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)
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 size_t me_mapsize; /**< size of the data memory map */
1114 off_t me_size; /**< current file size */
1115 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1116 MDB_dbx *me_dbxs; /**< array of static DB info */
1117 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1118 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1119 pthread_key_t me_txkey; /**< thread-key for readers */
1120 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1121 # define me_pglast me_pgstate.mf_pglast
1122 # define me_pghead me_pgstate.mf_pghead
1123 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1124 /** IDL of pages that became unused in a write txn */
1125 MDB_IDL me_free_pgs;
1126 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1127 MDB_ID2L me_dirty_list;
1128 /** Max number of freelist items that can fit in a single overflow page */
1130 /** Max size of a node on a page */
1131 unsigned int me_nodemax;
1132 #if !(MDB_MAXKEYSIZE)
1133 unsigned int me_maxkey; /**< max size of a key */
1135 int me_live_reader; /**< have liveness lock in reader table */
1137 int me_pidquery; /**< Used in OpenProcess */
1138 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1140 #elif defined(MDB_USE_POSIX_SEM)
1141 sem_t *me_rmutex; /* Shared mutexes are not supported */
1144 void *me_userctx; /**< User-settable context */
1145 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1148 /** Nested transaction */
1149 typedef struct MDB_ntxn {
1150 MDB_txn mnt_txn; /**< the transaction */
1151 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1154 /** max number of pages to commit in one writev() call */
1155 #define MDB_COMMIT_PAGES 64
1156 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1157 #undef MDB_COMMIT_PAGES
1158 #define MDB_COMMIT_PAGES IOV_MAX
1161 /** max bytes to write in one call */
1162 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1164 /** Check \b txn and \b dbi arguments to a function */
1165 #define TXN_DBI_EXIST(txn, dbi) \
1166 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1168 /** Check for misused \b dbi handles */
1169 #define TXN_DBI_CHANGED(txn, dbi) \
1170 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1172 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1173 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1174 static int mdb_page_touch(MDB_cursor *mc);
1176 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1177 static int mdb_page_search_root(MDB_cursor *mc,
1178 MDB_val *key, int modify);
1179 #define MDB_PS_MODIFY 1
1180 #define MDB_PS_ROOTONLY 2
1181 #define MDB_PS_FIRST 4
1182 #define MDB_PS_LAST 8
1183 static int mdb_page_search(MDB_cursor *mc,
1184 MDB_val *key, int flags);
1185 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1187 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1188 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1189 pgno_t newpgno, unsigned int nflags);
1191 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1192 static int mdb_env_pick_meta(const MDB_env *env);
1193 static int mdb_env_write_meta(MDB_txn *txn);
1194 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1195 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1197 static void mdb_env_close0(MDB_env *env, int excl);
1199 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1200 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1201 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1202 static void mdb_node_del(MDB_cursor *mc, int ksize);
1203 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1204 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1205 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1206 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1207 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1209 static int mdb_rebalance(MDB_cursor *mc);
1210 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1212 static void mdb_cursor_pop(MDB_cursor *mc);
1213 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1215 static int mdb_cursor_del0(MDB_cursor *mc);
1216 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1217 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1218 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1219 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1220 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1222 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1223 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1225 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1226 static void mdb_xcursor_init0(MDB_cursor *mc);
1227 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1229 static int mdb_drop0(MDB_cursor *mc, int subs);
1230 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1233 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1237 static SECURITY_DESCRIPTOR mdb_null_sd;
1238 static SECURITY_ATTRIBUTES mdb_all_sa;
1239 static int mdb_sec_inited;
1242 /** Return the library version info. */
1244 mdb_version(int *major, int *minor, int *patch)
1246 if (major) *major = MDB_VERSION_MAJOR;
1247 if (minor) *minor = MDB_VERSION_MINOR;
1248 if (patch) *patch = MDB_VERSION_PATCH;
1249 return MDB_VERSION_STRING;
1252 /** Table of descriptions for LMDB @ref errors */
1253 static char *const mdb_errstr[] = {
1254 "MDB_KEYEXIST: Key/data pair already exists",
1255 "MDB_NOTFOUND: No matching key/data pair found",
1256 "MDB_PAGE_NOTFOUND: Requested page not found",
1257 "MDB_CORRUPTED: Located page was wrong type",
1258 "MDB_PANIC: Update of meta page failed",
1259 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1260 "MDB_INVALID: File is not an LMDB file",
1261 "MDB_MAP_FULL: Environment mapsize limit reached",
1262 "MDB_DBS_FULL: Environment maxdbs limit reached",
1263 "MDB_READERS_FULL: Environment maxreaders limit reached",
1264 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1265 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1266 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1267 "MDB_PAGE_FULL: Internal error - page has no more space",
1268 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1269 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1270 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1271 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1272 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1273 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1277 mdb_strerror(int err)
1280 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1281 * This works as long as no function between the call to mdb_strerror
1282 * and the actual use of the message uses more than 4K of stack.
1285 char buf[1024], *ptr = buf;
1289 return ("Successful return: 0");
1291 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1292 i = err - MDB_KEYEXIST;
1293 return mdb_errstr[i];
1297 /* These are the C-runtime error codes we use. The comment indicates
1298 * their numeric value, and the Win32 error they would correspond to
1299 * if the error actually came from a Win32 API. A major mess, we should
1300 * have used LMDB-specific error codes for everything.
1303 case ENOENT: /* 2, FILE_NOT_FOUND */
1304 case EIO: /* 5, ACCESS_DENIED */
1305 case ENOMEM: /* 12, INVALID_ACCESS */
1306 case EACCES: /* 13, INVALID_DATA */
1307 case EBUSY: /* 16, CURRENT_DIRECTORY */
1308 case EINVAL: /* 22, BAD_COMMAND */
1309 case ENOSPC: /* 28, OUT_OF_PAPER */
1310 return strerror(err);
1315 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1316 FORMAT_MESSAGE_IGNORE_INSERTS,
1317 NULL, err, 0, ptr, sizeof(buf), pad);
1320 return strerror(err);
1324 /** assert(3) variant in cursor context */
1325 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1326 /** assert(3) variant in transaction context */
1327 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1328 /** assert(3) variant in environment context */
1329 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1332 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1333 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1336 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1337 const char *func, const char *file, int line)
1340 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1341 file, line, expr_txt, func);
1342 if (env->me_assert_func)
1343 env->me_assert_func(env, buf);
1344 fprintf(stderr, "%s\n", buf);
1348 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1352 /** Return the page number of \b mp which may be sub-page, for debug output */
1354 mdb_dbg_pgno(MDB_page *mp)
1357 COPY_PGNO(ret, mp->mp_pgno);
1361 /** Display a key in hexadecimal and return the address of the result.
1362 * @param[in] key the key to display
1363 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1364 * @return The key in hexadecimal form.
1367 mdb_dkey(MDB_val *key, char *buf)
1370 unsigned char *c = key->mv_data;
1376 if (key->mv_size > DKBUF_MAXKEYSIZE)
1377 return "MDB_MAXKEYSIZE";
1378 /* may want to make this a dynamic check: if the key is mostly
1379 * printable characters, print it as-is instead of converting to hex.
1383 for (i=0; i<key->mv_size; i++)
1384 ptr += sprintf(ptr, "%02x", *c++);
1386 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1392 mdb_leafnode_type(MDB_node *n)
1394 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1395 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1396 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1399 /** Display all the keys in the page. */
1401 mdb_page_list(MDB_page *mp)
1403 pgno_t pgno = mdb_dbg_pgno(mp);
1404 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1406 unsigned int i, nkeys, nsize, total = 0;
1410 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1411 case P_BRANCH: type = "Branch page"; break;
1412 case P_LEAF: type = "Leaf page"; break;
1413 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1414 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1415 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1417 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1418 pgno, mp->mp_pages, state);
1421 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1422 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1425 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1429 nkeys = NUMKEYS(mp);
1430 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1432 for (i=0; i<nkeys; i++) {
1433 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1434 key.mv_size = nsize = mp->mp_pad;
1435 key.mv_data = LEAF2KEY(mp, i, nsize);
1437 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1440 node = NODEPTR(mp, i);
1441 key.mv_size = node->mn_ksize;
1442 key.mv_data = node->mn_data;
1443 nsize = NODESIZE + key.mv_size;
1444 if (IS_BRANCH(mp)) {
1445 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1449 if (F_ISSET(node->mn_flags, F_BIGDATA))
1450 nsize += sizeof(pgno_t);
1452 nsize += NODEDSZ(node);
1454 nsize += sizeof(indx_t);
1455 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1456 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1458 total = EVEN(total);
1460 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1461 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1465 mdb_cursor_chk(MDB_cursor *mc)
1471 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1472 for (i=0; i<mc->mc_top; i++) {
1474 node = NODEPTR(mp, mc->mc_ki[i]);
1475 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1478 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1484 /** Count all the pages in each DB and in the freelist
1485 * and make sure it matches the actual number of pages
1487 * All named DBs must be open for a correct count.
1489 static void mdb_audit(MDB_txn *txn)
1493 MDB_ID freecount, count;
1498 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1499 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1500 freecount += *(MDB_ID *)data.mv_data;
1501 mdb_tassert(txn, rc == MDB_NOTFOUND);
1504 for (i = 0; i<txn->mt_numdbs; i++) {
1506 if (!(txn->mt_dbflags[i] & DB_VALID))
1508 mdb_cursor_init(&mc, txn, i, &mx);
1509 if (txn->mt_dbs[i].md_root == P_INVALID)
1511 count += txn->mt_dbs[i].md_branch_pages +
1512 txn->mt_dbs[i].md_leaf_pages +
1513 txn->mt_dbs[i].md_overflow_pages;
1514 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1515 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1516 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1519 mp = mc.mc_pg[mc.mc_top];
1520 for (j=0; j<NUMKEYS(mp); j++) {
1521 MDB_node *leaf = NODEPTR(mp, j);
1522 if (leaf->mn_flags & F_SUBDATA) {
1524 memcpy(&db, NODEDATA(leaf), sizeof(db));
1525 count += db.md_branch_pages + db.md_leaf_pages +
1526 db.md_overflow_pages;
1530 mdb_tassert(txn, rc == MDB_NOTFOUND);
1533 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1534 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1535 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1541 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1543 return txn->mt_dbxs[dbi].md_cmp(a, b);
1547 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1549 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1552 /** Allocate memory for a page.
1553 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1556 mdb_page_malloc(MDB_txn *txn, unsigned num)
1558 MDB_env *env = txn->mt_env;
1559 MDB_page *ret = env->me_dpages;
1560 size_t psize = env->me_psize, sz = psize, off;
1561 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1562 * For a single page alloc, we init everything after the page header.
1563 * For multi-page, we init the final page; if the caller needed that
1564 * many pages they will be filling in at least up to the last page.
1568 VGMEMP_ALLOC(env, ret, sz);
1569 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1570 env->me_dpages = ret->mp_next;
1573 psize -= off = PAGEHDRSZ;
1578 if ((ret = malloc(sz)) != NULL) {
1579 VGMEMP_ALLOC(env, ret, sz);
1580 if (!(env->me_flags & MDB_NOMEMINIT)) {
1581 memset((char *)ret + off, 0, psize);
1585 txn->mt_flags |= MDB_TXN_ERROR;
1589 /** Free a single page.
1590 * Saves single pages to a list, for future reuse.
1591 * (This is not used for multi-page overflow pages.)
1594 mdb_page_free(MDB_env *env, MDB_page *mp)
1596 mp->mp_next = env->me_dpages;
1597 VGMEMP_FREE(env, mp);
1598 env->me_dpages = mp;
1601 /** Free a dirty page */
1603 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1605 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1606 mdb_page_free(env, dp);
1608 /* large pages just get freed directly */
1609 VGMEMP_FREE(env, dp);
1614 /** Return all dirty pages to dpage list */
1616 mdb_dlist_free(MDB_txn *txn)
1618 MDB_env *env = txn->mt_env;
1619 MDB_ID2L dl = txn->mt_u.dirty_list;
1620 unsigned i, n = dl[0].mid;
1622 for (i = 1; i <= n; i++) {
1623 mdb_dpage_free(env, dl[i].mptr);
1628 /** Loosen or free a single page.
1629 * Saves single pages to a list for future reuse
1630 * in this same txn. It has been pulled from the freeDB
1631 * and already resides on the dirty list, but has been
1632 * deleted. Use these pages first before pulling again
1635 * If the page wasn't dirtied in this txn, just add it
1636 * to this txn's free list.
1639 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1642 pgno_t pgno = mp->mp_pgno;
1643 MDB_txn *txn = mc->mc_txn;
1645 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1646 if (txn->mt_parent) {
1647 MDB_ID2 *dl = txn->mt_u.dirty_list;
1648 /* If txn has a parent, make sure the page is in our
1652 unsigned x = mdb_mid2l_search(dl, pgno);
1653 if (x <= dl[0].mid && dl[x].mid == pgno) {
1654 if (mp != dl[x].mptr) { /* bad cursor? */
1655 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1656 txn->mt_flags |= MDB_TXN_ERROR;
1657 return MDB_CORRUPTED;
1664 /* no parent txn, so it's just ours */
1669 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1671 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1672 txn->mt_loose_pgs = mp;
1673 txn->mt_loose_count++;
1674 mp->mp_flags |= P_LOOSE;
1676 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1684 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1685 * @param[in] mc A cursor handle for the current operation.
1686 * @param[in] pflags Flags of the pages to update:
1687 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1688 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1689 * @return 0 on success, non-zero on failure.
1692 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1694 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1695 MDB_txn *txn = mc->mc_txn;
1701 int rc = MDB_SUCCESS, level;
1703 /* Mark pages seen by cursors */
1704 if (mc->mc_flags & C_UNTRACK)
1705 mc = NULL; /* will find mc in mt_cursors */
1706 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1707 for (; mc; mc=mc->mc_next) {
1708 if (!(mc->mc_flags & C_INITIALIZED))
1710 for (m3 = mc;; m3 = &mx->mx_cursor) {
1712 for (j=0; j<m3->mc_snum; j++) {
1714 if ((mp->mp_flags & Mask) == pflags)
1715 mp->mp_flags ^= P_KEEP;
1717 mx = m3->mc_xcursor;
1718 /* Proceed to mx if it is at a sub-database */
1719 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1721 if (! (mp && (mp->mp_flags & P_LEAF)))
1723 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1724 if (!(leaf->mn_flags & F_SUBDATA))
1733 /* Mark dirty root pages */
1734 for (i=0; i<txn->mt_numdbs; i++) {
1735 if (txn->mt_dbflags[i] & DB_DIRTY) {
1736 pgno_t pgno = txn->mt_dbs[i].md_root;
1737 if (pgno == P_INVALID)
1739 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1741 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1742 dp->mp_flags ^= P_KEEP;
1750 static int mdb_page_flush(MDB_txn *txn, int keep);
1752 /** Spill pages from the dirty list back to disk.
1753 * This is intended to prevent running into #MDB_TXN_FULL situations,
1754 * but note that they may still occur in a few cases:
1755 * 1) our estimate of the txn size could be too small. Currently this
1756 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1757 * 2) child txns may run out of space if their parents dirtied a
1758 * lot of pages and never spilled them. TODO: we probably should do
1759 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1760 * the parent's dirty_room is below a given threshold.
1762 * Otherwise, if not using nested txns, it is expected that apps will
1763 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1764 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1765 * If the txn never references them again, they can be left alone.
1766 * If the txn only reads them, they can be used without any fuss.
1767 * If the txn writes them again, they can be dirtied immediately without
1768 * going thru all of the work of #mdb_page_touch(). Such references are
1769 * handled by #mdb_page_unspill().
1771 * Also note, we never spill DB root pages, nor pages of active cursors,
1772 * because we'll need these back again soon anyway. And in nested txns,
1773 * we can't spill a page in a child txn if it was already spilled in a
1774 * parent txn. That would alter the parent txns' data even though
1775 * the child hasn't committed yet, and we'd have no way to undo it if
1776 * the child aborted.
1778 * @param[in] m0 cursor A cursor handle identifying the transaction and
1779 * database for which we are checking space.
1780 * @param[in] key For a put operation, the key being stored.
1781 * @param[in] data For a put operation, the data being stored.
1782 * @return 0 on success, non-zero on failure.
1785 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1787 MDB_txn *txn = m0->mc_txn;
1789 MDB_ID2L dl = txn->mt_u.dirty_list;
1790 unsigned int i, j, need;
1793 if (m0->mc_flags & C_SUB)
1796 /* Estimate how much space this op will take */
1797 i = m0->mc_db->md_depth;
1798 /* Named DBs also dirty the main DB */
1799 if (m0->mc_dbi > MAIN_DBI)
1800 i += txn->mt_dbs[MAIN_DBI].md_depth;
1801 /* For puts, roughly factor in the key+data size */
1803 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1804 i += i; /* double it for good measure */
1807 if (txn->mt_dirty_room > i)
1810 if (!txn->mt_spill_pgs) {
1811 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1812 if (!txn->mt_spill_pgs)
1815 /* purge deleted slots */
1816 MDB_IDL sl = txn->mt_spill_pgs;
1817 unsigned int num = sl[0];
1819 for (i=1; i<=num; i++) {
1826 /* Preserve pages which may soon be dirtied again */
1827 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1830 /* Less aggressive spill - we originally spilled the entire dirty list,
1831 * with a few exceptions for cursor pages and DB root pages. But this
1832 * turns out to be a lot of wasted effort because in a large txn many
1833 * of those pages will need to be used again. So now we spill only 1/8th
1834 * of the dirty pages. Testing revealed this to be a good tradeoff,
1835 * better than 1/2, 1/4, or 1/10.
1837 if (need < MDB_IDL_UM_MAX / 8)
1838 need = MDB_IDL_UM_MAX / 8;
1840 /* Save the page IDs of all the pages we're flushing */
1841 /* flush from the tail forward, this saves a lot of shifting later on. */
1842 for (i=dl[0].mid; i && need; i--) {
1843 MDB_ID pn = dl[i].mid << 1;
1845 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1847 /* Can't spill twice, make sure it's not already in a parent's
1850 if (txn->mt_parent) {
1852 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1853 if (tx2->mt_spill_pgs) {
1854 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1855 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1856 dp->mp_flags |= P_KEEP;
1864 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1868 mdb_midl_sort(txn->mt_spill_pgs);
1870 /* Flush the spilled part of dirty list */
1871 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1874 /* Reset any dirty pages we kept that page_flush didn't see */
1875 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1878 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1882 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1884 mdb_find_oldest(MDB_txn *txn)
1887 txnid_t mr, oldest = txn->mt_txnid - 1;
1888 if (txn->mt_env->me_txns) {
1889 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1890 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1901 /** Add a page to the txn's dirty list */
1903 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1906 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1908 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1909 insert = mdb_mid2l_append;
1911 insert = mdb_mid2l_insert;
1913 mid.mid = mp->mp_pgno;
1915 rc = insert(txn->mt_u.dirty_list, &mid);
1916 mdb_tassert(txn, rc == 0);
1917 txn->mt_dirty_room--;
1920 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1921 * me_pghead and mt_next_pgno.
1923 * If there are free pages available from older transactions, they
1924 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1925 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1926 * and move me_pglast to say which records were consumed. Only this
1927 * function can create me_pghead and move me_pglast/mt_next_pgno.
1928 * @param[in] mc cursor A cursor handle identifying the transaction and
1929 * database for which we are allocating.
1930 * @param[in] num the number of pages to allocate.
1931 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1932 * will always be satisfied by a single contiguous chunk of memory.
1933 * @return 0 on success, non-zero on failure.
1936 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1938 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1939 /* Get at most <Max_retries> more freeDB records once me_pghead
1940 * has enough pages. If not enough, use new pages from the map.
1941 * If <Paranoid> and mc is updating the freeDB, only get new
1942 * records if me_pghead is empty. Then the freelist cannot play
1943 * catch-up with itself by growing while trying to save it.
1945 enum { Paranoid = 1, Max_retries = 500 };
1947 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1949 int rc, retry = num * 20;
1950 MDB_txn *txn = mc->mc_txn;
1951 MDB_env *env = txn->mt_env;
1952 pgno_t pgno, *mop = env->me_pghead;
1953 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1955 txnid_t oldest = 0, last;
1959 /* If there are any loose pages, just use them */
1960 if (num == 1 && txn->mt_loose_pgs) {
1961 np = txn->mt_loose_pgs;
1962 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1963 txn->mt_loose_count--;
1964 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1972 /* If our dirty list is already full, we can't do anything */
1973 if (txn->mt_dirty_room == 0) {
1978 for (op = MDB_FIRST;; op = MDB_NEXT) {
1983 /* Seek a big enough contiguous page range. Prefer
1984 * pages at the tail, just truncating the list.
