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)) {
2631 size += env->me_maxdbs * sizeof(MDB_cursor *);
2632 /* child txns use parent's dbiseqs */
2634 size += env->me_maxdbs * sizeof(unsigned int);
2637 if ((txn = calloc(1, size)) == NULL) {
2638 DPRINTF(("calloc: %s", strerror(errno)));
2641 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2642 if (flags & MDB_RDONLY) {
2643 txn->mt_flags |= MDB_TXN_RDONLY;
2644 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2645 txn->mt_dbiseqs = env->me_dbiseqs;
2647 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2649 txn->mt_dbiseqs = parent->mt_dbiseqs;
2650 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2652 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2653 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2661 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2662 if (!txn->mt_u.dirty_list ||
2663 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2665 free(txn->mt_u.dirty_list);
2669 txn->mt_txnid = parent->mt_txnid;
2670 txn->mt_dirty_room = parent->mt_dirty_room;
2671 txn->mt_u.dirty_list[0].mid = 0;
2672 txn->mt_spill_pgs = NULL;
2673 txn->mt_next_pgno = parent->mt_next_pgno;
2674 parent->mt_child = txn;
2675 txn->mt_parent = parent;
2676 txn->mt_numdbs = parent->mt_numdbs;
2677 txn->mt_flags = parent->mt_flags;
2678 txn->mt_dbxs = parent->mt_dbxs;
2679 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2680 /* Copy parent's mt_dbflags, but clear DB_NEW */
2681 for (i=0; i<txn->mt_numdbs; i++)
2682 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2684 ntxn = (MDB_ntxn *)txn;
2685 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2686 if (env->me_pghead) {
2687 size = MDB_IDL_SIZEOF(env->me_pghead);
2688 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2690 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2695 rc = mdb_cursor_shadow(parent, txn);
2697 mdb_txn_reset0(txn, "beginchild-fail");
2699 rc = mdb_txn_renew0(txn);
2702 if (txn != env->me_txn0)
2706 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2707 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2708 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2715 mdb_txn_env(MDB_txn *txn)
2717 if(!txn) return NULL;
2721 /** Export or close DBI handles opened in this txn. */
2723 mdb_dbis_update(MDB_txn *txn, int keep)
2726 MDB_dbi n = txn->mt_numdbs;
2727 MDB_env *env = txn->mt_env;
2728 unsigned char *tdbflags = txn->mt_dbflags;
2730 for (i = n; --i >= 2;) {
2731 if (tdbflags[i] & DB_NEW) {
2733 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2735 char *ptr = env->me_dbxs[i].md_name.mv_data;
2737 env->me_dbxs[i].md_name.mv_data = NULL;
2738 env->me_dbxs[i].md_name.mv_size = 0;
2739 env->me_dbflags[i] = 0;
2740 env->me_dbiseqs[i]++;
2746 if (keep && env->me_numdbs < n)
2750 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2751 * May be called twice for readonly txns: First reset it, then abort.
2752 * @param[in] txn the transaction handle to reset
2753 * @param[in] act why the transaction is being reset
2756 mdb_txn_reset0(MDB_txn *txn, const char *act)
2758 MDB_env *env = txn->mt_env;
2760 /* Close any DBI handles opened in this txn */
2761 mdb_dbis_update(txn, 0);
2763 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2764 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2765 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2767 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2768 if (txn->mt_u.reader) {
2769 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2770 if (!(env->me_flags & MDB_NOTLS))
2771 txn->mt_u.reader = NULL; /* txn does not own reader */
2773 txn->mt_numdbs = 0; /* close nothing if called again */
2774 txn->mt_dbxs = NULL; /* mark txn as reset */
2776 mdb_cursors_close(txn, 0);
2778 if (!(env->me_flags & MDB_WRITEMAP)) {
2779 mdb_dlist_free(txn);
2781 mdb_midl_free(env->me_pghead);
2783 if (txn->mt_parent) {
2784 txn->mt_parent->mt_child = NULL;
2785 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2786 mdb_midl_free(txn->mt_free_pgs);
2787 mdb_midl_free(txn->mt_spill_pgs);
2788 free(txn->mt_u.dirty_list);
2792 if (mdb_midl_shrink(&txn->mt_free_pgs))
2793 env->me_free_pgs = txn->mt_free_pgs;
2794 env->me_pghead = NULL;
2798 /* The writer mutex was locked in mdb_txn_begin. */
2800 UNLOCK_MUTEX_W(env);
2805 mdb_txn_reset(MDB_txn *txn)
2810 /* This call is only valid for read-only txns */
2811 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2814 mdb_txn_reset0(txn, "reset");
2818 mdb_txn_abort(MDB_txn *txn)
2824 mdb_txn_abort(txn->mt_child);
2826 mdb_txn_reset0(txn, "abort");
2827 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2828 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2829 txn->mt_u.reader->mr_pid = 0;
2831 if (txn != txn->mt_env->me_txn0)
2835 /** Save the freelist as of this transaction to the freeDB.
2836 * This changes the freelist. Keep trying until it stabilizes.
2839 mdb_freelist_save(MDB_txn *txn)
2841 /* env->me_pghead[] can grow and shrink during this call.
2842 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2843 * Page numbers cannot disappear from txn->mt_free_pgs[].
2846 MDB_env *env = txn->mt_env;
2847 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2848 txnid_t pglast = 0, head_id = 0;
2849 pgno_t freecnt = 0, *free_pgs, *mop;
2850 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2852 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2854 if (env->me_pghead) {
2855 /* Make sure first page of freeDB is touched and on freelist */
2856 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2857 if (rc && rc != MDB_NOTFOUND)
2861 if (!env->me_pghead && txn->mt_loose_pgs) {
2862 /* Put loose page numbers in mt_free_pgs, since
2863 * we may be unable to return them to me_pghead.
2865 MDB_page *mp = txn->mt_loose_pgs;
2866 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2868 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2869 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2870 txn->mt_loose_pgs = NULL;
2871 txn->mt_loose_count = 0;
2874 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2875 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2876 ? SSIZE_MAX : maxfree_1pg;
2879 /* Come back here after each Put() in case freelist changed */
2884 /* If using records from freeDB which we have not yet
2885 * deleted, delete them and any we reserved for me_pghead.
2887 while (pglast < env->me_pglast) {
2888 rc = mdb_cursor_first(&mc, &key, NULL);
2891 pglast = head_id = *(txnid_t *)key.mv_data;
2892 total_room = head_room = 0;
2893 mdb_tassert(txn, pglast <= env->me_pglast);
2894 rc = mdb_cursor_del(&mc, 0);
2899 /* Save the IDL of pages freed by this txn, to a single record */
2900 if (freecnt < txn->mt_free_pgs[0]) {
2902 /* Make sure last page of freeDB is touched and on freelist */
2903 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2904 if (rc && rc != MDB_NOTFOUND)
2907 free_pgs = txn->mt_free_pgs;
2908 /* Write to last page of freeDB */
2909 key.mv_size = sizeof(txn->mt_txnid);
2910 key.mv_data = &txn->mt_txnid;
2912 freecnt = free_pgs[0];
2913 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2914 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2917 /* Retry if mt_free_pgs[] grew during the Put() */
2918 free_pgs = txn->mt_free_pgs;
2919 } while (freecnt < free_pgs[0]);
2920 mdb_midl_sort(free_pgs);
2921 memcpy(data.mv_data, free_pgs, data.mv_size);
2924 unsigned int i = free_pgs[0];
2925 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2926 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2928 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2934 mop = env->me_pghead;
2935 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2937 /* Reserve records for me_pghead[]. Split it if multi-page,
2938 * to avoid searching freeDB for a page range. Use keys in
2939 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2941 if (total_room >= mop_len) {
2942 if (total_room == mop_len || --more < 0)
2944 } else if (head_room >= maxfree_1pg && head_id > 1) {
2945 /* Keep current record (overflow page), add a new one */
2949 /* (Re)write {key = head_id, IDL length = head_room} */
2950 total_room -= head_room;
2951 head_room = mop_len - total_room;
2952 if (head_room > maxfree_1pg && head_id > 1) {
2953 /* Overflow multi-page for part of me_pghead */
2954 head_room /= head_id; /* amortize page sizes */
2955 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2956 } else if (head_room < 0) {
2957 /* Rare case, not bothering to delete this record */
2960 key.mv_size = sizeof(head_id);
2961 key.mv_data = &head_id;
2962 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2963 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2966 /* IDL is initially empty, zero out at least the length */
2967 pgs = (pgno_t *)data.mv_data;
2968 j = head_room > clean_limit ? head_room : 0;
2972 total_room += head_room;
2975 /* Return loose page numbers to me_pghead, though usually none are
2976 * left at this point. The pages themselves remain in dirty_list.
2978 if (txn->mt_loose_pgs) {
2979 MDB_page *mp = txn->mt_loose_pgs;
2980 unsigned count = txn->mt_loose_count;
2982 /* Room for loose pages + temp IDL with same */
2983 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
2985 mop = env->me_pghead;
2986 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
2987 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
2988 loose[ ++count ] = mp->mp_pgno;
2990 mdb_midl_sort(loose);
2991 mdb_midl_xmerge(mop, loose);
2992 txn->mt_loose_pgs = NULL;
2993 txn->mt_loose_count = 0;
2997 /* Fill in the reserved me_pghead records */
3003 rc = mdb_cursor_first(&mc, &key, &data);
3004 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3005 txnid_t id = *(txnid_t *)key.mv_data;
3006 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3009 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3011 if (len > mop_len) {
3013 data.mv_size = (len + 1) * sizeof(MDB_ID);
3015 data.mv_data = mop -= len;
3018 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3020 if (rc || !(mop_len -= len))
3027 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3028 * @param[in] txn the transaction that's being committed
3029 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3030 * @return 0 on success, non-zero on failure.
3033 mdb_page_flush(MDB_txn *txn, int keep)
3035 MDB_env *env = txn->mt_env;
3036 MDB_ID2L dl = txn->mt_u.dirty_list;
3037 unsigned psize = env->me_psize, j;
3038 int i, pagecount = dl[0].mid, rc;
3039 size_t size = 0, pos = 0;
3041 MDB_page *dp = NULL;
3045 struct iovec iov[MDB_COMMIT_PAGES];
3046 ssize_t wpos = 0, wsize = 0, wres;
3047 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3053 if (env->me_flags & MDB_WRITEMAP) {
3054 /* Clear dirty flags */
3055 while (++i <= pagecount) {
3057 /* Don't flush this page yet */
3058 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3059 dp->mp_flags &= ~P_KEEP;
3063 dp->mp_flags &= ~P_DIRTY;
3068 /* Write the pages */
3070 if (++i <= pagecount) {
3072 /* Don't flush this page yet */
3073 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3074 dp->mp_flags &= ~P_KEEP;
3079 /* clear dirty flag */
3080 dp->mp_flags &= ~P_DIRTY;
3083 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3088 /* Windows actually supports scatter/gather I/O, but only on
3089 * unbuffered file handles. Since we're relying on the OS page
3090 * cache for all our data, that's self-defeating. So we just
3091 * write pages one at a time. We use the ov structure to set
3092 * the write offset, to at least save the overhead of a Seek
3095 DPRINTF(("committing page %"Z"u", pgno));
3096 memset(&ov, 0, sizeof(ov));
3097 ov.Offset = pos & 0xffffffff;
3098 ov.OffsetHigh = pos >> 16 >> 16;
3099 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3101 DPRINTF(("WriteFile: %d", rc));
3105 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3106 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3108 /* Write previous page(s) */
3109 #ifdef MDB_USE_PWRITEV
3110 wres = pwritev(env->me_fd, iov, n, wpos);
3113 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3115 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3117 DPRINTF(("lseek: %s", strerror(rc)));
3120 wres = writev(env->me_fd, iov, n);
3123 if (wres != wsize) {
3126 DPRINTF(("Write error: %s", strerror(rc)));
3128 rc = EIO; /* TODO: Use which error code? */
3129 DPUTS("short write, filesystem full?");
3140 DPRINTF(("committing page %"Z"u", pgno));
3141 next_pos = pos + size;
3142 iov[n].iov_len = size;
3143 iov[n].iov_base = (char *)dp;
3149 /* MIPS has cache coherency issues, this is a no-op everywhere else
3150 * Note: for any size >= on-chip cache size, entire on-chip cache is
3153 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3155 for (i = keep; ++i <= pagecount; ) {
3157 /* This is a page we skipped above */
3160 dl[j].mid = dp->mp_pgno;
3163 mdb_dpage_free(env, dp);
3168 txn->mt_dirty_room += i - j;
3174 mdb_txn_commit(MDB_txn *txn)
3180 if (txn == NULL || txn->mt_env == NULL)
3183 if (txn->mt_child) {
3184 rc = mdb_txn_commit(txn->mt_child);
3185 txn->mt_child = NULL;
3192 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3193 mdb_dbis_update(txn, 1);
3194 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3199 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3200 DPUTS("error flag is set, can't commit");
3202 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3207 if (txn->mt_parent) {
3208 MDB_txn *parent = txn->mt_parent;
3212 unsigned x, y, len, ps_len;
3214 /* Append our free list to parent's */
3215 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3218 mdb_midl_free(txn->mt_free_pgs);
3219 /* Failures after this must either undo the changes
3220 * to the parent or set MDB_TXN_ERROR in the parent.
3223 parent->mt_next_pgno = txn->mt_next_pgno;
3224 parent->mt_flags = txn->mt_flags;
3226 /* Merge our cursors into parent's and close them */
3227 mdb_cursors_close(txn, 1);
3229 /* Update parent's DB table. */
3230 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3231 parent->mt_numdbs = txn->mt_numdbs;
3232 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3233 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3234 for (i=2; i<txn->mt_numdbs; i++) {
3235 /* preserve parent's DB_NEW status */
3236 x = parent->mt_dbflags[i] & DB_NEW;
3237 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3240 dst = parent->mt_u.dirty_list;
3241 src = txn->mt_u.dirty_list;
3242 /* Remove anything in our dirty list from parent's spill list */
3243 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3245 pspill[0] = (pgno_t)-1;
3246 /* Mark our dirty pages as deleted in parent spill list */
3247 for (i=0, len=src[0].mid; ++i <= len; ) {
3248 MDB_ID pn = src[i].mid << 1;
3249 while (pn > pspill[x])
3251 if (pn == pspill[x]) {
3256 /* Squash deleted pagenums if we deleted any */
3257 for (x=y; ++x <= ps_len; )
3258 if (!(pspill[x] & 1))
3259 pspill[++y] = pspill[x];
3263 /* Find len = length of merging our dirty list with parent's */
3265 dst[0].mid = 0; /* simplify loops */
3266 if (parent->mt_parent) {
3267 len = x + src[0].mid;
3268 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3269 for (i = x; y && i; y--) {
3270 pgno_t yp = src[y].mid;
3271 while (yp < dst[i].mid)
3273 if (yp == dst[i].mid) {
3278 } else { /* Simplify the above for single-ancestor case */
3279 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3281 /* Merge our dirty list with parent's */
3283 for (i = len; y; dst[i--] = src[y--]) {
3284 pgno_t yp = src[y].mid;
3285 while (yp < dst[x].mid)
3286 dst[i--] = dst[x--];
3287 if (yp == dst[x].mid)
3288 free(dst[x--].mptr);
3290 mdb_tassert(txn, i == x);
3292 free(txn->mt_u.dirty_list);
3293 parent->mt_dirty_room = txn->mt_dirty_room;
3294 if (txn->mt_spill_pgs) {
3295 if (parent->mt_spill_pgs) {
3296 /* TODO: Prevent failure here, so parent does not fail */
3297 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3299 parent->mt_flags |= MDB_TXN_ERROR;
3300 mdb_midl_free(txn->mt_spill_pgs);
3301 mdb_midl_sort(parent->mt_spill_pgs);
3303 parent->mt_spill_pgs = txn->mt_spill_pgs;
3307 /* Append our loose page list to parent's */
3308 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3310 *lp = txn->mt_loose_pgs;
3311 parent->mt_loose_count += txn->mt_loose_count;
3313 parent->mt_child = NULL;
3314 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3319 if (txn != env->me_txn) {
3320 DPUTS("attempt to commit unknown transaction");
3325 mdb_cursors_close(txn, 0);
3327 if (!txn->mt_u.dirty_list[0].mid &&
3328 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3331 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3332 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3334 /* Update DB root pointers */
3335 if (txn->mt_numdbs > 2) {
3339 data.mv_size = sizeof(MDB_db);
3341 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3342 for (i = 2; i < txn->mt_numdbs; i++) {
3343 if (txn->mt_dbflags[i] & DB_DIRTY) {
3344 if (TXN_DBI_CHANGED(txn, i)) {
3348 data.mv_data = &txn->mt_dbs[i];
3349 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3356 rc = mdb_freelist_save(txn);
3360 mdb_midl_free(env->me_pghead);
3361 env->me_pghead = NULL;
3362 if (mdb_midl_shrink(&txn->mt_free_pgs))
3363 env->me_free_pgs = txn->mt_free_pgs;
3369 if ((rc = mdb_page_flush(txn, 0)) ||
3370 (rc = mdb_env_sync(env, 0)) ||
3371 (rc = mdb_env_write_meta(txn)))
3374 /* Free P_LOOSE pages left behind in dirty_list */
3375 if (!(env->me_flags & MDB_WRITEMAP))
3376 mdb_dlist_free(txn);
3381 mdb_dbis_update(txn, 1);
3384 UNLOCK_MUTEX_W(env);
3385 if (txn != env->me_txn0)
3395 /** Read the environment parameters of a DB environment before
3396 * mapping it into memory.
