */
8631 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8633 MDB_txn *txn = NULL;
8639 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8643 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8646 /* Do the lock/unlock of the reader mutex before starting the
8647 * write txn. Otherwise other read txns could block writers.
8649 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8654 /* We must start the actual read txn after blocking writers */
8655 mdb_txn_reset0(txn, "reset-stage1");
8657 /* Temporarily block writers until we snapshot the meta pages */
8660 rc = mdb_txn_renew0(txn);
8662 UNLOCK_MUTEX_W(env);
8667 wsize = env->me_psize * 2;
8671 DO_WRITE(rc, fd, ptr, w2, len);
8675 } else if (len > 0) {
8681 /* Non-blocking or async handles are not supported */
8687 UNLOCK_MUTEX_W(env);
8692 w2 = txn->mt_next_pgno * env->me_psize;
8695 LARGE_INTEGER fsize;
8697 if (!GetFileSizeEx(env->me_fd, &fsize)) {
8702 if (w2 > fsize.QuadPart)
8703 w2 = fsize.QuadPart;
8709 if ((rc = fstat(env->me_fd, &st))) {
8713 if (w2 > (size_t)st.st_size)
8719 if (wsize > MAX_WRITE)
8723 DO_WRITE(rc, fd, ptr, w2, len);
8727 } else if (len > 0) {
8744 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8746 if (flags & MDB_CP_COMPACT)
8747 return mdb_env_copyfd1(env, fd);
8749 return mdb_env_copyfd0(env, fd);
8753 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8755 return mdb_env_copyfd2(env, fd, 0);
8759 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8763 HANDLE newfd = INVALID_HANDLE_VALUE;
8765 if (env->me_flags & MDB_NOSUBDIR) {
8766 lpath = (char *)path;
8769 len += sizeof(DATANAME);
8770 lpath = malloc(len);
8773 sprintf(lpath, "%s" DATANAME, path);
8776 /* The destination path must exist, but the destination file must not.
8777 * We don't want the OS to cache the writes, since the source data is
8778 * already in the OS cache.
8781 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8782 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8784 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8786 if (newfd == INVALID_HANDLE_VALUE) {
8791 if (env->me_psize >= env->me_os_psize) {
8793 /* Set O_DIRECT if the file system supports it */
8794 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8795 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8797 #ifdef F_NOCACHE /* __APPLE__ */
8798 rc = fcntl(newfd, F_NOCACHE, 1);
8806 rc = mdb_env_copyfd2(env, newfd, flags);
8809 if (!(env->me_flags & MDB_NOSUBDIR))
8811 if (newfd != INVALID_HANDLE_VALUE)
8812 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8819 mdb_env_copy(MDB_env *env, const char *path)
8821 return mdb_env_copy2(env, path, 0);
8825 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8827 if ((flag & CHANGEABLE) != flag)
8830 env->me_flags |= flag;
8832 env->me_flags &= ~flag;
8837 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8842 *arg = env->me_flags;
8847 mdb_env_set_userctx(MDB_env *env, void *ctx)
8851 env->me_userctx = ctx;
8856 mdb_env_get_userctx(MDB_env *env)
8858 return env ? env->me_userctx : NULL;
8862 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8867 env->me_assert_func = func;
8873 mdb_env_get_path(MDB_env *env, const char **arg)
8878 *arg = env->me_path;
8883 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8892 /** Common code for #mdb_stat() and #mdb_env_stat().
8893 * @param[in] env the environment to operate in.
8894 * @param[in] db the #MDB_db record containing the stats to return.
8895 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8896 * @return 0, this function always succeeds.
8899 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8901 arg->ms_psize = env->me_psize;
8902 arg->ms_depth = db->md_depth;
8903 arg->ms_branch_pages = db->md_branch_pages;
8904 arg->ms_leaf_pages = db->md_leaf_pages;
8905 arg->ms_overflow_pages = db->md_overflow_pages;
8906 arg->ms_entries = db->md_entries;
8912 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8916 if (env == NULL || arg == NULL)
8919 toggle = mdb_env_pick_meta(env);
8921 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8925 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8929 if (env == NULL || arg == NULL)
8932 toggle = mdb_env_pick_meta(env);
8933 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8934 arg->me_mapsize = env->me_mapsize;
8935 arg->me_maxreaders = env->me_maxreaders;
8937 /* me_numreaders may be zero if this process never used any readers. Use
8938 * the shared numreader count if it exists.
8940 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8942 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8943 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8947 /** Set the default comparison functions for a database.
8948 * Called immediately after a database is opened to set the defaults.