1990 if (mop[i-n2] == pgno+n2)
1997 if (op == MDB_FIRST) { /* 1st iteration */
1998 /* Prepare to fetch more and coalesce */
1999 oldest = mdb_find_oldest(txn);
2000 last = env->me_pglast;
2001 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2004 key.mv_data = &last; /* will look up last+1 */
2005 key.mv_size = sizeof(last);
2007 if (Paranoid && mc->mc_dbi == FREE_DBI)
2010 if (Paranoid && retry < 0 && mop_len)
2014 /* Do not fetch more if the record will be too recent */
2017 rc = mdb_cursor_get(&m2, &key, NULL, op);
2019 if (rc == MDB_NOTFOUND)
2023 last = *(txnid_t*)key.mv_data;
2026 np = m2.mc_pg[m2.mc_top];
2027 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2028 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2031 idl = (MDB_ID *) data.mv_data;
2034 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2039 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2041 mop = env->me_pghead;
2043 env->me_pglast = last;
2045 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2046 last, txn->mt_dbs[FREE_DBI].md_root, i));
2048 DPRINTF(("IDL %"Z"u", idl[j]));
2050 /* Merge in descending sorted order */
2051 mdb_midl_xmerge(mop, idl);
2055 /* Use new pages from the map when nothing suitable in the freeDB */
2057 pgno = txn->mt_next_pgno;
2058 if (pgno + num >= env->me_maxpg) {
2059 DPUTS("DB size maxed out");
2065 if (env->me_flags & MDB_WRITEMAP) {
2066 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2068 if (!(np = mdb_page_malloc(txn, num))) {
2074 mop[0] = mop_len -= num;
2075 /* Move any stragglers down */
2076 for (j = i-num; j < mop_len; )
2077 mop[++j] = mop[++i];
2079 txn->mt_next_pgno = pgno + num;
2082 mdb_page_dirty(txn, np);
2088 txn->mt_flags |= MDB_TXN_ERROR;
2092 /** Copy the used portions of a non-overflow page.
2093 * @param[in] dst page to copy into
2094 * @param[in] src page to copy from
2095 * @param[in] psize size of a page
2098 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2100 enum { Align = sizeof(pgno_t) };
2101 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2103 /* If page isn't full, just copy the used portion. Adjust
2104 * alignment so memcpy may copy words instead of bytes.
2106 if ((unused &= -Align) && !IS_LEAF2(src)) {
2107 upper = (upper + PAGEBASE) & -Align;
2108 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2109 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2112 memcpy(dst, src, psize - unused);
2116 /** Pull a page off the txn's spill list, if present.
2117 * If a page being referenced was spilled to disk in this txn, bring
2118 * it back and make it dirty/writable again.
2119 * @param[in] txn the transaction handle.
2120 * @param[in] mp the page being referenced. It must not be dirty.
2121 * @param[out] ret the writable page, if any. ret is unchanged if
2122 * mp wasn't spilled.
2125 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2127 MDB_env *env = txn->mt_env;
2130 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2132 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2133 if (!tx2->mt_spill_pgs)
2135 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2136 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2139 if (txn->mt_dirty_room == 0)
2140 return MDB_TXN_FULL;
2141 if (IS_OVERFLOW(mp))
2145 if (env->me_flags & MDB_WRITEMAP) {
2148 np = mdb_page_malloc(txn, num);
2152 memcpy(np, mp, num * env->me_psize);
2154 mdb_page_copy(np, mp, env->me_psize);
2157 /* If in current txn, this page is no longer spilled.
2158 * If it happens to be the last page, truncate the spill list.
2159 * Otherwise mark it as deleted by setting the LSB.
2161 if (x == txn->mt_spill_pgs[0])
2162 txn->mt_spill_pgs[0]--;
2164 txn->mt_spill_pgs[x] |= 1;
2165 } /* otherwise, if belonging to a parent txn, the
2166 * page remains spilled until child commits
2169 mdb_page_dirty(txn, np);
2170 np->mp_flags |= P_DIRTY;
2178 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2179 * @param[in] mc cursor pointing to the page to be touched
2180 * @return 0 on success, non-zero on failure.
2183 mdb_page_touch(MDB_cursor *mc)
2185 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2186 MDB_txn *txn = mc->mc_txn;
2187 MDB_cursor *m2, *m3;
2191 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2192 if (txn->mt_flags & MDB_TXN_SPILLS) {
2194 rc = mdb_page_unspill(txn, mp, &np);
2200 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2201 (rc = mdb_page_alloc(mc, 1, &np)))
2204 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2205 mp->mp_pgno, pgno));
2206 mdb_cassert(mc, mp->mp_pgno != pgno);
2207 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2208 /* Update the parent page, if any, to point to the new page */
2210 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2211 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2212 SETPGNO(node, pgno);
2214 mc->mc_db->md_root = pgno;
2216 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2217 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2219 /* If txn has a parent, make sure the page is in our
2223 unsigned x = mdb_mid2l_search(dl, pgno);
2224 if (x <= dl[0].mid && dl[x].mid == pgno) {
2225 if (mp != dl[x].mptr) { /* bad cursor? */
2226 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2227 txn->mt_flags |= MDB_TXN_ERROR;
2228 return MDB_CORRUPTED;
2233 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2235 np = mdb_page_malloc(txn, 1);
2240 rc = mdb_mid2l_insert(dl, &mid);
2241 mdb_cassert(mc, rc == 0);
2246 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2248 np->mp_flags |= P_DIRTY;
2251 /* Adjust cursors pointing to mp */
2252 mc->mc_pg[mc->mc_top] = np;
2253 m2 = txn->mt_cursors[mc->mc_dbi];
2254 if (mc->mc_flags & C_SUB) {
2255 for (; m2; m2=m2->mc_next) {
2256 m3 = &m2->mc_xcursor->mx_cursor;
2257 if (m3->mc_snum < mc->mc_snum) continue;
2258 if (m3->mc_pg[mc->mc_top] == mp)
2259 m3->mc_pg[mc->mc_top] = np;
2262 for (; m2; m2=m2->mc_next) {
2263 if (m2->mc_snum < mc->mc_snum) continue;
2264 if (m2->mc_pg[mc->mc_top] == mp) {
2265 m2->mc_pg[mc->mc_top] = np;
2266 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2268 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2270 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2271 if (!(leaf->mn_flags & F_SUBDATA))
2272 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2280 txn->mt_flags |= MDB_TXN_ERROR;
2285 mdb_env_sync(MDB_env *env, int force)
2288 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2289 if (env->me_flags & MDB_WRITEMAP) {
2290 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2291 ? MS_ASYNC : MS_SYNC;
2292 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2295 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2299 if (MDB_FDATASYNC(env->me_fd))
2306 /** Back up parent txn's cursors, then grab the originals for tracking */
2308 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2310 MDB_cursor *mc, *bk;
2315 for (i = src->mt_numdbs; --i >= 0; ) {
2316 if ((mc = src->mt_cursors[i]) != NULL) {
2317 size = sizeof(MDB_cursor);
2319 size += sizeof(MDB_xcursor);
2320 for (; mc; mc = bk->mc_next) {
2326 mc->mc_db = &dst->mt_dbs[i];
2327 /* Kill pointers into src - and dst to reduce abuse: The
2328 * user may not use mc until dst ends. Otherwise we'd...
2330 mc->mc_txn = NULL; /* ...set this to dst */
2331 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2332 if ((mx = mc->mc_xcursor) != NULL) {
2333 *(MDB_xcursor *)(bk+1) = *mx;
2334 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2336 mc->mc_next = dst->mt_cursors[i];
2337 dst->mt_cursors[i] = mc;
2344 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2345 * @param[in] txn the transaction handle.
2346 * @param[in] merge true to keep changes to parent cursors, false to revert.
2347 * @return 0 on success, non-zero on failure.
2350 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2352 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2356 for (i = txn->mt_numdbs; --i >= 0; ) {
2357 for (mc = cursors[i]; mc; mc = next) {
2359 if ((bk = mc->mc_backup) != NULL) {
2361 /* Commit changes to parent txn */
2362 mc->mc_next = bk->mc_next;
2363 mc->mc_backup = bk->mc_backup;
2364 mc->mc_txn = bk->mc_txn;
2365 mc->mc_db = bk->mc_db;
2366 mc->mc_dbflag = bk->mc_dbflag;
2367 if ((mx = mc->mc_xcursor) != NULL)
2368 mx->mx_cursor.mc_txn = bk->mc_txn;
2370 /* Abort nested txn */
2372 if ((mx = mc->mc_xcursor) != NULL)
2373 *mx = *(MDB_xcursor *)(bk+1);
2377 /* Only malloced cursors are permanently tracked. */
2385 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2388 mdb_txn_reset0(MDB_txn *txn, const char *act);
2390 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2396 Pidset = F_SETLK, Pidcheck = F_GETLK
2400 /** Set or check a pid lock. Set returns 0 on success.
2401 * Check returns 0 if the process is certainly dead, nonzero if it may
2402 * be alive (the lock exists or an error happened so we do not know).
2404 * On Windows Pidset is a no-op, we merely check for the existence
2405 * of the process with the given pid. On POSIX we use a single byte
2406 * lock on the lockfile, set at an offset equal to the pid.
2409 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2411 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2414 if (op == Pidcheck) {
2415 h = OpenProcess(env->me_pidquery, FALSE, pid);
2416 /* No documented "no such process" code, but other program use this: */
2418 return ErrCode() != ERROR_INVALID_PARAMETER;
2419 /* A process exists until all handles to it close. Has it exited? */
2420 ret = WaitForSingleObject(h, 0) != 0;
2427 struct flock lock_info;
2428 memset(&lock_info, 0, sizeof(lock_info));
2429 lock_info.l_type = F_WRLCK;
2430 lock_info.l_whence = SEEK_SET;
2431 lock_info.l_start = pid;
2432 lock_info.l_len = 1;
2433 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2434 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2436 } else if ((rc = ErrCode()) == EINTR) {
2444 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2445 * @param[in] txn the transaction handle to initialize
2446 * @return 0 on success, non-zero on failure.
2449 mdb_txn_renew0(MDB_txn *txn)
2451 MDB_env *env = txn->mt_env;
2452 MDB_txninfo *ti = env->me_txns;
2456 int rc, new_notls = 0;
2459 txn->mt_numdbs = env->me_numdbs;
2460 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2462 if (txn->mt_flags & MDB_TXN_RDONLY) {
2464 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2465 txn->mt_txnid = meta->mm_txnid;
2466 txn->mt_u.reader = NULL;
2468 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2469 pthread_getspecific(env->me_txkey);
2471 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2472 return MDB_BAD_RSLOT;
2474 MDB_PID_T pid = env->me_pid;
2475 MDB_THR_T tid = pthread_self();
2477 if (!env->me_live_reader) {
2478 rc = mdb_reader_pid(env, Pidset, pid);
2481 env->me_live_reader = 1;
2485 nr = ti->mti_numreaders;
2486 for (i=0; i<nr; i++)
2487 if (ti->mti_readers[i].mr_pid == 0)
2489 if (i == env->me_maxreaders) {
2490 UNLOCK_MUTEX_R(env);
2491 return MDB_READERS_FULL;
2493 ti->mti_readers[i].mr_pid = pid;
2494 ti->mti_readers[i].mr_tid = tid;
2496 ti->mti_numreaders = ++nr;
2497 /* Save numreaders for un-mutexed mdb_env_close() */
2498 env->me_numreaders = nr;
2499 UNLOCK_MUTEX_R(env);
2501 r = &ti->mti_readers[i];
2502 new_notls = (env->me_flags & MDB_NOTLS);
2503 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2508 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2509 txn->mt_u.reader = r;
2510 meta = env->me_metas[txn->mt_txnid & 1];
2516 txn->mt_txnid = ti->mti_txnid;
2517 meta = env->me_metas[txn->mt_txnid & 1];
2519 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2520 txn->mt_txnid = meta->mm_txnid;
2524 if (txn->mt_txnid == mdb_debug_start)
2527 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2528 txn->mt_u.dirty_list = env->me_dirty_list;
2529 txn->mt_u.dirty_list[0].mid = 0;
2530 txn->mt_free_pgs = env->me_free_pgs;
2531 txn->mt_free_pgs[0] = 0;
2532 txn->mt_spill_pgs = NULL;
2534 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2537 /* Copy the DB info and flags */
2538 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2540 /* Moved to here to avoid a data race in read TXNs */
2541 txn->mt_next_pgno = meta->mm_last_pg+1;
2543 for (i=2; i<txn->mt_numdbs; i++) {
2544 x = env->me_dbflags[i];
2545 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2546 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2548 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2550 if (env->me_maxpg < txn->mt_next_pgno) {
2551 mdb_txn_reset0(txn, "renew0-mapfail");
2553 txn->mt_u.reader->mr_pid = 0;
2554 txn->mt_u.reader = NULL;
2556 return MDB_MAP_RESIZED;
2563 mdb_txn_renew(MDB_txn *txn)
2567 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2570 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2571 DPUTS("environment had fatal error, must shutdown!");
2575 rc = mdb_txn_renew0(txn);
2576 if (rc == MDB_SUCCESS) {
2577 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2578 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2579 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2585 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2589 int rc, size, tsize = sizeof(MDB_txn);
2591 if (env->me_flags & MDB_FATAL_ERROR) {
2592 DPUTS("environment had fatal error, must shutdown!");
2595 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2598 /* Nested transactions: Max 1 child, write txns only, no writemap */
2599 if (parent->mt_child ||
2600 (flags & MDB_RDONLY) ||
2601 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2602 (env->me_flags & MDB_WRITEMAP))
2604 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2606 tsize = sizeof(MDB_ntxn);
2608 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2609 if (!(flags & MDB_RDONLY)) {
2610 size += env->me_maxdbs * sizeof(MDB_cursor *);
2611 /* child txns use parent's dbiseqs */
2613 size += env->me_maxdbs * sizeof(unsigned int);
2616 if ((txn = calloc(1, size)) == NULL) {
2617 DPRINTF(("calloc: %s", strerror(errno)));
2620 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2621 if (flags & MDB_RDONLY) {
2622 txn->mt_flags |= MDB_TXN_RDONLY;
2623 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2624 txn->mt_dbiseqs = env->me_dbiseqs;
2626 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2628 txn->mt_dbiseqs = parent->mt_dbiseqs;
2629 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2631 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2632 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2639 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2640 if (!txn->mt_u.dirty_list ||
2641 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2643 free(txn->mt_u.dirty_list);
2647 txn->mt_txnid = parent->mt_txnid;
2648 txn->mt_dirty_room = parent->mt_dirty_room;
2649 txn->mt_u.dirty_list[0].mid = 0;
2650 txn->mt_spill_pgs = NULL;
2651 txn->mt_next_pgno = parent->mt_next_pgno;
2652 parent->mt_child = txn;
2653 txn->mt_parent = parent;
2654 txn->mt_numdbs = parent->mt_numdbs;
2655 txn->mt_flags = parent->mt_flags;
2656 txn->mt_dbxs = parent->mt_dbxs;
2657 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2658 /* Copy parent's mt_dbflags, but clear DB_NEW */
2659 for (i=0; i<txn->mt_numdbs; i++)
2660 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2662 ntxn = (MDB_ntxn *)txn;
2663 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2664 if (env->me_pghead) {
2665 size = MDB_IDL_SIZEOF(env->me_pghead);
2666 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2668 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2673 rc = mdb_cursor_shadow(parent, txn);
2675 mdb_txn_reset0(txn, "beginchild-fail");
2677 rc = mdb_txn_renew0(txn);
2683 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2684 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2685 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2692 mdb_txn_env(MDB_txn *txn)
2694 if(!txn) return NULL;
2698 /** Export or close DBI handles opened in this txn. */
2700 mdb_dbis_update(MDB_txn *txn, int keep)
2703 MDB_dbi n = txn->mt_numdbs;
2704 MDB_env *env = txn->mt_env;
2705 unsigned char *tdbflags = txn->mt_dbflags;
2707 for (i = n; --i >= 2;) {
2708 if (tdbflags[i] & DB_NEW) {
2710 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2712 char *ptr = env->me_dbxs[i].md_name.mv_data;
2714 env->me_dbxs[i].md_name.mv_data = NULL;
2715 env->me_dbxs[i].md_name.mv_size = 0;
2716 env->me_dbflags[i] = 0;
2717 env->me_dbiseqs[i]++;
2723 if (keep && env->me_numdbs < n)
2727 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2728 * May be called twice for readonly txns: First reset it, then abort.
2729 * @param[in] txn the transaction handle to reset
2730 * @param[in] act why the transaction is being reset
2733 mdb_txn_reset0(MDB_txn *txn, const char *act)
2735 MDB_env *env = txn->mt_env;
2737 /* Close any DBI handles opened in this txn */
2738 mdb_dbis_update(txn, 0);
2740 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2741 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2742 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2744 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2745 if (txn->mt_u.reader) {
2746 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2747 if (!(env->me_flags & MDB_NOTLS))
2748 txn->mt_u.reader = NULL; /* txn does not own reader */
2750 txn->mt_numdbs = 0; /* close nothing if called again */
2751 txn->mt_dbxs = NULL; /* mark txn as reset */
2753 mdb_cursors_close(txn, 0);
2755 if (!(env->me_flags & MDB_WRITEMAP)) {
2756 mdb_dlist_free(txn);
2758 mdb_midl_free(env->me_pghead);
2760 if (txn->mt_parent) {
2761 txn->mt_parent->mt_child = NULL;
2762 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2763 mdb_midl_free(txn->mt_free_pgs);
2764 mdb_midl_free(txn->mt_spill_pgs);
2765 free(txn->mt_u.dirty_list);
2769 if (mdb_midl_shrink(&txn->mt_free_pgs))
2770 env->me_free_pgs = txn->mt_free_pgs;
2771 env->me_pghead = NULL;
2775 /* The writer mutex was locked in mdb_txn_begin. */
2777 UNLOCK_MUTEX_W(env);
2782 mdb_txn_reset(MDB_txn *txn)
2787 /* This call is only valid for read-only txns */
2788 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2791 mdb_txn_reset0(txn, "reset");
2795 mdb_txn_abort(MDB_txn *txn)
2801 mdb_txn_abort(txn->mt_child);
2803 mdb_txn_reset0(txn, "abort");
2804 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2805 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2806 txn->mt_u.reader->mr_pid = 0;
2811 /** Save the freelist as of this transaction to the freeDB.
2812 * This changes the freelist. Keep trying until it stabilizes.
2815 mdb_freelist_save(MDB_txn *txn)
2817 /* env->me_pghead[] can grow and shrink during this call.
2818 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2819 * Page numbers cannot disappear from txn->mt_free_pgs[].
2822 MDB_env *env = txn->mt_env;
2823 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2824 txnid_t pglast = 0, head_id = 0;
2825 pgno_t freecnt = 0, *free_pgs, *mop;
2826 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2828 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2830 if (env->me_pghead) {
2831 /* Make sure first page of freeDB is touched and on freelist */
2832 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2833 if (rc && rc != MDB_NOTFOUND)
2837 if (!env->me_pghead && txn->mt_loose_pgs) {
2838 /* Put loose page numbers in mt_free_pgs, since
2839 * we may be unable to return them to me_pghead.
2841 MDB_page *mp = txn->mt_loose_pgs;
2842 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2844 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2845 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2846 txn->mt_loose_pgs = NULL;
2847 txn->mt_loose_count = 0;
2850 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2851 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2852 ? SSIZE_MAX : maxfree_1pg;
2855 /* Come back here after each Put() in case freelist changed */
2860 /* If using records from freeDB which we have not yet
2861 * deleted, delete them and any we reserved for me_pghead.
2863 while (pglast < env->me_pglast) {
2864 rc = mdb_cursor_first(&mc, &key, NULL);
2867 pglast = head_id = *(txnid_t *)key.mv_data;
2868 total_room = head_room = 0;
2869 mdb_tassert(txn, pglast <= env->me_pglast);
2870 rc = mdb_cursor_del(&mc, 0);
2875 /* Save the IDL of pages freed by this txn, to a single record */
2876 if (freecnt < txn->mt_free_pgs[0]) {
2878 /* Make sure last page of freeDB is touched and on freelist */
2879 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2880 if (rc && rc != MDB_NOTFOUND)
2883 free_pgs = txn->mt_free_pgs;
2884 /* Write to last page of freeDB */
2885 key.mv_size = sizeof(txn->mt_txnid);
2886 key.mv_data = &txn->mt_txnid;
2888 freecnt = free_pgs[0];
2889 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2890 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2893 /* Retry if mt_free_pgs[] grew during the Put() */
2894 free_pgs = txn->mt_free_pgs;
2895 } while (freecnt < free_pgs[0]);
2896 mdb_midl_sort(free_pgs);
2897 memcpy(data.mv_data, free_pgs, data.mv_size);
2900 unsigned int i = free_pgs[0];
2901 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2902 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2904 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2910 mop = env->me_pghead;
2911 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2913 /* Reserve records for me_pghead[]. Split it if multi-page,
2914 * to avoid searching freeDB for a page range. Use keys in
2915 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2917 if (total_room >= mop_len) {
2918 if (total_room == mop_len || --more < 0)
2920 } else if (head_room >= maxfree_1pg && head_id > 1) {
2921 /* Keep current record (overflow page), add a new one */
2925 /* (Re)write {key = head_id, IDL length = head_room} */
2926 total_room -= head_room;
2927 head_room = mop_len - total_room;
2928 if (head_room > maxfree_1pg && head_id > 1) {
2929 /* Overflow multi-page for part of me_pghead */
2930 head_room /= head_id; /* amortize page sizes */
2931 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2932 } else if (head_room < 0) {
2933 /* Rare case, not bothering to delete this record */
2936 key.mv_size = sizeof(head_id);
2937 key.mv_data = &head_id;
2938 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2939 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2942 /* IDL is initially empty, zero out at least the length */
2943 pgs = (pgno_t *)data.mv_data;
2944 j = head_room > clean_limit ? head_room : 0;
2948 total_room += head_room;
2951 /* Return loose page numbers to me_pghead, though usually none are
2952 * left at this point. The pages themselves remain in dirty_list.