3397 * @param[in] env the environment handle
3398 * @param[out] meta address of where to store the meta information
3399 * @return 0 on success, non-zero on failure.
3402 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3408 enum { Size = sizeof(pbuf) };
3410 /* We don't know the page size yet, so use a minimum value.
3411 * Read both meta pages so we can use the latest one.
3414 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3418 memset(&ov, 0, sizeof(ov));
3420 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3421 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3424 rc = pread(env->me_fd, &pbuf, Size, off);
3427 if (rc == 0 && off == 0)
3429 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3430 DPRINTF(("read: %s", mdb_strerror(rc)));
3434 p = (MDB_page *)&pbuf;
3436 if (!F_ISSET(p->mp_flags, P_META)) {
3437 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3442 if (m->mm_magic != MDB_MAGIC) {
3443 DPUTS("meta has invalid magic");
3447 if (m->mm_version != MDB_DATA_VERSION) {
3448 DPRINTF(("database is version %u, expected version %u",
3449 m->mm_version, MDB_DATA_VERSION));
3450 return MDB_VERSION_MISMATCH;
3453 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3460 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3462 meta->mm_magic = MDB_MAGIC;
3463 meta->mm_version = MDB_DATA_VERSION;
3464 meta->mm_mapsize = env->me_mapsize;
3465 meta->mm_psize = env->me_psize;
3466 meta->mm_last_pg = 1;
3467 meta->mm_flags = env->me_flags & 0xffff;
3468 meta->mm_flags |= MDB_INTEGERKEY;
3469 meta->mm_dbs[0].md_root = P_INVALID;
3470 meta->mm_dbs[1].md_root = P_INVALID;
3473 /** Write the environment parameters of a freshly created DB environment.
3474 * @param[in] env the environment handle
3475 * @param[out] meta address of where to store the meta information
3476 * @return 0 on success, non-zero on failure.
3479 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3487 memset(&ov, 0, sizeof(ov));
3488 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3490 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3493 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3494 len = pwrite(fd, ptr, size, pos); \
3495 rc = (len >= 0); } while(0)
3498 DPUTS("writing new meta page");
3500 psize = env->me_psize;
3502 mdb_env_init_meta0(env, meta);
3504 p = calloc(2, psize);
3506 p->mp_flags = P_META;
3507 *(MDB_meta *)METADATA(p) = *meta;
3509 q = (MDB_page *)((char *)p + psize);
3511 q->mp_flags = P_META;
3512 *(MDB_meta *)METADATA(q) = *meta;
3514 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3517 else if ((unsigned) len == psize * 2)
3525 /** Update the environment info to commit a transaction.
3526 * @param[in] txn the transaction that's being committed
3527 * @return 0 on success, non-zero on failure.
3530 mdb_env_write_meta(MDB_txn *txn)
3533 MDB_meta meta, metab, *mp;
3536 int rc, len, toggle;
3545 toggle = txn->mt_txnid & 1;
3546 DPRINTF(("writing meta page %d for root page %"Z"u",
3547 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3550 mp = env->me_metas[toggle];
3551 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3552 /* Persist any increases of mapsize config */
3553 if (mapsize < env->me_mapsize)
3554 mapsize = env->me_mapsize;
3556 if (env->me_flags & MDB_WRITEMAP) {
3557 mp->mm_mapsize = mapsize;
3558 mp->mm_dbs[0] = txn->mt_dbs[0];
3559 mp->mm_dbs[1] = txn->mt_dbs[1];
3560 mp->mm_last_pg = txn->mt_next_pgno - 1;
3561 mp->mm_txnid = txn->mt_txnid;
3562 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3563 unsigned meta_size = env->me_psize;
3564 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3567 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3568 if (meta_size < env->me_os_psize)
3569 meta_size += meta_size;
3574 if (MDB_MSYNC(ptr, meta_size, rc)) {
3581 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3582 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3584 meta.mm_mapsize = mapsize;
3585 meta.mm_dbs[0] = txn->mt_dbs[0];
3586 meta.mm_dbs[1] = txn->mt_dbs[1];
3587 meta.mm_last_pg = txn->mt_next_pgno - 1;
3588 meta.mm_txnid = txn->mt_txnid;
3590 off = offsetof(MDB_meta, mm_mapsize);
3591 ptr = (char *)&meta + off;
3592 len = sizeof(MDB_meta) - off;
3594 off += env->me_psize;
3597 /* Write to the SYNC fd */
3598 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3599 env->me_fd : env->me_mfd;
3602 memset(&ov, 0, sizeof(ov));
3604 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3608 rc = pwrite(mfd, ptr, len, off);
3611 rc = rc < 0 ? ErrCode() : EIO;
3612 DPUTS("write failed, disk error?");
3613 /* On a failure, the pagecache still contains the new data.
3614 * Write some old data back, to prevent it from being used.
3615 * Use the non-SYNC fd; we know it will fail anyway.
3617 meta.mm_last_pg = metab.mm_last_pg;
3618 meta.mm_txnid = metab.mm_txnid;
3620 memset(&ov, 0, sizeof(ov));
3622 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3624 r2 = pwrite(env->me_fd, ptr, len, off);
3625 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3628 env->me_flags |= MDB_FATAL_ERROR;
3631 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3632 CACHEFLUSH(env->me_map + off, len, DCACHE);
3634 /* Memory ordering issues are irrelevant; since the entire writer
3635 * is wrapped by wmutex, all of these changes will become visible
3636 * after the wmutex is unlocked. Since the DB is multi-version,
3637 * readers will get consistent data regardless of how fresh or
3638 * how stale their view of these values is.
3641 env->me_txns->mti_txnid = txn->mt_txnid;
3646 /** Check both meta pages to see which one is newer.
3647 * @param[in] env the environment handle
3648 * @return meta toggle (0 or 1).
3651 mdb_env_pick_meta(const MDB_env *env)
3653 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3657 mdb_env_create(MDB_env **env)
3661 e = calloc(1, sizeof(MDB_env));
3665 e->me_maxreaders = DEFAULT_READERS;
3666 e->me_maxdbs = e->me_numdbs = 2;
3667 e->me_fd = INVALID_HANDLE_VALUE;
3668 e->me_lfd = INVALID_HANDLE_VALUE;
3669 e->me_mfd = INVALID_HANDLE_VALUE;
3670 #ifdef MDB_USE_POSIX_SEM
3671 e->me_rmutex = SEM_FAILED;
3672 e->me_wmutex = SEM_FAILED;
3674 e->me_pid = getpid();
3675 GET_PAGESIZE(e->me_os_psize);
3676 VGMEMP_CREATE(e,0,0);
3682 mdb_env_map(MDB_env *env, void *addr)
3685 unsigned int flags = env->me_flags;
3689 LONG sizelo, sizehi;
3692 if (flags & MDB_RDONLY) {
3693 /* Don't set explicit map size, use whatever exists */
3698 msize = env->me_mapsize;
3699 sizelo = msize & 0xffffffff;
3700 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3702 /* Windows won't create mappings for zero length files.
3703 * and won't map more than the file size.
3704 * Just set the maxsize right now.
3706 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3707 || !SetEndOfFile(env->me_fd)
3708 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3712 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3713 PAGE_READWRITE : PAGE_READONLY,
3714 sizehi, sizelo, NULL);
3717 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3718 FILE_MAP_WRITE : FILE_MAP_READ,
3720 rc = env->me_map ? 0 : ErrCode();
3725 int prot = PROT_READ;
3726 if (flags & MDB_WRITEMAP) {
3728 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3731 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3733 if (env->me_map == MAP_FAILED) {
3738 if (flags & MDB_NORDAHEAD) {
3739 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3741 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3743 #ifdef POSIX_MADV_RANDOM
3744 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3745 #endif /* POSIX_MADV_RANDOM */
3746 #endif /* MADV_RANDOM */
3750 /* Can happen because the address argument to mmap() is just a
3751 * hint. mmap() can pick another, e.g. if the range is in use.
3752 * The MAP_FIXED flag would prevent that, but then mmap could
3753 * instead unmap existing pages to make room for the new map.
3755 if (addr && env->me_map != addr)
3756 return EBUSY; /* TODO: Make a new MDB_* error code? */
3758 p = (MDB_page *)env->me_map;
3759 env->me_metas[0] = METADATA(p);
3760 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3766 mdb_env_set_mapsize(MDB_env *env, size_t size)
3768 /* If env is already open, caller is responsible for making
3769 * sure there are no active txns.
3777 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3778 else if (size < env->me_mapsize) {
3779 /* If the configured size is smaller, make sure it's
3780 * still big enough. Silently round up to minimum if not.
3782 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3786 munmap(env->me_map, env->me_mapsize);
3787 env->me_mapsize = size;
3788 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3789 rc = mdb_env_map(env, old);
3793 env->me_mapsize = size;
3795 env->me_maxpg = env->me_mapsize / env->me_psize;
3800 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3804 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3809 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3811 if (env->me_map || readers < 1)
3813 env->me_maxreaders = readers;
3818 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3820 if (!env || !readers)
3822 *readers = env->me_maxreaders;
3826 /** Further setup required for opening an LMDB environment
3829 mdb_env_open2(MDB_env *env)
3831 unsigned int flags = env->me_flags;
3832 int i, newenv = 0, rc;
3836 /* See if we should use QueryLimited */
3838 if ((rc & 0xff) > 5)
3839 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3841 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3844 memset(&meta, 0, sizeof(meta));
3846 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3849 DPUTS("new mdbenv");
3851 env->me_psize = env->me_os_psize;
3852 if (env->me_psize > MAX_PAGESIZE)
3853 env->me_psize = MAX_PAGESIZE;
3855 env->me_psize = meta.mm_psize;
3858 /* Was a mapsize configured? */
3859 if (!env->me_mapsize) {
3860 /* If this is a new environment, take the default,
3861 * else use the size recorded in the existing env.
3863 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3864 } else if (env->me_mapsize < meta.mm_mapsize) {
3865 /* If the configured size is smaller, make sure it's
3866 * still big enough. Silently round up to minimum if not.
3868 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3869 if (env->me_mapsize < minsize)
3870 env->me_mapsize = minsize;
3873 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3878 if (flags & MDB_FIXEDMAP)
3879 meta.mm_address = env->me_map;
3880 i = mdb_env_init_meta(env, &meta);
3881 if (i != MDB_SUCCESS) {
3886 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3887 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3889 #if !(MDB_MAXKEYSIZE)
3890 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3892 env->me_maxpg = env->me_mapsize / env->me_psize;
3896 int toggle = mdb_env_pick_meta(env);
3897 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3899 DPRINTF(("opened database version %u, pagesize %u",
3900 env->me_metas[0]->mm_version, env->me_psize));
3901 DPRINTF(("using meta page %d", toggle));
3902 DPRINTF(("depth: %u", db->md_depth));
3903 DPRINTF(("entries: %"Z"u", db->md_entries));
3904 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3905 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3906 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3907 DPRINTF(("root: %"Z"u", db->md_root));
3915 /** Release a reader thread's slot in the reader lock table.
3916 * This function is called automatically when a thread exits.
3917 * @param[in] ptr This points to the slot in the reader lock table.
3920 mdb_env_reader_dest(void *ptr)
3922 MDB_reader *reader = ptr;
3928 /** Junk for arranging thread-specific callbacks on Windows. This is
3929 * necessarily platform and compiler-specific. Windows supports up
3930 * to 1088 keys. Let's assume nobody opens more than 64 environments
3931 * in a single process, for now. They can override this if needed.
3933 #ifndef MAX_TLS_KEYS
3934 #define MAX_TLS_KEYS 64
3936 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3937 static int mdb_tls_nkeys;
3939 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3943 case DLL_PROCESS_ATTACH: break;
3944 case DLL_THREAD_ATTACH: break;
3945 case DLL_THREAD_DETACH:
3946 for (i=0; i<mdb_tls_nkeys; i++) {
3947 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3949 mdb_env_reader_dest(r);
3953 case DLL_PROCESS_DETACH: break;
3958 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3960 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3964 /* Force some symbol references.
3965 * _tls_used forces the linker to create the TLS directory if not already done
3966 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3968 #pragma comment(linker, "/INCLUDE:_tls_used")
3969 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3970 #pragma const_seg(".CRT$XLB")
3971 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3972 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3975 #pragma comment(linker, "/INCLUDE:__tls_used")
3976 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3977 #pragma data_seg(".CRT$XLB")
3978 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3980 #endif /* WIN 32/64 */
3981 #endif /* !__GNUC__ */
3984 /** Downgrade the exclusive lock on the region back to shared */
3986 mdb_env_share_locks(MDB_env *env, int *excl)
3988 int rc = 0, toggle = mdb_env_pick_meta(env);
3990 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3995 /* First acquire a shared lock. The Unlock will
3996 * then release the existing exclusive lock.
3998 memset(&ov, 0, sizeof(ov));
3999 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4002 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4008 struct flock lock_info;
4009 /* The shared lock replaces the existing lock */
4010 memset((void *)&lock_info, 0, sizeof(lock_info));
4011 lock_info.l_type = F_RDLCK;
4012 lock_info.l_whence = SEEK_SET;
4013 lock_info.l_start = 0;
4014 lock_info.l_len = 1;
4015 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4016 (rc = ErrCode()) == EINTR) ;
4017 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4024 /** Try to get exlusive lock, otherwise shared.
4025 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4028 mdb_env_excl_lock(MDB_env *env, int *excl)
4032 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4036 memset(&ov, 0, sizeof(ov));
4037 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4044 struct flock lock_info;
4045 memset((void *)&lock_info, 0, sizeof(lock_info));
4046 lock_info.l_type = F_WRLCK;
4047 lock_info.l_whence = SEEK_SET;
4048 lock_info.l_start = 0;
4049 lock_info.l_len = 1;
4050 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4051 (rc = ErrCode()) == EINTR) ;
4055 # ifdef MDB_USE_POSIX_SEM
4056 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4059 lock_info.l_type = F_RDLCK;
4060 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4061 (rc = ErrCode()) == EINTR) ;
4071 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4073 * @(#) $Revision: 5.1 $
4074 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4075 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4077 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4081 * Please do not copyright this code. This code is in the public domain.
4083 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4084 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4085 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4086 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4087 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4088 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4089 * PERFORMANCE OF THIS SOFTWARE.
4092 * chongo <Landon Curt Noll> /\oo/\
4093 * http://www.isthe.com/chongo/
4095 * Share and Enjoy! :-)
4098 typedef unsigned long long mdb_hash_t;
4099 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4101 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4102 * @param[in] val value to hash
4103 * @param[in] hval initial value for hash
4104 * @return 64 bit hash
4106 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4107 * hval arg on the first call.
4110 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4112 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4113 unsigned char *end = s + val->mv_size;
4115 * FNV-1a hash each octet of the string
4118 /* xor the bottom with the current octet */
4119 hval ^= (mdb_hash_t)*s++;
4121 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4122 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4123 (hval << 7) + (hval << 8) + (hval << 40);
4125 /* return our new hash value */
4129 /** Hash the string and output the encoded hash.
4130 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4131 * very short name limits. We don't care about the encoding being reversible,
4132 * we just want to preserve as many bits of the input as possible in a
4133 * small printable string.
4134 * @param[in] str string to hash
4135 * @param[out] encbuf an array of 11 chars to hold the hash
4137 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4140 mdb_pack85(unsigned long l, char *out)
4144 for (i=0; i<5; i++) {
4145 *out++ = mdb_a85[l % 85];
4151 mdb_hash_enc(MDB_val *val, char *encbuf)
4153 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4155 mdb_pack85(h, encbuf);
4156 mdb_pack85(h>>32, encbuf+5);
4161 /** Open and/or initialize the lock region for the environment.
4162 * @param[in] env The LMDB environment.
4163 * @param[in] lpath The pathname of the file used for the lock region.
4164 * @param[in] mode The Unix permissions for the file, if we create it.
4165 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4166 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4167 * @return 0 on success, non-zero on failure.
4170 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4173 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4175 # define MDB_ERRCODE_ROFS EROFS
4176 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4177 # define MDB_CLOEXEC O_CLOEXEC
4180 # define MDB_CLOEXEC 0
4187 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4188 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4189 FILE_ATTRIBUTE_NORMAL, NULL);
4191 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4193 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4195 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4200 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4201 /* Lose record locks when exec*() */
4202 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4203 fcntl(env->me_lfd, F_SETFD, fdflags);
4206 if (!(env->me_flags & MDB_NOTLS)) {
4207 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4210 env->me_flags |= MDB_ENV_TXKEY;
4212 /* Windows TLS callbacks need help finding their TLS info. */
4213 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4217 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4221 /* Try to get exclusive lock. If we succeed, then
4222 * nobody is using the lock region and we should initialize it.