8949 * The user can then override them with #mdb_set_compare() or
8950 * #mdb_set_dupsort().
8951 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8952 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8955 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8957 uint16_t f = txn->mt_dbs[dbi].md_flags;
8959 txn->mt_dbxs[dbi].md_cmp =
8960 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8961 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8963 txn->mt_dbxs[dbi].md_dcmp =
8964 !(f & MDB_DUPSORT) ? 0 :
8965 ((f & MDB_INTEGERDUP)
8966 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8967 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8970 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8976 int rc, dbflag, exact;
8977 unsigned int unused = 0, seq;
8980 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8981 mdb_default_cmp(txn, FREE_DBI);
8984 if ((flags & VALID_FLAGS) != flags)
8986 if (txn->mt_flags & MDB_TXN_ERROR)
8992 if (flags & PERSISTENT_FLAGS) {
8993 uint16_t f2 = flags & PERSISTENT_FLAGS;
8994 /* make sure flag changes get committed */
8995 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8996 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8997 txn->mt_flags |= MDB_TXN_DIRTY;
9000 mdb_default_cmp(txn, MAIN_DBI);
9004 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9005 mdb_default_cmp(txn, MAIN_DBI);
9008 /* Is the DB already open? */
9010 for (i=2; i<txn->mt_numdbs; i++) {
9011 if (!txn->mt_dbxs[i].md_name.mv_size) {
9012 /* Remember this free slot */
9013 if (!unused) unused = i;
9016 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9017 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9023 /* If no free slot and max hit, fail */
9024 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9025 return MDB_DBS_FULL;
9027 /* Cannot mix named databases with some mainDB flags */
9028 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9029 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9031 /* Find the DB info */
9032 dbflag = DB_NEW|DB_VALID;
9035 key.mv_data = (void *)name;
9036 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9037 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9038 if (rc == MDB_SUCCESS) {
9039 /* make sure this is actually a DB */
9040 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9041 if (!(node->mn_flags & F_SUBDATA))
9042 return MDB_INCOMPATIBLE;
9043 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9044 /* Create if requested */
9045 data.mv_size = sizeof(MDB_db);
9046 data.mv_data = &dummy;
9047 memset(&dummy, 0, sizeof(dummy));
9048 dummy.md_root = P_INVALID;
9049 dummy.md_flags = flags & PERSISTENT_FLAGS;
9050 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9054 /* OK, got info, add to table */
9055 if (rc == MDB_SUCCESS) {
9056 unsigned int slot = unused ? unused : txn->mt_numdbs;
9057 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9058 txn->mt_dbxs[slot].md_name.mv_size = len;
9059 txn->mt_dbxs[slot].md_rel = NULL;
9060 txn->mt_dbflags[slot] = dbflag;
9061 /* txn-> and env-> are the same in read txns, use
9062 * tmp variable to avoid undefined assignment
9064 seq = ++txn->mt_env->me_dbiseqs[slot];
9065 txn->mt_dbiseqs[slot] = seq;
9067 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9069 mdb_default_cmp(txn, slot);
9078 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9080 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9083 if (txn->mt_flags & MDB_TXN_ERROR)
9086 if (txn->mt_dbflags[dbi] & DB_STALE) {
9089 /* Stale, must read the DB's root. cursor_init does it for us. */
9090 mdb_cursor_init(&mc, txn, dbi, &mx);
9092 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9095 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9098 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9100 ptr = env->me_dbxs[dbi].md_name.mv_data;
9101 /* If there was no name, this was already closed */
9103 env->me_dbxs[dbi].md_name.mv_data = NULL;
9104 env->me_dbxs[dbi].md_name.mv_size = 0;
9105 env->me_dbflags[dbi] = 0;
9106 env->me_dbiseqs[dbi]++;
9111 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9113 /* We could return the flags for the FREE_DBI too but what's the point? */
9114 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9116 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9120 /** Add all the DB's pages to the free list.
9121 * @param[in] mc Cursor on the DB to free.
9122 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9123 * @return 0 on success, non-zero on failure.