2954 if (txn->mt_loose_pgs) {
2955 MDB_page *mp = txn->mt_loose_pgs;
2956 unsigned count = txn->mt_loose_count;
2958 /* Room for loose pages + temp IDL with same */
2959 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
2961 mop = env->me_pghead;
2962 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
2963 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
2964 loose[ ++count ] = mp->mp_pgno;
2966 mdb_midl_sort(loose);
2967 mdb_midl_xmerge(mop, loose);
2968 txn->mt_loose_pgs = NULL;
2969 txn->mt_loose_count = 0;
2973 /* Fill in the reserved me_pghead records */
2979 rc = mdb_cursor_first(&mc, &key, &data);
2980 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2981 txnid_t id = *(txnid_t *)key.mv_data;
2982 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2985 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2987 if (len > mop_len) {
2989 data.mv_size = (len + 1) * sizeof(MDB_ID);
2991 data.mv_data = mop -= len;
2994 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2996 if (rc || !(mop_len -= len))
3003 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3004 * @param[in] txn the transaction that's being committed
3005 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3006 * @return 0 on success, non-zero on failure.
3009 mdb_page_flush(MDB_txn *txn, int keep)
3011 MDB_env *env = txn->mt_env;
3012 MDB_ID2L dl = txn->mt_u.dirty_list;
3013 unsigned psize = env->me_psize, j;
3014 int i, pagecount = dl[0].mid, rc;
3015 size_t size = 0, pos = 0;
3017 MDB_page *dp = NULL;
3021 struct iovec iov[MDB_COMMIT_PAGES];
3022 ssize_t wpos = 0, wsize = 0, wres;
3023 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3029 if (env->me_flags & MDB_WRITEMAP) {
3030 /* Clear dirty flags */
3031 while (++i <= pagecount) {
3033 /* Don't flush this page yet */
3034 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3035 dp->mp_flags &= ~P_KEEP;
3039 dp->mp_flags &= ~P_DIRTY;
3044 /* Write the pages */
3046 if (++i <= pagecount) {
3048 /* Don't flush this page yet */
3049 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3050 dp->mp_flags &= ~P_KEEP;
3055 /* clear dirty flag */
3056 dp->mp_flags &= ~P_DIRTY;
3059 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3064 /* Windows actually supports scatter/gather I/O, but only on
3065 * unbuffered file handles. Since we're relying on the OS page
3066 * cache for all our data, that's self-defeating. So we just
3067 * write pages one at a time. We use the ov structure to set
3068 * the write offset, to at least save the overhead of a Seek
3071 DPRINTF(("committing page %"Z"u", pgno));
3072 memset(&ov, 0, sizeof(ov));
3073 ov.Offset = pos & 0xffffffff;
3074 ov.OffsetHigh = pos >> 16 >> 16;
3075 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3077 DPRINTF(("WriteFile: %d", rc));
3081 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3082 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3084 /* Write previous page(s) */
3085 #ifdef MDB_USE_PWRITEV
3086 wres = pwritev(env->me_fd, iov, n, wpos);
3089 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3091 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3093 DPRINTF(("lseek: %s", strerror(rc)));
3096 wres = writev(env->me_fd, iov, n);
3099 if (wres != wsize) {
3102 DPRINTF(("Write error: %s", strerror(rc)));
3104 rc = EIO; /* TODO: Use which error code? */
3105 DPUTS("short write, filesystem full?");
3116 DPRINTF(("committing page %"Z"u", pgno));
3117 next_pos = pos + size;
3118 iov[n].iov_len = size;
3119 iov[n].iov_base = (char *)dp;
3125 /* MIPS has cache coherency issues, this is a no-op everywhere else
3126 * Note: for any size >= on-chip cache size, entire on-chip cache is
3129 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3131 for (i = keep; ++i <= pagecount; ) {
3133 /* This is a page we skipped above */
3136 dl[j].mid = dp->mp_pgno;
3139 mdb_dpage_free(env, dp);
3144 txn->mt_dirty_room += i - j;
3150 mdb_txn_commit(MDB_txn *txn)
3156 if (txn == NULL || txn->mt_env == NULL)
3159 if (txn->mt_child) {
3160 rc = mdb_txn_commit(txn->mt_child);
3161 txn->mt_child = NULL;
3168 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3169 mdb_dbis_update(txn, 1);
3170 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3175 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3176 DPUTS("error flag is set, can't commit");
3178 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3183 if (txn->mt_parent) {
3184 MDB_txn *parent = txn->mt_parent;
3188 unsigned x, y, len, ps_len;
3190 /* Append our free list to parent's */
3191 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3194 mdb_midl_free(txn->mt_free_pgs);
3195 /* Failures after this must either undo the changes
3196 * to the parent or set MDB_TXN_ERROR in the parent.
3199 parent->mt_next_pgno = txn->mt_next_pgno;
3200 parent->mt_flags = txn->mt_flags;
3202 /* Merge our cursors into parent's and close them */
3203 mdb_cursors_close(txn, 1);
3205 /* Update parent's DB table. */
3206 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3207 parent->mt_numdbs = txn->mt_numdbs;
3208 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3209 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3210 for (i=2; i<txn->mt_numdbs; i++) {
3211 /* preserve parent's DB_NEW status */
3212 x = parent->mt_dbflags[i] & DB_NEW;
3213 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3216 dst = parent->mt_u.dirty_list;
3217 src = txn->mt_u.dirty_list;
3218 /* Remove anything in our dirty list from parent's spill list */
3219 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3221 pspill[0] = (pgno_t)-1;
3222 /* Mark our dirty pages as deleted in parent spill list */
3223 for (i=0, len=src[0].mid; ++i <= len; ) {
3224 MDB_ID pn = src[i].mid << 1;
3225 while (pn > pspill[x])
3227 if (pn == pspill[x]) {
3232 /* Squash deleted pagenums if we deleted any */
3233 for (x=y; ++x <= ps_len; )
3234 if (!(pspill[x] & 1))
3235 pspill[++y] = pspill[x];
3239 /* Find len = length of merging our dirty list with parent's */
3241 dst[0].mid = 0; /* simplify loops */
3242 if (parent->mt_parent) {
3243 len = x + src[0].mid;
3244 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3245 for (i = x; y && i; y--) {
3246 pgno_t yp = src[y].mid;
3247 while (yp < dst[i].mid)
3249 if (yp == dst[i].mid) {
3254 } else { /* Simplify the above for single-ancestor case */
3255 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3257 /* Merge our dirty list with parent's */
3259 for (i = len; y; dst[i--] = src[y--]) {
3260 pgno_t yp = src[y].mid;
3261 while (yp < dst[x].mid)
3262 dst[i--] = dst[x--];
3263 if (yp == dst[x].mid)
3264 free(dst[x--].mptr);
3266 mdb_tassert(txn, i == x);
3268 free(txn->mt_u.dirty_list);
3269 parent->mt_dirty_room = txn->mt_dirty_room;
3270 if (txn->mt_spill_pgs) {
3271 if (parent->mt_spill_pgs) {
3272 /* TODO: Prevent failure here, so parent does not fail */
3273 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3275 parent->mt_flags |= MDB_TXN_ERROR;
3276 mdb_midl_free(txn->mt_spill_pgs);
3277 mdb_midl_sort(parent->mt_spill_pgs);
3279 parent->mt_spill_pgs = txn->mt_spill_pgs;
3283 /* Append our loose page list to parent's */
3284 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3286 *lp = txn->mt_loose_pgs;
3287 parent->mt_loose_count += txn->mt_loose_count;
3289 parent->mt_child = NULL;
3290 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3295 if (txn != env->me_txn) {
3296 DPUTS("attempt to commit unknown transaction");
3301 mdb_cursors_close(txn, 0);
3303 if (!txn->mt_u.dirty_list[0].mid &&
3304 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3307 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3308 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3310 /* Update DB root pointers */
3311 if (txn->mt_numdbs > 2) {
3315 data.mv_size = sizeof(MDB_db);
3317 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3318 for (i = 2; i < txn->mt_numdbs; i++) {
3319 if (txn->mt_dbflags[i] & DB_DIRTY) {
3320 if (TXN_DBI_CHANGED(txn, i)) {
3324 data.mv_data = &txn->mt_dbs[i];
3325 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3332 rc = mdb_freelist_save(txn);
3336 mdb_midl_free(env->me_pghead);
3337 env->me_pghead = NULL;
3338 if (mdb_midl_shrink(&txn->mt_free_pgs))
3339 env->me_free_pgs = txn->mt_free_pgs;
3345 if ((rc = mdb_page_flush(txn, 0)) ||
3346 (rc = mdb_env_sync(env, 0)) ||
3347 (rc = mdb_env_write_meta(txn)))
3350 /* Free P_LOOSE pages left behind in dirty_list */
3351 if (!(env->me_flags & MDB_WRITEMAP))
3352 mdb_dlist_free(txn);
3357 mdb_dbis_update(txn, 1);
3360 UNLOCK_MUTEX_W(env);
3370 /** Read the environment parameters of a DB environment before
3371 * mapping it into memory.
3372 * @param[in] env the environment handle
3373 * @param[out] meta address of where to store the meta information
3374 * @return 0 on success, non-zero on failure.
3377 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3383 enum { Size = sizeof(pbuf) };
3385 /* We don't know the page size yet, so use a minimum value.
3386 * Read both meta pages so we can use the latest one.
3389 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3393 memset(&ov, 0, sizeof(ov));
3395 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3396 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3399 rc = pread(env->me_fd, &pbuf, Size, off);
3402 if (rc == 0 && off == 0)
3404 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3405 DPRINTF(("read: %s", mdb_strerror(rc)));
3409 p = (MDB_page *)&pbuf;
3411 if (!F_ISSET(p->mp_flags, P_META)) {
3412 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3417 if (m->mm_magic != MDB_MAGIC) {
3418 DPUTS("meta has invalid magic");
3422 if (m->mm_version != MDB_DATA_VERSION) {
3423 DPRINTF(("database is version %u, expected version %u",
3424 m->mm_version, MDB_DATA_VERSION));
3425 return MDB_VERSION_MISMATCH;
3428 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3435 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3437 meta->mm_magic = MDB_MAGIC;
3438 meta->mm_version = MDB_DATA_VERSION;
3439 meta->mm_mapsize = env->me_mapsize;
3440 meta->mm_psize = env->me_psize;
3441 meta->mm_last_pg = 1;
3442 meta->mm_flags = env->me_flags & 0xffff;
3443 meta->mm_flags |= MDB_INTEGERKEY;
3444 meta->mm_dbs[0].md_root = P_INVALID;
3445 meta->mm_dbs[1].md_root = P_INVALID;
3448 /** Write the environment parameters of a freshly created DB environment.
3449 * @param[in] env the environment handle
3450 * @param[out] meta address of where to store the meta information
3451 * @return 0 on success, non-zero on failure.
3454 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3462 memset(&ov, 0, sizeof(ov));
3463 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3465 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3468 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3469 len = pwrite(fd, ptr, size, pos); \
3470 rc = (len >= 0); } while(0)
3473 DPUTS("writing new meta page");
3475 psize = env->me_psize;
3477 mdb_env_init_meta0(env, meta);
3479 p = calloc(2, psize);
3481 p->mp_flags = P_META;
3482 *(MDB_meta *)METADATA(p) = *meta;
3484 q = (MDB_page *)((char *)p + psize);
3486 q->mp_flags = P_META;
3487 *(MDB_meta *)METADATA(q) = *meta;
3489 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3492 else if ((unsigned) len == psize * 2)
3500 /** Update the environment info to commit a transaction.
3501 * @param[in] txn the transaction that's being committed
3502 * @return 0 on success, non-zero on failure.
3505 mdb_env_write_meta(MDB_txn *txn)
3508 MDB_meta meta, metab, *mp;
3511 int rc, len, toggle;
3520 toggle = txn->mt_txnid & 1;
3521 DPRINTF(("writing meta page %d for root page %"Z"u",
3522 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3525 mp = env->me_metas[toggle];
3526 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3527 /* Persist any increases of mapsize config */
3528 if (mapsize < env->me_mapsize)
3529 mapsize = env->me_mapsize;
3531 if (env->me_flags & MDB_WRITEMAP) {
3532 mp->mm_mapsize = mapsize;
3533 mp->mm_dbs[0] = txn->mt_dbs[0];
3534 mp->mm_dbs[1] = txn->mt_dbs[1];
3535 mp->mm_last_pg = txn->mt_next_pgno - 1;
3536 mp->mm_txnid = txn->mt_txnid;
3537 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3538 unsigned meta_size = env->me_psize;
3539 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3542 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3543 if (meta_size < env->me_os_psize)
3544 meta_size += meta_size;
3549 if (MDB_MSYNC(ptr, meta_size, rc)) {
3556 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3557 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3559 meta.mm_mapsize = mapsize;
3560 meta.mm_dbs[0] = txn->mt_dbs[0];
3561 meta.mm_dbs[1] = txn->mt_dbs[1];
3562 meta.mm_last_pg = txn->mt_next_pgno - 1;
3563 meta.mm_txnid = txn->mt_txnid;
3565 off = offsetof(MDB_meta, mm_mapsize);
3566 ptr = (char *)&meta + off;
3567 len = sizeof(MDB_meta) - off;
3569 off += env->me_psize;
3572 /* Write to the SYNC fd */
3573 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3574 env->me_fd : env->me_mfd;
3577 memset(&ov, 0, sizeof(ov));
3579 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3583 rc = pwrite(mfd, ptr, len, off);
3586 rc = rc < 0 ? ErrCode() : EIO;
3587 DPUTS("write failed, disk error?");
3588 /* On a failure, the pagecache still contains the new data.
3589 * Write some old data back, to prevent it from being used.
3590 * Use the non-SYNC fd; we know it will fail anyway.
3592 meta.mm_last_pg = metab.mm_last_pg;
3593 meta.mm_txnid = metab.mm_txnid;
3595 memset(&ov, 0, sizeof(ov));
3597 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3599 r2 = pwrite(env->me_fd, ptr, len, off);
3600 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3603 env->me_flags |= MDB_FATAL_ERROR;
3606 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3607 CACHEFLUSH(env->me_map + off, len, DCACHE);
3609 /* Memory ordering issues are irrelevant; since the entire writer
3610 * is wrapped by wmutex, all of these changes will become visible
3611 * after the wmutex is unlocked. Since the DB is multi-version,
3612 * readers will get consistent data regardless of how fresh or
3613 * how stale their view of these values is.
3616 env->me_txns->mti_txnid = txn->mt_txnid;
3621 /** Check both meta pages to see which one is newer.
3622 * @param[in] env the environment handle
3623 * @return meta toggle (0 or 1).
3626 mdb_env_pick_meta(const MDB_env *env)
3628 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3632 mdb_env_create(MDB_env **env)
3636 e = calloc(1, sizeof(MDB_env));
3640 e->me_maxreaders = DEFAULT_READERS;
3641 e->me_maxdbs = e->me_numdbs = 2;
3642 e->me_fd = INVALID_HANDLE_VALUE;
3643 e->me_lfd = INVALID_HANDLE_VALUE;
3644 e->me_mfd = INVALID_HANDLE_VALUE;
3645 #ifdef MDB_USE_POSIX_SEM
3646 e->me_rmutex = SEM_FAILED;
3647 e->me_wmutex = SEM_FAILED;
3649 e->me_pid = getpid();
3650 GET_PAGESIZE(e->me_os_psize);
3651 VGMEMP_CREATE(e,0,0);
3657 mdb_env_map(MDB_env *env, void *addr)
3660 unsigned int flags = env->me_flags;
3664 LONG sizelo, sizehi;
3667 if (flags & MDB_RDONLY) {
3668 /* Don't set explicit map size, use whatever exists */
3673 msize = env->me_mapsize;
3674 sizelo = msize & 0xffffffff;
3675 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3677 /* Windows won't create mappings for zero length files.
3678 * and won't map more than the file size.
3679 * Just set the maxsize right now.
3681 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3682 || !SetEndOfFile(env->me_fd)
3683 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3687 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3688 PAGE_READWRITE : PAGE_READONLY,
3689 sizehi, sizelo, NULL);
3692 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3693 FILE_MAP_WRITE : FILE_MAP_READ,
3695 rc = env->me_map ? 0 : ErrCode();
3700 int prot = PROT_READ;
3701 if (flags & MDB_WRITEMAP) {
3703 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3706 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3708 if (env->me_map == MAP_FAILED) {
3713 if (flags & MDB_NORDAHEAD) {
3714 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3716 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3718 #ifdef POSIX_MADV_RANDOM
3719 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3720 #endif /* POSIX_MADV_RANDOM */
3721 #endif /* MADV_RANDOM */
3725 /* Can happen because the address argument to mmap() is just a
3726 * hint. mmap() can pick another, e.g. if the range is in use.
3727 * The MAP_FIXED flag would prevent that, but then mmap could
3728 * instead unmap existing pages to make room for the new map.
3730 if (addr && env->me_map != addr)
3731 return EBUSY; /* TODO: Make a new MDB_* error code? */
3733 p = (MDB_page *)env->me_map;
3734 env->me_metas[0] = METADATA(p);
3735 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3741 mdb_env_set_mapsize(MDB_env *env, size_t size)
3743 /* If env is already open, caller is responsible for making
3744 * sure there are no active txns.
3752 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3753 else if (size < env->me_mapsize) {
3754 /* If the configured size is smaller, make sure it's
3755 * still big enough. Silently round up to minimum if not.
3757 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3761 munmap(env->me_map, env->me_mapsize);
3762 env->me_mapsize = size;
3763 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3764 rc = mdb_env_map(env, old);
3768 env->me_mapsize = size;
3770 env->me_maxpg = env->me_mapsize / env->me_psize;
3775 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3779 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3784 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3786 if (env->me_map || readers < 1)
3788 env->me_maxreaders = readers;
3793 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3795 if (!env || !readers)
3797 *readers = env->me_maxreaders;
3801 /** Further setup required for opening an LMDB environment
3804 mdb_env_open2(MDB_env *env)
3806 unsigned int flags = env->me_flags;
3807 int i, newenv = 0, rc;
3811 /* See if we should use QueryLimited */
3813 if ((rc & 0xff) > 5)
3814 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3816 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3819 memset(&meta, 0, sizeof(meta));
3821 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3824 DPUTS("new mdbenv");
3826 env->me_psize = env->me_os_psize;
3827 if (env->me_psize > MAX_PAGESIZE)
3828 env->me_psize = MAX_PAGESIZE;
3830 env->me_psize = meta.mm_psize;
3833 /* Was a mapsize configured? */
3834 if (!env->me_mapsize) {
3835 /* If this is a new environment, take the default,
3836 * else use the size recorded in the existing env.
3838 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3839 } else if (env->me_mapsize < meta.mm_mapsize) {
3840 /* If the configured size is smaller, make sure it's
3841 * still big enough. Silently round up to minimum if not.
3843 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3844 if (env->me_mapsize < minsize)
3845 env->me_mapsize = minsize;
3848 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3853 if (flags & MDB_FIXEDMAP)
3854 meta.mm_address = env->me_map;
3855 i = mdb_env_init_meta(env, &meta);
3856 if (i != MDB_SUCCESS) {
3861 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3862 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3864 #if !(MDB_MAXKEYSIZE)
3865 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3867 env->me_maxpg = env->me_mapsize / env->me_psize;
3871 int toggle = mdb_env_pick_meta(env);
3872 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3874 DPRINTF(("opened database version %u, pagesize %u",
3875 env->me_metas[0]->mm_version, env->me_psize));
3876 DPRINTF(("using meta page %d", toggle));
3877 DPRINTF(("depth: %u", db->md_depth));
3878 DPRINTF(("entries: %"Z"u", db->md_entries));
3879 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3880 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3881 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3882 DPRINTF(("root: %"Z"u", db->md_root));
3890 /** Release a reader thread's slot in the reader lock table.
3891 * This function is called automatically when a thread exits.
3892 * @param[in] ptr This points to the slot in the reader lock table.
3895 mdb_env_reader_dest(void *ptr)
3897 MDB_reader *reader = ptr;
3903 /** Junk for arranging thread-specific callbacks on Windows. This is
3904 * necessarily platform and compiler-specific. Windows supports up
3905 * to 1088 keys. Let's assume nobody opens more than 64 environments
3906 * in a single process, for now. They can override this if needed.
3908 #ifndef MAX_TLS_KEYS
3909 #define MAX_TLS_KEYS 64
3911 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3912 static int mdb_tls_nkeys;
3914 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3918 case DLL_PROCESS_ATTACH: break;
3919 case DLL_THREAD_ATTACH: break;
3920 case DLL_THREAD_DETACH:
3921 for (i=0; i<mdb_tls_nkeys; i++) {
3922 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3924 mdb_env_reader_dest(r);
3928 case DLL_PROCESS_DETACH: break;
3933 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3935 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3939 /* Force some symbol references.
3940 * _tls_used forces the linker to create the TLS directory if not already done
3941 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3943 #pragma comment(linker, "/INCLUDE:_tls_used")
3944 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3945 #pragma const_seg(".CRT$XLB")
3946 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3947 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3950 #pragma comment(linker, "/INCLUDE:__tls_used")
3951 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3952 #pragma data_seg(".CRT$XLB")
3953 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3955 #endif /* WIN 32/64 */
3956 #endif /* !__GNUC__ */
3959 /** Downgrade the exclusive lock on the region back to shared */
3961 mdb_env_share_locks(MDB_env *env, int *excl)
3963 int rc = 0, toggle = mdb_env_pick_meta(env);
3965 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3970 /* First acquire a shared lock. The Unlock will
3971 * then release the existing exclusive lock.
3973 memset(&ov, 0, sizeof(ov));
3974 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3977 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3983 struct flock lock_info;
3984 /* The shared lock replaces the existing lock */
3985 memset((void *)&lock_info, 0, sizeof(lock_info));
3986 lock_info.l_type = F_RDLCK;
3987 lock_info.l_whence = SEEK_SET;
3988 lock_info.l_start = 0;
3989 lock_info.l_len = 1;
3990 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3991 (rc = ErrCode()) == EINTR) ;
3992 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3999 /** Try to get exlusive lock, otherwise shared.