4224 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4227 size = GetFileSize(env->me_lfd, NULL);
4229 size = lseek(env->me_lfd, 0, SEEK_END);
4230 if (size == -1) goto fail_errno;
4232 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4233 if (size < rsize && *excl > 0) {
4235 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4236 || !SetEndOfFile(env->me_lfd))
4239 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4243 size = rsize - sizeof(MDB_txninfo);
4244 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4249 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4251 if (!mh) goto fail_errno;
4252 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4254 if (!env->me_txns) goto fail_errno;
4256 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4258 if (m == MAP_FAILED) goto fail_errno;
4264 BY_HANDLE_FILE_INFORMATION stbuf;
4273 if (!mdb_sec_inited) {
4274 InitializeSecurityDescriptor(&mdb_null_sd,
4275 SECURITY_DESCRIPTOR_REVISION);
4276 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4277 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4278 mdb_all_sa.bInheritHandle = FALSE;
4279 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4282 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4283 idbuf.volume = stbuf.dwVolumeSerialNumber;
4284 idbuf.nhigh = stbuf.nFileIndexHigh;
4285 idbuf.nlow = stbuf.nFileIndexLow;
4286 val.mv_data = &idbuf;
4287 val.mv_size = sizeof(idbuf);
4288 mdb_hash_enc(&val, encbuf);
4289 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4290 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4291 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4292 if (!env->me_rmutex) goto fail_errno;
4293 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4294 if (!env->me_wmutex) goto fail_errno;
4295 #elif defined(MDB_USE_POSIX_SEM)
4304 #if defined(__NetBSD__)
4305 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4307 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4308 idbuf.dev = stbuf.st_dev;
4309 idbuf.ino = stbuf.st_ino;
4310 val.mv_data = &idbuf;
4311 val.mv_size = sizeof(idbuf);
4312 mdb_hash_enc(&val, encbuf);
4313 #ifdef MDB_SHORT_SEMNAMES
4314 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4316 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4317 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4318 /* Clean up after a previous run, if needed: Try to
4319 * remove both semaphores before doing anything else.
4321 sem_unlink(env->me_txns->mti_rmname);
4322 sem_unlink(env->me_txns->mti_wmname);
4323 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4324 O_CREAT|O_EXCL, mode, 1);
4325 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4326 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4327 O_CREAT|O_EXCL, mode, 1);
4328 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4329 #else /* MDB_USE_POSIX_SEM */
4330 pthread_mutexattr_t mattr;
4332 if ((rc = pthread_mutexattr_init(&mattr))
4333 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4334 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4335 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4337 pthread_mutexattr_destroy(&mattr);
4338 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4340 env->me_txns->mti_magic = MDB_MAGIC;
4341 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4342 env->me_txns->mti_txnid = 0;
4343 env->me_txns->mti_numreaders = 0;
4346 if (env->me_txns->mti_magic != MDB_MAGIC) {
4347 DPUTS("lock region has invalid magic");
4351 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4352 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4353 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4354 rc = MDB_VERSION_MISMATCH;
4358 if (rc && rc != EACCES && rc != EAGAIN) {
4362 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4363 if (!env->me_rmutex) goto fail_errno;
4364 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4365 if (!env->me_wmutex) goto fail_errno;
4366 #elif defined(MDB_USE_POSIX_SEM)
4367 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4368 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4369 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4370 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4381 /** The name of the lock file in the DB environment */
4382 #define LOCKNAME "/lock.mdb"
4383 /** The name of the data file in the DB environment */
4384 #define DATANAME "/data.mdb"
4385 /** The suffix of the lock file when no subdir is used */
4386 #define LOCKSUFF "-lock"
4387 /** Only a subset of the @ref mdb_env flags can be changed
4388 * at runtime. Changing other flags requires closing the
4389 * environment and re-opening it with the new flags.
4391 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4392 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4393 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4395 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4396 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4400 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4402 int oflags, rc, len, excl = -1;
4403 char *lpath, *dpath;
4405 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4409 if (flags & MDB_NOSUBDIR) {
4410 rc = len + sizeof(LOCKSUFF) + len + 1;
4412 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4417 if (flags & MDB_NOSUBDIR) {
4418 dpath = lpath + len + sizeof(LOCKSUFF);
4419 sprintf(lpath, "%s" LOCKSUFF, path);
4420 strcpy(dpath, path);
4422 dpath = lpath + len + sizeof(LOCKNAME);
4423 sprintf(lpath, "%s" LOCKNAME, path);
4424 sprintf(dpath, "%s" DATANAME, path);
4428 flags |= env->me_flags;
4429 if (flags & MDB_RDONLY) {
4430 /* silently ignore WRITEMAP when we're only getting read access */
4431 flags &= ~MDB_WRITEMAP;
4433 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4434 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4437 env->me_flags = flags |= MDB_ENV_ACTIVE;
4441 env->me_path = strdup(path);
4442 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4443 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4444 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4445 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4450 /* For RDONLY, get lockfile after we know datafile exists */
4451 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4452 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4458 if (F_ISSET(flags, MDB_RDONLY)) {
4459 oflags = GENERIC_READ;
4460 len = OPEN_EXISTING;
4462 oflags = GENERIC_READ|GENERIC_WRITE;
4465 mode = FILE_ATTRIBUTE_NORMAL;
4466 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4467 NULL, len, mode, NULL);
4469 if (F_ISSET(flags, MDB_RDONLY))
4472 oflags = O_RDWR | O_CREAT;
4474 env->me_fd = open(dpath, oflags, mode);
4476 if (env->me_fd == INVALID_HANDLE_VALUE) {
4481 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4482 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4487 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4488 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4489 env->me_mfd = env->me_fd;
4491 /* Synchronous fd for meta writes. Needed even with
4492 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4495 len = OPEN_EXISTING;
4496 env->me_mfd = CreateFile(dpath, oflags,
4497 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4498 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4501 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4503 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4508 DPRINTF(("opened dbenv %p", (void *) env));
4510 rc = mdb_env_share_locks(env, &excl);
4514 if (!((flags & MDB_RDONLY) ||
4515 (env->me_pbuf = calloc(1, env->me_psize))))
4517 if (!(flags & MDB_RDONLY)) {
4519 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4520 (sizeof(MDB_db)+sizeof(MDB_cursor)+sizeof(unsigned int)+1);
4521 txn = calloc(1, size);
4523 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4524 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4525 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4526 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4537 mdb_env_close0(env, excl);
4543 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4545 mdb_env_close0(MDB_env *env, int excl)
4549 if (!(env->me_flags & MDB_ENV_ACTIVE))
4552 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4553 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4554 free(env->me_dbxs[i].md_name.mv_data);
4557 free(env->me_dbiseqs);
4558 free(env->me_dbflags);
4561 free(env->me_dirty_list);
4563 mdb_midl_free(env->me_free_pgs);
4565 if (env->me_flags & MDB_ENV_TXKEY) {
4566 pthread_key_delete(env->me_txkey);
4568 /* Delete our key from the global list */
4569 for (i=0; i<mdb_tls_nkeys; i++)
4570 if (mdb_tls_keys[i] == env->me_txkey) {
4571 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4579 munmap(env->me_map, env->me_mapsize);
4581 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4582 (void) close(env->me_mfd);
4583 if (env->me_fd != INVALID_HANDLE_VALUE)
4584 (void) close(env->me_fd);
4586 MDB_PID_T pid = env->me_pid;
4587 /* Clearing readers is done in this function because
4588 * me_txkey with its destructor must be disabled first.
4590 for (i = env->me_numreaders; --i >= 0; )
4591 if (env->me_txns->mti_readers[i].mr_pid == pid)
4592 env->me_txns->mti_readers[i].mr_pid = 0;
4594 if (env->me_rmutex) {
4595 CloseHandle(env->me_rmutex);
4596 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4598 /* Windows automatically destroys the mutexes when
4599 * the last handle closes.
4601 #elif defined(MDB_USE_POSIX_SEM)
4602 if (env->me_rmutex != SEM_FAILED) {
4603 sem_close(env->me_rmutex);
4604 if (env->me_wmutex != SEM_FAILED)
4605 sem_close(env->me_wmutex);
4606 /* If we have the filelock: If we are the
4607 * only remaining user, clean up semaphores.
4610 mdb_env_excl_lock(env, &excl);
4612 sem_unlink(env->me_txns->mti_rmname);
4613 sem_unlink(env->me_txns->mti_wmname);
4617 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4619 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4622 /* Unlock the lockfile. Windows would have unlocked it
4623 * after closing anyway, but not necessarily at once.
4625 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4628 (void) close(env->me_lfd);
4631 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4636 mdb_env_close(MDB_env *env)
4643 VGMEMP_DESTROY(env);
4644 while ((dp = env->me_dpages) != NULL) {
4645 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4646 env->me_dpages = dp->mp_next;
4650 mdb_env_close0(env, 0);
4654 /** Compare two items pointing at aligned size_t's */
4656 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4658 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4659 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4662 /** Compare two items pointing at aligned unsigned int's */
4664 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4666 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4667 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4670 /** Compare two items pointing at unsigned ints of unknown alignment.
4671 * Nodes and keys are guaranteed to be 2-byte aligned.
4674 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4676 #if BYTE_ORDER == LITTLE_ENDIAN
4677 unsigned short *u, *c;
4680 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4681 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4684 } while(!x && u > (unsigned short *)a->mv_data);
4687 unsigned short *u, *c, *end;
4690 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4691 u = (unsigned short *)a->mv_data;
4692 c = (unsigned short *)b->mv_data;
4695 } while(!x && u < end);
4700 /** Compare two items pointing at size_t's of unknown alignment. */
4701 #ifdef MISALIGNED_OK
4702 # define mdb_cmp_clong mdb_cmp_long
4704 # define mdb_cmp_clong mdb_cmp_cint
4707 /** Compare two items lexically */
4709 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4716 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4722 diff = memcmp(a->mv_data, b->mv_data, len);
4723 return diff ? diff : len_diff<0 ? -1 : len_diff;
4726 /** Compare two items in reverse byte order */
4728 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4730 const unsigned char *p1, *p2, *p1_lim;
4734 p1_lim = (const unsigned char *)a->mv_data;
4735 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4736 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4738 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4744 while (p1 > p1_lim) {
4745 diff = *--p1 - *--p2;
4749 return len_diff<0 ? -1 : len_diff;
4752 /** Search for key within a page, using binary search.
4753 * Returns the smallest entry larger or equal to the key.
4754 * If exactp is non-null, stores whether the found entry was an exact match
4755 * in *exactp (1 or 0).
4756 * Updates the cursor index with the index of the found entry.
4757 * If no entry larger or equal to the key is found, returns NULL.
4760 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4762 unsigned int i = 0, nkeys;
4765 MDB_page *mp = mc->mc_pg[mc->mc_top];
4766 MDB_node *node = NULL;
4771 nkeys = NUMKEYS(mp);
4773 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4774 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4777 low = IS_LEAF(mp) ? 0 : 1;
4779 cmp = mc->mc_dbx->md_cmp;
4781 /* Branch pages have no data, so if using integer keys,
4782 * alignment is guaranteed. Use faster mdb_cmp_int.
4784 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4785 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4792 nodekey.mv_size = mc->mc_db->md_pad;
4793 node = NODEPTR(mp, 0); /* fake */
4794 while (low <= high) {
4795 i = (low + high) >> 1;
4796 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4797 rc = cmp(key, &nodekey);
4798 DPRINTF(("found leaf index %u [%s], rc = %i",
4799 i, DKEY(&nodekey), rc));
4808 while (low <= high) {
4809 i = (low + high) >> 1;
4811 node = NODEPTR(mp, i);
4812 nodekey.mv_size = NODEKSZ(node);
4813 nodekey.mv_data = NODEKEY(node);
4815 rc = cmp(key, &nodekey);
4818 DPRINTF(("found leaf index %u [%s], rc = %i",
4819 i, DKEY(&nodekey), rc));
4821 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4822 i, DKEY(&nodekey), NODEPGNO(node), rc));
4833 if (rc > 0) { /* Found entry is less than the key. */
4834 i++; /* Skip to get the smallest entry larger than key. */
4836 node = NODEPTR(mp, i);
4839 *exactp = (rc == 0 && nkeys > 0);
4840 /* store the key index */
4841 mc->mc_ki[mc->mc_top] = i;
4843 /* There is no entry larger or equal to the key. */
4846 /* nodeptr is fake for LEAF2 */
4852 mdb_cursor_adjust(MDB_cursor *mc, func)
4856 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4857 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4864 /** Pop a page off the top of the cursor's stack. */
4866 mdb_cursor_pop(MDB_cursor *mc)
4870 MDB_page *top = mc->mc_pg[mc->mc_top];
4876 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4877 DDBI(mc), (void *) mc));
4881 /** Push a page onto the top of the cursor's stack. */
4883 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4885 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4886 DDBI(mc), (void *) mc));
4888 if (mc->mc_snum >= CURSOR_STACK) {
4889 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4890 return MDB_CURSOR_FULL;
4893 mc->mc_top = mc->mc_snum++;
4894 mc->mc_pg[mc->mc_top] = mp;
4895 mc->mc_ki[mc->mc_top] = 0;
4900 /** Find the address of the page corresponding to a given page number.
4901 * @param[in] txn the transaction for this access.
4902 * @param[in] pgno the page number for the page to retrieve.
4903 * @param[out] ret address of a pointer where the page's address will be stored.
4904 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4905 * @return 0 on success, non-zero on failure.
4908 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4910 MDB_env *env = txn->mt_env;
4914 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4918 MDB_ID2L dl = tx2->mt_u.dirty_list;
4920 /* Spilled pages were dirtied in this txn and flushed
4921 * because the dirty list got full. Bring this page
4922 * back in from the map (but don't unspill it here,
4923 * leave that unless page_touch happens again).
4925 if (tx2->mt_spill_pgs) {
4926 MDB_ID pn = pgno << 1;
4927 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4928 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4929 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4934 unsigned x = mdb_mid2l_search(dl, pgno);
4935 if (x <= dl[0].mid && dl[x].mid == pgno) {
4941 } while ((tx2 = tx2->mt_parent) != NULL);
4944 if (pgno < txn->mt_next_pgno) {
4946 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4948 DPRINTF(("page %"Z"u not found", pgno));
4949 txn->mt_flags |= MDB_TXN_ERROR;
4950 return MDB_PAGE_NOTFOUND;
4960 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4961 * The cursor is at the root page, set up the rest of it.
4964 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4966 MDB_page *mp = mc->mc_pg[mc->mc_top];
4970 while (IS_BRANCH(mp)) {
4974 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4975 mdb_cassert(mc, NUMKEYS(mp) > 1);
4976 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4978 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4980 if (flags & MDB_PS_LAST)
4981 i = NUMKEYS(mp) - 1;
4984 node = mdb_node_search(mc, key, &exact);
4986 i = NUMKEYS(mp) - 1;
4988 i = mc->mc_ki[mc->mc_top];
4990 mdb_cassert(mc, i > 0);
4994 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4997 mdb_cassert(mc, i < NUMKEYS(mp));
4998 node = NODEPTR(mp, i);
5000 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5003 mc->mc_ki[mc->mc_top] = i;
5004 if ((rc = mdb_cursor_push(mc, mp)))
5007 if (flags & MDB_PS_MODIFY) {
5008 if ((rc = mdb_page_touch(mc)) != 0)
5010 mp = mc->mc_pg[mc->mc_top];
5015 DPRINTF(("internal error, index points to a %02X page!?",
5017 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5018 return MDB_CORRUPTED;
5021 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5022 key ? DKEY(key) : "null"));
5023 mc->mc_flags |= C_INITIALIZED;
5024 mc->mc_flags &= ~C_EOF;
5029 /** Search for the lowest key under the current branch page.
5030 * This just bypasses a NUMKEYS check in the current page
5031 * before calling mdb_page_search_root(), because the callers
5032 * are all in situations where the current page is known to
5036 mdb_page_search_lowest(MDB_cursor *mc)
5038 MDB_page *mp = mc->mc_pg[mc->mc_top];
5039 MDB_node *node = NODEPTR(mp, 0);
5042 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5045 mc->mc_ki[mc->mc_top] = 0;
5046 if ((rc = mdb_cursor_push(mc, mp)))
5048 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5051 /** Search for the page a given key should be in.
5052 * Push it and its parent pages on the cursor stack.
5053 * @param[in,out] mc the cursor for this operation.
5054 * @param[in] key the key to search for, or NULL for first/last page.
5055 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5056 * are touched (updated with new page numbers).
5057 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5058 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5059 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5060 * @return 0 on success, non-zero on failure.
5063 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5068 /* Make sure the txn is still viable, then find the root from
5069 * the txn's db table and set it as the root of the cursor's stack.
5071 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5072 DPUTS("transaction has failed, must abort");
5075 /* Make sure we're using an up-to-date root */
5076 if (*mc->mc_dbflag & DB_STALE) {
5078 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5080 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5081 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5088 MDB_node *leaf = mdb_node_search(&mc2,
5089 &mc->mc_dbx->md_name, &exact);
5091 return MDB_NOTFOUND;
5092 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5095 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5097 /* The txn may not know this DBI, or another process may
5098 * have dropped and recreated the DB with other flags.
5100 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5101 return MDB_INCOMPATIBLE;
5102 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5104 *mc->mc_dbflag &= ~DB_STALE;
5106 root = mc->mc_db->md_root;
5108 if (root == P_INVALID) { /* Tree is empty. */
5109 DPUTS("tree is empty");
5110 return MDB_NOTFOUND;
5114 mdb_cassert(mc, root > 1);
5115 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5116 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5122 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5123 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5125 if (flags & MDB_PS_MODIFY) {
5126 if ((rc = mdb_page_touch(mc)))
5130 if (flags & MDB_PS_ROOTONLY)
5133 return mdb_page_search_root(mc, key, flags);
5137 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5139 MDB_txn *txn = mc->mc_txn;
5140 pgno_t pg = mp->mp_pgno;
5141 unsigned x = 0, ovpages = mp->mp_pages;
5142 MDB_env *env = txn->mt_env;
5143 MDB_IDL sl = txn->mt_spill_pgs;
5144 MDB_ID pn = pg << 1;
5147 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5148 /* If the page is dirty or on the spill list we just acquired it,
5149 * so we should give it back to our current free list, if any.
