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 2011-2015 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)
82 #if defined(__linux) && !defined(MDB_FDATASYNC_WORKS)
83 /** fdatasync is broken on ext3/ext4fs on older kernels, see
84 * description in #mdb_env_open2 comments. You can safely
85 * define MDB_FDATASYNC_WORKS if this code will only be run
86 * on kernels 3.6 and newer.
88 #define BROKEN_FDATASYNC
101 #if defined(__sun) || defined(ANDROID)
102 /* Most platforms have posix_memalign, older may only have memalign */
103 #define HAVE_MEMALIGN 1
107 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
108 #include <netinet/in.h>
109 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
112 #if defined(__APPLE__) || defined (BSD)
113 # define MDB_USE_POSIX_SEM 1
114 # define MDB_FDATASYNC fsync
115 #elif defined(ANDROID)
116 # define MDB_FDATASYNC fsync
121 #ifdef MDB_USE_POSIX_SEM
122 # define MDB_USE_HASH 1
123 #include <semaphore.h>
128 #include <valgrind/memcheck.h>
129 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
130 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
131 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
132 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
133 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
135 #define VGMEMP_CREATE(h,r,z)
136 #define VGMEMP_ALLOC(h,a,s)
137 #define VGMEMP_FREE(h,a)
138 #define VGMEMP_DESTROY(h)
139 #define VGMEMP_DEFINED(a,s)
143 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
144 /* Solaris just defines one or the other */
145 # define LITTLE_ENDIAN 1234
146 # define BIG_ENDIAN 4321
147 # ifdef _LITTLE_ENDIAN
148 # define BYTE_ORDER LITTLE_ENDIAN
150 # define BYTE_ORDER BIG_ENDIAN
153 # define BYTE_ORDER __BYTE_ORDER
157 #ifndef LITTLE_ENDIAN
158 #define LITTLE_ENDIAN __LITTLE_ENDIAN
161 #define BIG_ENDIAN __BIG_ENDIAN
164 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
165 #define MISALIGNED_OK 1
171 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
172 # error "Unknown or unsupported endianness (BYTE_ORDER)"
173 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
174 # error "Two's complement, reasonably sized integer types, please"
178 /** Put infrequently used env functions in separate section */
180 # define ESECT __attribute__ ((section("__TEXT,text_env")))
182 # define ESECT __attribute__ ((section("text_env")))
189 #define CALL_CONV WINAPI
194 /** @defgroup internal LMDB Internals
197 /** @defgroup compat Compatibility Macros
198 * A bunch of macros to minimize the amount of platform-specific ifdefs
199 * needed throughout the rest of the code. When the features this library
200 * needs are similar enough to POSIX to be hidden in a one-or-two line
201 * replacement, this macro approach is used.
205 /** Features under development */
210 /** Wrapper around __func__, which is a C99 feature */
211 #if __STDC_VERSION__ >= 199901L
212 # define mdb_func_ __func__
213 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
214 # define mdb_func_ __FUNCTION__
216 /* If a debug message says <mdb_unknown>(), update the #if statements above */
217 # define mdb_func_ "<mdb_unknown>"
221 #define MDB_USE_HASH 1
222 #define MDB_PIDLOCK 0
223 #define THREAD_RET DWORD
224 #define pthread_t HANDLE
225 #define pthread_mutex_t HANDLE
226 #define pthread_cond_t HANDLE
227 #define pthread_key_t DWORD
228 #define pthread_self() GetCurrentThreadId()
229 #define pthread_key_create(x,y) \
230 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
231 #define pthread_key_delete(x) TlsFree(x)
232 #define pthread_getspecific(x) TlsGetValue(x)
233 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
234 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
235 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
236 #define pthread_cond_signal(x) SetEvent(*x)
237 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
238 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
239 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
240 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
241 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
242 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
243 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
244 #define getpid() GetCurrentProcessId()
245 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
246 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
247 #define ErrCode() GetLastError()
248 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
249 #define close(fd) (CloseHandle(fd) ? 0 : -1)
250 #define munmap(ptr,len) UnmapViewOfFile(ptr)
251 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
252 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
254 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
258 #define THREAD_RET void *
259 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
260 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
261 #define Z "z" /**< printf format modifier for size_t */
263 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
264 #define MDB_PIDLOCK 1
266 #ifdef MDB_USE_POSIX_SEM
268 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
269 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
270 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
271 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
274 mdb_sem_wait(sem_t *sem)
277 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
282 /** Lock the reader mutex.
284 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
285 /** Unlock the reader mutex.
287 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
289 /** Lock the writer mutex.
290 * Only a single write transaction is allowed at a time. Other writers
291 * will block waiting for this mutex.
293 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
294 /** Unlock the writer mutex.
296 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
297 #endif /* MDB_USE_POSIX_SEM */
299 /** Get the error code for the last failed system function.
301 #define ErrCode() errno
303 /** An abstraction for a file handle.
304 * On POSIX systems file handles are small integers. On Windows
305 * they're opaque pointers.
309 /** A value for an invalid file handle.
310 * Mainly used to initialize file variables and signify that they are
313 #define INVALID_HANDLE_VALUE (-1)
315 /** Get the size of a memory page for the system.
316 * This is the basic size that the platform's memory manager uses, and is
317 * fundamental to the use of memory-mapped files.
319 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
322 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
325 #define MNAME_LEN (sizeof(pthread_mutex_t))
331 /** A flag for opening a file and requesting synchronous data writes.
332 * This is only used when writing a meta page. It's not strictly needed;
333 * we could just do a normal write and then immediately perform a flush.
334 * But if this flag is available it saves us an extra system call.
336 * @note If O_DSYNC is undefined but exists in /usr/include,
337 * preferably set some compiler flag to get the definition.
338 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
341 # define MDB_DSYNC O_DSYNC
345 /** Function for flushing the data of a file. Define this to fsync
346 * if fdatasync() is not supported.
348 #ifndef MDB_FDATASYNC
349 # define MDB_FDATASYNC fdatasync
353 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
364 /** A page number in the database.
365 * Note that 64 bit page numbers are overkill, since pages themselves
366 * already represent 12-13 bits of addressable memory, and the OS will
367 * always limit applications to a maximum of 63 bits of address space.
369 * @note In the #MDB_node structure, we only store 48 bits of this value,
370 * which thus limits us to only 60 bits of addressable data.
372 typedef MDB_ID pgno_t;
374 /** A transaction ID.
375 * See struct MDB_txn.mt_txnid for details.
377 typedef MDB_ID txnid_t;
379 /** @defgroup debug Debug Macros
383 /** Enable debug output. Needs variable argument macros (a C99 feature).
384 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
385 * read from and written to the database (used for free space management).
391 static int mdb_debug;
392 static txnid_t mdb_debug_start;
394 /** Print a debug message with printf formatting.
395 * Requires double parenthesis around 2 or more args.
397 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
398 # define DPRINTF0(fmt, ...) \
399 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
401 # define DPRINTF(args) ((void) 0)
403 /** Print a debug string.
404 * The string is printed literally, with no format processing.
406 #define DPUTS(arg) DPRINTF(("%s", arg))
407 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
409 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
412 /** @brief The maximum size of a database page.
414 * It is 32k or 64k, since value-PAGEBASE must fit in
415 * #MDB_page.%mp_upper.
417 * LMDB will use database pages < OS pages if needed.
418 * That causes more I/O in write transactions: The OS must
419 * know (read) the whole page before writing a partial page.
421 * Note that we don't currently support Huge pages. On Linux,
422 * regular data files cannot use Huge pages, and in general
423 * Huge pages aren't actually pageable. We rely on the OS
424 * demand-pager to read our data and page it out when memory
425 * pressure from other processes is high. So until OSs have
426 * actual paging support for Huge pages, they're not viable.
428 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
430 /** The minimum number of keys required in a database page.
431 * Setting this to a larger value will place a smaller bound on the
432 * maximum size of a data item. Data items larger than this size will
433 * be pushed into overflow pages instead of being stored directly in
434 * the B-tree node. This value used to default to 4. With a page size
435 * of 4096 bytes that meant that any item larger than 1024 bytes would
436 * go into an overflow page. That also meant that on average 2-3KB of
437 * each overflow page was wasted space. The value cannot be lower than
438 * 2 because then there would no longer be a tree structure. With this
439 * value, items larger than 2KB will go into overflow pages, and on
440 * average only 1KB will be wasted.
442 #define MDB_MINKEYS 2
444 /** A stamp that identifies a file as an LMDB file.
445 * There's nothing special about this value other than that it is easily
446 * recognizable, and it will reflect any byte order mismatches.
448 #define MDB_MAGIC 0xBEEFC0DE
450 /** The version number for a database's datafile format. */
451 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
452 /** The version number for a database's lockfile format. */
453 #define MDB_LOCK_VERSION 1
455 /** @brief The max size of a key we can write, or 0 for computed max.
457 * This macro should normally be left alone or set to 0.
458 * Note that a database with big keys or dupsort data cannot be
459 * reliably modified by a liblmdb which uses a smaller max.
460 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
462 * Other values are allowed, for backwards compat. However:
463 * A value bigger than the computed max can break if you do not
464 * know what you are doing, and liblmdb <= 0.9.10 can break when
465 * modifying a DB with keys/dupsort data bigger than its max.
467 * Data items in an #MDB_DUPSORT database are also limited to
468 * this size, since they're actually keys of a sub-DB. Keys and
469 * #MDB_DUPSORT data items must fit on a node in a regular page.
471 #ifndef MDB_MAXKEYSIZE
472 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
475 /** The maximum size of a key we can write to the environment. */
477 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
479 #define ENV_MAXKEY(env) ((env)->me_maxkey)
482 /** @brief The maximum size of a data item.
484 * We only store a 32 bit value for node sizes.
486 #define MAXDATASIZE 0xffffffffUL
489 /** Key size which fits in a #DKBUF.
492 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
495 * This is used for printing a hex dump of a key's contents.
497 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
498 /** Display a key in hex.
500 * Invoke a function to display a key in hex.
502 #define DKEY(x) mdb_dkey(x, kbuf)
508 /** An invalid page number.
509 * Mainly used to denote an empty tree.
511 #define P_INVALID (~(pgno_t)0)
513 /** Test if the flags \b f are set in a flag word \b w. */
514 #define F_ISSET(w, f) (((w) & (f)) == (f))
516 /** Round \b n up to an even number. */
517 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
519 /** Used for offsets within a single page.
520 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
523 typedef uint16_t indx_t;
525 /** Default size of memory map.
526 * This is certainly too small for any actual applications. Apps should always set
527 * the size explicitly using #mdb_env_set_mapsize().
529 #define DEFAULT_MAPSIZE 1048576
531 /** @defgroup readers Reader Lock Table
532 * Readers don't acquire any locks for their data access. Instead, they
533 * simply record their transaction ID in the reader table. The reader
534 * mutex is needed just to find an empty slot in the reader table. The
535 * slot's address is saved in thread-specific data so that subsequent read
536 * transactions started by the same thread need no further locking to proceed.
538 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
540 * No reader table is used if the database is on a read-only filesystem, or
541 * if #MDB_NOLOCK is set.
543 * Since the database uses multi-version concurrency control, readers don't
544 * actually need any locking. This table is used to keep track of which
545 * readers are using data from which old transactions, so that we'll know
546 * when a particular old transaction is no longer in use. Old transactions
547 * that have discarded any data pages can then have those pages reclaimed
548 * for use by a later write transaction.
550 * The lock table is constructed such that reader slots are aligned with the
551 * processor's cache line size. Any slot is only ever used by one thread.
552 * This alignment guarantees that there will be no contention or cache
553 * thrashing as threads update their own slot info, and also eliminates
554 * any need for locking when accessing a slot.
556 * A writer thread will scan every slot in the table to determine the oldest
557 * outstanding reader transaction. Any freed pages older than this will be
558 * reclaimed by the writer. The writer doesn't use any locks when scanning
559 * this table. This means that there's no guarantee that the writer will
560 * see the most up-to-date reader info, but that's not required for correct
561 * operation - all we need is to know the upper bound on the oldest reader,
562 * we don't care at all about the newest reader. So the only consequence of
563 * reading stale information here is that old pages might hang around a
564 * while longer before being reclaimed. That's actually good anyway, because
565 * the longer we delay reclaiming old pages, the more likely it is that a
566 * string of contiguous pages can be found after coalescing old pages from
567 * many old transactions together.
570 /** Number of slots in the reader table.
571 * This value was chosen somewhat arbitrarily. 126 readers plus a
572 * couple mutexes fit exactly into 8KB on my development machine.
573 * Applications should set the table size using #mdb_env_set_maxreaders().
575 #define DEFAULT_READERS 126
577 /** The size of a CPU cache line in bytes. We want our lock structures
578 * aligned to this size to avoid false cache line sharing in the
580 * This value works for most CPUs. For Itanium this should be 128.
586 /** The information we store in a single slot of the reader table.
587 * In addition to a transaction ID, we also record the process and
588 * thread ID that owns a slot, so that we can detect stale information,
589 * e.g. threads or processes that went away without cleaning up.
590 * @note We currently don't check for stale records. We simply re-init
591 * the table when we know that we're the only process opening the
594 typedef struct MDB_rxbody {
595 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
596 * Multiple readers that start at the same time will probably have the
597 * same ID here. Again, it's not important to exclude them from
598 * anything; all we need to know is which version of the DB they
599 * started from so we can avoid overwriting any data used in that
600 * particular version.
602 volatile txnid_t mrb_txnid;
603 /** The process ID of the process owning this reader txn. */
604 volatile MDB_PID_T mrb_pid;
605 /** The thread ID of the thread owning this txn. */
606 volatile MDB_THR_T mrb_tid;
609 /** The actual reader record, with cacheline padding. */
610 typedef struct MDB_reader {
613 /** shorthand for mrb_txnid */
614 #define mr_txnid mru.mrx.mrb_txnid
615 #define mr_pid mru.mrx.mrb_pid
616 #define mr_tid mru.mrx.mrb_tid
617 /** cache line alignment */
618 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
622 /** The header for the reader table.
623 * The table resides in a memory-mapped file. (This is a different file
624 * than is used for the main database.)
626 * For POSIX the actual mutexes reside in the shared memory of this
627 * mapped file. On Windows, mutexes are named objects allocated by the
628 * kernel; we store the mutex names in this mapped file so that other
629 * processes can grab them. This same approach is also used on
630 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
631 * process-shared POSIX mutexes. For these cases where a named object
632 * is used, the object name is derived from a 64 bit FNV hash of the
633 * environment pathname. As such, naming collisions are extremely
634 * unlikely. If a collision occurs, the results are unpredictable.
636 typedef struct MDB_txbody {
637 /** Stamp identifying this as an LMDB file. It must be set
640 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
642 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
643 char mtb_rmname[MNAME_LEN];
645 /** Mutex protecting access to this table.
646 * This is the reader lock that #LOCK_MUTEX_R acquires.
648 pthread_mutex_t mtb_mutex;
650 /** The ID of the last transaction committed to the database.
651 * This is recorded here only for convenience; the value can always
652 * be determined by reading the main database meta pages.
654 volatile txnid_t mtb_txnid;
655 /** The number of slots that have been used in the reader table.
656 * This always records the maximum count, it is not decremented
657 * when readers release their slots.
659 volatile unsigned mtb_numreaders;
662 /** The actual reader table definition. */
663 typedef struct MDB_txninfo {
666 #define mti_magic mt1.mtb.mtb_magic
667 #define mti_format mt1.mtb.mtb_format
668 #define mti_mutex mt1.mtb.mtb_mutex
669 #define mti_rmname mt1.mtb.mtb_rmname
670 #define mti_txnid mt1.mtb.mtb_txnid
671 #define mti_numreaders mt1.mtb.mtb_numreaders
672 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
675 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
676 char mt2_wmname[MNAME_LEN];
677 #define mti_wmname mt2.mt2_wmname
679 pthread_mutex_t mt2_wmutex;
680 #define mti_wmutex mt2.mt2_wmutex
682 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
684 MDB_reader mti_readers[1];
687 /** Lockfile format signature: version, features and field layout */
688 #define MDB_LOCK_FORMAT \
690 ((MDB_LOCK_VERSION) \
691 /* Flags which describe functionality */ \
692 + (((MDB_PIDLOCK) != 0) << 16)))
695 /** Common header for all page types.
696 * Overflow records occupy a number of contiguous pages with no
697 * headers on any page after the first.
699 typedef struct MDB_page {
700 #define mp_pgno mp_p.p_pgno
701 #define mp_next mp_p.p_next
703 pgno_t p_pgno; /**< page number */
704 struct MDB_page *p_next; /**< for in-memory list of freed pages */
707 /** @defgroup mdb_page Page Flags
709 * Flags for the page headers.
712 #define P_BRANCH 0x01 /**< branch page */
713 #define P_LEAF 0x02 /**< leaf page */
714 #define P_OVERFLOW 0x04 /**< overflow page */
715 #define P_META 0x08 /**< meta page */
716 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
717 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
718 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
719 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
720 #define P_KEEP 0x8000 /**< leave this page alone during spill */
722 uint16_t mp_flags; /**< @ref mdb_page */
723 #define mp_lower mp_pb.pb.pb_lower
724 #define mp_upper mp_pb.pb.pb_upper
725 #define mp_pages mp_pb.pb_pages
728 indx_t pb_lower; /**< lower bound of free space */
729 indx_t pb_upper; /**< upper bound of free space */
731 uint32_t pb_pages; /**< number of overflow pages */
733 indx_t mp_ptrs[1]; /**< dynamic size */
736 /** Size of the page header, excluding dynamic data at the end */
737 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
739 /** Address of first usable data byte in a page, after the header */
740 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
742 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
743 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
745 /** Number of nodes on a page */
746 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
748 /** The amount of space remaining in the page */
749 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
751 /** The percentage of space used in the page, in tenths of a percent. */
752 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
753 ((env)->me_psize - PAGEHDRSZ))
754 /** The minimum page fill factor, in tenths of a percent.
755 * Pages emptier than this are candidates for merging.
757 #define FILL_THRESHOLD 250
759 /** Test if a page is a leaf page */
760 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
761 /** Test if a page is a LEAF2 page */
762 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
763 /** Test if a page is a branch page */
764 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
765 /** Test if a page is an overflow page */
766 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
767 /** Test if a page is a sub page */
768 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
770 /** The number of overflow pages needed to store the given size. */
771 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
773 /** Link in #MDB_txn.%mt_loose_pgs list */
774 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
776 /** Header for a single key/data pair within a page.
777 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
778 * We guarantee 2-byte alignment for 'MDB_node's.
780 typedef struct MDB_node {
781 /** lo and hi are used for data size on leaf nodes and for
782 * child pgno on branch nodes. On 64 bit platforms, flags
783 * is also used for pgno. (Branch nodes have no flags).
784 * They are in host byte order in case that lets some
785 * accesses be optimized into a 32-bit word access.
787 #if BYTE_ORDER == LITTLE_ENDIAN
788 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
790 unsigned short mn_hi, mn_lo;
792 /** @defgroup mdb_node Node Flags
794 * Flags for node headers.
797 #define F_BIGDATA 0x01 /**< data put on overflow page */
798 #define F_SUBDATA 0x02 /**< data is a sub-database */
799 #define F_DUPDATA 0x04 /**< data has duplicates */
801 /** valid flags for #mdb_node_add() */
802 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
805 unsigned short mn_flags; /**< @ref mdb_node */
806 unsigned short mn_ksize; /**< key size */
807 char mn_data[1]; /**< key and data are appended here */
810 /** Size of the node header, excluding dynamic data at the end */
811 #define NODESIZE offsetof(MDB_node, mn_data)
813 /** Bit position of top word in page number, for shifting mn_flags */
814 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
816 /** Size of a node in a branch page with a given key.
817 * This is just the node header plus the key, there is no data.
819 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
821 /** Size of a node in a leaf page with a given key and data.
822 * This is node header plus key plus data size.
824 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
826 /** Address of node \b i in page \b p */
827 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
829 /** Address of the key for the node */
830 #define NODEKEY(node) (void *)((node)->mn_data)
832 /** Address of the data for a node */
833 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
835 /** Get the page number pointed to by a branch node */
836 #define NODEPGNO(node) \
837 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
838 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
839 /** Set the page number in a branch node */
840 #define SETPGNO(node,pgno) do { \
841 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
842 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
844 /** Get the size of the data in a leaf node */
845 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
846 /** Set the size of the data for a leaf node */
847 #define SETDSZ(node,size) do { \
848 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
849 /** The size of a key in a node */
850 #define NODEKSZ(node) ((node)->mn_ksize)
852 /** Copy a page number from src to dst */
854 #define COPY_PGNO(dst,src) dst = src
856 #if SIZE_MAX > 4294967295UL
857 #define COPY_PGNO(dst,src) do { \
858 unsigned short *s, *d; \
859 s = (unsigned short *)&(src); \
860 d = (unsigned short *)&(dst); \
867 #define COPY_PGNO(dst,src) do { \
868 unsigned short *s, *d; \
869 s = (unsigned short *)&(src); \
870 d = (unsigned short *)&(dst); \
876 /** The address of a key in a LEAF2 page.
877 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
878 * There are no node headers, keys are stored contiguously.
880 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
882 /** Set the \b node's key into \b keyptr, if requested. */
883 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
884 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
886 /** Set the \b node's key into \b key. */
887 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
889 /** Information about a single database in the environment. */
890 typedef struct MDB_db {
891 uint32_t md_pad; /**< also ksize for LEAF2 pages */
892 uint16_t md_flags; /**< @ref mdb_dbi_open */
893 uint16_t md_depth; /**< depth of this tree */
894 pgno_t md_branch_pages; /**< number of internal pages */
895 pgno_t md_leaf_pages; /**< number of leaf pages */
896 pgno_t md_overflow_pages; /**< number of overflow pages */
897 size_t md_entries; /**< number of data items */
898 pgno_t md_root; /**< the root page of this tree */
901 /** mdb_dbi_open flags */
902 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
903 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
904 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
905 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
907 /** Handle for the DB used to track free pages. */
909 /** Handle for the default DB. */
912 /** Meta page content.
913 * A meta page is the start point for accessing a database snapshot.
914 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
916 typedef struct MDB_meta {
917 /** Stamp identifying this as an LMDB file. It must be set
920 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
922 void *mm_address; /**< address for fixed mapping */
923 size_t mm_mapsize; /**< size of mmap region */
924 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
925 /** The size of pages used in this DB */
926 #define mm_psize mm_dbs[0].md_pad
927 /** Any persistent environment flags. @ref mdb_env */
928 #define mm_flags mm_dbs[0].md_flags
929 pgno_t mm_last_pg; /**< last used page in file */
930 volatile txnid_t mm_txnid; /**< txnid that committed this page */
933 /** Buffer for a stack-allocated meta page.
934 * The members define size and alignment, and silence type
935 * aliasing warnings. They are not used directly; that could
936 * mean incorrectly using several union members in parallel.
938 typedef union MDB_metabuf {
941 char mm_pad[PAGEHDRSZ];
946 /** Auxiliary DB info.
947 * The information here is mostly static/read-only. There is
948 * only a single copy of this record in the environment.
950 typedef struct MDB_dbx {
951 MDB_val md_name; /**< name of the database */
952 MDB_cmp_func *md_cmp; /**< function for comparing keys */
953 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
954 MDB_rel_func *md_rel; /**< user relocate function */
955 void *md_relctx; /**< user-provided context for md_rel */
958 /** A database transaction.
