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 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1114 size_t me_mapsize; /**< size of the data memory map */
1115 off_t me_size; /**< current file size */
1116 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1117 MDB_dbx *me_dbxs; /**< array of static DB info */
1118 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1119 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1120 pthread_key_t me_txkey; /**< thread-key for readers */
1121 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1122 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1123 # define me_pglast me_pgstate.mf_pglast
1124 # define me_pghead me_pgstate.mf_pghead
1125 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1126 /** IDL of pages that became unused in a write txn */
1127 MDB_IDL me_free_pgs;
1128 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1129 MDB_ID2L me_dirty_list;
1130 /** Max number of freelist items that can fit in a single overflow page */
1132 /** Max size of a node on a page */
1133 unsigned int me_nodemax;
1134 #if !(MDB_MAXKEYSIZE)
1135 unsigned int me_maxkey; /**< max size of a key */
1137 int me_live_reader; /**< have liveness lock in reader table */
1139 int me_pidquery; /**< Used in OpenProcess */
1140 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1142 #elif defined(MDB_USE_POSIX_SEM)
1143 sem_t *me_rmutex; /* Shared mutexes are not supported */
1146 void *me_userctx; /**< User-settable context */
1147 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1150 /** Nested transaction */
1151 typedef struct MDB_ntxn {
1152 MDB_txn mnt_txn; /**< the transaction */
1153 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1156 /** max number of pages to commit in one writev() call */
1157 #define MDB_COMMIT_PAGES 64
1158 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1159 #undef MDB_COMMIT_PAGES
1160 #define MDB_COMMIT_PAGES IOV_MAX
1163 /** max bytes to write in one call */
1164 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1166 /** Check \b txn and \b dbi arguments to a function */
1167 #define TXN_DBI_EXIST(txn, dbi) \
1168 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1170 /** Check for misused \b dbi handles */
1171 #define TXN_DBI_CHANGED(txn, dbi) \
1172 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1174 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1175 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1176 static int mdb_page_touch(MDB_cursor *mc);
1178 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1179 static int mdb_page_search_root(MDB_cursor *mc,
1180 MDB_val *key, int modify);
1181 #define MDB_PS_MODIFY 1
1182 #define MDB_PS_ROOTONLY 2
1183 #define MDB_PS_FIRST 4
1184 #define MDB_PS_LAST 8
1185 static int mdb_page_search(MDB_cursor *mc,
1186 MDB_val *key, int flags);
1187 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1189 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1190 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1191 pgno_t newpgno, unsigned int nflags);
1193 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1194 static int mdb_env_pick_meta(const MDB_env *env);
1195 static int mdb_env_write_meta(MDB_txn *txn);
1196 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1197 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1199 static void mdb_env_close0(MDB_env *env, int excl);
1201 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1202 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1203 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1204 static void mdb_node_del(MDB_cursor *mc, int ksize);
1205 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1206 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1207 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1208 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1209 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1211 static int mdb_rebalance(MDB_cursor *mc);
1212 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1214 static void mdb_cursor_pop(MDB_cursor *mc);
1215 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1217 static int mdb_cursor_del0(MDB_cursor *mc);
1218 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1219 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1220 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1221 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1222 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1224 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1225 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1227 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1228 static void mdb_xcursor_init0(MDB_cursor *mc);
1229 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1231 static int mdb_drop0(MDB_cursor *mc, int subs);
1232 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1235 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1239 static SECURITY_DESCRIPTOR mdb_null_sd;
1240 static SECURITY_ATTRIBUTES mdb_all_sa;
1241 static int mdb_sec_inited;
1244 /** Return the library version info. */
1246 mdb_version(int *major, int *minor, int *patch)
1248 if (major) *major = MDB_VERSION_MAJOR;
1249 if (minor) *minor = MDB_VERSION_MINOR;
1250 if (patch) *patch = MDB_VERSION_PATCH;
1251 return MDB_VERSION_STRING;
1254 /** Table of descriptions for LMDB @ref errors */
1255 static char *const mdb_errstr[] = {
1256 "MDB_KEYEXIST: Key/data pair already exists",
1257 "MDB_NOTFOUND: No matching key/data pair found",
1258 "MDB_PAGE_NOTFOUND: Requested page not found",
1259 "MDB_CORRUPTED: Located page was wrong type",
1260 "MDB_PANIC: Update of meta page failed",
1261 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1262 "MDB_INVALID: File is not an LMDB file",
1263 "MDB_MAP_FULL: Environment mapsize limit reached",
1264 "MDB_DBS_FULL: Environment maxdbs limit reached",
1265 "MDB_READERS_FULL: Environment maxreaders limit reached",
1266 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1267 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1268 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1269 "MDB_PAGE_FULL: Internal error - page has no more space",
1270 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1271 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1272 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1273 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1274 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1275 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1279 mdb_strerror(int err)
1282 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1283 * This works as long as no function between the call to mdb_strerror
1284 * and the actual use of the message uses more than 4K of stack.
1287 char buf[1024], *ptr = buf;
1291 return ("Successful return: 0");
1293 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1294 i = err - MDB_KEYEXIST;
1295 return mdb_errstr[i];
1299 /* These are the C-runtime error codes we use. The comment indicates
1300 * their numeric value, and the Win32 error they would correspond to
1301 * if the error actually came from a Win32 API. A major mess, we should
1302 * have used LMDB-specific error codes for everything.
1305 case ENOENT: /* 2, FILE_NOT_FOUND */
1306 case EIO: /* 5, ACCESS_DENIED */
1307 case ENOMEM: /* 12, INVALID_ACCESS */
1308 case EACCES: /* 13, INVALID_DATA */
1309 case EBUSY: /* 16, CURRENT_DIRECTORY */
1310 case EINVAL: /* 22, BAD_COMMAND */
1311 case ENOSPC: /* 28, OUT_OF_PAPER */
1312 return strerror(err);
1317 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1318 FORMAT_MESSAGE_IGNORE_INSERTS,
1319 NULL, err, 0, ptr, sizeof(buf), pad);
1322 return strerror(err);
1326 /** assert(3) variant in cursor context */
1327 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1328 /** assert(3) variant in transaction context */
1329 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1330 /** assert(3) variant in environment context */
1331 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1334 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1335 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1338 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1339 const char *func, const char *file, int line)
1342 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1343 file, line, expr_txt, func);
1344 if (env->me_assert_func)
1345 env->me_assert_func(env, buf);
1346 fprintf(stderr, "%s\n", buf);
1350 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1354 /** Return the page number of \b mp which may be sub-page, for debug output */
1356 mdb_dbg_pgno(MDB_page *mp)
1359 COPY_PGNO(ret, mp->mp_pgno);
1363 /** Display a key in hexadecimal and return the address of the result.
1364 * @param[in] key the key to display
1365 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1366 * @return The key in hexadecimal form.
1369 mdb_dkey(MDB_val *key, char *buf)
1372 unsigned char *c = key->mv_data;
1378 if (key->mv_size > DKBUF_MAXKEYSIZE)
1379 return "MDB_MAXKEYSIZE";
1380 /* may want to make this a dynamic check: if the key is mostly
1381 * printable characters, print it as-is instead of converting to hex.
1385 for (i=0; i<key->mv_size; i++)
1386 ptr += sprintf(ptr, "%02x", *c++);
1388 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1394 mdb_leafnode_type(MDB_node *n)
1396 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1397 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1398 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1401 /** Display all the keys in the page. */
1403 mdb_page_list(MDB_page *mp)
1405 pgno_t pgno = mdb_dbg_pgno(mp);
1406 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1408 unsigned int i, nkeys, nsize, total = 0;
1412 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1413 case P_BRANCH: type = "Branch page"; break;
1414 case P_LEAF: type = "Leaf page"; break;
1415 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1416 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1417 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1419 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1420 pgno, mp->mp_pages, state);
1423 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1424 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1427 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1431 nkeys = NUMKEYS(mp);
1432 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1434 for (i=0; i<nkeys; i++) {
1435 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1436 key.mv_size = nsize = mp->mp_pad;
1437 key.mv_data = LEAF2KEY(mp, i, nsize);
1439 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1442 node = NODEPTR(mp, i);
1443 key.mv_size = node->mn_ksize;
1444 key.mv_data = node->mn_data;
1445 nsize = NODESIZE + key.mv_size;
1446 if (IS_BRANCH(mp)) {
1447 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1451 if (F_ISSET(node->mn_flags, F_BIGDATA))
1452 nsize += sizeof(pgno_t);
1454 nsize += NODEDSZ(node);
1456 nsize += sizeof(indx_t);
1457 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1458 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1460 total = EVEN(total);
1462 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1463 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1467 mdb_cursor_chk(MDB_cursor *mc)
1473 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1474 for (i=0; i<mc->mc_top; i++) {
1476 node = NODEPTR(mp, mc->mc_ki[i]);
1477 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1480 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1486 /** Count all the pages in each DB and in the freelist
1487 * and make sure it matches the actual number of pages
1489 * All named DBs must be open for a correct count.
1491 static void mdb_audit(MDB_txn *txn)
1495 MDB_ID freecount, count;
1500 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1501 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1502 freecount += *(MDB_ID *)data.mv_data;
1503 mdb_tassert(txn, rc == MDB_NOTFOUND);
1506 for (i = 0; i<txn->mt_numdbs; i++) {
1508 if (!(txn->mt_dbflags[i] & DB_VALID))
1510 mdb_cursor_init(&mc, txn, i, &mx);
1511 if (txn->mt_dbs[i].md_root == P_INVALID)
1513 count += txn->mt_dbs[i].md_branch_pages +
1514 txn->mt_dbs[i].md_leaf_pages +
1515 txn->mt_dbs[i].md_overflow_pages;
1516 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1517 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1518 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1521 mp = mc.mc_pg[mc.mc_top];
1522 for (j=0; j<NUMKEYS(mp); j++) {
1523 MDB_node *leaf = NODEPTR(mp, j);
1524 if (leaf->mn_flags & F_SUBDATA) {
1526 memcpy(&db, NODEDATA(leaf), sizeof(db));
1527 count += db.md_branch_pages + db.md_leaf_pages +
1528 db.md_overflow_pages;
1532 mdb_tassert(txn, rc == MDB_NOTFOUND);
1535 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1536 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1537 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1543 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1545 return txn->mt_dbxs[dbi].md_cmp(a, b);
1549 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1551 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1554 /** Allocate memory for a page.
1555 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1558 mdb_page_malloc(MDB_txn *txn, unsigned num)
1560 MDB_env *env = txn->mt_env;
1561 MDB_page *ret = env->me_dpages;
1562 size_t psize = env->me_psize, sz = psize, off;
1563 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1564 * For a single page alloc, we init everything after the page header.
1565 * For multi-page, we init the final page; if the caller needed that
1566 * many pages they will be filling in at least up to the last page.
1570 VGMEMP_ALLOC(env, ret, sz);
1571 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1572 env->me_dpages = ret->mp_next;
1575 psize -= off = PAGEHDRSZ;
1580 if ((ret = malloc(sz)) != NULL) {
1581 VGMEMP_ALLOC(env, ret, sz);
1582 if (!(env->me_flags & MDB_NOMEMINIT)) {
1583 memset((char *)ret + off, 0, psize);
1587 txn->mt_flags |= MDB_TXN_ERROR;
1591 /** Free a single page.
1592 * Saves single pages to a list, for future reuse.
1593 * (This is not used for multi-page overflow pages.)
1596 mdb_page_free(MDB_env *env, MDB_page *mp)
1598 mp->mp_next = env->me_dpages;
1599 VGMEMP_FREE(env, mp);
1600 env->me_dpages = mp;
1603 /** Free a dirty page */
1605 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1607 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1608 mdb_page_free(env, dp);
1610 /* large pages just get freed directly */
1611 VGMEMP_FREE(env, dp);
1616 /** Return all dirty pages to dpage list */
1618 mdb_dlist_free(MDB_txn *txn)
1620 MDB_env *env = txn->mt_env;
1621 MDB_ID2L dl = txn->mt_u.dirty_list;
1622 unsigned i, n = dl[0].mid;
1624 for (i = 1; i <= n; i++) {
1625 mdb_dpage_free(env, dl[i].mptr);
1630 /** Loosen or free a single page.
1631 * Saves single pages to a list for future reuse
1632 * in this same txn. It has been pulled from the freeDB
1633 * and already resides on the dirty list, but has been
1634 * deleted. Use these pages first before pulling again
1637 * If the page wasn't dirtied in this txn, just add it
1638 * to this txn's free list.
1641 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1644 pgno_t pgno = mp->mp_pgno;
1645 MDB_txn *txn = mc->mc_txn;
1647 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1648 if (txn->mt_parent) {
1649 MDB_ID2 *dl = txn->mt_u.dirty_list;
1650 /* If txn has a parent, make sure the page is in our
1654 unsigned x = mdb_mid2l_search(dl, pgno);
1655 if (x <= dl[0].mid && dl[x].mid == pgno) {
1656 if (mp != dl[x].mptr) { /* bad cursor? */
1657 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1658 txn->mt_flags |= MDB_TXN_ERROR;
1659 return MDB_CORRUPTED;
1666 /* no parent txn, so it's just ours */
1671 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1673 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1674 txn->mt_loose_pgs = mp;
1675 txn->mt_loose_count++;
1676 mp->mp_flags |= P_LOOSE;
1678 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1686 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1687 * @param[in] mc A cursor handle for the current operation.
1688 * @param[in] pflags Flags of the pages to update:
1689 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1690 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1691 * @return 0 on success, non-zero on failure.
1694 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1696 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1697 MDB_txn *txn = mc->mc_txn;
1703 int rc = MDB_SUCCESS, level;
1705 /* Mark pages seen by cursors */
1706 if (mc->mc_flags & C_UNTRACK)
1707 mc = NULL; /* will find mc in mt_cursors */
1708 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1709 for (; mc; mc=mc->mc_next) {
1710 if (!(mc->mc_flags & C_INITIALIZED))
1712 for (m3 = mc;; m3 = &mx->mx_cursor) {
1714 for (j=0; j<m3->mc_snum; j++) {
1716 if ((mp->mp_flags & Mask) == pflags)
1717 mp->mp_flags ^= P_KEEP;
1719 mx = m3->mc_xcursor;
1720 /* Proceed to mx if it is at a sub-database */
1721 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1723 if (! (mp && (mp->mp_flags & P_LEAF)))
1725 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1726 if (!(leaf->mn_flags & F_SUBDATA))
1735 /* Mark dirty root pages */
1736 for (i=0; i<txn->mt_numdbs; i++) {
1737 if (txn->mt_dbflags[i] & DB_DIRTY) {
1738 pgno_t pgno = txn->mt_dbs[i].md_root;
1739 if (pgno == P_INVALID)
1741 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1743 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1744 dp->mp_flags ^= P_KEEP;
1752 static int mdb_page_flush(MDB_txn *txn, int keep);
1754 /** Spill pages from the dirty list back to disk.
1755 * This is intended to prevent running into #MDB_TXN_FULL situations,
1756 * but note that they may still occur in a few cases:
1757 * 1) our estimate of the txn size could be too small. Currently this
1758 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1759 * 2) child txns may run out of space if their parents dirtied a
1760 * lot of pages and never spilled them. TODO: we probably should do
1761 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1762 * the parent's dirty_room is below a given threshold.
1764 * Otherwise, if not using nested txns, it is expected that apps will
1765 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1766 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1767 * If the txn never references them again, they can be left alone.
1768 * If the txn only reads them, they can be used without any fuss.
1769 * If the txn writes them again, they can be dirtied immediately without
1770 * going thru all of the work of #mdb_page_touch(). Such references are
1771 * handled by #mdb_page_unspill().
1773 * Also note, we never spill DB root pages, nor pages of active cursors,
1774 * because we'll need these back again soon anyway. And in nested txns,
1775 * we can't spill a page in a child txn if it was already spilled in a
1776 * parent txn. That would alter the parent txns' data even though
1777 * the child hasn't committed yet, and we'd have no way to undo it if
1778 * the child aborted.
1780 * @param[in] m0 cursor A cursor handle identifying the transaction and
1781 * database for which we are checking space.
1782 * @param[in] key For a put operation, the key being stored.
1783 * @param[in] data For a put operation, the data being stored.
1784 * @return 0 on success, non-zero on failure.
1787 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1789 MDB_txn *txn = m0->mc_txn;
1791 MDB_ID2L dl = txn->mt_u.dirty_list;
1792 unsigned int i, j, need;
1795 if (m0->mc_flags & C_SUB)
1798 /* Estimate how much space this op will take */
1799 i = m0->mc_db->md_depth;
1800 /* Named DBs also dirty the main DB */
1801 if (m0->mc_dbi > MAIN_DBI)
1802 i += txn->mt_dbs[MAIN_DBI].md_depth;
1803 /* For puts, roughly factor in the key+data size */
1805 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1806 i += i; /* double it for good measure */
1809 if (txn->mt_dirty_room > i)
1812 if (!txn->mt_spill_pgs) {
1813 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1814 if (!txn->mt_spill_pgs)
1817 /* purge deleted slots */
1818 MDB_IDL sl = txn->mt_spill_pgs;
1819 unsigned int num = sl[0];
1821 for (i=1; i<=num; i++) {
1828 /* Preserve pages which may soon be dirtied again */
1829 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1832 /* Less aggressive spill - we originally spilled the entire dirty list,
1833 * with a few exceptions for cursor pages and DB root pages. But this
1834 * turns out to be a lot of wasted effort because in a large txn many
1835 * of those pages will need to be used again. So now we spill only 1/8th
1836 * of the dirty pages. Testing revealed this to be a good tradeoff,
1837 * better than 1/2, 1/4, or 1/10.
1839 if (need < MDB_IDL_UM_MAX / 8)
1840 need = MDB_IDL_UM_MAX / 8;
1842 /* Save the page IDs of all the pages we're flushing */
1843 /* flush from the tail forward, this saves a lot of shifting later on. */
1844 for (i=dl[0].mid; i && need; i--) {
1845 MDB_ID pn = dl[i].mid << 1;
1847 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1849 /* Can't spill twice, make sure it's not already in a parent's
1852 if (txn->mt_parent) {
1854 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1855 if (tx2->mt_spill_pgs) {
1856 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1857 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1858 dp->mp_flags |= P_KEEP;
1866 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1870 mdb_midl_sort(txn->mt_spill_pgs);
1872 /* Flush the spilled part of dirty list */
1873 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1876 /* Reset any dirty pages we kept that page_flush didn't see */
1877 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1880 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1884 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1886 mdb_find_oldest(MDB_txn *txn)
1889 txnid_t mr, oldest = txn->mt_txnid - 1;
1890 if (txn->mt_env->me_txns) {
1891 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1892 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1903 /** Add a page to the txn's dirty list */
1905 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1908 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1910 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1911 insert = mdb_mid2l_append;
1913 insert = mdb_mid2l_insert;
1915 mid.mid = mp->mp_pgno;
1917 rc = insert(txn->mt_u.dirty_list, &mid);
1918 mdb_tassert(txn, rc == 0);
1919 txn->mt_dirty_room--;
1922 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1923 * me_pghead and mt_next_pgno.
1925 * If there are free pages available from older transactions, they
1926 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1927 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1928 * and move me_pglast to say which records were consumed. Only this
1929 * function can create me_pghead and move me_pglast/mt_next_pgno.
1930 * @param[in] mc cursor A cursor handle identifying the transaction and
1931 * database for which we are allocating.
1932 * @param[in] num the number of pages to allocate.
1933 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1934 * will always be satisfied by a single contiguous chunk of memory.
1935 * @return 0 on success, non-zero on failure.
1938 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1940 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1941 /* Get at most <Max_retries> more freeDB records once me_pghead
1942 * has enough pages. If not enough, use new pages from the map.
1943 * If <Paranoid> and mc is updating the freeDB, only get new
1944 * records if me_pghead is empty. Then the freelist cannot play
1945 * catch-up with itself by growing while trying to save it.
1947 enum { Paranoid = 1, Max_retries = 500 };
1949 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1951 int rc, retry = num * 60;
1952 MDB_txn *txn = mc->mc_txn;
1953 MDB_env *env = txn->mt_env;
1954 pgno_t pgno, *mop = env->me_pghead;
1955 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1957 txnid_t oldest = 0, last;
1962 /* If there are any loose pages, just use them */
1963 if (num == 1 && txn->mt_loose_pgs) {
1964 np = txn->mt_loose_pgs;
1965 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1966 txn->mt_loose_count--;
1967 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1975 /* If our dirty list is already full, we can't do anything */
1976 if (txn->mt_dirty_room == 0) {
1981 for (op = MDB_FIRST;; op = MDB_NEXT) {
1986 /* Seek a big enough contiguous page range. Prefer
1987 * pages at the tail, just truncating the list.