5150 * Otherwise put it onto the list of pages we freed in this txn.
5152 * Won't create me_pghead: me_pglast must be inited along with it.
5153 * Unsupported in nested txns: They would need to hide the page
5154 * range in ancestor txns' dirty and spilled lists.
5156 if (env->me_pghead &&
5158 ((mp->mp_flags & P_DIRTY) ||
5159 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5163 MDB_ID2 *dl, ix, iy;
5164 rc = mdb_midl_need(&env->me_pghead, ovpages);
5167 if (!(mp->mp_flags & P_DIRTY)) {
5168 /* This page is no longer spilled */
5175 /* Remove from dirty list */
5176 dl = txn->mt_u.dirty_list;
5178 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5184 mdb_cassert(mc, x > 1);
5186 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5187 txn->mt_flags |= MDB_TXN_ERROR;
5188 return MDB_CORRUPTED;
5191 if (!(env->me_flags & MDB_WRITEMAP))
5192 mdb_dpage_free(env, mp);
5194 /* Insert in me_pghead */
5195 mop = env->me_pghead;
5196 j = mop[0] + ovpages;
5197 for (i = mop[0]; i && mop[i] < pg; i--)
5203 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5207 mc->mc_db->md_overflow_pages -= ovpages;
5211 /** Return the data associated with a given node.
5212 * @param[in] txn The transaction for this operation.
5213 * @param[in] leaf The node being read.
5214 * @param[out] data Updated to point to the node's data.
5215 * @return 0 on success, non-zero on failure.
5218 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5220 MDB_page *omp; /* overflow page */
5224 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5225 data->mv_size = NODEDSZ(leaf);
5226 data->mv_data = NODEDATA(leaf);
5230 /* Read overflow data.
5232 data->mv_size = NODEDSZ(leaf);
5233 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5234 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5235 DPRINTF(("read overflow page %"Z"u failed", pgno));
5238 data->mv_data = METADATA(omp);
5244 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5245 MDB_val *key, MDB_val *data)
5252 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5254 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5257 if (txn->mt_flags & MDB_TXN_ERROR)
5260 mdb_cursor_init(&mc, txn, dbi, &mx);
5261 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5264 /** Find a sibling for a page.
5265 * Replaces the page at the top of the cursor's stack with the
5266 * specified sibling, if one exists.
5267 * @param[in] mc The cursor for this operation.
5268 * @param[in] move_right Non-zero if the right sibling is requested,
5269 * otherwise the left sibling.
5270 * @return 0 on success, non-zero on failure.
5273 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5279 if (mc->mc_snum < 2) {
5280 return MDB_NOTFOUND; /* root has no siblings */
5284 DPRINTF(("parent page is page %"Z"u, index %u",
5285 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5287 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5288 : (mc->mc_ki[mc->mc_top] == 0)) {
5289 DPRINTF(("no more keys left, moving to %s sibling",
5290 move_right ? "right" : "left"));
5291 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5292 /* undo cursor_pop before returning */
5299 mc->mc_ki[mc->mc_top]++;
5301 mc->mc_ki[mc->mc_top]--;
5302 DPRINTF(("just moving to %s index key %u",
5303 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5305 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5307 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5308 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5309 /* mc will be inconsistent if caller does mc_snum++ as above */
5310 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5314 mdb_cursor_push(mc, mp);
5316 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5321 /** Move the cursor to the next data item. */
5323 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5329 if (mc->mc_flags & C_EOF) {
5330 return MDB_NOTFOUND;
5333 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5335 mp = mc->mc_pg[mc->mc_top];
5337 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5338 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5339 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5340 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5341 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5342 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5343 if (rc == MDB_SUCCESS)
5344 MDB_GET_KEY(leaf, key);
5349 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5350 if (op == MDB_NEXT_DUP)
5351 return MDB_NOTFOUND;
5355 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5356 mdb_dbg_pgno(mp), (void *) mc));
5357 if (mc->mc_flags & C_DEL)
5360 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5361 DPUTS("=====> move to next sibling page");
5362 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5363 mc->mc_flags |= C_EOF;
5366 mp = mc->mc_pg[mc->mc_top];
5367 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5369 mc->mc_ki[mc->mc_top]++;
5372 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5373 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5376 key->mv_size = mc->mc_db->md_pad;
5377 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5381 mdb_cassert(mc, IS_LEAF(mp));
5382 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5384 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5385 mdb_xcursor_init1(mc, leaf);
5388 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5391 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5392 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5393 if (rc != MDB_SUCCESS)
5398 MDB_GET_KEY(leaf, key);
5402 /** Move the cursor to the previous data item. */
5404 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5410 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5412 mp = mc->mc_pg[mc->mc_top];
5414 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5415 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5416 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5417 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5418 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5419 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5420 if (rc == MDB_SUCCESS) {
5421 MDB_GET_KEY(leaf, key);
5422 mc->mc_flags &= ~C_EOF;
5427 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5428 if (op == MDB_PREV_DUP)
5429 return MDB_NOTFOUND;
5434 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5435 mdb_dbg_pgno(mp), (void *) mc));
5437 if (mc->mc_ki[mc->mc_top] == 0) {
5438 DPUTS("=====> move to prev sibling page");
5439 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5442 mp = mc->mc_pg[mc->mc_top];
5443 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5444 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5446 mc->mc_ki[mc->mc_top]--;
5448 mc->mc_flags &= ~C_EOF;
5450 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5451 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5454 key->mv_size = mc->mc_db->md_pad;
5455 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5459 mdb_cassert(mc, IS_LEAF(mp));
5460 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5462 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5463 mdb_xcursor_init1(mc, leaf);
5466 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5469 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5470 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5471 if (rc != MDB_SUCCESS)
5476 MDB_GET_KEY(leaf, key);
5480 /** Set the cursor on a specific data item. */
5482 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5483 MDB_cursor_op op, int *exactp)
5487 MDB_node *leaf = NULL;
5490 if (key->mv_size == 0)
5491 return MDB_BAD_VALSIZE;
5494 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5496 /* See if we're already on the right page */
5497 if (mc->mc_flags & C_INITIALIZED) {
5500 mp = mc->mc_pg[mc->mc_top];
5502 mc->mc_ki[mc->mc_top] = 0;
5503 return MDB_NOTFOUND;
5505 if (mp->mp_flags & P_LEAF2) {
5506 nodekey.mv_size = mc->mc_db->md_pad;
5507 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5509 leaf = NODEPTR(mp, 0);
5510 MDB_GET_KEY2(leaf, nodekey);
5512 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5514 /* Probably happens rarely, but first node on the page
5515 * was the one we wanted.
5517 mc->mc_ki[mc->mc_top] = 0;
5524 unsigned int nkeys = NUMKEYS(mp);
5526 if (mp->mp_flags & P_LEAF2) {
5527 nodekey.mv_data = LEAF2KEY(mp,
5528 nkeys-1, nodekey.mv_size);
5530 leaf = NODEPTR(mp, nkeys-1);
5531 MDB_GET_KEY2(leaf, nodekey);
5533 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5535 /* last node was the one we wanted */
5536 mc->mc_ki[mc->mc_top] = nkeys-1;
5542 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5543 /* This is definitely the right page, skip search_page */
5544 if (mp->mp_flags & P_LEAF2) {
5545 nodekey.mv_data = LEAF2KEY(mp,
5546 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5548 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5549 MDB_GET_KEY2(leaf, nodekey);
5551 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5553 /* current node was the one we wanted */
5563 /* If any parents have right-sibs, search.
5564 * Otherwise, there's nothing further.
5566 for (i=0; i<mc->mc_top; i++)
5568 NUMKEYS(mc->mc_pg[i])-1)
5570 if (i == mc->mc_top) {
5571 /* There are no other pages */
5572 mc->mc_ki[mc->mc_top] = nkeys;
5573 return MDB_NOTFOUND;
5577 /* There are no other pages */
5578 mc->mc_ki[mc->mc_top] = 0;
5579 if (op == MDB_SET_RANGE && !exactp) {
5583 return MDB_NOTFOUND;
5587 rc = mdb_page_search(mc, key, 0);
5588 if (rc != MDB_SUCCESS)
5591 mp = mc->mc_pg[mc->mc_top];
5592 mdb_cassert(mc, IS_LEAF(mp));
5595 leaf = mdb_node_search(mc, key, exactp);
5596 if (exactp != NULL && !*exactp) {
5597 /* MDB_SET specified and not an exact match. */
5598 return MDB_NOTFOUND;
5602 DPUTS("===> inexact leaf not found, goto sibling");
5603 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5604 return rc; /* no entries matched */
5605 mp = mc->mc_pg[mc->mc_top];
5606 mdb_cassert(mc, IS_LEAF(mp));
5607 leaf = NODEPTR(mp, 0);
5611 mc->mc_flags |= C_INITIALIZED;
5612 mc->mc_flags &= ~C_EOF;
5615 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5616 key->mv_size = mc->mc_db->md_pad;
5617 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5622 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5623 mdb_xcursor_init1(mc, leaf);
5626 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5627 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5628 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5631 if (op == MDB_GET_BOTH) {
5637 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5638 if (rc != MDB_SUCCESS)
5641 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5643 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5645 rc = mc->mc_dbx->md_dcmp(data, &d2);
5647 if (op == MDB_GET_BOTH || rc > 0)
5648 return MDB_NOTFOUND;
5655 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5656 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5661 /* The key already matches in all other cases */
5662 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5663 MDB_GET_KEY(leaf, key);
5664 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5669 /** Move the cursor to the first item in the database. */
5671 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5677 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5679 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5680 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5681 if (rc != MDB_SUCCESS)
5684 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5686 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5687 mc->mc_flags |= C_INITIALIZED;
5688 mc->mc_flags &= ~C_EOF;
5690 mc->mc_ki[mc->mc_top] = 0;
5692 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5693 key->mv_size = mc->mc_db->md_pad;
5694 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5699 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5700 mdb_xcursor_init1(mc, leaf);
5701 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5705 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5709 MDB_GET_KEY(leaf, key);
5713 /** Move the cursor to the last item in the database. */
5715 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5721 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5723 if (!(mc->mc_flags & C_EOF)) {
5725 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5726 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5727 if (rc != MDB_SUCCESS)
5730 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5733 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5734 mc->mc_flags |= C_INITIALIZED|C_EOF;
5735 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5737 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5738 key->mv_size = mc->mc_db->md_pad;
5739 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5744 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5745 mdb_xcursor_init1(mc, leaf);
5746 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5750 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5755 MDB_GET_KEY(leaf, key);
5760 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5765 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5770 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5774 case MDB_GET_CURRENT:
5775 if (!(mc->mc_flags & C_INITIALIZED)) {
5778 MDB_page *mp = mc->mc_pg[mc->mc_top];
5779 int nkeys = NUMKEYS(mp);
5780 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5781 mc->mc_ki[mc->mc_top] = nkeys;
5787 key->mv_size = mc->mc_db->md_pad;
5788 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5790 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5791 MDB_GET_KEY(leaf, key);
5793 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5794 if (mc->mc_flags & C_DEL)
5795 mdb_xcursor_init1(mc, leaf);
5796 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5798 rc = mdb_node_read(mc->mc_txn, leaf, data);
5805 case MDB_GET_BOTH_RANGE:
5810 if (mc->mc_xcursor == NULL) {
5811 rc = MDB_INCOMPATIBLE;
5821 rc = mdb_cursor_set(mc, key, data, op,
5822 op == MDB_SET_RANGE ? NULL : &exact);
5825 case MDB_GET_MULTIPLE:
5826 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5830 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5831 rc = MDB_INCOMPATIBLE;
5835 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5836 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5839 case MDB_NEXT_MULTIPLE:
5844 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5845 rc = MDB_INCOMPATIBLE;
5848 if (!(mc->mc_flags & C_INITIALIZED))
5849 rc = mdb_cursor_first(mc, key, data);
5851 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5852 if (rc == MDB_SUCCESS) {
5853 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5856 mx = &mc->mc_xcursor->mx_cursor;
5857 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5859 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5860 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5868 case MDB_NEXT_NODUP:
5869 if (!(mc->mc_flags & C_INITIALIZED))
5870 rc = mdb_cursor_first(mc, key, data);
5872 rc = mdb_cursor_next(mc, key, data, op);
5876 case MDB_PREV_NODUP:
5877 if (!(mc->mc_flags & C_INITIALIZED)) {
5878 rc = mdb_cursor_last(mc, key, data);
5881 mc->mc_flags |= C_INITIALIZED;
5882 mc->mc_ki[mc->mc_top]++;
5884 rc = mdb_cursor_prev(mc, key, data, op);
5887 rc = mdb_cursor_first(mc, key, data);
5890 mfunc = mdb_cursor_first;
5892 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5896 if (mc->mc_xcursor == NULL) {
5897 rc = MDB_INCOMPATIBLE;
5901 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5902 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5903 MDB_GET_KEY(leaf, key);
5904 rc = mdb_node_read(mc->mc_txn, leaf, data);
5908 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5912 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5915 rc = mdb_cursor_last(mc, key, data);
5918 mfunc = mdb_cursor_last;
5921 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5926 if (mc->mc_flags & C_DEL)
5927 mc->mc_flags ^= C_DEL;
5932 /** Touch all the pages in the cursor stack. Set mc_top.
5933 * Makes sure all the pages are writable, before attempting a write operation.
5934 * @param[in] mc The cursor to operate on.
5937 mdb_cursor_touch(MDB_cursor *mc)
5939 int rc = MDB_SUCCESS;
5941 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5944 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5946 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5947 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5950 *mc->mc_dbflag |= DB_DIRTY;
5955 rc = mdb_page_touch(mc);
5956 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5957 mc->mc_top = mc->mc_snum-1;
5962 /** Do not spill pages to disk if txn is getting full, may fail instead */
5963 #define MDB_NOSPILL 0x8000
5966 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5969 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5971 MDB_node *leaf = NULL;
5974 MDB_val xdata, *rdata, dkey, olddata;
5976 int do_sub = 0, insert_key, insert_data;
5977 unsigned int mcount = 0, dcount = 0, nospill;
5980 unsigned int nflags;
5983 if (mc == NULL || key == NULL)
5986 env = mc->mc_txn->mt_env;
5988 /* Check this first so counter will always be zero on any
5991 if (flags & MDB_MULTIPLE) {
5992 dcount = data[1].mv_size;
5993 data[1].mv_size = 0;
5994 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5995 return MDB_INCOMPATIBLE;
5998 nospill = flags & MDB_NOSPILL;
5999 flags &= ~MDB_NOSPILL;
6001 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6002 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6004 if (key->mv_size-1 >= ENV_MAXKEY(env))
6005 return MDB_BAD_VALSIZE;
6007 #if SIZE_MAX > MAXDATASIZE
6008 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6009 return MDB_BAD_VALSIZE;
6011 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6012 return MDB_BAD_VALSIZE;
6015 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6016 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6020 if (flags == MDB_CURRENT) {
6021 if (!(mc->mc_flags & C_INITIALIZED))
6024 } else if (mc->mc_db->md_root == P_INVALID) {
6025 /* new database, cursor has nothing to point to */
6028 mc->mc_flags &= ~C_INITIALIZED;
6033 if (flags & MDB_APPEND) {
6035 rc = mdb_cursor_last(mc, &k2, &d2);
6037 rc = mc->mc_dbx->md_cmp(key, &k2);
6040 mc->mc_ki[mc->mc_top]++;
6042 /* new key is <= last key */
6047 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6049 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6050 DPRINTF(("duplicate key [%s]", DKEY(key)));
6052 return MDB_KEYEXIST;
6054 if (rc && rc != MDB_NOTFOUND)
6058 if (mc->mc_flags & C_DEL)
6059 mc->mc_flags ^= C_DEL;
6061 /* Cursor is positioned, check for room in the dirty list */
6063 if (flags & MDB_MULTIPLE) {
6065 xdata.mv_size = data->mv_size * dcount;
6069 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6073 if (rc == MDB_NO_ROOT) {
6075 /* new database, write a root leaf page */
6076 DPUTS("allocating new root leaf page");
6077 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6080 mdb_cursor_push(mc, np);
6081 mc->mc_db->md_root = np->mp_pgno;
6082 mc->mc_db->md_depth++;
6083 *mc->mc_dbflag |= DB_DIRTY;
6084 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6086 np->mp_flags |= P_LEAF2;
6087 mc->mc_flags |= C_INITIALIZED;
6089 /* make sure all cursor pages are writable */
6090 rc2 = mdb_cursor_touch(mc);
6095 insert_key = insert_data = rc;
6097 /* The key does not exist */
6098 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6099 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6100 LEAFSIZE(key, data) > env->me_nodemax)
6102 /* Too big for a node, insert in sub-DB. Set up an empty
6103 * "old sub-page" for prep_subDB to expand to a full page.
6105 fp_flags = P_LEAF|P_DIRTY;
6107 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6108 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6109 olddata.mv_size = PAGEHDRSZ;
6113 /* there's only a key anyway, so this is a no-op */
6114 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6116 unsigned int ksize = mc->mc_db->md_pad;
6117 if (key->mv_size != ksize)
6118 return MDB_BAD_VALSIZE;
6119 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6120 memcpy(ptr, key->mv_data, ksize);
6122 /* if overwriting slot 0 of leaf, need to
6123 * update branch key if there is a parent page
6125 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6126 unsigned short top = mc->mc_top;
6128 /* slot 0 is always an empty key, find real slot */
6129 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6131 if (mc->mc_ki[mc->mc_top])
6132 rc2 = mdb_update_key(mc, key);
6143 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6144 olddata.mv_size = NODEDSZ(leaf);
6145 olddata.mv_data = NODEDATA(leaf);
6148 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6149 /* Prepare (sub-)page/sub-DB to accept the new item,
6150 * if needed. fp: old sub-page or a header faking
6151 * it. mp: new (sub-)page. offset: growth in page
6152 * size. xdata: node data with new page or DB.