959 * Every operation requires a transaction handle.
962 MDB_txn *mt_parent; /**< parent of a nested txn */
963 MDB_txn *mt_child; /**< nested txn under this txn */
964 pgno_t mt_next_pgno; /**< next unallocated page */
965 /** The ID of this transaction. IDs are integers incrementing from 1.
966 * Only committed write transactions increment the ID. If a transaction
967 * aborts, the ID may be re-used by the next writer.
970 MDB_env *mt_env; /**< the DB environment */
971 /** The list of pages that became unused during this transaction.
974 /** The list of loose pages that became unused and may be reused
975 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
977 MDB_page *mt_loose_pgs;
978 /* #Number of loose pages (#mt_loose_pgs) */
980 /** The sorted list of dirty pages we temporarily wrote to disk
981 * because the dirty list was full. page numbers in here are
982 * shifted left by 1, deleted slots have the LSB set.
984 MDB_IDL mt_spill_pgs;
986 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
988 /** For read txns: This thread/txn's reader table slot, or NULL. */
991 /** Array of records for each DB known in the environment. */
993 /** Array of MDB_db records for each known DB */
995 /** Array of sequence numbers for each DB handle */
996 unsigned int *mt_dbiseqs;
997 /** @defgroup mt_dbflag Transaction DB Flags
1001 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
1002 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
1003 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
1004 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1006 /** In write txns, array of cursors for each DB */
1007 MDB_cursor **mt_cursors;
1008 /** Array of flags for each DB */
1009 unsigned char *mt_dbflags;
1010 /** Number of DB records in use. This number only ever increments;
1011 * we don't decrement it when individual DB handles are closed.
1015 /** @defgroup mdb_txn Transaction Flags
1019 /** #mdb_txn_begin() flags */
1020 #define MDB_TXN_BEGIN_FLAGS MDB_RDONLY
1021 #define MDB_TXN_RDONLY MDB_RDONLY /**< read-only transaction */
1022 /* internal txn flags */
1023 #define MDB_TXN_WRITEMAP MDB_WRITEMAP /**< copy of #MDB_env flag in writers */
1024 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1025 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1026 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1028 unsigned int mt_flags; /**< @ref mdb_txn */
1029 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1030 * Includes ancestor txns' dirty pages not hidden by other txns'
1031 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1032 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1034 unsigned int mt_dirty_room;
1037 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1038 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1039 * raise this on a 64 bit machine.
1041 #define CURSOR_STACK 32
1045 /** Cursors are used for all DB operations.
1046 * A cursor holds a path of (page pointer, key index) from the DB
1047 * root to a position in the DB, plus other state. #MDB_DUPSORT
1048 * cursors include an xcursor to the current data item. Write txns
1049 * track their cursors and keep them up to date when data moves.
1050 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1051 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1054 /** Next cursor on this DB in this txn */
1055 MDB_cursor *mc_next;
1056 /** Backup of the original cursor if this cursor is a shadow */
1057 MDB_cursor *mc_backup;
1058 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1059 struct MDB_xcursor *mc_xcursor;
1060 /** The transaction that owns this cursor */
1062 /** The database handle this cursor operates on */
1064 /** The database record for this cursor */
1066 /** The database auxiliary record for this cursor */
1068 /** The @ref mt_dbflag for this database */
1069 unsigned char *mc_dbflag;
1070 unsigned short mc_snum; /**< number of pushed pages */
1071 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1072 /** @defgroup mdb_cursor Cursor Flags
1074 * Cursor state flags.
1077 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1078 #define C_EOF 0x02 /**< No more data */
1079 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1080 #define C_DEL 0x08 /**< last op was a cursor_del */
1081 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1082 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1084 unsigned int mc_flags; /**< @ref mdb_cursor */
1085 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1086 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1089 /** Context for sorted-dup records.
1090 * We could have gone to a fully recursive design, with arbitrarily
1091 * deep nesting of sub-databases. But for now we only handle these
1092 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1094 typedef struct MDB_xcursor {
1095 /** A sub-cursor for traversing the Dup DB */
1096 MDB_cursor mx_cursor;
1097 /** The database record for this Dup DB */
1099 /** The auxiliary DB record for this Dup DB */
1101 /** The @ref mt_dbflag for this Dup DB */
1102 unsigned char mx_dbflag;
1105 /** State of FreeDB old pages, stored in the MDB_env */
1106 typedef struct MDB_pgstate {
1107 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1108 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1111 /** The database environment. */
1113 HANDLE me_fd; /**< The main data file */
1114 HANDLE me_lfd; /**< The lock file */
1115 HANDLE me_mfd; /**< just for writing the meta pages */
1116 /** Failed to update the meta page. Probably an I/O error. */
1117 #define MDB_FATAL_ERROR 0x80000000U
1118 /** Some fields are initialized. */
1119 #define MDB_ENV_ACTIVE 0x20000000U
1120 /** me_txkey is set */
1121 #define MDB_ENV_TXKEY 0x10000000U
1122 /** fdatasync is unreliable */
1123 #define MDB_FSYNCONLY 0x08000000U
1124 uint32_t me_flags; /**< @ref mdb_env */
1125 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1126 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1127 unsigned int me_maxreaders; /**< size of the reader table */
1128 /** Max #MDB_txninfo.%mti_numreaders of interest to #mdb_env_close() */
1129 volatile int me_close_readers;
1130 MDB_dbi me_numdbs; /**< number of DBs opened */
1131 MDB_dbi me_maxdbs; /**< size of the DB table */
1132 MDB_PID_T me_pid; /**< process ID of this env */
1133 char *me_path; /**< path to the DB files */
1134 char *me_map; /**< the memory map of the data file */
1135 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1136 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1137 void *me_pbuf; /**< scratch area for DUPSORT put() */
1138 MDB_txn *me_txn; /**< current write transaction */
1139 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1140 size_t me_mapsize; /**< size of the data memory map */
1141 off_t me_size; /**< current file size */
1142 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1143 MDB_dbx *me_dbxs; /**< array of static DB info */
1144 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1145 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1146 pthread_key_t me_txkey; /**< thread-key for readers */
1147 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1148 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1149 # define me_pglast me_pgstate.mf_pglast
1150 # define me_pghead me_pgstate.mf_pghead
1151 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1152 /** IDL of pages that became unused in a write txn */
1153 MDB_IDL me_free_pgs;
1154 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1155 MDB_ID2L me_dirty_list;
1156 /** Max number of freelist items that can fit in a single overflow page */
1158 /** Max size of a node on a page */
1159 unsigned int me_nodemax;
1160 #if !(MDB_MAXKEYSIZE)
1161 unsigned int me_maxkey; /**< max size of a key */
1163 int me_live_reader; /**< have liveness lock in reader table */
1165 int me_pidquery; /**< Used in OpenProcess */
1166 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1168 #elif defined(MDB_USE_POSIX_SEM)
1169 sem_t *me_rmutex; /* Shared mutexes are not supported */
1172 void *me_userctx; /**< User-settable context */
1173 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1176 /** Nested transaction */
1177 typedef struct MDB_ntxn {
1178 MDB_txn mnt_txn; /**< the transaction */
1179 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1182 /** max number of pages to commit in one writev() call */
1183 #define MDB_COMMIT_PAGES 64
1184 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1185 #undef MDB_COMMIT_PAGES
1186 #define MDB_COMMIT_PAGES IOV_MAX
1189 /** max bytes to write in one call */
1190 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1192 /** Check \b txn and \b dbi arguments to a function */
1193 #define TXN_DBI_EXIST(txn, dbi) \
1194 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1196 /** Check for misused \b dbi handles */
1197 #define TXN_DBI_CHANGED(txn, dbi) \
1198 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1200 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1201 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1202 static int mdb_page_touch(MDB_cursor *mc);
1204 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1205 static int mdb_page_search_root(MDB_cursor *mc,
1206 MDB_val *key, int modify);
1207 #define MDB_PS_MODIFY 1
1208 #define MDB_PS_ROOTONLY 2
1209 #define MDB_PS_FIRST 4
1210 #define MDB_PS_LAST 8
1211 static int mdb_page_search(MDB_cursor *mc,
1212 MDB_val *key, int flags);
1213 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1215 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1216 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1217 pgno_t newpgno, unsigned int nflags);
1219 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1220 static int mdb_env_pick_meta(const MDB_env *env);
1221 static int mdb_env_write_meta(MDB_txn *txn);
1222 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1223 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1225 static void mdb_env_close0(MDB_env *env, int excl);
1227 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1228 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1229 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1230 static void mdb_node_del(MDB_cursor *mc, int ksize);
1231 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1232 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1233 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1234 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1235 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1237 static int mdb_rebalance(MDB_cursor *mc);
1238 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1240 static void mdb_cursor_pop(MDB_cursor *mc);
1241 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1243 static int mdb_cursor_del0(MDB_cursor *mc);
1244 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1245 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1246 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1247 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1248 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1250 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1251 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1253 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1254 static void mdb_xcursor_init0(MDB_cursor *mc);
1255 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1256 static void mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int force);
1258 static int mdb_drop0(MDB_cursor *mc, int subs);
1259 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1262 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1265 /** Compare two items pointing at size_t's of unknown alignment. */
1266 #ifdef MISALIGNED_OK
1267 # define mdb_cmp_clong mdb_cmp_long
1269 # define mdb_cmp_clong mdb_cmp_cint
1273 static SECURITY_DESCRIPTOR mdb_null_sd;
1274 static SECURITY_ATTRIBUTES mdb_all_sa;
1275 static int mdb_sec_inited;
1278 /** Return the library version info. */
1280 mdb_version(int *major, int *minor, int *patch)
1282 if (major) *major = MDB_VERSION_MAJOR;
1283 if (minor) *minor = MDB_VERSION_MINOR;
1284 if (patch) *patch = MDB_VERSION_PATCH;
1285 return MDB_VERSION_STRING;
1288 /** Table of descriptions for LMDB @ref errors */
1289 static char *const mdb_errstr[] = {
1290 "MDB_KEYEXIST: Key/data pair already exists",
1291 "MDB_NOTFOUND: No matching key/data pair found",
1292 "MDB_PAGE_NOTFOUND: Requested page not found",
1293 "MDB_CORRUPTED: Located page was wrong type",
1294 "MDB_PANIC: Update of meta page failed",
1295 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1296 "MDB_INVALID: File is not an LMDB file",
1297 "MDB_MAP_FULL: Environment mapsize limit reached",
1298 "MDB_DBS_FULL: Environment maxdbs limit reached",
1299 "MDB_READERS_FULL: Environment maxreaders limit reached",
1300 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1301 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1302 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1303 "MDB_PAGE_FULL: Internal error - page has no more space",
1304 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1305 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1306 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1307 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1308 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1309 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1313 mdb_strerror(int err)
1316 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1317 * This works as long as no function between the call to mdb_strerror
1318 * and the actual use of the message uses more than 4K of stack.
1321 char buf[1024], *ptr = buf;
1325 return ("Successful return: 0");
1327 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1328 i = err - MDB_KEYEXIST;
1329 return mdb_errstr[i];
1333 /* These are the C-runtime error codes we use. The comment indicates
1334 * their numeric value, and the Win32 error they would correspond to
1335 * if the error actually came from a Win32 API. A major mess, we should
1336 * have used LMDB-specific error codes for everything.
1339 case ENOENT: /* 2, FILE_NOT_FOUND */
1340 case EIO: /* 5, ACCESS_DENIED */
1341 case ENOMEM: /* 12, INVALID_ACCESS */
1342 case EACCES: /* 13, INVALID_DATA */
1343 case EBUSY: /* 16, CURRENT_DIRECTORY */
1344 case EINVAL: /* 22, BAD_COMMAND */
1345 case ENOSPC: /* 28, OUT_OF_PAPER */
1346 return strerror(err);
1351 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1352 FORMAT_MESSAGE_IGNORE_INSERTS,
1353 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1356 return strerror(err);
1360 /** assert(3) variant in cursor context */
1361 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1362 /** assert(3) variant in transaction context */
1363 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1364 /** assert(3) variant in environment context */
1365 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1368 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1369 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1372 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1373 const char *func, const char *file, int line)
1376 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1377 file, line, expr_txt, func);
1378 if (env->me_assert_func)
1379 env->me_assert_func(env, buf);
1380 fprintf(stderr, "%s\n", buf);
1384 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1388 /** Return the page number of \b mp which may be sub-page, for debug output */
1390 mdb_dbg_pgno(MDB_page *mp)
1393 COPY_PGNO(ret, mp->mp_pgno);
1397 /** Display a key in hexadecimal and return the address of the result.
1398 * @param[in] key the key to display
1399 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1400 * @return The key in hexadecimal form.
1403 mdb_dkey(MDB_val *key, char *buf)
1406 unsigned char *c = key->mv_data;
1412 if (key->mv_size > DKBUF_MAXKEYSIZE)
1413 return "MDB_MAXKEYSIZE";
1414 /* may want to make this a dynamic check: if the key is mostly
1415 * printable characters, print it as-is instead of converting to hex.
1419 for (i=0; i<key->mv_size; i++)
1420 ptr += sprintf(ptr, "%02x", *c++);
1422 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1428 mdb_leafnode_type(MDB_node *n)
1430 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1431 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1432 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1435 /** Display all the keys in the page. */
1437 mdb_page_list(MDB_page *mp)
1439 pgno_t pgno = mdb_dbg_pgno(mp);
1440 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1442 unsigned int i, nkeys, nsize, total = 0;
1446 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1447 case P_BRANCH: type = "Branch page"; break;
1448 case P_LEAF: type = "Leaf page"; break;
1449 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1450 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1451 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1453 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1454 pgno, mp->mp_pages, state);
1457 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1458 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1461 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1465 nkeys = NUMKEYS(mp);
1466 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1468 for (i=0; i<nkeys; i++) {
1469 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1470 key.mv_size = nsize = mp->mp_pad;
1471 key.mv_data = LEAF2KEY(mp, i, nsize);
1473 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1476 node = NODEPTR(mp, i);
1477 key.mv_size = node->mn_ksize;
1478 key.mv_data = node->mn_data;
1479 nsize = NODESIZE + key.mv_size;
1480 if (IS_BRANCH(mp)) {
1481 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1485 if (F_ISSET(node->mn_flags, F_BIGDATA))
1486 nsize += sizeof(pgno_t);
1488 nsize += NODEDSZ(node);
1490 nsize += sizeof(indx_t);
1491 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1492 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1494 total = EVEN(total);
1496 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1497 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1501 mdb_cursor_chk(MDB_cursor *mc)
1507 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1508 for (i=0; i<mc->mc_top; i++) {
1510 node = NODEPTR(mp, mc->mc_ki[i]);
1511 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1514 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1520 /** Count all the pages in each DB and in the freelist
1521 * and make sure it matches the actual number of pages
1523 * All named DBs must be open for a correct count.
1525 static void mdb_audit(MDB_txn *txn)
1529 MDB_ID freecount, count;
1534 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1535 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1536 freecount += *(MDB_ID *)data.mv_data;
1537 mdb_tassert(txn, rc == MDB_NOTFOUND);
1540 for (i = 0; i<txn->mt_numdbs; i++) {
1542 if (!(txn->mt_dbflags[i] & DB_VALID))
1544 mdb_cursor_init(&mc, txn, i, &mx);
1545 if (txn->mt_dbs[i].md_root == P_INVALID)
1547 count += txn->mt_dbs[i].md_branch_pages +
1548 txn->mt_dbs[i].md_leaf_pages +
1549 txn->mt_dbs[i].md_overflow_pages;
1550 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1551 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1552 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1555 mp = mc.mc_pg[mc.mc_top];
1556 for (j=0; j<NUMKEYS(mp); j++) {
1557 MDB_node *leaf = NODEPTR(mp, j);
1558 if (leaf->mn_flags & F_SUBDATA) {
1560 memcpy(&db, NODEDATA(leaf), sizeof(db));
1561 count += db.md_branch_pages + db.md_leaf_pages +
1562 db.md_overflow_pages;
1566 mdb_tassert(txn, rc == MDB_NOTFOUND);
1569 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1570 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1571 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1577 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1579 return txn->mt_dbxs[dbi].md_cmp(a, b);
1583 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1585 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1586 #if UINT_MAX < SIZE_MAX
1587 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1588 dcmp = mdb_cmp_clong;
1593 /** Allocate memory for a page.
1594 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1597 mdb_page_malloc(MDB_txn *txn, unsigned num)
1599 MDB_env *env = txn->mt_env;
1600 MDB_page *ret = env->me_dpages;
1601 size_t psize = env->me_psize, sz = psize, off;
1602 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1603 * For a single page alloc, we init everything after the page header.
1604 * For multi-page, we init the final page; if the caller needed that
1605 * many pages they will be filling in at least up to the last page.
1609 VGMEMP_ALLOC(env, ret, sz);
1610 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1611 env->me_dpages = ret->mp_next;
1614 psize -= off = PAGEHDRSZ;
1619 if ((ret = malloc(sz)) != NULL) {
1620 VGMEMP_ALLOC(env, ret, sz);
1621 if (!(env->me_flags & MDB_NOMEMINIT)) {
1622 memset((char *)ret + off, 0, psize);
1626 txn->mt_flags |= MDB_TXN_ERROR;
1630 /** Free a single page.
1631 * Saves single pages to a list, for future reuse.
1632 * (This is not used for multi-page overflow pages.)
1635 mdb_page_free(MDB_env *env, MDB_page *mp)
1637 mp->mp_next = env->me_dpages;
1638 VGMEMP_FREE(env, mp);
1639 env->me_dpages = mp;
1642 /** Free a dirty page */
1644 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1646 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1647 mdb_page_free(env, dp);
1649 /* large pages just get freed directly */
1650 VGMEMP_FREE(env, dp);
1655 /** Return all dirty pages to dpage list */
1657 mdb_dlist_free(MDB_txn *txn)
1659 MDB_env *env = txn->mt_env;
1660 MDB_ID2L dl = txn->mt_u.dirty_list;
1661 unsigned i, n = dl[0].mid;
1663 for (i = 1; i <= n; i++) {
1664 mdb_dpage_free(env, dl[i].mptr);
1669 /** Loosen or free a single page.
1670 * Saves single pages to a list for future reuse
1671 * in this same txn. It has been pulled from the freeDB
1672 * and already resides on the dirty list, but has been
1673 * deleted. Use these pages first before pulling again
1676 * If the page wasn't dirtied in this txn, just add it
1677 * to this txn's free list.
1680 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1683 pgno_t pgno = mp->mp_pgno;
1684 MDB_txn *txn = mc->mc_txn;
1686 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1687 if (txn->mt_parent) {
1688 MDB_ID2 *dl = txn->mt_u.dirty_list;
1689 /* If txn has a parent, make sure the page is in our
1693 unsigned x = mdb_mid2l_search(dl, pgno);
1694 if (x <= dl[0].mid && dl[x].mid == pgno) {
1695 if (mp != dl[x].mptr) { /* bad cursor? */
1696 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1697 txn->mt_flags |= MDB_TXN_ERROR;
1698 return MDB_CORRUPTED;
1705 /* no parent txn, so it's just ours */
1710 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1712 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1713 txn->mt_loose_pgs = mp;
1714 txn->mt_loose_count++;
1715 mp->mp_flags |= P_LOOSE;
1717 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1725 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1726 * @param[in] mc A cursor handle for the current operation.
1727 * @param[in] pflags Flags of the pages to update:
1728 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1729 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1730 * @return 0 on success, non-zero on failure.
1733 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1735 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1736 MDB_txn *txn = mc->mc_txn;
1742 int rc = MDB_SUCCESS, level;
1744 /* Mark pages seen by cursors */
1745 if (mc->mc_flags & C_UNTRACK)
1746 mc = NULL; /* will find mc in mt_cursors */
1747 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1748 for (; mc; mc=mc->mc_next) {
1749 if (!(mc->mc_flags & C_INITIALIZED))
1751 for (m3 = mc;; m3 = &mx->mx_cursor) {
1753 for (j=0; j<m3->mc_snum; j++) {
1755 if ((mp->mp_flags & Mask) == pflags)
1756 mp->mp_flags ^= P_KEEP;
1758 mx = m3->mc_xcursor;
1759 /* Proceed to mx if it is at a sub-database */
1760 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1762 if (! (mp && (mp->mp_flags & P_LEAF)))
1764 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1765 if (!(leaf->mn_flags & F_SUBDATA))
1774 /* Mark dirty root pages */
1775 for (i=0; i<txn->mt_numdbs; i++) {
1776 if (txn->mt_dbflags[i] & DB_DIRTY) {
1777 pgno_t pgno = txn->mt_dbs[i].md_root;
1778 if (pgno == P_INVALID)
1780 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1782 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1783 dp->mp_flags ^= P_KEEP;
1791 static int mdb_page_flush(MDB_txn *txn, int keep);
1793 /** Spill pages from the dirty list back to disk.
1794 * This is intended to prevent running into #MDB_TXN_FULL situations,
1795 * but note that they may still occur in a few cases:
1796 * 1) our estimate of the txn size could be too small. Currently this
1797 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1798 * 2) child txns may run out of space if their parents dirtied a
1799 * lot of pages and never spilled them. TODO: we probably should do
1800 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1801 * the parent's dirty_room is below a given threshold.
1803 * Otherwise, if not using nested txns, it is expected that apps will
1804 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1805 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1806 * If the txn never references them again, they can be left alone.
1807 * If the txn only reads them, they can be used without any fuss.
1808 * If the txn writes them again, they can be dirtied immediately without
1809 * going thru all of the work of #mdb_page_touch(). Such references are
1810 * handled by #mdb_page_unspill().
1812 * Also note, we never spill DB root pages, nor pages of active cursors,
1813 * because we'll need these back again soon anyway. And in nested txns,
1814 * we can't spill a page in a child txn if it was already spilled in a
1815 * parent txn. That would alter the parent txns' data even though
1816 * the child hasn't committed yet, and we'd have no way to undo it if
1817 * the child aborted.
1819 * @param[in] m0 cursor A cursor handle identifying the transaction and
1820 * database for which we are checking space.
1821 * @param[in] key For a put operation, the key being stored.
1822 * @param[in] data For a put operation, the data being stored.
1823 * @return 0 on success, non-zero on failure.
1826 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1828 MDB_txn *txn = m0->mc_txn;
1830 MDB_ID2L dl = txn->mt_u.dirty_list;
1831 unsigned int i, j, need;
1834 if (m0->mc_flags & C_SUB)
1837 /* Estimate how much space this op will take */
1838 i = m0->mc_db->md_depth;
1839 /* Named DBs also dirty the main DB */
1840 if (m0->mc_dbi > MAIN_DBI)
1841 i += txn->mt_dbs[MAIN_DBI].md_depth;
1842 /* For puts, roughly factor in the key+data size */
1844 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1845 i += i; /* double it for good measure */
1848 if (txn->mt_dirty_room > i)
1851 if (!txn->mt_spill_pgs) {
1852 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1853 if (!txn->mt_spill_pgs)
1856 /* purge deleted slots */
1857 MDB_IDL sl = txn->mt_spill_pgs;
1858 unsigned int num = sl[0];
1860 for (i=1; i<=num; i++) {
1867 /* Preserve pages which may soon be dirtied again */
1868 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1871 /* Less aggressive spill - we originally spilled the entire dirty list,
1872 * with a few exceptions for cursor pages and DB root pages. But this
1873 * turns out to be a lot of wasted effort because in a large txn many
1874 * of those pages will need to be used again. So now we spill only 1/8th
1875 * of the dirty pages. Testing revealed this to be a good tradeoff,
1876 * better than 1/2, 1/4, or 1/10.