4000 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4003 mdb_env_excl_lock(MDB_env *env, int *excl)
4007 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4011 memset(&ov, 0, sizeof(ov));
4012 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4019 struct flock lock_info;
4020 memset((void *)&lock_info, 0, sizeof(lock_info));
4021 lock_info.l_type = F_WRLCK;
4022 lock_info.l_whence = SEEK_SET;
4023 lock_info.l_start = 0;
4024 lock_info.l_len = 1;
4025 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4026 (rc = ErrCode()) == EINTR) ;
4030 # ifdef MDB_USE_POSIX_SEM
4031 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4034 lock_info.l_type = F_RDLCK;
4035 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4036 (rc = ErrCode()) == EINTR) ;
4046 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4048 * @(#) $Revision: 5.1 $
4049 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4050 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4052 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4056 * Please do not copyright this code. This code is in the public domain.
4058 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4059 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4060 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4061 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4062 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4063 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4064 * PERFORMANCE OF THIS SOFTWARE.
4067 * chongo <Landon Curt Noll> /\oo/\
4068 * http://www.isthe.com/chongo/
4070 * Share and Enjoy! :-)
4073 typedef unsigned long long mdb_hash_t;
4074 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4076 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4077 * @param[in] val value to hash
4078 * @param[in] hval initial value for hash
4079 * @return 64 bit hash
4081 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4082 * hval arg on the first call.
4085 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4087 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4088 unsigned char *end = s + val->mv_size;
4090 * FNV-1a hash each octet of the string
4093 /* xor the bottom with the current octet */
4094 hval ^= (mdb_hash_t)*s++;
4096 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4097 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4098 (hval << 7) + (hval << 8) + (hval << 40);
4100 /* return our new hash value */
4104 /** Hash the string and output the encoded hash.
4105 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4106 * very short name limits. We don't care about the encoding being reversible,
4107 * we just want to preserve as many bits of the input as possible in a
4108 * small printable string.
4109 * @param[in] str string to hash
4110 * @param[out] encbuf an array of 11 chars to hold the hash
4112 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4115 mdb_pack85(unsigned long l, char *out)
4119 for (i=0; i<5; i++) {
4120 *out++ = mdb_a85[l % 85];
4126 mdb_hash_enc(MDB_val *val, char *encbuf)
4128 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4130 mdb_pack85(h, encbuf);
4131 mdb_pack85(h>>32, encbuf+5);
4136 /** Open and/or initialize the lock region for the environment.
4137 * @param[in] env The LMDB environment.
4138 * @param[in] lpath The pathname of the file used for the lock region.
4139 * @param[in] mode The Unix permissions for the file, if we create it.
4140 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4141 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4142 * @return 0 on success, non-zero on failure.
4145 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4148 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4150 # define MDB_ERRCODE_ROFS EROFS
4151 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4152 # define MDB_CLOEXEC O_CLOEXEC
4155 # define MDB_CLOEXEC 0
4162 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4163 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4164 FILE_ATTRIBUTE_NORMAL, NULL);
4166 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4168 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4170 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4175 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4176 /* Lose record locks when exec*() */
4177 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4178 fcntl(env->me_lfd, F_SETFD, fdflags);
4181 if (!(env->me_flags & MDB_NOTLS)) {
4182 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4185 env->me_flags |= MDB_ENV_TXKEY;
4187 /* Windows TLS callbacks need help finding their TLS info. */
4188 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4192 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4196 /* Try to get exclusive lock. If we succeed, then
4197 * nobody is using the lock region and we should initialize it.
4199 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4202 size = GetFileSize(env->me_lfd, NULL);
4204 size = lseek(env->me_lfd, 0, SEEK_END);
4205 if (size == -1) goto fail_errno;
4207 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4208 if (size < rsize && *excl > 0) {
4210 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4211 || !SetEndOfFile(env->me_lfd))
4214 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4218 size = rsize - sizeof(MDB_txninfo);
4219 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4224 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4226 if (!mh) goto fail_errno;
4227 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4229 if (!env->me_txns) goto fail_errno;
4231 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4233 if (m == MAP_FAILED) goto fail_errno;
4239 BY_HANDLE_FILE_INFORMATION stbuf;
4248 if (!mdb_sec_inited) {
4249 InitializeSecurityDescriptor(&mdb_null_sd,
4250 SECURITY_DESCRIPTOR_REVISION);
4251 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4252 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4253 mdb_all_sa.bInheritHandle = FALSE;
4254 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4257 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4258 idbuf.volume = stbuf.dwVolumeSerialNumber;
4259 idbuf.nhigh = stbuf.nFileIndexHigh;
4260 idbuf.nlow = stbuf.nFileIndexLow;
4261 val.mv_data = &idbuf;
4262 val.mv_size = sizeof(idbuf);
4263 mdb_hash_enc(&val, encbuf);
4264 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4265 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4266 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4267 if (!env->me_rmutex) goto fail_errno;
4268 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4269 if (!env->me_wmutex) goto fail_errno;
4270 #elif defined(MDB_USE_POSIX_SEM)
4279 #if defined(__NetBSD__)
4280 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4282 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4283 idbuf.dev = stbuf.st_dev;
4284 idbuf.ino = stbuf.st_ino;
4285 val.mv_data = &idbuf;
4286 val.mv_size = sizeof(idbuf);
4287 mdb_hash_enc(&val, encbuf);
4288 #ifdef MDB_SHORT_SEMNAMES
4289 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4291 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4292 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4293 /* Clean up after a previous run, if needed: Try to
4294 * remove both semaphores before doing anything else.
4296 sem_unlink(env->me_txns->mti_rmname);
4297 sem_unlink(env->me_txns->mti_wmname);
4298 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4299 O_CREAT|O_EXCL, mode, 1);
4300 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4301 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4302 O_CREAT|O_EXCL, mode, 1);
4303 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4304 #else /* MDB_USE_POSIX_SEM */
4305 pthread_mutexattr_t mattr;
4307 if ((rc = pthread_mutexattr_init(&mattr))
4308 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4309 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4310 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4312 pthread_mutexattr_destroy(&mattr);
4313 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4315 env->me_txns->mti_magic = MDB_MAGIC;
4316 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4317 env->me_txns->mti_txnid = 0;
4318 env->me_txns->mti_numreaders = 0;
4321 if (env->me_txns->mti_magic != MDB_MAGIC) {
4322 DPUTS("lock region has invalid magic");
4326 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4327 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4328 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4329 rc = MDB_VERSION_MISMATCH;
4333 if (rc && rc != EACCES && rc != EAGAIN) {
4337 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4338 if (!env->me_rmutex) goto fail_errno;
4339 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4340 if (!env->me_wmutex) goto fail_errno;
4341 #elif defined(MDB_USE_POSIX_SEM)
4342 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4343 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4344 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4345 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4356 /** The name of the lock file in the DB environment */
4357 #define LOCKNAME "/lock.mdb"
4358 /** The name of the data file in the DB environment */
4359 #define DATANAME "/data.mdb"
4360 /** The suffix of the lock file when no subdir is used */
4361 #define LOCKSUFF "-lock"
4362 /** Only a subset of the @ref mdb_env flags can be changed
4363 * at runtime. Changing other flags requires closing the
4364 * environment and re-opening it with the new flags.
4366 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4367 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4368 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4370 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4371 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4375 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4377 int oflags, rc, len, excl = -1;
4378 char *lpath, *dpath;
4380 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4384 if (flags & MDB_NOSUBDIR) {
4385 rc = len + sizeof(LOCKSUFF) + len + 1;
4387 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4392 if (flags & MDB_NOSUBDIR) {
4393 dpath = lpath + len + sizeof(LOCKSUFF);
4394 sprintf(lpath, "%s" LOCKSUFF, path);
4395 strcpy(dpath, path);
4397 dpath = lpath + len + sizeof(LOCKNAME);
4398 sprintf(lpath, "%s" LOCKNAME, path);
4399 sprintf(dpath, "%s" DATANAME, path);
4403 flags |= env->me_flags;
4404 if (flags & MDB_RDONLY) {
4405 /* silently ignore WRITEMAP when we're only getting read access */
4406 flags &= ~MDB_WRITEMAP;
4408 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4409 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4412 env->me_flags = flags |= MDB_ENV_ACTIVE;
4416 env->me_path = strdup(path);
4417 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4418 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4419 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4420 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4425 /* For RDONLY, get lockfile after we know datafile exists */
4426 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4427 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4433 if (F_ISSET(flags, MDB_RDONLY)) {
4434 oflags = GENERIC_READ;
4435 len = OPEN_EXISTING;
4437 oflags = GENERIC_READ|GENERIC_WRITE;
4440 mode = FILE_ATTRIBUTE_NORMAL;
4441 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4442 NULL, len, mode, NULL);
4444 if (F_ISSET(flags, MDB_RDONLY))
4447 oflags = O_RDWR | O_CREAT;
4449 env->me_fd = open(dpath, oflags, mode);
4451 if (env->me_fd == INVALID_HANDLE_VALUE) {
4456 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4457 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4462 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4463 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4464 env->me_mfd = env->me_fd;
4466 /* Synchronous fd for meta writes. Needed even with
4467 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4470 len = OPEN_EXISTING;
4471 env->me_mfd = CreateFile(dpath, oflags,
4472 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4473 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4476 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4478 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4483 DPRINTF(("opened dbenv %p", (void *) env));
4485 rc = mdb_env_share_locks(env, &excl);
4489 if (!((flags & MDB_RDONLY) ||
4490 (env->me_pbuf = calloc(1, env->me_psize))))
4496 mdb_env_close0(env, excl);
4502 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4504 mdb_env_close0(MDB_env *env, int excl)
4508 if (!(env->me_flags & MDB_ENV_ACTIVE))
4511 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4512 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4513 free(env->me_dbxs[i].md_name.mv_data);
4516 free(env->me_dbiseqs);
4517 free(env->me_dbflags);
4520 free(env->me_dirty_list);
4521 mdb_midl_free(env->me_free_pgs);
4523 if (env->me_flags & MDB_ENV_TXKEY) {
4524 pthread_key_delete(env->me_txkey);
4526 /* Delete our key from the global list */
4527 for (i=0; i<mdb_tls_nkeys; i++)
4528 if (mdb_tls_keys[i] == env->me_txkey) {
4529 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4537 munmap(env->me_map, env->me_mapsize);
4539 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4540 (void) close(env->me_mfd);
4541 if (env->me_fd != INVALID_HANDLE_VALUE)
4542 (void) close(env->me_fd);
4544 MDB_PID_T pid = env->me_pid;
4545 /* Clearing readers is done in this function because
4546 * me_txkey with its destructor must be disabled first.
4548 for (i = env->me_numreaders; --i >= 0; )
4549 if (env->me_txns->mti_readers[i].mr_pid == pid)
4550 env->me_txns->mti_readers[i].mr_pid = 0;
4552 if (env->me_rmutex) {
4553 CloseHandle(env->me_rmutex);
4554 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4556 /* Windows automatically destroys the mutexes when
4557 * the last handle closes.
4559 #elif defined(MDB_USE_POSIX_SEM)
4560 if (env->me_rmutex != SEM_FAILED) {
4561 sem_close(env->me_rmutex);
4562 if (env->me_wmutex != SEM_FAILED)
4563 sem_close(env->me_wmutex);
4564 /* If we have the filelock: If we are the
4565 * only remaining user, clean up semaphores.
4568 mdb_env_excl_lock(env, &excl);
4570 sem_unlink(env->me_txns->mti_rmname);
4571 sem_unlink(env->me_txns->mti_wmname);
4575 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4577 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4580 /* Unlock the lockfile. Windows would have unlocked it
4581 * after closing anyway, but not necessarily at once.
4583 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4586 (void) close(env->me_lfd);
4589 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4594 mdb_env_close(MDB_env *env)
4601 VGMEMP_DESTROY(env);
4602 while ((dp = env->me_dpages) != NULL) {
4603 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4604 env->me_dpages = dp->mp_next;
4608 mdb_env_close0(env, 0);
4612 /** Compare two items pointing at aligned size_t's */
4614 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4616 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4617 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4620 /** Compare two items pointing at aligned unsigned int's */
4622 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4624 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4625 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4628 /** Compare two items pointing at unsigned ints of unknown alignment.
4629 * Nodes and keys are guaranteed to be 2-byte aligned.
4632 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4634 #if BYTE_ORDER == LITTLE_ENDIAN
4635 unsigned short *u, *c;
4638 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4639 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4642 } while(!x && u > (unsigned short *)a->mv_data);
4645 unsigned short *u, *c, *end;
4648 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4649 u = (unsigned short *)a->mv_data;
4650 c = (unsigned short *)b->mv_data;
4653 } while(!x && u < end);
4658 /** Compare two items pointing at size_t's of unknown alignment. */
4659 #ifdef MISALIGNED_OK
4660 # define mdb_cmp_clong mdb_cmp_long
4662 # define mdb_cmp_clong mdb_cmp_cint
4665 /** Compare two items lexically */
4667 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4674 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4680 diff = memcmp(a->mv_data, b->mv_data, len);
4681 return diff ? diff : len_diff<0 ? -1 : len_diff;
4684 /** Compare two items in reverse byte order */
4686 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4688 const unsigned char *p1, *p2, *p1_lim;
4692 p1_lim = (const unsigned char *)a->mv_data;
4693 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4694 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4696 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4702 while (p1 > p1_lim) {
4703 diff = *--p1 - *--p2;
4707 return len_diff<0 ? -1 : len_diff;
4710 /** Search for key within a page, using binary search.
4711 * Returns the smallest entry larger or equal to the key.
4712 * If exactp is non-null, stores whether the found entry was an exact match
4713 * in *exactp (1 or 0).
4714 * Updates the cursor index with the index of the found entry.
4715 * If no entry larger or equal to the key is found, returns NULL.
4718 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4720 unsigned int i = 0, nkeys;
4723 MDB_page *mp = mc->mc_pg[mc->mc_top];
4724 MDB_node *node = NULL;
4729 nkeys = NUMKEYS(mp);
4731 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4732 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4735 low = IS_LEAF(mp) ? 0 : 1;
4737 cmp = mc->mc_dbx->md_cmp;
4739 /* Branch pages have no data, so if using integer keys,
4740 * alignment is guaranteed. Use faster mdb_cmp_int.
4742 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4743 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4750 nodekey.mv_size = mc->mc_db->md_pad;
4751 node = NODEPTR(mp, 0); /* fake */
4752 while (low <= high) {
4753 i = (low + high) >> 1;
4754 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4755 rc = cmp(key, &nodekey);
4756 DPRINTF(("found leaf index %u [%s], rc = %i",
4757 i, DKEY(&nodekey), rc));
4766 while (low <= high) {
4767 i = (low + high) >> 1;
4769 node = NODEPTR(mp, i);
4770 nodekey.mv_size = NODEKSZ(node);
4771 nodekey.mv_data = NODEKEY(node);
4773 rc = cmp(key, &nodekey);
4776 DPRINTF(("found leaf index %u [%s], rc = %i",
4777 i, DKEY(&nodekey), rc));
4779 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4780 i, DKEY(&nodekey), NODEPGNO(node), rc));
4791 if (rc > 0) { /* Found entry is less than the key. */
4792 i++; /* Skip to get the smallest entry larger than key. */
4794 node = NODEPTR(mp, i);
4797 *exactp = (rc == 0 && nkeys > 0);
4798 /* store the key index */
4799 mc->mc_ki[mc->mc_top] = i;
4801 /* There is no entry larger or equal to the key. */
4804 /* nodeptr is fake for LEAF2 */
4810 mdb_cursor_adjust(MDB_cursor *mc, func)
4814 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4815 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4822 /** Pop a page off the top of the cursor's stack. */
4824 mdb_cursor_pop(MDB_cursor *mc)
4828 MDB_page *top = mc->mc_pg[mc->mc_top];
4834 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4835 DDBI(mc), (void *) mc));
4839 /** Push a page onto the top of the cursor's stack. */
4841 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4843 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4844 DDBI(mc), (void *) mc));
4846 if (mc->mc_snum >= CURSOR_STACK) {
4847 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4848 return MDB_CURSOR_FULL;
4851 mc->mc_top = mc->mc_snum++;
4852 mc->mc_pg[mc->mc_top] = mp;
4853 mc->mc_ki[mc->mc_top] = 0;
4858 /** Find the address of the page corresponding to a given page number.
4859 * @param[in] txn the transaction for this access.
4860 * @param[in] pgno the page number for the page to retrieve.
4861 * @param[out] ret address of a pointer where the page's address will be stored.
4862 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4863 * @return 0 on success, non-zero on failure.
4866 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4868 MDB_env *env = txn->mt_env;
4872 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4876 MDB_ID2L dl = tx2->mt_u.dirty_list;
4878 /* Spilled pages were dirtied in this txn and flushed
4879 * because the dirty list got full. Bring this page
4880 * back in from the map (but don't unspill it here,
4881 * leave that unless page_touch happens again).
4883 if (tx2->mt_spill_pgs) {
4884 MDB_ID pn = pgno << 1;
4885 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4886 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4887 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4892 unsigned x = mdb_mid2l_search(dl, pgno);
4893 if (x <= dl[0].mid && dl[x].mid == pgno) {
4899 } while ((tx2 = tx2->mt_parent) != NULL);
4902 if (pgno < txn->mt_next_pgno) {
4904 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4906 DPRINTF(("page %"Z"u not found", pgno));
4907 txn->mt_flags |= MDB_TXN_ERROR;
4908 return MDB_PAGE_NOTFOUND;
4918 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4919 * The cursor is at the root page, set up the rest of it.
4922 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4924 MDB_page *mp = mc->mc_pg[mc->mc_top];
4928 while (IS_BRANCH(mp)) {
4932 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4933 mdb_cassert(mc, NUMKEYS(mp) > 1);
4934 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4936 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4938 if (flags & MDB_PS_LAST)
4939 i = NUMKEYS(mp) - 1;
4942 node = mdb_node_search(mc, key, &exact);
4944 i = NUMKEYS(mp) - 1;
4946 i = mc->mc_ki[mc->mc_top];
4948 mdb_cassert(mc, i > 0);
4952 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4955 mdb_cassert(mc, i < NUMKEYS(mp));
4956 node = NODEPTR(mp, i);
4958 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4961 mc->mc_ki[mc->mc_top] = i;
4962 if ((rc = mdb_cursor_push(mc, mp)))
4965 if (flags & MDB_PS_MODIFY) {
4966 if ((rc = mdb_page_touch(mc)) != 0)
4968 mp = mc->mc_pg[mc->mc_top];
4973 DPRINTF(("internal error, index points to a %02X page!?",
4975 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4976 return MDB_CORRUPTED;
4979 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4980 key ? DKEY(key) : "null"));
4981 mc->mc_flags |= C_INITIALIZED;
4982 mc->mc_flags &= ~C_EOF;
4987 /** Search for the lowest key under the current branch page.
4988 * This just bypasses a NUMKEYS check in the current page
4989 * before calling mdb_page_search_root(), because the callers
4990 * are all in situations where the current page is known to
4994 mdb_page_search_lowest(MDB_cursor *mc)
4996 MDB_page *mp = mc->mc_pg[mc->mc_top];
4997 MDB_node *node = NODEPTR(mp, 0);
5000 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5003 mc->mc_ki[mc->mc_top] = 0;
5004 if ((rc = mdb_cursor_push(mc, mp)))
5006 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5009 /** Search for the page a given key should be in.
5010 * Push it and its parent pages on the cursor stack.
5011 * @param[in,out] mc the cursor for this operation.
5012 * @param[in] key the key to search for, or NULL for first/last page.
5013 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5014 * are touched (updated with new page numbers).
5015 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5016 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5017 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5018 * @return 0 on success, non-zero on failure.
5021 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5026 /* Make sure the txn is still viable, then find the root from
5027 * the txn's db table and set it as the root of the cursor's stack.
5029 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5030 DPUTS("transaction has failed, must abort");
5033 /* Make sure we're using an up-to-date root */
5034 if (*mc->mc_dbflag & DB_STALE) {
5036 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5038 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5039 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5046 MDB_node *leaf = mdb_node_search(&mc2,
5047 &mc->mc_dbx->md_name, &exact);
5049 return MDB_NOTFOUND;
5050 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5053 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5055 /* The txn may not know this DBI, or another process may
5056 * have dropped and recreated the DB with other flags.
5058 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5059 return MDB_INCOMPATIBLE;
5060 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5062 *mc->mc_dbflag &= ~DB_STALE;
5064 root = mc->mc_db->md_root;
5066 if (root == P_INVALID) { /* Tree is empty. */
5067 DPUTS("tree is empty");
5068 return MDB_NOTFOUND;
5072 mdb_cassert(mc, root > 1);
5073 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5074 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5080 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5081 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5083 if (flags & MDB_PS_MODIFY) {
5084 if ((rc = mdb_page_touch(mc)))
5088 if (flags & MDB_PS_ROOTONLY)
5091 return mdb_page_search_root(mc, key, flags);
5095 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5097 MDB_txn *txn = mc->mc_txn;
5098 pgno_t pg = mp->mp_pgno;
5099 unsigned x = 0, ovpages = mp->mp_pages;
5100 MDB_env *env = txn->mt_env;
5101 MDB_IDL sl = txn->mt_spill_pgs;
5102 MDB_ID pn = pg << 1;
5105 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5106 /* If the page is dirty or on the spill list we just acquired it,
5107 * so we should give it back to our current free list, if any.
5108 * Otherwise put it onto the list of pages we freed in this txn.
5110 * Won't create me_pghead: me_pglast must be inited along with it.
5111 * Unsupported in nested txns: They would need to hide the page
5112 * range in ancestor txns' dirty and spilled lists.