6154 unsigned i, offset = 0;
6155 mp = fp = xdata.mv_data = env->me_pbuf;
6156 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6158 /* Was a single item before, must convert now */
6159 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6160 /* Just overwrite the current item */
6161 if (flags == MDB_CURRENT)
6164 #if UINT_MAX < SIZE_MAX
6165 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6166 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6168 /* does data match? */
6169 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6170 if (flags & MDB_NODUPDATA)
6171 return MDB_KEYEXIST;
6176 /* Back up original data item */
6177 dkey.mv_size = olddata.mv_size;
6178 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6180 /* Make sub-page header for the dup items, with dummy body */
6181 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6182 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6183 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6184 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6185 fp->mp_flags |= P_LEAF2;
6186 fp->mp_pad = data->mv_size;
6187 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6189 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6190 (dkey.mv_size & 1) + (data->mv_size & 1);
6192 fp->mp_upper = xdata.mv_size - PAGEBASE;
6193 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6194 } else if (leaf->mn_flags & F_SUBDATA) {
6195 /* Data is on sub-DB, just store it */
6196 flags |= F_DUPDATA|F_SUBDATA;
6199 /* Data is on sub-page */
6200 fp = olddata.mv_data;
6203 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6204 offset = EVEN(NODESIZE + sizeof(indx_t) +
6208 offset = fp->mp_pad;
6209 if (SIZELEFT(fp) < offset) {
6210 offset *= 4; /* space for 4 more */
6213 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6215 fp->mp_flags |= P_DIRTY;
6216 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6217 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6221 xdata.mv_size = olddata.mv_size + offset;
6224 fp_flags = fp->mp_flags;
6225 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6226 /* Too big for a sub-page, convert to sub-DB */
6227 fp_flags &= ~P_SUBP;
6229 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6230 fp_flags |= P_LEAF2;
6231 dummy.md_pad = fp->mp_pad;
6232 dummy.md_flags = MDB_DUPFIXED;
6233 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6234 dummy.md_flags |= MDB_INTEGERKEY;
6240 dummy.md_branch_pages = 0;
6241 dummy.md_leaf_pages = 1;
6242 dummy.md_overflow_pages = 0;
6243 dummy.md_entries = NUMKEYS(fp);
6244 xdata.mv_size = sizeof(MDB_db);
6245 xdata.mv_data = &dummy;
6246 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6248 offset = env->me_psize - olddata.mv_size;
6249 flags |= F_DUPDATA|F_SUBDATA;
6250 dummy.md_root = mp->mp_pgno;
6253 mp->mp_flags = fp_flags | P_DIRTY;
6254 mp->mp_pad = fp->mp_pad;
6255 mp->mp_lower = fp->mp_lower;
6256 mp->mp_upper = fp->mp_upper + offset;
6257 if (fp_flags & P_LEAF2) {
6258 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6260 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6261 olddata.mv_size - fp->mp_upper - PAGEBASE);
6262 for (i=0; i<NUMKEYS(fp); i++)
6263 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6271 mdb_node_del(mc, 0);
6275 /* overflow page overwrites need special handling */
6276 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6279 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6281 memcpy(&pg, olddata.mv_data, sizeof(pg));
6282 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6284 ovpages = omp->mp_pages;
6286 /* Is the ov page large enough? */
6287 if (ovpages >= dpages) {
6288 if (!(omp->mp_flags & P_DIRTY) &&
6289 (level || (env->me_flags & MDB_WRITEMAP)))
6291 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6294 level = 0; /* dirty in this txn or clean */
6297 if (omp->mp_flags & P_DIRTY) {
6298 /* yes, overwrite it. Note in this case we don't
6299 * bother to try shrinking the page if the new data
6300 * is smaller than the overflow threshold.
6303 /* It is writable only in a parent txn */
6304 size_t sz = (size_t) env->me_psize * ovpages, off;
6305 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6311 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6312 mdb_cassert(mc, rc2 == 0);
6313 if (!(flags & MDB_RESERVE)) {
6314 /* Copy end of page, adjusting alignment so
6315 * compiler may copy words instead of bytes.
6317 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6318 memcpy((size_t *)((char *)np + off),
6319 (size_t *)((char *)omp + off), sz - off);
6322 memcpy(np, omp, sz); /* Copy beginning of page */
6325 SETDSZ(leaf, data->mv_size);
6326 if (F_ISSET(flags, MDB_RESERVE))
6327 data->mv_data = METADATA(omp);
6329 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6333 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6335 } else if (data->mv_size == olddata.mv_size) {
6336 /* same size, just replace it. Note that we could
6337 * also reuse this node if the new data is smaller,
6338 * but instead we opt to shrink the node in that case.
6340 if (F_ISSET(flags, MDB_RESERVE))
6341 data->mv_data = olddata.mv_data;
6342 else if (!(mc->mc_flags & C_SUB))
6343 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6345 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6350 mdb_node_del(mc, 0);
6356 nflags = flags & NODE_ADD_FLAGS;
6357 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6358 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6359 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6360 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6362 nflags |= MDB_SPLIT_REPLACE;
6363 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6365 /* There is room already in this leaf page. */
6366 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6367 if (rc == 0 && insert_key) {
6368 /* Adjust other cursors pointing to mp */
6369 MDB_cursor *m2, *m3;
6370 MDB_dbi dbi = mc->mc_dbi;
6371 unsigned i = mc->mc_top;
6372 MDB_page *mp = mc->mc_pg[i];
6374 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6375 if (mc->mc_flags & C_SUB)
6376 m3 = &m2->mc_xcursor->mx_cursor;
6379 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6380 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6387 if (rc == MDB_SUCCESS) {
6388 /* Now store the actual data in the child DB. Note that we're
6389 * storing the user data in the keys field, so there are strict
6390 * size limits on dupdata. The actual data fields of the child
6391 * DB are all zero size.
6399 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6400 if (flags & MDB_CURRENT) {
6401 xflags = MDB_CURRENT|MDB_NOSPILL;
6403 mdb_xcursor_init1(mc, leaf);
6404 xflags = (flags & MDB_NODUPDATA) ?
6405 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6407 /* converted, write the original data first */
6409 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6413 /* Adjust other cursors pointing to mp */
6415 unsigned i = mc->mc_top;
6416 MDB_page *mp = mc->mc_pg[i];
6418 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6419 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6420 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6421 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6422 mdb_xcursor_init1(m2, leaf);
6426 /* we've done our job */
6429 ecount = mc->mc_xcursor->mx_db.md_entries;
6430 if (flags & MDB_APPENDDUP)
6431 xflags |= MDB_APPEND;
6432 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6433 if (flags & F_SUBDATA) {
6434 void *db = NODEDATA(leaf);
6435 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6437 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6439 /* Increment count unless we just replaced an existing item. */
6441 mc->mc_db->md_entries++;
6443 /* Invalidate txn if we created an empty sub-DB */
6446 /* If we succeeded and the key didn't exist before,
6447 * make sure the cursor is marked valid.
6449 mc->mc_flags |= C_INITIALIZED;
6451 if (flags & MDB_MULTIPLE) {
6454 /* let caller know how many succeeded, if any */
6455 data[1].mv_size = mcount;
6456 if (mcount < dcount) {
6457 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6458 insert_key = insert_data = 0;
6465 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6468 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6473 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6479 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6480 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6482 if (!(mc->mc_flags & C_INITIALIZED))
6485 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6486 return MDB_NOTFOUND;
6488 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6491 rc = mdb_cursor_touch(mc);
6495 mp = mc->mc_pg[mc->mc_top];
6498 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6500 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6501 if (flags & MDB_NODUPDATA) {
6502 /* mdb_cursor_del0() will subtract the final entry */
6503 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6505 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6506 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6508 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6511 /* If sub-DB still has entries, we're done */
6512 if (mc->mc_xcursor->mx_db.md_entries) {
6513 if (leaf->mn_flags & F_SUBDATA) {
6514 /* update subDB info */
6515 void *db = NODEDATA(leaf);
6516 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6519 /* shrink fake page */
6520 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6521 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6522 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6523 /* fix other sub-DB cursors pointed at this fake page */
6524 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6525 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6526 if (m2->mc_pg[mc->mc_top] == mp &&
6527 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6528 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6531 mc->mc_db->md_entries--;
6532 mc->mc_flags |= C_DEL;
6535 /* otherwise fall thru and delete the sub-DB */
6538 if (leaf->mn_flags & F_SUBDATA) {
6539 /* add all the child DB's pages to the free list */
6540 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6546 /* add overflow pages to free list */
6547 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6551 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6552 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6553 (rc = mdb_ovpage_free(mc, omp)))
6558 return mdb_cursor_del0(mc);
6561 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6565 /** Allocate and initialize new pages for a database.
6566 * @param[in] mc a cursor on the database being added to.
6567 * @param[in] flags flags defining what type of page is being allocated.
6568 * @param[in] num the number of pages to allocate. This is usually 1,
6569 * unless allocating overflow pages for a large record.
6570 * @param[out] mp Address of a page, or NULL on failure.
6571 * @return 0 on success, non-zero on failure.
6574 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6579 if ((rc = mdb_page_alloc(mc, num, &np)))
6581 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6582 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6583 np->mp_flags = flags | P_DIRTY;
6584 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6585 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6588 mc->mc_db->md_branch_pages++;
6589 else if (IS_LEAF(np))
6590 mc->mc_db->md_leaf_pages++;
6591 else if (IS_OVERFLOW(np)) {
6592 mc->mc_db->md_overflow_pages += num;
6600 /** Calculate the size of a leaf node.
6601 * The size depends on the environment's page size; if a data item
6602 * is too large it will be put onto an overflow page and the node
6603 * size will only include the key and not the data. Sizes are always
6604 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6605 * of the #MDB_node headers.
6606 * @param[in] env The environment handle.
6607 * @param[in] key The key for the node.
6608 * @param[in] data The data for the node.
6609 * @return The number of bytes needed to store the node.
6612 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6616 sz = LEAFSIZE(key, data);
6617 if (sz > env->me_nodemax) {
6618 /* put on overflow page */
6619 sz -= data->mv_size - sizeof(pgno_t);
6622 return EVEN(sz + sizeof(indx_t));
6625 /** Calculate the size of a branch node.
6626 * The size should depend on the environment's page size but since
6627 * we currently don't support spilling large keys onto overflow
6628 * pages, it's simply the size of the #MDB_node header plus the
6629 * size of the key. Sizes are always rounded up to an even number
6630 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6631 * @param[in] env The environment handle.
6632 * @param[in] key The key for the node.
6633 * @return The number of bytes needed to store the node.
6636 mdb_branch_size(MDB_env *env, MDB_val *key)
6641 if (sz > env->me_nodemax) {
6642 /* put on overflow page */
6643 /* not implemented */
6644 /* sz -= key->size - sizeof(pgno_t); */
6647 return sz + sizeof(indx_t);
6650 /** Add a node to the page pointed to by the cursor.
6651 * @param[in] mc The cursor for this operation.
6652 * @param[in] indx The index on the page where the new node should be added.
6653 * @param[in] key The key for the new node.
6654 * @param[in] data The data for the new node, if any.
6655 * @param[in] pgno The page number, if adding a branch node.
6656 * @param[in] flags Flags for the node.
6657 * @return 0 on success, non-zero on failure. Possible errors are:
6659 * <li>ENOMEM - failed to allocate overflow pages for the node.
6660 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6661 * should never happen since all callers already calculate the
6662 * page's free space before calling this function.
6666 mdb_node_add(MDB_cursor *mc, indx_t indx,
6667 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6670 size_t node_size = NODESIZE;
6674 MDB_page *mp = mc->mc_pg[mc->mc_top];
6675 MDB_page *ofp = NULL; /* overflow page */
6678 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6680 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6681 IS_LEAF(mp) ? "leaf" : "branch",
6682 IS_SUBP(mp) ? "sub-" : "",
6683 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6684 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6687 /* Move higher keys up one slot. */
6688 int ksize = mc->mc_db->md_pad, dif;
6689 char *ptr = LEAF2KEY(mp, indx, ksize);
6690 dif = NUMKEYS(mp) - indx;
6692 memmove(ptr+ksize, ptr, dif*ksize);
6693 /* insert new key */
6694 memcpy(ptr, key->mv_data, ksize);
6696 /* Just using these for counting */
6697 mp->mp_lower += sizeof(indx_t);
6698 mp->mp_upper -= ksize - sizeof(indx_t);
6702 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6704 node_size += key->mv_size;
6706 mdb_cassert(mc, data);
6707 if (F_ISSET(flags, F_BIGDATA)) {
6708 /* Data already on overflow page. */
6709 node_size += sizeof(pgno_t);
6710 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6711 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6713 /* Put data on overflow page. */
6714 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6715 data->mv_size, node_size+data->mv_size));
6716 node_size = EVEN(node_size + sizeof(pgno_t));
6717 if ((ssize_t)node_size > room)
6719 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6721 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6725 node_size += data->mv_size;
6728 node_size = EVEN(node_size);
6729 if ((ssize_t)node_size > room)
6733 /* Move higher pointers up one slot. */
6734 for (i = NUMKEYS(mp); i > indx; i--)
6735 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6737 /* Adjust free space offsets. */
6738 ofs = mp->mp_upper - node_size;
6739 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6740 mp->mp_ptrs[indx] = ofs;
6742 mp->mp_lower += sizeof(indx_t);
6744 /* Write the node data. */
6745 node = NODEPTR(mp, indx);
6746 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6747 node->mn_flags = flags;
6749 SETDSZ(node,data->mv_size);
6754 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6757 mdb_cassert(mc, key);
6759 if (F_ISSET(flags, F_BIGDATA))
6760 memcpy(node->mn_data + key->mv_size, data->mv_data,
6762 else if (F_ISSET(flags, MDB_RESERVE))
6763 data->mv_data = node->mn_data + key->mv_size;
6765 memcpy(node->mn_data + key->mv_size, data->mv_data,
6768 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6770 if (F_ISSET(flags, MDB_RESERVE))
6771 data->mv_data = METADATA(ofp);
6773 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6780 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6781 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6782 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6783 DPRINTF(("node size = %"Z"u", node_size));
6784 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6785 return MDB_PAGE_FULL;
6788 /** Delete the specified node from a page.
6789 * @param[in] mc Cursor pointing to the node to delete.
6790 * @param[in] ksize The size of a node. Only used if the page is
6791 * part of a #MDB_DUPFIXED database.
6794 mdb_node_del(MDB_cursor *mc, int ksize)
6796 MDB_page *mp = mc->mc_pg[mc->mc_top];
6797 indx_t indx = mc->mc_ki[mc->mc_top];
6799 indx_t i, j, numkeys, ptr;
6803 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6804 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6805 numkeys = NUMKEYS(mp);
6806 mdb_cassert(mc, indx < numkeys);
6809 int x = numkeys - 1 - indx;
6810 base = LEAF2KEY(mp, indx, ksize);
6812 memmove(base, base + ksize, x * ksize);
6813 mp->mp_lower -= sizeof(indx_t);
6814 mp->mp_upper += ksize - sizeof(indx_t);
6818 node = NODEPTR(mp, indx);
6819 sz = NODESIZE + node->mn_ksize;
6821 if (F_ISSET(node->mn_flags, F_BIGDATA))
6822 sz += sizeof(pgno_t);
6824 sz += NODEDSZ(node);
6828 ptr = mp->mp_ptrs[indx];
6829 for (i = j = 0; i < numkeys; i++) {
6831 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6832 if (mp->mp_ptrs[i] < ptr)
6833 mp->mp_ptrs[j] += sz;
6838 base = (char *)mp + mp->mp_upper + PAGEBASE;
6839 memmove(base + sz, base, ptr - mp->mp_upper);
6841 mp->mp_lower -= sizeof(indx_t);
6845 /** Compact the main page after deleting a node on a subpage.
6846 * @param[in] mp The main page to operate on.
6847 * @param[in] indx The index of the subpage on the main page.
6850 mdb_node_shrink(MDB_page *mp, indx_t indx)
6856 indx_t i, numkeys, ptr;
6858 node = NODEPTR(mp, indx);
6859 sp = (MDB_page *)NODEDATA(node);
6860 delta = SIZELEFT(sp);
6861 xp = (MDB_page *)((char *)sp + delta);
6863 /* shift subpage upward */
6865 nsize = NUMKEYS(sp) * sp->mp_pad;
6867 return; /* do not make the node uneven-sized */
6868 memmove(METADATA(xp), METADATA(sp), nsize);
6871 numkeys = NUMKEYS(sp);
6872 for (i=numkeys-1; i>=0; i--)
6873 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6875 xp->mp_upper = sp->mp_lower;
6876 xp->mp_lower = sp->mp_lower;
6877 xp->mp_flags = sp->mp_flags;
6878 xp->mp_pad = sp->mp_pad;
6879 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6881 nsize = NODEDSZ(node) - delta;
6882 SETDSZ(node, nsize);
6884 /* shift lower nodes upward */
6885 ptr = mp->mp_ptrs[indx];
6886 numkeys = NUMKEYS(mp);
6887 for (i = 0; i < numkeys; i++) {
6888 if (mp->mp_ptrs[i] <= ptr)
6889 mp->mp_ptrs[i] += delta;
6892 base = (char *)mp + mp->mp_upper + PAGEBASE;
6893 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6894 mp->mp_upper += delta;
6897 /** Initial setup of a sorted-dups cursor.