1993 if (mop[i-n2] == pgno+n2)
2000 if (op == MDB_FIRST) { /* 1st iteration */
2001 /* Prepare to fetch more and coalesce */
2002 last = env->me_pglast;
2003 oldest = env->me_pgoldest;
2004 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2007 key.mv_data = &last; /* will look up last+1 */
2008 key.mv_size = sizeof(last);
2010 if (Paranoid && mc->mc_dbi == FREE_DBI)
2013 if (Paranoid && retry < 0 && mop_len)
2017 /* Do not fetch more if the record will be too recent */
2018 if (oldest <= last) {
2020 oldest = mdb_find_oldest(txn);
2021 env->me_pgoldest = oldest;
2027 rc = mdb_cursor_get(&m2, &key, NULL, op);
2029 if (rc == MDB_NOTFOUND)
2033 last = *(txnid_t*)key.mv_data;
2034 if (oldest <= last) {
2036 oldest = mdb_find_oldest(txn);
2037 env->me_pgoldest = oldest;
2043 np = m2.mc_pg[m2.mc_top];
2044 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2045 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2048 idl = (MDB_ID *) data.mv_data;
2051 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2056 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2058 mop = env->me_pghead;
2060 env->me_pglast = last;
2062 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2063 last, txn->mt_dbs[FREE_DBI].md_root, i));
2065 DPRINTF(("IDL %"Z"u", idl[j]));
2067 /* Merge in descending sorted order */
2068 mdb_midl_xmerge(mop, idl);
2072 /* Use new pages from the map when nothing suitable in the freeDB */
2074 pgno = txn->mt_next_pgno;
2075 if (pgno + num >= env->me_maxpg) {
2076 DPUTS("DB size maxed out");
2082 if (env->me_flags & MDB_WRITEMAP) {
2083 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2085 if (!(np = mdb_page_malloc(txn, num))) {
2091 mop[0] = mop_len -= num;
2092 /* Move any stragglers down */
2093 for (j = i-num; j < mop_len; )
2094 mop[++j] = mop[++i];
2096 txn->mt_next_pgno = pgno + num;
2099 mdb_page_dirty(txn, np);
2105 txn->mt_flags |= MDB_TXN_ERROR;
2109 /** Copy the used portions of a non-overflow page.
2110 * @param[in] dst page to copy into
2111 * @param[in] src page to copy from
2112 * @param[in] psize size of a page
2115 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2117 enum { Align = sizeof(pgno_t) };
2118 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2120 /* If page isn't full, just copy the used portion. Adjust
2121 * alignment so memcpy may copy words instead of bytes.
2123 if ((unused &= -Align) && !IS_LEAF2(src)) {
2124 upper = (upper + PAGEBASE) & -Align;
2125 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2126 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2129 memcpy(dst, src, psize - unused);
2133 /** Pull a page off the txn's spill list, if present.
2134 * If a page being referenced was spilled to disk in this txn, bring
2135 * it back and make it dirty/writable again.
2136 * @param[in] txn the transaction handle.
2137 * @param[in] mp the page being referenced. It must not be dirty.
2138 * @param[out] ret the writable page, if any. ret is unchanged if
2139 * mp wasn't spilled.
2142 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2144 MDB_env *env = txn->mt_env;
2147 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2149 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2150 if (!tx2->mt_spill_pgs)
2152 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2153 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2156 if (txn->mt_dirty_room == 0)
2157 return MDB_TXN_FULL;
2158 if (IS_OVERFLOW(mp))
2162 if (env->me_flags & MDB_WRITEMAP) {
2165 np = mdb_page_malloc(txn, num);
2169 memcpy(np, mp, num * env->me_psize);
2171 mdb_page_copy(np, mp, env->me_psize);
2174 /* If in current txn, this page is no longer spilled.
2175 * If it happens to be the last page, truncate the spill list.
2176 * Otherwise mark it as deleted by setting the LSB.
2178 if (x == txn->mt_spill_pgs[0])
2179 txn->mt_spill_pgs[0]--;
2181 txn->mt_spill_pgs[x] |= 1;
2182 } /* otherwise, if belonging to a parent txn, the
2183 * page remains spilled until child commits
2186 mdb_page_dirty(txn, np);
2187 np->mp_flags |= P_DIRTY;
2195 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2196 * @param[in] mc cursor pointing to the page to be touched
2197 * @return 0 on success, non-zero on failure.
2200 mdb_page_touch(MDB_cursor *mc)
2202 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2203 MDB_txn *txn = mc->mc_txn;
2204 MDB_cursor *m2, *m3;
2208 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2209 if (txn->mt_flags & MDB_TXN_SPILLS) {
2211 rc = mdb_page_unspill(txn, mp, &np);
2217 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2218 (rc = mdb_page_alloc(mc, 1, &np)))
2221 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2222 mp->mp_pgno, pgno));
2223 mdb_cassert(mc, mp->mp_pgno != pgno);
2224 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2225 /* Update the parent page, if any, to point to the new page */
2227 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2228 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2229 SETPGNO(node, pgno);
2231 mc->mc_db->md_root = pgno;
2233 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2234 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2236 /* If txn has a parent, make sure the page is in our
2240 unsigned x = mdb_mid2l_search(dl, pgno);
2241 if (x <= dl[0].mid && dl[x].mid == pgno) {
2242 if (mp != dl[x].mptr) { /* bad cursor? */
2243 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2244 txn->mt_flags |= MDB_TXN_ERROR;
2245 return MDB_CORRUPTED;
2250 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2252 np = mdb_page_malloc(txn, 1);
2257 rc = mdb_mid2l_insert(dl, &mid);
2258 mdb_cassert(mc, rc == 0);
2263 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2265 np->mp_flags |= P_DIRTY;
2268 /* Adjust cursors pointing to mp */
2269 mc->mc_pg[mc->mc_top] = np;
2270 m2 = txn->mt_cursors[mc->mc_dbi];
2271 if (mc->mc_flags & C_SUB) {
2272 for (; m2; m2=m2->mc_next) {
2273 m3 = &m2->mc_xcursor->mx_cursor;
2274 if (m3->mc_snum < mc->mc_snum) continue;
2275 if (m3->mc_pg[mc->mc_top] == mp)
2276 m3->mc_pg[mc->mc_top] = np;
2279 for (; m2; m2=m2->mc_next) {
2280 if (m2->mc_snum < mc->mc_snum) continue;
2281 if (m2->mc_pg[mc->mc_top] == mp) {
2282 m2->mc_pg[mc->mc_top] = np;
2283 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2285 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2287 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2288 if (!(leaf->mn_flags & F_SUBDATA))
2289 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2297 txn->mt_flags |= MDB_TXN_ERROR;
2302 mdb_env_sync(MDB_env *env, int force)
2305 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2306 if (env->me_flags & MDB_WRITEMAP) {
2307 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2308 ? MS_ASYNC : MS_SYNC;
2309 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2312 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2316 if (MDB_FDATASYNC(env->me_fd))
2323 /** Back up parent txn's cursors, then grab the originals for tracking */
2325 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2327 MDB_cursor *mc, *bk;
2332 for (i = src->mt_numdbs; --i >= 0; ) {
2333 if ((mc = src->mt_cursors[i]) != NULL) {
2334 size = sizeof(MDB_cursor);
2336 size += sizeof(MDB_xcursor);
2337 for (; mc; mc = bk->mc_next) {
2343 mc->mc_db = &dst->mt_dbs[i];
2344 /* Kill pointers into src - and dst to reduce abuse: The
2345 * user may not use mc until dst ends. Otherwise we'd...
2347 mc->mc_txn = NULL; /* ...set this to dst */
2348 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2349 if ((mx = mc->mc_xcursor) != NULL) {
2350 *(MDB_xcursor *)(bk+1) = *mx;
2351 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2353 mc->mc_next = dst->mt_cursors[i];
2354 dst->mt_cursors[i] = mc;
2361 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2362 * @param[in] txn the transaction handle.
2363 * @param[in] merge true to keep changes to parent cursors, false to revert.
2364 * @return 0 on success, non-zero on failure.
2367 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2369 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2373 for (i = txn->mt_numdbs; --i >= 0; ) {
2374 for (mc = cursors[i]; mc; mc = next) {
2376 if ((bk = mc->mc_backup) != NULL) {
2378 /* Commit changes to parent txn */
2379 mc->mc_next = bk->mc_next;
2380 mc->mc_backup = bk->mc_backup;
2381 mc->mc_txn = bk->mc_txn;
2382 mc->mc_db = bk->mc_db;
2383 mc->mc_dbflag = bk->mc_dbflag;
2384 if ((mx = mc->mc_xcursor) != NULL)
2385 mx->mx_cursor.mc_txn = bk->mc_txn;
2387 /* Abort nested txn */
2389 if ((mx = mc->mc_xcursor) != NULL)
2390 *mx = *(MDB_xcursor *)(bk+1);
2394 /* Only malloced cursors are permanently tracked. */
2402 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2405 mdb_txn_reset0(MDB_txn *txn, const char *act);
2407 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2413 Pidset = F_SETLK, Pidcheck = F_GETLK
2417 /** Set or check a pid lock. Set returns 0 on success.
2418 * Check returns 0 if the process is certainly dead, nonzero if it may
2419 * be alive (the lock exists or an error happened so we do not know).
2421 * On Windows Pidset is a no-op, we merely check for the existence
2422 * of the process with the given pid. On POSIX we use a single byte
2423 * lock on the lockfile, set at an offset equal to the pid.
2426 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2428 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2431 if (op == Pidcheck) {
2432 h = OpenProcess(env->me_pidquery, FALSE, pid);
2433 /* No documented "no such process" code, but other program use this: */
2435 return ErrCode() != ERROR_INVALID_PARAMETER;
2436 /* A process exists until all handles to it close. Has it exited? */
2437 ret = WaitForSingleObject(h, 0) != 0;
2444 struct flock lock_info;
2445 memset(&lock_info, 0, sizeof(lock_info));
2446 lock_info.l_type = F_WRLCK;
2447 lock_info.l_whence = SEEK_SET;
2448 lock_info.l_start = pid;
2449 lock_info.l_len = 1;
2450 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2451 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2453 } else if ((rc = ErrCode()) == EINTR) {
2461 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2462 * @param[in] txn the transaction handle to initialize
2463 * @return 0 on success, non-zero on failure.
2466 mdb_txn_renew0(MDB_txn *txn)
2468 MDB_env *env = txn->mt_env;
2469 MDB_txninfo *ti = env->me_txns;
2473 int rc, new_notls = 0;
2476 txn->mt_numdbs = env->me_numdbs;
2477 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2479 if (txn->mt_flags & MDB_TXN_RDONLY) {
2481 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2482 txn->mt_txnid = meta->mm_txnid;
2483 txn->mt_u.reader = NULL;
2485 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2486 pthread_getspecific(env->me_txkey);
2488 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2489 return MDB_BAD_RSLOT;
2491 MDB_PID_T pid = env->me_pid;
2492 MDB_THR_T tid = pthread_self();
2494 if (!env->me_live_reader) {
2495 rc = mdb_reader_pid(env, Pidset, pid);
2498 env->me_live_reader = 1;
2502 nr = ti->mti_numreaders;
2503 for (i=0; i<nr; i++)
2504 if (ti->mti_readers[i].mr_pid == 0)
2506 if (i == env->me_maxreaders) {
2507 UNLOCK_MUTEX_R(env);
2508 return MDB_READERS_FULL;
2510 ti->mti_readers[i].mr_pid = pid;
2511 ti->mti_readers[i].mr_tid = tid;
2513 ti->mti_numreaders = ++nr;
2514 /* Save numreaders for un-mutexed mdb_env_close() */
2515 env->me_numreaders = nr;
2516 UNLOCK_MUTEX_R(env);
2518 r = &ti->mti_readers[i];
2519 new_notls = (env->me_flags & MDB_NOTLS);
2520 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2525 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2526 txn->mt_u.reader = r;
2527 meta = env->me_metas[txn->mt_txnid & 1];
2533 txn->mt_txnid = ti->mti_txnid;
2534 meta = env->me_metas[txn->mt_txnid & 1];
2536 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2537 txn->mt_txnid = meta->mm_txnid;
2541 if (txn->mt_txnid == mdb_debug_start)
2544 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2545 txn->mt_u.dirty_list = env->me_dirty_list;
2546 txn->mt_u.dirty_list[0].mid = 0;
2547 txn->mt_free_pgs = env->me_free_pgs;
2548 txn->mt_free_pgs[0] = 0;
2549 txn->mt_spill_pgs = NULL;
2551 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2554 /* Copy the DB info and flags */
2555 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2557 /* Moved to here to avoid a data race in read TXNs */
2558 txn->mt_next_pgno = meta->mm_last_pg+1;
2560 for (i=2; i<txn->mt_numdbs; i++) {
2561 x = env->me_dbflags[i];
2562 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2563 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2565 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2567 if (env->me_maxpg < txn->mt_next_pgno) {
2568 mdb_txn_reset0(txn, "renew0-mapfail");
2570 txn->mt_u.reader->mr_pid = 0;
2571 txn->mt_u.reader = NULL;
2573 return MDB_MAP_RESIZED;
2580 mdb_txn_renew(MDB_txn *txn)
2584 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2587 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2588 DPUTS("environment had fatal error, must shutdown!");
2592 rc = mdb_txn_renew0(txn);
2593 if (rc == MDB_SUCCESS) {
2594 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2595 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2596 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2602 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2606 int rc, size, tsize = sizeof(MDB_txn);
2608 if (env->me_flags & MDB_FATAL_ERROR) {
2609 DPUTS("environment had fatal error, must shutdown!");
2612 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2615 /* Nested transactions: Max 1 child, write txns only, no writemap */
2616 if (parent->mt_child ||
2617 (flags & MDB_RDONLY) ||
2618 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2619 (env->me_flags & MDB_WRITEMAP))
2621 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2623 tsize = sizeof(MDB_ntxn);
2625 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2626 if (!(flags & MDB_RDONLY)) {
2632 size += env->me_maxdbs * sizeof(MDB_cursor *);
2633 /* child txns use parent's dbiseqs */
2635 size += env->me_maxdbs * sizeof(unsigned int);
2638 if ((txn = calloc(1, size)) == NULL) {
2639 DPRINTF(("calloc: %s", strerror(errno)));
2642 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2643 if (flags & MDB_RDONLY) {
2644 txn->mt_flags |= MDB_TXN_RDONLY;
2645 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2646 txn->mt_dbiseqs = env->me_dbiseqs;
2648 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2650 txn->mt_dbiseqs = parent->mt_dbiseqs;
2651 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2653 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2654 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2662 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2663 if (!txn->mt_u.dirty_list ||
2664 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2666 free(txn->mt_u.dirty_list);
2670 txn->mt_txnid = parent->mt_txnid;
2671 txn->mt_dirty_room = parent->mt_dirty_room;
2672 txn->mt_u.dirty_list[0].mid = 0;
2673 txn->mt_spill_pgs = NULL;
2674 txn->mt_next_pgno = parent->mt_next_pgno;
2675 parent->mt_child = txn;
2676 txn->mt_parent = parent;
2677 txn->mt_numdbs = parent->mt_numdbs;
2678 txn->mt_flags = parent->mt_flags;
2679 txn->mt_dbxs = parent->mt_dbxs;
2680 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2681 /* Copy parent's mt_dbflags, but clear DB_NEW */
2682 for (i=0; i<txn->mt_numdbs; i++)
2683 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2685 ntxn = (MDB_ntxn *)txn;
2686 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2687 if (env->me_pghead) {
2688 size = MDB_IDL_SIZEOF(env->me_pghead);
2689 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2691 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2696 rc = mdb_cursor_shadow(parent, txn);
2698 mdb_txn_reset0(txn, "beginchild-fail");
2700 rc = mdb_txn_renew0(txn);
2703 if (txn != env->me_txn0)
2707 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2708 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2709 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2716 mdb_txn_env(MDB_txn *txn)
2718 if(!txn) return NULL;
2722 /** Export or close DBI handles opened in this txn. */
2724 mdb_dbis_update(MDB_txn *txn, int keep)
2727 MDB_dbi n = txn->mt_numdbs;
2728 MDB_env *env = txn->mt_env;
2729 unsigned char *tdbflags = txn->mt_dbflags;
2731 for (i = n; --i >= 2;) {
2732 if (tdbflags[i] & DB_NEW) {
2734 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2736 char *ptr = env->me_dbxs[i].md_name.mv_data;
2738 env->me_dbxs[i].md_name.mv_data = NULL;
2739 env->me_dbxs[i].md_name.mv_size = 0;
2740 env->me_dbflags[i] = 0;
2741 env->me_dbiseqs[i]++;
2747 if (keep && env->me_numdbs < n)
2751 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2752 * May be called twice for readonly txns: First reset it, then abort.
2753 * @param[in] txn the transaction handle to reset
2754 * @param[in] act why the transaction is being reset
2757 mdb_txn_reset0(MDB_txn *txn, const char *act)
2759 MDB_env *env = txn->mt_env;
2761 /* Close any DBI handles opened in this txn */
2762 mdb_dbis_update(txn, 0);
2764 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2765 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2766 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2768 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2769 if (txn->mt_u.reader) {
2770 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2771 if (!(env->me_flags & MDB_NOTLS))
2772 txn->mt_u.reader = NULL; /* txn does not own reader */
2774 txn->mt_numdbs = 0; /* close nothing if called again */
2775 txn->mt_dbxs = NULL; /* mark txn as reset */
2777 mdb_cursors_close(txn, 0);
2779 if (!(env->me_flags & MDB_WRITEMAP)) {
2780 mdb_dlist_free(txn);
2782 mdb_midl_free(env->me_pghead);
2784 if (txn->mt_parent) {
2785 txn->mt_parent->mt_child = NULL;
2786 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2787 mdb_midl_free(txn->mt_free_pgs);
2788 mdb_midl_free(txn->mt_spill_pgs);
2789 free(txn->mt_u.dirty_list);
2793 if (mdb_midl_shrink(&txn->mt_free_pgs))
2794 env->me_free_pgs = txn->mt_free_pgs;
2795 env->me_pghead = NULL;
2799 /* The writer mutex was locked in mdb_txn_begin. */
2801 UNLOCK_MUTEX_W(env);
2806 mdb_txn_reset(MDB_txn *txn)
2811 /* This call is only valid for read-only txns */
2812 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2815 mdb_txn_reset0(txn, "reset");
2819 mdb_txn_abort(MDB_txn *txn)
2825 mdb_txn_abort(txn->mt_child);
2827 mdb_txn_reset0(txn, "abort");
2828 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2829 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2830 txn->mt_u.reader->mr_pid = 0;
2832 if (txn != txn->mt_env->me_txn0)
2836 /** Save the freelist as of this transaction to the freeDB.
2837 * This changes the freelist. Keep trying until it stabilizes.
2840 mdb_freelist_save(MDB_txn *txn)
2842 /* env->me_pghead[] can grow and shrink during this call.
2843 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2844 * Page numbers cannot disappear from txn->mt_free_pgs[].
2847 MDB_env *env = txn->mt_env;
2848 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2849 txnid_t pglast = 0, head_id = 0;
2850 pgno_t freecnt = 0, *free_pgs, *mop;
2851 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2853 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2855 if (env->me_pghead) {
2856 /* Make sure first page of freeDB is touched and on freelist */
2857 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2858 if (rc && rc != MDB_NOTFOUND)
2862 if (!env->me_pghead && txn->mt_loose_pgs) {
2863 /* Put loose page numbers in mt_free_pgs, since
2864 * we may be unable to return them to me_pghead.
2866 MDB_page *mp = txn->mt_loose_pgs;
2867 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2869 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2870 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2871 txn->mt_loose_pgs = NULL;
2872 txn->mt_loose_count = 0;
2875 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2876 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2877 ? SSIZE_MAX : maxfree_1pg;
2880 /* Come back here after each Put() in case freelist changed */
2885 /* If using records from freeDB which we have not yet
2886 * deleted, delete them and any we reserved for me_pghead.
2888 while (pglast < env->me_pglast) {
2889 rc = mdb_cursor_first(&mc, &key, NULL);
2892 pglast = head_id = *(txnid_t *)key.mv_data;
2893 total_room = head_room = 0;
2894 mdb_tassert(txn, pglast <= env->me_pglast);
2895 rc = mdb_cursor_del(&mc, 0);
2900 /* Save the IDL of pages freed by this txn, to a single record */
2901 if (freecnt < txn->mt_free_pgs[0]) {
2903 /* Make sure last page of freeDB is touched and on freelist */
2904 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2905 if (rc && rc != MDB_NOTFOUND)
2908 free_pgs = txn->mt_free_pgs;
2909 /* Write to last page of freeDB */
2910 key.mv_size = sizeof(txn->mt_txnid);
2911 key.mv_data = &txn->mt_txnid;
2913 freecnt = free_pgs[0];
2914 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2915 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2918 /* Retry if mt_free_pgs[] grew during the Put() */
2919 free_pgs = txn->mt_free_pgs;
2920 } while (freecnt < free_pgs[0]);
2921 mdb_midl_sort(free_pgs);
2922 memcpy(data.mv_data, free_pgs, data.mv_size);
2925 unsigned int i = free_pgs[0];
2926 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2927 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2929 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2935 mop = env->me_pghead;
2936 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2938 /* Reserve records for me_pghead[]. Split it if multi-page,
2939 * to avoid searching freeDB for a page range. Use keys in
2940 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2942 if (total_room >= mop_len) {
2943 if (total_room == mop_len || --more < 0)
2945 } else if (head_room >= maxfree_1pg && head_id > 1) {
2946 /* Keep current record (overflow page), add a new one */
2950 /* (Re)write {key = head_id, IDL length = head_room} */
2951 total_room -= head_room;
2952 head_room = mop_len - total_room;
2953 if (head_room > maxfree_1pg && head_id > 1) {
2954 /* Overflow multi-page for part of me_pghead */
2955 head_room /= head_id; /* amortize page sizes */
2956 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2957 } else if (head_room < 0) {
2958 /* Rare case, not bothering to delete this record */
2961 key.mv_size = sizeof(head_id);
2962 key.mv_data = &head_id;
2963 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2964 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2967 /* IDL is initially empty, zero out at least the length */
2968 pgs = (pgno_t *)data.mv_data;
2969 j = head_room > clean_limit ? head_room : 0;
2973 total_room += head_room;
2976 /* Return loose page numbers to me_pghead, though usually none are
2977 * left at this point. The pages themselves remain in dirty_list.