1878 if (need < MDB_IDL_UM_MAX / 8)
1879 need = MDB_IDL_UM_MAX / 8;
1881 /* Save the page IDs of all the pages we're flushing */
1882 /* flush from the tail forward, this saves a lot of shifting later on. */
1883 for (i=dl[0].mid; i && need; i--) {
1884 MDB_ID pn = dl[i].mid << 1;
1886 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1888 /* Can't spill twice, make sure it's not already in a parent's
1891 if (txn->mt_parent) {
1893 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1894 if (tx2->mt_spill_pgs) {
1895 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1896 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1897 dp->mp_flags |= P_KEEP;
1905 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1909 mdb_midl_sort(txn->mt_spill_pgs);
1911 /* Flush the spilled part of dirty list */
1912 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1915 /* Reset any dirty pages we kept that page_flush didn't see */
1916 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1919 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1923 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1925 mdb_find_oldest(MDB_txn *txn)
1928 txnid_t mr, oldest = txn->mt_txnid - 1;
1929 if (txn->mt_env->me_txns) {
1930 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1931 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1942 /** Add a page to the txn's dirty list */
1944 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1947 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1949 if (txn->mt_flags & MDB_TXN_WRITEMAP) {
1950 insert = mdb_mid2l_append;
1952 insert = mdb_mid2l_insert;
1954 mid.mid = mp->mp_pgno;
1956 rc = insert(txn->mt_u.dirty_list, &mid);
1957 mdb_tassert(txn, rc == 0);
1958 txn->mt_dirty_room--;
1961 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1962 * me_pghead and mt_next_pgno.
1964 * If there are free pages available from older transactions, they
1965 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1966 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1967 * and move me_pglast to say which records were consumed. Only this
1968 * function can create me_pghead and move me_pglast/mt_next_pgno.
1969 * @param[in] mc cursor A cursor handle identifying the transaction and
1970 * database for which we are allocating.
1971 * @param[in] num the number of pages to allocate.
1972 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1973 * will always be satisfied by a single contiguous chunk of memory.
1974 * @return 0 on success, non-zero on failure.
1977 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1979 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1980 /* Get at most <Max_retries> more freeDB records once me_pghead
1981 * has enough pages. If not enough, use new pages from the map.
1982 * If <Paranoid> and mc is updating the freeDB, only get new
1983 * records if me_pghead is empty. Then the freelist cannot play
1984 * catch-up with itself by growing while trying to save it.
1986 enum { Paranoid = 1, Max_retries = 500 };
1988 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1990 int rc, retry = num * 60;
1991 MDB_txn *txn = mc->mc_txn;
1992 MDB_env *env = txn->mt_env;
1993 pgno_t pgno, *mop = env->me_pghead;
1994 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1996 txnid_t oldest = 0, last;
2001 /* If there are any loose pages, just use them */
2002 if (num == 1 && txn->mt_loose_pgs) {
2003 np = txn->mt_loose_pgs;
2004 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
2005 txn->mt_loose_count--;
2006 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2014 /* If our dirty list is already full, we can't do anything */
2015 if (txn->mt_dirty_room == 0) {
2020 for (op = MDB_FIRST;; op = MDB_NEXT) {
2025 /* Seek a big enough contiguous page range. Prefer
2026 * pages at the tail, just truncating the list.
2032 if (mop[i-n2] == pgno+n2)
2039 if (op == MDB_FIRST) { /* 1st iteration */
2040 /* Prepare to fetch more and coalesce */
2041 last = env->me_pglast;
2042 oldest = env->me_pgoldest;
2043 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2046 key.mv_data = &last; /* will look up last+1 */
2047 key.mv_size = sizeof(last);
2049 if (Paranoid && mc->mc_dbi == FREE_DBI)
2052 if (Paranoid && retry < 0 && mop_len)
2056 /* Do not fetch more if the record will be too recent */
2057 if (oldest <= last) {
2059 oldest = mdb_find_oldest(txn);
2060 env->me_pgoldest = oldest;
2066 rc = mdb_cursor_get(&m2, &key, NULL, op);
2068 if (rc == MDB_NOTFOUND)
2072 last = *(txnid_t*)key.mv_data;
2073 if (oldest <= last) {
2075 oldest = mdb_find_oldest(txn);
2076 env->me_pgoldest = oldest;
2082 np = m2.mc_pg[m2.mc_top];
2083 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2084 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2087 idl = (MDB_ID *) data.mv_data;
2090 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2095 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2097 mop = env->me_pghead;
2099 env->me_pglast = last;
2101 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2102 last, txn->mt_dbs[FREE_DBI].md_root, i));
2104 DPRINTF(("IDL %"Z"u", idl[j]));
2106 /* Merge in descending sorted order */
2107 mdb_midl_xmerge(mop, idl);
2111 /* Use new pages from the map when nothing suitable in the freeDB */
2113 pgno = txn->mt_next_pgno;
2114 if (pgno + num >= env->me_maxpg) {
2115 DPUTS("DB size maxed out");
2121 if (env->me_flags & MDB_WRITEMAP) {
2122 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2124 if (!(np = mdb_page_malloc(txn, num))) {
2130 mop[0] = mop_len -= num;
2131 /* Move any stragglers down */
2132 for (j = i-num; j < mop_len; )
2133 mop[++j] = mop[++i];
2135 txn->mt_next_pgno = pgno + num;
2138 mdb_page_dirty(txn, np);
2144 txn->mt_flags |= MDB_TXN_ERROR;
2148 /** Copy the used portions of a non-overflow page.
2149 * @param[in] dst page to copy into
2150 * @param[in] src page to copy from
2151 * @param[in] psize size of a page
2154 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2156 enum { Align = sizeof(pgno_t) };
2157 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2159 /* If page isn't full, just copy the used portion. Adjust
2160 * alignment so memcpy may copy words instead of bytes.
2162 if ((unused &= -Align) && !IS_LEAF2(src)) {
2163 upper = (upper + PAGEBASE) & -Align;
2164 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2165 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2168 memcpy(dst, src, psize - unused);
2172 /** Pull a page off the txn's spill list, if present.
2173 * If a page being referenced was spilled to disk in this txn, bring
2174 * it back and make it dirty/writable again.
2175 * @param[in] txn the transaction handle.
2176 * @param[in] mp the page being referenced. It must not be dirty.
2177 * @param[out] ret the writable page, if any. ret is unchanged if
2178 * mp wasn't spilled.
2181 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2183 MDB_env *env = txn->mt_env;
2186 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2188 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2189 if (!tx2->mt_spill_pgs)
2191 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2192 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2195 if (txn->mt_dirty_room == 0)
2196 return MDB_TXN_FULL;
2197 if (IS_OVERFLOW(mp))
2201 if (env->me_flags & MDB_WRITEMAP) {
2204 np = mdb_page_malloc(txn, num);
2208 memcpy(np, mp, num * env->me_psize);
2210 mdb_page_copy(np, mp, env->me_psize);
2213 /* If in current txn, this page is no longer spilled.
2214 * If it happens to be the last page, truncate the spill list.
2215 * Otherwise mark it as deleted by setting the LSB.
2217 if (x == txn->mt_spill_pgs[0])
2218 txn->mt_spill_pgs[0]--;
2220 txn->mt_spill_pgs[x] |= 1;
2221 } /* otherwise, if belonging to a parent txn, the
2222 * page remains spilled until child commits
2225 mdb_page_dirty(txn, np);
2226 np->mp_flags |= P_DIRTY;
2234 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2235 * @param[in] mc cursor pointing to the page to be touched
2236 * @return 0 on success, non-zero on failure.
2239 mdb_page_touch(MDB_cursor *mc)
2241 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2242 MDB_txn *txn = mc->mc_txn;
2243 MDB_cursor *m2, *m3;
2247 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2248 if (txn->mt_flags & MDB_TXN_SPILLS) {
2250 rc = mdb_page_unspill(txn, mp, &np);
2256 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2257 (rc = mdb_page_alloc(mc, 1, &np)))
2260 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2261 mp->mp_pgno, pgno));
2262 mdb_cassert(mc, mp->mp_pgno != pgno);
2263 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2264 /* Update the parent page, if any, to point to the new page */
2266 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2267 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2268 SETPGNO(node, pgno);
2270 mc->mc_db->md_root = pgno;
2272 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2273 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2275 /* If txn has a parent, make sure the page is in our
2279 unsigned x = mdb_mid2l_search(dl, pgno);
2280 if (x <= dl[0].mid && dl[x].mid == pgno) {
2281 if (mp != dl[x].mptr) { /* bad cursor? */
2282 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2283 txn->mt_flags |= MDB_TXN_ERROR;
2284 return MDB_CORRUPTED;
2289 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2291 np = mdb_page_malloc(txn, 1);
2296 rc = mdb_mid2l_insert(dl, &mid);
2297 mdb_cassert(mc, rc == 0);
2302 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2304 np->mp_flags |= P_DIRTY;
2307 /* Adjust cursors pointing to mp */
2308 mc->mc_pg[mc->mc_top] = np;
2309 m2 = txn->mt_cursors[mc->mc_dbi];
2310 if (mc->mc_flags & C_SUB) {
2311 for (; m2; m2=m2->mc_next) {
2312 m3 = &m2->mc_xcursor->mx_cursor;
2313 if (m3->mc_snum < mc->mc_snum) continue;
2314 if (m3->mc_pg[mc->mc_top] == mp)
2315 m3->mc_pg[mc->mc_top] = np;
2318 for (; m2; m2=m2->mc_next) {
2319 if (m2->mc_snum < mc->mc_snum) continue;
2320 if (m2->mc_pg[mc->mc_top] == mp) {
2321 m2->mc_pg[mc->mc_top] = np;
2322 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2324 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2326 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2327 if (!(leaf->mn_flags & F_SUBDATA))
2328 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2336 txn->mt_flags |= MDB_TXN_ERROR;
2341 mdb_env_sync(MDB_env *env, int force)
2344 if (env->me_flags & MDB_RDONLY)
2346 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2347 if (env->me_flags & MDB_WRITEMAP) {
2348 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2349 ? MS_ASYNC : MS_SYNC;
2350 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2353 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2357 #ifdef BROKEN_FDATASYNC
2358 if (env->me_flags & MDB_FSYNCONLY) {
2359 if (fsync(env->me_fd))
2363 if (MDB_FDATASYNC(env->me_fd))
2370 /** Back up parent txn's cursors, then grab the originals for tracking */
2372 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2374 MDB_cursor *mc, *bk;
2379 for (i = src->mt_numdbs; --i >= 0; ) {
2380 if ((mc = src->mt_cursors[i]) != NULL) {
2381 size = sizeof(MDB_cursor);
2383 size += sizeof(MDB_xcursor);
2384 for (; mc; mc = bk->mc_next) {
2390 mc->mc_db = &dst->mt_dbs[i];
2391 /* Kill pointers into src - and dst to reduce abuse: The
2392 * user may not use mc until dst ends. Otherwise we'd...
2394 mc->mc_txn = NULL; /* ...set this to dst */
2395 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2396 if ((mx = mc->mc_xcursor) != NULL) {
2397 *(MDB_xcursor *)(bk+1) = *mx;
2398 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2400 mc->mc_next = dst->mt_cursors[i];
2401 dst->mt_cursors[i] = mc;
2408 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2409 * @param[in] txn the transaction handle.
2410 * @param[in] merge true to keep changes to parent cursors, false to revert.
2411 * @return 0 on success, non-zero on failure.
2414 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2416 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2420 for (i = txn->mt_numdbs; --i >= 0; ) {
2421 for (mc = cursors[i]; mc; mc = next) {
2423 if ((bk = mc->mc_backup) != NULL) {
2425 /* Commit changes to parent txn */
2426 mc->mc_next = bk->mc_next;
2427 mc->mc_backup = bk->mc_backup;
2428 mc->mc_txn = bk->mc_txn;
2429 mc->mc_db = bk->mc_db;
2430 mc->mc_dbflag = bk->mc_dbflag;
2431 if ((mx = mc->mc_xcursor) != NULL)
2432 mx->mx_cursor.mc_txn = bk->mc_txn;
2434 /* Abort nested txn */
2436 if ((mx = mc->mc_xcursor) != NULL)
2437 *mx = *(MDB_xcursor *)(bk+1);
2441 /* Only malloced cursors are permanently tracked. */
2449 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2452 mdb_txn_reset0(MDB_txn *txn, const char *act);
2454 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2460 Pidset = F_SETLK, Pidcheck = F_GETLK
2464 /** Set or check a pid lock. Set returns 0 on success.
2465 * Check returns 0 if the process is certainly dead, nonzero if it may
2466 * be alive (the lock exists or an error happened so we do not know).
2468 * On Windows Pidset is a no-op, we merely check for the existence
2469 * of the process with the given pid. On POSIX we use a single byte
2470 * lock on the lockfile, set at an offset equal to the pid.
2473 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2475 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2478 if (op == Pidcheck) {
2479 h = OpenProcess(env->me_pidquery, FALSE, pid);
2480 /* No documented "no such process" code, but other program use this: */
2482 return ErrCode() != ERROR_INVALID_PARAMETER;
2483 /* A process exists until all handles to it close. Has it exited? */
2484 ret = WaitForSingleObject(h, 0) != 0;
2491 struct flock lock_info;
2492 memset(&lock_info, 0, sizeof(lock_info));
2493 lock_info.l_type = F_WRLCK;
2494 lock_info.l_whence = SEEK_SET;
2495 lock_info.l_start = pid;
2496 lock_info.l_len = 1;
2497 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2498 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2500 } else if ((rc = ErrCode()) == EINTR) {
2508 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2509 * @param[in] txn the transaction handle to initialize
2510 * @return 0 on success, non-zero on failure.
2513 mdb_txn_renew0(MDB_txn *txn)
2515 MDB_env *env = txn->mt_env;
2516 MDB_txninfo *ti = env->me_txns;
2518 unsigned int i, nr, flags = txn->mt_flags;
2520 int rc, new_notls = 0;
2522 if ((flags &= MDB_TXN_RDONLY) != 0) {
2524 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2525 txn->mt_txnid = meta->mm_txnid;
2526 txn->mt_u.reader = NULL;
2528 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2529 pthread_getspecific(env->me_txkey);
2531 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2532 return MDB_BAD_RSLOT;
2534 MDB_PID_T pid = env->me_pid;
2535 MDB_THR_T tid = pthread_self();
2537 if (!env->me_live_reader) {
2538 rc = mdb_reader_pid(env, Pidset, pid);
2541 env->me_live_reader = 1;
2545 nr = ti->mti_numreaders;
2546 for (i=0; i<nr; i++)
2547 if (ti->mti_readers[i].mr_pid == 0)
2549 if (i == env->me_maxreaders) {
2550 UNLOCK_MUTEX_R(env);
2551 return MDB_READERS_FULL;
2553 r = &ti->mti_readers[i];
2554 /* Claim the reader slot, carefully since other code
2555 * uses the reader table un-mutexed: First reset the
2556 * slot, next publish it in mti_numreaders. After
2557 * that, it is safe for mdb_env_close() to touch it.
2558 * When it will be closed, we can finally claim it.
2561 r->mr_txnid = (txnid_t)-1;
2564 ti->mti_numreaders = ++nr;
2565 env->me_close_readers = nr;
2567 UNLOCK_MUTEX_R(env);
2569 new_notls = (env->me_flags & MDB_NOTLS);
2570 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2575 do /* LY: Retry on a race, ITS#7970. */
2576 r->mr_txnid = ti->mti_txnid;
2577 while(r->mr_txnid != ti->mti_txnid);
2578 txn->mt_txnid = r->mr_txnid;
2579 txn->mt_u.reader = r;
2580 meta = env->me_metas[txn->mt_txnid & 1];
2582 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2584 /* Not yet touching txn == env->me_txn0, it may be active */
2588 txn->mt_txnid = ti->mti_txnid;
2589 meta = env->me_metas[txn->mt_txnid & 1];
2591 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2592 txn->mt_txnid = meta->mm_txnid;
2596 if (txn->mt_txnid == mdb_debug_start)
2599 txn->mt_child = NULL;
2600 txn->mt_loose_pgs = NULL;
2601 txn->mt_loose_count = 0;
2602 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2603 txn->mt_u.dirty_list = env->me_dirty_list;
2604 txn->mt_u.dirty_list[0].mid = 0;
2605 txn->mt_free_pgs = env->me_free_pgs;
2606 txn->mt_free_pgs[0] = 0;
2607 txn->mt_spill_pgs = NULL;
2609 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2612 /* Copy the DB info and flags */
2613 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2615 /* Moved to here to avoid a data race in read TXNs */
2616 txn->mt_next_pgno = meta->mm_last_pg+1;
2618 txn->mt_flags = flags;
2621 txn->mt_numdbs = env->me_numdbs;
2622 for (i=2; i<txn->mt_numdbs; i++) {
2623 x = env->me_dbflags[i];
2624 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2625 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2627 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2629 if (env->me_maxpg < txn->mt_next_pgno) {
2630 mdb_txn_reset0(txn, "renew0-mapfail");
2632 txn->mt_u.reader->mr_pid = 0;
2633 txn->mt_u.reader = NULL;
2635 return MDB_MAP_RESIZED;
2642 mdb_txn_renew(MDB_txn *txn)
2646 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2649 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2650 DPUTS("environment had fatal error, must shutdown!");
2654 rc = mdb_txn_renew0(txn);
2655 if (rc == MDB_SUCCESS) {
2656 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2657 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2658 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2664 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2668 int rc, size, tsize;
2670 flags &= MDB_TXN_BEGIN_FLAGS;
2671 flags |= env->me_flags & MDB_WRITEMAP;
2673 if (env->me_flags & MDB_FATAL_ERROR) {
2674 DPUTS("environment had fatal error, must shutdown!");
2677 if (env->me_flags & MDB_RDONLY & ~flags) /* write txn in RDONLY env */
2680 size = tsize = sizeof(MDB_txn);
2682 /* Nested transactions: Max 1 child, write txns only, no writemap */
2683 flags |= parent->mt_flags;
2684 if (parent->mt_child ||
2685 (flags & (MDB_RDONLY|MDB_WRITEMAP|MDB_TXN_ERROR)))
2687 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2689 /* Child txns save MDB_pgstate and use own copy of cursors */
2690 size = tsize = sizeof(MDB_ntxn);
2691 size += env->me_maxdbs * sizeof(MDB_cursor *);
2692 } else if (!(flags & MDB_RDONLY)) {
2693 /* Reuse preallocated write txn. However, do not touch it until
2694 * mdb_txn_renew0() succeeds, since it currently may be active.
2699 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2701 if ((txn = calloc(1, size)) == NULL) {
2702 DPRINTF(("calloc: %s", strerror(errno)));
2705 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2706 if (flags & MDB_RDONLY) {
2707 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2708 txn->mt_dbiseqs = env->me_dbiseqs;
2710 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2712 txn->mt_dbiseqs = parent->mt_dbiseqs;
2713 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2715 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2716 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2719 txn->mt_flags = flags;
2724 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2725 if (!txn->mt_u.dirty_list ||
2726 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2728 free(txn->mt_u.dirty_list);
2732 txn->mt_txnid = parent->mt_txnid;
2733 txn->mt_dirty_room = parent->mt_dirty_room;
2734 txn->mt_u.dirty_list[0].mid = 0;
2735 txn->mt_spill_pgs = NULL;
2736 txn->mt_next_pgno = parent->mt_next_pgno;
2737 parent->mt_child = txn;
2738 txn->mt_parent = parent;
2739 txn->mt_numdbs = parent->mt_numdbs;
2740 txn->mt_dbxs = parent->mt_dbxs;
2741 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2742 /* Copy parent's mt_dbflags, but clear DB_NEW */
2743 for (i=0; i<txn->mt_numdbs; i++)
2744 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2746 ntxn = (MDB_ntxn *)txn;
2747 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2748 if (env->me_pghead) {
2749 size = MDB_IDL_SIZEOF(env->me_pghead);
2750 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2752 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2757 rc = mdb_cursor_shadow(parent, txn);
2759 mdb_txn_reset0(txn, "beginchild-fail");
2762 rc = mdb_txn_renew0(txn);
2765 if (txn != env->me_txn0)
2768 txn->mt_flags |= flags; /* for txn==me_txn0, no effect otherwise */
2770 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2771 txn->mt_txnid, (flags & MDB_RDONLY) ? 'r' : 'w',
2772 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2779 mdb_txn_env(MDB_txn *txn)
2781 if(!txn) return NULL;
2786 mdb_txn_id(MDB_txn *txn)
2789 return txn->mt_txnid;
2792 /** Export or close DBI handles opened in this txn. */
2794 mdb_dbis_update(MDB_txn *txn, int keep)
2797 MDB_dbi n = txn->mt_numdbs;
2798 MDB_env *env = txn->mt_env;
2799 unsigned char *tdbflags = txn->mt_dbflags;
2801 for (i = n; --i >= 2;) {
2802 if (tdbflags[i] & DB_NEW) {
2804 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2806 char *ptr = env->me_dbxs[i].md_name.mv_data;
2808 env->me_dbxs[i].md_name.mv_data = NULL;
2809 env->me_dbxs[i].md_name.mv_size = 0;
2810 env->me_dbflags[i] = 0;
2811 env->me_dbiseqs[i]++;
2817 if (keep && env->me_numdbs < n)
2821 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2822 * May be called twice for readonly txns: First reset it, then abort.
2823 * @param[in] txn the transaction handle to reset
2824 * @param[in] act why the transaction is being reset
2827 mdb_txn_reset0(MDB_txn *txn, const char *act)
2829 MDB_env *env = txn->mt_env;
2831 /* Close any DBI handles opened in this txn */
2832 mdb_dbis_update(txn, 0);
2834 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2835 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2836 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2838 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2839 if (txn->mt_u.reader) {
2840 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2841 if (!(env->me_flags & MDB_NOTLS))
2842 txn->mt_u.reader = NULL; /* txn does not own reader */
2844 txn->mt_numdbs = 0; /* close nothing if called again */
2845 txn->mt_dbxs = NULL; /* mark txn as reset */
2847 pgno_t *pghead = env->me_pghead;
2849 mdb_cursors_close(txn, 0);
2850 if (!(env->me_flags & MDB_WRITEMAP)) {
2851 mdb_dlist_free(txn);
2854 if (!txn->mt_parent) {
2855 mdb_midl_shrink(&txn->mt_free_pgs);
2856 env->me_free_pgs = txn->mt_free_pgs;
2858 env->me_pghead = NULL;
2862 /* The writer mutex was locked in mdb_txn_begin. */
2864 UNLOCK_MUTEX_W(env);
2866 txn->mt_parent->mt_child = NULL;
2867 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2868 mdb_midl_free(txn->mt_free_pgs);
2869 mdb_midl_free(txn->mt_spill_pgs);
2870 free(txn->mt_u.dirty_list);
2873 mdb_midl_free(pghead);
2878 mdb_txn_reset(MDB_txn *txn)
2883 /* This call is only valid for read-only txns */
2884 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2887 mdb_txn_reset0(txn, "reset");
2891 mdb_txn_abort(MDB_txn *txn)
2897 mdb_txn_abort(txn->mt_child);
2899 mdb_txn_reset0(txn, "abort");
2900 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2901 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2902 txn->mt_u.reader->mr_pid = 0;
2904 if (txn != txn->mt_env->me_txn0)
2908 /** Save the freelist as of this transaction to the freeDB.