6898 * Sorted duplicates are implemented as a sub-database for the given key.
6899 * The duplicate data items are actually keys of the sub-database.
6900 * Operations on the duplicate data items are performed using a sub-cursor
6901 * initialized when the sub-database is first accessed. This function does
6902 * the preliminary setup of the sub-cursor, filling in the fields that
6903 * depend only on the parent DB.
6904 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6907 mdb_xcursor_init0(MDB_cursor *mc)
6909 MDB_xcursor *mx = mc->mc_xcursor;
6911 mx->mx_cursor.mc_xcursor = NULL;
6912 mx->mx_cursor.mc_txn = mc->mc_txn;
6913 mx->mx_cursor.mc_db = &mx->mx_db;
6914 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6915 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6916 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6917 mx->mx_cursor.mc_snum = 0;
6918 mx->mx_cursor.mc_top = 0;
6919 mx->mx_cursor.mc_flags = C_SUB;
6920 mx->mx_dbx.md_name.mv_size = 0;
6921 mx->mx_dbx.md_name.mv_data = NULL;
6922 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6923 mx->mx_dbx.md_dcmp = NULL;
6924 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6927 /** Final setup of a sorted-dups cursor.
6928 * Sets up the fields that depend on the data from the main cursor.
6929 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6930 * @param[in] node The data containing the #MDB_db record for the
6931 * sorted-dup database.
6934 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6936 MDB_xcursor *mx = mc->mc_xcursor;
6938 if (node->mn_flags & F_SUBDATA) {
6939 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6940 mx->mx_cursor.mc_pg[0] = 0;
6941 mx->mx_cursor.mc_snum = 0;
6942 mx->mx_cursor.mc_top = 0;
6943 mx->mx_cursor.mc_flags = C_SUB;
6945 MDB_page *fp = NODEDATA(node);
6946 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6947 mx->mx_db.md_flags = 0;
6948 mx->mx_db.md_depth = 1;
6949 mx->mx_db.md_branch_pages = 0;
6950 mx->mx_db.md_leaf_pages = 1;
6951 mx->mx_db.md_overflow_pages = 0;
6952 mx->mx_db.md_entries = NUMKEYS(fp);
6953 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6954 mx->mx_cursor.mc_snum = 1;
6955 mx->mx_cursor.mc_top = 0;
6956 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6957 mx->mx_cursor.mc_pg[0] = fp;
6958 mx->mx_cursor.mc_ki[0] = 0;
6959 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6960 mx->mx_db.md_flags = MDB_DUPFIXED;
6961 mx->mx_db.md_pad = fp->mp_pad;
6962 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6963 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6966 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6967 mx->mx_db.md_root));
6968 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6969 #if UINT_MAX < SIZE_MAX
6970 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6971 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6975 /** Initialize a cursor for a given transaction and database. */
6977 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6980 mc->mc_backup = NULL;
6983 mc->mc_db = &txn->mt_dbs[dbi];
6984 mc->mc_dbx = &txn->mt_dbxs[dbi];
6985 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6990 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6991 mdb_tassert(txn, mx != NULL);
6992 mc->mc_xcursor = mx;
6993 mdb_xcursor_init0(mc);
6995 mc->mc_xcursor = NULL;
6997 if (*mc->mc_dbflag & DB_STALE) {
6998 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7003 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7006 size_t size = sizeof(MDB_cursor);
7008 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7011 if (txn->mt_flags & MDB_TXN_ERROR)
7014 /* Allow read access to the freelist */
7015 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7018 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7019 size += sizeof(MDB_xcursor);
7021 if ((mc = malloc(size)) != NULL) {
7022 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7023 if (txn->mt_cursors) {
7024 mc->mc_next = txn->mt_cursors[dbi];
7025 txn->mt_cursors[dbi] = mc;
7026 mc->mc_flags |= C_UNTRACK;
7038 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7040 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7043 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7046 if (txn->mt_flags & MDB_TXN_ERROR)
7049 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7053 /* Return the count of duplicate data items for the current key */
7055 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7059 if (mc == NULL || countp == NULL)
7062 if (mc->mc_xcursor == NULL)
7063 return MDB_INCOMPATIBLE;
7065 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7068 if (!(mc->mc_flags & C_INITIALIZED))
7071 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7072 return MDB_NOTFOUND;
7074 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7075 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7078 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7081 *countp = mc->mc_xcursor->mx_db.md_entries;
7087 mdb_cursor_close(MDB_cursor *mc)
7089 if (mc && !mc->mc_backup) {
7090 /* remove from txn, if tracked */
7091 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7092 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7093 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7095 *prev = mc->mc_next;
7102 mdb_cursor_txn(MDB_cursor *mc)
7104 if (!mc) return NULL;
7109 mdb_cursor_dbi(MDB_cursor *mc)
7114 /** Replace the key for a branch node with a new key.
7115 * @param[in] mc Cursor pointing to the node to operate on.
7116 * @param[in] key The new key to use.
7117 * @return 0 on success, non-zero on failure.
7120 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7126 int delta, ksize, oksize;
7127 indx_t ptr, i, numkeys, indx;
7130 indx = mc->mc_ki[mc->mc_top];
7131 mp = mc->mc_pg[mc->mc_top];
7132 node = NODEPTR(mp, indx);
7133 ptr = mp->mp_ptrs[indx];
7137 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7138 k2.mv_data = NODEKEY(node);
7139 k2.mv_size = node->mn_ksize;
7140 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7142 mdb_dkey(&k2, kbuf2),
7148 /* Sizes must be 2-byte aligned. */
7149 ksize = EVEN(key->mv_size);
7150 oksize = EVEN(node->mn_ksize);
7151 delta = ksize - oksize;
7153 /* Shift node contents if EVEN(key length) changed. */
7155 if (delta > 0 && SIZELEFT(mp) < delta) {
7157 /* not enough space left, do a delete and split */
7158 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7159 pgno = NODEPGNO(node);
7160 mdb_node_del(mc, 0);
7161 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7164 numkeys = NUMKEYS(mp);
7165 for (i = 0; i < numkeys; i++) {
7166 if (mp->mp_ptrs[i] <= ptr)
7167 mp->mp_ptrs[i] -= delta;
7170 base = (char *)mp + mp->mp_upper + PAGEBASE;
7171 len = ptr - mp->mp_upper + NODESIZE;
7172 memmove(base - delta, base, len);
7173 mp->mp_upper -= delta;
7175 node = NODEPTR(mp, indx);
7178 /* But even if no shift was needed, update ksize */
7179 if (node->mn_ksize != key->mv_size)
7180 node->mn_ksize = key->mv_size;
7183 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7189 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7191 /** Move a node from csrc to cdst.
7194 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7201 unsigned short flags;
7205 /* Mark src and dst as dirty. */
7206 if ((rc = mdb_page_touch(csrc)) ||
7207 (rc = mdb_page_touch(cdst)))
7210 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7211 key.mv_size = csrc->mc_db->md_pad;
7212 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7214 data.mv_data = NULL;
7218 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7219 mdb_cassert(csrc, !((size_t)srcnode & 1));
7220 srcpg = NODEPGNO(srcnode);
7221 flags = srcnode->mn_flags;
7222 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7223 unsigned int snum = csrc->mc_snum;
7225 /* must find the lowest key below src */
7226 rc = mdb_page_search_lowest(csrc);
7229 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7230 key.mv_size = csrc->mc_db->md_pad;
7231 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7233 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7234 key.mv_size = NODEKSZ(s2);
7235 key.mv_data = NODEKEY(s2);
7237 csrc->mc_snum = snum--;
7238 csrc->mc_top = snum;
7240 key.mv_size = NODEKSZ(srcnode);
7241 key.mv_data = NODEKEY(srcnode);
7243 data.mv_size = NODEDSZ(srcnode);
7244 data.mv_data = NODEDATA(srcnode);
7246 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7247 unsigned int snum = cdst->mc_snum;
7250 /* must find the lowest key below dst */
7251 mdb_cursor_copy(cdst, &mn);
7252 rc = mdb_page_search_lowest(&mn);
7255 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7256 bkey.mv_size = mn.mc_db->md_pad;
7257 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7259 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7260 bkey.mv_size = NODEKSZ(s2);
7261 bkey.mv_data = NODEKEY(s2);
7263 mn.mc_snum = snum--;
7266 rc = mdb_update_key(&mn, &bkey);
7271 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7272 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7273 csrc->mc_ki[csrc->mc_top],
7275 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7276 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7278 /* Add the node to the destination page.
7280 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7281 if (rc != MDB_SUCCESS)
7284 /* Delete the node from the source page.
7286 mdb_node_del(csrc, key.mv_size);
7289 /* Adjust other cursors pointing to mp */
7290 MDB_cursor *m2, *m3;
7291 MDB_dbi dbi = csrc->mc_dbi;
7292 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7294 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7295 if (csrc->mc_flags & C_SUB)
7296 m3 = &m2->mc_xcursor->mx_cursor;
7299 if (m3 == csrc) continue;
7300 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7301 csrc->mc_ki[csrc->mc_top]) {
7302 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7303 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7308 /* Update the parent separators.
7310 if (csrc->mc_ki[csrc->mc_top] == 0) {
7311 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7312 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7313 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7315 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7316 key.mv_size = NODEKSZ(srcnode);
7317 key.mv_data = NODEKEY(srcnode);
7319 DPRINTF(("update separator for source page %"Z"u to [%s]",
7320 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7321 mdb_cursor_copy(csrc, &mn);
7324 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7327 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7329 indx_t ix = csrc->mc_ki[csrc->mc_top];
7330 nullkey.mv_size = 0;
7331 csrc->mc_ki[csrc->mc_top] = 0;
7332 rc = mdb_update_key(csrc, &nullkey);
7333 csrc->mc_ki[csrc->mc_top] = ix;
7334 mdb_cassert(csrc, rc == MDB_SUCCESS);
7338 if (cdst->mc_ki[cdst->mc_top] == 0) {
7339 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7340 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7341 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7343 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7344 key.mv_size = NODEKSZ(srcnode);
7345 key.mv_data = NODEKEY(srcnode);
7347 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7348 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7349 mdb_cursor_copy(cdst, &mn);
7352 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7355 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7357 indx_t ix = cdst->mc_ki[cdst->mc_top];
7358 nullkey.mv_size = 0;
7359 cdst->mc_ki[cdst->mc_top] = 0;
7360 rc = mdb_update_key(cdst, &nullkey);
7361 cdst->mc_ki[cdst->mc_top] = ix;
7362 mdb_cassert(csrc, rc == MDB_SUCCESS);
7369 /** Merge one page into another.
7370 * The nodes from the page pointed to by \b csrc will
7371 * be copied to the page pointed to by \b cdst and then
7372 * the \b csrc page will be freed.
7373 * @param[in] csrc Cursor pointing to the source page.
7374 * @param[in] cdst Cursor pointing to the destination page.
7375 * @return 0 on success, non-zero on failure.
7378 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7380 MDB_page *psrc, *pdst;
7387 psrc = csrc->mc_pg[csrc->mc_top];
7388 pdst = cdst->mc_pg[cdst->mc_top];
7390 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7392 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7393 mdb_cassert(csrc, cdst->mc_snum > 1);
7395 /* Mark dst as dirty. */
7396 if ((rc = mdb_page_touch(cdst)))
7399 /* Move all nodes from src to dst.
7401 j = nkeys = NUMKEYS(pdst);
7402 if (IS_LEAF2(psrc)) {
7403 key.mv_size = csrc->mc_db->md_pad;
7404 key.mv_data = METADATA(psrc);
7405 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7406 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7407 if (rc != MDB_SUCCESS)
7409 key.mv_data = (char *)key.mv_data + key.mv_size;
7412 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7413 srcnode = NODEPTR(psrc, i);
7414 if (i == 0 && IS_BRANCH(psrc)) {
7417 mdb_cursor_copy(csrc, &mn);
7418 /* must find the lowest key below src */
7419 rc = mdb_page_search_lowest(&mn);
7422 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7423 key.mv_size = mn.mc_db->md_pad;
7424 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7426 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7427 key.mv_size = NODEKSZ(s2);
7428 key.mv_data = NODEKEY(s2);
7431 key.mv_size = srcnode->mn_ksize;
7432 key.mv_data = NODEKEY(srcnode);
7435 data.mv_size = NODEDSZ(srcnode);
7436 data.mv_data = NODEDATA(srcnode);
7437 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7438 if (rc != MDB_SUCCESS)
7443 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7444 pdst->mp_pgno, NUMKEYS(pdst),
7445 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7447 /* Unlink the src page from parent and add to free list.
7450 mdb_node_del(csrc, 0);
7451 if (csrc->mc_ki[csrc->mc_top] == 0) {
7453 rc = mdb_update_key(csrc, &key);
7461 psrc = csrc->mc_pg[csrc->mc_top];
7462 /* If not operating on FreeDB, allow this page to be reused
7463 * in this txn. Otherwise just add to free list.
7465 rc = mdb_page_loose(csrc, psrc);
7469 csrc->mc_db->md_leaf_pages--;
7471 csrc->mc_db->md_branch_pages--;
7473 /* Adjust other cursors pointing to mp */
7474 MDB_cursor *m2, *m3;
7475 MDB_dbi dbi = csrc->mc_dbi;
7477 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7478 if (csrc->mc_flags & C_SUB)
7479 m3 = &m2->mc_xcursor->mx_cursor;
7482 if (m3 == csrc) continue;
7483 if (m3->mc_snum < csrc->mc_snum) continue;
7484 if (m3->mc_pg[csrc->mc_top] == psrc) {
7485 m3->mc_pg[csrc->mc_top] = pdst;
7486 m3->mc_ki[csrc->mc_top] += nkeys;
7491 unsigned int snum = cdst->mc_snum;
7492 uint16_t depth = cdst->mc_db->md_depth;
7493 mdb_cursor_pop(cdst);
7494 rc = mdb_rebalance(cdst);
7495 /* Did the tree shrink? */
7496 if (depth > cdst->mc_db->md_depth)
7498 cdst->mc_snum = snum;
7499 cdst->mc_top = snum-1;
7504 /** Copy the contents of a cursor.
7505 * @param[in] csrc The cursor to copy from.
7506 * @param[out] cdst The cursor to copy to.
7509 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7513 cdst->mc_txn = csrc->mc_txn;
7514 cdst->mc_dbi = csrc->mc_dbi;
7515 cdst->mc_db = csrc->mc_db;
7516 cdst->mc_dbx = csrc->mc_dbx;
7517 cdst->mc_snum = csrc->mc_snum;
7518 cdst->mc_top = csrc->mc_top;
7519 cdst->mc_flags = csrc->mc_flags;
7521 for (i=0; i<csrc->mc_snum; i++) {
7522 cdst->mc_pg[i] = csrc->mc_pg[i];
7523 cdst->mc_ki[i] = csrc->mc_ki[i];
7527 /** Rebalance the tree after a delete operation.
7528 * @param[in] mc Cursor pointing to the page where rebalancing
7530 * @return 0 on success, non-zero on failure.
7533 mdb_rebalance(MDB_cursor *mc)
7537 unsigned int ptop, minkeys;
7541 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7542 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7543 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7544 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7545 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7547 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7548 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7549 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7550 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7554 if (mc->mc_snum < 2) {
7555 MDB_page *mp = mc->mc_pg[0];
7557 DPUTS("Can't rebalance a subpage, ignoring");
7560 if (NUMKEYS(mp) == 0) {
7561 DPUTS("tree is completely empty");
7562 mc->mc_db->md_root = P_INVALID;
7563 mc->mc_db->md_depth = 0;
7564 mc->mc_db->md_leaf_pages = 0;
7565 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7568 /* Adjust cursors pointing to mp */
7571 mc->mc_flags &= ~C_INITIALIZED;
7573 MDB_cursor *m2, *m3;
7574 MDB_dbi dbi = mc->mc_dbi;
7576 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7577 if (mc->mc_flags & C_SUB)
7578 m3 = &m2->mc_xcursor->mx_cursor;
7581 if (m3->mc_snum < mc->mc_snum) continue;
7582 if (m3->mc_pg[0] == mp) {
7585 m3->mc_flags &= ~C_INITIALIZED;
7589 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7591 DPUTS("collapsing root page!");
7592 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7595 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7596 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7599 mc->mc_db->md_depth--;
7600 mc->mc_db->md_branch_pages--;
7601 mc->mc_ki[0] = mc->mc_ki[1];
7602 for (i = 1; i<mc->mc_db->md_depth; i++) {
7603 mc->mc_pg[i] = mc->mc_pg[i+1];
7604 mc->mc_ki[i] = mc->mc_ki[i+1];
7607 /* Adjust other cursors pointing to mp */
7608 MDB_cursor *m2, *m3;
7609 MDB_dbi dbi = mc->mc_dbi;
7611 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7612 if (mc->mc_flags & C_SUB)
7613 m3 = &m2->mc_xcursor->mx_cursor;
7616 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7617 if (m3->mc_pg[0] == mp) {
7620 for (i=0; i<m3->mc_snum; i++) {
7621 m3->mc_pg[i] = m3->mc_pg[i+1];
7622 m3->mc_ki[i] = m3->mc_ki[i+1];
7628 DPUTS("root page doesn't need rebalancing");
7632 /* The parent (branch page) must have at least 2 pointers,
7633 * otherwise the tree is invalid.