2979 if (txn->mt_loose_pgs) {
2980 MDB_page *mp = txn->mt_loose_pgs;
2981 unsigned count = txn->mt_loose_count;
2983 /* Room for loose pages + temp IDL with same */
2984 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
2986 mop = env->me_pghead;
2987 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
2988 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
2989 loose[ ++count ] = mp->mp_pgno;
2991 mdb_midl_sort(loose);
2992 mdb_midl_xmerge(mop, loose);
2993 txn->mt_loose_pgs = NULL;
2994 txn->mt_loose_count = 0;
2998 /* Fill in the reserved me_pghead records */
3004 rc = mdb_cursor_first(&mc, &key, &data);
3005 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3006 txnid_t id = *(txnid_t *)key.mv_data;
3007 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3010 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3012 if (len > mop_len) {
3014 data.mv_size = (len + 1) * sizeof(MDB_ID);
3016 data.mv_data = mop -= len;
3019 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3021 if (rc || !(mop_len -= len))
3028 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3029 * @param[in] txn the transaction that's being committed
3030 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3031 * @return 0 on success, non-zero on failure.
3034 mdb_page_flush(MDB_txn *txn, int keep)
3036 MDB_env *env = txn->mt_env;
3037 MDB_ID2L dl = txn->mt_u.dirty_list;
3038 unsigned psize = env->me_psize, j;
3039 int i, pagecount = dl[0].mid, rc;
3040 size_t size = 0, pos = 0;
3042 MDB_page *dp = NULL;
3046 struct iovec iov[MDB_COMMIT_PAGES];
3047 ssize_t wpos = 0, wsize = 0, wres;
3048 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3054 if (env->me_flags & MDB_WRITEMAP) {
3055 /* Clear dirty flags */
3056 while (++i <= pagecount) {
3058 /* Don't flush this page yet */
3059 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3060 dp->mp_flags &= ~P_KEEP;
3064 dp->mp_flags &= ~P_DIRTY;
3069 /* Write the pages */
3071 if (++i <= pagecount) {
3073 /* Don't flush this page yet */
3074 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3075 dp->mp_flags &= ~P_KEEP;
3080 /* clear dirty flag */
3081 dp->mp_flags &= ~P_DIRTY;
3084 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3089 /* Windows actually supports scatter/gather I/O, but only on
3090 * unbuffered file handles. Since we're relying on the OS page
3091 * cache for all our data, that's self-defeating. So we just
3092 * write pages one at a time. We use the ov structure to set
3093 * the write offset, to at least save the overhead of a Seek
3096 DPRINTF(("committing page %"Z"u", pgno));
3097 memset(&ov, 0, sizeof(ov));
3098 ov.Offset = pos & 0xffffffff;
3099 ov.OffsetHigh = pos >> 16 >> 16;
3100 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3102 DPRINTF(("WriteFile: %d", rc));
3106 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3107 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3109 /* Write previous page(s) */
3110 #ifdef MDB_USE_PWRITEV
3111 wres = pwritev(env->me_fd, iov, n, wpos);
3114 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3116 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3118 DPRINTF(("lseek: %s", strerror(rc)));
3121 wres = writev(env->me_fd, iov, n);
3124 if (wres != wsize) {
3127 DPRINTF(("Write error: %s", strerror(rc)));
3129 rc = EIO; /* TODO: Use which error code? */
3130 DPUTS("short write, filesystem full?");
3141 DPRINTF(("committing page %"Z"u", pgno));
3142 next_pos = pos + size;
3143 iov[n].iov_len = size;
3144 iov[n].iov_base = (char *)dp;
3150 /* MIPS has cache coherency issues, this is a no-op everywhere else
3151 * Note: for any size >= on-chip cache size, entire on-chip cache is
3154 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3156 for (i = keep; ++i <= pagecount; ) {
3158 /* This is a page we skipped above */
3161 dl[j].mid = dp->mp_pgno;
3164 mdb_dpage_free(env, dp);
3169 txn->mt_dirty_room += i - j;
3175 mdb_txn_commit(MDB_txn *txn)
3181 if (txn == NULL || txn->mt_env == NULL)
3184 if (txn->mt_child) {
3185 rc = mdb_txn_commit(txn->mt_child);
3186 txn->mt_child = NULL;
3193 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3194 mdb_dbis_update(txn, 1);
3195 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3200 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3201 DPUTS("error flag is set, can't commit");
3203 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3208 if (txn->mt_parent) {
3209 MDB_txn *parent = txn->mt_parent;
3213 unsigned x, y, len, ps_len;
3215 /* Append our free list to parent's */
3216 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3219 mdb_midl_free(txn->mt_free_pgs);
3220 /* Failures after this must either undo the changes
3221 * to the parent or set MDB_TXN_ERROR in the parent.
3224 parent->mt_next_pgno = txn->mt_next_pgno;
3225 parent->mt_flags = txn->mt_flags;
3227 /* Merge our cursors into parent's and close them */
3228 mdb_cursors_close(txn, 1);
3230 /* Update parent's DB table. */
3231 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3232 parent->mt_numdbs = txn->mt_numdbs;
3233 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3234 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3235 for (i=2; i<txn->mt_numdbs; i++) {
3236 /* preserve parent's DB_NEW status */
3237 x = parent->mt_dbflags[i] & DB_NEW;
3238 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3241 dst = parent->mt_u.dirty_list;
3242 src = txn->mt_u.dirty_list;
3243 /* Remove anything in our dirty list from parent's spill list */
3244 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3246 pspill[0] = (pgno_t)-1;
3247 /* Mark our dirty pages as deleted in parent spill list */
3248 for (i=0, len=src[0].mid; ++i <= len; ) {
3249 MDB_ID pn = src[i].mid << 1;
3250 while (pn > pspill[x])
3252 if (pn == pspill[x]) {
3257 /* Squash deleted pagenums if we deleted any */
3258 for (x=y; ++x <= ps_len; )
3259 if (!(pspill[x] & 1))
3260 pspill[++y] = pspill[x];
3264 /* Find len = length of merging our dirty list with parent's */
3266 dst[0].mid = 0; /* simplify loops */
3267 if (parent->mt_parent) {
3268 len = x + src[0].mid;
3269 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3270 for (i = x; y && i; y--) {
3271 pgno_t yp = src[y].mid;
3272 while (yp < dst[i].mid)
3274 if (yp == dst[i].mid) {
3279 } else { /* Simplify the above for single-ancestor case */
3280 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3282 /* Merge our dirty list with parent's */
3284 for (i = len; y; dst[i--] = src[y--]) {
3285 pgno_t yp = src[y].mid;
3286 while (yp < dst[x].mid)
3287 dst[i--] = dst[x--];
3288 if (yp == dst[x].mid)
3289 free(dst[x--].mptr);
3291 mdb_tassert(txn, i == x);
3293 free(txn->mt_u.dirty_list);
3294 parent->mt_dirty_room = txn->mt_dirty_room;
3295 if (txn->mt_spill_pgs) {
3296 if (parent->mt_spill_pgs) {
3297 /* TODO: Prevent failure here, so parent does not fail */
3298 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3300 parent->mt_flags |= MDB_TXN_ERROR;
3301 mdb_midl_free(txn->mt_spill_pgs);
3302 mdb_midl_sort(parent->mt_spill_pgs);
3304 parent->mt_spill_pgs = txn->mt_spill_pgs;
3308 /* Append our loose page list to parent's */
3309 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3311 *lp = txn->mt_loose_pgs;
3312 parent->mt_loose_count += txn->mt_loose_count;
3314 parent->mt_child = NULL;
3315 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3320 if (txn != env->me_txn) {
3321 DPUTS("attempt to commit unknown transaction");
3326 mdb_cursors_close(txn, 0);
3328 if (!txn->mt_u.dirty_list[0].mid &&
3329 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3332 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3333 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3335 /* Update DB root pointers */
3336 if (txn->mt_numdbs > 2) {
3340 data.mv_size = sizeof(MDB_db);
3342 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3343 for (i = 2; i < txn->mt_numdbs; i++) {
3344 if (txn->mt_dbflags[i] & DB_DIRTY) {
3345 if (TXN_DBI_CHANGED(txn, i)) {
3349 data.mv_data = &txn->mt_dbs[i];
3350 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3357 rc = mdb_freelist_save(txn);
3361 mdb_midl_free(env->me_pghead);
3362 env->me_pghead = NULL;
3363 if (mdb_midl_shrink(&txn->mt_free_pgs))
3364 env->me_free_pgs = txn->mt_free_pgs;
3370 if ((rc = mdb_page_flush(txn, 0)) ||
3371 (rc = mdb_env_sync(env, 0)) ||
3372 (rc = mdb_env_write_meta(txn)))
3375 /* Free P_LOOSE pages left behind in dirty_list */
3376 if (!(env->me_flags & MDB_WRITEMAP))
3377 mdb_dlist_free(txn);
3382 mdb_dbis_update(txn, 1);
3385 UNLOCK_MUTEX_W(env);
3386 if (txn != env->me_txn0)
3396 /** Read the environment parameters of a DB environment before
3397 * mapping it into memory.
3398 * @param[in] env the environment handle
3399 * @param[out] meta address of where to store the meta information
3400 * @return 0 on success, non-zero on failure.
3403 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3409 enum { Size = sizeof(pbuf) };
3411 /* We don't know the page size yet, so use a minimum value.
3412 * Read both meta pages so we can use the latest one.
3415 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3419 memset(&ov, 0, sizeof(ov));
3421 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3422 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3425 rc = pread(env->me_fd, &pbuf, Size, off);
3428 if (rc == 0 && off == 0)
3430 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3431 DPRINTF(("read: %s", mdb_strerror(rc)));
3435 p = (MDB_page *)&pbuf;
3437 if (!F_ISSET(p->mp_flags, P_META)) {
3438 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3443 if (m->mm_magic != MDB_MAGIC) {
3444 DPUTS("meta has invalid magic");
3448 if (m->mm_version != MDB_DATA_VERSION) {
3449 DPRINTF(("database is version %u, expected version %u",
3450 m->mm_version, MDB_DATA_VERSION));
3451 return MDB_VERSION_MISMATCH;
3454 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3461 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3463 meta->mm_magic = MDB_MAGIC;
3464 meta->mm_version = MDB_DATA_VERSION;
3465 meta->mm_mapsize = env->me_mapsize;
3466 meta->mm_psize = env->me_psize;
3467 meta->mm_last_pg = 1;
3468 meta->mm_flags = env->me_flags & 0xffff;
3469 meta->mm_flags |= MDB_INTEGERKEY;
3470 meta->mm_dbs[0].md_root = P_INVALID;
3471 meta->mm_dbs[1].md_root = P_INVALID;
3474 /** Write the environment parameters of a freshly created DB environment.
3475 * @param[in] env the environment handle
3476 * @param[out] meta address of where to store the meta information
3477 * @return 0 on success, non-zero on failure.
3480 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3488 memset(&ov, 0, sizeof(ov));
3489 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3491 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3494 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3495 len = pwrite(fd, ptr, size, pos); \
3496 rc = (len >= 0); } while(0)
3499 DPUTS("writing new meta page");
3501 psize = env->me_psize;
3503 mdb_env_init_meta0(env, meta);
3505 p = calloc(2, psize);
3507 p->mp_flags = P_META;
3508 *(MDB_meta *)METADATA(p) = *meta;
3510 q = (MDB_page *)((char *)p + psize);
3512 q->mp_flags = P_META;
3513 *(MDB_meta *)METADATA(q) = *meta;
3515 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3518 else if ((unsigned) len == psize * 2)
3526 /** Update the environment info to commit a transaction.
3527 * @param[in] txn the transaction that's being committed
3528 * @return 0 on success, non-zero on failure.
3531 mdb_env_write_meta(MDB_txn *txn)
3534 MDB_meta meta, metab, *mp;
3537 int rc, len, toggle;
3546 toggle = txn->mt_txnid & 1;
3547 DPRINTF(("writing meta page %d for root page %"Z"u",
3548 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3551 mp = env->me_metas[toggle];
3552 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3553 /* Persist any increases of mapsize config */
3554 if (mapsize < env->me_mapsize)
3555 mapsize = env->me_mapsize;
3557 if (env->me_flags & MDB_WRITEMAP) {
3558 mp->mm_mapsize = mapsize;
3559 mp->mm_dbs[0] = txn->mt_dbs[0];
3560 mp->mm_dbs[1] = txn->mt_dbs[1];
3561 mp->mm_last_pg = txn->mt_next_pgno - 1;
3562 mp->mm_txnid = txn->mt_txnid;
3563 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3564 unsigned meta_size = env->me_psize;
3565 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3568 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3569 if (meta_size < env->me_os_psize)
3570 meta_size += meta_size;
3575 if (MDB_MSYNC(ptr, meta_size, rc)) {
3582 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3583 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3585 meta.mm_mapsize = mapsize;
3586 meta.mm_dbs[0] = txn->mt_dbs[0];
3587 meta.mm_dbs[1] = txn->mt_dbs[1];
3588 meta.mm_last_pg = txn->mt_next_pgno - 1;
3589 meta.mm_txnid = txn->mt_txnid;
3591 off = offsetof(MDB_meta, mm_mapsize);
3592 ptr = (char *)&meta + off;
3593 len = sizeof(MDB_meta) - off;
3595 off += env->me_psize;
3598 /* Write to the SYNC fd */
3599 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3600 env->me_fd : env->me_mfd;
3603 memset(&ov, 0, sizeof(ov));
3605 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3609 rc = pwrite(mfd, ptr, len, off);
3612 rc = rc < 0 ? ErrCode() : EIO;
3613 DPUTS("write failed, disk error?");
3614 /* On a failure, the pagecache still contains the new data.
3615 * Write some old data back, to prevent it from being used.
3616 * Use the non-SYNC fd; we know it will fail anyway.
3618 meta.mm_last_pg = metab.mm_last_pg;
3619 meta.mm_txnid = metab.mm_txnid;
3621 memset(&ov, 0, sizeof(ov));
3623 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3625 r2 = pwrite(env->me_fd, ptr, len, off);
3626 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3629 env->me_flags |= MDB_FATAL_ERROR;
3632 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3633 CACHEFLUSH(env->me_map + off, len, DCACHE);
3635 /* Memory ordering issues are irrelevant; since the entire writer
3636 * is wrapped by wmutex, all of these changes will become visible
3637 * after the wmutex is unlocked. Since the DB is multi-version,
3638 * readers will get consistent data regardless of how fresh or
3639 * how stale their view of these values is.
3642 env->me_txns->mti_txnid = txn->mt_txnid;
3647 /** Check both meta pages to see which one is newer.
3648 * @param[in] env the environment handle
3649 * @return meta toggle (0 or 1).
3652 mdb_env_pick_meta(const MDB_env *env)
3654 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3658 mdb_env_create(MDB_env **env)
3662 e = calloc(1, sizeof(MDB_env));
3666 e->me_maxreaders = DEFAULT_READERS;
3667 e->me_maxdbs = e->me_numdbs = 2;
3668 e->me_fd = INVALID_HANDLE_VALUE;
3669 e->me_lfd = INVALID_HANDLE_VALUE;
3670 e->me_mfd = INVALID_HANDLE_VALUE;
3671 #ifdef MDB_USE_POSIX_SEM
3672 e->me_rmutex = SEM_FAILED;
3673 e->me_wmutex = SEM_FAILED;
3675 e->me_pid = getpid();
3676 GET_PAGESIZE(e->me_os_psize);
3677 VGMEMP_CREATE(e,0,0);
3683 mdb_env_map(MDB_env *env, void *addr)
3686 unsigned int flags = env->me_flags;
3690 LONG sizelo, sizehi;
3693 if (flags & MDB_RDONLY) {
3694 /* Don't set explicit map size, use whatever exists */
3699 msize = env->me_mapsize;
3700 sizelo = msize & 0xffffffff;
3701 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3703 /* Windows won't create mappings for zero length files.
3704 * and won't map more than the file size.
3705 * Just set the maxsize right now.
3707 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3708 || !SetEndOfFile(env->me_fd)
3709 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3713 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3714 PAGE_READWRITE : PAGE_READONLY,
3715 sizehi, sizelo, NULL);
3718 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3719 FILE_MAP_WRITE : FILE_MAP_READ,
3721 rc = env->me_map ? 0 : ErrCode();
3726 int prot = PROT_READ;
3727 if (flags & MDB_WRITEMAP) {
3729 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3732 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3734 if (env->me_map == MAP_FAILED) {
3739 if (flags & MDB_NORDAHEAD) {
3740 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3742 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3744 #ifdef POSIX_MADV_RANDOM
3745 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3746 #endif /* POSIX_MADV_RANDOM */
3747 #endif /* MADV_RANDOM */
3751 /* Can happen because the address argument to mmap() is just a
3752 * hint. mmap() can pick another, e.g. if the range is in use.
3753 * The MAP_FIXED flag would prevent that, but then mmap could
3754 * instead unmap existing pages to make room for the new map.
3756 if (addr && env->me_map != addr)
3757 return EBUSY; /* TODO: Make a new MDB_* error code? */
3759 p = (MDB_page *)env->me_map;
3760 env->me_metas[0] = METADATA(p);
3761 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3767 mdb_env_set_mapsize(MDB_env *env, size_t size)
3769 /* If env is already open, caller is responsible for making
3770 * sure there are no active txns.
3778 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3779 else if (size < env->me_mapsize) {
3780 /* If the configured size is smaller, make sure it's
3781 * still big enough. Silently round up to minimum if not.
3783 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3787 munmap(env->me_map, env->me_mapsize);
3788 env->me_mapsize = size;
3789 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3790 rc = mdb_env_map(env, old);
3794 env->me_mapsize = size;
3796 env->me_maxpg = env->me_mapsize / env->me_psize;
3801 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3805 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3810 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3812 if (env->me_map || readers < 1)
3814 env->me_maxreaders = readers;
3819 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3821 if (!env || !readers)
3823 *readers = env->me_maxreaders;
3827 /** Further setup required for opening an LMDB environment
3830 mdb_env_open2(MDB_env *env)
3832 unsigned int flags = env->me_flags;
3833 int i, newenv = 0, rc;
3837 /* See if we should use QueryLimited */
3839 if ((rc & 0xff) > 5)
3840 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3842 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3845 memset(&meta, 0, sizeof(meta));
3847 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3850 DPUTS("new mdbenv");
3852 env->me_psize = env->me_os_psize;
3853 if (env->me_psize > MAX_PAGESIZE)
3854 env->me_psize = MAX_PAGESIZE;
3856 env->me_psize = meta.mm_psize;
3859 /* Was a mapsize configured? */
3860 if (!env->me_mapsize) {
3861 /* If this is a new environment, take the default,
3862 * else use the size recorded in the existing env.
3864 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3865 } else if (env->me_mapsize < meta.mm_mapsize) {
3866 /* If the configured size is smaller, make sure it's
3867 * still big enough. Silently round up to minimum if not.
3869 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3870 if (env->me_mapsize < minsize)
3871 env->me_mapsize = minsize;
3874 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3879 if (flags & MDB_FIXEDMAP)
3880 meta.mm_address = env->me_map;
3881 i = mdb_env_init_meta(env, &meta);
3882 if (i != MDB_SUCCESS) {
3887 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3888 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3890 #if !(MDB_MAXKEYSIZE)
3891 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3893 env->me_maxpg = env->me_mapsize / env->me_psize;
3897 int toggle = mdb_env_pick_meta(env);
3898 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3900 DPRINTF(("opened database version %u, pagesize %u",
3901 env->me_metas[0]->mm_version, env->me_psize));
3902 DPRINTF(("using meta page %d", toggle));
3903 DPRINTF(("depth: %u", db->md_depth));
3904 DPRINTF(("entries: %"Z"u", db->md_entries));
3905 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3906 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3907 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3908 DPRINTF(("root: %"Z"u", db->md_root));
3916 /** Release a reader thread's slot in the reader lock table.
3917 * This function is called automatically when a thread exits.
3918 * @param[in] ptr This points to the slot in the reader lock table.
3921 mdb_env_reader_dest(void *ptr)
3923 MDB_reader *reader = ptr;
3929 /** Junk for arranging thread-specific callbacks on Windows. This is
3930 * necessarily platform and compiler-specific. Windows supports up
3931 * to 1088 keys. Let's assume nobody opens more than 64 environments
3932 * in a single process, for now. They can override this if needed.
3934 #ifndef MAX_TLS_KEYS
3935 #define MAX_TLS_KEYS 64
3937 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3938 static int mdb_tls_nkeys;
3940 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3944 case DLL_PROCESS_ATTACH: break;
3945 case DLL_THREAD_ATTACH: break;
3946 case DLL_THREAD_DETACH:
3947 for (i=0; i<mdb_tls_nkeys; i++) {
3948 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3950 mdb_env_reader_dest(r);
3954 case DLL_PROCESS_DETACH: break;
3959 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3961 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3965 /* Force some symbol references.
3966 * _tls_used forces the linker to create the TLS directory if not already done
3967 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3969 #pragma comment(linker, "/INCLUDE:_tls_used")
3970 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3971 #pragma const_seg(".CRT$XLB")
3972 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3973 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3976 #pragma comment(linker, "/INCLUDE:__tls_used")
3977 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3978 #pragma data_seg(".CRT$XLB")
3979 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3981 #endif /* WIN 32/64 */
3982 #endif /* !__GNUC__ */
3985 /** Downgrade the exclusive lock on the region back to shared */
3987 mdb_env_share_locks(MDB_env *env, int *excl)
3989 int rc = 0, toggle = mdb_env_pick_meta(env);
3991 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3996 /* First acquire a shared lock. The Unlock will
3997 * then release the existing exclusive lock.
3999 memset(&ov, 0, sizeof(ov));
4000 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4003 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4009 struct flock lock_info;
4010 /* The shared lock replaces the existing lock */
4011 memset((void *)&lock_info, 0, sizeof(lock_info));
4012 lock_info.l_type = F_RDLCK;
4013 lock_info.l_whence = SEEK_SET;
4014 lock_info.l_start = 0;
4015 lock_info.l_len = 1;
4016 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4017 (rc = ErrCode()) == EINTR) ;
4018 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4025 /** Try to get exlusive lock, otherwise shared.