2909 * This changes the freelist. Keep trying until it stabilizes.
2912 mdb_freelist_save(MDB_txn *txn)
2914 /* env->me_pghead[] can grow and shrink during this call.
2915 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2916 * Page numbers cannot disappear from txn->mt_free_pgs[].
2919 MDB_env *env = txn->mt_env;
2920 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2921 txnid_t pglast = 0, head_id = 0;
2922 pgno_t freecnt = 0, *free_pgs, *mop;
2923 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2925 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2927 if (env->me_pghead) {
2928 /* Make sure first page of freeDB is touched and on freelist */
2929 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2930 if (rc && rc != MDB_NOTFOUND)
2934 if (!env->me_pghead && txn->mt_loose_pgs) {
2935 /* Put loose page numbers in mt_free_pgs, since
2936 * we may be unable to return them to me_pghead.
2938 MDB_page *mp = txn->mt_loose_pgs;
2939 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2941 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2942 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2943 txn->mt_loose_pgs = NULL;
2944 txn->mt_loose_count = 0;
2947 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2948 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2949 ? SSIZE_MAX : maxfree_1pg;
2952 /* Come back here after each Put() in case freelist changed */
2957 /* If using records from freeDB which we have not yet
2958 * deleted, delete them and any we reserved for me_pghead.
2960 while (pglast < env->me_pglast) {
2961 rc = mdb_cursor_first(&mc, &key, NULL);
2964 pglast = head_id = *(txnid_t *)key.mv_data;
2965 total_room = head_room = 0;
2966 mdb_tassert(txn, pglast <= env->me_pglast);
2967 rc = mdb_cursor_del(&mc, 0);
2972 /* Save the IDL of pages freed by this txn, to a single record */
2973 if (freecnt < txn->mt_free_pgs[0]) {
2975 /* Make sure last page of freeDB is touched and on freelist */
2976 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2977 if (rc && rc != MDB_NOTFOUND)
2980 free_pgs = txn->mt_free_pgs;
2981 /* Write to last page of freeDB */
2982 key.mv_size = sizeof(txn->mt_txnid);
2983 key.mv_data = &txn->mt_txnid;
2985 freecnt = free_pgs[0];
2986 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2987 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2990 /* Retry if mt_free_pgs[] grew during the Put() */
2991 free_pgs = txn->mt_free_pgs;
2992 } while (freecnt < free_pgs[0]);
2993 mdb_midl_sort(free_pgs);
2994 memcpy(data.mv_data, free_pgs, data.mv_size);
2997 unsigned int i = free_pgs[0];
2998 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2999 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
3001 DPRINTF(("IDL %"Z"u", free_pgs[i]));
3007 mop = env->me_pghead;
3008 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
3010 /* Reserve records for me_pghead[]. Split it if multi-page,
3011 * to avoid searching freeDB for a page range. Use keys in
3012 * range [1,me_pglast]: Smaller than txnid of oldest reader.
3014 if (total_room >= mop_len) {
3015 if (total_room == mop_len || --more < 0)
3017 } else if (head_room >= maxfree_1pg && head_id > 1) {
3018 /* Keep current record (overflow page), add a new one */
3022 /* (Re)write {key = head_id, IDL length = head_room} */
3023 total_room -= head_room;
3024 head_room = mop_len - total_room;
3025 if (head_room > maxfree_1pg && head_id > 1) {
3026 /* Overflow multi-page for part of me_pghead */
3027 head_room /= head_id; /* amortize page sizes */
3028 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
3029 } else if (head_room < 0) {
3030 /* Rare case, not bothering to delete this record */
3033 key.mv_size = sizeof(head_id);
3034 key.mv_data = &head_id;
3035 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3036 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3039 /* IDL is initially empty, zero out at least the length */
3040 pgs = (pgno_t *)data.mv_data;
3041 j = head_room > clean_limit ? head_room : 0;
3045 total_room += head_room;
3048 /* Return loose page numbers to me_pghead, though usually none are
3049 * left at this point. The pages themselves remain in dirty_list.
3051 if (txn->mt_loose_pgs) {
3052 MDB_page *mp = txn->mt_loose_pgs;
3053 unsigned count = txn->mt_loose_count;
3055 /* Room for loose pages + temp IDL with same */
3056 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3058 mop = env->me_pghead;
3059 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3060 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3061 loose[ ++count ] = mp->mp_pgno;
3063 mdb_midl_sort(loose);
3064 mdb_midl_xmerge(mop, loose);
3065 txn->mt_loose_pgs = NULL;
3066 txn->mt_loose_count = 0;
3070 /* Fill in the reserved me_pghead records */
3076 rc = mdb_cursor_first(&mc, &key, &data);
3077 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3078 txnid_t id = *(txnid_t *)key.mv_data;
3079 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3082 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3084 if (len > mop_len) {
3086 data.mv_size = (len + 1) * sizeof(MDB_ID);
3088 data.mv_data = mop -= len;
3091 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3093 if (rc || !(mop_len -= len))
3100 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3101 * @param[in] txn the transaction that's being committed
3102 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3103 * @return 0 on success, non-zero on failure.
3106 mdb_page_flush(MDB_txn *txn, int keep)
3108 MDB_env *env = txn->mt_env;
3109 MDB_ID2L dl = txn->mt_u.dirty_list;
3110 unsigned psize = env->me_psize, j;
3111 int i, pagecount = dl[0].mid, rc;
3112 size_t size = 0, pos = 0;
3114 MDB_page *dp = NULL;
3118 struct iovec iov[MDB_COMMIT_PAGES];
3119 ssize_t wpos = 0, wsize = 0, wres;
3120 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3126 if (env->me_flags & MDB_WRITEMAP) {
3127 /* Clear dirty flags */
3128 while (++i <= pagecount) {
3130 /* Don't flush this page yet */
3131 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3132 dp->mp_flags &= ~P_KEEP;
3136 dp->mp_flags &= ~P_DIRTY;
3141 /* Write the pages */
3143 if (++i <= pagecount) {
3145 /* Don't flush this page yet */
3146 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3147 dp->mp_flags &= ~P_KEEP;
3152 /* clear dirty flag */
3153 dp->mp_flags &= ~P_DIRTY;
3156 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3161 /* Windows actually supports scatter/gather I/O, but only on
3162 * unbuffered file handles. Since we're relying on the OS page
3163 * cache for all our data, that's self-defeating. So we just
3164 * write pages one at a time. We use the ov structure to set
3165 * the write offset, to at least save the overhead of a Seek
3168 DPRINTF(("committing page %"Z"u", pgno));
3169 memset(&ov, 0, sizeof(ov));
3170 ov.Offset = pos & 0xffffffff;
3171 ov.OffsetHigh = pos >> 16 >> 16;
3172 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3174 DPRINTF(("WriteFile: %d", rc));
3178 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3179 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3182 /* Write previous page(s) */
3183 #ifdef MDB_USE_PWRITEV
3184 wres = pwritev(env->me_fd, iov, n, wpos);
3187 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3190 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3194 DPRINTF(("lseek: %s", strerror(rc)));
3197 wres = writev(env->me_fd, iov, n);
3200 if (wres != wsize) {
3205 DPRINTF(("Write error: %s", strerror(rc)));
3207 rc = EIO; /* TODO: Use which error code? */
3208 DPUTS("short write, filesystem full?");
3219 DPRINTF(("committing page %"Z"u", pgno));
3220 next_pos = pos + size;
3221 iov[n].iov_len = size;
3222 iov[n].iov_base = (char *)dp;
3228 /* MIPS has cache coherency issues, this is a no-op everywhere else
3229 * Note: for any size >= on-chip cache size, entire on-chip cache is
3232 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3234 for (i = keep; ++i <= pagecount; ) {
3236 /* This is a page we skipped above */
3239 dl[j].mid = dp->mp_pgno;
3242 mdb_dpage_free(env, dp);
3247 txn->mt_dirty_room += i - j;
3253 mdb_txn_commit(MDB_txn *txn)
3259 if (txn == NULL || txn->mt_env == NULL)
3262 if (txn->mt_child) {
3263 rc = mdb_txn_commit(txn->mt_child);
3264 txn->mt_child = NULL;
3271 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3272 mdb_dbis_update(txn, 1);
3273 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3278 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3279 DPUTS("error flag is set, can't commit");
3281 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3286 if (txn->mt_parent) {
3287 MDB_txn *parent = txn->mt_parent;
3291 unsigned x, y, len, ps_len;
3293 /* Append our free list to parent's */
3294 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3297 mdb_midl_free(txn->mt_free_pgs);
3298 /* Failures after this must either undo the changes
3299 * to the parent or set MDB_TXN_ERROR in the parent.
3302 parent->mt_next_pgno = txn->mt_next_pgno;
3303 parent->mt_flags = txn->mt_flags;
3305 /* Merge our cursors into parent's and close them */
3306 mdb_cursors_close(txn, 1);
3308 /* Update parent's DB table. */
3309 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3310 parent->mt_numdbs = txn->mt_numdbs;
3311 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3312 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3313 for (i=2; i<txn->mt_numdbs; i++) {
3314 /* preserve parent's DB_NEW status */
3315 x = parent->mt_dbflags[i] & DB_NEW;
3316 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3319 dst = parent->mt_u.dirty_list;
3320 src = txn->mt_u.dirty_list;
3321 /* Remove anything in our dirty list from parent's spill list */
3322 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3324 pspill[0] = (pgno_t)-1;
3325 /* Mark our dirty pages as deleted in parent spill list */
3326 for (i=0, len=src[0].mid; ++i <= len; ) {
3327 MDB_ID pn = src[i].mid << 1;
3328 while (pn > pspill[x])
3330 if (pn == pspill[x]) {
3335 /* Squash deleted pagenums if we deleted any */
3336 for (x=y; ++x <= ps_len; )
3337 if (!(pspill[x] & 1))
3338 pspill[++y] = pspill[x];
3342 /* Find len = length of merging our dirty list with parent's */
3344 dst[0].mid = 0; /* simplify loops */
3345 if (parent->mt_parent) {
3346 len = x + src[0].mid;
3347 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3348 for (i = x; y && i; y--) {
3349 pgno_t yp = src[y].mid;
3350 while (yp < dst[i].mid)
3352 if (yp == dst[i].mid) {
3357 } else { /* Simplify the above for single-ancestor case */
3358 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3360 /* Merge our dirty list with parent's */
3362 for (i = len; y; dst[i--] = src[y--]) {
3363 pgno_t yp = src[y].mid;
3364 while (yp < dst[x].mid)
3365 dst[i--] = dst[x--];
3366 if (yp == dst[x].mid)
3367 free(dst[x--].mptr);
3369 mdb_tassert(txn, i == x);
3371 free(txn->mt_u.dirty_list);
3372 parent->mt_dirty_room = txn->mt_dirty_room;
3373 if (txn->mt_spill_pgs) {
3374 if (parent->mt_spill_pgs) {
3375 /* TODO: Prevent failure here, so parent does not fail */
3376 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3378 parent->mt_flags |= MDB_TXN_ERROR;
3379 mdb_midl_free(txn->mt_spill_pgs);
3380 mdb_midl_sort(parent->mt_spill_pgs);
3382 parent->mt_spill_pgs = txn->mt_spill_pgs;
3386 /* Append our loose page list to parent's */
3387 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3389 *lp = txn->mt_loose_pgs;
3390 parent->mt_loose_count += txn->mt_loose_count;
3392 parent->mt_child = NULL;
3393 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3398 if (txn != env->me_txn) {
3399 DPUTS("attempt to commit unknown transaction");
3404 mdb_cursors_close(txn, 0);
3406 if (!txn->mt_u.dirty_list[0].mid &&
3407 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3410 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3411 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3413 /* Update DB root pointers */
3414 if (txn->mt_numdbs > 2) {
3418 data.mv_size = sizeof(MDB_db);
3420 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3421 for (i = 2; i < txn->mt_numdbs; i++) {
3422 if (txn->mt_dbflags[i] & DB_DIRTY) {
3423 if (TXN_DBI_CHANGED(txn, i)) {
3427 data.mv_data = &txn->mt_dbs[i];
3428 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3436 rc = mdb_freelist_save(txn);
3440 mdb_midl_free(env->me_pghead);
3441 env->me_pghead = NULL;
3442 mdb_midl_shrink(&txn->mt_free_pgs);
3443 env->me_free_pgs = txn->mt_free_pgs;
3449 if ((rc = mdb_page_flush(txn, 0)) ||
3450 (rc = mdb_env_sync(env, 0)) ||
3451 (rc = mdb_env_write_meta(txn)))
3454 /* Free P_LOOSE pages left behind in dirty_list */
3455 if (!(env->me_flags & MDB_WRITEMAP))
3456 mdb_dlist_free(txn);
3461 mdb_dbis_update(txn, 1);
3464 UNLOCK_MUTEX_W(env);
3465 if (txn != env->me_txn0)
3475 /** Read the environment parameters of a DB environment before
3476 * mapping it into memory.
3477 * @param[in] env the environment handle
3478 * @param[out] meta address of where to store the meta information
3479 * @return 0 on success, non-zero on failure.
3482 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3488 enum { Size = sizeof(pbuf) };
3490 /* We don't know the page size yet, so use a minimum value.
3491 * Read both meta pages so we can use the latest one.
3494 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3498 memset(&ov, 0, sizeof(ov));
3500 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3501 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3504 rc = pread(env->me_fd, &pbuf, Size, off);
3507 if (rc == 0 && off == 0)
3509 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3510 DPRINTF(("read: %s", mdb_strerror(rc)));
3514 p = (MDB_page *)&pbuf;
3516 if (!F_ISSET(p->mp_flags, P_META)) {
3517 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3522 if (m->mm_magic != MDB_MAGIC) {
3523 DPUTS("meta has invalid magic");
3527 if (m->mm_version != MDB_DATA_VERSION) {
3528 DPRINTF(("database is version %u, expected version %u",
3529 m->mm_version, MDB_DATA_VERSION));
3530 return MDB_VERSION_MISMATCH;
3533 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3539 /** Fill in most of the zeroed #MDB_meta for an empty database environment */
3541 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3543 meta->mm_magic = MDB_MAGIC;
3544 meta->mm_version = MDB_DATA_VERSION;
3545 meta->mm_mapsize = env->me_mapsize;
3546 meta->mm_psize = env->me_psize;
3547 meta->mm_last_pg = 1;
3548 meta->mm_flags = env->me_flags & 0xffff;
3549 meta->mm_flags |= MDB_INTEGERKEY;
3550 meta->mm_dbs[0].md_root = P_INVALID;
3551 meta->mm_dbs[1].md_root = P_INVALID;
3554 /** Write the environment parameters of a freshly created DB environment.
3555 * @param[in] env the environment handle
3556 * @param[in] meta the #MDB_meta to write
3557 * @return 0 on success, non-zero on failure.
3560 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3568 memset(&ov, 0, sizeof(ov));
3569 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3571 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3574 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3575 len = pwrite(fd, ptr, size, pos); \
3576 if (len == -1 && ErrCode() == EINTR) continue; \
3577 rc = (len >= 0); break; } while(1)
3580 DPUTS("writing new meta page");
3582 psize = env->me_psize;
3584 p = calloc(2, psize);
3589 p->mp_flags = P_META;
3590 *(MDB_meta *)METADATA(p) = *meta;
3592 q = (MDB_page *)((char *)p + psize);
3594 q->mp_flags = P_META;
3595 *(MDB_meta *)METADATA(q) = *meta;
3597 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3600 else if ((unsigned) len == psize * 2)
3608 /** Update the environment info to commit a transaction.
3609 * @param[in] txn the transaction that's being committed
3610 * @return 0 on success, non-zero on failure.
3613 mdb_env_write_meta(MDB_txn *txn)
3616 MDB_meta meta, metab, *mp;
3620 int rc, len, toggle;
3629 toggle = txn->mt_txnid & 1;
3630 DPRINTF(("writing meta page %d for root page %"Z"u",
3631 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3634 flags = env->me_flags;
3635 mp = env->me_metas[toggle];
3636 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3637 /* Persist any increases of mapsize config */
3638 if (mapsize < env->me_mapsize)
3639 mapsize = env->me_mapsize;
3641 if (flags & MDB_WRITEMAP) {
3642 mp->mm_mapsize = mapsize;
3643 mp->mm_dbs[0] = txn->mt_dbs[0];
3644 mp->mm_dbs[1] = txn->mt_dbs[1];
3645 mp->mm_last_pg = txn->mt_next_pgno - 1;
3646 #if (__GNUC__ * 100 + __GNUC_MINOR__ >= 404) && /* TODO: portability */ \
3647 !(defined(__i386__) || defined(__x86_64__))
3648 /* LY: issue a memory barrier, if not x86. ITS#7969 */
3649 __sync_synchronize();
3651 mp->mm_txnid = txn->mt_txnid;
3652 if (!(flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3653 unsigned meta_size = env->me_psize;
3654 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3657 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3658 if (meta_size < env->me_os_psize)
3659 meta_size += meta_size;
3664 if (MDB_MSYNC(ptr, meta_size, rc)) {
3671 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3672 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3674 meta.mm_mapsize = mapsize;
3675 meta.mm_dbs[0] = txn->mt_dbs[0];
3676 meta.mm_dbs[1] = txn->mt_dbs[1];
3677 meta.mm_last_pg = txn->mt_next_pgno - 1;
3678 meta.mm_txnid = txn->mt_txnid;
3680 off = offsetof(MDB_meta, mm_mapsize);
3681 ptr = (char *)&meta + off;
3682 len = sizeof(MDB_meta) - off;
3684 off += env->me_psize;
3687 /* Write to the SYNC fd */
3688 mfd = (flags & (MDB_NOSYNC|MDB_NOMETASYNC)) ? env->me_fd : env->me_mfd;
3691 memset(&ov, 0, sizeof(ov));
3693 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3698 rc = pwrite(mfd, ptr, len, off);
3701 rc = rc < 0 ? ErrCode() : EIO;
3706 DPUTS("write failed, disk error?");
3707 /* On a failure, the pagecache still contains the new data.
3708 * Write some old data back, to prevent it from being used.
3709 * Use the non-SYNC fd; we know it will fail anyway.
3711 meta.mm_last_pg = metab.mm_last_pg;
3712 meta.mm_txnid = metab.mm_txnid;
3714 memset(&ov, 0, sizeof(ov));
3716 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3718 r2 = pwrite(env->me_fd, ptr, len, off);
3719 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3722 env->me_flags |= MDB_FATAL_ERROR;
3725 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3726 CACHEFLUSH(env->me_map + off, len, DCACHE);
3728 /* Memory ordering issues are irrelevant; since the entire writer
3729 * is wrapped by wmutex, all of these changes will become visible
3730 * after the wmutex is unlocked. Since the DB is multi-version,
3731 * readers will get consistent data regardless of how fresh or
3732 * how stale their view of these values is.
3735 env->me_txns->mti_txnid = txn->mt_txnid;
3740 /** Check both meta pages to see which one is newer.
3741 * @param[in] env the environment handle
3742 * @return meta toggle (0 or 1).
3745 mdb_env_pick_meta(const MDB_env *env)
3747 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3751 mdb_env_create(MDB_env **env)
3755 e = calloc(1, sizeof(MDB_env));
3759 e->me_maxreaders = DEFAULT_READERS;
3760 e->me_maxdbs = e->me_numdbs = 2;
3761 e->me_fd = INVALID_HANDLE_VALUE;
3762 e->me_lfd = INVALID_HANDLE_VALUE;
3763 e->me_mfd = INVALID_HANDLE_VALUE;
3764 #ifdef MDB_USE_POSIX_SEM
3765 e->me_rmutex = SEM_FAILED;
3766 e->me_wmutex = SEM_FAILED;
3768 e->me_pid = getpid();
3769 GET_PAGESIZE(e->me_os_psize);
3770 VGMEMP_CREATE(e,0,0);
3776 mdb_env_map(MDB_env *env, void *addr)
3779 unsigned int flags = env->me_flags;
3783 LONG sizelo, sizehi;
3786 if (flags & MDB_RDONLY) {
3787 /* Don't set explicit map size, use whatever exists */
3792 msize = env->me_mapsize;
3793 sizelo = msize & 0xffffffff;
3794 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3796 /* Windows won't create mappings for zero length files.
3797 * and won't map more than the file size.
3798 * Just set the maxsize right now.
3800 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3801 || !SetEndOfFile(env->me_fd)
3802 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3806 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3807 PAGE_READWRITE : PAGE_READONLY,
3808 sizehi, sizelo, NULL);
3811 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3812 FILE_MAP_WRITE : FILE_MAP_READ,
3814 rc = env->me_map ? 0 : ErrCode();
3819 int prot = PROT_READ;
3820 if (flags & MDB_WRITEMAP) {
3822 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3825 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3827 if (env->me_map == MAP_FAILED) {
3832 if (flags & MDB_NORDAHEAD) {
3833 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3835 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3837 #ifdef POSIX_MADV_RANDOM
3838 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3839 #endif /* POSIX_MADV_RANDOM */
3840 #endif /* MADV_RANDOM */
3844 /* Can happen because the address argument to mmap() is just a
3845 * hint. mmap() can pick another, e.g. if the range is in use.
3846 * The MAP_FIXED flag would prevent that, but then mmap could
3847 * instead unmap existing pages to make room for the new map.
3849 if (addr && env->me_map != addr)
3850 return EBUSY; /* TODO: Make a new MDB_* error code? */
3852 p = (MDB_page *)env->me_map;
3853 env->me_metas[0] = METADATA(p);
3854 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3860 mdb_env_set_mapsize(MDB_env *env, size_t size)
3862 /* If env is already open, caller is responsible for making
3863 * sure there are no active txns.
3871 meta = env->me_metas[mdb_env_pick_meta(env)];
3873 size = meta->mm_mapsize;
3875 /* Silently round up to minimum if the size is too small */
3876 size_t minsize = (meta->mm_last_pg + 1) * env->me_psize;
3880 munmap(env->me_map, env->me_mapsize);
3881 env->me_mapsize = size;
3882 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3883 rc = mdb_env_map(env, old);
3887 env->me_mapsize = size;
3889 env->me_maxpg = env->me_mapsize / env->me_psize;
3894 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3898 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3903 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3905 if (env->me_map || readers < 1)
3907 env->me_maxreaders = readers;
3912 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3914 if (!env || !readers)
3916 *readers = env->me_maxreaders;
3921 mdb_fsize(HANDLE fd, size_t *size)
3924 LARGE_INTEGER fsize;
3926 if (!GetFileSizeEx(fd, &fsize))
3929 *size = fsize.QuadPart;
3941 #ifdef BROKEN_FDATASYNC
3942 #include <sys/utsname.h>
3943 #include <sys/vfs.h>
3946 /** Further setup required for opening an LMDB environment
3949 mdb_env_open2(MDB_env *env)
3951 unsigned int flags = env->me_flags;
3952 int i, newenv = 0, rc;
3956 /* See if we should use QueryLimited */
3958 if ((rc & 0xff) > 5)
3959 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3961 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3964 #ifdef BROKEN_FDATASYNC
3965 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
3966 * https://lkml.org/lkml/2012/9/3/83
3967 * Kernels after 3.6-rc6 are known good.