5114 if (env->me_pghead &&
5116 ((mp->mp_flags & P_DIRTY) ||
5117 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5121 MDB_ID2 *dl, ix, iy;
5122 rc = mdb_midl_need(&env->me_pghead, ovpages);
5125 if (!(mp->mp_flags & P_DIRTY)) {
5126 /* This page is no longer spilled */
5133 /* Remove from dirty list */
5134 dl = txn->mt_u.dirty_list;
5136 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5142 mdb_cassert(mc, x > 1);
5144 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5145 txn->mt_flags |= MDB_TXN_ERROR;
5146 return MDB_CORRUPTED;
5149 if (!(env->me_flags & MDB_WRITEMAP))
5150 mdb_dpage_free(env, mp);
5152 /* Insert in me_pghead */
5153 mop = env->me_pghead;
5154 j = mop[0] + ovpages;
5155 for (i = mop[0]; i && mop[i] < pg; i--)
5161 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5165 mc->mc_db->md_overflow_pages -= ovpages;
5169 /** Return the data associated with a given node.
5170 * @param[in] txn The transaction for this operation.
5171 * @param[in] leaf The node being read.
5172 * @param[out] data Updated to point to the node's data.
5173 * @return 0 on success, non-zero on failure.
5176 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5178 MDB_page *omp; /* overflow page */
5182 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5183 data->mv_size = NODEDSZ(leaf);
5184 data->mv_data = NODEDATA(leaf);
5188 /* Read overflow data.
5190 data->mv_size = NODEDSZ(leaf);
5191 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5192 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5193 DPRINTF(("read overflow page %"Z"u failed", pgno));
5196 data->mv_data = METADATA(omp);
5202 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5203 MDB_val *key, MDB_val *data)
5210 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5212 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5215 if (txn->mt_flags & MDB_TXN_ERROR)
5218 mdb_cursor_init(&mc, txn, dbi, &mx);
5219 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5222 /** Find a sibling for a page.
5223 * Replaces the page at the top of the cursor's stack with the
5224 * specified sibling, if one exists.
5225 * @param[in] mc The cursor for this operation.
5226 * @param[in] move_right Non-zero if the right sibling is requested,
5227 * otherwise the left sibling.
5228 * @return 0 on success, non-zero on failure.
5231 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5237 if (mc->mc_snum < 2) {
5238 return MDB_NOTFOUND; /* root has no siblings */
5242 DPRINTF(("parent page is page %"Z"u, index %u",
5243 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5245 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5246 : (mc->mc_ki[mc->mc_top] == 0)) {
5247 DPRINTF(("no more keys left, moving to %s sibling",
5248 move_right ? "right" : "left"));
5249 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5250 /* undo cursor_pop before returning */
5257 mc->mc_ki[mc->mc_top]++;
5259 mc->mc_ki[mc->mc_top]--;
5260 DPRINTF(("just moving to %s index key %u",
5261 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5263 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5265 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5266 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5267 /* mc will be inconsistent if caller does mc_snum++ as above */
5268 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5272 mdb_cursor_push(mc, mp);
5274 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5279 /** Move the cursor to the next data item. */
5281 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5287 if (mc->mc_flags & C_EOF) {
5288 return MDB_NOTFOUND;
5291 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5293 mp = mc->mc_pg[mc->mc_top];
5295 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5296 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5297 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5298 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5299 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5300 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5301 if (rc == MDB_SUCCESS)
5302 MDB_GET_KEY(leaf, key);
5307 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5308 if (op == MDB_NEXT_DUP)
5309 return MDB_NOTFOUND;
5313 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5314 mdb_dbg_pgno(mp), (void *) mc));
5315 if (mc->mc_flags & C_DEL)
5318 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5319 DPUTS("=====> move to next sibling page");
5320 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5321 mc->mc_flags |= C_EOF;
5324 mp = mc->mc_pg[mc->mc_top];
5325 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5327 mc->mc_ki[mc->mc_top]++;
5330 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5331 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5334 key->mv_size = mc->mc_db->md_pad;
5335 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5339 mdb_cassert(mc, IS_LEAF(mp));
5340 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5342 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5343 mdb_xcursor_init1(mc, leaf);
5346 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5349 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5350 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5351 if (rc != MDB_SUCCESS)
5356 MDB_GET_KEY(leaf, key);
5360 /** Move the cursor to the previous data item. */
5362 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5368 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5370 mp = mc->mc_pg[mc->mc_top];
5372 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5373 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5374 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5375 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5376 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5377 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5378 if (rc == MDB_SUCCESS) {
5379 MDB_GET_KEY(leaf, key);
5380 mc->mc_flags &= ~C_EOF;
5385 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5386 if (op == MDB_PREV_DUP)
5387 return MDB_NOTFOUND;
5392 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5393 mdb_dbg_pgno(mp), (void *) mc));
5395 if (mc->mc_ki[mc->mc_top] == 0) {
5396 DPUTS("=====> move to prev sibling page");
5397 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5400 mp = mc->mc_pg[mc->mc_top];
5401 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5402 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5404 mc->mc_ki[mc->mc_top]--;
5406 mc->mc_flags &= ~C_EOF;
5408 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5409 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5412 key->mv_size = mc->mc_db->md_pad;
5413 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5417 mdb_cassert(mc, IS_LEAF(mp));
5418 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5420 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5421 mdb_xcursor_init1(mc, leaf);
5424 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5427 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5428 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5429 if (rc != MDB_SUCCESS)
5434 MDB_GET_KEY(leaf, key);
5438 /** Set the cursor on a specific data item. */
5440 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5441 MDB_cursor_op op, int *exactp)
5445 MDB_node *leaf = NULL;
5448 if (key->mv_size == 0)
5449 return MDB_BAD_VALSIZE;
5452 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5454 /* See if we're already on the right page */
5455 if (mc->mc_flags & C_INITIALIZED) {
5458 mp = mc->mc_pg[mc->mc_top];
5460 mc->mc_ki[mc->mc_top] = 0;
5461 return MDB_NOTFOUND;
5463 if (mp->mp_flags & P_LEAF2) {
5464 nodekey.mv_size = mc->mc_db->md_pad;
5465 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5467 leaf = NODEPTR(mp, 0);
5468 MDB_GET_KEY2(leaf, nodekey);
5470 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5472 /* Probably happens rarely, but first node on the page
5473 * was the one we wanted.
5475 mc->mc_ki[mc->mc_top] = 0;
5482 unsigned int nkeys = NUMKEYS(mp);
5484 if (mp->mp_flags & P_LEAF2) {
5485 nodekey.mv_data = LEAF2KEY(mp,
5486 nkeys-1, nodekey.mv_size);
5488 leaf = NODEPTR(mp, nkeys-1);
5489 MDB_GET_KEY2(leaf, nodekey);
5491 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5493 /* last node was the one we wanted */
5494 mc->mc_ki[mc->mc_top] = nkeys-1;
5500 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5501 /* This is definitely the right page, skip search_page */
5502 if (mp->mp_flags & P_LEAF2) {
5503 nodekey.mv_data = LEAF2KEY(mp,
5504 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5506 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5507 MDB_GET_KEY2(leaf, nodekey);
5509 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5511 /* current node was the one we wanted */
5521 /* If any parents have right-sibs, search.
5522 * Otherwise, there's nothing further.
5524 for (i=0; i<mc->mc_top; i++)
5526 NUMKEYS(mc->mc_pg[i])-1)
5528 if (i == mc->mc_top) {
5529 /* There are no other pages */
5530 mc->mc_ki[mc->mc_top] = nkeys;
5531 return MDB_NOTFOUND;
5535 /* There are no other pages */
5536 mc->mc_ki[mc->mc_top] = 0;
5537 if (op == MDB_SET_RANGE && !exactp) {
5541 return MDB_NOTFOUND;
5545 rc = mdb_page_search(mc, key, 0);
5546 if (rc != MDB_SUCCESS)
5549 mp = mc->mc_pg[mc->mc_top];
5550 mdb_cassert(mc, IS_LEAF(mp));
5553 leaf = mdb_node_search(mc, key, exactp);
5554 if (exactp != NULL && !*exactp) {
5555 /* MDB_SET specified and not an exact match. */
5556 return MDB_NOTFOUND;
5560 DPUTS("===> inexact leaf not found, goto sibling");
5561 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5562 return rc; /* no entries matched */
5563 mp = mc->mc_pg[mc->mc_top];
5564 mdb_cassert(mc, IS_LEAF(mp));
5565 leaf = NODEPTR(mp, 0);
5569 mc->mc_flags |= C_INITIALIZED;
5570 mc->mc_flags &= ~C_EOF;
5573 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5574 key->mv_size = mc->mc_db->md_pad;
5575 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5580 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5581 mdb_xcursor_init1(mc, leaf);
5584 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5585 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5586 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5589 if (op == MDB_GET_BOTH) {
5595 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5596 if (rc != MDB_SUCCESS)
5599 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5601 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5603 rc = mc->mc_dbx->md_dcmp(data, &d2);
5605 if (op == MDB_GET_BOTH || rc > 0)
5606 return MDB_NOTFOUND;
5613 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5614 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5619 /* The key already matches in all other cases */
5620 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5621 MDB_GET_KEY(leaf, key);
5622 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5627 /** Move the cursor to the first item in the database. */
5629 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5635 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5637 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5638 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5639 if (rc != MDB_SUCCESS)
5642 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5644 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5645 mc->mc_flags |= C_INITIALIZED;
5646 mc->mc_flags &= ~C_EOF;
5648 mc->mc_ki[mc->mc_top] = 0;
5650 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5651 key->mv_size = mc->mc_db->md_pad;
5652 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5657 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5658 mdb_xcursor_init1(mc, leaf);
5659 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5663 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5667 MDB_GET_KEY(leaf, key);
5671 /** Move the cursor to the last item in the database. */
5673 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5679 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5681 if (!(mc->mc_flags & C_EOF)) {
5683 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5684 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5685 if (rc != MDB_SUCCESS)
5688 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5691 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5692 mc->mc_flags |= C_INITIALIZED|C_EOF;
5693 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5695 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5696 key->mv_size = mc->mc_db->md_pad;
5697 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5702 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5703 mdb_xcursor_init1(mc, leaf);
5704 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5708 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5713 MDB_GET_KEY(leaf, key);
5718 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5723 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5728 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5732 case MDB_GET_CURRENT:
5733 if (!(mc->mc_flags & C_INITIALIZED)) {
5736 MDB_page *mp = mc->mc_pg[mc->mc_top];
5737 int nkeys = NUMKEYS(mp);
5738 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5739 mc->mc_ki[mc->mc_top] = nkeys;
5745 key->mv_size = mc->mc_db->md_pad;
5746 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5748 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5749 MDB_GET_KEY(leaf, key);
5751 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5752 if (mc->mc_flags & C_DEL)
5753 mdb_xcursor_init1(mc, leaf);
5754 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5756 rc = mdb_node_read(mc->mc_txn, leaf, data);
5763 case MDB_GET_BOTH_RANGE:
5768 if (mc->mc_xcursor == NULL) {
5769 rc = MDB_INCOMPATIBLE;
5779 rc = mdb_cursor_set(mc, key, data, op,
5780 op == MDB_SET_RANGE ? NULL : &exact);
5783 case MDB_GET_MULTIPLE:
5784 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5788 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5789 rc = MDB_INCOMPATIBLE;
5793 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5794 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5797 case MDB_NEXT_MULTIPLE:
5802 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5803 rc = MDB_INCOMPATIBLE;
5806 if (!(mc->mc_flags & C_INITIALIZED))
5807 rc = mdb_cursor_first(mc, key, data);
5809 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5810 if (rc == MDB_SUCCESS) {
5811 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5814 mx = &mc->mc_xcursor->mx_cursor;
5815 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5817 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5818 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5826 case MDB_NEXT_NODUP:
5827 if (!(mc->mc_flags & C_INITIALIZED))
5828 rc = mdb_cursor_first(mc, key, data);
5830 rc = mdb_cursor_next(mc, key, data, op);
5834 case MDB_PREV_NODUP:
5835 if (!(mc->mc_flags & C_INITIALIZED)) {
5836 rc = mdb_cursor_last(mc, key, data);
5839 mc->mc_flags |= C_INITIALIZED;
5840 mc->mc_ki[mc->mc_top]++;
5842 rc = mdb_cursor_prev(mc, key, data, op);
5845 rc = mdb_cursor_first(mc, key, data);
5848 mfunc = mdb_cursor_first;
5850 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5854 if (mc->mc_xcursor == NULL) {
5855 rc = MDB_INCOMPATIBLE;
5859 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5860 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5861 MDB_GET_KEY(leaf, key);
5862 rc = mdb_node_read(mc->mc_txn, leaf, data);
5866 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5870 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5873 rc = mdb_cursor_last(mc, key, data);
5876 mfunc = mdb_cursor_last;
5879 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5884 if (mc->mc_flags & C_DEL)
5885 mc->mc_flags ^= C_DEL;
5890 /** Touch all the pages in the cursor stack. Set mc_top.
5891 * Makes sure all the pages are writable, before attempting a write operation.
5892 * @param[in] mc The cursor to operate on.
5895 mdb_cursor_touch(MDB_cursor *mc)
5897 int rc = MDB_SUCCESS;
5899 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5902 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5904 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5905 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5908 *mc->mc_dbflag |= DB_DIRTY;
5913 rc = mdb_page_touch(mc);
5914 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5915 mc->mc_top = mc->mc_snum-1;
5920 /** Do not spill pages to disk if txn is getting full, may fail instead */
5921 #define MDB_NOSPILL 0x8000
5924 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5927 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5929 MDB_node *leaf = NULL;
5932 MDB_val xdata, *rdata, dkey, olddata;
5934 int do_sub = 0, insert_key, insert_data;
5935 unsigned int mcount = 0, dcount = 0, nospill;
5938 unsigned int nflags;
5941 if (mc == NULL || key == NULL)
5944 env = mc->mc_txn->mt_env;
5946 /* Check this first so counter will always be zero on any
5949 if (flags & MDB_MULTIPLE) {
5950 dcount = data[1].mv_size;
5951 data[1].mv_size = 0;
5952 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5953 return MDB_INCOMPATIBLE;
5956 nospill = flags & MDB_NOSPILL;
5957 flags &= ~MDB_NOSPILL;
5959 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5960 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5962 if (key->mv_size-1 >= ENV_MAXKEY(env))
5963 return MDB_BAD_VALSIZE;
5965 #if SIZE_MAX > MAXDATASIZE
5966 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5967 return MDB_BAD_VALSIZE;
5969 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5970 return MDB_BAD_VALSIZE;
5973 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5974 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5978 if (flags == MDB_CURRENT) {
5979 if (!(mc->mc_flags & C_INITIALIZED))
5982 } else if (mc->mc_db->md_root == P_INVALID) {
5983 /* new database, cursor has nothing to point to */
5986 mc->mc_flags &= ~C_INITIALIZED;
5991 if (flags & MDB_APPEND) {
5993 rc = mdb_cursor_last(mc, &k2, &d2);
5995 rc = mc->mc_dbx->md_cmp(key, &k2);
5998 mc->mc_ki[mc->mc_top]++;
6000 /* new key is <= last key */
6005 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6007 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6008 DPRINTF(("duplicate key [%s]", DKEY(key)));
6010 return MDB_KEYEXIST;
6012 if (rc && rc != MDB_NOTFOUND)
6016 if (mc->mc_flags & C_DEL)
6017 mc->mc_flags ^= C_DEL;
6019 /* Cursor is positioned, check for room in the dirty list */
6021 if (flags & MDB_MULTIPLE) {
6023 xdata.mv_size = data->mv_size * dcount;
6027 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6031 if (rc == MDB_NO_ROOT) {
6033 /* new database, write a root leaf page */
6034 DPUTS("allocating new root leaf page");
6035 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6038 mdb_cursor_push(mc, np);
6039 mc->mc_db->md_root = np->mp_pgno;
6040 mc->mc_db->md_depth++;
6041 *mc->mc_dbflag |= DB_DIRTY;
6042 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6044 np->mp_flags |= P_LEAF2;
6045 mc->mc_flags |= C_INITIALIZED;
6047 /* make sure all cursor pages are writable */
6048 rc2 = mdb_cursor_touch(mc);
6053 insert_key = insert_data = rc;
6055 /* The key does not exist */
6056 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6057 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6058 LEAFSIZE(key, data) > env->me_nodemax)
6060 /* Too big for a node, insert in sub-DB. Set up an empty
6061 * "old sub-page" for prep_subDB to expand to a full page.
6063 fp_flags = P_LEAF|P_DIRTY;
6065 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6066 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6067 olddata.mv_size = PAGEHDRSZ;
6071 /* there's only a key anyway, so this is a no-op */
6072 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6074 unsigned int ksize = mc->mc_db->md_pad;
6075 if (key->mv_size != ksize)
6076 return MDB_BAD_VALSIZE;
6077 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6078 memcpy(ptr, key->mv_data, ksize);
6080 /* if overwriting slot 0 of leaf, need to
6081 * update branch key if there is a parent page
6083 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6084 unsigned short top = mc->mc_top;
6086 /* slot 0 is always an empty key, find real slot */
6087 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6089 if (mc->mc_ki[mc->mc_top])
6090 rc2 = mdb_update_key(mc, key);
6101 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6102 olddata.mv_size = NODEDSZ(leaf);
6103 olddata.mv_data = NODEDATA(leaf);
6106 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6107 /* Prepare (sub-)page/sub-DB to accept the new item,
6108 * if needed. fp: old sub-page or a header faking
6109 * it. mp: new (sub-)page. offset: growth in page
6110 * size. xdata: node data with new page or DB.
6112 unsigned i, offset = 0;
6113 mp = fp = xdata.mv_data = env->me_pbuf;
6114 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6116 /* Was a single item before, must convert now */
6117 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6118 /* Just overwrite the current item */
6119 if (flags == MDB_CURRENT)
6122 #if UINT_MAX < SIZE_MAX
6123 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6124 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6126 /* does data match? */
6127 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6128 if (flags & MDB_NODUPDATA)
6129 return MDB_KEYEXIST;
6134 /* Back up original data item */
6135 dkey.mv_size = olddata.mv_size;
6136 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6138 /* Make sub-page header for the dup items, with dummy body */
6139 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6140 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6141 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6142 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6143 fp->mp_flags |= P_LEAF2;
6144 fp->mp_pad = data->mv_size;
6145 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6147 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6148 (dkey.mv_size & 1) + (data->mv_size & 1);
6150 fp->mp_upper = xdata.mv_size - PAGEBASE;
6151 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6152 } else if (leaf->mn_flags & F_SUBDATA) {
6153 /* Data is on sub-DB, just store it */
6154 flags |= F_DUPDATA|F_SUBDATA;
6157 /* Data is on sub-page */
6158 fp = olddata.mv_data;
6161 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6162 offset = EVEN(NODESIZE + sizeof(indx_t) +
6166 offset = fp->mp_pad;
6167 if (SIZELEFT(fp) < offset) {
6168 offset *= 4; /* space for 4 more */
6171 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6173 fp->mp_flags |= P_DIRTY;
6174 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6175 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6179 xdata.mv_size = olddata.mv_size + offset;
6182 fp_flags = fp->mp_flags;
6183 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6184 /* Too big for a sub-page, convert to sub-DB */
6185 fp_flags &= ~P_SUBP;
6187 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6188 fp_flags |= P_LEAF2;
6189 dummy.md_pad = fp->mp_pad;
6190 dummy.md_flags = MDB_DUPFIXED;
6191 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6192 dummy.md_flags |= MDB_INTEGERKEY;
6198 dummy.md_branch_pages = 0;
6199 dummy.md_leaf_pages = 1;
6200 dummy.md_overflow_pages = 0;
6201 dummy.md_entries = NUMKEYS(fp);
6202 xdata.mv_size = sizeof(MDB_db);
6203 xdata.mv_data = &dummy;
6204 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6206 offset = env->me_psize - olddata.mv_size;
6207 flags |= F_DUPDATA|F_SUBDATA;
6208 dummy.md_root = mp->mp_pgno;
6211 mp->mp_flags = fp_flags | P_DIRTY;
6212 mp->mp_pad = fp->mp_pad;
6213 mp->mp_lower = fp->mp_lower;
6214 mp->mp_upper = fp->mp_upper + offset;
6215 if (fp_flags & P_LEAF2) {
6216 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6218 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6219 olddata.mv_size - fp->mp_upper - PAGEBASE);
6220 for (i=0; i<NUMKEYS(fp); i++)
6221 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6229 mdb_node_del(mc, 0);
6233 /* overflow page overwrites need special handling */
6234 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6237 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6239 memcpy(&pg, olddata.mv_data, sizeof(pg));
6240 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6242 ovpages = omp->mp_pages;
6244 /* Is the ov page large enough? */
6245 if (ovpages >= dpages) {
6246 if (!(omp->mp_flags & P_DIRTY) &&
6247 (level || (env->me_flags & MDB_WRITEMAP)))
6249 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6252 level = 0; /* dirty in this txn or clean */
6255 if (omp->mp_flags & P_DIRTY) {
6256 /* yes, overwrite it. Note in this case we don't
6257 * bother to try shrinking the page if the new data
6258 * is smaller than the overflow threshold.
6261 /* It is writable only in a parent txn */
6262 size_t sz = (size_t) env->me_psize * ovpages, off;
6263 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6269 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6270 mdb_cassert(mc, rc2 == 0);
6271 if (!(flags & MDB_RESERVE)) {
6272 /* Copy end of page, adjusting alignment so
6273 * compiler may copy words instead of bytes.