4026 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4029 mdb_env_excl_lock(MDB_env *env, int *excl)
4033 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4037 memset(&ov, 0, sizeof(ov));
4038 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4045 struct flock lock_info;
4046 memset((void *)&lock_info, 0, sizeof(lock_info));
4047 lock_info.l_type = F_WRLCK;
4048 lock_info.l_whence = SEEK_SET;
4049 lock_info.l_start = 0;
4050 lock_info.l_len = 1;
4051 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4052 (rc = ErrCode()) == EINTR) ;
4056 # ifdef MDB_USE_POSIX_SEM
4057 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4060 lock_info.l_type = F_RDLCK;
4061 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4062 (rc = ErrCode()) == EINTR) ;
4072 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4074 * @(#) $Revision: 5.1 $
4075 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4076 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4078 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4082 * Please do not copyright this code. This code is in the public domain.
4084 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4085 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4086 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4087 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4088 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4089 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4090 * PERFORMANCE OF THIS SOFTWARE.
4093 * chongo <Landon Curt Noll> /\oo/\
4094 * http://www.isthe.com/chongo/
4096 * Share and Enjoy! :-)
4099 typedef unsigned long long mdb_hash_t;
4100 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4102 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4103 * @param[in] val value to hash
4104 * @param[in] hval initial value for hash
4105 * @return 64 bit hash
4107 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4108 * hval arg on the first call.
4111 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4113 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4114 unsigned char *end = s + val->mv_size;
4116 * FNV-1a hash each octet of the string
4119 /* xor the bottom with the current octet */
4120 hval ^= (mdb_hash_t)*s++;
4122 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4123 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4124 (hval << 7) + (hval << 8) + (hval << 40);
4126 /* return our new hash value */
4130 /** Hash the string and output the encoded hash.
4131 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4132 * very short name limits. We don't care about the encoding being reversible,
4133 * we just want to preserve as many bits of the input as possible in a
4134 * small printable string.
4135 * @param[in] str string to hash
4136 * @param[out] encbuf an array of 11 chars to hold the hash
4138 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4141 mdb_pack85(unsigned long l, char *out)
4145 for (i=0; i<5; i++) {
4146 *out++ = mdb_a85[l % 85];
4152 mdb_hash_enc(MDB_val *val, char *encbuf)
4154 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4156 mdb_pack85(h, encbuf);
4157 mdb_pack85(h>>32, encbuf+5);
4162 /** Open and/or initialize the lock region for the environment.
4163 * @param[in] env The LMDB environment.
4164 * @param[in] lpath The pathname of the file used for the lock region.
4165 * @param[in] mode The Unix permissions for the file, if we create it.
4166 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4167 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4168 * @return 0 on success, non-zero on failure.
4171 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4174 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4176 # define MDB_ERRCODE_ROFS EROFS
4177 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4178 # define MDB_CLOEXEC O_CLOEXEC
4181 # define MDB_CLOEXEC 0
4188 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4189 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4190 FILE_ATTRIBUTE_NORMAL, NULL);
4192 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4194 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4196 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4201 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4202 /* Lose record locks when exec*() */
4203 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4204 fcntl(env->me_lfd, F_SETFD, fdflags);
4207 if (!(env->me_flags & MDB_NOTLS)) {
4208 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4211 env->me_flags |= MDB_ENV_TXKEY;
4213 /* Windows TLS callbacks need help finding their TLS info. */
4214 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4218 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4222 /* Try to get exclusive lock. If we succeed, then
4223 * nobody is using the lock region and we should initialize it.
4225 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4228 size = GetFileSize(env->me_lfd, NULL);
4230 size = lseek(env->me_lfd, 0, SEEK_END);
4231 if (size == -1) goto fail_errno;
4233 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4234 if (size < rsize && *excl > 0) {
4236 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4237 || !SetEndOfFile(env->me_lfd))
4240 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4244 size = rsize - sizeof(MDB_txninfo);
4245 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4250 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4252 if (!mh) goto fail_errno;
4253 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4255 if (!env->me_txns) goto fail_errno;
4257 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4259 if (m == MAP_FAILED) goto fail_errno;
4265 BY_HANDLE_FILE_INFORMATION stbuf;
4274 if (!mdb_sec_inited) {
4275 InitializeSecurityDescriptor(&mdb_null_sd,
4276 SECURITY_DESCRIPTOR_REVISION);
4277 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4278 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4279 mdb_all_sa.bInheritHandle = FALSE;
4280 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4283 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4284 idbuf.volume = stbuf.dwVolumeSerialNumber;
4285 idbuf.nhigh = stbuf.nFileIndexHigh;
4286 idbuf.nlow = stbuf.nFileIndexLow;
4287 val.mv_data = &idbuf;
4288 val.mv_size = sizeof(idbuf);
4289 mdb_hash_enc(&val, encbuf);
4290 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4291 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4292 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4293 if (!env->me_rmutex) goto fail_errno;
4294 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4295 if (!env->me_wmutex) goto fail_errno;
4296 #elif defined(MDB_USE_POSIX_SEM)
4305 #if defined(__NetBSD__)
4306 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4308 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4309 idbuf.dev = stbuf.st_dev;
4310 idbuf.ino = stbuf.st_ino;
4311 val.mv_data = &idbuf;
4312 val.mv_size = sizeof(idbuf);
4313 mdb_hash_enc(&val, encbuf);
4314 #ifdef MDB_SHORT_SEMNAMES
4315 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4317 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4318 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4319 /* Clean up after a previous run, if needed: Try to
4320 * remove both semaphores before doing anything else.
4322 sem_unlink(env->me_txns->mti_rmname);
4323 sem_unlink(env->me_txns->mti_wmname);
4324 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4325 O_CREAT|O_EXCL, mode, 1);
4326 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4327 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4328 O_CREAT|O_EXCL, mode, 1);
4329 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4330 #else /* MDB_USE_POSIX_SEM */
4331 pthread_mutexattr_t mattr;
4333 if ((rc = pthread_mutexattr_init(&mattr))
4334 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4335 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4336 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4338 pthread_mutexattr_destroy(&mattr);
4339 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4341 env->me_txns->mti_magic = MDB_MAGIC;
4342 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4343 env->me_txns->mti_txnid = 0;
4344 env->me_txns->mti_numreaders = 0;
4347 if (env->me_txns->mti_magic != MDB_MAGIC) {
4348 DPUTS("lock region has invalid magic");
4352 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4353 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4354 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4355 rc = MDB_VERSION_MISMATCH;
4359 if (rc && rc != EACCES && rc != EAGAIN) {
4363 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4364 if (!env->me_rmutex) goto fail_errno;
4365 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4366 if (!env->me_wmutex) goto fail_errno;
4367 #elif defined(MDB_USE_POSIX_SEM)
4368 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4369 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4370 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4371 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4382 /** The name of the lock file in the DB environment */
4383 #define LOCKNAME "/lock.mdb"
4384 /** The name of the data file in the DB environment */
4385 #define DATANAME "/data.mdb"
4386 /** The suffix of the lock file when no subdir is used */
4387 #define LOCKSUFF "-lock"
4388 /** Only a subset of the @ref mdb_env flags can be changed
4389 * at runtime. Changing other flags requires closing the
4390 * environment and re-opening it with the new flags.
4392 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4393 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4394 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4396 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4397 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4401 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4403 int oflags, rc, len, excl = -1;
4404 char *lpath, *dpath;
4406 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4410 if (flags & MDB_NOSUBDIR) {
4411 rc = len + sizeof(LOCKSUFF) + len + 1;
4413 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4418 if (flags & MDB_NOSUBDIR) {
4419 dpath = lpath + len + sizeof(LOCKSUFF);
4420 sprintf(lpath, "%s" LOCKSUFF, path);
4421 strcpy(dpath, path);
4423 dpath = lpath + len + sizeof(LOCKNAME);
4424 sprintf(lpath, "%s" LOCKNAME, path);
4425 sprintf(dpath, "%s" DATANAME, path);
4429 flags |= env->me_flags;
4430 if (flags & MDB_RDONLY) {
4431 /* silently ignore WRITEMAP when we're only getting read access */
4432 flags &= ~MDB_WRITEMAP;
4434 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4435 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4438 env->me_flags = flags |= MDB_ENV_ACTIVE;
4442 env->me_path = strdup(path);
4443 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4444 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4445 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4446 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4451 /* For RDONLY, get lockfile after we know datafile exists */
4452 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4453 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4459 if (F_ISSET(flags, MDB_RDONLY)) {
4460 oflags = GENERIC_READ;
4461 len = OPEN_EXISTING;
4463 oflags = GENERIC_READ|GENERIC_WRITE;
4466 mode = FILE_ATTRIBUTE_NORMAL;
4467 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4468 NULL, len, mode, NULL);
4470 if (F_ISSET(flags, MDB_RDONLY))
4473 oflags = O_RDWR | O_CREAT;
4475 env->me_fd = open(dpath, oflags, mode);
4477 if (env->me_fd == INVALID_HANDLE_VALUE) {
4482 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4483 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4488 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4489 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4490 env->me_mfd = env->me_fd;
4492 /* Synchronous fd for meta writes. Needed even with
4493 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4496 len = OPEN_EXISTING;
4497 env->me_mfd = CreateFile(dpath, oflags,
4498 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4499 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4502 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4504 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4509 DPRINTF(("opened dbenv %p", (void *) env));
4511 rc = mdb_env_share_locks(env, &excl);
4515 if (!((flags & MDB_RDONLY) ||
4516 (env->me_pbuf = calloc(1, env->me_psize))))
4518 if (!(flags & MDB_RDONLY)) {
4520 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4521 (sizeof(MDB_db)+sizeof(MDB_cursor)+sizeof(unsigned int)+1);
4522 txn = calloc(1, size);
4524 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4525 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4526 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4527 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4538 mdb_env_close0(env, excl);
4544 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4546 mdb_env_close0(MDB_env *env, int excl)
4550 if (!(env->me_flags & MDB_ENV_ACTIVE))
4553 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4554 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4555 free(env->me_dbxs[i].md_name.mv_data);
4558 free(env->me_dbiseqs);
4559 free(env->me_dbflags);
4562 free(env->me_dirty_list);
4564 mdb_midl_free(env->me_free_pgs);
4566 if (env->me_flags & MDB_ENV_TXKEY) {
4567 pthread_key_delete(env->me_txkey);
4569 /* Delete our key from the global list */
4570 for (i=0; i<mdb_tls_nkeys; i++)
4571 if (mdb_tls_keys[i] == env->me_txkey) {
4572 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4580 munmap(env->me_map, env->me_mapsize);
4582 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4583 (void) close(env->me_mfd);
4584 if (env->me_fd != INVALID_HANDLE_VALUE)
4585 (void) close(env->me_fd);
4587 MDB_PID_T pid = env->me_pid;
4588 /* Clearing readers is done in this function because
4589 * me_txkey with its destructor must be disabled first.
4591 for (i = env->me_numreaders; --i >= 0; )
4592 if (env->me_txns->mti_readers[i].mr_pid == pid)
4593 env->me_txns->mti_readers[i].mr_pid = 0;
4595 if (env->me_rmutex) {
4596 CloseHandle(env->me_rmutex);
4597 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4599 /* Windows automatically destroys the mutexes when
4600 * the last handle closes.
4602 #elif defined(MDB_USE_POSIX_SEM)
4603 if (env->me_rmutex != SEM_FAILED) {
4604 sem_close(env->me_rmutex);
4605 if (env->me_wmutex != SEM_FAILED)
4606 sem_close(env->me_wmutex);
4607 /* If we have the filelock: If we are the
4608 * only remaining user, clean up semaphores.
4611 mdb_env_excl_lock(env, &excl);
4613 sem_unlink(env->me_txns->mti_rmname);
4614 sem_unlink(env->me_txns->mti_wmname);
4618 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4620 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4623 /* Unlock the lockfile. Windows would have unlocked it
4624 * after closing anyway, but not necessarily at once.
4626 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4629 (void) close(env->me_lfd);
4632 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4637 mdb_env_close(MDB_env *env)
4644 VGMEMP_DESTROY(env);
4645 while ((dp = env->me_dpages) != NULL) {
4646 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4647 env->me_dpages = dp->mp_next;
4651 mdb_env_close0(env, 0);
4655 /** Compare two items pointing at aligned size_t's */
4657 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4659 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4660 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4663 /** Compare two items pointing at aligned unsigned int's */
4665 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4667 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4668 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4671 /** Compare two items pointing at unsigned ints of unknown alignment.
4672 * Nodes and keys are guaranteed to be 2-byte aligned.
4675 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4677 #if BYTE_ORDER == LITTLE_ENDIAN
4678 unsigned short *u, *c;
4681 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4682 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4685 } while(!x && u > (unsigned short *)a->mv_data);
4688 unsigned short *u, *c, *end;
4691 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4692 u = (unsigned short *)a->mv_data;
4693 c = (unsigned short *)b->mv_data;
4696 } while(!x && u < end);
4701 /** Compare two items pointing at size_t's of unknown alignment. */
4702 #ifdef MISALIGNED_OK
4703 # define mdb_cmp_clong mdb_cmp_long
4705 # define mdb_cmp_clong mdb_cmp_cint
4708 /** Compare two items lexically */
4710 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4717 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4723 diff = memcmp(a->mv_data, b->mv_data, len);
4724 return diff ? diff : len_diff<0 ? -1 : len_diff;
4727 /** Compare two items in reverse byte order */
4729 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4731 const unsigned char *p1, *p2, *p1_lim;
4735 p1_lim = (const unsigned char *)a->mv_data;
4736 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4737 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4739 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4745 while (p1 > p1_lim) {
4746 diff = *--p1 - *--p2;
4750 return len_diff<0 ? -1 : len_diff;
4753 /** Search for key within a page, using binary search.
4754 * Returns the smallest entry larger or equal to the key.
4755 * If exactp is non-null, stores whether the found entry was an exact match
4756 * in *exactp (1 or 0).
4757 * Updates the cursor index with the index of the found entry.
4758 * If no entry larger or equal to the key is found, returns NULL.
4761 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4763 unsigned int i = 0, nkeys;
4766 MDB_page *mp = mc->mc_pg[mc->mc_top];
4767 MDB_node *node = NULL;
4772 nkeys = NUMKEYS(mp);
4774 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4775 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4778 low = IS_LEAF(mp) ? 0 : 1;
4780 cmp = mc->mc_dbx->md_cmp;
4782 /* Branch pages have no data, so if using integer keys,
4783 * alignment is guaranteed. Use faster mdb_cmp_int.
4785 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4786 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4793 nodekey.mv_size = mc->mc_db->md_pad;
4794 node = NODEPTR(mp, 0); /* fake */
4795 while (low <= high) {
4796 i = (low + high) >> 1;
4797 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4798 rc = cmp(key, &nodekey);
4799 DPRINTF(("found leaf index %u [%s], rc = %i",
4800 i, DKEY(&nodekey), rc));
4809 while (low <= high) {
4810 i = (low + high) >> 1;
4812 node = NODEPTR(mp, i);
4813 nodekey.mv_size = NODEKSZ(node);
4814 nodekey.mv_data = NODEKEY(node);
4816 rc = cmp(key, &nodekey);
4819 DPRINTF(("found leaf index %u [%s], rc = %i",
4820 i, DKEY(&nodekey), rc));
4822 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4823 i, DKEY(&nodekey), NODEPGNO(node), rc));
4834 if (rc > 0) { /* Found entry is less than the key. */
4835 i++; /* Skip to get the smallest entry larger than key. */
4837 node = NODEPTR(mp, i);
4840 *exactp = (rc == 0 && nkeys > 0);
4841 /* store the key index */
4842 mc->mc_ki[mc->mc_top] = i;
4844 /* There is no entry larger or equal to the key. */
4847 /* nodeptr is fake for LEAF2 */
4853 mdb_cursor_adjust(MDB_cursor *mc, func)
4857 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4858 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4865 /** Pop a page off the top of the cursor's stack. */
4867 mdb_cursor_pop(MDB_cursor *mc)
4871 MDB_page *top = mc->mc_pg[mc->mc_top];
4877 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4878 DDBI(mc), (void *) mc));
4882 /** Push a page onto the top of the cursor's stack. */
4884 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4886 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4887 DDBI(mc), (void *) mc));
4889 if (mc->mc_snum >= CURSOR_STACK) {
4890 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4891 return MDB_CURSOR_FULL;
4894 mc->mc_top = mc->mc_snum++;
4895 mc->mc_pg[mc->mc_top] = mp;
4896 mc->mc_ki[mc->mc_top] = 0;
4901 /** Find the address of the page corresponding to a given page number.
4902 * @param[in] txn the transaction for this access.
4903 * @param[in] pgno the page number for the page to retrieve.
4904 * @param[out] ret address of a pointer where the page's address will be stored.
4905 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4906 * @return 0 on success, non-zero on failure.
4909 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4911 MDB_env *env = txn->mt_env;
4915 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4919 MDB_ID2L dl = tx2->mt_u.dirty_list;
4921 /* Spilled pages were dirtied in this txn and flushed
4922 * because the dirty list got full. Bring this page
4923 * back in from the map (but don't unspill it here,
4924 * leave that unless page_touch happens again).
4926 if (tx2->mt_spill_pgs) {
4927 MDB_ID pn = pgno << 1;
4928 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4929 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4930 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4935 unsigned x = mdb_mid2l_search(dl, pgno);
4936 if (x <= dl[0].mid && dl[x].mid == pgno) {
4942 } while ((tx2 = tx2->mt_parent) != NULL);
4945 if (pgno < txn->mt_next_pgno) {
4947 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4949 DPRINTF(("page %"Z"u not found", pgno));
4950 txn->mt_flags |= MDB_TXN_ERROR;
4951 return MDB_PAGE_NOTFOUND;
4961 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4962 * The cursor is at the root page, set up the rest of it.
4965 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4967 MDB_page *mp = mc->mc_pg[mc->mc_top];
4971 while (IS_BRANCH(mp)) {
4975 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4976 mdb_cassert(mc, NUMKEYS(mp) > 1);
4977 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4979 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4981 if (flags & MDB_PS_LAST)
4982 i = NUMKEYS(mp) - 1;
4985 node = mdb_node_search(mc, key, &exact);
4987 i = NUMKEYS(mp) - 1;
4989 i = mc->mc_ki[mc->mc_top];
4991 mdb_cassert(mc, i > 0);
4995 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4998 mdb_cassert(mc, i < NUMKEYS(mp));
4999 node = NODEPTR(mp, i);
5001 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5004 mc->mc_ki[mc->mc_top] = i;
5005 if ((rc = mdb_cursor_push(mc, mp)))
5008 if (flags & MDB_PS_MODIFY) {
5009 if ((rc = mdb_page_touch(mc)) != 0)
5011 mp = mc->mc_pg[mc->mc_top];
5016 DPRINTF(("internal error, index points to a %02X page!?",
5018 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5019 return MDB_CORRUPTED;
5022 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5023 key ? DKEY(key) : "null"));
5024 mc->mc_flags |= C_INITIALIZED;
5025 mc->mc_flags &= ~C_EOF;
5030 /** Search for the lowest key under the current branch page.
5031 * This just bypasses a NUMKEYS check in the current page
5032 * before calling mdb_page_search_root(), because the callers
5033 * are all in situations where the current page is known to
5037 mdb_page_search_lowest(MDB_cursor *mc)
5039 MDB_page *mp = mc->mc_pg[mc->mc_top];
5040 MDB_node *node = NODEPTR(mp, 0);
5043 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5046 mc->mc_ki[mc->mc_top] = 0;
5047 if ((rc = mdb_cursor_push(mc, mp)))
5049 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5052 /** Search for the page a given key should be in.
5053 * Push it and its parent pages on the cursor stack.
5054 * @param[in,out] mc the cursor for this operation.
5055 * @param[in] key the key to search for, or NULL for first/last page.
5056 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5057 * are touched (updated with new page numbers).
5058 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5059 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5060 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5061 * @return 0 on success, non-zero on failure.
5064 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5069 /* Make sure the txn is still viable, then find the root from
5070 * the txn's db table and set it as the root of the cursor's stack.
5072 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5073 DPUTS("transaction has failed, must abort");
5076 /* Make sure we're using an up-to-date root */
5077 if (*mc->mc_dbflag & DB_STALE) {
5079 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5081 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5082 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5089 MDB_node *leaf = mdb_node_search(&mc2,
5090 &mc->mc_dbx->md_name, &exact);
5092 return MDB_NOTFOUND;
5093 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5096 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5098 /* The txn may not know this DBI, or another process may
5099 * have dropped and recreated the DB with other flags.
5101 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5102 return MDB_INCOMPATIBLE;
5103 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5105 *mc->mc_dbflag &= ~DB_STALE;
5107 root = mc->mc_db->md_root;
5109 if (root == P_INVALID) { /* Tree is empty. */
5110 DPUTS("tree is empty");
5111 return MDB_NOTFOUND;
5115 mdb_cassert(mc, root > 1);
5116 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5117 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5123 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5124 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5126 if (flags & MDB_PS_MODIFY) {
5127 if ((rc = mdb_page_touch(mc)))
5131 if (flags & MDB_PS_ROOTONLY)
5134 return mdb_page_search_root(mc, key, flags);
5138 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5140 MDB_txn *txn = mc->mc_txn;
5141 pgno_t pg = mp->mp_pgno;
5142 unsigned x = 0, ovpages = mp->mp_pages;
5143 MDB_env *env = txn->mt_env;
5144 MDB_IDL sl = txn->mt_spill_pgs;
5145 MDB_ID pn = pg << 1;
5148 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5149 /* If the page is dirty or on the spill list we just acquired it,
5150 * so we should give it back to our current free list, if any.