3968 * https://lkml.org/lkml/2012/9/10/556
3969 * See if the DB is on ext3/ext4, then check for new enough kernel
3970 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
3975 fstatfs(env->me_fd, &st);
3976 while (st.f_type == 0xEF53) {
3980 if (uts.release[0] < '3') {
3981 if (!strncmp(uts.release, "2.6.32.", 7)) {
3982 i = atoi(uts.release+7);
3984 break; /* 2.6.32.60 and newer is OK */
3985 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
3986 i = atoi(uts.release+7);
3988 break; /* 2.6.34.15 and newer is OK */
3990 } else if (uts.release[0] == '3') {
3991 i = atoi(uts.release+2);
3993 break; /* 3.6 and newer is OK */
3995 i = atoi(uts.release+4);
3997 break; /* 3.5.4 and newer is OK */
3998 } else if (i == 2) {
3999 i = atoi(uts.release+4);
4001 break; /* 3.2.30 and newer is OK */
4003 } else { /* 4.x and newer is OK */
4006 env->me_flags |= MDB_FSYNCONLY;
4012 if ((i = mdb_env_read_header(env, &meta)) != 0) {
4015 DPUTS("new mdbenv");
4017 env->me_psize = env->me_os_psize;
4018 if (env->me_psize > MAX_PAGESIZE)
4019 env->me_psize = MAX_PAGESIZE;
4020 memset(&meta, 0, sizeof(meta));
4021 mdb_env_init_meta0(env, &meta);
4022 meta.mm_mapsize = DEFAULT_MAPSIZE;
4024 env->me_psize = meta.mm_psize;
4027 /* Was a mapsize configured? */
4028 if (!env->me_mapsize) {
4029 env->me_mapsize = meta.mm_mapsize;
4032 /* Make sure mapsize >= committed data size. Even when using
4033 * mm_mapsize, which could be broken in old files (ITS#7789).
4035 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
4036 if (env->me_mapsize < minsize)
4037 env->me_mapsize = minsize;
4039 meta.mm_mapsize = env->me_mapsize;
4041 if (newenv && !(flags & MDB_FIXEDMAP)) {
4042 /* mdb_env_map() may grow the datafile. Write the metapages
4043 * first, so the file will be valid if initialization fails.
4044 * Except with FIXEDMAP, since we do not yet know mm_address.
4045 * We could fill in mm_address later, but then a different
4046 * program might end up doing that - one with a memory layout
4047 * and map address which does not suit the main program.
4049 rc = mdb_env_init_meta(env, &meta);
4055 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4060 if (flags & MDB_FIXEDMAP)
4061 meta.mm_address = env->me_map;
4062 i = mdb_env_init_meta(env, &meta);
4063 if (i != MDB_SUCCESS) {
4068 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4069 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4071 #if !(MDB_MAXKEYSIZE)
4072 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4074 env->me_maxpg = env->me_mapsize / env->me_psize;
4078 int toggle = mdb_env_pick_meta(env);
4079 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4081 DPRINTF(("opened database version %u, pagesize %u",
4082 env->me_metas[0]->mm_version, env->me_psize));
4083 DPRINTF(("using meta page %d", toggle));
4084 DPRINTF(("depth: %u", db->md_depth));
4085 DPRINTF(("entries: %"Z"u", db->md_entries));
4086 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4087 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4088 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4089 DPRINTF(("root: %"Z"u", db->md_root));
4097 /** Release a reader thread's slot in the reader lock table.
4098 * This function is called automatically when a thread exits.
4099 * @param[in] ptr This points to the slot in the reader lock table.
4102 mdb_env_reader_dest(void *ptr)
4104 MDB_reader *reader = ptr;
4110 /** Junk for arranging thread-specific callbacks on Windows. This is
4111 * necessarily platform and compiler-specific. Windows supports up
4112 * to 1088 keys. Let's assume nobody opens more than 64 environments
4113 * in a single process, for now. They can override this if needed.
4115 #ifndef MAX_TLS_KEYS
4116 #define MAX_TLS_KEYS 64
4118 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4119 static int mdb_tls_nkeys;
4121 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4125 case DLL_PROCESS_ATTACH: break;
4126 case DLL_THREAD_ATTACH: break;
4127 case DLL_THREAD_DETACH:
4128 for (i=0; i<mdb_tls_nkeys; i++) {
4129 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4131 mdb_env_reader_dest(r);
4135 case DLL_PROCESS_DETACH: break;
4140 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4142 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4146 /* Force some symbol references.
4147 * _tls_used forces the linker to create the TLS directory if not already done
4148 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4150 #pragma comment(linker, "/INCLUDE:_tls_used")
4151 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4152 #pragma const_seg(".CRT$XLB")
4153 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4154 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4157 #pragma comment(linker, "/INCLUDE:__tls_used")
4158 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4159 #pragma data_seg(".CRT$XLB")
4160 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4162 #endif /* WIN 32/64 */
4163 #endif /* !__GNUC__ */
4166 /** Downgrade the exclusive lock on the region back to shared */
4168 mdb_env_share_locks(MDB_env *env, int *excl)
4170 int rc = 0, toggle = mdb_env_pick_meta(env);
4172 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4177 /* First acquire a shared lock. The Unlock will
4178 * then release the existing exclusive lock.
4180 memset(&ov, 0, sizeof(ov));
4181 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4184 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4190 struct flock lock_info;
4191 /* The shared lock replaces the existing lock */
4192 memset((void *)&lock_info, 0, sizeof(lock_info));
4193 lock_info.l_type = F_RDLCK;
4194 lock_info.l_whence = SEEK_SET;
4195 lock_info.l_start = 0;
4196 lock_info.l_len = 1;
4197 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4198 (rc = ErrCode()) == EINTR) ;
4199 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4206 /** Try to get exclusive lock, otherwise shared.
4207 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4210 mdb_env_excl_lock(MDB_env *env, int *excl)
4214 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4218 memset(&ov, 0, sizeof(ov));
4219 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4226 struct flock lock_info;
4227 memset((void *)&lock_info, 0, sizeof(lock_info));
4228 lock_info.l_type = F_WRLCK;
4229 lock_info.l_whence = SEEK_SET;
4230 lock_info.l_start = 0;
4231 lock_info.l_len = 1;
4232 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4233 (rc = ErrCode()) == EINTR) ;
4237 # ifdef MDB_USE_POSIX_SEM
4238 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4241 lock_info.l_type = F_RDLCK;
4242 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4243 (rc = ErrCode()) == EINTR) ;
4253 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4255 * @(#) $Revision: 5.1 $
4256 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4257 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4259 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4263 * Please do not copyright this code. This code is in the public domain.
4265 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4266 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4267 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4268 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4269 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4270 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4271 * PERFORMANCE OF THIS SOFTWARE.
4274 * chongo <Landon Curt Noll> /\oo/\
4275 * http://www.isthe.com/chongo/
4277 * Share and Enjoy! :-)
4280 typedef unsigned long long mdb_hash_t;
4281 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4283 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4284 * @param[in] val value to hash
4285 * @param[in] hval initial value for hash
4286 * @return 64 bit hash
4288 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4289 * hval arg on the first call.
4292 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4294 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4295 unsigned char *end = s + val->mv_size;
4297 * FNV-1a hash each octet of the string
4300 /* xor the bottom with the current octet */
4301 hval ^= (mdb_hash_t)*s++;
4303 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4304 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4305 (hval << 7) + (hval << 8) + (hval << 40);
4307 /* return our new hash value */
4311 /** Hash the string and output the encoded hash.
4312 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4313 * very short name limits. We don't care about the encoding being reversible,
4314 * we just want to preserve as many bits of the input as possible in a
4315 * small printable string.
4316 * @param[in] str string to hash
4317 * @param[out] encbuf an array of 11 chars to hold the hash
4319 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4322 mdb_pack85(unsigned long l, char *out)
4326 for (i=0; i<5; i++) {
4327 *out++ = mdb_a85[l % 85];
4333 mdb_hash_enc(MDB_val *val, char *encbuf)
4335 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4337 mdb_pack85(h, encbuf);
4338 mdb_pack85(h>>32, encbuf+5);
4343 /** Open and/or initialize the lock region for the environment.
4344 * @param[in] env The LMDB environment.
4345 * @param[in] lpath The pathname of the file used for the lock region.
4346 * @param[in] mode The Unix permissions for the file, if we create it.
4347 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4348 * @return 0 on success, non-zero on failure.
4351 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4354 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4356 # define MDB_ERRCODE_ROFS EROFS
4357 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4358 # define MDB_CLOEXEC O_CLOEXEC
4361 # define MDB_CLOEXEC 0
4368 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4369 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4370 FILE_ATTRIBUTE_NORMAL, NULL);
4372 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4374 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4376 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4381 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4382 /* Lose record locks when exec*() */
4383 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4384 fcntl(env->me_lfd, F_SETFD, fdflags);
4387 if (!(env->me_flags & MDB_NOTLS)) {
4388 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4391 env->me_flags |= MDB_ENV_TXKEY;
4393 /* Windows TLS callbacks need help finding their TLS info. */
4394 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4398 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4402 /* Try to get exclusive lock. If we succeed, then
4403 * nobody is using the lock region and we should initialize it.
4405 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4408 size = GetFileSize(env->me_lfd, NULL);
4410 size = lseek(env->me_lfd, 0, SEEK_END);
4411 if (size == -1) goto fail_errno;
4413 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4414 if (size < rsize && *excl > 0) {
4416 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4417 || !SetEndOfFile(env->me_lfd))
4420 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4424 size = rsize - sizeof(MDB_txninfo);
4425 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4430 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4432 if (!mh) goto fail_errno;
4433 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4435 if (!env->me_txns) goto fail_errno;
4437 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4439 if (m == MAP_FAILED) goto fail_errno;
4445 BY_HANDLE_FILE_INFORMATION stbuf;
4454 if (!mdb_sec_inited) {
4455 InitializeSecurityDescriptor(&mdb_null_sd,
4456 SECURITY_DESCRIPTOR_REVISION);
4457 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4458 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4459 mdb_all_sa.bInheritHandle = FALSE;
4460 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4463 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4464 idbuf.volume = stbuf.dwVolumeSerialNumber;
4465 idbuf.nhigh = stbuf.nFileIndexHigh;
4466 idbuf.nlow = stbuf.nFileIndexLow;
4467 val.mv_data = &idbuf;
4468 val.mv_size = sizeof(idbuf);
4469 mdb_hash_enc(&val, encbuf);
4470 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4471 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4472 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4473 if (!env->me_rmutex) goto fail_errno;
4474 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4475 if (!env->me_wmutex) goto fail_errno;
4476 #elif defined(MDB_USE_POSIX_SEM)
4485 #if defined(__NetBSD__)
4486 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4488 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4489 idbuf.dev = stbuf.st_dev;
4490 idbuf.ino = stbuf.st_ino;
4491 val.mv_data = &idbuf;
4492 val.mv_size = sizeof(idbuf);
4493 mdb_hash_enc(&val, encbuf);
4494 #ifdef MDB_SHORT_SEMNAMES
4495 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4497 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4498 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4499 /* Clean up after a previous run, if needed: Try to
4500 * remove both semaphores before doing anything else.
4502 sem_unlink(env->me_txns->mti_rmname);
4503 sem_unlink(env->me_txns->mti_wmname);
4504 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4505 O_CREAT|O_EXCL, mode, 1);
4506 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4507 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4508 O_CREAT|O_EXCL, mode, 1);
4509 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4510 #else /* MDB_USE_POSIX_SEM */
4511 pthread_mutexattr_t mattr;
4513 if ((rc = pthread_mutexattr_init(&mattr))
4514 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4515 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4516 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4518 pthread_mutexattr_destroy(&mattr);
4519 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4521 env->me_txns->mti_magic = MDB_MAGIC;
4522 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4523 env->me_txns->mti_txnid = 0;
4524 env->me_txns->mti_numreaders = 0;
4527 if (env->me_txns->mti_magic != MDB_MAGIC) {
4528 DPUTS("lock region has invalid magic");
4532 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4533 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4534 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4535 rc = MDB_VERSION_MISMATCH;
4539 if (rc && rc != EACCES && rc != EAGAIN) {
4543 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4544 if (!env->me_rmutex) goto fail_errno;
4545 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4546 if (!env->me_wmutex) goto fail_errno;
4547 #elif defined(MDB_USE_POSIX_SEM)
4548 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4549 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4550 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4551 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4562 /** The name of the lock file in the DB environment */
4563 #define LOCKNAME "/lock.mdb"
4564 /** The name of the data file in the DB environment */
4565 #define DATANAME "/data.mdb"
4566 /** The suffix of the lock file when no subdir is used */
4567 #define LOCKSUFF "-lock"
4568 /** Only a subset of the @ref mdb_env flags can be changed
4569 * at runtime. Changing other flags requires closing the
4570 * environment and re-opening it with the new flags.
4572 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4573 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4574 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4576 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4577 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4581 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4583 int oflags, rc, len, excl = -1;
4584 char *lpath, *dpath;
4586 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4590 if (flags & MDB_NOSUBDIR) {
4591 rc = len + sizeof(LOCKSUFF) + len + 1;
4593 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4598 if (flags & MDB_NOSUBDIR) {
4599 dpath = lpath + len + sizeof(LOCKSUFF);
4600 sprintf(lpath, "%s" LOCKSUFF, path);
4601 strcpy(dpath, path);
4603 dpath = lpath + len + sizeof(LOCKNAME);
4604 sprintf(lpath, "%s" LOCKNAME, path);
4605 sprintf(dpath, "%s" DATANAME, path);
4609 flags |= env->me_flags;
4610 if (flags & MDB_RDONLY) {
4611 /* silently ignore WRITEMAP when we're only getting read access */
4612 flags &= ~MDB_WRITEMAP;
4614 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4615 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4618 env->me_flags = flags |= MDB_ENV_ACTIVE;
4622 env->me_path = strdup(path);
4623 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4624 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4625 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4626 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4631 /* For RDONLY, get lockfile after we know datafile exists */
4632 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4633 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4639 if (F_ISSET(flags, MDB_RDONLY)) {
4640 oflags = GENERIC_READ;
4641 len = OPEN_EXISTING;
4643 oflags = GENERIC_READ|GENERIC_WRITE;
4646 mode = FILE_ATTRIBUTE_NORMAL;
4647 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4648 NULL, len, mode, NULL);
4650 if (F_ISSET(flags, MDB_RDONLY))
4653 oflags = O_RDWR | O_CREAT;
4655 env->me_fd = open(dpath, oflags, mode);
4657 if (env->me_fd == INVALID_HANDLE_VALUE) {
4662 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4663 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4668 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4669 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4670 env->me_mfd = env->me_fd;
4672 /* Synchronous fd for meta writes. Needed even with
4673 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4676 len = OPEN_EXISTING;
4677 env->me_mfd = CreateFile(dpath, oflags,
4678 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4679 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4682 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4684 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4689 DPRINTF(("opened dbenv %p", (void *) env));
4691 rc = mdb_env_share_locks(env, &excl);
4695 if (!(flags & MDB_RDONLY)) {
4697 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4698 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4699 if ((env->me_pbuf = calloc(1, env->me_psize)) &&
4700 (txn = calloc(1, size)))
4702 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4703 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4704 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4705 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4707 txn->mt_dbxs = env->me_dbxs;
4717 mdb_env_close0(env, excl);
4723 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4725 mdb_env_close0(MDB_env *env, int excl)
4729 if (!(env->me_flags & MDB_ENV_ACTIVE))
4732 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4734 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4735 free(env->me_dbxs[i].md_name.mv_data);
4740 free(env->me_dbiseqs);
4741 free(env->me_dbflags);
4743 free(env->me_dirty_list);
4745 mdb_midl_free(env->me_free_pgs);
4747 if (env->me_flags & MDB_ENV_TXKEY) {
4748 pthread_key_delete(env->me_txkey);
4750 /* Delete our key from the global list */
4751 for (i=0; i<mdb_tls_nkeys; i++)
4752 if (mdb_tls_keys[i] == env->me_txkey) {
4753 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4761 munmap(env->me_map, env->me_mapsize);
4763 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4764 (void) close(env->me_mfd);
4765 if (env->me_fd != INVALID_HANDLE_VALUE)
4766 (void) close(env->me_fd);
4768 MDB_PID_T pid = env->me_pid;
4769 /* Clearing readers is done in this function because
4770 * me_txkey with its destructor must be disabled first.
4772 * We skip the the reader mutex, so we touch only
4773 * data owned by this process (me_close_readers and
4774 * our readers), and clear each reader atomically.
4776 for (i = env->me_close_readers; --i >= 0; )
4777 if (env->me_txns->mti_readers[i].mr_pid == pid)
4778 env->me_txns->mti_readers[i].mr_pid = 0;
4780 if (env->me_rmutex) {
4781 CloseHandle(env->me_rmutex);
4782 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4784 /* Windows automatically destroys the mutexes when
4785 * the last handle closes.
4787 #elif defined(MDB_USE_POSIX_SEM)
4788 if (env->me_rmutex != SEM_FAILED) {
4789 sem_close(env->me_rmutex);
4790 if (env->me_wmutex != SEM_FAILED)
4791 sem_close(env->me_wmutex);
4792 /* If we have the filelock: If we are the
4793 * only remaining user, clean up semaphores.
4796 mdb_env_excl_lock(env, &excl);
4798 sem_unlink(env->me_txns->mti_rmname);
4799 sem_unlink(env->me_txns->mti_wmname);
4803 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4805 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4808 /* Unlock the lockfile. Windows would have unlocked it
4809 * after closing anyway, but not necessarily at once.
4811 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4814 (void) close(env->me_lfd);
4817 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4821 mdb_env_close(MDB_env *env)
4828 VGMEMP_DESTROY(env);
4829 while ((dp = env->me_dpages) != NULL) {
4830 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4831 env->me_dpages = dp->mp_next;
4835 mdb_env_close0(env, 0);
4839 /** Compare two items pointing at aligned size_t's */
4841 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4843 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4844 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4847 /** Compare two items pointing at aligned unsigned int's.
4849 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4850 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4853 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4855 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4856 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4859 /** Compare two items pointing at unsigned ints of unknown alignment.
4860 * Nodes and keys are guaranteed to be 2-byte aligned.
4863 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4865 #if BYTE_ORDER == LITTLE_ENDIAN
4866 unsigned short *u, *c;
4869 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4870 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4873 } while(!x && u > (unsigned short *)a->mv_data);
4876 unsigned short *u, *c, *end;
4879 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4880 u = (unsigned short *)a->mv_data;
4881 c = (unsigned short *)b->mv_data;
4884 } while(!x && u < end);
4889 /** Compare two items lexically */
4891 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4898 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4904 diff = memcmp(a->mv_data, b->mv_data, len);
4905 return diff ? diff : len_diff<0 ? -1 : len_diff;
4908 /** Compare two items in reverse byte order */
4910 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4912 const unsigned char *p1, *p2, *p1_lim;
4916 p1_lim = (const unsigned char *)a->mv_data;
4917 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4918 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4920 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4926 while (p1 > p1_lim) {
4927 diff = *--p1 - *--p2;
4931 return len_diff<0 ? -1 : len_diff;
4934 /** Search for key within a page, using binary search.
4935 * Returns the smallest entry larger or equal to the key.
4936 * If exactp is non-null, stores whether the found entry was an exact match
4937 * in *exactp (1 or 0).
4938 * Updates the cursor index with the index of the found entry.
4939 * If no entry larger or equal to the key is found, returns NULL.
4942 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4944 unsigned int i = 0, nkeys;
4947 MDB_page *mp = mc->mc_pg[mc->mc_top];
4948 MDB_node *node = NULL;
4953 nkeys = NUMKEYS(mp);
4955 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4956 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4959 low = IS_LEAF(mp) ? 0 : 1;
4961 cmp = mc->mc_dbx->md_cmp;
4963 /* Branch pages have no data, so if using integer keys,
4964 * alignment is guaranteed. Use faster mdb_cmp_int.
4966 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4967 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4974 nodekey.mv_size = mc->mc_db->md_pad;
4975 node = NODEPTR(mp, 0); /* fake */
4976 while (low <= high) {
4977 i = (low + high) >> 1;
4978 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4979 rc = cmp(key, &nodekey);
4980 DPRINTF(("found leaf index %u [%s], rc = %i",
4981 i, DKEY(&nodekey), rc));
4990 while (low <= high) {
4991 i = (low + high) >> 1;
4993 node = NODEPTR(mp, i);
4994 nodekey.mv_size = NODEKSZ(node);
4995 nodekey.mv_data = NODEKEY(node);
4997 rc = cmp(key, &nodekey);
5000 DPRINTF(("found leaf index %u [%s], rc = %i",
5001 i, DKEY(&nodekey), rc));
5003 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
5004 i, DKEY(&nodekey), NODEPGNO(node), rc));
5015 if (rc > 0) { /* Found entry is less than the key. */
5016 i++; /* Skip to get the smallest entry larger than key. */
5018 node = NODEPTR(mp, i);
5021 *exactp = (rc == 0 && nkeys > 0);
5022 /* store the key index */
5023 mc->mc_ki[mc->mc_top] = i;
5025 /* There is no entry larger or equal to the key. */
5028 /* nodeptr is fake for LEAF2 */
5034 mdb_cursor_adjust(MDB_cursor *mc, func)
5038 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
5039 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
5046 /** Pop a page off the top of the cursor's stack. */
5048 mdb_cursor_pop(MDB_cursor *mc)
5052 MDB_page *top = mc->mc_pg[mc->mc_top];
5058 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
5059 DDBI(mc), (void *) mc));
5063 /** Push a page onto the top of the cursor's stack. */
5065 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5067 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5068 DDBI(mc), (void *) mc));
5070 if (mc->mc_snum >= CURSOR_STACK) {
5071 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5072 return MDB_CURSOR_FULL;
5075 mc->mc_top = mc->mc_snum++;
5076 mc->mc_pg[mc->mc_top] = mp;
5077 mc->mc_ki[mc->mc_top] = 0;
5082 /** Find the address of the page corresponding to a given page number.
5083 * @param[in] txn the transaction for this access.
5084 * @param[in] pgno the page number for the page to retrieve.
5085 * @param[out] ret address of a pointer where the page's address will be stored.
5086 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5087 * @return 0 on success, non-zero on failure.
5090 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5092 MDB_env *env = txn->mt_env;
5096 if (! (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_WRITEMAP))) {
5100 MDB_ID2L dl = tx2->mt_u.dirty_list;
5102 /* Spilled pages were dirtied in this txn and flushed
5103 * because the dirty list got full. Bring this page
5104 * back in from the map (but don't unspill it here,
5105 * leave that unless page_touch happens again).
5107 if (tx2->mt_spill_pgs) {
5108 MDB_ID pn = pgno << 1;
5109 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5110 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5111 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5116 unsigned x = mdb_mid2l_search(dl, pgno);
5117 if (x <= dl[0].mid && dl[x].mid == pgno) {
5123 } while ((tx2 = tx2->mt_parent) != NULL);
5126 if (pgno < txn->mt_next_pgno) {
5128 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5130 DPRINTF(("page %"Z"u not found", pgno));
5131 txn->mt_flags |= MDB_TXN_ERROR;
5132 return MDB_PAGE_NOTFOUND;
5142 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5143 * The cursor is at the root page, set up the rest of it.