6275 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6276 memcpy((size_t *)((char *)np + off),
6277 (size_t *)((char *)omp + off), sz - off);
6280 memcpy(np, omp, sz); /* Copy beginning of page */
6283 SETDSZ(leaf, data->mv_size);
6284 if (F_ISSET(flags, MDB_RESERVE))
6285 data->mv_data = METADATA(omp);
6287 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6291 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6293 } else if (data->mv_size == olddata.mv_size) {
6294 /* same size, just replace it. Note that we could
6295 * also reuse this node if the new data is smaller,
6296 * but instead we opt to shrink the node in that case.
6298 if (F_ISSET(flags, MDB_RESERVE))
6299 data->mv_data = olddata.mv_data;
6300 else if (!(mc->mc_flags & C_SUB))
6301 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6303 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6308 mdb_node_del(mc, 0);
6314 nflags = flags & NODE_ADD_FLAGS;
6315 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6316 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6317 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6318 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6320 nflags |= MDB_SPLIT_REPLACE;
6321 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6323 /* There is room already in this leaf page. */
6324 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6325 if (rc == 0 && insert_key) {
6326 /* Adjust other cursors pointing to mp */
6327 MDB_cursor *m2, *m3;
6328 MDB_dbi dbi = mc->mc_dbi;
6329 unsigned i = mc->mc_top;
6330 MDB_page *mp = mc->mc_pg[i];
6332 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6333 if (mc->mc_flags & C_SUB)
6334 m3 = &m2->mc_xcursor->mx_cursor;
6337 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6338 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6345 if (rc == MDB_SUCCESS) {
6346 /* Now store the actual data in the child DB. Note that we're
6347 * storing the user data in the keys field, so there are strict
6348 * size limits on dupdata. The actual data fields of the child
6349 * DB are all zero size.
6357 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6358 if (flags & MDB_CURRENT) {
6359 xflags = MDB_CURRENT|MDB_NOSPILL;
6361 mdb_xcursor_init1(mc, leaf);
6362 xflags = (flags & MDB_NODUPDATA) ?
6363 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6365 /* converted, write the original data first */
6367 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6371 /* Adjust other cursors pointing to mp */
6373 unsigned i = mc->mc_top;
6374 MDB_page *mp = mc->mc_pg[i];
6376 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6377 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6378 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6379 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6380 mdb_xcursor_init1(m2, leaf);
6384 /* we've done our job */
6387 ecount = mc->mc_xcursor->mx_db.md_entries;
6388 if (flags & MDB_APPENDDUP)
6389 xflags |= MDB_APPEND;
6390 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6391 if (flags & F_SUBDATA) {
6392 void *db = NODEDATA(leaf);
6393 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6395 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6397 /* Increment count unless we just replaced an existing item. */
6399 mc->mc_db->md_entries++;
6401 /* Invalidate txn if we created an empty sub-DB */
6404 /* If we succeeded and the key didn't exist before,
6405 * make sure the cursor is marked valid.
6407 mc->mc_flags |= C_INITIALIZED;
6409 if (flags & MDB_MULTIPLE) {
6412 /* let caller know how many succeeded, if any */
6413 data[1].mv_size = mcount;
6414 if (mcount < dcount) {
6415 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6416 insert_key = insert_data = 0;
6423 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6426 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6431 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6437 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6438 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6440 if (!(mc->mc_flags & C_INITIALIZED))
6443 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6444 return MDB_NOTFOUND;
6446 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6449 rc = mdb_cursor_touch(mc);
6453 mp = mc->mc_pg[mc->mc_top];
6456 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6458 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6459 if (flags & MDB_NODUPDATA) {
6460 /* mdb_cursor_del0() will subtract the final entry */
6461 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6463 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6464 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6466 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6469 /* If sub-DB still has entries, we're done */
6470 if (mc->mc_xcursor->mx_db.md_entries) {
6471 if (leaf->mn_flags & F_SUBDATA) {
6472 /* update subDB info */
6473 void *db = NODEDATA(leaf);
6474 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6477 /* shrink fake page */
6478 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6479 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6480 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6481 /* fix other sub-DB cursors pointed at this fake page */
6482 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6483 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6484 if (m2->mc_pg[mc->mc_top] == mp &&
6485 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6486 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6489 mc->mc_db->md_entries--;
6490 mc->mc_flags |= C_DEL;
6493 /* otherwise fall thru and delete the sub-DB */
6496 if (leaf->mn_flags & F_SUBDATA) {
6497 /* add all the child DB's pages to the free list */
6498 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6504 /* add overflow pages to free list */
6505 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6509 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6510 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6511 (rc = mdb_ovpage_free(mc, omp)))
6516 return mdb_cursor_del0(mc);
6519 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6523 /** Allocate and initialize new pages for a database.
6524 * @param[in] mc a cursor on the database being added to.
6525 * @param[in] flags flags defining what type of page is being allocated.
6526 * @param[in] num the number of pages to allocate. This is usually 1,
6527 * unless allocating overflow pages for a large record.
6528 * @param[out] mp Address of a page, or NULL on failure.
6529 * @return 0 on success, non-zero on failure.
6532 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6537 if ((rc = mdb_page_alloc(mc, num, &np)))
6539 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6540 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6541 np->mp_flags = flags | P_DIRTY;
6542 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6543 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6546 mc->mc_db->md_branch_pages++;
6547 else if (IS_LEAF(np))
6548 mc->mc_db->md_leaf_pages++;
6549 else if (IS_OVERFLOW(np)) {
6550 mc->mc_db->md_overflow_pages += num;
6558 /** Calculate the size of a leaf node.
6559 * The size depends on the environment's page size; if a data item
6560 * is too large it will be put onto an overflow page and the node
6561 * size will only include the key and not the data. Sizes are always
6562 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6563 * of the #MDB_node headers.
6564 * @param[in] env The environment handle.
6565 * @param[in] key The key for the node.
6566 * @param[in] data The data for the node.
6567 * @return The number of bytes needed to store the node.
6570 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6574 sz = LEAFSIZE(key, data);
6575 if (sz > env->me_nodemax) {
6576 /* put on overflow page */
6577 sz -= data->mv_size - sizeof(pgno_t);
6580 return EVEN(sz + sizeof(indx_t));
6583 /** Calculate the size of a branch node.
6584 * The size should depend on the environment's page size but since
6585 * we currently don't support spilling large keys onto overflow
6586 * pages, it's simply the size of the #MDB_node header plus the
6587 * size of the key. Sizes are always rounded up to an even number
6588 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6589 * @param[in] env The environment handle.
6590 * @param[in] key The key for the node.
6591 * @return The number of bytes needed to store the node.
6594 mdb_branch_size(MDB_env *env, MDB_val *key)
6599 if (sz > env->me_nodemax) {
6600 /* put on overflow page */
6601 /* not implemented */
6602 /* sz -= key->size - sizeof(pgno_t); */
6605 return sz + sizeof(indx_t);
6608 /** Add a node to the page pointed to by the cursor.
6609 * @param[in] mc The cursor for this operation.
6610 * @param[in] indx The index on the page where the new node should be added.
6611 * @param[in] key The key for the new node.
6612 * @param[in] data The data for the new node, if any.
6613 * @param[in] pgno The page number, if adding a branch node.
6614 * @param[in] flags Flags for the node.
6615 * @return 0 on success, non-zero on failure. Possible errors are:
6617 * <li>ENOMEM - failed to allocate overflow pages for the node.
6618 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6619 * should never happen since all callers already calculate the
6620 * page's free space before calling this function.
6624 mdb_node_add(MDB_cursor *mc, indx_t indx,
6625 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6628 size_t node_size = NODESIZE;
6632 MDB_page *mp = mc->mc_pg[mc->mc_top];
6633 MDB_page *ofp = NULL; /* overflow page */
6636 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6638 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6639 IS_LEAF(mp) ? "leaf" : "branch",
6640 IS_SUBP(mp) ? "sub-" : "",
6641 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6642 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6645 /* Move higher keys up one slot. */
6646 int ksize = mc->mc_db->md_pad, dif;
6647 char *ptr = LEAF2KEY(mp, indx, ksize);
6648 dif = NUMKEYS(mp) - indx;
6650 memmove(ptr+ksize, ptr, dif*ksize);
6651 /* insert new key */
6652 memcpy(ptr, key->mv_data, ksize);
6654 /* Just using these for counting */
6655 mp->mp_lower += sizeof(indx_t);
6656 mp->mp_upper -= ksize - sizeof(indx_t);
6660 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6662 node_size += key->mv_size;
6664 mdb_cassert(mc, data);
6665 if (F_ISSET(flags, F_BIGDATA)) {
6666 /* Data already on overflow page. */
6667 node_size += sizeof(pgno_t);
6668 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6669 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6671 /* Put data on overflow page. */
6672 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6673 data->mv_size, node_size+data->mv_size));
6674 node_size = EVEN(node_size + sizeof(pgno_t));
6675 if ((ssize_t)node_size > room)
6677 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6679 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6683 node_size += data->mv_size;
6686 node_size = EVEN(node_size);
6687 if ((ssize_t)node_size > room)
6691 /* Move higher pointers up one slot. */
6692 for (i = NUMKEYS(mp); i > indx; i--)
6693 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6695 /* Adjust free space offsets. */
6696 ofs = mp->mp_upper - node_size;
6697 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6698 mp->mp_ptrs[indx] = ofs;
6700 mp->mp_lower += sizeof(indx_t);
6702 /* Write the node data. */
6703 node = NODEPTR(mp, indx);
6704 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6705 node->mn_flags = flags;
6707 SETDSZ(node,data->mv_size);
6712 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6715 mdb_cassert(mc, key);
6717 if (F_ISSET(flags, F_BIGDATA))
6718 memcpy(node->mn_data + key->mv_size, data->mv_data,
6720 else if (F_ISSET(flags, MDB_RESERVE))
6721 data->mv_data = node->mn_data + key->mv_size;
6723 memcpy(node->mn_data + key->mv_size, data->mv_data,
6726 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6728 if (F_ISSET(flags, MDB_RESERVE))
6729 data->mv_data = METADATA(ofp);
6731 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6738 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6739 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6740 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6741 DPRINTF(("node size = %"Z"u", node_size));
6742 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6743 return MDB_PAGE_FULL;
6746 /** Delete the specified node from a page.
6747 * @param[in] mc Cursor pointing to the node to delete.
6748 * @param[in] ksize The size of a node. Only used if the page is
6749 * part of a #MDB_DUPFIXED database.
6752 mdb_node_del(MDB_cursor *mc, int ksize)
6754 MDB_page *mp = mc->mc_pg[mc->mc_top];
6755 indx_t indx = mc->mc_ki[mc->mc_top];
6757 indx_t i, j, numkeys, ptr;
6761 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6762 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6763 numkeys = NUMKEYS(mp);
6764 mdb_cassert(mc, indx < numkeys);
6767 int x = numkeys - 1 - indx;
6768 base = LEAF2KEY(mp, indx, ksize);
6770 memmove(base, base + ksize, x * ksize);
6771 mp->mp_lower -= sizeof(indx_t);
6772 mp->mp_upper += ksize - sizeof(indx_t);
6776 node = NODEPTR(mp, indx);
6777 sz = NODESIZE + node->mn_ksize;
6779 if (F_ISSET(node->mn_flags, F_BIGDATA))
6780 sz += sizeof(pgno_t);
6782 sz += NODEDSZ(node);
6786 ptr = mp->mp_ptrs[indx];
6787 for (i = j = 0; i < numkeys; i++) {
6789 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6790 if (mp->mp_ptrs[i] < ptr)
6791 mp->mp_ptrs[j] += sz;
6796 base = (char *)mp + mp->mp_upper + PAGEBASE;
6797 memmove(base + sz, base, ptr - mp->mp_upper);
6799 mp->mp_lower -= sizeof(indx_t);
6803 /** Compact the main page after deleting a node on a subpage.
6804 * @param[in] mp The main page to operate on.
6805 * @param[in] indx The index of the subpage on the main page.
6808 mdb_node_shrink(MDB_page *mp, indx_t indx)
6814 indx_t i, numkeys, ptr;
6816 node = NODEPTR(mp, indx);
6817 sp = (MDB_page *)NODEDATA(node);
6818 delta = SIZELEFT(sp);
6819 xp = (MDB_page *)((char *)sp + delta);
6821 /* shift subpage upward */
6823 nsize = NUMKEYS(sp) * sp->mp_pad;
6825 return; /* do not make the node uneven-sized */
6826 memmove(METADATA(xp), METADATA(sp), nsize);
6829 numkeys = NUMKEYS(sp);
6830 for (i=numkeys-1; i>=0; i--)
6831 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6833 xp->mp_upper = sp->mp_lower;
6834 xp->mp_lower = sp->mp_lower;
6835 xp->mp_flags = sp->mp_flags;
6836 xp->mp_pad = sp->mp_pad;
6837 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6839 nsize = NODEDSZ(node) - delta;
6840 SETDSZ(node, nsize);
6842 /* shift lower nodes upward */
6843 ptr = mp->mp_ptrs[indx];
6844 numkeys = NUMKEYS(mp);
6845 for (i = 0; i < numkeys; i++) {
6846 if (mp->mp_ptrs[i] <= ptr)
6847 mp->mp_ptrs[i] += delta;
6850 base = (char *)mp + mp->mp_upper + PAGEBASE;
6851 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6852 mp->mp_upper += delta;
6855 /** Initial setup of a sorted-dups cursor.
6856 * Sorted duplicates are implemented as a sub-database for the given key.
6857 * The duplicate data items are actually keys of the sub-database.
6858 * Operations on the duplicate data items are performed using a sub-cursor
6859 * initialized when the sub-database is first accessed. This function does
6860 * the preliminary setup of the sub-cursor, filling in the fields that
6861 * depend only on the parent DB.
6862 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6865 mdb_xcursor_init0(MDB_cursor *mc)
6867 MDB_xcursor *mx = mc->mc_xcursor;
6869 mx->mx_cursor.mc_xcursor = NULL;
6870 mx->mx_cursor.mc_txn = mc->mc_txn;
6871 mx->mx_cursor.mc_db = &mx->mx_db;
6872 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6873 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6874 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6875 mx->mx_cursor.mc_snum = 0;
6876 mx->mx_cursor.mc_top = 0;
6877 mx->mx_cursor.mc_flags = C_SUB;
6878 mx->mx_dbx.md_name.mv_size = 0;
6879 mx->mx_dbx.md_name.mv_data = NULL;
6880 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6881 mx->mx_dbx.md_dcmp = NULL;
6882 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6885 /** Final setup of a sorted-dups cursor.
6886 * Sets up the fields that depend on the data from the main cursor.
6887 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6888 * @param[in] node The data containing the #MDB_db record for the
6889 * sorted-dup database.
6892 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6894 MDB_xcursor *mx = mc->mc_xcursor;
6896 if (node->mn_flags & F_SUBDATA) {
6897 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6898 mx->mx_cursor.mc_pg[0] = 0;
6899 mx->mx_cursor.mc_snum = 0;
6900 mx->mx_cursor.mc_top = 0;
6901 mx->mx_cursor.mc_flags = C_SUB;
6903 MDB_page *fp = NODEDATA(node);
6904 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6905 mx->mx_db.md_flags = 0;
6906 mx->mx_db.md_depth = 1;
6907 mx->mx_db.md_branch_pages = 0;
6908 mx->mx_db.md_leaf_pages = 1;
6909 mx->mx_db.md_overflow_pages = 0;
6910 mx->mx_db.md_entries = NUMKEYS(fp);
6911 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6912 mx->mx_cursor.mc_snum = 1;
6913 mx->mx_cursor.mc_top = 0;
6914 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6915 mx->mx_cursor.mc_pg[0] = fp;
6916 mx->mx_cursor.mc_ki[0] = 0;
6917 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6918 mx->mx_db.md_flags = MDB_DUPFIXED;
6919 mx->mx_db.md_pad = fp->mp_pad;
6920 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6921 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6924 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6925 mx->mx_db.md_root));
6926 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6927 #if UINT_MAX < SIZE_MAX
6928 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6929 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6933 /** Initialize a cursor for a given transaction and database. */
6935 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6938 mc->mc_backup = NULL;
6941 mc->mc_db = &txn->mt_dbs[dbi];
6942 mc->mc_dbx = &txn->mt_dbxs[dbi];
6943 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6948 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6949 mdb_tassert(txn, mx != NULL);
6950 mc->mc_xcursor = mx;
6951 mdb_xcursor_init0(mc);
6953 mc->mc_xcursor = NULL;
6955 if (*mc->mc_dbflag & DB_STALE) {
6956 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6961 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6964 size_t size = sizeof(MDB_cursor);
6966 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6969 if (txn->mt_flags & MDB_TXN_ERROR)
6972 /* Allow read access to the freelist */
6973 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6976 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6977 size += sizeof(MDB_xcursor);
6979 if ((mc = malloc(size)) != NULL) {
6980 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6981 if (txn->mt_cursors) {
6982 mc->mc_next = txn->mt_cursors[dbi];
6983 txn->mt_cursors[dbi] = mc;
6984 mc->mc_flags |= C_UNTRACK;
6996 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6998 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7001 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7004 if (txn->mt_flags & MDB_TXN_ERROR)
7007 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7011 /* Return the count of duplicate data items for the current key */
7013 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7017 if (mc == NULL || countp == NULL)
7020 if (mc->mc_xcursor == NULL)
7021 return MDB_INCOMPATIBLE;
7023 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7026 if (!(mc->mc_flags & C_INITIALIZED))
7029 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7030 return MDB_NOTFOUND;
7032 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7033 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7036 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7039 *countp = mc->mc_xcursor->mx_db.md_entries;
7045 mdb_cursor_close(MDB_cursor *mc)
7047 if (mc && !mc->mc_backup) {
7048 /* remove from txn, if tracked */
7049 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7050 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7051 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7053 *prev = mc->mc_next;
7060 mdb_cursor_txn(MDB_cursor *mc)
7062 if (!mc) return NULL;
7067 mdb_cursor_dbi(MDB_cursor *mc)
7072 /** Replace the key for a branch node with a new key.
7073 * @param[in] mc Cursor pointing to the node to operate on.
7074 * @param[in] key The new key to use.
7075 * @return 0 on success, non-zero on failure.
7078 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7084 int delta, ksize, oksize;
7085 indx_t ptr, i, numkeys, indx;
7088 indx = mc->mc_ki[mc->mc_top];
7089 mp = mc->mc_pg[mc->mc_top];
7090 node = NODEPTR(mp, indx);
7091 ptr = mp->mp_ptrs[indx];
7095 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7096 k2.mv_data = NODEKEY(node);
7097 k2.mv_size = node->mn_ksize;
7098 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7100 mdb_dkey(&k2, kbuf2),
7106 /* Sizes must be 2-byte aligned. */
7107 ksize = EVEN(key->mv_size);
7108 oksize = EVEN(node->mn_ksize);
7109 delta = ksize - oksize;
7111 /* Shift node contents if EVEN(key length) changed. */
7113 if (delta > 0 && SIZELEFT(mp) < delta) {
7115 /* not enough space left, do a delete and split */
7116 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7117 pgno = NODEPGNO(node);
7118 mdb_node_del(mc, 0);
7119 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7122 numkeys = NUMKEYS(mp);
7123 for (i = 0; i < numkeys; i++) {
7124 if (mp->mp_ptrs[i] <= ptr)
7125 mp->mp_ptrs[i] -= delta;
7128 base = (char *)mp + mp->mp_upper + PAGEBASE;
7129 len = ptr - mp->mp_upper + NODESIZE;
7130 memmove(base - delta, base, len);
7131 mp->mp_upper -= delta;
7133 node = NODEPTR(mp, indx);
7136 /* But even if no shift was needed, update ksize */
7137 if (node->mn_ksize != key->mv_size)
7138 node->mn_ksize = key->mv_size;
7141 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7147 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7149 /** Move a node from csrc to cdst.
7152 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7159 unsigned short flags;
7163 /* Mark src and dst as dirty. */
7164 if ((rc = mdb_page_touch(csrc)) ||
7165 (rc = mdb_page_touch(cdst)))
7168 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7169 key.mv_size = csrc->mc_db->md_pad;
7170 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7172 data.mv_data = NULL;
7176 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7177 mdb_cassert(csrc, !((size_t)srcnode & 1));
7178 srcpg = NODEPGNO(srcnode);
7179 flags = srcnode->mn_flags;
7180 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7181 unsigned int snum = csrc->mc_snum;
7183 /* must find the lowest key below src */
7184 rc = mdb_page_search_lowest(csrc);
7187 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7188 key.mv_size = csrc->mc_db->md_pad;
7189 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7191 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7192 key.mv_size = NODEKSZ(s2);
7193 key.mv_data = NODEKEY(s2);
7195 csrc->mc_snum = snum--;
7196 csrc->mc_top = snum;
7198 key.mv_size = NODEKSZ(srcnode);
7199 key.mv_data = NODEKEY(srcnode);
7201 data.mv_size = NODEDSZ(srcnode);
7202 data.mv_data = NODEDATA(srcnode);
7204 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7205 unsigned int snum = cdst->mc_snum;
7208 /* must find the lowest key below dst */
7209 mdb_cursor_copy(cdst, &mn);
7210 rc = mdb_page_search_lowest(&mn);
7213 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7214 bkey.mv_size = mn.mc_db->md_pad;
7215 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7217 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7218 bkey.mv_size = NODEKSZ(s2);
7219 bkey.mv_data = NODEKEY(s2);
7221 mn.mc_snum = snum--;
7224 rc = mdb_update_key(&mn, &bkey);
7229 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7230 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7231 csrc->mc_ki[csrc->mc_top],
7233 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7234 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7236 /* Add the node to the destination page.
7238 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7239 if (rc != MDB_SUCCESS)
7242 /* Delete the node from the source page.