5151 * Otherwise put it onto the list of pages we freed in this txn.
5153 * Won't create me_pghead: me_pglast must be inited along with it.
5154 * Unsupported in nested txns: They would need to hide the page
5155 * range in ancestor txns' dirty and spilled lists.
5157 if (env->me_pghead &&
5159 ((mp->mp_flags & P_DIRTY) ||
5160 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5164 MDB_ID2 *dl, ix, iy;
5165 rc = mdb_midl_need(&env->me_pghead, ovpages);
5168 if (!(mp->mp_flags & P_DIRTY)) {
5169 /* This page is no longer spilled */
5176 /* Remove from dirty list */
5177 dl = txn->mt_u.dirty_list;
5179 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5185 mdb_cassert(mc, x > 1);
5187 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5188 txn->mt_flags |= MDB_TXN_ERROR;
5189 return MDB_CORRUPTED;
5192 if (!(env->me_flags & MDB_WRITEMAP))
5193 mdb_dpage_free(env, mp);
5195 /* Insert in me_pghead */
5196 mop = env->me_pghead;
5197 j = mop[0] + ovpages;
5198 for (i = mop[0]; i && mop[i] < pg; i--)
5204 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5208 mc->mc_db->md_overflow_pages -= ovpages;
5212 /** Return the data associated with a given node.
5213 * @param[in] txn The transaction for this operation.
5214 * @param[in] leaf The node being read.
5215 * @param[out] data Updated to point to the node's data.
5216 * @return 0 on success, non-zero on failure.
5219 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5221 MDB_page *omp; /* overflow page */
5225 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5226 data->mv_size = NODEDSZ(leaf);
5227 data->mv_data = NODEDATA(leaf);
5231 /* Read overflow data.
5233 data->mv_size = NODEDSZ(leaf);
5234 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5235 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5236 DPRINTF(("read overflow page %"Z"u failed", pgno));
5239 data->mv_data = METADATA(omp);
5245 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5246 MDB_val *key, MDB_val *data)
5253 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5255 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5258 if (txn->mt_flags & MDB_TXN_ERROR)
5261 mdb_cursor_init(&mc, txn, dbi, &mx);
5262 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5265 /** Find a sibling for a page.
5266 * Replaces the page at the top of the cursor's stack with the
5267 * specified sibling, if one exists.
5268 * @param[in] mc The cursor for this operation.
5269 * @param[in] move_right Non-zero if the right sibling is requested,
5270 * otherwise the left sibling.
5271 * @return 0 on success, non-zero on failure.
5274 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5280 if (mc->mc_snum < 2) {
5281 return MDB_NOTFOUND; /* root has no siblings */
5285 DPRINTF(("parent page is page %"Z"u, index %u",
5286 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5288 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5289 : (mc->mc_ki[mc->mc_top] == 0)) {
5290 DPRINTF(("no more keys left, moving to %s sibling",
5291 move_right ? "right" : "left"));
5292 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5293 /* undo cursor_pop before returning */
5300 mc->mc_ki[mc->mc_top]++;
5302 mc->mc_ki[mc->mc_top]--;
5303 DPRINTF(("just moving to %s index key %u",
5304 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5306 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5308 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5309 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5310 /* mc will be inconsistent if caller does mc_snum++ as above */
5311 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5315 mdb_cursor_push(mc, mp);
5317 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5322 /** Move the cursor to the next data item. */
5324 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5330 if (mc->mc_flags & C_EOF) {
5331 return MDB_NOTFOUND;
5334 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5336 mp = mc->mc_pg[mc->mc_top];
5338 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5339 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5340 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5341 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5342 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5343 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5344 if (rc == MDB_SUCCESS)
5345 MDB_GET_KEY(leaf, key);
5350 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5351 if (op == MDB_NEXT_DUP)
5352 return MDB_NOTFOUND;
5356 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5357 mdb_dbg_pgno(mp), (void *) mc));
5358 if (mc->mc_flags & C_DEL)
5361 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5362 DPUTS("=====> move to next sibling page");
5363 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5364 mc->mc_flags |= C_EOF;
5367 mp = mc->mc_pg[mc->mc_top];
5368 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5370 mc->mc_ki[mc->mc_top]++;
5373 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5374 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5377 key->mv_size = mc->mc_db->md_pad;
5378 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5382 mdb_cassert(mc, IS_LEAF(mp));
5383 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5385 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5386 mdb_xcursor_init1(mc, leaf);
5389 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5392 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5393 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5394 if (rc != MDB_SUCCESS)
5399 MDB_GET_KEY(leaf, key);
5403 /** Move the cursor to the previous data item. */
5405 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5411 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5413 mp = mc->mc_pg[mc->mc_top];
5415 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5416 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5417 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5418 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5419 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5420 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5421 if (rc == MDB_SUCCESS) {
5422 MDB_GET_KEY(leaf, key);
5423 mc->mc_flags &= ~C_EOF;
5428 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5429 if (op == MDB_PREV_DUP)
5430 return MDB_NOTFOUND;
5435 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5436 mdb_dbg_pgno(mp), (void *) mc));
5438 if (mc->mc_ki[mc->mc_top] == 0) {
5439 DPUTS("=====> move to prev sibling page");
5440 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5443 mp = mc->mc_pg[mc->mc_top];
5444 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5445 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5447 mc->mc_ki[mc->mc_top]--;
5449 mc->mc_flags &= ~C_EOF;
5451 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5452 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5455 key->mv_size = mc->mc_db->md_pad;
5456 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5460 mdb_cassert(mc, IS_LEAF(mp));
5461 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5463 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5464 mdb_xcursor_init1(mc, leaf);
5467 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5470 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5471 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5472 if (rc != MDB_SUCCESS)
5477 MDB_GET_KEY(leaf, key);
5481 /** Set the cursor on a specific data item. */
5483 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5484 MDB_cursor_op op, int *exactp)
5488 MDB_node *leaf = NULL;
5491 if (key->mv_size == 0)
5492 return MDB_BAD_VALSIZE;
5495 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5497 /* See if we're already on the right page */
5498 if (mc->mc_flags & C_INITIALIZED) {
5501 mp = mc->mc_pg[mc->mc_top];
5503 mc->mc_ki[mc->mc_top] = 0;
5504 return MDB_NOTFOUND;
5506 if (mp->mp_flags & P_LEAF2) {
5507 nodekey.mv_size = mc->mc_db->md_pad;
5508 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5510 leaf = NODEPTR(mp, 0);
5511 MDB_GET_KEY2(leaf, nodekey);
5513 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5515 /* Probably happens rarely, but first node on the page
5516 * was the one we wanted.
5518 mc->mc_ki[mc->mc_top] = 0;
5525 unsigned int nkeys = NUMKEYS(mp);
5527 if (mp->mp_flags & P_LEAF2) {
5528 nodekey.mv_data = LEAF2KEY(mp,
5529 nkeys-1, nodekey.mv_size);
5531 leaf = NODEPTR(mp, nkeys-1);
5532 MDB_GET_KEY2(leaf, nodekey);
5534 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5536 /* last node was the one we wanted */
5537 mc->mc_ki[mc->mc_top] = nkeys-1;
5543 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5544 /* This is definitely the right page, skip search_page */
5545 if (mp->mp_flags & P_LEAF2) {
5546 nodekey.mv_data = LEAF2KEY(mp,
5547 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5549 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5550 MDB_GET_KEY2(leaf, nodekey);
5552 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5554 /* current node was the one we wanted */
5564 /* If any parents have right-sibs, search.
5565 * Otherwise, there's nothing further.
5567 for (i=0; i<mc->mc_top; i++)
5569 NUMKEYS(mc->mc_pg[i])-1)
5571 if (i == mc->mc_top) {
5572 /* There are no other pages */
5573 mc->mc_ki[mc->mc_top] = nkeys;
5574 return MDB_NOTFOUND;
5578 /* There are no other pages */
5579 mc->mc_ki[mc->mc_top] = 0;
5580 if (op == MDB_SET_RANGE && !exactp) {
5584 return MDB_NOTFOUND;
5588 rc = mdb_page_search(mc, key, 0);
5589 if (rc != MDB_SUCCESS)
5592 mp = mc->mc_pg[mc->mc_top];
5593 mdb_cassert(mc, IS_LEAF(mp));
5596 leaf = mdb_node_search(mc, key, exactp);
5597 if (exactp != NULL && !*exactp) {
5598 /* MDB_SET specified and not an exact match. */
5599 return MDB_NOTFOUND;
5603 DPUTS("===> inexact leaf not found, goto sibling");
5604 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5605 return rc; /* no entries matched */
5606 mp = mc->mc_pg[mc->mc_top];
5607 mdb_cassert(mc, IS_LEAF(mp));
5608 leaf = NODEPTR(mp, 0);
5612 mc->mc_flags |= C_INITIALIZED;
5613 mc->mc_flags &= ~C_EOF;
5616 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5617 key->mv_size = mc->mc_db->md_pad;
5618 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5623 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5624 mdb_xcursor_init1(mc, leaf);
5627 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5628 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5629 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5632 if (op == MDB_GET_BOTH) {
5638 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5639 if (rc != MDB_SUCCESS)
5642 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5644 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5646 rc = mc->mc_dbx->md_dcmp(data, &d2);
5648 if (op == MDB_GET_BOTH || rc > 0)
5649 return MDB_NOTFOUND;
5656 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5657 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5662 /* The key already matches in all other cases */
5663 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5664 MDB_GET_KEY(leaf, key);
5665 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5670 /** Move the cursor to the first item in the database. */
5672 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5678 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5680 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5681 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5682 if (rc != MDB_SUCCESS)
5685 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5687 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5688 mc->mc_flags |= C_INITIALIZED;
5689 mc->mc_flags &= ~C_EOF;
5691 mc->mc_ki[mc->mc_top] = 0;
5693 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5694 key->mv_size = mc->mc_db->md_pad;
5695 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5700 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5701 mdb_xcursor_init1(mc, leaf);
5702 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5706 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5710 MDB_GET_KEY(leaf, key);
5714 /** Move the cursor to the last item in the database. */
5716 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5722 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5724 if (!(mc->mc_flags & C_EOF)) {
5726 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5727 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5728 if (rc != MDB_SUCCESS)
5731 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5734 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5735 mc->mc_flags |= C_INITIALIZED|C_EOF;
5736 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5738 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5739 key->mv_size = mc->mc_db->md_pad;
5740 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5745 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5746 mdb_xcursor_init1(mc, leaf);
5747 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5751 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5756 MDB_GET_KEY(leaf, key);
5761 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5766 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5771 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5775 case MDB_GET_CURRENT:
5776 if (!(mc->mc_flags & C_INITIALIZED)) {
5779 MDB_page *mp = mc->mc_pg[mc->mc_top];
5780 int nkeys = NUMKEYS(mp);
5781 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5782 mc->mc_ki[mc->mc_top] = nkeys;
5788 key->mv_size = mc->mc_db->md_pad;
5789 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5791 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5792 MDB_GET_KEY(leaf, key);
5794 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5795 if (mc->mc_flags & C_DEL)
5796 mdb_xcursor_init1(mc, leaf);
5797 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5799 rc = mdb_node_read(mc->mc_txn, leaf, data);
5806 case MDB_GET_BOTH_RANGE:
5811 if (mc->mc_xcursor == NULL) {
5812 rc = MDB_INCOMPATIBLE;
5822 rc = mdb_cursor_set(mc, key, data, op,
5823 op == MDB_SET_RANGE ? NULL : &exact);
5826 case MDB_GET_MULTIPLE:
5827 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5831 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5832 rc = MDB_INCOMPATIBLE;
5836 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5837 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5840 case MDB_NEXT_MULTIPLE:
5845 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5846 rc = MDB_INCOMPATIBLE;
5849 if (!(mc->mc_flags & C_INITIALIZED))
5850 rc = mdb_cursor_first(mc, key, data);
5852 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5853 if (rc == MDB_SUCCESS) {
5854 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5857 mx = &mc->mc_xcursor->mx_cursor;
5858 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5860 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5861 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5869 case MDB_NEXT_NODUP:
5870 if (!(mc->mc_flags & C_INITIALIZED))
5871 rc = mdb_cursor_first(mc, key, data);
5873 rc = mdb_cursor_next(mc, key, data, op);
5877 case MDB_PREV_NODUP:
5878 if (!(mc->mc_flags & C_INITIALIZED)) {
5879 rc = mdb_cursor_last(mc, key, data);
5882 mc->mc_flags |= C_INITIALIZED;
5883 mc->mc_ki[mc->mc_top]++;
5885 rc = mdb_cursor_prev(mc, key, data, op);
5888 rc = mdb_cursor_first(mc, key, data);
5891 mfunc = mdb_cursor_first;
5893 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5897 if (mc->mc_xcursor == NULL) {
5898 rc = MDB_INCOMPATIBLE;
5902 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5903 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5904 MDB_GET_KEY(leaf, key);
5905 rc = mdb_node_read(mc->mc_txn, leaf, data);
5909 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5913 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5916 rc = mdb_cursor_last(mc, key, data);
5919 mfunc = mdb_cursor_last;
5922 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5927 if (mc->mc_flags & C_DEL)
5928 mc->mc_flags ^= C_DEL;
5933 /** Touch all the pages in the cursor stack. Set mc_top.
5934 * Makes sure all the pages are writable, before attempting a write operation.
5935 * @param[in] mc The cursor to operate on.
5938 mdb_cursor_touch(MDB_cursor *mc)
5940 int rc = MDB_SUCCESS;
5942 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5945 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5947 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5948 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5951 *mc->mc_dbflag |= DB_DIRTY;
5956 rc = mdb_page_touch(mc);
5957 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5958 mc->mc_top = mc->mc_snum-1;
5963 /** Do not spill pages to disk if txn is getting full, may fail instead */
5964 #define MDB_NOSPILL 0x8000
5967 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5970 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5972 MDB_node *leaf = NULL;
5975 MDB_val xdata, *rdata, dkey, olddata;
5977 int do_sub = 0, insert_key, insert_data;
5978 unsigned int mcount = 0, dcount = 0, nospill;
5981 unsigned int nflags;
5984 if (mc == NULL || key == NULL)
5987 env = mc->mc_txn->mt_env;
5989 /* Check this first so counter will always be zero on any
5992 if (flags & MDB_MULTIPLE) {
5993 dcount = data[1].mv_size;
5994 data[1].mv_size = 0;
5995 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5996 return MDB_INCOMPATIBLE;
5999 nospill = flags & MDB_NOSPILL;
6000 flags &= ~MDB_NOSPILL;
6002 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6003 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6005 if (key->mv_size-1 >= ENV_MAXKEY(env))
6006 return MDB_BAD_VALSIZE;
6008 #if SIZE_MAX > MAXDATASIZE
6009 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6010 return MDB_BAD_VALSIZE;
6012 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6013 return MDB_BAD_VALSIZE;
6016 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6017 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6021 if (flags == MDB_CURRENT) {
6022 if (!(mc->mc_flags & C_INITIALIZED))
6025 } else if (mc->mc_db->md_root == P_INVALID) {
6026 /* new database, cursor has nothing to point to */
6029 mc->mc_flags &= ~C_INITIALIZED;
6034 if (flags & MDB_APPEND) {
6036 rc = mdb_cursor_last(mc, &k2, &d2);
6038 rc = mc->mc_dbx->md_cmp(key, &k2);
6041 mc->mc_ki[mc->mc_top]++;
6043 /* new key is <= last key */
6048 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6050 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6051 DPRINTF(("duplicate key [%s]", DKEY(key)));
6053 return MDB_KEYEXIST;
6055 if (rc && rc != MDB_NOTFOUND)
6059 if (mc->mc_flags & C_DEL)
6060 mc->mc_flags ^= C_DEL;
6062 /* Cursor is positioned, check for room in the dirty list */
6064 if (flags & MDB_MULTIPLE) {
6066 xdata.mv_size = data->mv_size * dcount;
6070 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6074 if (rc == MDB_NO_ROOT) {
6076 /* new database, write a root leaf page */
6077 DPUTS("allocating new root leaf page");
6078 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6081 mdb_cursor_push(mc, np);
6082 mc->mc_db->md_root = np->mp_pgno;
6083 mc->mc_db->md_depth++;
6084 *mc->mc_dbflag |= DB_DIRTY;
6085 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6087 np->mp_flags |= P_LEAF2;
6088 mc->mc_flags |= C_INITIALIZED;
6090 /* make sure all cursor pages are writable */
6091 rc2 = mdb_cursor_touch(mc);
6096 insert_key = insert_data = rc;
6098 /* The key does not exist */
6099 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6100 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6101 LEAFSIZE(key, data) > env->me_nodemax)
6103 /* Too big for a node, insert in sub-DB. Set up an empty
6104 * "old sub-page" for prep_subDB to expand to a full page.
6106 fp_flags = P_LEAF|P_DIRTY;
6108 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6109 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6110 olddata.mv_size = PAGEHDRSZ;
6114 /* there's only a key anyway, so this is a no-op */
6115 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6117 unsigned int ksize = mc->mc_db->md_pad;
6118 if (key->mv_size != ksize)
6119 return MDB_BAD_VALSIZE;
6120 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6121 memcpy(ptr, key->mv_data, ksize);
6123 /* if overwriting slot 0 of leaf, need to
6124 * update branch key if there is a parent page
6126 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6127 unsigned short top = mc->mc_top;
6129 /* slot 0 is always an empty key, find real slot */
6130 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6132 if (mc->mc_ki[mc->mc_top])
6133 rc2 = mdb_update_key(mc, key);
6144 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6145 olddata.mv_size = NODEDSZ(leaf);
6146 olddata.mv_data = NODEDATA(leaf);
6149 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6150 /* Prepare (sub-)page/sub-DB to accept the new item,
6151 * if needed. fp: old sub-page or a header faking
6152 * it. mp: new (sub-)page. offset: growth in page
6153 * size. xdata: node data with new page or DB.
6155 unsigned i, offset = 0;
6156 mp = fp = xdata.mv_data = env->me_pbuf;
6157 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6159 /* Was a single item before, must convert now */
6160 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6161 /* Just overwrite the current item */
6162 if (flags == MDB_CURRENT)
6165 #if UINT_MAX < SIZE_MAX
6166 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6167 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6169 /* does data match? */
6170 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6171 if (flags & MDB_NODUPDATA)
6172 return MDB_KEYEXIST;
6177 /* Back up original data item */
6178 dkey.mv_size = olddata.mv_size;
6179 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6181 /* Make sub-page header for the dup items, with dummy body */
6182 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6183 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6184 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6185 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6186 fp->mp_flags |= P_LEAF2;
6187 fp->mp_pad = data->mv_size;
6188 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6190 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6191 (dkey.mv_size & 1) + (data->mv_size & 1);
6193 fp->mp_upper = xdata.mv_size - PAGEBASE;
6194 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6195 } else if (leaf->mn_flags & F_SUBDATA) {
6196 /* Data is on sub-DB, just store it */
6197 flags |= F_DUPDATA|F_SUBDATA;
6200 /* Data is on sub-page */
6201 fp = olddata.mv_data;
6204 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6205 offset = EVEN(NODESIZE + sizeof(indx_t) +
6209 offset = fp->mp_pad;
6210 if (SIZELEFT(fp) < offset) {
6211 offset *= 4; /* space for 4 more */
6214 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6216 fp->mp_flags |= P_DIRTY;
6217 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6218 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6222 xdata.mv_size = olddata.mv_size + offset;
6225 fp_flags = fp->mp_flags;
6226 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6227 /* Too big for a sub-page, convert to sub-DB */
6228 fp_flags &= ~P_SUBP;
6230 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6231 fp_flags |= P_LEAF2;
6232 dummy.md_pad = fp->mp_pad;
6233 dummy.md_flags = MDB_DUPFIXED;
6234 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6235 dummy.md_flags |= MDB_INTEGERKEY;
6241 dummy.md_branch_pages = 0;
6242 dummy.md_leaf_pages = 1;
6243 dummy.md_overflow_pages = 0;
6244 dummy.md_entries = NUMKEYS(fp);
6245 xdata.mv_size = sizeof(MDB_db);
6246 xdata.mv_data = &dummy;
6247 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6249 offset = env->me_psize - olddata.mv_size;
6250 flags |= F_DUPDATA|F_SUBDATA;
6251 dummy.md_root = mp->mp_pgno;
6254 mp->mp_flags = fp_flags | P_DIRTY;
6255 mp->mp_pad = fp->mp_pad;
6256 mp->mp_lower = fp->mp_lower;
6257 mp->mp_upper = fp->mp_upper + offset;
6258 if (fp_flags & P_LEAF2) {
6259 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6261 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6262 olddata.mv_size - fp->mp_upper - PAGEBASE);
6263 for (i=0; i<NUMKEYS(fp); i++)
6264 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6272 mdb_node_del(mc, 0);
6276 /* overflow page overwrites need special handling */
6277 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6280 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6282 memcpy(&pg, olddata.mv_data, sizeof(pg));
6283 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6285 ovpages = omp->mp_pages;
6287 /* Is the ov page large enough? */
6288 if (ovpages >= dpages) {
6289 if (!(omp->mp_flags & P_DIRTY) &&
6290 (level || (env->me_flags & MDB_WRITEMAP)))
6292 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6295 level = 0; /* dirty in this txn or clean */
6298 if (omp->mp_flags & P_DIRTY) {
6299 /* yes, overwrite it. Note in this case we don't
6300 * bother to try shrinking the page if the new data
6301 * is smaller than the overflow threshold.
6304 /* It is writable only in a parent txn */
6305 size_t sz = (size_t) env->me_psize * ovpages, off;
6306 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6312 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6313 mdb_cassert(mc, rc2 == 0);
6314 if (!(flags & MDB_RESERVE)) {
6315 /* Copy end of page, adjusting alignment so
6316 * compiler may copy words instead of bytes.