5146 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5148 MDB_page *mp = mc->mc_pg[mc->mc_top];
5152 while (IS_BRANCH(mp)) {
5156 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5157 mdb_cassert(mc, NUMKEYS(mp) > 1);
5158 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5160 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5162 if (flags & MDB_PS_LAST)
5163 i = NUMKEYS(mp) - 1;
5166 node = mdb_node_search(mc, key, &exact);
5168 i = NUMKEYS(mp) - 1;
5170 i = mc->mc_ki[mc->mc_top];
5172 mdb_cassert(mc, i > 0);
5176 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5179 mdb_cassert(mc, i < NUMKEYS(mp));
5180 node = NODEPTR(mp, i);
5182 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5185 mc->mc_ki[mc->mc_top] = i;
5186 if ((rc = mdb_cursor_push(mc, mp)))
5189 if (flags & MDB_PS_MODIFY) {
5190 if ((rc = mdb_page_touch(mc)) != 0)
5192 mp = mc->mc_pg[mc->mc_top];
5197 DPRINTF(("internal error, index points to a %02X page!?",
5199 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5200 return MDB_CORRUPTED;
5203 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5204 key ? DKEY(key) : "null"));
5205 mc->mc_flags |= C_INITIALIZED;
5206 mc->mc_flags &= ~C_EOF;
5211 /** Search for the lowest key under the current branch page.
5212 * This just bypasses a NUMKEYS check in the current page
5213 * before calling mdb_page_search_root(), because the callers
5214 * are all in situations where the current page is known to
5218 mdb_page_search_lowest(MDB_cursor *mc)
5220 MDB_page *mp = mc->mc_pg[mc->mc_top];
5221 MDB_node *node = NODEPTR(mp, 0);
5224 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5227 mc->mc_ki[mc->mc_top] = 0;
5228 if ((rc = mdb_cursor_push(mc, mp)))
5230 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5233 /** Search for the page a given key should be in.
5234 * Push it and its parent pages on the cursor stack.
5235 * @param[in,out] mc the cursor for this operation.
5236 * @param[in] key the key to search for, or NULL for first/last page.
5237 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5238 * are touched (updated with new page numbers).
5239 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5240 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5241 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5242 * @return 0 on success, non-zero on failure.
5245 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5250 /* Make sure the txn is still viable, then find the root from
5251 * the txn's db table and set it as the root of the cursor's stack.
5253 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5254 DPUTS("transaction has failed, must abort");
5257 /* Make sure we're using an up-to-date root */
5258 if (*mc->mc_dbflag & DB_STALE) {
5260 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5262 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5263 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5270 MDB_node *leaf = mdb_node_search(&mc2,
5271 &mc->mc_dbx->md_name, &exact);
5273 return MDB_NOTFOUND;
5274 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5275 return MDB_INCOMPATIBLE; /* not a named DB */
5276 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5279 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5281 /* The txn may not know this DBI, or another process may
5282 * have dropped and recreated the DB with other flags.
5284 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5285 return MDB_INCOMPATIBLE;
5286 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5288 *mc->mc_dbflag &= ~DB_STALE;
5290 root = mc->mc_db->md_root;
5292 if (root == P_INVALID) { /* Tree is empty. */
5293 DPUTS("tree is empty");
5294 return MDB_NOTFOUND;
5298 mdb_cassert(mc, root > 1);
5299 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5300 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5306 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5307 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5309 if (flags & MDB_PS_MODIFY) {
5310 if ((rc = mdb_page_touch(mc)))
5314 if (flags & MDB_PS_ROOTONLY)
5317 return mdb_page_search_root(mc, key, flags);
5321 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5323 MDB_txn *txn = mc->mc_txn;
5324 pgno_t pg = mp->mp_pgno;
5325 unsigned x = 0, ovpages = mp->mp_pages;
5326 MDB_env *env = txn->mt_env;
5327 MDB_IDL sl = txn->mt_spill_pgs;
5328 MDB_ID pn = pg << 1;
5331 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5332 /* If the page is dirty or on the spill list we just acquired it,
5333 * so we should give it back to our current free list, if any.
5334 * Otherwise put it onto the list of pages we freed in this txn.
5336 * Won't create me_pghead: me_pglast must be inited along with it.
5337 * Unsupported in nested txns: They would need to hide the page
5338 * range in ancestor txns' dirty and spilled lists.
5340 if (env->me_pghead &&
5342 ((mp->mp_flags & P_DIRTY) ||
5343 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5347 MDB_ID2 *dl, ix, iy;
5348 rc = mdb_midl_need(&env->me_pghead, ovpages);
5351 if (!(mp->mp_flags & P_DIRTY)) {
5352 /* This page is no longer spilled */
5359 /* Remove from dirty list */
5360 dl = txn->mt_u.dirty_list;
5362 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5368 mdb_cassert(mc, x > 1);
5370 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5371 txn->mt_flags |= MDB_TXN_ERROR;
5372 return MDB_CORRUPTED;
5375 if (!(env->me_flags & MDB_WRITEMAP))
5376 mdb_dpage_free(env, mp);
5378 /* Insert in me_pghead */
5379 mop = env->me_pghead;
5380 j = mop[0] + ovpages;
5381 for (i = mop[0]; i && mop[i] < pg; i--)
5387 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5391 mc->mc_db->md_overflow_pages -= ovpages;
5395 /** Return the data associated with a given node.
5396 * @param[in] txn The transaction for this operation.
5397 * @param[in] leaf The node being read.
5398 * @param[out] data Updated to point to the node's data.
5399 * @return 0 on success, non-zero on failure.
5402 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5404 MDB_page *omp; /* overflow page */
5408 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5409 data->mv_size = NODEDSZ(leaf);
5410 data->mv_data = NODEDATA(leaf);
5414 /* Read overflow data.
5416 data->mv_size = NODEDSZ(leaf);
5417 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5418 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5419 DPRINTF(("read overflow page %"Z"u failed", pgno));
5422 data->mv_data = METADATA(omp);
5428 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5429 MDB_val *key, MDB_val *data)
5436 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5438 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5441 if (txn->mt_flags & MDB_TXN_ERROR)
5444 mdb_cursor_init(&mc, txn, dbi, &mx);
5445 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5448 /** Find a sibling for a page.
5449 * Replaces the page at the top of the cursor's stack with the
5450 * specified sibling, if one exists.
5451 * @param[in] mc The cursor for this operation.
5452 * @param[in] move_right Non-zero if the right sibling is requested,
5453 * otherwise the left sibling.
5454 * @return 0 on success, non-zero on failure.
5457 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5463 if (mc->mc_snum < 2) {
5464 return MDB_NOTFOUND; /* root has no siblings */
5468 DPRINTF(("parent page is page %"Z"u, index %u",
5469 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5471 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5472 : (mc->mc_ki[mc->mc_top] == 0)) {
5473 DPRINTF(("no more keys left, moving to %s sibling",
5474 move_right ? "right" : "left"));
5475 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5476 /* undo cursor_pop before returning */
5483 mc->mc_ki[mc->mc_top]++;
5485 mc->mc_ki[mc->mc_top]--;
5486 DPRINTF(("just moving to %s index key %u",
5487 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5489 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5491 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5492 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5493 /* mc will be inconsistent if caller does mc_snum++ as above */
5494 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5498 mdb_cursor_push(mc, mp);
5500 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5505 /** Move the cursor to the next data item. */
5507 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5513 if (mc->mc_flags & C_EOF) {
5514 return MDB_NOTFOUND;
5517 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5519 mp = mc->mc_pg[mc->mc_top];
5521 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5522 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5523 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5524 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5525 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5526 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5527 if (rc == MDB_SUCCESS)
5528 MDB_GET_KEY(leaf, key);
5533 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5534 if (op == MDB_NEXT_DUP)
5535 return MDB_NOTFOUND;
5539 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5540 mdb_dbg_pgno(mp), (void *) mc));
5541 if (mc->mc_flags & C_DEL)
5544 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5545 DPUTS("=====> move to next sibling page");
5546 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5547 mc->mc_flags |= C_EOF;
5550 mp = mc->mc_pg[mc->mc_top];
5551 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5553 mc->mc_ki[mc->mc_top]++;
5556 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5557 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5560 key->mv_size = mc->mc_db->md_pad;
5561 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5565 mdb_cassert(mc, IS_LEAF(mp));
5566 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5568 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5569 mdb_xcursor_init1(mc, leaf);
5572 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5575 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5576 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5577 if (rc != MDB_SUCCESS)
5582 MDB_GET_KEY(leaf, key);
5586 /** Move the cursor to the previous data item. */
5588 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5594 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5596 mp = mc->mc_pg[mc->mc_top];
5598 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5599 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5600 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5601 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5602 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5603 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5604 if (rc == MDB_SUCCESS) {
5605 MDB_GET_KEY(leaf, key);
5606 mc->mc_flags &= ~C_EOF;
5612 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5613 if (op == MDB_PREV_DUP)
5614 return MDB_NOTFOUND;
5618 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5619 mdb_dbg_pgno(mp), (void *) mc));
5621 if (mc->mc_ki[mc->mc_top] == 0) {
5622 DPUTS("=====> move to prev sibling page");
5623 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5626 mp = mc->mc_pg[mc->mc_top];
5627 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5628 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5630 mc->mc_ki[mc->mc_top]--;
5632 mc->mc_flags &= ~C_EOF;
5634 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5635 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5638 key->mv_size = mc->mc_db->md_pad;
5639 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5643 mdb_cassert(mc, IS_LEAF(mp));
5644 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5646 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5647 mdb_xcursor_init1(mc, leaf);
5650 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5653 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5654 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5655 if (rc != MDB_SUCCESS)
5660 MDB_GET_KEY(leaf, key);
5664 /** Set the cursor on a specific data item. */
5666 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5667 MDB_cursor_op op, int *exactp)
5671 MDB_node *leaf = NULL;
5674 if (key->mv_size == 0)
5675 return MDB_BAD_VALSIZE;
5678 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5680 /* See if we're already on the right page */
5681 if (mc->mc_flags & C_INITIALIZED) {
5684 mp = mc->mc_pg[mc->mc_top];
5686 mc->mc_ki[mc->mc_top] = 0;
5687 return MDB_NOTFOUND;
5689 if (mp->mp_flags & P_LEAF2) {
5690 nodekey.mv_size = mc->mc_db->md_pad;
5691 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5693 leaf = NODEPTR(mp, 0);
5694 MDB_GET_KEY2(leaf, nodekey);
5696 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5698 /* Probably happens rarely, but first node on the page
5699 * was the one we wanted.
5701 mc->mc_ki[mc->mc_top] = 0;
5708 unsigned int nkeys = NUMKEYS(mp);
5710 if (mp->mp_flags & P_LEAF2) {
5711 nodekey.mv_data = LEAF2KEY(mp,
5712 nkeys-1, nodekey.mv_size);
5714 leaf = NODEPTR(mp, nkeys-1);
5715 MDB_GET_KEY2(leaf, nodekey);
5717 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5719 /* last node was the one we wanted */
5720 mc->mc_ki[mc->mc_top] = nkeys-1;
5726 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5727 /* This is definitely the right page, skip search_page */
5728 if (mp->mp_flags & P_LEAF2) {
5729 nodekey.mv_data = LEAF2KEY(mp,
5730 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5732 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5733 MDB_GET_KEY2(leaf, nodekey);
5735 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5737 /* current node was the one we wanted */
5747 /* If any parents have right-sibs, search.
5748 * Otherwise, there's nothing further.
5750 for (i=0; i<mc->mc_top; i++)
5752 NUMKEYS(mc->mc_pg[i])-1)
5754 if (i == mc->mc_top) {
5755 /* There are no other pages */
5756 mc->mc_ki[mc->mc_top] = nkeys;
5757 return MDB_NOTFOUND;
5761 /* There are no other pages */
5762 mc->mc_ki[mc->mc_top] = 0;
5763 if (op == MDB_SET_RANGE && !exactp) {
5767 return MDB_NOTFOUND;
5771 rc = mdb_page_search(mc, key, 0);
5772 if (rc != MDB_SUCCESS)
5775 mp = mc->mc_pg[mc->mc_top];
5776 mdb_cassert(mc, IS_LEAF(mp));
5779 leaf = mdb_node_search(mc, key, exactp);
5780 if (exactp != NULL && !*exactp) {
5781 /* MDB_SET specified and not an exact match. */
5782 return MDB_NOTFOUND;
5786 DPUTS("===> inexact leaf not found, goto sibling");
5787 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5788 mc->mc_flags |= C_EOF;
5789 return rc; /* no entries matched */
5791 mp = mc->mc_pg[mc->mc_top];
5792 mdb_cassert(mc, IS_LEAF(mp));
5793 leaf = NODEPTR(mp, 0);
5797 mc->mc_flags |= C_INITIALIZED;
5798 mc->mc_flags &= ~C_EOF;
5801 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5802 key->mv_size = mc->mc_db->md_pad;
5803 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5808 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5809 mdb_xcursor_init1(mc, leaf);
5812 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5813 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5814 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5817 if (op == MDB_GET_BOTH) {
5823 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5824 if (rc != MDB_SUCCESS)
5827 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5830 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5832 dcmp = mc->mc_dbx->md_dcmp;
5833 #if UINT_MAX < SIZE_MAX
5834 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5835 dcmp = mdb_cmp_clong;
5837 rc = dcmp(data, &olddata);
5839 if (op == MDB_GET_BOTH || rc > 0)
5840 return MDB_NOTFOUND;
5847 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5848 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5853 /* The key already matches in all other cases */
5854 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5855 MDB_GET_KEY(leaf, key);
5856 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5861 /** Move the cursor to the first item in the database. */
5863 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5869 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5871 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5872 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5873 if (rc != MDB_SUCCESS)
5876 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5878 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5879 mc->mc_flags |= C_INITIALIZED;
5880 mc->mc_flags &= ~C_EOF;
5882 mc->mc_ki[mc->mc_top] = 0;
5884 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5885 key->mv_size = mc->mc_db->md_pad;
5886 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5891 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5892 mdb_xcursor_init1(mc, leaf);
5893 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5897 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5901 MDB_GET_KEY(leaf, key);
5905 /** Move the cursor to the last item in the database. */
5907 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5913 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5915 if (!(mc->mc_flags & C_EOF)) {
5917 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5918 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5919 if (rc != MDB_SUCCESS)
5922 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5925 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5926 mc->mc_flags |= C_INITIALIZED|C_EOF;
5927 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5929 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5930 key->mv_size = mc->mc_db->md_pad;
5931 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5936 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5937 mdb_xcursor_init1(mc, leaf);
5938 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5942 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5947 MDB_GET_KEY(leaf, key);
5952 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5957 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5962 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5966 case MDB_GET_CURRENT:
5967 if (!(mc->mc_flags & C_INITIALIZED)) {
5970 MDB_page *mp = mc->mc_pg[mc->mc_top];
5971 int nkeys = NUMKEYS(mp);
5972 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5973 mc->mc_ki[mc->mc_top] = nkeys;
5979 key->mv_size = mc->mc_db->md_pad;
5980 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5982 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5983 MDB_GET_KEY(leaf, key);
5985 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5986 if (mc->mc_flags & C_DEL)
5987 mdb_xcursor_init1(mc, leaf);
5988 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5990 rc = mdb_node_read(mc->mc_txn, leaf, data);
5997 case MDB_GET_BOTH_RANGE:
6002 if (mc->mc_xcursor == NULL) {
6003 rc = MDB_INCOMPATIBLE;
6013 rc = mdb_cursor_set(mc, key, data, op,
6014 op == MDB_SET_RANGE ? NULL : &exact);
6017 case MDB_GET_MULTIPLE:
6018 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6022 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6023 rc = MDB_INCOMPATIBLE;
6027 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
6028 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
6031 case MDB_NEXT_MULTIPLE:
6036 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6037 rc = MDB_INCOMPATIBLE;
6040 if (!(mc->mc_flags & C_INITIALIZED))
6041 rc = mdb_cursor_first(mc, key, data);
6043 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
6044 if (rc == MDB_SUCCESS) {
6045 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
6048 mx = &mc->mc_xcursor->mx_cursor;
6049 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
6051 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
6052 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6060 case MDB_NEXT_NODUP:
6061 if (!(mc->mc_flags & C_INITIALIZED))
6062 rc = mdb_cursor_first(mc, key, data);
6064 rc = mdb_cursor_next(mc, key, data, op);
6068 case MDB_PREV_NODUP:
6069 if (!(mc->mc_flags & C_INITIALIZED)) {
6070 rc = mdb_cursor_last(mc, key, data);
6073 mc->mc_flags |= C_INITIALIZED;
6074 mc->mc_ki[mc->mc_top]++;
6076 rc = mdb_cursor_prev(mc, key, data, op);
6079 rc = mdb_cursor_first(mc, key, data);
6082 mfunc = mdb_cursor_first;
6084 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6088 if (mc->mc_xcursor == NULL) {
6089 rc = MDB_INCOMPATIBLE;
6093 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6094 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6095 MDB_GET_KEY(leaf, key);
6096 rc = mdb_node_read(mc->mc_txn, leaf, data);
6100 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6104 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6107 rc = mdb_cursor_last(mc, key, data);
6110 mfunc = mdb_cursor_last;
6113 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6118 if (mc->mc_flags & C_DEL)
6119 mc->mc_flags ^= C_DEL;
6124 /** Touch all the pages in the cursor stack. Set mc_top.
6125 * Makes sure all the pages are writable, before attempting a write operation.
6126 * @param[in] mc The cursor to operate on.
6129 mdb_cursor_touch(MDB_cursor *mc)
6131 int rc = MDB_SUCCESS;
6133 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6136 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6138 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6139 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6142 *mc->mc_dbflag |= DB_DIRTY;
6147 rc = mdb_page_touch(mc);
6148 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6149 mc->mc_top = mc->mc_snum-1;
6154 /** Do not spill pages to disk if txn is getting full, may fail instead */
6155 #define MDB_NOSPILL 0x8000
6158 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6161 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
6163 MDB_node *leaf = NULL;
6164 MDB_page *fp, *mp, *sub_root = NULL;
6166 MDB_val xdata, *rdata, dkey, olddata;
6168 int do_sub = 0, insert_key, insert_data;
6169 unsigned int mcount = 0, dcount = 0, nospill;
6172 unsigned int nflags;
6175 if (mc == NULL || key == NULL)
6178 env = mc->mc_txn->mt_env;
6180 /* Check this first so counter will always be zero on any
6183 if (flags & MDB_MULTIPLE) {
6184 dcount = data[1].mv_size;
6185 data[1].mv_size = 0;
6186 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6187 return MDB_INCOMPATIBLE;
6190 nospill = flags & MDB_NOSPILL;
6191 flags &= ~MDB_NOSPILL;
6193 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6194 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6196 if (key->mv_size-1 >= ENV_MAXKEY(env))
6197 return MDB_BAD_VALSIZE;
6199 #if SIZE_MAX > MAXDATASIZE
6200 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6201 return MDB_BAD_VALSIZE;
6203 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6204 return MDB_BAD_VALSIZE;
6207 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6208 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6212 if (flags == MDB_CURRENT) {
6213 if (!(mc->mc_flags & C_INITIALIZED))
6216 } else if (mc->mc_db->md_root == P_INVALID) {
6217 /* new database, cursor has nothing to point to */
6220 mc->mc_flags &= ~C_INITIALIZED;
6225 if (flags & MDB_APPEND) {
6227 rc = mdb_cursor_last(mc, &k2, &d2);
6229 rc = mc->mc_dbx->md_cmp(key, &k2);
6232 mc->mc_ki[mc->mc_top]++;
6234 /* new key is <= last key */
6239 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6241 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6242 DPRINTF(("duplicate key [%s]", DKEY(key)));
6244 return MDB_KEYEXIST;
6246 if (rc && rc != MDB_NOTFOUND)
6250 if (mc->mc_flags & C_DEL)
6251 mc->mc_flags ^= C_DEL;
6253 /* Cursor is positioned, check for room in the dirty list */
6255 if (flags & MDB_MULTIPLE) {
6257 xdata.mv_size = data->mv_size * dcount;
6261 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6265 if (rc == MDB_NO_ROOT) {
6267 /* new database, write a root leaf page */
6268 DPUTS("allocating new root leaf page");
6269 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6272 mdb_cursor_push(mc, np);
6273 mc->mc_db->md_root = np->mp_pgno;
6274 mc->mc_db->md_depth++;
6275 *mc->mc_dbflag |= DB_DIRTY;
6276 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6278 np->mp_flags |= P_LEAF2;
6279 mc->mc_flags |= C_INITIALIZED;
6281 /* make sure all cursor pages are writable */
6282 rc2 = mdb_cursor_touch(mc);
6287 insert_key = insert_data = rc;
6289 /* The key does not exist */
6290 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6291 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6292 LEAFSIZE(key, data) > env->me_nodemax)
6294 /* Too big for a node, insert in sub-DB. Set up an empty
6295 * "old sub-page" for prep_subDB to expand to a full page.
6297 fp_flags = P_LEAF|P_DIRTY;
6299 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6300 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6301 olddata.mv_size = PAGEHDRSZ;
6305 /* there's only a key anyway, so this is a no-op */
6306 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6308 unsigned int ksize = mc->mc_db->md_pad;
6309 if (key->mv_size != ksize)
6310 return MDB_BAD_VALSIZE;
6311 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6312 memcpy(ptr, key->mv_data, ksize);
6314 /* if overwriting slot 0 of leaf, need to
6315 * update branch key if there is a parent page
6317 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6318 unsigned short top = mc->mc_top;
6320 /* slot 0 is always an empty key, find real slot */
6321 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6323 if (mc->mc_ki[mc->mc_top])
6324 rc2 = mdb_update_key(mc, key);
6335 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6336 olddata.mv_size = NODEDSZ(leaf);
6337 olddata.mv_data = NODEDATA(leaf);
6340 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6341 /* Prepare (sub-)page/sub-DB to accept the new item,
6342 * if needed. fp: old sub-page or a header faking
6343 * it. mp: new (sub-)page. offset: growth in page
6344 * size. xdata: node data with new page or DB.