7635 ptop = mc->mc_top-1;
7636 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7638 /* Leaf page fill factor is below the threshold.
7639 * Try to move keys from left or right neighbor, or
7640 * merge with a neighbor page.
7645 mdb_cursor_copy(mc, &mn);
7646 mn.mc_xcursor = NULL;
7648 oldki = mc->mc_ki[mc->mc_top];
7649 if (mc->mc_ki[ptop] == 0) {
7650 /* We're the leftmost leaf in our parent.
7652 DPUTS("reading right neighbor");
7654 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7655 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7658 mn.mc_ki[mn.mc_top] = 0;
7659 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7661 /* There is at least one neighbor to the left.
7663 DPUTS("reading left neighbor");
7665 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7666 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7669 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7670 mc->mc_ki[mc->mc_top] = 0;
7673 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7674 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7675 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7677 /* If the neighbor page is above threshold and has enough keys,
7678 * move one key from it. Otherwise we should try to merge them.
7679 * (A branch page must never have less than 2 keys.)
7681 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7682 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7683 rc = mdb_node_move(&mn, mc);
7684 if (mc->mc_ki[ptop]) {
7688 if (mc->mc_ki[ptop] == 0) {
7689 rc = mdb_page_merge(&mn, mc);
7691 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7692 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7693 rc = mdb_page_merge(mc, &mn);
7694 mdb_cursor_copy(&mn, mc);
7696 mc->mc_flags &= ~C_EOF;
7698 mc->mc_ki[mc->mc_top] = oldki;
7702 /** Complete a delete operation started by #mdb_cursor_del(). */
7704 mdb_cursor_del0(MDB_cursor *mc)
7711 ki = mc->mc_ki[mc->mc_top];
7712 mdb_node_del(mc, mc->mc_db->md_pad);
7713 mc->mc_db->md_entries--;
7714 rc = mdb_rebalance(mc);
7716 if (rc == MDB_SUCCESS) {
7717 MDB_cursor *m2, *m3;
7718 MDB_dbi dbi = mc->mc_dbi;
7720 mp = mc->mc_pg[mc->mc_top];
7721 nkeys = NUMKEYS(mp);
7723 /* if mc points past last node in page, find next sibling */
7724 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7725 rc = mdb_cursor_sibling(mc, 1);
7726 if (rc == MDB_NOTFOUND) {
7727 mc->mc_flags |= C_EOF;
7732 /* Adjust other cursors pointing to mp */
7733 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7734 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7735 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7737 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7739 if (m3->mc_pg[mc->mc_top] == mp) {
7740 if (m3->mc_ki[mc->mc_top] >= ki) {
7741 m3->mc_flags |= C_DEL;
7742 if (m3->mc_ki[mc->mc_top] > ki)
7743 m3->mc_ki[mc->mc_top]--;
7744 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7745 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7747 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7748 rc = mdb_cursor_sibling(m3, 1);
7749 if (rc == MDB_NOTFOUND) {
7750 m3->mc_flags |= C_EOF;
7756 mc->mc_flags |= C_DEL;
7760 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7765 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7766 MDB_val *key, MDB_val *data)
7768 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7771 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7772 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7774 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7775 /* must ignore any data */
7779 return mdb_del0(txn, dbi, key, data, 0);
7783 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7784 MDB_val *key, MDB_val *data, unsigned flags)
7789 MDB_val rdata, *xdata;
7793 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7795 mdb_cursor_init(&mc, txn, dbi, &mx);
7804 flags |= MDB_NODUPDATA;
7806 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7808 /* let mdb_page_split know about this cursor if needed:
7809 * delete will trigger a rebalance; if it needs to move
7810 * a node from one page to another, it will have to
7811 * update the parent's separator key(s). If the new sepkey
7812 * is larger than the current one, the parent page may
7813 * run out of space, triggering a split. We need this
7814 * cursor to be consistent until the end of the rebalance.
7816 mc.mc_flags |= C_UNTRACK;
7817 mc.mc_next = txn->mt_cursors[dbi];
7818 txn->mt_cursors[dbi] = &mc;
7819 rc = mdb_cursor_del(&mc, flags);
7820 txn->mt_cursors[dbi] = mc.mc_next;
7825 /** Split a page and insert a new node.
7826 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7827 * The cursor will be updated to point to the actual page and index where
7828 * the node got inserted after the split.
7829 * @param[in] newkey The key for the newly inserted node.
7830 * @param[in] newdata The data for the newly inserted node.
7831 * @param[in] newpgno The page number, if the new node is a branch node.
7832 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7833 * @return 0 on success, non-zero on failure.
7836 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7837 unsigned int nflags)
7840 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7843 int i, j, split_indx, nkeys, pmax;
7844 MDB_env *env = mc->mc_txn->mt_env;
7846 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7847 MDB_page *copy = NULL;
7848 MDB_page *mp, *rp, *pp;
7853 mp = mc->mc_pg[mc->mc_top];
7854 newindx = mc->mc_ki[mc->mc_top];
7855 nkeys = NUMKEYS(mp);
7857 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7858 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7859 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7861 /* Create a right sibling. */
7862 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7864 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7866 if (mc->mc_snum < 2) {
7867 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7869 /* shift current top to make room for new parent */
7870 mc->mc_pg[1] = mc->mc_pg[0];
7871 mc->mc_ki[1] = mc->mc_ki[0];
7874 mc->mc_db->md_root = pp->mp_pgno;
7875 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7876 mc->mc_db->md_depth++;
7879 /* Add left (implicit) pointer. */
7880 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7881 /* undo the pre-push */
7882 mc->mc_pg[0] = mc->mc_pg[1];
7883 mc->mc_ki[0] = mc->mc_ki[1];
7884 mc->mc_db->md_root = mp->mp_pgno;
7885 mc->mc_db->md_depth--;
7892 ptop = mc->mc_top-1;
7893 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7896 mc->mc_flags |= C_SPLITTING;
7897 mdb_cursor_copy(mc, &mn);
7898 mn.mc_pg[mn.mc_top] = rp;
7899 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7901 if (nflags & MDB_APPEND) {
7902 mn.mc_ki[mn.mc_top] = 0;
7904 split_indx = newindx;
7908 split_indx = (nkeys+1) / 2;
7913 unsigned int lsize, rsize, ksize;
7914 /* Move half of the keys to the right sibling */
7915 x = mc->mc_ki[mc->mc_top] - split_indx;
7916 ksize = mc->mc_db->md_pad;
7917 split = LEAF2KEY(mp, split_indx, ksize);
7918 rsize = (nkeys - split_indx) * ksize;
7919 lsize = (nkeys - split_indx) * sizeof(indx_t);
7920 mp->mp_lower -= lsize;
7921 rp->mp_lower += lsize;
7922 mp->mp_upper += rsize - lsize;
7923 rp->mp_upper -= rsize - lsize;
7924 sepkey.mv_size = ksize;
7925 if (newindx == split_indx) {
7926 sepkey.mv_data = newkey->mv_data;
7928 sepkey.mv_data = split;
7931 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7932 memcpy(rp->mp_ptrs, split, rsize);
7933 sepkey.mv_data = rp->mp_ptrs;
7934 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7935 memcpy(ins, newkey->mv_data, ksize);
7936 mp->mp_lower += sizeof(indx_t);
7937 mp->mp_upper -= ksize - sizeof(indx_t);
7940 memcpy(rp->mp_ptrs, split, x * ksize);
7941 ins = LEAF2KEY(rp, x, ksize);
7942 memcpy(ins, newkey->mv_data, ksize);
7943 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7944 rp->mp_lower += sizeof(indx_t);
7945 rp->mp_upper -= ksize - sizeof(indx_t);
7946 mc->mc_ki[mc->mc_top] = x;
7947 mc->mc_pg[mc->mc_top] = rp;
7950 int psize, nsize, k;
7951 /* Maximum free space in an empty page */
7952 pmax = env->me_psize - PAGEHDRSZ;
7954 nsize = mdb_leaf_size(env, newkey, newdata);
7956 nsize = mdb_branch_size(env, newkey);
7957 nsize = EVEN(nsize);
7959 /* grab a page to hold a temporary copy */
7960 copy = mdb_page_malloc(mc->mc_txn, 1);
7965 copy->mp_pgno = mp->mp_pgno;
7966 copy->mp_flags = mp->mp_flags;
7967 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7968 copy->mp_upper = env->me_psize - PAGEBASE;
7970 /* prepare to insert */
7971 for (i=0, j=0; i<nkeys; i++) {
7973 copy->mp_ptrs[j++] = 0;
7975 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7978 /* When items are relatively large the split point needs
7979 * to be checked, because being off-by-one will make the
7980 * difference between success or failure in mdb_node_add.
7982 * It's also relevant if a page happens to be laid out
7983 * such that one half of its nodes are all "small" and
7984 * the other half of its nodes are "large." If the new
7985 * item is also "large" and falls on the half with
7986 * "large" nodes, it also may not fit.
7988 * As a final tweak, if the new item goes on the last
7989 * spot on the page (and thus, onto the new page), bias
7990 * the split so the new page is emptier than the old page.
7991 * This yields better packing during sequential inserts.
7993 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7994 /* Find split point */
7996 if (newindx <= split_indx || newindx >= nkeys) {
7998 k = newindx >= nkeys ? nkeys : split_indx+2;
8003 for (; i!=k; i+=j) {
8008 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8009 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8011 if (F_ISSET(node->mn_flags, F_BIGDATA))
8012 psize += sizeof(pgno_t);
8014 psize += NODEDSZ(node);
8016 psize = EVEN(psize);
8018 if (psize > pmax || i == k-j) {
8019 split_indx = i + (j<0);
8024 if (split_indx == newindx) {
8025 sepkey.mv_size = newkey->mv_size;
8026 sepkey.mv_data = newkey->mv_data;
8028 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8029 sepkey.mv_size = node->mn_ksize;
8030 sepkey.mv_data = NODEKEY(node);
8035 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8037 /* Copy separator key to the parent.
8039 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8043 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8048 if (mn.mc_snum == mc->mc_snum) {
8049 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8050 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8051 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8052 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8057 /* Right page might now have changed parent.
8058 * Check if left page also changed parent.
8060 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8061 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8062 for (i=0; i<ptop; i++) {
8063 mc->mc_pg[i] = mn.mc_pg[i];
8064 mc->mc_ki[i] = mn.mc_ki[i];
8066 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8067 if (mn.mc_ki[ptop]) {
8068 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8070 /* find right page's left sibling */
8071 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8072 mdb_cursor_sibling(mc, 0);
8077 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8080 mc->mc_flags ^= C_SPLITTING;
8081 if (rc != MDB_SUCCESS) {
8084 if (nflags & MDB_APPEND) {
8085 mc->mc_pg[mc->mc_top] = rp;
8086 mc->mc_ki[mc->mc_top] = 0;
8087 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8090 for (i=0; i<mc->mc_top; i++)
8091 mc->mc_ki[i] = mn.mc_ki[i];
8092 } else if (!IS_LEAF2(mp)) {
8094 mc->mc_pg[mc->mc_top] = rp;
8099 rkey.mv_data = newkey->mv_data;
8100 rkey.mv_size = newkey->mv_size;
8106 /* Update index for the new key. */
8107 mc->mc_ki[mc->mc_top] = j;
8109 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8110 rkey.mv_data = NODEKEY(node);
8111 rkey.mv_size = node->mn_ksize;
8113 xdata.mv_data = NODEDATA(node);
8114 xdata.mv_size = NODEDSZ(node);
8117 pgno = NODEPGNO(node);
8118 flags = node->mn_flags;
8121 if (!IS_LEAF(mp) && j == 0) {
8122 /* First branch index doesn't need key data. */
8126 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8132 mc->mc_pg[mc->mc_top] = copy;
8137 } while (i != split_indx);
8139 nkeys = NUMKEYS(copy);
8140 for (i=0; i<nkeys; i++)
8141 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8142 mp->mp_lower = copy->mp_lower;
8143 mp->mp_upper = copy->mp_upper;
8144 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8145 env->me_psize - copy->mp_upper - PAGEBASE);
8147 /* reset back to original page */
8148 if (newindx < split_indx) {
8149 mc->mc_pg[mc->mc_top] = mp;
8150 if (nflags & MDB_RESERVE) {
8151 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8152 if (!(node->mn_flags & F_BIGDATA))
8153 newdata->mv_data = NODEDATA(node);
8156 mc->mc_pg[mc->mc_top] = rp;
8158 /* Make sure mc_ki is still valid.
8160 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8161 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8162 for (i=0; i<=ptop; i++) {
8163 mc->mc_pg[i] = mn.mc_pg[i];
8164 mc->mc_ki[i] = mn.mc_ki[i];
8171 /* Adjust other cursors pointing to mp */
8172 MDB_cursor *m2, *m3;
8173 MDB_dbi dbi = mc->mc_dbi;
8174 int fixup = NUMKEYS(mp);
8176 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8177 if (mc->mc_flags & C_SUB)
8178 m3 = &m2->mc_xcursor->mx_cursor;
8183 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8185 if (m3->mc_flags & C_SPLITTING)
8190 for (k=m3->mc_top; k>=0; k--) {
8191 m3->mc_ki[k+1] = m3->mc_ki[k];
8192 m3->mc_pg[k+1] = m3->mc_pg[k];
8194 if (m3->mc_ki[0] >= split_indx) {
8199 m3->mc_pg[0] = mc->mc_pg[0];
8203 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8204 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8205 m3->mc_ki[mc->mc_top]++;
8206 if (m3->mc_ki[mc->mc_top] >= fixup) {
8207 m3->mc_pg[mc->mc_top] = rp;
8208 m3->mc_ki[mc->mc_top] -= fixup;
8209 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8211 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8212 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8217 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8220 if (copy) /* tmp page */
8221 mdb_page_free(env, copy);
8223 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8228 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8229 MDB_val *key, MDB_val *data, unsigned int flags)
8234 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8237 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8240 mdb_cursor_init(&mc, txn, dbi, &mx);
8241 return mdb_cursor_put(&mc, key, data, flags);
8245 #define MDB_WBUF (1024*1024)
8248 /** State needed for a compacting copy. */
8249 typedef struct mdb_copy {
8250 pthread_mutex_t mc_mutex;
8251 pthread_cond_t mc_cond;
8258 pgno_t mc_next_pgno;
8261 volatile int mc_new;
8266 /** Dedicated writer thread for compacting copy. */
8267 static THREAD_RET ESECT
8268 mdb_env_copythr(void *arg)
8272 int toggle = 0, wsize, rc;
8275 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8278 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8281 pthread_mutex_lock(&my->mc_mutex);
8283 pthread_cond_signal(&my->mc_cond);
8286 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8287 if (my->mc_new < 0) {
8292 wsize = my->mc_wlen[toggle];
8293 ptr = my->mc_wbuf[toggle];
8296 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8300 } else if (len > 0) {
8314 /* If there's an overflow page tail, write it too */
8315 if (my->mc_olen[toggle]) {
8316 wsize = my->mc_olen[toggle];
8317 ptr = my->mc_over[toggle];
8318 my->mc_olen[toggle] = 0;
8321 my->mc_wlen[toggle] = 0;
8323 pthread_cond_signal(&my->mc_cond);
8325 pthread_cond_signal(&my->mc_cond);
8326 pthread_mutex_unlock(&my->mc_mutex);
8327 return (THREAD_RET)0;
8331 /** Tell the writer thread there's a buffer ready to write */
8333 mdb_env_cthr_toggle(mdb_copy *my, int st)
8335 int toggle = my->mc_toggle ^ 1;
8336 pthread_mutex_lock(&my->mc_mutex);
8337 if (my->mc_status) {
8338 pthread_mutex_unlock(&my->mc_mutex);
8339 return my->mc_status;
8341 while (my->mc_new == 1)
8342 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8344 my->mc_toggle = toggle;
8345 pthread_cond_signal(&my->mc_cond);
8346 pthread_mutex_unlock(&my->mc_mutex);
8350 /** Depth-first tree traversal for compacting copy. */
8352 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8355 MDB_txn *txn = my->mc_txn;
8357 MDB_page *mo, *mp, *leaf;
8362 /* Empty DB, nothing to do */
8363 if (*pg == P_INVALID)
8370 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8373 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8377 /* Make cursor pages writable */
8378 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8382 for (i=0; i<mc.mc_top; i++) {
8383 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8384 mc.mc_pg[i] = (MDB_page *)ptr;
8385 ptr += my->mc_env->me_psize;
8388 /* This is writable space for a leaf page. Usually not needed. */
8389 leaf = (MDB_page *)ptr;
8391 toggle = my->mc_toggle;
8392 while (mc.mc_snum > 0) {
8394 mp = mc.mc_pg[mc.mc_top];
8398 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8399 for (i=0; i<n; i++) {
8400 ni = NODEPTR(mp, i);
8401 if (ni->mn_flags & F_BIGDATA) {
8405 /* Need writable leaf */
8407 mc.mc_pg[mc.mc_top] = leaf;
8408 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8410 ni = NODEPTR(mp, i);
8413 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8414 rc = mdb_page_get(txn, pg, &omp, NULL);
8417 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8418 rc = mdb_env_cthr_toggle(my, 1);
8421 toggle = my->mc_toggle;
8423 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8424 memcpy(mo, omp, my->mc_env->me_psize);
8425 mo->mp_pgno = my->mc_next_pgno;
8426 my->mc_next_pgno += omp->mp_pages;
8427 my->mc_wlen[toggle] += my->mc_env->me_psize;
8428 if (omp->mp_pages > 1) {
8429 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8430 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8431 rc = mdb_env_cthr_toggle(my, 1);
8434 toggle = my->mc_toggle;
8436 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8437 } else if (ni->mn_flags & F_SUBDATA) {
8440 /* Need writable leaf */
8442 mc.mc_pg[mc.mc_top] = leaf;
8443 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8445 ni = NODEPTR(mp, i);
8448 memcpy(&db, NODEDATA(ni), sizeof(db));
8449 my->mc_toggle = toggle;
8450 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8453 toggle = my->mc_toggle;
8454 memcpy(NODEDATA(ni), &db, sizeof(db));
8459 mc.mc_ki[mc.mc_top]++;
8460 if (mc.mc_ki[mc.mc_top] < n) {
8463 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8465 rc = mdb_page_get(txn, pg, &mp, NULL);
8470 mc.mc_ki[mc.mc_top] = 0;
8471 if (IS_BRANCH(mp)) {
8472 /* Whenever we advance to a sibling branch page,
8473 * we must proceed all the way down to its first leaf.