6318 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6319 memcpy((size_t *)((char *)np + off),
6320 (size_t *)((char *)omp + off), sz - off);
6323 memcpy(np, omp, sz); /* Copy beginning of page */
6326 SETDSZ(leaf, data->mv_size);
6327 if (F_ISSET(flags, MDB_RESERVE))
6328 data->mv_data = METADATA(omp);
6330 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6334 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6336 } else if (data->mv_size == olddata.mv_size) {
6337 /* same size, just replace it. Note that we could
6338 * also reuse this node if the new data is smaller,
6339 * but instead we opt to shrink the node in that case.
6341 if (F_ISSET(flags, MDB_RESERVE))
6342 data->mv_data = olddata.mv_data;
6343 else if (!(mc->mc_flags & C_SUB))
6344 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6346 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6351 mdb_node_del(mc, 0);
6357 nflags = flags & NODE_ADD_FLAGS;
6358 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6359 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6360 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6361 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6363 nflags |= MDB_SPLIT_REPLACE;
6364 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6366 /* There is room already in this leaf page. */
6367 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6368 if (rc == 0 && insert_key) {
6369 /* Adjust other cursors pointing to mp */
6370 MDB_cursor *m2, *m3;
6371 MDB_dbi dbi = mc->mc_dbi;
6372 unsigned i = mc->mc_top;
6373 MDB_page *mp = mc->mc_pg[i];
6375 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6376 if (mc->mc_flags & C_SUB)
6377 m3 = &m2->mc_xcursor->mx_cursor;
6380 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6381 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6388 if (rc == MDB_SUCCESS) {
6389 /* Now store the actual data in the child DB. Note that we're
6390 * storing the user data in the keys field, so there are strict
6391 * size limits on dupdata. The actual data fields of the child
6392 * DB are all zero size.
6400 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6401 if (flags & MDB_CURRENT) {
6402 xflags = MDB_CURRENT|MDB_NOSPILL;
6404 mdb_xcursor_init1(mc, leaf);
6405 xflags = (flags & MDB_NODUPDATA) ?
6406 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6408 /* converted, write the original data first */
6410 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6414 /* Adjust other cursors pointing to mp */
6416 unsigned i = mc->mc_top;
6417 MDB_page *mp = mc->mc_pg[i];
6419 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6420 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6421 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6422 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6423 mdb_xcursor_init1(m2, leaf);
6427 /* we've done our job */
6430 ecount = mc->mc_xcursor->mx_db.md_entries;
6431 if (flags & MDB_APPENDDUP)
6432 xflags |= MDB_APPEND;
6433 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6434 if (flags & F_SUBDATA) {
6435 void *db = NODEDATA(leaf);
6436 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6438 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6440 /* Increment count unless we just replaced an existing item. */
6442 mc->mc_db->md_entries++;
6444 /* Invalidate txn if we created an empty sub-DB */
6447 /* If we succeeded and the key didn't exist before,
6448 * make sure the cursor is marked valid.
6450 mc->mc_flags |= C_INITIALIZED;
6452 if (flags & MDB_MULTIPLE) {
6455 /* let caller know how many succeeded, if any */
6456 data[1].mv_size = mcount;
6457 if (mcount < dcount) {
6458 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6459 insert_key = insert_data = 0;
6466 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6469 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6474 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6480 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6481 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6483 if (!(mc->mc_flags & C_INITIALIZED))
6486 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6487 return MDB_NOTFOUND;
6489 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6492 rc = mdb_cursor_touch(mc);
6496 mp = mc->mc_pg[mc->mc_top];
6499 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6501 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6502 if (flags & MDB_NODUPDATA) {
6503 /* mdb_cursor_del0() will subtract the final entry */
6504 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6506 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6507 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6509 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6512 /* If sub-DB still has entries, we're done */
6513 if (mc->mc_xcursor->mx_db.md_entries) {
6514 if (leaf->mn_flags & F_SUBDATA) {
6515 /* update subDB info */
6516 void *db = NODEDATA(leaf);
6517 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6520 /* shrink fake page */
6521 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6522 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6523 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6524 /* fix other sub-DB cursors pointed at this fake page */
6525 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6526 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6527 if (m2->mc_pg[mc->mc_top] == mp &&
6528 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6529 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6532 mc->mc_db->md_entries--;
6533 mc->mc_flags |= C_DEL;
6536 /* otherwise fall thru and delete the sub-DB */
6539 if (leaf->mn_flags & F_SUBDATA) {
6540 /* add all the child DB's pages to the free list */
6541 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6547 /* add overflow pages to free list */
6548 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6552 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6553 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6554 (rc = mdb_ovpage_free(mc, omp)))
6559 return mdb_cursor_del0(mc);
6562 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6566 /** Allocate and initialize new pages for a database.
6567 * @param[in] mc a cursor on the database being added to.
6568 * @param[in] flags flags defining what type of page is being allocated.
6569 * @param[in] num the number of pages to allocate. This is usually 1,
6570 * unless allocating overflow pages for a large record.
6571 * @param[out] mp Address of a page, or NULL on failure.
6572 * @return 0 on success, non-zero on failure.
6575 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6580 if ((rc = mdb_page_alloc(mc, num, &np)))
6582 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6583 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6584 np->mp_flags = flags | P_DIRTY;
6585 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6586 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6589 mc->mc_db->md_branch_pages++;
6590 else if (IS_LEAF(np))
6591 mc->mc_db->md_leaf_pages++;
6592 else if (IS_OVERFLOW(np)) {
6593 mc->mc_db->md_overflow_pages += num;
6601 /** Calculate the size of a leaf node.
6602 * The size depends on the environment's page size; if a data item
6603 * is too large it will be put onto an overflow page and the node
6604 * size will only include the key and not the data. Sizes are always
6605 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6606 * of the #MDB_node headers.
6607 * @param[in] env The environment handle.
6608 * @param[in] key The key for the node.
6609 * @param[in] data The data for the node.
6610 * @return The number of bytes needed to store the node.
6613 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6617 sz = LEAFSIZE(key, data);
6618 if (sz > env->me_nodemax) {
6619 /* put on overflow page */
6620 sz -= data->mv_size - sizeof(pgno_t);
6623 return EVEN(sz + sizeof(indx_t));
6626 /** Calculate the size of a branch node.
6627 * The size should depend on the environment's page size but since
6628 * we currently don't support spilling large keys onto overflow
6629 * pages, it's simply the size of the #MDB_node header plus the
6630 * size of the key. Sizes are always rounded up to an even number
6631 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6632 * @param[in] env The environment handle.
6633 * @param[in] key The key for the node.
6634 * @return The number of bytes needed to store the node.
6637 mdb_branch_size(MDB_env *env, MDB_val *key)
6642 if (sz > env->me_nodemax) {
6643 /* put on overflow page */
6644 /* not implemented */
6645 /* sz -= key->size - sizeof(pgno_t); */
6648 return sz + sizeof(indx_t);
6651 /** Add a node to the page pointed to by the cursor.
6652 * @param[in] mc The cursor for this operation.
6653 * @param[in] indx The index on the page where the new node should be added.
6654 * @param[in] key The key for the new node.
6655 * @param[in] data The data for the new node, if any.
6656 * @param[in] pgno The page number, if adding a branch node.
6657 * @param[in] flags Flags for the node.
6658 * @return 0 on success, non-zero on failure. Possible errors are:
6660 * <li>ENOMEM - failed to allocate overflow pages for the node.
6661 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6662 * should never happen since all callers already calculate the
6663 * page's free space before calling this function.
6667 mdb_node_add(MDB_cursor *mc, indx_t indx,
6668 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6671 size_t node_size = NODESIZE;
6675 MDB_page *mp = mc->mc_pg[mc->mc_top];
6676 MDB_page *ofp = NULL; /* overflow page */
6679 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6681 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6682 IS_LEAF(mp) ? "leaf" : "branch",
6683 IS_SUBP(mp) ? "sub-" : "",
6684 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6685 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6688 /* Move higher keys up one slot. */
6689 int ksize = mc->mc_db->md_pad, dif;
6690 char *ptr = LEAF2KEY(mp, indx, ksize);
6691 dif = NUMKEYS(mp) - indx;
6693 memmove(ptr+ksize, ptr, dif*ksize);
6694 /* insert new key */
6695 memcpy(ptr, key->mv_data, ksize);
6697 /* Just using these for counting */
6698 mp->mp_lower += sizeof(indx_t);
6699 mp->mp_upper -= ksize - sizeof(indx_t);
6703 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6705 node_size += key->mv_size;
6707 mdb_cassert(mc, data);
6708 if (F_ISSET(flags, F_BIGDATA)) {
6709 /* Data already on overflow page. */
6710 node_size += sizeof(pgno_t);
6711 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6712 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6714 /* Put data on overflow page. */
6715 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6716 data->mv_size, node_size+data->mv_size));
6717 node_size = EVEN(node_size + sizeof(pgno_t));
6718 if ((ssize_t)node_size > room)
6720 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6722 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6726 node_size += data->mv_size;
6729 node_size = EVEN(node_size);
6730 if ((ssize_t)node_size > room)
6734 /* Move higher pointers up one slot. */
6735 for (i = NUMKEYS(mp); i > indx; i--)
6736 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6738 /* Adjust free space offsets. */
6739 ofs = mp->mp_upper - node_size;
6740 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6741 mp->mp_ptrs[indx] = ofs;
6743 mp->mp_lower += sizeof(indx_t);
6745 /* Write the node data. */
6746 node = NODEPTR(mp, indx);
6747 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6748 node->mn_flags = flags;
6750 SETDSZ(node,data->mv_size);
6755 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6758 mdb_cassert(mc, key);
6760 if (F_ISSET(flags, F_BIGDATA))
6761 memcpy(node->mn_data + key->mv_size, data->mv_data,
6763 else if (F_ISSET(flags, MDB_RESERVE))
6764 data->mv_data = node->mn_data + key->mv_size;
6766 memcpy(node->mn_data + key->mv_size, data->mv_data,
6769 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6771 if (F_ISSET(flags, MDB_RESERVE))
6772 data->mv_data = METADATA(ofp);
6774 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6781 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6782 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6783 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6784 DPRINTF(("node size = %"Z"u", node_size));
6785 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6786 return MDB_PAGE_FULL;
6789 /** Delete the specified node from a page.
6790 * @param[in] mc Cursor pointing to the node to delete.
6791 * @param[in] ksize The size of a node. Only used if the page is
6792 * part of a #MDB_DUPFIXED database.
6795 mdb_node_del(MDB_cursor *mc, int ksize)
6797 MDB_page *mp = mc->mc_pg[mc->mc_top];
6798 indx_t indx = mc->mc_ki[mc->mc_top];
6800 indx_t i, j, numkeys, ptr;
6804 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6805 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6806 numkeys = NUMKEYS(mp);
6807 mdb_cassert(mc, indx < numkeys);
6810 int x = numkeys - 1 - indx;
6811 base = LEAF2KEY(mp, indx, ksize);
6813 memmove(base, base + ksize, x * ksize);
6814 mp->mp_lower -= sizeof(indx_t);
6815 mp->mp_upper += ksize - sizeof(indx_t);
6819 node = NODEPTR(mp, indx);
6820 sz = NODESIZE + node->mn_ksize;
6822 if (F_ISSET(node->mn_flags, F_BIGDATA))
6823 sz += sizeof(pgno_t);
6825 sz += NODEDSZ(node);
6829 ptr = mp->mp_ptrs[indx];
6830 for (i = j = 0; i < numkeys; i++) {
6832 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6833 if (mp->mp_ptrs[i] < ptr)
6834 mp->mp_ptrs[j] += sz;
6839 base = (char *)mp + mp->mp_upper + PAGEBASE;
6840 memmove(base + sz, base, ptr - mp->mp_upper);
6842 mp->mp_lower -= sizeof(indx_t);
6846 /** Compact the main page after deleting a node on a subpage.
6847 * @param[in] mp The main page to operate on.
6848 * @param[in] indx The index of the subpage on the main page.
6851 mdb_node_shrink(MDB_page *mp, indx_t indx)
6857 indx_t i, numkeys, ptr;
6859 node = NODEPTR(mp, indx);
6860 sp = (MDB_page *)NODEDATA(node);
6861 delta = SIZELEFT(sp);
6862 xp = (MDB_page *)((char *)sp + delta);
6864 /* shift subpage upward */
6866 nsize = NUMKEYS(sp) * sp->mp_pad;
6868 return; /* do not make the node uneven-sized */
6869 memmove(METADATA(xp), METADATA(sp), nsize);
6872 numkeys = NUMKEYS(sp);
6873 for (i=numkeys-1; i>=0; i--)
6874 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6876 xp->mp_upper = sp->mp_lower;
6877 xp->mp_lower = sp->mp_lower;
6878 xp->mp_flags = sp->mp_flags;
6879 xp->mp_pad = sp->mp_pad;
6880 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6882 nsize = NODEDSZ(node) - delta;
6883 SETDSZ(node, nsize);
6885 /* shift lower nodes upward */
6886 ptr = mp->mp_ptrs[indx];
6887 numkeys = NUMKEYS(mp);
6888 for (i = 0; i < numkeys; i++) {
6889 if (mp->mp_ptrs[i] <= ptr)
6890 mp->mp_ptrs[i] += delta;
6893 base = (char *)mp + mp->mp_upper + PAGEBASE;
6894 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6895 mp->mp_upper += delta;
6898 /** Initial setup of a sorted-dups cursor.
6899 * Sorted duplicates are implemented as a sub-database for the given key.
6900 * The duplicate data items are actually keys of the sub-database.
6901 * Operations on the duplicate data items are performed using a sub-cursor
6902 * initialized when the sub-database is first accessed. This function does
6903 * the preliminary setup of the sub-cursor, filling in the fields that
6904 * depend only on the parent DB.
6905 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6908 mdb_xcursor_init0(MDB_cursor *mc)
6910 MDB_xcursor *mx = mc->mc_xcursor;
6912 mx->mx_cursor.mc_xcursor = NULL;
6913 mx->mx_cursor.mc_txn = mc->mc_txn;
6914 mx->mx_cursor.mc_db = &mx->mx_db;
6915 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6916 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6917 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6918 mx->mx_cursor.mc_snum = 0;
6919 mx->mx_cursor.mc_top = 0;
6920 mx->mx_cursor.mc_flags = C_SUB;
6921 mx->mx_dbx.md_name.mv_size = 0;
6922 mx->mx_dbx.md_name.mv_data = NULL;
6923 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6924 mx->mx_dbx.md_dcmp = NULL;
6925 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6928 /** Final setup of a sorted-dups cursor.
6929 * Sets up the fields that depend on the data from the main cursor.
6930 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6931 * @param[in] node The data containing the #MDB_db record for the
6932 * sorted-dup database.
6935 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6937 MDB_xcursor *mx = mc->mc_xcursor;
6939 if (node->mn_flags & F_SUBDATA) {
6940 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6941 mx->mx_cursor.mc_pg[0] = 0;
6942 mx->mx_cursor.mc_snum = 0;
6943 mx->mx_cursor.mc_top = 0;
6944 mx->mx_cursor.mc_flags = C_SUB;
6946 MDB_page *fp = NODEDATA(node);
6947 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6948 mx->mx_db.md_flags = 0;
6949 mx->mx_db.md_depth = 1;
6950 mx->mx_db.md_branch_pages = 0;
6951 mx->mx_db.md_leaf_pages = 1;
6952 mx->mx_db.md_overflow_pages = 0;
6953 mx->mx_db.md_entries = NUMKEYS(fp);
6954 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6955 mx->mx_cursor.mc_snum = 1;
6956 mx->mx_cursor.mc_top = 0;
6957 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6958 mx->mx_cursor.mc_pg[0] = fp;
6959 mx->mx_cursor.mc_ki[0] = 0;
6960 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6961 mx->mx_db.md_flags = MDB_DUPFIXED;
6962 mx->mx_db.md_pad = fp->mp_pad;
6963 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6964 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6967 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6968 mx->mx_db.md_root));
6969 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6970 #if UINT_MAX < SIZE_MAX
6971 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6972 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6976 /** Initialize a cursor for a given transaction and database. */
6978 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6981 mc->mc_backup = NULL;
6984 mc->mc_db = &txn->mt_dbs[dbi];
6985 mc->mc_dbx = &txn->mt_dbxs[dbi];
6986 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6991 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6992 mdb_tassert(txn, mx != NULL);
6993 mc->mc_xcursor = mx;
6994 mdb_xcursor_init0(mc);
6996 mc->mc_xcursor = NULL;
6998 if (*mc->mc_dbflag & DB_STALE) {
6999 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7004 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7007 size_t size = sizeof(MDB_cursor);
7009 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7012 if (txn->mt_flags & MDB_TXN_ERROR)
7015 /* Allow read access to the freelist */
7016 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7019 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7020 size += sizeof(MDB_xcursor);
7022 if ((mc = malloc(size)) != NULL) {
7023 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7024 if (txn->mt_cursors) {
7025 mc->mc_next = txn->mt_cursors[dbi];
7026 txn->mt_cursors[dbi] = mc;
7027 mc->mc_flags |= C_UNTRACK;
7039 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7041 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7044 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7047 if (txn->mt_flags & MDB_TXN_ERROR)
7050 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7054 /* Return the count of duplicate data items for the current key */
7056 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7060 if (mc == NULL || countp == NULL)
7063 if (mc->mc_xcursor == NULL)
7064 return MDB_INCOMPATIBLE;
7066 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7069 if (!(mc->mc_flags & C_INITIALIZED))
7072 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7073 return MDB_NOTFOUND;
7075 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7076 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7079 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7082 *countp = mc->mc_xcursor->mx_db.md_entries;
7088 mdb_cursor_close(MDB_cursor *mc)
7090 if (mc && !mc->mc_backup) {
7091 /* remove from txn, if tracked */
7092 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7093 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7094 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7096 *prev = mc->mc_next;
7103 mdb_cursor_txn(MDB_cursor *mc)
7105 if (!mc) return NULL;
7110 mdb_cursor_dbi(MDB_cursor *mc)
7115 /** Replace the key for a branch node with a new key.
7116 * @param[in] mc Cursor pointing to the node to operate on.
7117 * @param[in] key The new key to use.
7118 * @return 0 on success, non-zero on failure.
7121 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7127 int delta, ksize, oksize;
7128 indx_t ptr, i, numkeys, indx;
7131 indx = mc->mc_ki[mc->mc_top];
7132 mp = mc->mc_pg[mc->mc_top];
7133 node = NODEPTR(mp, indx);
7134 ptr = mp->mp_ptrs[indx];
7138 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7139 k2.mv_data = NODEKEY(node);
7140 k2.mv_size = node->mn_ksize;
7141 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7143 mdb_dkey(&k2, kbuf2),
7149 /* Sizes must be 2-byte aligned. */
7150 ksize = EVEN(key->mv_size);
7151 oksize = EVEN(node->mn_ksize);
7152 delta = ksize - oksize;
7154 /* Shift node contents if EVEN(key length) changed. */
7156 if (delta > 0 && SIZELEFT(mp) < delta) {
7158 /* not enough space left, do a delete and split */
7159 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7160 pgno = NODEPGNO(node);
7161 mdb_node_del(mc, 0);
7162 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7165 numkeys = NUMKEYS(mp);
7166 for (i = 0; i < numkeys; i++) {
7167 if (mp->mp_ptrs[i] <= ptr)
7168 mp->mp_ptrs[i] -= delta;
7171 base = (char *)mp + mp->mp_upper + PAGEBASE;
7172 len = ptr - mp->mp_upper + NODESIZE;
7173 memmove(base - delta, base, len);
7174 mp->mp_upper -= delta;
7176 node = NODEPTR(mp, indx);
7179 /* But even if no shift was needed, update ksize */
7180 if (node->mn_ksize != key->mv_size)
7181 node->mn_ksize = key->mv_size;
7184 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7190 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7192 /** Move a node from csrc to cdst.
7195 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7202 unsigned short flags;
7206 /* Mark src and dst as dirty. */
7207 if ((rc = mdb_page_touch(csrc)) ||
7208 (rc = mdb_page_touch(cdst)))
7211 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7212 key.mv_size = csrc->mc_db->md_pad;
7213 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7215 data.mv_data = NULL;
7219 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7220 mdb_cassert(csrc, !((size_t)srcnode & 1));
7221 srcpg = NODEPGNO(srcnode);
7222 flags = srcnode->mn_flags;
7223 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7224 unsigned int snum = csrc->mc_snum;
7226 /* must find the lowest key below src */
7227 rc = mdb_page_search_lowest(csrc);
7230 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7231 key.mv_size = csrc->mc_db->md_pad;
7232 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7234 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7235 key.mv_size = NODEKSZ(s2);
7236 key.mv_data = NODEKEY(s2);
7238 csrc->mc_snum = snum--;
7239 csrc->mc_top = snum;
7241 key.mv_size = NODEKSZ(srcnode);
7242 key.mv_data = NODEKEY(srcnode);
7244 data.mv_size = NODEDSZ(srcnode);
7245 data.mv_data = NODEDATA(srcnode);
7247 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7248 unsigned int snum = cdst->mc_snum;
7251 /* must find the lowest key below dst */
7252 mdb_cursor_copy(cdst, &mn);
7253 rc = mdb_page_search_lowest(&mn);
7256 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7257 bkey.mv_size = mn.mc_db->md_pad;
7258 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7260 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7261 bkey.mv_size = NODEKSZ(s2);
7262 bkey.mv_data = NODEKEY(s2);
7264 mn.mc_snum = snum--;
7267 rc = mdb_update_key(&mn, &bkey);
7272 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7273 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7274 csrc->mc_ki[csrc->mc_top],
7276 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7277 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7279 /* Add the node to the destination page.