6346 unsigned i, offset = 0;
6347 mp = fp = xdata.mv_data = env->me_pbuf;
6348 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6350 /* Was a single item before, must convert now */
6351 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6353 /* Just overwrite the current item */
6354 if (flags == MDB_CURRENT)
6356 dcmp = mc->mc_dbx->md_dcmp;
6357 #if UINT_MAX < SIZE_MAX
6358 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6359 dcmp = mdb_cmp_clong;
6361 /* does data match? */
6362 if (!dcmp(data, &olddata)) {
6363 if (flags & MDB_NODUPDATA)
6364 return MDB_KEYEXIST;
6369 /* Back up original data item */
6370 dkey.mv_size = olddata.mv_size;
6371 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6373 /* Make sub-page header for the dup items, with dummy body */
6374 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6375 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6376 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6377 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6378 fp->mp_flags |= P_LEAF2;
6379 fp->mp_pad = data->mv_size;
6380 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6382 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6383 (dkey.mv_size & 1) + (data->mv_size & 1);
6385 fp->mp_upper = xdata.mv_size - PAGEBASE;
6386 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6387 } else if (leaf->mn_flags & F_SUBDATA) {
6388 /* Data is on sub-DB, just store it */
6389 flags |= F_DUPDATA|F_SUBDATA;
6392 /* Data is on sub-page */
6393 fp = olddata.mv_data;
6396 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6397 offset = EVEN(NODESIZE + sizeof(indx_t) +
6401 offset = fp->mp_pad;
6402 if (SIZELEFT(fp) < offset) {
6403 offset *= 4; /* space for 4 more */
6406 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6408 fp->mp_flags |= P_DIRTY;
6409 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6410 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6414 xdata.mv_size = olddata.mv_size + offset;
6417 fp_flags = fp->mp_flags;
6418 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6419 /* Too big for a sub-page, convert to sub-DB */
6420 fp_flags &= ~P_SUBP;
6422 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6423 fp_flags |= P_LEAF2;
6424 dummy.md_pad = fp->mp_pad;
6425 dummy.md_flags = MDB_DUPFIXED;
6426 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6427 dummy.md_flags |= MDB_INTEGERKEY;
6433 dummy.md_branch_pages = 0;
6434 dummy.md_leaf_pages = 1;
6435 dummy.md_overflow_pages = 0;
6436 dummy.md_entries = NUMKEYS(fp);
6437 xdata.mv_size = sizeof(MDB_db);
6438 xdata.mv_data = &dummy;
6439 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6441 offset = env->me_psize - olddata.mv_size;
6442 flags |= F_DUPDATA|F_SUBDATA;
6443 dummy.md_root = mp->mp_pgno;
6447 mp->mp_flags = fp_flags | P_DIRTY;
6448 mp->mp_pad = fp->mp_pad;
6449 mp->mp_lower = fp->mp_lower;
6450 mp->mp_upper = fp->mp_upper + offset;
6451 if (fp_flags & P_LEAF2) {
6452 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6454 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6455 olddata.mv_size - fp->mp_upper - PAGEBASE);
6456 for (i=0; i<NUMKEYS(fp); i++)
6457 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6465 mdb_node_del(mc, 0);
6469 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6470 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6471 return MDB_INCOMPATIBLE;
6472 /* overflow page overwrites need special handling */
6473 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6476 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6478 memcpy(&pg, olddata.mv_data, sizeof(pg));
6479 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6481 ovpages = omp->mp_pages;
6483 /* Is the ov page large enough? */
6484 if (ovpages >= dpages) {
6485 if (!(omp->mp_flags & P_DIRTY) &&
6486 (level || (env->me_flags & MDB_WRITEMAP)))
6488 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6491 level = 0; /* dirty in this txn or clean */
6494 if (omp->mp_flags & P_DIRTY) {
6495 /* yes, overwrite it. Note in this case we don't
6496 * bother to try shrinking the page if the new data
6497 * is smaller than the overflow threshold.
6500 /* It is writable only in a parent txn */
6501 size_t sz = (size_t) env->me_psize * ovpages, off;
6502 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6508 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6509 mdb_cassert(mc, rc2 == 0);
6510 if (!(flags & MDB_RESERVE)) {
6511 /* Copy end of page, adjusting alignment so
6512 * compiler may copy words instead of bytes.
6514 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6515 memcpy((size_t *)((char *)np + off),
6516 (size_t *)((char *)omp + off), sz - off);
6519 memcpy(np, omp, sz); /* Copy beginning of page */
6522 SETDSZ(leaf, data->mv_size);
6523 if (F_ISSET(flags, MDB_RESERVE))
6524 data->mv_data = METADATA(omp);
6526 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6530 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6532 } else if (data->mv_size == olddata.mv_size) {
6533 /* same size, just replace it. Note that we could
6534 * also reuse this node if the new data is smaller,
6535 * but instead we opt to shrink the node in that case.
6537 if (F_ISSET(flags, MDB_RESERVE))
6538 data->mv_data = olddata.mv_data;
6539 else if (!(mc->mc_flags & C_SUB))
6540 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6542 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6547 mdb_node_del(mc, 0);
6553 nflags = flags & NODE_ADD_FLAGS;
6554 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6555 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6556 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6557 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6559 nflags |= MDB_SPLIT_REPLACE;
6560 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6562 /* There is room already in this leaf page. */
6563 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6564 if (rc == 0 && insert_key) {
6565 /* Adjust other cursors pointing to mp */
6566 MDB_cursor *m2, *m3;
6567 MDB_dbi dbi = mc->mc_dbi;
6568 unsigned i = mc->mc_top;
6569 MDB_page *mp = mc->mc_pg[i];
6571 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6572 if (mc->mc_flags & C_SUB)
6573 m3 = &m2->mc_xcursor->mx_cursor;
6576 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6577 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6584 if (rc == MDB_SUCCESS) {
6585 /* Now store the actual data in the child DB. Note that we're
6586 * storing the user data in the keys field, so there are strict
6587 * size limits on dupdata. The actual data fields of the child
6588 * DB are all zero size.
6591 int xflags, new_dupdata;
6596 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6597 if (flags & MDB_CURRENT) {
6598 xflags = MDB_CURRENT|MDB_NOSPILL;
6600 mdb_xcursor_init1(mc, leaf);
6601 xflags = (flags & MDB_NODUPDATA) ?
6602 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6605 mc->mc_xcursor->mx_cursor.mc_pg[0] = sub_root;
6606 new_dupdata = (int)dkey.mv_size;
6607 /* converted, write the original data first */
6609 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6612 /* we've done our job */
6615 if (!(leaf->mn_flags & F_SUBDATA) || sub_root) {
6616 /* Adjust other cursors pointing to mp */
6618 MDB_xcursor *mx = mc->mc_xcursor;
6619 unsigned i = mc->mc_top;
6620 MDB_page *mp = mc->mc_pg[i];
6622 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6623 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6624 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6625 if (m2->mc_pg[i] == mp) {
6626 if (m2->mc_ki[i] == mc->mc_ki[i]) {
6627 mdb_xcursor_init2(m2, mx, new_dupdata);
6628 } else if (!insert_key) {
6629 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[i]);
6630 if (!(n2->mn_flags & F_SUBDATA))
6631 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6636 ecount = mc->mc_xcursor->mx_db.md_entries;
6637 if (flags & MDB_APPENDDUP)
6638 xflags |= MDB_APPEND;
6639 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6640 if (flags & F_SUBDATA) {
6641 void *db = NODEDATA(leaf);
6642 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6644 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6646 /* Increment count unless we just replaced an existing item. */
6648 mc->mc_db->md_entries++;
6650 /* Invalidate txn if we created an empty sub-DB */
6653 /* If we succeeded and the key didn't exist before,
6654 * make sure the cursor is marked valid.
6656 mc->mc_flags |= C_INITIALIZED;
6658 if (flags & MDB_MULTIPLE) {
6661 /* let caller know how many succeeded, if any */
6662 data[1].mv_size = mcount;
6663 if (mcount < dcount) {
6664 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6665 insert_key = insert_data = 0;
6672 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6675 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6680 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6686 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6687 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6689 if (!(mc->mc_flags & C_INITIALIZED))
6692 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6693 return MDB_NOTFOUND;
6695 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6698 rc = mdb_cursor_touch(mc);
6702 mp = mc->mc_pg[mc->mc_top];
6705 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6707 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6708 if (flags & MDB_NODUPDATA) {
6709 /* mdb_cursor_del0() will subtract the final entry */
6710 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6712 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6713 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6715 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6718 /* If sub-DB still has entries, we're done */
6719 if (mc->mc_xcursor->mx_db.md_entries) {
6720 if (leaf->mn_flags & F_SUBDATA) {
6721 /* update subDB info */
6722 void *db = NODEDATA(leaf);
6723 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6726 /* shrink fake page */
6727 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6728 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6729 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6730 /* fix other sub-DB cursors pointed at fake pages on this page */
6731 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6732 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6733 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6734 if (m2->mc_pg[mc->mc_top] == mp) {
6735 if (m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top]) {
6736 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6738 MDB_node *n2 = NODEPTR(mp, m2->mc_ki[mc->mc_top]);
6739 if (!(n2->mn_flags & F_SUBDATA))
6740 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(n2);
6745 mc->mc_db->md_entries--;
6746 mc->mc_flags |= C_DEL;
6749 /* otherwise fall thru and delete the sub-DB */
6752 if (leaf->mn_flags & F_SUBDATA) {
6753 /* add all the child DB's pages to the free list */
6754 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6759 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6760 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6761 rc = MDB_INCOMPATIBLE;
6765 /* add overflow pages to free list */
6766 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6770 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6771 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6772 (rc = mdb_ovpage_free(mc, omp)))
6777 return mdb_cursor_del0(mc);
6780 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6784 /** Allocate and initialize new pages for a database.
6785 * @param[in] mc a cursor on the database being added to.
6786 * @param[in] flags flags defining what type of page is being allocated.
6787 * @param[in] num the number of pages to allocate. This is usually 1,
6788 * unless allocating overflow pages for a large record.
6789 * @param[out] mp Address of a page, or NULL on failure.
6790 * @return 0 on success, non-zero on failure.
6793 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6798 if ((rc = mdb_page_alloc(mc, num, &np)))
6800 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6801 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6802 np->mp_flags = flags | P_DIRTY;
6803 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6804 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6807 mc->mc_db->md_branch_pages++;
6808 else if (IS_LEAF(np))
6809 mc->mc_db->md_leaf_pages++;
6810 else if (IS_OVERFLOW(np)) {
6811 mc->mc_db->md_overflow_pages += num;
6819 /** Calculate the size of a leaf node.
6820 * The size depends on the environment's page size; if a data item
6821 * is too large it will be put onto an overflow page and the node
6822 * size will only include the key and not the data. Sizes are always
6823 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6824 * of the #MDB_node headers.
6825 * @param[in] env The environment handle.
6826 * @param[in] key The key for the node.
6827 * @param[in] data The data for the node.
6828 * @return The number of bytes needed to store the node.
6831 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6835 sz = LEAFSIZE(key, data);
6836 if (sz > env->me_nodemax) {
6837 /* put on overflow page */
6838 sz -= data->mv_size - sizeof(pgno_t);
6841 return EVEN(sz + sizeof(indx_t));
6844 /** Calculate the size of a branch node.
6845 * The size should depend on the environment's page size but since
6846 * we currently don't support spilling large keys onto overflow
6847 * pages, it's simply the size of the #MDB_node header plus the
6848 * size of the key. Sizes are always rounded up to an even number
6849 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6850 * @param[in] env The environment handle.
6851 * @param[in] key The key for the node.
6852 * @return The number of bytes needed to store the node.
6855 mdb_branch_size(MDB_env *env, MDB_val *key)
6860 if (sz > env->me_nodemax) {
6861 /* put on overflow page */
6862 /* not implemented */
6863 /* sz -= key->size - sizeof(pgno_t); */
6866 return sz + sizeof(indx_t);
6869 /** Add a node to the page pointed to by the cursor.
6870 * @param[in] mc The cursor for this operation.
6871 * @param[in] indx The index on the page where the new node should be added.
6872 * @param[in] key The key for the new node.
6873 * @param[in] data The data for the new node, if any.
6874 * @param[in] pgno The page number, if adding a branch node.
6875 * @param[in] flags Flags for the node.
6876 * @return 0 on success, non-zero on failure. Possible errors are:
6878 * <li>ENOMEM - failed to allocate overflow pages for the node.
6879 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6880 * should never happen since all callers already calculate the
6881 * page's free space before calling this function.
6885 mdb_node_add(MDB_cursor *mc, indx_t indx,
6886 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6889 size_t node_size = NODESIZE;
6893 MDB_page *mp = mc->mc_pg[mc->mc_top];
6894 MDB_page *ofp = NULL; /* overflow page */
6897 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6899 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6900 IS_LEAF(mp) ? "leaf" : "branch",
6901 IS_SUBP(mp) ? "sub-" : "",
6902 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6903 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6906 /* Move higher keys up one slot. */
6907 int ksize = mc->mc_db->md_pad, dif;
6908 char *ptr = LEAF2KEY(mp, indx, ksize);
6909 dif = NUMKEYS(mp) - indx;
6911 memmove(ptr+ksize, ptr, dif*ksize);
6912 /* insert new key */
6913 memcpy(ptr, key->mv_data, ksize);
6915 /* Just using these for counting */
6916 mp->mp_lower += sizeof(indx_t);
6917 mp->mp_upper -= ksize - sizeof(indx_t);
6921 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6923 node_size += key->mv_size;
6925 mdb_cassert(mc, data);
6926 if (F_ISSET(flags, F_BIGDATA)) {
6927 /* Data already on overflow page. */
6928 node_size += sizeof(pgno_t);
6929 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6930 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6932 /* Put data on overflow page. */
6933 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6934 data->mv_size, node_size+data->mv_size));
6935 node_size = EVEN(node_size + sizeof(pgno_t));
6936 if ((ssize_t)node_size > room)
6938 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6940 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6944 node_size += data->mv_size;
6947 node_size = EVEN(node_size);
6948 if ((ssize_t)node_size > room)
6952 /* Move higher pointers up one slot. */
6953 for (i = NUMKEYS(mp); i > indx; i--)
6954 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6956 /* Adjust free space offsets. */
6957 ofs = mp->mp_upper - node_size;
6958 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6959 mp->mp_ptrs[indx] = ofs;
6961 mp->mp_lower += sizeof(indx_t);
6963 /* Write the node data. */
6964 node = NODEPTR(mp, indx);
6965 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6966 node->mn_flags = flags;
6968 SETDSZ(node,data->mv_size);
6973 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6976 mdb_cassert(mc, key);
6978 if (F_ISSET(flags, F_BIGDATA))
6979 memcpy(node->mn_data + key->mv_size, data->mv_data,
6981 else if (F_ISSET(flags, MDB_RESERVE))
6982 data->mv_data = node->mn_data + key->mv_size;
6984 memcpy(node->mn_data + key->mv_size, data->mv_data,
6987 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6989 if (F_ISSET(flags, MDB_RESERVE))
6990 data->mv_data = METADATA(ofp);
6992 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6999 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
7000 mdb_dbg_pgno(mp), NUMKEYS(mp)));
7001 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
7002 DPRINTF(("node size = %"Z"u", node_size));
7003 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7004 return MDB_PAGE_FULL;
7007 /** Delete the specified node from a page.
7008 * @param[in] mc Cursor pointing to the node to delete.
7009 * @param[in] ksize The size of a node. Only used if the page is
7010 * part of a #MDB_DUPFIXED database.
7013 mdb_node_del(MDB_cursor *mc, int ksize)
7015 MDB_page *mp = mc->mc_pg[mc->mc_top];
7016 indx_t indx = mc->mc_ki[mc->mc_top];
7018 indx_t i, j, numkeys, ptr;
7022 DPRINTF(("delete node %u on %s page %"Z"u", indx,
7023 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
7024 numkeys = NUMKEYS(mp);
7025 mdb_cassert(mc, indx < numkeys);
7028 int x = numkeys - 1 - indx;
7029 base = LEAF2KEY(mp, indx, ksize);
7031 memmove(base, base + ksize, x * ksize);
7032 mp->mp_lower -= sizeof(indx_t);
7033 mp->mp_upper += ksize - sizeof(indx_t);
7037 node = NODEPTR(mp, indx);
7038 sz = NODESIZE + node->mn_ksize;
7040 if (F_ISSET(node->mn_flags, F_BIGDATA))
7041 sz += sizeof(pgno_t);
7043 sz += NODEDSZ(node);
7047 ptr = mp->mp_ptrs[indx];
7048 for (i = j = 0; i < numkeys; i++) {
7050 mp->mp_ptrs[j] = mp->mp_ptrs[i];
7051 if (mp->mp_ptrs[i] < ptr)
7052 mp->mp_ptrs[j] += sz;
7057 base = (char *)mp + mp->mp_upper + PAGEBASE;
7058 memmove(base + sz, base, ptr - mp->mp_upper);
7060 mp->mp_lower -= sizeof(indx_t);
7064 /** Compact the main page after deleting a node on a subpage.
7065 * @param[in] mp The main page to operate on.
7066 * @param[in] indx The index of the subpage on the main page.
7069 mdb_node_shrink(MDB_page *mp, indx_t indx)
7075 indx_t i, numkeys, ptr;
7077 node = NODEPTR(mp, indx);
7078 sp = (MDB_page *)NODEDATA(node);
7079 delta = SIZELEFT(sp);
7080 xp = (MDB_page *)((char *)sp + delta);
7082 /* shift subpage upward */
7084 nsize = NUMKEYS(sp) * sp->mp_pad;
7086 return; /* do not make the node uneven-sized */
7087 memmove(METADATA(xp), METADATA(sp), nsize);
7090 numkeys = NUMKEYS(sp);
7091 for (i=numkeys-1; i>=0; i--)
7092 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7094 xp->mp_upper = sp->mp_lower;
7095 xp->mp_lower = sp->mp_lower;
7096 xp->mp_flags = sp->mp_flags;
7097 xp->mp_pad = sp->mp_pad;
7098 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
7100 nsize = NODEDSZ(node) - delta;
7101 SETDSZ(node, nsize);
7103 /* shift lower nodes upward */
7104 ptr = mp->mp_ptrs[indx];
7105 numkeys = NUMKEYS(mp);
7106 for (i = 0; i < numkeys; i++) {
7107 if (mp->mp_ptrs[i] <= ptr)
7108 mp->mp_ptrs[i] += delta;
7111 base = (char *)mp + mp->mp_upper + PAGEBASE;
7112 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7113 mp->mp_upper += delta;
7116 /** Initial setup of a sorted-dups cursor.
7117 * Sorted duplicates are implemented as a sub-database for the given key.
7118 * The duplicate data items are actually keys of the sub-database.
7119 * Operations on the duplicate data items are performed using a sub-cursor
7120 * initialized when the sub-database is first accessed. This function does
7121 * the preliminary setup of the sub-cursor, filling in the fields that
7122 * depend only on the parent DB.
7123 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7126 mdb_xcursor_init0(MDB_cursor *mc)
7128 MDB_xcursor *mx = mc->mc_xcursor;
7130 mx->mx_cursor.mc_xcursor = NULL;
7131 mx->mx_cursor.mc_txn = mc->mc_txn;
7132 mx->mx_cursor.mc_db = &mx->mx_db;
7133 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7134 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7135 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7136 mx->mx_cursor.mc_snum = 0;
7137 mx->mx_cursor.mc_top = 0;
7138 mx->mx_cursor.mc_flags = C_SUB;
7139 mx->mx_dbx.md_name.mv_size = 0;
7140 mx->mx_dbx.md_name.mv_data = NULL;
7141 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7142 mx->mx_dbx.md_dcmp = NULL;
7143 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7146 /** Final setup of a sorted-dups cursor.
7147 * Sets up the fields that depend on the data from the main cursor.
7148 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7149 * @param[in] node The data containing the #MDB_db record for the
7150 * sorted-dup database.
7153 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7155 MDB_xcursor *mx = mc->mc_xcursor;
7157 if (node->mn_flags & F_SUBDATA) {
7158 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7159 mx->mx_cursor.mc_pg[0] = 0;
7160 mx->mx_cursor.mc_snum = 0;
7161 mx->mx_cursor.mc_top = 0;
7162 mx->mx_cursor.mc_flags = C_SUB;
7164 MDB_page *fp = NODEDATA(node);
7165 mx->mx_db.md_pad = 0;
7166 mx->mx_db.md_flags = 0;
7167 mx->mx_db.md_depth = 1;
7168 mx->mx_db.md_branch_pages = 0;
7169 mx->mx_db.md_leaf_pages = 1;
7170 mx->mx_db.md_overflow_pages = 0;
7171 mx->mx_db.md_entries = NUMKEYS(fp);
7172 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7173 mx->mx_cursor.mc_snum = 1;
7174 mx->mx_cursor.mc_top = 0;
7175 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7176 mx->mx_cursor.mc_pg[0] = fp;
7177 mx->mx_cursor.mc_ki[0] = 0;
7178 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7179 mx->mx_db.md_flags = MDB_DUPFIXED;
7180 mx->mx_db.md_pad = fp->mp_pad;
7181 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7182 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7185 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7186 mx->mx_db.md_root));
7187 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7188 #if UINT_MAX < SIZE_MAX
7189 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7190 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7195 /** Fixup a sorted-dups cursor due to underlying update.
7196 * Sets up some fields that depend on the data from the main cursor.
7197 * Almost the same as init1, but skips initialization steps if the
7198 * xcursor had already been used.
7199 * @param[in] mc The main cursor whose sorted-dups cursor is to be fixed up.
7200 * @param[in] src_mx The xcursor of an up-to-date cursor.
7201 * @param[in] new_dupdata True if converting from a non-#F_DUPDATA item.
7204 mdb_xcursor_init2(MDB_cursor *mc, MDB_xcursor *src_mx, int new_dupdata)
7206 MDB_xcursor *mx = mc->mc_xcursor;
7209 mx->mx_cursor.mc_snum = 1;
7210 mx->mx_cursor.mc_top = 0;
7211 mx->mx_cursor.mc_flags |= C_INITIALIZED;
7212 mx->mx_cursor.mc_ki[0] = 0;
7213 mx->mx_dbflag = DB_VALID|DB_USRVALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7214 #if UINT_MAX < SIZE_MAX
7215 mx->mx_dbx.md_cmp = src_mx->mx_dbx.md_cmp;
7217 } else if (!(mx->mx_cursor.mc_flags & C_INITIALIZED)) {
7220 mx->mx_db = src_mx->mx_db;
7221 mx->mx_cursor.mc_pg[0] = src_mx->mx_cursor.mc_pg[0];
7222 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7223 mx->mx_db.md_root));
7226 /** Initialize a cursor for a given transaction and database. */
7228 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7231 mc->mc_backup = NULL;
7234 mc->mc_db = &txn->mt_dbs[dbi];
7235 mc->mc_dbx = &txn->mt_dbxs[dbi];
7236 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7242 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7243 mdb_tassert(txn, mx != NULL);
7244 mc->mc_xcursor = mx;
7245 mdb_xcursor_init0(mc);
7247 mc->mc_xcursor = NULL;
7249 if (*mc->mc_dbflag & DB_STALE) {
7250 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7255 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7258 size_t size = sizeof(MDB_cursor);
7260 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7263 if (txn->mt_flags & MDB_TXN_ERROR)
7266 /* Allow read access to the freelist */
7267 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7270 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7271 size += sizeof(MDB_xcursor);
7273 if ((mc = malloc(size)) != NULL) {
7274 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7275 if (txn->mt_cursors) {
7276 mc->mc_next = txn->mt_cursors[dbi];
7277 txn->mt_cursors[dbi] = mc;
7278 mc->mc_flags |= C_UNTRACK;
7290 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7292 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7295 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7298 if (txn->mt_flags & MDB_TXN_ERROR)
7301 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7305 /* Return the count of duplicate data items for the current key */
7307 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7311 if (mc == NULL || countp == NULL)
7314 if (mc->mc_xcursor == NULL)
7315 return MDB_INCOMPATIBLE;
7317 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7320 if (!(mc->mc_flags & C_INITIALIZED))
7323 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7324 return MDB_NOTFOUND;
7326 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7327 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7330 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7333 *countp = mc->mc_xcursor->mx_db.md_entries;
7339 mdb_cursor_close(MDB_cursor *mc)
7341 if (mc && !mc->mc_backup) {
7342 /* remove from txn, if tracked */
7343 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7344 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7345 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7347 *prev = mc->mc_next;
7354 mdb_cursor_txn(MDB_cursor *mc)
7356 if (!mc) return NULL;
7361 mdb_cursor_dbi(MDB_cursor *mc)
7366 /** Replace the key for a branch node with a new key.