7244 mdb_node_del(csrc, key.mv_size);
7247 /* Adjust other cursors pointing to mp */
7248 MDB_cursor *m2, *m3;
7249 MDB_dbi dbi = csrc->mc_dbi;
7250 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7252 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7253 if (csrc->mc_flags & C_SUB)
7254 m3 = &m2->mc_xcursor->mx_cursor;
7257 if (m3 == csrc) continue;
7258 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7259 csrc->mc_ki[csrc->mc_top]) {
7260 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7261 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7266 /* Update the parent separators.
7268 if (csrc->mc_ki[csrc->mc_top] == 0) {
7269 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7270 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7271 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7273 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7274 key.mv_size = NODEKSZ(srcnode);
7275 key.mv_data = NODEKEY(srcnode);
7277 DPRINTF(("update separator for source page %"Z"u to [%s]",
7278 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7279 mdb_cursor_copy(csrc, &mn);
7282 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7285 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7287 indx_t ix = csrc->mc_ki[csrc->mc_top];
7288 nullkey.mv_size = 0;
7289 csrc->mc_ki[csrc->mc_top] = 0;
7290 rc = mdb_update_key(csrc, &nullkey);
7291 csrc->mc_ki[csrc->mc_top] = ix;
7292 mdb_cassert(csrc, rc == MDB_SUCCESS);
7296 if (cdst->mc_ki[cdst->mc_top] == 0) {
7297 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7298 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7299 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7301 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7302 key.mv_size = NODEKSZ(srcnode);
7303 key.mv_data = NODEKEY(srcnode);
7305 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7306 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7307 mdb_cursor_copy(cdst, &mn);
7310 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7313 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7315 indx_t ix = cdst->mc_ki[cdst->mc_top];
7316 nullkey.mv_size = 0;
7317 cdst->mc_ki[cdst->mc_top] = 0;
7318 rc = mdb_update_key(cdst, &nullkey);
7319 cdst->mc_ki[cdst->mc_top] = ix;
7320 mdb_cassert(csrc, rc == MDB_SUCCESS);
7327 /** Merge one page into another.
7328 * The nodes from the page pointed to by \b csrc will
7329 * be copied to the page pointed to by \b cdst and then
7330 * the \b csrc page will be freed.
7331 * @param[in] csrc Cursor pointing to the source page.
7332 * @param[in] cdst Cursor pointing to the destination page.
7333 * @return 0 on success, non-zero on failure.
7336 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7338 MDB_page *psrc, *pdst;
7345 psrc = csrc->mc_pg[csrc->mc_top];
7346 pdst = cdst->mc_pg[cdst->mc_top];
7348 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7350 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7351 mdb_cassert(csrc, cdst->mc_snum > 1);
7353 /* Mark dst as dirty. */
7354 if ((rc = mdb_page_touch(cdst)))
7357 /* Move all nodes from src to dst.
7359 j = nkeys = NUMKEYS(pdst);
7360 if (IS_LEAF2(psrc)) {
7361 key.mv_size = csrc->mc_db->md_pad;
7362 key.mv_data = METADATA(psrc);
7363 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7364 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7365 if (rc != MDB_SUCCESS)
7367 key.mv_data = (char *)key.mv_data + key.mv_size;
7370 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7371 srcnode = NODEPTR(psrc, i);
7372 if (i == 0 && IS_BRANCH(psrc)) {
7375 mdb_cursor_copy(csrc, &mn);
7376 /* must find the lowest key below src */
7377 rc = mdb_page_search_lowest(&mn);
7380 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7381 key.mv_size = mn.mc_db->md_pad;
7382 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7384 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7385 key.mv_size = NODEKSZ(s2);
7386 key.mv_data = NODEKEY(s2);
7389 key.mv_size = srcnode->mn_ksize;
7390 key.mv_data = NODEKEY(srcnode);
7393 data.mv_size = NODEDSZ(srcnode);
7394 data.mv_data = NODEDATA(srcnode);
7395 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7396 if (rc != MDB_SUCCESS)
7401 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7402 pdst->mp_pgno, NUMKEYS(pdst),
7403 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7405 /* Unlink the src page from parent and add to free list.
7408 mdb_node_del(csrc, 0);
7409 if (csrc->mc_ki[csrc->mc_top] == 0) {
7411 rc = mdb_update_key(csrc, &key);
7419 psrc = csrc->mc_pg[csrc->mc_top];
7420 /* If not operating on FreeDB, allow this page to be reused
7421 * in this txn. Otherwise just add to free list.
7423 rc = mdb_page_loose(csrc, psrc);
7427 csrc->mc_db->md_leaf_pages--;
7429 csrc->mc_db->md_branch_pages--;
7431 /* Adjust other cursors pointing to mp */
7432 MDB_cursor *m2, *m3;
7433 MDB_dbi dbi = csrc->mc_dbi;
7435 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7436 if (csrc->mc_flags & C_SUB)
7437 m3 = &m2->mc_xcursor->mx_cursor;
7440 if (m3 == csrc) continue;
7441 if (m3->mc_snum < csrc->mc_snum) continue;
7442 if (m3->mc_pg[csrc->mc_top] == psrc) {
7443 m3->mc_pg[csrc->mc_top] = pdst;
7444 m3->mc_ki[csrc->mc_top] += nkeys;
7449 unsigned int snum = cdst->mc_snum;
7450 uint16_t depth = cdst->mc_db->md_depth;
7451 mdb_cursor_pop(cdst);
7452 rc = mdb_rebalance(cdst);
7453 /* Did the tree shrink? */
7454 if (depth > cdst->mc_db->md_depth)
7456 cdst->mc_snum = snum;
7457 cdst->mc_top = snum-1;
7462 /** Copy the contents of a cursor.
7463 * @param[in] csrc The cursor to copy from.
7464 * @param[out] cdst The cursor to copy to.
7467 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7471 cdst->mc_txn = csrc->mc_txn;
7472 cdst->mc_dbi = csrc->mc_dbi;
7473 cdst->mc_db = csrc->mc_db;
7474 cdst->mc_dbx = csrc->mc_dbx;
7475 cdst->mc_snum = csrc->mc_snum;
7476 cdst->mc_top = csrc->mc_top;
7477 cdst->mc_flags = csrc->mc_flags;
7479 for (i=0; i<csrc->mc_snum; i++) {
7480 cdst->mc_pg[i] = csrc->mc_pg[i];
7481 cdst->mc_ki[i] = csrc->mc_ki[i];
7485 /** Rebalance the tree after a delete operation.
7486 * @param[in] mc Cursor pointing to the page where rebalancing
7488 * @return 0 on success, non-zero on failure.
7491 mdb_rebalance(MDB_cursor *mc)
7495 unsigned int ptop, minkeys;
7499 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7500 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7501 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7502 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7503 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7505 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7506 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7507 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7508 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7512 if (mc->mc_snum < 2) {
7513 MDB_page *mp = mc->mc_pg[0];
7515 DPUTS("Can't rebalance a subpage, ignoring");
7518 if (NUMKEYS(mp) == 0) {
7519 DPUTS("tree is completely empty");
7520 mc->mc_db->md_root = P_INVALID;
7521 mc->mc_db->md_depth = 0;
7522 mc->mc_db->md_leaf_pages = 0;
7523 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7526 /* Adjust cursors pointing to mp */
7529 mc->mc_flags &= ~C_INITIALIZED;
7531 MDB_cursor *m2, *m3;
7532 MDB_dbi dbi = mc->mc_dbi;
7534 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7535 if (mc->mc_flags & C_SUB)
7536 m3 = &m2->mc_xcursor->mx_cursor;
7539 if (m3->mc_snum < mc->mc_snum) continue;
7540 if (m3->mc_pg[0] == mp) {
7543 m3->mc_flags &= ~C_INITIALIZED;
7547 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7549 DPUTS("collapsing root page!");
7550 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7553 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7554 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7557 mc->mc_db->md_depth--;
7558 mc->mc_db->md_branch_pages--;
7559 mc->mc_ki[0] = mc->mc_ki[1];
7560 for (i = 1; i<mc->mc_db->md_depth; i++) {
7561 mc->mc_pg[i] = mc->mc_pg[i+1];
7562 mc->mc_ki[i] = mc->mc_ki[i+1];
7565 /* Adjust other cursors pointing to mp */
7566 MDB_cursor *m2, *m3;
7567 MDB_dbi dbi = mc->mc_dbi;
7569 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7570 if (mc->mc_flags & C_SUB)
7571 m3 = &m2->mc_xcursor->mx_cursor;
7574 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7575 if (m3->mc_pg[0] == mp) {
7578 for (i=0; i<m3->mc_snum; i++) {
7579 m3->mc_pg[i] = m3->mc_pg[i+1];
7580 m3->mc_ki[i] = m3->mc_ki[i+1];
7586 DPUTS("root page doesn't need rebalancing");
7590 /* The parent (branch page) must have at least 2 pointers,
7591 * otherwise the tree is invalid.
7593 ptop = mc->mc_top-1;
7594 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7596 /* Leaf page fill factor is below the threshold.
7597 * Try to move keys from left or right neighbor, or
7598 * merge with a neighbor page.
7603 mdb_cursor_copy(mc, &mn);
7604 mn.mc_xcursor = NULL;
7606 oldki = mc->mc_ki[mc->mc_top];
7607 if (mc->mc_ki[ptop] == 0) {
7608 /* We're the leftmost leaf in our parent.
7610 DPUTS("reading right neighbor");
7612 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7613 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7616 mn.mc_ki[mn.mc_top] = 0;
7617 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7619 /* There is at least one neighbor to the left.
7621 DPUTS("reading left neighbor");
7623 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7624 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7627 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7628 mc->mc_ki[mc->mc_top] = 0;
7631 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7632 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7633 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7635 /* If the neighbor page is above threshold and has enough keys,
7636 * move one key from it. Otherwise we should try to merge them.
7637 * (A branch page must never have less than 2 keys.)
7639 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7640 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7641 rc = mdb_node_move(&mn, mc);
7642 if (mc->mc_ki[ptop]) {
7646 if (mc->mc_ki[ptop] == 0) {
7647 rc = mdb_page_merge(&mn, mc);
7649 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7650 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7651 rc = mdb_page_merge(mc, &mn);
7652 mdb_cursor_copy(&mn, mc);
7654 mc->mc_flags &= ~C_EOF;
7656 mc->mc_ki[mc->mc_top] = oldki;
7660 /** Complete a delete operation started by #mdb_cursor_del(). */
7662 mdb_cursor_del0(MDB_cursor *mc)
7669 ki = mc->mc_ki[mc->mc_top];
7670 mdb_node_del(mc, mc->mc_db->md_pad);
7671 mc->mc_db->md_entries--;
7672 rc = mdb_rebalance(mc);
7674 if (rc == MDB_SUCCESS) {
7675 MDB_cursor *m2, *m3;
7676 MDB_dbi dbi = mc->mc_dbi;
7678 mp = mc->mc_pg[mc->mc_top];
7679 nkeys = NUMKEYS(mp);
7681 /* if mc points past last node in page, find next sibling */
7682 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7683 rc = mdb_cursor_sibling(mc, 1);
7684 if (rc == MDB_NOTFOUND) {
7685 mc->mc_flags |= C_EOF;
7690 /* Adjust other cursors pointing to mp */
7691 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7692 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7693 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7695 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7697 if (m3->mc_pg[mc->mc_top] == mp) {
7698 if (m3->mc_ki[mc->mc_top] >= ki) {
7699 m3->mc_flags |= C_DEL;
7700 if (m3->mc_ki[mc->mc_top] > ki)
7701 m3->mc_ki[mc->mc_top]--;
7702 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7703 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7705 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7706 rc = mdb_cursor_sibling(m3, 1);
7707 if (rc == MDB_NOTFOUND) {
7708 m3->mc_flags |= C_EOF;
7714 mc->mc_flags |= C_DEL;
7718 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7723 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7724 MDB_val *key, MDB_val *data)
7726 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7729 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7730 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7732 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7733 /* must ignore any data */
7737 return mdb_del0(txn, dbi, key, data, 0);
7741 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7742 MDB_val *key, MDB_val *data, unsigned flags)
7747 MDB_val rdata, *xdata;
7751 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7753 mdb_cursor_init(&mc, txn, dbi, &mx);
7762 flags |= MDB_NODUPDATA;
7764 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7766 /* let mdb_page_split know about this cursor if needed:
7767 * delete will trigger a rebalance; if it needs to move
7768 * a node from one page to another, it will have to
7769 * update the parent's separator key(s). If the new sepkey
7770 * is larger than the current one, the parent page may
7771 * run out of space, triggering a split. We need this
7772 * cursor to be consistent until the end of the rebalance.
7774 mc.mc_flags |= C_UNTRACK;
7775 mc.mc_next = txn->mt_cursors[dbi];
7776 txn->mt_cursors[dbi] = &mc;
7777 rc = mdb_cursor_del(&mc, flags);
7778 txn->mt_cursors[dbi] = mc.mc_next;
7783 /** Split a page and insert a new node.
7784 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7785 * The cursor will be updated to point to the actual page and index where
7786 * the node got inserted after the split.
7787 * @param[in] newkey The key for the newly inserted node.
7788 * @param[in] newdata The data for the newly inserted node.
7789 * @param[in] newpgno The page number, if the new node is a branch node.
7790 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7791 * @return 0 on success, non-zero on failure.
7794 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7795 unsigned int nflags)
7798 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7801 int i, j, split_indx, nkeys, pmax;
7802 MDB_env *env = mc->mc_txn->mt_env;
7804 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7805 MDB_page *copy = NULL;
7806 MDB_page *mp, *rp, *pp;
7811 mp = mc->mc_pg[mc->mc_top];
7812 newindx = mc->mc_ki[mc->mc_top];
7813 nkeys = NUMKEYS(mp);
7815 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7816 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7817 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7819 /* Create a right sibling. */
7820 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7822 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7824 if (mc->mc_snum < 2) {
7825 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7827 /* shift current top to make room for new parent */
7828 mc->mc_pg[1] = mc->mc_pg[0];
7829 mc->mc_ki[1] = mc->mc_ki[0];
7832 mc->mc_db->md_root = pp->mp_pgno;
7833 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7834 mc->mc_db->md_depth++;
7837 /* Add left (implicit) pointer. */
7838 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7839 /* undo the pre-push */
7840 mc->mc_pg[0] = mc->mc_pg[1];
7841 mc->mc_ki[0] = mc->mc_ki[1];
7842 mc->mc_db->md_root = mp->mp_pgno;
7843 mc->mc_db->md_depth--;
7850 ptop = mc->mc_top-1;
7851 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7854 mc->mc_flags |= C_SPLITTING;
7855 mdb_cursor_copy(mc, &mn);
7856 mn.mc_pg[mn.mc_top] = rp;
7857 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7859 if (nflags & MDB_APPEND) {
7860 mn.mc_ki[mn.mc_top] = 0;
7862 split_indx = newindx;
7866 split_indx = (nkeys+1) / 2;
7871 unsigned int lsize, rsize, ksize;
7872 /* Move half of the keys to the right sibling */
7873 x = mc->mc_ki[mc->mc_top] - split_indx;
7874 ksize = mc->mc_db->md_pad;
7875 split = LEAF2KEY(mp, split_indx, ksize);
7876 rsize = (nkeys - split_indx) * ksize;
7877 lsize = (nkeys - split_indx) * sizeof(indx_t);
7878 mp->mp_lower -= lsize;
7879 rp->mp_lower += lsize;
7880 mp->mp_upper += rsize - lsize;
7881 rp->mp_upper -= rsize - lsize;
7882 sepkey.mv_size = ksize;
7883 if (newindx == split_indx) {
7884 sepkey.mv_data = newkey->mv_data;
7886 sepkey.mv_data = split;
7889 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7890 memcpy(rp->mp_ptrs, split, rsize);
7891 sepkey.mv_data = rp->mp_ptrs;
7892 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7893 memcpy(ins, newkey->mv_data, ksize);
7894 mp->mp_lower += sizeof(indx_t);
7895 mp->mp_upper -= ksize - sizeof(indx_t);
7898 memcpy(rp->mp_ptrs, split, x * ksize);
7899 ins = LEAF2KEY(rp, x, ksize);
7900 memcpy(ins, newkey->mv_data, ksize);
7901 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7902 rp->mp_lower += sizeof(indx_t);
7903 rp->mp_upper -= ksize - sizeof(indx_t);
7904 mc->mc_ki[mc->mc_top] = x;
7905 mc->mc_pg[mc->mc_top] = rp;
7908 int psize, nsize, k;
7909 /* Maximum free space in an empty page */
7910 pmax = env->me_psize - PAGEHDRSZ;
7912 nsize = mdb_leaf_size(env, newkey, newdata);
7914 nsize = mdb_branch_size(env, newkey);
7915 nsize = EVEN(nsize);
7917 /* grab a page to hold a temporary copy */
7918 copy = mdb_page_malloc(mc->mc_txn, 1);
7923 copy->mp_pgno = mp->mp_pgno;
7924 copy->mp_flags = mp->mp_flags;
7925 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7926 copy->mp_upper = env->me_psize - PAGEBASE;
7928 /* prepare to insert */
7929 for (i=0, j=0; i<nkeys; i++) {
7931 copy->mp_ptrs[j++] = 0;
7933 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7936 /* When items are relatively large the split point needs
7937 * to be checked, because being off-by-one will make the
7938 * difference between success or failure in mdb_node_add.
7940 * It's also relevant if a page happens to be laid out
7941 * such that one half of its nodes are all "small" and
7942 * the other half of its nodes are "large." If the new
7943 * item is also "large" and falls on the half with
7944 * "large" nodes, it also may not fit.
7946 * As a final tweak, if the new item goes on the last
7947 * spot on the page (and thus, onto the new page), bias
7948 * the split so the new page is emptier than the old page.
7949 * This yields better packing during sequential inserts.
7951 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7952 /* Find split point */
7954 if (newindx <= split_indx || newindx >= nkeys) {
7956 k = newindx >= nkeys ? nkeys : split_indx+2;
7961 for (; i!=k; i+=j) {
7966 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
7967 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7969 if (F_ISSET(node->mn_flags, F_BIGDATA))
7970 psize += sizeof(pgno_t);
7972 psize += NODEDSZ(node);
7974 psize = EVEN(psize);
7976 if (psize > pmax || i == k-j) {
7977 split_indx = i + (j<0);
7982 if (split_indx == newindx) {
7983 sepkey.mv_size = newkey->mv_size;
7984 sepkey.mv_data = newkey->mv_data;
7986 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
7987 sepkey.mv_size = node->mn_ksize;
7988 sepkey.mv_data = NODEKEY(node);
7993 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7995 /* Copy separator key to the parent.
7997 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8001 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8006 if (mn.mc_snum == mc->mc_snum) {
8007 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8008 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8009 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8010 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8015 /* Right page might now have changed parent.
8016 * Check if left page also changed parent.
8018 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8019 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8020 for (i=0; i<ptop; i++) {
8021 mc->mc_pg[i] = mn.mc_pg[i];
8022 mc->mc_ki[i] = mn.mc_ki[i];
8024 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8025 if (mn.mc_ki[ptop]) {
8026 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8028 /* find right page's left sibling */
8029 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8030 mdb_cursor_sibling(mc, 0);
8035 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8038 mc->mc_flags ^= C_SPLITTING;
8039 if (rc != MDB_SUCCESS) {
8042 if (nflags & MDB_APPEND) {
8043 mc->mc_pg[mc->mc_top] = rp;
8044 mc->mc_ki[mc->mc_top] = 0;
8045 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8048 for (i=0; i<mc->mc_top; i++)
8049 mc->mc_ki[i] = mn.mc_ki[i];
8050 } else if (!IS_LEAF2(mp)) {
8052 mc->mc_pg[mc->mc_top] = rp;
8057 rkey.mv_data = newkey->mv_data;
8058 rkey.mv_size = newkey->mv_size;
8064 /* Update index for the new key. */
8065 mc->mc_ki[mc->mc_top] = j;
8067 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8068 rkey.mv_data = NODEKEY(node);
8069 rkey.mv_size = node->mn_ksize;
8071 xdata.mv_data = NODEDATA(node);
8072 xdata.mv_size = NODEDSZ(node);
8075 pgno = NODEPGNO(node);
8076 flags = node->mn_flags;
8079 if (!IS_LEAF(mp) && j == 0) {
8080 /* First branch index doesn't need key data. */
8084 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8090 mc->mc_pg[mc->mc_top] = copy;
8095 } while (i != split_indx);
8097 nkeys = NUMKEYS(copy);
8098 for (i=0; i<nkeys; i++)
8099 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8100 mp->mp_lower = copy->mp_lower;
8101 mp->mp_upper = copy->mp_upper;
8102 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8103 env->me_psize - copy->mp_upper - PAGEBASE);
8105 /* reset back to original page */
8106 if (newindx < split_indx) {
8107 mc->mc_pg[mc->mc_top] = mp;
8108 if (nflags & MDB_RESERVE) {
8109 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8110 if (!(node->mn_flags & F_BIGDATA))
8111 newdata->mv_data = NODEDATA(node);
8114 mc->mc_pg[mc->mc_top] = rp;
8116 /* Make sure mc_ki is still valid.