7281 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7282 if (rc != MDB_SUCCESS)
7285 /* Delete the node from the source page.
7287 mdb_node_del(csrc, key.mv_size);
7290 /* Adjust other cursors pointing to mp */
7291 MDB_cursor *m2, *m3;
7292 MDB_dbi dbi = csrc->mc_dbi;
7293 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7295 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7296 if (csrc->mc_flags & C_SUB)
7297 m3 = &m2->mc_xcursor->mx_cursor;
7300 if (m3 == csrc) continue;
7301 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7302 csrc->mc_ki[csrc->mc_top]) {
7303 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7304 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7309 /* Update the parent separators.
7311 if (csrc->mc_ki[csrc->mc_top] == 0) {
7312 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7313 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7314 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7316 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7317 key.mv_size = NODEKSZ(srcnode);
7318 key.mv_data = NODEKEY(srcnode);
7320 DPRINTF(("update separator for source page %"Z"u to [%s]",
7321 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7322 mdb_cursor_copy(csrc, &mn);
7325 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7328 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7330 indx_t ix = csrc->mc_ki[csrc->mc_top];
7331 nullkey.mv_size = 0;
7332 csrc->mc_ki[csrc->mc_top] = 0;
7333 rc = mdb_update_key(csrc, &nullkey);
7334 csrc->mc_ki[csrc->mc_top] = ix;
7335 mdb_cassert(csrc, rc == MDB_SUCCESS);
7339 if (cdst->mc_ki[cdst->mc_top] == 0) {
7340 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7341 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7342 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7344 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7345 key.mv_size = NODEKSZ(srcnode);
7346 key.mv_data = NODEKEY(srcnode);
7348 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7349 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7350 mdb_cursor_copy(cdst, &mn);
7353 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7356 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7358 indx_t ix = cdst->mc_ki[cdst->mc_top];
7359 nullkey.mv_size = 0;
7360 cdst->mc_ki[cdst->mc_top] = 0;
7361 rc = mdb_update_key(cdst, &nullkey);
7362 cdst->mc_ki[cdst->mc_top] = ix;
7363 mdb_cassert(csrc, rc == MDB_SUCCESS);
7370 /** Merge one page into another.
7371 * The nodes from the page pointed to by \b csrc will
7372 * be copied to the page pointed to by \b cdst and then
7373 * the \b csrc page will be freed.
7374 * @param[in] csrc Cursor pointing to the source page.
7375 * @param[in] cdst Cursor pointing to the destination page.
7376 * @return 0 on success, non-zero on failure.
7379 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7381 MDB_page *psrc, *pdst;
7388 psrc = csrc->mc_pg[csrc->mc_top];
7389 pdst = cdst->mc_pg[cdst->mc_top];
7391 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7393 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7394 mdb_cassert(csrc, cdst->mc_snum > 1);
7396 /* Mark dst as dirty. */
7397 if ((rc = mdb_page_touch(cdst)))
7400 /* Move all nodes from src to dst.
7402 j = nkeys = NUMKEYS(pdst);
7403 if (IS_LEAF2(psrc)) {
7404 key.mv_size = csrc->mc_db->md_pad;
7405 key.mv_data = METADATA(psrc);
7406 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7407 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7408 if (rc != MDB_SUCCESS)
7410 key.mv_data = (char *)key.mv_data + key.mv_size;
7413 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7414 srcnode = NODEPTR(psrc, i);
7415 if (i == 0 && IS_BRANCH(psrc)) {
7418 mdb_cursor_copy(csrc, &mn);
7419 /* must find the lowest key below src */
7420 rc = mdb_page_search_lowest(&mn);
7423 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7424 key.mv_size = mn.mc_db->md_pad;
7425 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7427 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7428 key.mv_size = NODEKSZ(s2);
7429 key.mv_data = NODEKEY(s2);
7432 key.mv_size = srcnode->mn_ksize;
7433 key.mv_data = NODEKEY(srcnode);
7436 data.mv_size = NODEDSZ(srcnode);
7437 data.mv_data = NODEDATA(srcnode);
7438 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7439 if (rc != MDB_SUCCESS)
7444 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7445 pdst->mp_pgno, NUMKEYS(pdst),
7446 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7448 /* Unlink the src page from parent and add to free list.
7451 mdb_node_del(csrc, 0);
7452 if (csrc->mc_ki[csrc->mc_top] == 0) {
7454 rc = mdb_update_key(csrc, &key);
7462 psrc = csrc->mc_pg[csrc->mc_top];
7463 /* If not operating on FreeDB, allow this page to be reused
7464 * in this txn. Otherwise just add to free list.
7466 rc = mdb_page_loose(csrc, psrc);
7470 csrc->mc_db->md_leaf_pages--;
7472 csrc->mc_db->md_branch_pages--;
7474 /* Adjust other cursors pointing to mp */
7475 MDB_cursor *m2, *m3;
7476 MDB_dbi dbi = csrc->mc_dbi;
7478 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7479 if (csrc->mc_flags & C_SUB)
7480 m3 = &m2->mc_xcursor->mx_cursor;
7483 if (m3 == csrc) continue;
7484 if (m3->mc_snum < csrc->mc_snum) continue;
7485 if (m3->mc_pg[csrc->mc_top] == psrc) {
7486 m3->mc_pg[csrc->mc_top] = pdst;
7487 m3->mc_ki[csrc->mc_top] += nkeys;
7492 unsigned int snum = cdst->mc_snum;
7493 uint16_t depth = cdst->mc_db->md_depth;
7494 mdb_cursor_pop(cdst);
7495 rc = mdb_rebalance(cdst);
7496 /* Did the tree shrink? */
7497 if (depth > cdst->mc_db->md_depth)
7499 cdst->mc_snum = snum;
7500 cdst->mc_top = snum-1;
7505 /** Copy the contents of a cursor.
7506 * @param[in] csrc The cursor to copy from.
7507 * @param[out] cdst The cursor to copy to.
7510 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7514 cdst->mc_txn = csrc->mc_txn;
7515 cdst->mc_dbi = csrc->mc_dbi;
7516 cdst->mc_db = csrc->mc_db;
7517 cdst->mc_dbx = csrc->mc_dbx;
7518 cdst->mc_snum = csrc->mc_snum;
7519 cdst->mc_top = csrc->mc_top;
7520 cdst->mc_flags = csrc->mc_flags;
7522 for (i=0; i<csrc->mc_snum; i++) {
7523 cdst->mc_pg[i] = csrc->mc_pg[i];
7524 cdst->mc_ki[i] = csrc->mc_ki[i];
7528 /** Rebalance the tree after a delete operation.
7529 * @param[in] mc Cursor pointing to the page where rebalancing
7531 * @return 0 on success, non-zero on failure.
7534 mdb_rebalance(MDB_cursor *mc)
7538 unsigned int ptop, minkeys;
7542 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7543 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7544 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7545 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7546 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7548 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7549 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7550 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7551 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7555 if (mc->mc_snum < 2) {
7556 MDB_page *mp = mc->mc_pg[0];
7558 DPUTS("Can't rebalance a subpage, ignoring");
7561 if (NUMKEYS(mp) == 0) {
7562 DPUTS("tree is completely empty");
7563 mc->mc_db->md_root = P_INVALID;
7564 mc->mc_db->md_depth = 0;
7565 mc->mc_db->md_leaf_pages = 0;
7566 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7569 /* Adjust cursors pointing to mp */
7572 mc->mc_flags &= ~C_INITIALIZED;
7574 MDB_cursor *m2, *m3;
7575 MDB_dbi dbi = mc->mc_dbi;
7577 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7578 if (mc->mc_flags & C_SUB)
7579 m3 = &m2->mc_xcursor->mx_cursor;
7582 if (m3->mc_snum < mc->mc_snum) continue;
7583 if (m3->mc_pg[0] == mp) {
7586 m3->mc_flags &= ~C_INITIALIZED;
7590 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7592 DPUTS("collapsing root page!");
7593 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7596 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7597 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7600 mc->mc_db->md_depth--;
7601 mc->mc_db->md_branch_pages--;
7602 mc->mc_ki[0] = mc->mc_ki[1];
7603 for (i = 1; i<mc->mc_db->md_depth; i++) {
7604 mc->mc_pg[i] = mc->mc_pg[i+1];
7605 mc->mc_ki[i] = mc->mc_ki[i+1];
7608 /* Adjust other cursors pointing to mp */
7609 MDB_cursor *m2, *m3;
7610 MDB_dbi dbi = mc->mc_dbi;
7612 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7613 if (mc->mc_flags & C_SUB)
7614 m3 = &m2->mc_xcursor->mx_cursor;
7617 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7618 if (m3->mc_pg[0] == mp) {
7621 for (i=0; i<m3->mc_snum; i++) {
7622 m3->mc_pg[i] = m3->mc_pg[i+1];
7623 m3->mc_ki[i] = m3->mc_ki[i+1];
7629 DPUTS("root page doesn't need rebalancing");
7633 /* The parent (branch page) must have at least 2 pointers,
7634 * otherwise the tree is invalid.
7636 ptop = mc->mc_top-1;
7637 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7639 /* Leaf page fill factor is below the threshold.
7640 * Try to move keys from left or right neighbor, or
7641 * merge with a neighbor page.
7646 mdb_cursor_copy(mc, &mn);
7647 mn.mc_xcursor = NULL;
7649 oldki = mc->mc_ki[mc->mc_top];
7650 if (mc->mc_ki[ptop] == 0) {
7651 /* We're the leftmost leaf in our parent.
7653 DPUTS("reading right neighbor");
7655 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7656 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7659 mn.mc_ki[mn.mc_top] = 0;
7660 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7662 /* There is at least one neighbor to the left.
7664 DPUTS("reading left neighbor");
7666 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7667 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7670 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7671 mc->mc_ki[mc->mc_top] = 0;
7674 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7675 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7676 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7678 /* If the neighbor page is above threshold and has enough keys,
7679 * move one key from it. Otherwise we should try to merge them.
7680 * (A branch page must never have less than 2 keys.)
7682 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7683 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7684 rc = mdb_node_move(&mn, mc);
7685 if (mc->mc_ki[ptop]) {
7689 if (mc->mc_ki[ptop] == 0) {
7690 rc = mdb_page_merge(&mn, mc);
7692 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7693 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7694 rc = mdb_page_merge(mc, &mn);
7695 mdb_cursor_copy(&mn, mc);
7697 mc->mc_flags &= ~C_EOF;
7699 mc->mc_ki[mc->mc_top] = oldki;
7703 /** Complete a delete operation started by #mdb_cursor_del(). */
7705 mdb_cursor_del0(MDB_cursor *mc)
7712 ki = mc->mc_ki[mc->mc_top];
7713 mdb_node_del(mc, mc->mc_db->md_pad);
7714 mc->mc_db->md_entries--;
7715 rc = mdb_rebalance(mc);
7717 if (rc == MDB_SUCCESS) {
7718 MDB_cursor *m2, *m3;
7719 MDB_dbi dbi = mc->mc_dbi;
7721 mp = mc->mc_pg[mc->mc_top];
7722 nkeys = NUMKEYS(mp);
7724 /* if mc points past last node in page, find next sibling */
7725 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7726 rc = mdb_cursor_sibling(mc, 1);
7727 if (rc == MDB_NOTFOUND) {
7728 mc->mc_flags |= C_EOF;
7733 /* Adjust other cursors pointing to mp */
7734 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7735 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7736 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7738 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7740 if (m3->mc_pg[mc->mc_top] == mp) {
7741 if (m3->mc_ki[mc->mc_top] >= ki) {
7742 m3->mc_flags |= C_DEL;
7743 if (m3->mc_ki[mc->mc_top] > ki)
7744 m3->mc_ki[mc->mc_top]--;
7745 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7746 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7748 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7749 rc = mdb_cursor_sibling(m3, 1);
7750 if (rc == MDB_NOTFOUND) {
7751 m3->mc_flags |= C_EOF;
7757 mc->mc_flags |= C_DEL;
7761 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7766 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7767 MDB_val *key, MDB_val *data)
7769 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7772 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7773 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7775 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7776 /* must ignore any data */
7780 return mdb_del0(txn, dbi, key, data, 0);
7784 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7785 MDB_val *key, MDB_val *data, unsigned flags)
7790 MDB_val rdata, *xdata;
7794 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7796 mdb_cursor_init(&mc, txn, dbi, &mx);
7805 flags |= MDB_NODUPDATA;
7807 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7809 /* let mdb_page_split know about this cursor if needed:
7810 * delete will trigger a rebalance; if it needs to move
7811 * a node from one page to another, it will have to
7812 * update the parent's separator key(s). If the new sepkey
7813 * is larger than the current one, the parent page may
7814 * run out of space, triggering a split. We need this
7815 * cursor to be consistent until the end of the rebalance.
7817 mc.mc_flags |= C_UNTRACK;
7818 mc.mc_next = txn->mt_cursors[dbi];
7819 txn->mt_cursors[dbi] = &mc;
7820 rc = mdb_cursor_del(&mc, flags);
7821 txn->mt_cursors[dbi] = mc.mc_next;
7826 /** Split a page and insert a new node.
7827 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7828 * The cursor will be updated to point to the actual page and index where
7829 * the node got inserted after the split.
7830 * @param[in] newkey The key for the newly inserted node.
7831 * @param[in] newdata The data for the newly inserted node.
7832 * @param[in] newpgno The page number, if the new node is a branch node.
7833 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7834 * @return 0 on success, non-zero on failure.
7837 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7838 unsigned int nflags)
7841 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7844 int i, j, split_indx, nkeys, pmax;
7845 MDB_env *env = mc->mc_txn->mt_env;
7847 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7848 MDB_page *copy = NULL;
7849 MDB_page *mp, *rp, *pp;
7854 mp = mc->mc_pg[mc->mc_top];
7855 newindx = mc->mc_ki[mc->mc_top];
7856 nkeys = NUMKEYS(mp);
7858 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7859 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7860 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7862 /* Create a right sibling. */
7863 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7865 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7867 if (mc->mc_snum < 2) {
7868 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7870 /* shift current top to make room for new parent */
7871 mc->mc_pg[1] = mc->mc_pg[0];
7872 mc->mc_ki[1] = mc->mc_ki[0];
7875 mc->mc_db->md_root = pp->mp_pgno;
7876 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7877 mc->mc_db->md_depth++;
7880 /* Add left (implicit) pointer. */
7881 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7882 /* undo the pre-push */
7883 mc->mc_pg[0] = mc->mc_pg[1];
7884 mc->mc_ki[0] = mc->mc_ki[1];
7885 mc->mc_db->md_root = mp->mp_pgno;
7886 mc->mc_db->md_depth--;
7893 ptop = mc->mc_top-1;
7894 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7897 mc->mc_flags |= C_SPLITTING;
7898 mdb_cursor_copy(mc, &mn);
7899 mn.mc_pg[mn.mc_top] = rp;
7900 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7902 if (nflags & MDB_APPEND) {
7903 mn.mc_ki[mn.mc_top] = 0;
7905 split_indx = newindx;
7909 split_indx = (nkeys+1) / 2;
7914 unsigned int lsize, rsize, ksize;
7915 /* Move half of the keys to the right sibling */
7916 x = mc->mc_ki[mc->mc_top] - split_indx;
7917 ksize = mc->mc_db->md_pad;
7918 split = LEAF2KEY(mp, split_indx, ksize);
7919 rsize = (nkeys - split_indx) * ksize;
7920 lsize = (nkeys - split_indx) * sizeof(indx_t);
7921 mp->mp_lower -= lsize;
7922 rp->mp_lower += lsize;
7923 mp->mp_upper += rsize - lsize;
7924 rp->mp_upper -= rsize - lsize;
7925 sepkey.mv_size = ksize;
7926 if (newindx == split_indx) {
7927 sepkey.mv_data = newkey->mv_data;
7929 sepkey.mv_data = split;
7932 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7933 memcpy(rp->mp_ptrs, split, rsize);
7934 sepkey.mv_data = rp->mp_ptrs;
7935 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7936 memcpy(ins, newkey->mv_data, ksize);
7937 mp->mp_lower += sizeof(indx_t);
7938 mp->mp_upper -= ksize - sizeof(indx_t);
7941 memcpy(rp->mp_ptrs, split, x * ksize);
7942 ins = LEAF2KEY(rp, x, ksize);
7943 memcpy(ins, newkey->mv_data, ksize);
7944 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7945 rp->mp_lower += sizeof(indx_t);
7946 rp->mp_upper -= ksize - sizeof(indx_t);
7947 mc->mc_ki[mc->mc_top] = x;
7948 mc->mc_pg[mc->mc_top] = rp;
7951 int psize, nsize, k;
7952 /* Maximum free space in an empty page */
7953 pmax = env->me_psize - PAGEHDRSZ;
7955 nsize = mdb_leaf_size(env, newkey, newdata);
7957 nsize = mdb_branch_size(env, newkey);
7958 nsize = EVEN(nsize);
7960 /* grab a page to hold a temporary copy */
7961 copy = mdb_page_malloc(mc->mc_txn, 1);
7966 copy->mp_pgno = mp->mp_pgno;
7967 copy->mp_flags = mp->mp_flags;
7968 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7969 copy->mp_upper = env->me_psize - PAGEBASE;
7971 /* prepare to insert */
7972 for (i=0, j=0; i<nkeys; i++) {
7974 copy->mp_ptrs[j++] = 0;
7976 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7979 /* When items are relatively large the split point needs
7980 * to be checked, because being off-by-one will make the
7981 * difference between success or failure in mdb_node_add.
7983 * It's also relevant if a page happens to be laid out
7984 * such that one half of its nodes are all "small" and
7985 * the other half of its nodes are "large." If the new
7986 * item is also "large" and falls on the half with
7987 * "large" nodes, it also may not fit.
7989 * As a final tweak, if the new item goes on the last
7990 * spot on the page (and thus, onto the new page), bias
7991 * the split so the new page is emptier than the old page.
7992 * This yields better packing during sequential inserts.
7994 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7995 /* Find split point */
7997 if (newindx <= split_indx || newindx >= nkeys) {
7999 k = newindx >= nkeys ? nkeys : split_indx+2;
8004 for (; i!=k; i+=j) {
8009 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8010 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8012 if (F_ISSET(node->mn_flags, F_BIGDATA))
8013 psize += sizeof(pgno_t);
8015 psize += NODEDSZ(node);
8017 psize = EVEN(psize);
8019 if (psize > pmax || i == k-j) {
8020 split_indx = i + (j<0);
8025 if (split_indx == newindx) {
8026 sepkey.mv_size = newkey->mv_size;
8027 sepkey.mv_data = newkey->mv_data;
8029 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8030 sepkey.mv_size = node->mn_ksize;
8031 sepkey.mv_data = NODEKEY(node);
8036 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8038 /* Copy separator key to the parent.
8040 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8044 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8049 if (mn.mc_snum == mc->mc_snum) {
8050 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8051 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8052 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8053 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8058 /* Right page might now have changed parent.
8059 * Check if left page also changed parent.
8061 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8062 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8063 for (i=0; i<ptop; i++) {
8064 mc->mc_pg[i] = mn.mc_pg[i];
8065 mc->mc_ki[i] = mn.mc_ki[i];
8067 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8068 if (mn.mc_ki[ptop]) {
8069 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8071 /* find right page's left sibling */
8072 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8073 mdb_cursor_sibling(mc, 0);
8078 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8081 mc->mc_flags ^= C_SPLITTING;
8082 if (rc != MDB_SUCCESS) {
8085 if (nflags & MDB_APPEND) {
8086 mc->mc_pg[mc->mc_top] = rp;
8087 mc->mc_ki[mc->mc_top] = 0;
8088 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8091 for (i=0; i<mc->mc_top; i++)
8092 mc->mc_ki[i] = mn.mc_ki[i];
8093 } else if (!IS_LEAF2(mp)) {
8095 mc->mc_pg[mc->mc_top] = rp;
8100 rkey.mv_data = newkey->mv_data;
8101 rkey.mv_size = newkey->mv_size;
8107 /* Update index for the new key. */
8108 mc->mc_ki[mc->mc_top] = j;
8110 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8111 rkey.mv_data = NODEKEY(node);
8112 rkey.mv_size = node->mn_ksize;
8114 xdata.mv_data = NODEDATA(node);
8115 xdata.mv_size = NODEDSZ(node);
8118 pgno = NODEPGNO(node);
8119 flags = node->mn_flags;
8122 if (!IS_LEAF(mp) && j == 0) {
8123 /* First branch index doesn't need key data. */
8127 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8133 mc->mc_pg[mc->mc_top] = copy;
8138 } while (i != split_indx);
8140 nkeys = NUMKEYS(copy);
8141 for (i=0; i<nkeys; i++)
8142 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8143 mp->mp_lower = copy->mp_lower;
8144 mp->mp_upper = copy->mp_upper;
8145 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8146 env->me_psize - copy->mp_upper - PAGEBASE);
8148 /* reset back to original page */
8149 if (newindx < split_indx) {
8150 mc->mc_pg[mc->mc_top] = mp;
8151 if (nflags & MDB_RESERVE) {
8152 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8153 if (!(node->mn_flags & F_BIGDATA))
8154 newdata->mv_data = NODEDATA(node);
8157 mc->mc_pg[mc->mc_top] = rp;
8159 /* Make sure mc_ki is still valid.