8475 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8478 mc.mc_pg[mc.mc_top] = mp;
8482 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8483 rc = mdb_env_cthr_toggle(my, 1);
8486 toggle = my->mc_toggle;
8488 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8489 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8490 mo->mp_pgno = my->mc_next_pgno++;
8491 my->mc_wlen[toggle] += my->mc_env->me_psize;
8493 /* Update parent if there is one */
8494 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8495 SETPGNO(ni, mo->mp_pgno);
8496 mdb_cursor_pop(&mc);
8498 /* Otherwise we're done */
8508 /** Copy environment with compaction. */
8510 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8515 MDB_txn *txn = NULL;
8520 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8521 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8522 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8523 if (my.mc_wbuf[0] == NULL)
8526 pthread_mutex_init(&my.mc_mutex, NULL);
8527 pthread_cond_init(&my.mc_cond, NULL);
8528 #ifdef HAVE_MEMALIGN
8529 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8530 if (my.mc_wbuf[0] == NULL)
8533 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8538 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8539 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8544 my.mc_next_pgno = 2;
8550 THREAD_CREATE(thr, mdb_env_copythr, &my);
8552 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8556 mp = (MDB_page *)my.mc_wbuf[0];
8557 memset(mp, 0, 2*env->me_psize);
8559 mp->mp_flags = P_META;
8560 mm = (MDB_meta *)METADATA(mp);
8561 mdb_env_init_meta0(env, mm);
8562 mm->mm_address = env->me_metas[0]->mm_address;
8564 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8566 mp->mp_flags = P_META;
8567 *(MDB_meta *)METADATA(mp) = *mm;
8568 mm = (MDB_meta *)METADATA(mp);
8570 /* Count the number of free pages, subtract from lastpg to find
8571 * number of active pages
8574 MDB_ID freecount = 0;
8577 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8578 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8579 freecount += *(MDB_ID *)data.mv_data;
8580 freecount += txn->mt_dbs[0].md_branch_pages +
8581 txn->mt_dbs[0].md_leaf_pages +
8582 txn->mt_dbs[0].md_overflow_pages;
8584 /* Set metapage 1 */
8585 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8586 mm->mm_dbs[1] = txn->mt_dbs[1];
8587 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8590 my.mc_wlen[0] = env->me_psize * 2;
8592 pthread_mutex_lock(&my.mc_mutex);
8594 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8595 pthread_mutex_unlock(&my.mc_mutex);
8596 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8597 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8598 rc = mdb_env_cthr_toggle(&my, 1);
8599 mdb_env_cthr_toggle(&my, -1);
8600 pthread_mutex_lock(&my.mc_mutex);
8602 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8603 pthread_mutex_unlock(&my.mc_mutex);
8608 CloseHandle(my.mc_cond);
8609 CloseHandle(my.mc_mutex);
8610 _aligned_free(my.mc_wbuf[0]);
8612 pthread_cond_destroy(&my.mc_cond);
8613 pthread_mutex_destroy(&my.mc_mutex);
8614 free(my.mc_wbuf[0]);
8619 /** Copy environment as-is. */
8621 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8623 MDB_txn *txn = NULL;
8629 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8633 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8636 /* Do the lock/unlock of the reader mutex before starting the
8637 * write txn. Otherwise other read txns could block writers.
8639 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8644 /* We must start the actual read txn after blocking writers */
8645 mdb_txn_reset0(txn, "reset-stage1");
8647 /* Temporarily block writers until we snapshot the meta pages */
8650 rc = mdb_txn_renew0(txn);
8652 UNLOCK_MUTEX_W(env);
8657 wsize = env->me_psize * 2;
8661 DO_WRITE(rc, fd, ptr, w2, len);
8665 } else if (len > 0) {
8671 /* Non-blocking or async handles are not supported */
8677 UNLOCK_MUTEX_W(env);
8682 w2 = txn->mt_next_pgno * env->me_psize;
8685 LARGE_INTEGER fsize;
8686 GetFileSizeEx(env->me_fd, &fsize);
8687 if (w2 > fsize.QuadPart)
8688 w2 = fsize.QuadPart;
8693 fstat(env->me_fd, &st);
8694 if (w2 > (size_t)st.st_size)
8700 if (wsize > MAX_WRITE)
8704 DO_WRITE(rc, fd, ptr, w2, len);
8708 } else if (len > 0) {
8725 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8727 if (flags & MDB_CP_COMPACT)
8728 return mdb_env_copyfd1(env, fd);
8730 return mdb_env_copyfd0(env, fd);
8734 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8736 return mdb_env_copyfd2(env, fd, 0);
8740 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8744 HANDLE newfd = INVALID_HANDLE_VALUE;
8746 if (env->me_flags & MDB_NOSUBDIR) {
8747 lpath = (char *)path;
8750 len += sizeof(DATANAME);
8751 lpath = malloc(len);
8754 sprintf(lpath, "%s" DATANAME, path);
8757 /* The destination path must exist, but the destination file must not.
8758 * We don't want the OS to cache the writes, since the source data is
8759 * already in the OS cache.
8762 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8763 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8765 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8767 if (newfd == INVALID_HANDLE_VALUE) {
8772 if (env->me_psize >= env->me_os_psize) {
8774 /* Set O_DIRECT if the file system supports it */
8775 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8776 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8778 #ifdef F_NOCACHE /* __APPLE__ */
8779 rc = fcntl(newfd, F_NOCACHE, 1);
8787 rc = mdb_env_copyfd2(env, newfd, flags);
8790 if (!(env->me_flags & MDB_NOSUBDIR))
8792 if (newfd != INVALID_HANDLE_VALUE)
8793 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8800 mdb_env_copy(MDB_env *env, const char *path)
8802 return mdb_env_copy2(env, path, 0);
8806 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8808 if ((flag & CHANGEABLE) != flag)
8811 env->me_flags |= flag;
8813 env->me_flags &= ~flag;
8818 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8823 *arg = env->me_flags;
8828 mdb_env_set_userctx(MDB_env *env, void *ctx)
8832 env->me_userctx = ctx;
8837 mdb_env_get_userctx(MDB_env *env)
8839 return env ? env->me_userctx : NULL;
8843 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8848 env->me_assert_func = func;
8854 mdb_env_get_path(MDB_env *env, const char **arg)
8859 *arg = env->me_path;
8864 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8873 /** Common code for #mdb_stat() and #mdb_env_stat().
8874 * @param[in] env the environment to operate in.
8875 * @param[in] db the #MDB_db record containing the stats to return.
8876 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8877 * @return 0, this function always succeeds.
8880 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8882 arg->ms_psize = env->me_psize;
8883 arg->ms_depth = db->md_depth;
8884 arg->ms_branch_pages = db->md_branch_pages;
8885 arg->ms_leaf_pages = db->md_leaf_pages;
8886 arg->ms_overflow_pages = db->md_overflow_pages;
8887 arg->ms_entries = db->md_entries;
8893 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8897 if (env == NULL || arg == NULL)
8900 toggle = mdb_env_pick_meta(env);
8902 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8906 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8910 if (env == NULL || arg == NULL)
8913 toggle = mdb_env_pick_meta(env);
8914 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8915 arg->me_mapsize = env->me_mapsize;
8916 arg->me_maxreaders = env->me_maxreaders;
8918 /* me_numreaders may be zero if this process never used any readers. Use
8919 * the shared numreader count if it exists.
8921 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8923 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8924 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8928 /** Set the default comparison functions for a database.
8929 * Called immediately after a database is opened to set the defaults.
8930 * The user can then override them with #mdb_set_compare() or
8931 * #mdb_set_dupsort().
8932 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8933 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8936 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8938 uint16_t f = txn->mt_dbs[dbi].md_flags;
8940 txn->mt_dbxs[dbi].md_cmp =
8941 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8942 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8944 txn->mt_dbxs[dbi].md_dcmp =
8945 !(f & MDB_DUPSORT) ? 0 :
8946 ((f & MDB_INTEGERDUP)
8947 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8948 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8951 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8957 int rc, dbflag, exact;
8958 unsigned int unused = 0, seq;
8961 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8962 mdb_default_cmp(txn, FREE_DBI);
8965 if ((flags & VALID_FLAGS) != flags)
8967 if (txn->mt_flags & MDB_TXN_ERROR)
8973 if (flags & PERSISTENT_FLAGS) {
8974 uint16_t f2 = flags & PERSISTENT_FLAGS;
8975 /* make sure flag changes get committed */
8976 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8977 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8978 txn->mt_flags |= MDB_TXN_DIRTY;
8981 mdb_default_cmp(txn, MAIN_DBI);
8985 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8986 mdb_default_cmp(txn, MAIN_DBI);
8989 /* Is the DB already open? */
8991 for (i=2; i<txn->mt_numdbs; i++) {
8992 if (!txn->mt_dbxs[i].md_name.mv_size) {
8993 /* Remember this free slot */
8994 if (!unused) unused = i;
8997 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8998 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9004 /* If no free slot and max hit, fail */
9005 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9006 return MDB_DBS_FULL;
9008 /* Cannot mix named databases with some mainDB flags */
9009 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9010 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9012 /* Find the DB info */
9013 dbflag = DB_NEW|DB_VALID;
9016 key.mv_data = (void *)name;
9017 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9018 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9019 if (rc == MDB_SUCCESS) {
9020 /* make sure this is actually a DB */
9021 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9022 if (!(node->mn_flags & F_SUBDATA))
9023 return MDB_INCOMPATIBLE;
9024 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9025 /* Create if requested */
9026 data.mv_size = sizeof(MDB_db);
9027 data.mv_data = &dummy;
9028 memset(&dummy, 0, sizeof(dummy));
9029 dummy.md_root = P_INVALID;
9030 dummy.md_flags = flags & PERSISTENT_FLAGS;
9031 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9035 /* OK, got info, add to table */
9036 if (rc == MDB_SUCCESS) {
9037 unsigned int slot = unused ? unused : txn->mt_numdbs;
9038 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9039 txn->mt_dbxs[slot].md_name.mv_size = len;
9040 txn->mt_dbxs[slot].md_rel = NULL;
9041 txn->mt_dbflags[slot] = dbflag;
9042 /* txn-> and env-> are the same in read txns, use
9043 * tmp variable to avoid undefined assignment
9045 seq = ++txn->mt_env->me_dbiseqs[slot];
9046 txn->mt_dbiseqs[slot] = seq;
9048 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9050 mdb_default_cmp(txn, slot);
9059 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9061 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9064 if (txn->mt_flags & MDB_TXN_ERROR)
9067 if (txn->mt_dbflags[dbi] & DB_STALE) {
9070 /* Stale, must read the DB's root. cursor_init does it for us. */
9071 mdb_cursor_init(&mc, txn, dbi, &mx);
9073 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9076 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9079 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9081 ptr = env->me_dbxs[dbi].md_name.mv_data;
9082 /* If there was no name, this was already closed */
9084 env->me_dbxs[dbi].md_name.mv_data = NULL;
9085 env->me_dbxs[dbi].md_name.mv_size = 0;
9086 env->me_dbflags[dbi] = 0;
9087 env->me_dbiseqs[dbi]++;
9092 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9094 /* We could return the flags for the FREE_DBI too but what's the point? */
9095 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9097 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9101 /** Add all the DB's pages to the free list.
9102 * @param[in] mc Cursor on the DB to free.
9103 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9104 * @return 0 on success, non-zero on failure.
9107 mdb_drop0(MDB_cursor *mc, int subs)
9111 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9112 if (rc == MDB_SUCCESS) {
9113 MDB_txn *txn = mc->mc_txn;
9118 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9119 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9122 mdb_cursor_copy(mc, &mx);
9123 while (mc->mc_snum > 0) {
9124 MDB_page *mp = mc->mc_pg[mc->mc_top];
9125 unsigned n = NUMKEYS(mp);
9127 for (i=0; i<n; i++) {
9128 ni = NODEPTR(mp, i);
9129 if (ni->mn_flags & F_BIGDATA) {
9132 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9133 rc = mdb_page_get(txn, pg, &omp, NULL);
9136 mdb_cassert(mc, IS_OVERFLOW(omp));
9137 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9141 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9142 mdb_xcursor_init1(mc, ni);
9143 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9149 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9151 for (i=0; i<n; i++) {
9153 ni = NODEPTR(mp, i);
9156 mdb_midl_xappend(txn->mt_free_pgs, pg);
9161 mc->mc_ki[mc->mc_top] = i;
9162 rc = mdb_cursor_sibling(mc, 1);
9164 if (rc != MDB_NOTFOUND)
9166 /* no more siblings, go back to beginning
9167 * of previous level.
9171 for (i=1; i<mc->mc_snum; i++) {
9173 mc->mc_pg[i] = mx.mc_pg[i];
9178 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9181 txn->mt_flags |= MDB_TXN_ERROR;
9182 } else if (rc == MDB_NOTFOUND) {
9188 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9190 MDB_cursor *mc, *m2;
9193 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9196 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9199 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9202 rc = mdb_cursor_open(txn, dbi, &mc);
9206 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9207 /* Invalidate the dropped DB's cursors */
9208 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9209 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9213 /* Can't delete the main DB */
9214 if (del && dbi > MAIN_DBI) {
9215 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9217 txn->mt_dbflags[dbi] = DB_STALE;
9218 mdb_dbi_close(txn->mt_env, dbi);
9220 txn->mt_flags |= MDB_TXN_ERROR;
9223 /* reset the DB record, mark it dirty */
9224 txn->mt_dbflags[dbi] |= DB_DIRTY;
9225 txn->mt_dbs[dbi].md_depth = 0;
9226 txn->mt_dbs[dbi].md_branch_pages = 0;
9227 txn->mt_dbs[dbi].md_leaf_pages = 0;
9228 txn->mt_dbs[dbi].md_overflow_pages = 0;
9229 txn->mt_dbs[dbi].md_entries = 0;
9230 txn->mt_dbs[dbi].md_root = P_INVALID;
9232 txn->mt_flags |= MDB_TXN_DIRTY;
9235 mdb_cursor_close(mc);
9239 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9241 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9244 txn->mt_dbxs[dbi].md_cmp = cmp;
9248 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9250 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9253 txn->mt_dbxs[dbi].md_dcmp = cmp;
9257 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9259 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9262 txn->mt_dbxs[dbi].md_rel = rel;
9266 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9268 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9271 txn->mt_dbxs[dbi].md_relctx = ctx;
9276 mdb_env_get_maxkeysize(MDB_env *env)
9278 return ENV_MAXKEY(env);
9282 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9284 unsigned int i, rdrs;
9287 int rc = 0, first = 1;
9291 if (!env->me_txns) {
9292 return func("(no reader locks)\n", ctx);
9294 rdrs = env->me_txns->mti_numreaders;
9295 mr = env->me_txns->mti_readers;
9296 for (i=0; i<rdrs; i++) {
9298 txnid_t txnid = mr[i].mr_txnid;
9299 sprintf(buf, txnid == (txnid_t)-1 ?
9300 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9301 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9304 rc = func(" pid thread txnid\n", ctx);
9308 rc = func(buf, ctx);
9314 rc = func("(no active readers)\n", ctx);
9319 /** Insert pid into list if not already present.
9320 * return -1 if already present.
9323 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9325 /* binary search of pid in list */
9327 unsigned cursor = 1;
9329 unsigned n = ids[0];
9332 unsigned pivot = n >> 1;
9333 cursor = base + pivot + 1;
9334 val = pid - ids[cursor];
9339 } else if ( val > 0 ) {
9344 /* found, so it's a duplicate */
9353 for (n = ids[0]; n > cursor; n--)
9360 mdb_reader_check(MDB_env *env, int *dead)
9362 unsigned int i, j, rdrs;
9364 MDB_PID_T *pids, pid;
9373 rdrs = env->me_txns->mti_numreaders;
9374 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9378 mr = env->me_txns->mti_readers;
9379 for (i=0; i<rdrs; i++) {
9380 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9382 if (mdb_pid_insert(pids, pid) == 0) {
9383 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9385 /* Recheck, a new process may have reused pid */
9386 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9387 for (j=i; j<rdrs; j++)
9388 if (mr[j].mr_pid == pid) {
9389 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9390 (unsigned) pid, mr[j].mr_txnid));
9395 UNLOCK_MUTEX_R(env);