8118 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8119 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8120 for (i=0; i<=ptop; i++) {
8121 mc->mc_pg[i] = mn.mc_pg[i];
8122 mc->mc_ki[i] = mn.mc_ki[i];
8129 /* Adjust other cursors pointing to mp */
8130 MDB_cursor *m2, *m3;
8131 MDB_dbi dbi = mc->mc_dbi;
8132 int fixup = NUMKEYS(mp);
8134 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8135 if (mc->mc_flags & C_SUB)
8136 m3 = &m2->mc_xcursor->mx_cursor;
8141 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8143 if (m3->mc_flags & C_SPLITTING)
8148 for (k=m3->mc_top; k>=0; k--) {
8149 m3->mc_ki[k+1] = m3->mc_ki[k];
8150 m3->mc_pg[k+1] = m3->mc_pg[k];
8152 if (m3->mc_ki[0] >= split_indx) {
8157 m3->mc_pg[0] = mc->mc_pg[0];
8161 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8162 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8163 m3->mc_ki[mc->mc_top]++;
8164 if (m3->mc_ki[mc->mc_top] >= fixup) {
8165 m3->mc_pg[mc->mc_top] = rp;
8166 m3->mc_ki[mc->mc_top] -= fixup;
8167 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8169 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8170 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8175 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8178 if (copy) /* tmp page */
8179 mdb_page_free(env, copy);
8181 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8186 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8187 MDB_val *key, MDB_val *data, unsigned int flags)
8192 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8195 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8198 mdb_cursor_init(&mc, txn, dbi, &mx);
8199 return mdb_cursor_put(&mc, key, data, flags);
8203 #define MDB_WBUF (1024*1024)
8206 /** State needed for a compacting copy. */
8207 typedef struct mdb_copy {
8208 pthread_mutex_t mc_mutex;
8209 pthread_cond_t mc_cond;
8216 pgno_t mc_next_pgno;
8219 volatile int mc_new;
8224 /** Dedicated writer thread for compacting copy. */
8225 static THREAD_RET ESECT
8226 mdb_env_copythr(void *arg)
8230 int toggle = 0, wsize, rc;
8233 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8236 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8239 pthread_mutex_lock(&my->mc_mutex);
8241 pthread_cond_signal(&my->mc_cond);
8244 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8245 if (my->mc_new < 0) {
8250 wsize = my->mc_wlen[toggle];
8251 ptr = my->mc_wbuf[toggle];
8254 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8258 } else if (len > 0) {
8272 /* If there's an overflow page tail, write it too */
8273 if (my->mc_olen[toggle]) {
8274 wsize = my->mc_olen[toggle];
8275 ptr = my->mc_over[toggle];
8276 my->mc_olen[toggle] = 0;
8279 my->mc_wlen[toggle] = 0;
8281 pthread_cond_signal(&my->mc_cond);
8283 pthread_cond_signal(&my->mc_cond);
8284 pthread_mutex_unlock(&my->mc_mutex);
8285 return (THREAD_RET)0;
8289 /** Tell the writer thread there's a buffer ready to write */
8291 mdb_env_cthr_toggle(mdb_copy *my, int st)
8293 int toggle = my->mc_toggle ^ 1;
8294 pthread_mutex_lock(&my->mc_mutex);
8295 if (my->mc_status) {
8296 pthread_mutex_unlock(&my->mc_mutex);
8297 return my->mc_status;
8299 while (my->mc_new == 1)
8300 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8302 my->mc_toggle = toggle;
8303 pthread_cond_signal(&my->mc_cond);
8304 pthread_mutex_unlock(&my->mc_mutex);
8308 /** Depth-first tree traversal for compacting copy. */
8310 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8313 MDB_txn *txn = my->mc_txn;
8315 MDB_page *mo, *mp, *leaf;
8320 /* Empty DB, nothing to do */
8321 if (*pg == P_INVALID)
8328 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8331 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8335 /* Make cursor pages writable */
8336 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8340 for (i=0; i<mc.mc_top; i++) {
8341 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8342 mc.mc_pg[i] = (MDB_page *)ptr;
8343 ptr += my->mc_env->me_psize;
8346 /* This is writable space for a leaf page. Usually not needed. */
8347 leaf = (MDB_page *)ptr;
8349 toggle = my->mc_toggle;
8350 while (mc.mc_snum > 0) {
8352 mp = mc.mc_pg[mc.mc_top];
8356 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8357 for (i=0; i<n; i++) {
8358 ni = NODEPTR(mp, i);
8359 if (ni->mn_flags & F_BIGDATA) {
8363 /* Need writable leaf */
8365 mc.mc_pg[mc.mc_top] = leaf;
8366 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8368 ni = NODEPTR(mp, i);
8371 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8372 rc = mdb_page_get(txn, pg, &omp, NULL);
8375 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8376 rc = mdb_env_cthr_toggle(my, 1);
8379 toggle = my->mc_toggle;
8381 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8382 memcpy(mo, omp, my->mc_env->me_psize);
8383 mo->mp_pgno = my->mc_next_pgno;
8384 my->mc_next_pgno += omp->mp_pages;
8385 my->mc_wlen[toggle] += my->mc_env->me_psize;
8386 if (omp->mp_pages > 1) {
8387 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8388 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8389 rc = mdb_env_cthr_toggle(my, 1);
8392 toggle = my->mc_toggle;
8394 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8395 } else if (ni->mn_flags & F_SUBDATA) {
8398 /* Need writable leaf */
8400 mc.mc_pg[mc.mc_top] = leaf;
8401 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8403 ni = NODEPTR(mp, i);
8406 memcpy(&db, NODEDATA(ni), sizeof(db));
8407 my->mc_toggle = toggle;
8408 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8411 toggle = my->mc_toggle;
8412 memcpy(NODEDATA(ni), &db, sizeof(db));
8417 mc.mc_ki[mc.mc_top]++;
8418 if (mc.mc_ki[mc.mc_top] < n) {
8421 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8423 rc = mdb_page_get(txn, pg, &mp, NULL);
8428 mc.mc_ki[mc.mc_top] = 0;
8429 if (IS_BRANCH(mp)) {
8430 /* Whenever we advance to a sibling branch page,
8431 * we must proceed all the way down to its first leaf.
8433 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8436 mc.mc_pg[mc.mc_top] = mp;
8440 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8441 rc = mdb_env_cthr_toggle(my, 1);
8444 toggle = my->mc_toggle;
8446 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8447 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8448 mo->mp_pgno = my->mc_next_pgno++;
8449 my->mc_wlen[toggle] += my->mc_env->me_psize;
8451 /* Update parent if there is one */
8452 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8453 SETPGNO(ni, mo->mp_pgno);
8454 mdb_cursor_pop(&mc);
8456 /* Otherwise we're done */
8466 /** Copy environment with compaction. */
8468 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8473 MDB_txn *txn = NULL;
8478 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8479 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8480 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8481 if (my.mc_wbuf[0] == NULL)
8484 pthread_mutex_init(&my.mc_mutex, NULL);
8485 pthread_cond_init(&my.mc_cond, NULL);
8486 #ifdef HAVE_MEMALIGN
8487 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8488 if (my.mc_wbuf[0] == NULL)
8491 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8496 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8497 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8502 my.mc_next_pgno = 2;
8508 THREAD_CREATE(thr, mdb_env_copythr, &my);
8510 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8514 mp = (MDB_page *)my.mc_wbuf[0];
8515 memset(mp, 0, 2*env->me_psize);
8517 mp->mp_flags = P_META;
8518 mm = (MDB_meta *)METADATA(mp);
8519 mdb_env_init_meta0(env, mm);
8520 mm->mm_address = env->me_metas[0]->mm_address;
8522 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8524 mp->mp_flags = P_META;
8525 *(MDB_meta *)METADATA(mp) = *mm;
8526 mm = (MDB_meta *)METADATA(mp);
8528 /* Count the number of free pages, subtract from lastpg to find
8529 * number of active pages
8532 MDB_ID freecount = 0;
8535 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8536 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8537 freecount += *(MDB_ID *)data.mv_data;
8538 freecount += txn->mt_dbs[0].md_branch_pages +
8539 txn->mt_dbs[0].md_leaf_pages +
8540 txn->mt_dbs[0].md_overflow_pages;
8542 /* Set metapage 1 */
8543 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8544 mm->mm_dbs[1] = txn->mt_dbs[1];
8545 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8548 my.mc_wlen[0] = env->me_psize * 2;
8550 pthread_mutex_lock(&my.mc_mutex);
8552 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8553 pthread_mutex_unlock(&my.mc_mutex);
8554 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8555 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8556 rc = mdb_env_cthr_toggle(&my, 1);
8557 mdb_env_cthr_toggle(&my, -1);
8558 pthread_mutex_lock(&my.mc_mutex);
8560 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8561 pthread_mutex_unlock(&my.mc_mutex);
8566 CloseHandle(my.mc_cond);
8567 CloseHandle(my.mc_mutex);
8568 _aligned_free(my.mc_wbuf[0]);
8570 pthread_cond_destroy(&my.mc_cond);
8571 pthread_mutex_destroy(&my.mc_mutex);
8572 free(my.mc_wbuf[0]);
8577 /** Copy environment as-is. */
8579 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8581 MDB_txn *txn = NULL;
8587 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8591 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8594 /* Do the lock/unlock of the reader mutex before starting the
8595 * write txn. Otherwise other read txns could block writers.
8597 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8602 /* We must start the actual read txn after blocking writers */
8603 mdb_txn_reset0(txn, "reset-stage1");
8605 /* Temporarily block writers until we snapshot the meta pages */
8608 rc = mdb_txn_renew0(txn);
8610 UNLOCK_MUTEX_W(env);
8615 wsize = env->me_psize * 2;
8619 DO_WRITE(rc, fd, ptr, w2, len);
8623 } else if (len > 0) {
8629 /* Non-blocking or async handles are not supported */
8635 UNLOCK_MUTEX_W(env);
8640 w2 = txn->mt_next_pgno * env->me_psize;
8643 LARGE_INTEGER fsize;
8644 GetFileSizeEx(env->me_fd, &fsize);
8645 if (w2 > fsize.QuadPart)
8646 w2 = fsize.QuadPart;
8651 fstat(env->me_fd, &st);
8652 if (w2 > (size_t)st.st_size)
8658 if (wsize > MAX_WRITE)
8662 DO_WRITE(rc, fd, ptr, w2, len);
8666 } else if (len > 0) {
8683 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8685 if (flags & MDB_CP_COMPACT)
8686 return mdb_env_copyfd1(env, fd);
8688 return mdb_env_copyfd0(env, fd);
8692 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8694 return mdb_env_copyfd2(env, fd, 0);
8698 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8702 HANDLE newfd = INVALID_HANDLE_VALUE;
8704 if (env->me_flags & MDB_NOSUBDIR) {
8705 lpath = (char *)path;
8708 len += sizeof(DATANAME);
8709 lpath = malloc(len);
8712 sprintf(lpath, "%s" DATANAME, path);
8715 /* The destination path must exist, but the destination file must not.
8716 * We don't want the OS to cache the writes, since the source data is
8717 * already in the OS cache.
8720 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8721 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8723 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8725 if (newfd == INVALID_HANDLE_VALUE) {
8730 if (env->me_psize >= env->me_os_psize) {
8732 /* Set O_DIRECT if the file system supports it */
8733 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8734 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8736 #ifdef F_NOCACHE /* __APPLE__ */
8737 rc = fcntl(newfd, F_NOCACHE, 1);
8745 rc = mdb_env_copyfd2(env, newfd, flags);
8748 if (!(env->me_flags & MDB_NOSUBDIR))
8750 if (newfd != INVALID_HANDLE_VALUE)
8751 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8758 mdb_env_copy(MDB_env *env, const char *path)
8760 return mdb_env_copy2(env, path, 0);
8764 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8766 if ((flag & CHANGEABLE) != flag)
8769 env->me_flags |= flag;
8771 env->me_flags &= ~flag;
8776 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8781 *arg = env->me_flags;
8786 mdb_env_set_userctx(MDB_env *env, void *ctx)
8790 env->me_userctx = ctx;
8795 mdb_env_get_userctx(MDB_env *env)
8797 return env ? env->me_userctx : NULL;
8801 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8806 env->me_assert_func = func;
8812 mdb_env_get_path(MDB_env *env, const char **arg)
8817 *arg = env->me_path;
8822 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8831 /** Common code for #mdb_stat() and #mdb_env_stat().
8832 * @param[in] env the environment to operate in.
8833 * @param[in] db the #MDB_db record containing the stats to return.
8834 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8835 * @return 0, this function always succeeds.
8838 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8840 arg->ms_psize = env->me_psize;
8841 arg->ms_depth = db->md_depth;
8842 arg->ms_branch_pages = db->md_branch_pages;
8843 arg->ms_leaf_pages = db->md_leaf_pages;
8844 arg->ms_overflow_pages = db->md_overflow_pages;
8845 arg->ms_entries = db->md_entries;
8851 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8855 if (env == NULL || arg == NULL)
8858 toggle = mdb_env_pick_meta(env);
8860 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8864 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8868 if (env == NULL || arg == NULL)
8871 toggle = mdb_env_pick_meta(env);
8872 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8873 arg->me_mapsize = env->me_mapsize;
8874 arg->me_maxreaders = env->me_maxreaders;
8876 /* me_numreaders may be zero if this process never used any readers. Use
8877 * the shared numreader count if it exists.
8879 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8881 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8882 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8886 /** Set the default comparison functions for a database.
8887 * Called immediately after a database is opened to set the defaults.
8888 * The user can then override them with #mdb_set_compare() or
8889 * #mdb_set_dupsort().
8890 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8891 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8894 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8896 uint16_t f = txn->mt_dbs[dbi].md_flags;
8898 txn->mt_dbxs[dbi].md_cmp =
8899 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8900 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8902 txn->mt_dbxs[dbi].md_dcmp =
8903 !(f & MDB_DUPSORT) ? 0 :
8904 ((f & MDB_INTEGERDUP)
8905 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8906 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8909 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8915 int rc, dbflag, exact;
8916 unsigned int unused = 0, seq;
8919 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8920 mdb_default_cmp(txn, FREE_DBI);
8923 if ((flags & VALID_FLAGS) != flags)
8925 if (txn->mt_flags & MDB_TXN_ERROR)
8931 if (flags & PERSISTENT_FLAGS) {
8932 uint16_t f2 = flags & PERSISTENT_FLAGS;
8933 /* make sure flag changes get committed */
8934 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8935 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8936 txn->mt_flags |= MDB_TXN_DIRTY;
8939 mdb_default_cmp(txn, MAIN_DBI);
8943 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8944 mdb_default_cmp(txn, MAIN_DBI);
8947 /* Is the DB already open? */
8949 for (i=2; i<txn->mt_numdbs; i++) {
8950 if (!txn->mt_dbxs[i].md_name.mv_size) {
8951 /* Remember this free slot */
8952 if (!unused) unused = i;
8955 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8956 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8962 /* If no free slot and max hit, fail */
8963 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8964 return MDB_DBS_FULL;
8966 /* Cannot mix named databases with some mainDB flags */
8967 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8968 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8970 /* Find the DB info */
8971 dbflag = DB_NEW|DB_VALID;
8974 key.mv_data = (void *)name;
8975 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8976 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8977 if (rc == MDB_SUCCESS) {
8978 /* make sure this is actually a DB */
8979 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8980 if (!(node->mn_flags & F_SUBDATA))
8981 return MDB_INCOMPATIBLE;
8982 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8983 /* Create if requested */
8984 data.mv_size = sizeof(MDB_db);
8985 data.mv_data = &dummy;
8986 memset(&dummy, 0, sizeof(dummy));
8987 dummy.md_root = P_INVALID;
8988 dummy.md_flags = flags & PERSISTENT_FLAGS;
8989 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8993 /* OK, got info, add to table */
8994 if (rc == MDB_SUCCESS) {
8995 unsigned int slot = unused ? unused : txn->mt_numdbs;
8996 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8997 txn->mt_dbxs[slot].md_name.mv_size = len;
8998 txn->mt_dbxs[slot].md_rel = NULL;
8999 txn->mt_dbflags[slot] = dbflag;
9000 /* txn-> and env-> are the same in read txns, use
9001 * tmp variable to avoid undefined assignment
9003 seq = ++txn->mt_env->me_dbiseqs[slot];
9004 txn->mt_dbiseqs[slot] = seq;
9006 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9008 mdb_default_cmp(txn, slot);
9017 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9019 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9022 if (txn->mt_flags & MDB_TXN_ERROR)
9025 if (txn->mt_dbflags[dbi] & DB_STALE) {
9028 /* Stale, must read the DB's root. cursor_init does it for us. */
9029 mdb_cursor_init(&mc, txn, dbi, &mx);
9031 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9034 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9037 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9039 ptr = env->me_dbxs[dbi].md_name.mv_data;
9040 /* If there was no name, this was already closed */
9042 env->me_dbxs[dbi].md_name.mv_data = NULL;
9043 env->me_dbxs[dbi].md_name.mv_size = 0;
9044 env->me_dbflags[dbi] = 0;
9045 env->me_dbiseqs[dbi]++;
9050 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9052 /* We could return the flags for the FREE_DBI too but what's the point? */
9053 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9055 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9059 /** Add all the DB's pages to the free list.
9060 * @param[in] mc Cursor on the DB to free.
9061 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9062 * @return 0 on success, non-zero on failure.
9065 mdb_drop0(MDB_cursor *mc, int subs)
9069 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9070 if (rc == MDB_SUCCESS) {
9071 MDB_txn *txn = mc->mc_txn;
9076 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9077 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9080 mdb_cursor_copy(mc, &mx);
9081 while (mc->mc_snum > 0) {
9082 MDB_page *mp = mc->mc_pg[mc->mc_top];
9083 unsigned n = NUMKEYS(mp);
9085 for (i=0; i<n; i++) {
9086 ni = NODEPTR(mp, i);
9087 if (ni->mn_flags & F_BIGDATA) {
9090 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9091 rc = mdb_page_get(txn, pg, &omp, NULL);
9094 mdb_cassert(mc, IS_OVERFLOW(omp));
9095 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9099 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9100 mdb_xcursor_init1(mc, ni);
9101 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9107 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9109 for (i=0; i<n; i++) {
9111 ni = NODEPTR(mp, i);
9114 mdb_midl_xappend(txn->mt_free_pgs, pg);
9119 mc->mc_ki[mc->mc_top] = i;
9120 rc = mdb_cursor_sibling(mc, 1);
9122 if (rc != MDB_NOTFOUND)
9124 /* no more siblings, go back to beginning
9125 * of previous level.
9129 for (i=1; i<mc->mc_snum; i++) {
9131 mc->mc_pg[i] = mx.mc_pg[i];
9136 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9139 txn->mt_flags |= MDB_TXN_ERROR;
9140 } else if (rc == MDB_NOTFOUND) {
9146 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9148 MDB_cursor *mc, *m2;
9151 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9154 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9157 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9160 rc = mdb_cursor_open(txn, dbi, &mc);
9164 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9165 /* Invalidate the dropped DB's cursors */
9166 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9167 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9171 /* Can't delete the main DB */
9172 if (del && dbi > MAIN_DBI) {
9173 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9175 txn->mt_dbflags[dbi] = DB_STALE;
9176 mdb_dbi_close(txn->mt_env, dbi);
9178 txn->mt_flags |= MDB_TXN_ERROR;
9181 /* reset the DB record, mark it dirty */
9182 txn->mt_dbflags[dbi] |= DB_DIRTY;
9183 txn->mt_dbs[dbi].md_depth = 0;
9184 txn->mt_dbs[dbi].md_branch_pages = 0;
9185 txn->mt_dbs[dbi].md_leaf_pages = 0;
9186 txn->mt_dbs[dbi].md_overflow_pages = 0;
9187 txn->mt_dbs[dbi].md_entries = 0;
9188 txn->mt_dbs[dbi].md_root = P_INVALID;
9190 txn->mt_flags |= MDB_TXN_DIRTY;
9193 mdb_cursor_close(mc);
9197 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9199 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9202 txn->mt_dbxs[dbi].md_cmp = cmp;
9206 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9208 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9211 txn->mt_dbxs[dbi].md_dcmp = cmp;
9215 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9217 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9220 txn->mt_dbxs[dbi].md_rel = rel;
9224 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9226 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9229 txn->mt_dbxs[dbi].md_relctx = ctx;
9234 mdb_env_get_maxkeysize(MDB_env *env)
9236 return ENV_MAXKEY(env);
9240 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9242 unsigned int i, rdrs;
9245 int rc = 0, first = 1;
9249 if (!env->me_txns) {
9250 return func("(no reader locks)\n", ctx);
9252 rdrs = env->me_txns->mti_numreaders;
9253 mr = env->me_txns->mti_readers;
9254 for (i=0; i<rdrs; i++) {
9256 txnid_t txnid = mr[i].mr_txnid;
9257 sprintf(buf, txnid == (txnid_t)-1 ?
9258 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9259 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9262 rc = func(" pid thread txnid\n", ctx);
9266 rc = func(buf, ctx);
9272 rc = func("(no active readers)\n", ctx);
9277 /** Insert pid into list if not already present.
9278 * return -1 if already present.
9281 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9283 /* binary search of pid in list */
9285 unsigned cursor = 1;
9287 unsigned n = ids[0];
9290 unsigned pivot = n >> 1;
9291 cursor = base + pivot + 1;
9292 val = pid - ids[cursor];
9297 } else if ( val > 0 ) {
9302 /* found, so it's a duplicate */
9311 for (n = ids[0]; n > cursor; n--)
9318 mdb_reader_check(MDB_env *env, int *dead)
9320 unsigned int i, j, rdrs;
9322 MDB_PID_T *pids, pid;
9331 rdrs = env->me_txns->mti_numreaders;
9332 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9336 mr = env->me_txns->mti_readers;
9337 for (i=0; i<rdrs; i++) {
9338 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9340 if (mdb_pid_insert(pids, pid) == 0) {
9341 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9343 /* Recheck, a new process may have reused pid */
9344 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9345 for (j=i; j<rdrs; j++)
9346 if (mr[j].mr_pid == pid) {
9347 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9348 (unsigned) pid, mr[j].mr_txnid));
9353 UNLOCK_MUTEX_R(env);