9126 mdb_drop0(MDB_cursor *mc, int subs)
9130 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9131 if (rc == MDB_SUCCESS) {
9132 MDB_txn *txn = mc->mc_txn;
9137 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9138 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9141 mdb_cursor_copy(mc, &mx);
9142 while (mc->mc_snum > 0) {
9143 MDB_page *mp = mc->mc_pg[mc->mc_top];
9144 unsigned n = NUMKEYS(mp);
9146 for (i=0; i<n; i++) {
9147 ni = NODEPTR(mp, i);
9148 if (ni->mn_flags & F_BIGDATA) {
9151 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9152 rc = mdb_page_get(txn, pg, &omp, NULL);
9155 mdb_cassert(mc, IS_OVERFLOW(omp));
9156 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9160 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9161 mdb_xcursor_init1(mc, ni);
9162 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9168 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9170 for (i=0; i<n; i++) {
9172 ni = NODEPTR(mp, i);
9175 mdb_midl_xappend(txn->mt_free_pgs, pg);
9180 mc->mc_ki[mc->mc_top] = i;
9181 rc = mdb_cursor_sibling(mc, 1);
9183 if (rc != MDB_NOTFOUND)
9185 /* no more siblings, go back to beginning
9186 * of previous level.
9190 for (i=1; i<mc->mc_snum; i++) {
9192 mc->mc_pg[i] = mx.mc_pg[i];
9197 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9200 txn->mt_flags |= MDB_TXN_ERROR;
9201 } else if (rc == MDB_NOTFOUND) {
9207 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9209 MDB_cursor *mc, *m2;
9212 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9215 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9218 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9221 rc = mdb_cursor_open(txn, dbi, &mc);
9225 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9226 /* Invalidate the dropped DB's cursors */
9227 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9228 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9232 /* Can't delete the main DB */
9233 if (del && dbi > MAIN_DBI) {
9234 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9236 txn->mt_dbflags[dbi] = DB_STALE;
9237 mdb_dbi_close(txn->mt_env, dbi);
9239 txn->mt_flags |= MDB_TXN_ERROR;
9242 /* reset the DB record, mark it dirty */
9243 txn->mt_dbflags[dbi] |= DB_DIRTY;
9244 txn->mt_dbs[dbi].md_depth = 0;
9245 txn->mt_dbs[dbi].md_branch_pages = 0;
9246 txn->mt_dbs[dbi].md_leaf_pages = 0;
9247 txn->mt_dbs[dbi].md_overflow_pages = 0;
9248 txn->mt_dbs[dbi].md_entries = 0;
9249 txn->mt_dbs[dbi].md_root = P_INVALID;
9251 txn->mt_flags |= MDB_TXN_DIRTY;
9254 mdb_cursor_close(mc);
9258 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9260 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9263 txn->mt_dbxs[dbi].md_cmp = cmp;
9267 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9269 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9272 txn->mt_dbxs[dbi].md_dcmp = cmp;
9276 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9278 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9281 txn->mt_dbxs[dbi].md_rel = rel;
9285 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9287 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9290 txn->mt_dbxs[dbi].md_relctx = ctx;
9295 mdb_env_get_maxkeysize(MDB_env *env)
9297 return ENV_MAXKEY(env);
9301 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9303 unsigned int i, rdrs;
9306 int rc = 0, first = 1;
9310 if (!env->me_txns) {
9311 return func("(no reader locks)\n", ctx);
9313 rdrs = env->me_txns->mti_numreaders;
9314 mr = env->me_txns->mti_readers;
9315 for (i=0; i<rdrs; i++) {
9317 txnid_t txnid = mr[i].mr_txnid;
9318 sprintf(buf, txnid == (txnid_t)-1 ?
9319 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9320 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9323 rc = func(" pid thread txnid\n", ctx);
9327 rc = func(buf, ctx);
9333 rc = func("(no active readers)\n", ctx);
9338 /** Insert pid into list if not already present.
9339 * return -1 if already present.
9342 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9344 /* binary search of pid in list */
9346 unsigned cursor = 1;
9348 unsigned n = ids[0];
9351 unsigned pivot = n >> 1;
9352 cursor = base + pivot + 1;
9353 val = pid - ids[cursor];
9358 } else if ( val > 0 ) {
9363 /* found, so it's a duplicate */
9372 for (n = ids[0]; n > cursor; n--)
9379 mdb_reader_check(MDB_env *env, int *dead)
9381 unsigned int i, j, rdrs;
9383 MDB_PID_T *pids, pid;
9392 rdrs = env->me_txns->mti_numreaders;
9393 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9397 mr = env->me_txns->mti_readers;
9398 for (i=0; i<rdrs; i++) {
9399 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9401 if (mdb_pid_insert(pids, pid) == 0) {
9402 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9404 /* Recheck, a new process may have reused pid */
9405 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9406 for (j=i; j<rdrs; j++)
9407 if (mr[j].mr_pid == pid) {
9408 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9409 (unsigned) pid, mr[j].mr_txnid));
9414 UNLOCK_MUTEX_R(env);