8161 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8162 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8163 for (i=0; i<=ptop; i++) {
8164 mc->mc_pg[i] = mn.mc_pg[i];
8165 mc->mc_ki[i] = mn.mc_ki[i];
8172 /* Adjust other cursors pointing to mp */
8173 MDB_cursor *m2, *m3;
8174 MDB_dbi dbi = mc->mc_dbi;
8175 int fixup = NUMKEYS(mp);
8177 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8178 if (mc->mc_flags & C_SUB)
8179 m3 = &m2->mc_xcursor->mx_cursor;
8184 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8186 if (m3->mc_flags & C_SPLITTING)
8191 for (k=m3->mc_top; k>=0; k--) {
8192 m3->mc_ki[k+1] = m3->mc_ki[k];
8193 m3->mc_pg[k+1] = m3->mc_pg[k];
8195 if (m3->mc_ki[0] >= split_indx) {
8200 m3->mc_pg[0] = mc->mc_pg[0];
8204 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8205 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8206 m3->mc_ki[mc->mc_top]++;
8207 if (m3->mc_ki[mc->mc_top] >= fixup) {
8208 m3->mc_pg[mc->mc_top] = rp;
8209 m3->mc_ki[mc->mc_top] -= fixup;
8210 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8212 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8213 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8218 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8221 if (copy) /* tmp page */
8222 mdb_page_free(env, copy);
8224 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8229 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8230 MDB_val *key, MDB_val *data, unsigned int flags)
8235 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8238 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8241 mdb_cursor_init(&mc, txn, dbi, &mx);
8242 return mdb_cursor_put(&mc, key, data, flags);
8246 #define MDB_WBUF (1024*1024)
8249 /** State needed for a compacting copy. */
8250 typedef struct mdb_copy {
8251 pthread_mutex_t mc_mutex;
8252 pthread_cond_t mc_cond;
8259 pgno_t mc_next_pgno;
8262 volatile int mc_new;
8267 /** Dedicated writer thread for compacting copy. */
8268 static THREAD_RET ESECT
8269 mdb_env_copythr(void *arg)
8273 int toggle = 0, wsize, rc;
8276 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8279 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8282 pthread_mutex_lock(&my->mc_mutex);
8284 pthread_cond_signal(&my->mc_cond);
8287 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8288 if (my->mc_new < 0) {
8293 wsize = my->mc_wlen[toggle];
8294 ptr = my->mc_wbuf[toggle];
8297 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8301 } else if (len > 0) {
8315 /* If there's an overflow page tail, write it too */
8316 if (my->mc_olen[toggle]) {
8317 wsize = my->mc_olen[toggle];
8318 ptr = my->mc_over[toggle];
8319 my->mc_olen[toggle] = 0;
8322 my->mc_wlen[toggle] = 0;
8324 pthread_cond_signal(&my->mc_cond);
8326 pthread_cond_signal(&my->mc_cond);
8327 pthread_mutex_unlock(&my->mc_mutex);
8328 return (THREAD_RET)0;
8332 /** Tell the writer thread there's a buffer ready to write */
8334 mdb_env_cthr_toggle(mdb_copy *my, int st)
8336 int toggle = my->mc_toggle ^ 1;
8337 pthread_mutex_lock(&my->mc_mutex);
8338 if (my->mc_status) {
8339 pthread_mutex_unlock(&my->mc_mutex);
8340 return my->mc_status;
8342 while (my->mc_new == 1)
8343 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8345 my->mc_toggle = toggle;
8346 pthread_cond_signal(&my->mc_cond);
8347 pthread_mutex_unlock(&my->mc_mutex);
8351 /** Depth-first tree traversal for compacting copy. */
8353 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8356 MDB_txn *txn = my->mc_txn;
8358 MDB_page *mo, *mp, *leaf;
8363 /* Empty DB, nothing to do */
8364 if (*pg == P_INVALID)
8371 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8374 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8378 /* Make cursor pages writable */
8379 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8383 for (i=0; i<mc.mc_top; i++) {
8384 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8385 mc.mc_pg[i] = (MDB_page *)ptr;
8386 ptr += my->mc_env->me_psize;
8389 /* This is writable space for a leaf page. Usually not needed. */
8390 leaf = (MDB_page *)ptr;
8392 toggle = my->mc_toggle;
8393 while (mc.mc_snum > 0) {
8395 mp = mc.mc_pg[mc.mc_top];
8399 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8400 for (i=0; i<n; i++) {
8401 ni = NODEPTR(mp, i);
8402 if (ni->mn_flags & F_BIGDATA) {
8406 /* Need writable leaf */
8408 mc.mc_pg[mc.mc_top] = leaf;
8409 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8411 ni = NODEPTR(mp, i);
8414 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8415 rc = mdb_page_get(txn, pg, &omp, NULL);
8418 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8419 rc = mdb_env_cthr_toggle(my, 1);
8422 toggle = my->mc_toggle;
8424 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8425 memcpy(mo, omp, my->mc_env->me_psize);
8426 mo->mp_pgno = my->mc_next_pgno;
8427 my->mc_next_pgno += omp->mp_pages;
8428 my->mc_wlen[toggle] += my->mc_env->me_psize;
8429 if (omp->mp_pages > 1) {
8430 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8431 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8432 rc = mdb_env_cthr_toggle(my, 1);
8435 toggle = my->mc_toggle;
8437 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8438 } else if (ni->mn_flags & F_SUBDATA) {
8441 /* Need writable leaf */
8443 mc.mc_pg[mc.mc_top] = leaf;
8444 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8446 ni = NODEPTR(mp, i);
8449 memcpy(&db, NODEDATA(ni), sizeof(db));
8450 my->mc_toggle = toggle;
8451 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8454 toggle = my->mc_toggle;
8455 memcpy(NODEDATA(ni), &db, sizeof(db));
8460 mc.mc_ki[mc.mc_top]++;
8461 if (mc.mc_ki[mc.mc_top] < n) {
8464 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8466 rc = mdb_page_get(txn, pg, &mp, NULL);
8471 mc.mc_ki[mc.mc_top] = 0;
8472 if (IS_BRANCH(mp)) {
8473 /* Whenever we advance to a sibling branch page,
8474 * we must proceed all the way down to its first leaf.
8476 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8479 mc.mc_pg[mc.mc_top] = mp;
8483 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8484 rc = mdb_env_cthr_toggle(my, 1);
8487 toggle = my->mc_toggle;
8489 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8490 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8491 mo->mp_pgno = my->mc_next_pgno++;
8492 my->mc_wlen[toggle] += my->mc_env->me_psize;
8494 /* Update parent if there is one */
8495 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8496 SETPGNO(ni, mo->mp_pgno);
8497 mdb_cursor_pop(&mc);
8499 /* Otherwise we're done */
8509 /** Copy environment with compaction. */
8511 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8516 MDB_txn *txn = NULL;
8521 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8522 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8523 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8524 if (my.mc_wbuf[0] == NULL)
8527 pthread_mutex_init(&my.mc_mutex, NULL);
8528 pthread_cond_init(&my.mc_cond, NULL);
8529 #ifdef HAVE_MEMALIGN
8530 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8531 if (my.mc_wbuf[0] == NULL)
8534 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8539 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8540 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8545 my.mc_next_pgno = 2;
8551 THREAD_CREATE(thr, mdb_env_copythr, &my);
8553 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8557 mp = (MDB_page *)my.mc_wbuf[0];
8558 memset(mp, 0, 2*env->me_psize);
8560 mp->mp_flags = P_META;
8561 mm = (MDB_meta *)METADATA(mp);
8562 mdb_env_init_meta0(env, mm);
8563 mm->mm_address = env->me_metas[0]->mm_address;
8565 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8567 mp->mp_flags = P_META;
8568 *(MDB_meta *)METADATA(mp) = *mm;
8569 mm = (MDB_meta *)METADATA(mp);
8571 /* Count the number of free pages, subtract from lastpg to find
8572 * number of active pages
8575 MDB_ID freecount = 0;
8578 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8579 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8580 freecount += *(MDB_ID *)data.mv_data;
8581 freecount += txn->mt_dbs[0].md_branch_pages +
8582 txn->mt_dbs[0].md_leaf_pages +
8583 txn->mt_dbs[0].md_overflow_pages;
8585 /* Set metapage 1 */
8586 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8587 mm->mm_dbs[1] = txn->mt_dbs[1];
8588 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8591 my.mc_wlen[0] = env->me_psize * 2;
8593 pthread_mutex_lock(&my.mc_mutex);
8595 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8596 pthread_mutex_unlock(&my.mc_mutex);
8597 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8598 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8599 rc = mdb_env_cthr_toggle(&my, 1);
8600 mdb_env_cthr_toggle(&my, -1);
8601 pthread_mutex_lock(&my.mc_mutex);
8603 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8604 pthread_mutex_unlock(&my.mc_mutex);
8609 CloseHandle(my.mc_cond);
8610 CloseHandle(my.mc_mutex);
8611 _aligned_free(my.mc_wbuf[0]);
8613 pthread_cond_destroy(&my.mc_cond);
8614 pthread_mutex_destroy(&my.mc_mutex);
8615 free(my.mc_wbuf[0]);
8620 /** Copy environment as-is. */
8622 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8624 MDB_txn *txn = NULL;
8630 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8634 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8637 /* Do the lock/unlock of the reader mutex before starting the
8638 * write txn. Otherwise other read txns could block writers.
8640 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8645 /* We must start the actual read txn after blocking writers */
8646 mdb_txn_reset0(txn, "reset-stage1");
8648 /* Temporarily block writers until we snapshot the meta pages */
8651 rc = mdb_txn_renew0(txn);
8653 UNLOCK_MUTEX_W(env);
8658 wsize = env->me_psize * 2;
8662 DO_WRITE(rc, fd, ptr, w2, len);
8666 } else if (len > 0) {
8672 /* Non-blocking or async handles are not supported */
8678 UNLOCK_MUTEX_W(env);
8683 w2 = txn->mt_next_pgno * env->me_psize;
8686 LARGE_INTEGER fsize;
8687 GetFileSizeEx(env->me_fd, &fsize);
8688 if (w2 > fsize.QuadPart)
8689 w2 = fsize.QuadPart;
8694 fstat(env->me_fd, &st);
8695 if (w2 > (size_t)st.st_size)
8701 if (wsize > MAX_WRITE)
8705 DO_WRITE(rc, fd, ptr, w2, len);
8709 } else if (len > 0) {
8726 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8728 if (flags & MDB_CP_COMPACT)
8729 return mdb_env_copyfd1(env, fd);
8731 return mdb_env_copyfd0(env, fd);
8735 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8737 return mdb_env_copyfd2(env, fd, 0);
8741 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8745 HANDLE newfd = INVALID_HANDLE_VALUE;
8747 if (env->me_flags & MDB_NOSUBDIR) {
8748 lpath = (char *)path;
8751 len += sizeof(DATANAME);
8752 lpath = malloc(len);
8755 sprintf(lpath, "%s" DATANAME, path);
8758 /* The destination path must exist, but the destination file must not.
8759 * We don't want the OS to cache the writes, since the source data is
8760 * already in the OS cache.
8763 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8764 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8766 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8768 if (newfd == INVALID_HANDLE_VALUE) {
8773 if (env->me_psize >= env->me_os_psize) {
8775 /* Set O_DIRECT if the file system supports it */
8776 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8777 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8779 #ifdef F_NOCACHE /* __APPLE__ */
8780 rc = fcntl(newfd, F_NOCACHE, 1);
8788 rc = mdb_env_copyfd2(env, newfd, flags);
8791 if (!(env->me_flags & MDB_NOSUBDIR))
8793 if (newfd != INVALID_HANDLE_VALUE)
8794 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8801 mdb_env_copy(MDB_env *env, const char *path)
8803 return mdb_env_copy2(env, path, 0);
8807 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8809 if ((flag & CHANGEABLE) != flag)
8812 env->me_flags |= flag;
8814 env->me_flags &= ~flag;
8819 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8824 *arg = env->me_flags;
8829 mdb_env_set_userctx(MDB_env *env, void *ctx)
8833 env->me_userctx = ctx;
8838 mdb_env_get_userctx(MDB_env *env)
8840 return env ? env->me_userctx : NULL;
8844 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8849 env->me_assert_func = func;
8855 mdb_env_get_path(MDB_env *env, const char **arg)
8860 *arg = env->me_path;
8865 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8874 /** Common code for #mdb_stat() and #mdb_env_stat().
8875 * @param[in] env the environment to operate in.
8876 * @param[in] db the #MDB_db record containing the stats to return.
8877 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8878 * @return 0, this function always succeeds.
8881 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8883 arg->ms_psize = env->me_psize;
8884 arg->ms_depth = db->md_depth;
8885 arg->ms_branch_pages = db->md_branch_pages;
8886 arg->ms_leaf_pages = db->md_leaf_pages;
8887 arg->ms_overflow_pages = db->md_overflow_pages;
8888 arg->ms_entries = db->md_entries;
8894 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8898 if (env == NULL || arg == NULL)
8901 toggle = mdb_env_pick_meta(env);
8903 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8907 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8911 if (env == NULL || arg == NULL)
8914 toggle = mdb_env_pick_meta(env);
8915 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8916 arg->me_mapsize = env->me_mapsize;
8917 arg->me_maxreaders = env->me_maxreaders;
8919 /* me_numreaders may be zero if this process never used any readers. Use
8920 * the shared numreader count if it exists.
8922 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8924 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8925 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8929 /** Set the default comparison functions for a database.
8930 * Called immediately after a database is opened to set the defaults.
8931 * The user can then override them with #mdb_set_compare() or
8932 * #mdb_set_dupsort().
8933 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8934 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8937 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8939 uint16_t f = txn->mt_dbs[dbi].md_flags;
8941 txn->mt_dbxs[dbi].md_cmp =
8942 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8943 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8945 txn->mt_dbxs[dbi].md_dcmp =
8946 !(f & MDB_DUPSORT) ? 0 :
8947 ((f & MDB_INTEGERDUP)
8948 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8949 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8952 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8958 int rc, dbflag, exact;
8959 unsigned int unused = 0, seq;
8962 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8963 mdb_default_cmp(txn, FREE_DBI);
8966 if ((flags & VALID_FLAGS) != flags)
8968 if (txn->mt_flags & MDB_TXN_ERROR)
8974 if (flags & PERSISTENT_FLAGS) {
8975 uint16_t f2 = flags & PERSISTENT_FLAGS;
8976 /* make sure flag changes get committed */
8977 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8978 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8979 txn->mt_flags |= MDB_TXN_DIRTY;
8982 mdb_default_cmp(txn, MAIN_DBI);
8986 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8987 mdb_default_cmp(txn, MAIN_DBI);
8990 /* Is the DB already open? */
8992 for (i=2; i<txn->mt_numdbs; i++) {
8993 if (!txn->mt_dbxs[i].md_name.mv_size) {
8994 /* Remember this free slot */
8995 if (!unused) unused = i;
8998 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8999 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9005 /* If no free slot and max hit, fail */
9006 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9007 return MDB_DBS_FULL;
9009 /* Cannot mix named databases with some mainDB flags */
9010 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9011 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9013 /* Find the DB info */
9014 dbflag = DB_NEW|DB_VALID;
9017 key.mv_data = (void *)name;
9018 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9019 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9020 if (rc == MDB_SUCCESS) {
9021 /* make sure this is actually a DB */
9022 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9023 if (!(node->mn_flags & F_SUBDATA))
9024 return MDB_INCOMPATIBLE;
9025 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9026 /* Create if requested */
9027 data.mv_size = sizeof(MDB_db);
9028 data.mv_data = &dummy;
9029 memset(&dummy, 0, sizeof(dummy));
9030 dummy.md_root = P_INVALID;
9031 dummy.md_flags = flags & PERSISTENT_FLAGS;
9032 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9036 /* OK, got info, add to table */
9037 if (rc == MDB_SUCCESS) {
9038 unsigned int slot = unused ? unused : txn->mt_numdbs;
9039 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9040 txn->mt_dbxs[slot].md_name.mv_size = len;
9041 txn->mt_dbxs[slot].md_rel = NULL;
9042 txn->mt_dbflags[slot] = dbflag;
9043 /* txn-> and env-> are the same in read txns, use
9044 * tmp variable to avoid undefined assignment
9046 seq = ++txn->mt_env->me_dbiseqs[slot];
9047 txn->mt_dbiseqs[slot] = seq;
9049 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9051 mdb_default_cmp(txn, slot);
9060 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9062 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9065 if (txn->mt_flags & MDB_TXN_ERROR)
9068 if (txn->mt_dbflags[dbi] & DB_STALE) {
9071 /* Stale, must read the DB's root. cursor_init does it for us. */
9072 mdb_cursor_init(&mc, txn, dbi, &mx);
9074 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9077 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9080 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9082 ptr = env->me_dbxs[dbi].md_name.mv_data;
9083 /* If there was no name, this was already closed */
9085 env->me_dbxs[dbi].md_name.mv_data = NULL;
9086 env->me_dbxs[dbi].md_name.mv_size = 0;
9087 env->me_dbflags[dbi] = 0;
9088 env->me_dbiseqs[dbi]++;
9093 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9095 /* We could return the flags for the FREE_DBI too but what's the point? */
9096 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9098 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9102 /** Add all the DB's pages to the free list.
9103 * @param[in] mc Cursor on the DB to free.
9104 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9105 * @return 0 on success, non-zero on failure.
9108 mdb_drop0(MDB_cursor *mc, int subs)
9112 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9113 if (rc == MDB_SUCCESS) {
9114 MDB_txn *txn = mc->mc_txn;
9119 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9120 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9123 mdb_cursor_copy(mc, &mx);
9124 while (mc->mc_snum > 0) {
9125 MDB_page *mp = mc->mc_pg[mc->mc_top];
9126 unsigned n = NUMKEYS(mp);
9128 for (i=0; i<n; i++) {
9129 ni = NODEPTR(mp, i);
9130 if (ni->mn_flags & F_BIGDATA) {
9133 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9134 rc = mdb_page_get(txn, pg, &omp, NULL);
9137 mdb_cassert(mc, IS_OVERFLOW(omp));
9138 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9142 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9143 mdb_xcursor_init1(mc, ni);
9144 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9150 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9152 for (i=0; i<n; i++) {
9154 ni = NODEPTR(mp, i);
9157 mdb_midl_xappend(txn->mt_free_pgs, pg);
9162 mc->mc_ki[mc->mc_top] = i;
9163 rc = mdb_cursor_sibling(mc, 1);
9165 if (rc != MDB_NOTFOUND)
9167 /* no more siblings, go back to beginning
9168 * of previous level.
9172 for (i=1; i<mc->mc_snum; i++) {
9174 mc->mc_pg[i] = mx.mc_pg[i];
9179 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9182 txn->mt_flags |= MDB_TXN_ERROR;
9183 } else if (rc == MDB_NOTFOUND) {
9189 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9191 MDB_cursor *mc, *m2;
9194 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9197 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9200 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9203 rc = mdb_cursor_open(txn, dbi, &mc);
9207 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9208 /* Invalidate the dropped DB's cursors */
9209 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9210 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9214 /* Can't delete the main DB */
9215 if (del && dbi > MAIN_DBI) {
9216 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9218 txn->mt_dbflags[dbi] = DB_STALE;
9219 mdb_dbi_close(txn->mt_env, dbi);
9221 txn->mt_flags |= MDB_TXN_ERROR;
9224 /* reset the DB record, mark it dirty */
9225 txn->mt_dbflags[dbi] |= DB_DIRTY;
9226 txn->mt_dbs[dbi].md_depth = 0;
9227 txn->mt_dbs[dbi].md_branch_pages = 0;
9228 txn->mt_dbs[dbi].md_leaf_pages = 0;
9229 txn->mt_dbs[dbi].md_overflow_pages = 0;
9230 txn->mt_dbs[dbi].md_entries = 0;
9231 txn->mt_dbs[dbi].md_root = P_INVALID;
9233 txn->mt_flags |= MDB_TXN_DIRTY;
9236 mdb_cursor_close(mc);
9240 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9242 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9245 txn->mt_dbxs[dbi].md_cmp = cmp;
9249 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9251 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9254 txn->mt_dbxs[dbi].md_dcmp = cmp;
9258 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9260 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9263 txn->mt_dbxs[dbi].md_rel = rel;
9267 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9269 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9272 txn->mt_dbxs[dbi].md_relctx = ctx;
9277 mdb_env_get_maxkeysize(MDB_env *env)
9279 return ENV_MAXKEY(env);
9283 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9285 unsigned int i, rdrs;
9288 int rc = 0, first = 1;
9292 if (!env->me_txns) {
9293 return func("(no reader locks)\n", ctx);
9295 rdrs = env->me_txns->mti_numreaders;
9296 mr = env->me_txns->mti_readers;
9297 for (i=0; i<rdrs; i++) {
9299 txnid_t txnid = mr[i].mr_txnid;
9300 sprintf(buf, txnid == (txnid_t)-1 ?
9301 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9302 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9305 rc = func(" pid thread txnid\n", ctx);
9309 rc = func(buf, ctx);
9315 rc = func("(no active readers)\n", ctx);
9320 /** Insert pid into list if not already present.
9321 * return -1 if already present.
9324 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9326 /* binary search of pid in list */
9328 unsigned cursor = 1;
9330 unsigned n = ids[0];
9333 unsigned pivot = n >> 1;
9334 cursor = base + pivot + 1;
9335 val = pid - ids[cursor];
9340 } else if ( val > 0 ) {
9345 /* found, so it's a duplicate */
9354 for (n = ids[0]; n > cursor; n--)
9361 mdb_reader_check(MDB_env *env, int *dead)
9363 unsigned int i, j, rdrs;
9365 MDB_PID_T *pids, pid;
9374 rdrs = env->me_txns->mti_numreaders;
9375 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9379 mr = env->me_txns->mti_readers;
9380 for (i=0; i<rdrs; i++) {
9381 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9383 if (mdb_pid_insert(pids, pid) == 0) {
9384 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9386 /* Recheck, a new process may have reused pid */
9387 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9388 for (j=i; j<rdrs; j++)
9389 if (mr[j].mr_pid == pid) {
9390 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9391 (unsigned) pid, mr[j].mr_txnid));
9396 UNLOCK_MUTEX_R(env);