7367 * @param[in] mc Cursor pointing to the node to operate on.
7368 * @param[in] key The new key to use.
7369 * @return 0 on success, non-zero on failure.
7372 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7378 int delta, ksize, oksize;
7379 indx_t ptr, i, numkeys, indx;
7382 indx = mc->mc_ki[mc->mc_top];
7383 mp = mc->mc_pg[mc->mc_top];
7384 node = NODEPTR(mp, indx);
7385 ptr = mp->mp_ptrs[indx];
7389 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7390 k2.mv_data = NODEKEY(node);
7391 k2.mv_size = node->mn_ksize;
7392 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7394 mdb_dkey(&k2, kbuf2),
7400 /* Sizes must be 2-byte aligned. */
7401 ksize = EVEN(key->mv_size);
7402 oksize = EVEN(node->mn_ksize);
7403 delta = ksize - oksize;
7405 /* Shift node contents if EVEN(key length) changed. */
7407 if (delta > 0 && SIZELEFT(mp) < delta) {
7409 /* not enough space left, do a delete and split */
7410 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7411 pgno = NODEPGNO(node);
7412 mdb_node_del(mc, 0);
7413 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7416 numkeys = NUMKEYS(mp);
7417 for (i = 0; i < numkeys; i++) {
7418 if (mp->mp_ptrs[i] <= ptr)
7419 mp->mp_ptrs[i] -= delta;
7422 base = (char *)mp + mp->mp_upper + PAGEBASE;
7423 len = ptr - mp->mp_upper + NODESIZE;
7424 memmove(base - delta, base, len);
7425 mp->mp_upper -= delta;
7427 node = NODEPTR(mp, indx);
7430 /* But even if no shift was needed, update ksize */
7431 if (node->mn_ksize != key->mv_size)
7432 node->mn_ksize = key->mv_size;
7435 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7441 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7443 /** Move a node from csrc to cdst.
7446 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7453 unsigned short flags;
7457 /* Mark src and dst as dirty. */
7458 if ((rc = mdb_page_touch(csrc)) ||
7459 (rc = mdb_page_touch(cdst)))
7462 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7463 key.mv_size = csrc->mc_db->md_pad;
7464 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7466 data.mv_data = NULL;
7470 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7471 mdb_cassert(csrc, !((size_t)srcnode & 1));
7472 srcpg = NODEPGNO(srcnode);
7473 flags = srcnode->mn_flags;
7474 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7475 unsigned int snum = csrc->mc_snum;
7477 /* must find the lowest key below src */
7478 rc = mdb_page_search_lowest(csrc);
7481 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7482 key.mv_size = csrc->mc_db->md_pad;
7483 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7485 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7486 key.mv_size = NODEKSZ(s2);
7487 key.mv_data = NODEKEY(s2);
7489 csrc->mc_snum = snum--;
7490 csrc->mc_top = snum;
7492 key.mv_size = NODEKSZ(srcnode);
7493 key.mv_data = NODEKEY(srcnode);
7495 data.mv_size = NODEDSZ(srcnode);
7496 data.mv_data = NODEDATA(srcnode);
7498 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7499 unsigned int snum = cdst->mc_snum;
7502 /* must find the lowest key below dst */
7503 mdb_cursor_copy(cdst, &mn);
7504 rc = mdb_page_search_lowest(&mn);
7507 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7508 bkey.mv_size = mn.mc_db->md_pad;
7509 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7511 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7512 bkey.mv_size = NODEKSZ(s2);
7513 bkey.mv_data = NODEKEY(s2);
7515 mn.mc_snum = snum--;
7518 rc = mdb_update_key(&mn, &bkey);
7523 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7524 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7525 csrc->mc_ki[csrc->mc_top],
7527 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7528 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7530 /* Add the node to the destination page.
7532 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7533 if (rc != MDB_SUCCESS)
7536 /* Delete the node from the source page.
7538 mdb_node_del(csrc, key.mv_size);
7541 /* Adjust other cursors pointing to mp */
7542 MDB_cursor *m2, *m3;
7543 MDB_dbi dbi = csrc->mc_dbi;
7546 mp = cdst->mc_pg[csrc->mc_top];
7547 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7548 if (csrc->mc_flags & C_SUB)
7549 m3 = &m2->mc_xcursor->mx_cursor;
7552 if (m3 == cdst) continue;
7553 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] >=
7554 cdst->mc_ki[csrc->mc_top]) {
7555 m3->mc_ki[csrc->mc_top]++;
7559 mp = csrc->mc_pg[csrc->mc_top];
7560 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7561 if (csrc->mc_flags & C_SUB)
7562 m3 = &m2->mc_xcursor->mx_cursor;
7565 if (m3 == csrc) continue;
7566 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7567 csrc->mc_ki[csrc->mc_top]) {
7568 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7569 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7574 /* Update the parent separators.
7576 if (csrc->mc_ki[csrc->mc_top] == 0) {
7577 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7578 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7579 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7581 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7582 key.mv_size = NODEKSZ(srcnode);
7583 key.mv_data = NODEKEY(srcnode);
7585 DPRINTF(("update separator for source page %"Z"u to [%s]",
7586 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7587 mdb_cursor_copy(csrc, &mn);
7590 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7593 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7595 indx_t ix = csrc->mc_ki[csrc->mc_top];
7596 nullkey.mv_size = 0;
7597 csrc->mc_ki[csrc->mc_top] = 0;
7598 rc = mdb_update_key(csrc, &nullkey);
7599 csrc->mc_ki[csrc->mc_top] = ix;
7600 mdb_cassert(csrc, rc == MDB_SUCCESS);
7604 if (cdst->mc_ki[cdst->mc_top] == 0) {
7605 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7606 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7607 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7609 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7610 key.mv_size = NODEKSZ(srcnode);
7611 key.mv_data = NODEKEY(srcnode);
7613 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7614 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7615 mdb_cursor_copy(cdst, &mn);
7618 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7621 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7623 indx_t ix = cdst->mc_ki[cdst->mc_top];
7624 nullkey.mv_size = 0;
7625 cdst->mc_ki[cdst->mc_top] = 0;
7626 rc = mdb_update_key(cdst, &nullkey);
7627 cdst->mc_ki[cdst->mc_top] = ix;
7628 mdb_cassert(cdst, rc == MDB_SUCCESS);
7635 /** Merge one page into another.
7636 * The nodes from the page pointed to by \b csrc will
7637 * be copied to the page pointed to by \b cdst and then
7638 * the \b csrc page will be freed.
7639 * @param[in] csrc Cursor pointing to the source page.
7640 * @param[in] cdst Cursor pointing to the destination page.
7641 * @return 0 on success, non-zero on failure.
7644 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7646 MDB_page *psrc, *pdst;
7653 psrc = csrc->mc_pg[csrc->mc_top];
7654 pdst = cdst->mc_pg[cdst->mc_top];
7656 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7658 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7659 mdb_cassert(csrc, cdst->mc_snum > 1);
7661 /* Mark dst as dirty. */
7662 if ((rc = mdb_page_touch(cdst)))
7665 /* Move all nodes from src to dst.
7667 j = nkeys = NUMKEYS(pdst);
7668 if (IS_LEAF2(psrc)) {
7669 key.mv_size = csrc->mc_db->md_pad;
7670 key.mv_data = METADATA(psrc);
7671 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7672 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7673 if (rc != MDB_SUCCESS)
7675 key.mv_data = (char *)key.mv_data + key.mv_size;
7678 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7679 srcnode = NODEPTR(psrc, i);
7680 if (i == 0 && IS_BRANCH(psrc)) {
7683 mdb_cursor_copy(csrc, &mn);
7684 /* must find the lowest key below src */
7685 rc = mdb_page_search_lowest(&mn);
7688 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7689 key.mv_size = mn.mc_db->md_pad;
7690 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7692 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7693 key.mv_size = NODEKSZ(s2);
7694 key.mv_data = NODEKEY(s2);
7697 key.mv_size = srcnode->mn_ksize;
7698 key.mv_data = NODEKEY(srcnode);
7701 data.mv_size = NODEDSZ(srcnode);
7702 data.mv_data = NODEDATA(srcnode);
7703 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7704 if (rc != MDB_SUCCESS)
7709 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7710 pdst->mp_pgno, NUMKEYS(pdst),
7711 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7713 /* Unlink the src page from parent and add to free list.
7716 mdb_node_del(csrc, 0);
7717 if (csrc->mc_ki[csrc->mc_top] == 0) {
7719 rc = mdb_update_key(csrc, &key);
7727 psrc = csrc->mc_pg[csrc->mc_top];
7728 /* If not operating on FreeDB, allow this page to be reused
7729 * in this txn. Otherwise just add to free list.
7731 rc = mdb_page_loose(csrc, psrc);
7735 csrc->mc_db->md_leaf_pages--;
7737 csrc->mc_db->md_branch_pages--;
7739 /* Adjust other cursors pointing to mp */
7740 MDB_cursor *m2, *m3;
7741 MDB_dbi dbi = csrc->mc_dbi;
7743 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7744 if (csrc->mc_flags & C_SUB)
7745 m3 = &m2->mc_xcursor->mx_cursor;
7748 if (m3 == csrc) continue;
7749 if (m3->mc_snum < csrc->mc_snum) continue;
7750 if (m3->mc_pg[csrc->mc_top] == psrc) {
7751 m3->mc_pg[csrc->mc_top] = pdst;
7752 m3->mc_ki[csrc->mc_top] += nkeys;
7757 unsigned int snum = cdst->mc_snum;
7758 uint16_t depth = cdst->mc_db->md_depth;
7759 mdb_cursor_pop(cdst);
7760 rc = mdb_rebalance(cdst);
7761 /* Did the tree height change? */
7762 if (depth != cdst->mc_db->md_depth)
7763 snum += cdst->mc_db->md_depth - depth;
7764 cdst->mc_snum = snum;
7765 cdst->mc_top = snum-1;
7770 /** Copy the contents of a cursor.
7771 * @param[in] csrc The cursor to copy from.
7772 * @param[out] cdst The cursor to copy to.
7775 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7779 cdst->mc_txn = csrc->mc_txn;
7780 cdst->mc_dbi = csrc->mc_dbi;
7781 cdst->mc_db = csrc->mc_db;
7782 cdst->mc_dbx = csrc->mc_dbx;
7783 cdst->mc_snum = csrc->mc_snum;
7784 cdst->mc_top = csrc->mc_top;
7785 cdst->mc_flags = csrc->mc_flags;
7787 for (i=0; i<csrc->mc_snum; i++) {
7788 cdst->mc_pg[i] = csrc->mc_pg[i];
7789 cdst->mc_ki[i] = csrc->mc_ki[i];
7793 /** Rebalance the tree after a delete operation.
7794 * @param[in] mc Cursor pointing to the page where rebalancing
7796 * @return 0 on success, non-zero on failure.
7799 mdb_rebalance(MDB_cursor *mc)
7803 unsigned int ptop, minkeys, thresh;
7807 if (IS_BRANCH(mc->mc_pg[mc->mc_top])) {
7812 thresh = FILL_THRESHOLD;
7814 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7815 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7816 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7817 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7819 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= thresh &&
7820 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7821 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7822 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7826 if (mc->mc_snum < 2) {
7827 MDB_page *mp = mc->mc_pg[0];
7829 DPUTS("Can't rebalance a subpage, ignoring");
7832 if (NUMKEYS(mp) == 0) {
7833 DPUTS("tree is completely empty");
7834 mc->mc_db->md_root = P_INVALID;
7835 mc->mc_db->md_depth = 0;
7836 mc->mc_db->md_leaf_pages = 0;
7837 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7840 /* Adjust cursors pointing to mp */
7843 mc->mc_flags &= ~C_INITIALIZED;
7845 MDB_cursor *m2, *m3;
7846 MDB_dbi dbi = mc->mc_dbi;
7848 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7849 if (mc->mc_flags & C_SUB)
7850 m3 = &m2->mc_xcursor->mx_cursor;
7853 if (m3->mc_snum < mc->mc_snum) continue;
7854 if (m3->mc_pg[0] == mp) {
7857 m3->mc_flags &= ~C_INITIALIZED;
7861 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7863 DPUTS("collapsing root page!");
7864 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7867 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7868 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7871 mc->mc_db->md_depth--;
7872 mc->mc_db->md_branch_pages--;
7873 mc->mc_ki[0] = mc->mc_ki[1];
7874 for (i = 1; i<mc->mc_db->md_depth; i++) {
7875 mc->mc_pg[i] = mc->mc_pg[i+1];
7876 mc->mc_ki[i] = mc->mc_ki[i+1];
7879 /* Adjust other cursors pointing to mp */
7880 MDB_cursor *m2, *m3;
7881 MDB_dbi dbi = mc->mc_dbi;
7883 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7884 if (mc->mc_flags & C_SUB)
7885 m3 = &m2->mc_xcursor->mx_cursor;
7888 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7889 if (m3->mc_pg[0] == mp) {
7890 for (i=0; i<m3->mc_snum; i++) {
7891 m3->mc_pg[i] = m3->mc_pg[i+1];
7892 m3->mc_ki[i] = m3->mc_ki[i+1];
7900 DPUTS("root page doesn't need rebalancing");
7904 /* The parent (branch page) must have at least 2 pointers,
7905 * otherwise the tree is invalid.
7907 ptop = mc->mc_top-1;
7908 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7910 /* Leaf page fill factor is below the threshold.
7911 * Try to move keys from left or right neighbor, or
7912 * merge with a neighbor page.
7917 mdb_cursor_copy(mc, &mn);
7918 mn.mc_xcursor = NULL;
7920 oldki = mc->mc_ki[mc->mc_top];
7921 if (mc->mc_ki[ptop] == 0) {
7922 /* We're the leftmost leaf in our parent.
7924 DPUTS("reading right neighbor");
7926 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7927 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7930 mn.mc_ki[mn.mc_top] = 0;
7931 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7933 /* There is at least one neighbor to the left.
7935 DPUTS("reading left neighbor");
7937 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7938 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7941 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7942 mc->mc_ki[mc->mc_top] = 0;
7945 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7946 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7947 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7949 /* If the neighbor page is above threshold and has enough keys,
7950 * move one key from it. Otherwise we should try to merge them.
7951 * (A branch page must never have less than 2 keys.)
7953 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= thresh && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7954 rc = mdb_node_move(&mn, mc);
7955 if (mc->mc_ki[mc->mc_top-1]) {
7959 if (mc->mc_ki[ptop] == 0) {
7960 rc = mdb_page_merge(&mn, mc);
7963 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7964 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7965 /* We want mdb_rebalance to find mn when doing fixups */
7966 if (mc->mc_flags & C_SUB) {
7967 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7968 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
7969 dummy.mc_xcursor = (MDB_xcursor *)&mn;
7971 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7972 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
7974 rc = mdb_page_merge(mc, &mn);
7975 if (mc->mc_flags & C_SUB)
7976 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
7978 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
7979 mdb_cursor_copy(&mn, mc);
7981 mc->mc_flags &= ~C_EOF;
7983 mc->mc_ki[mc->mc_top] = oldki;
7987 /** Complete a delete operation started by #mdb_cursor_del(). */
7989 mdb_cursor_del0(MDB_cursor *mc)
7995 MDB_cursor *m2, *m3;
7996 MDB_dbi dbi = mc->mc_dbi;
7998 ki = mc->mc_ki[mc->mc_top];
7999 mp = mc->mc_pg[mc->mc_top];
8000 mdb_node_del(mc, mc->mc_db->md_pad);
8001 mc->mc_db->md_entries--;
8003 /* Adjust other cursors pointing to mp */
8004 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8005 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8006 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8008 if (m3 == mc || m3->mc_snum < mc->mc_snum)
8010 if (m3->mc_pg[mc->mc_top] == mp) {
8011 if (m3->mc_ki[mc->mc_top] >= ki) {
8012 m3->mc_flags |= C_DEL;
8013 if (m3->mc_ki[mc->mc_top] > ki)
8014 m3->mc_ki[mc->mc_top]--;
8015 else if (mc->mc_db->md_flags & MDB_DUPSORT)
8016 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
8021 rc = mdb_rebalance(mc);
8023 if (rc == MDB_SUCCESS) {
8024 /* DB is totally empty now, just bail out.
8025 * Other cursors adjustments were already done
8026 * by mdb_rebalance and aren't needed here.
8031 mp = mc->mc_pg[mc->mc_top];
8032 nkeys = NUMKEYS(mp);
8034 /* Adjust other cursors pointing to mp */
8035 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
8036 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
8037 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8039 if (m3->mc_snum < mc->mc_snum)
8041 if (m3->mc_pg[mc->mc_top] == mp) {
8042 /* if m3 points past last node in page, find next sibling */
8043 if (m3->mc_ki[mc->mc_top] >= nkeys) {
8044 rc = mdb_cursor_sibling(m3, 1);
8045 if (rc == MDB_NOTFOUND) {
8046 m3->mc_flags |= C_EOF;
8052 mc->mc_flags |= C_DEL;
8056 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8061 mdb_del(MDB_txn *txn, MDB_dbi dbi,
8062 MDB_val *key, MDB_val *data)
8064 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8067 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
8068 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
8070 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
8071 /* must ignore any data */
8075 return mdb_del0(txn, dbi, key, data, 0);
8079 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
8080 MDB_val *key, MDB_val *data, unsigned flags)
8085 MDB_val rdata, *xdata;
8089 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
8091 mdb_cursor_init(&mc, txn, dbi, &mx);
8100 flags |= MDB_NODUPDATA;
8102 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
8104 /* let mdb_page_split know about this cursor if needed:
8105 * delete will trigger a rebalance; if it needs to move
8106 * a node from one page to another, it will have to
8107 * update the parent's separator key(s). If the new sepkey
8108 * is larger than the current one, the parent page may
8109 * run out of space, triggering a split. We need this
8110 * cursor to be consistent until the end of the rebalance.
8112 mc.mc_flags |= C_UNTRACK;
8113 mc.mc_next = txn->mt_cursors[dbi];
8114 txn->mt_cursors[dbi] = &mc;
8115 rc = mdb_cursor_del(&mc, flags);
8116 txn->mt_cursors[dbi] = mc.mc_next;
8121 /** Split a page and insert a new node.
8122 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
8123 * The cursor will be updated to point to the actual page and index where
8124 * the node got inserted after the split.
8125 * @param[in] newkey The key for the newly inserted node.
8126 * @param[in] newdata The data for the newly inserted node.
8127 * @param[in] newpgno The page number, if the new node is a branch node.
8128 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8129 * @return 0 on success, non-zero on failure.
8132 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8133 unsigned int nflags)
8136 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8139 int i, j, split_indx, nkeys, pmax;
8140 MDB_env *env = mc->mc_txn->mt_env;
8142 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8143 MDB_page *copy = NULL;
8144 MDB_page *mp, *rp, *pp;
8149 mp = mc->mc_pg[mc->mc_top];
8150 newindx = mc->mc_ki[mc->mc_top];
8151 nkeys = NUMKEYS(mp);
8153 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8154 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8155 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8157 /* Create a right sibling. */
8158 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8160 rp->mp_pad = mp->mp_pad;
8161 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8163 if (mc->mc_snum < 2) {
8164 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8166 /* shift current top to make room for new parent */
8167 mc->mc_pg[1] = mc->mc_pg[0];
8168 mc->mc_ki[1] = mc->mc_ki[0];
8171 mc->mc_db->md_root = pp->mp_pgno;
8172 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8173 new_root = mc->mc_db->md_depth++;
8175 /* Add left (implicit) pointer. */
8176 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8177 /* undo the pre-push */
8178 mc->mc_pg[0] = mc->mc_pg[1];
8179 mc->mc_ki[0] = mc->mc_ki[1];
8180 mc->mc_db->md_root = mp->mp_pgno;
8181 mc->mc_db->md_depth--;
8188 ptop = mc->mc_top-1;
8189 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8192 mc->mc_flags |= C_SPLITTING;
8193 mdb_cursor_copy(mc, &mn);
8194 mn.mc_pg[mn.mc_top] = rp;
8195 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8197 if (nflags & MDB_APPEND) {
8198 mn.mc_ki[mn.mc_top] = 0;
8200 split_indx = newindx;
8204 split_indx = (nkeys+1) / 2;
8209 unsigned int lsize, rsize, ksize;
8210 /* Move half of the keys to the right sibling */
8211 x = mc->mc_ki[mc->mc_top] - split_indx;
8212 ksize = mc->mc_db->md_pad;
8213 split = LEAF2KEY(mp, split_indx, ksize);
8214 rsize = (nkeys - split_indx) * ksize;
8215 lsize = (nkeys - split_indx) * sizeof(indx_t);
8216 mp->mp_lower -= lsize;
8217 rp->mp_lower += lsize;
8218 mp->mp_upper += rsize - lsize;
8219 rp->mp_upper -= rsize - lsize;
8220 sepkey.mv_size = ksize;
8221 if (newindx == split_indx) {
8222 sepkey.mv_data = newkey->mv_data;
8224 sepkey.mv_data = split;
8227 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8228 memcpy(rp->mp_ptrs, split, rsize);
8229 sepkey.mv_data = rp->mp_ptrs;
8230 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8231 memcpy(ins, newkey->mv_data, ksize);
8232 mp->mp_lower += sizeof(indx_t);
8233 mp->mp_upper -= ksize - sizeof(indx_t);
8236 memcpy(rp->mp_ptrs, split, x * ksize);
8237 ins = LEAF2KEY(rp, x, ksize);
8238 memcpy(ins, newkey->mv_data, ksize);
8239 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8240 rp->mp_lower += sizeof(indx_t);
8241 rp->mp_upper -= ksize - sizeof(indx_t);
8242 mc->mc_ki[mc->mc_top] = x;
8243 mc->mc_pg[mc->mc_top] = rp;
8246 int psize, nsize, k;
8247 /* Maximum free space in an empty page */
8248 pmax = env->me_psize - PAGEHDRSZ;
8250 nsize = mdb_leaf_size(env, newkey, newdata);
8252 nsize = mdb_branch_size(env, newkey);
8253 nsize = EVEN(nsize);
8255 /* grab a page to hold a temporary copy */
8256 copy = mdb_page_malloc(mc->mc_txn, 1);
8261 copy->mp_pgno = mp->mp_pgno;
8262 copy->mp_flags = mp->mp_flags;
8263 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8264 copy->mp_upper = env->me_psize - PAGEBASE;
8266 /* prepare to insert */
8267 for (i=0, j=0; i<nkeys; i++) {
8269 copy->mp_ptrs[j++] = 0;
8271 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8274 /* When items are relatively large the split point needs
8275 * to be checked, because being off-by-one will make the
8276 * difference between success or failure in mdb_node_add.
8278 * It's also relevant if a page happens to be laid out
8279 * such that one half of its nodes are all "small" and
8280 * the other half of its nodes are "large." If the new
8281 * item is also "large" and falls on the half with
8282 * "large" nodes, it also may not fit.
8284 * As a final tweak, if the new item goes on the last
8285 * spot on the page (and thus, onto the new page), bias
8286 * the split so the new page is emptier than the old page.
8287 * This yields better packing during sequential inserts.
8289 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8290 /* Find split point */
8292 if (newindx <= split_indx || newindx >= nkeys) {
8294 k = newindx >= nkeys ? nkeys : split_indx+1+IS_LEAF(mp);
8299 for (; i!=k; i+=j) {
8304 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8305 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8307 if (F_ISSET(node->mn_flags, F_BIGDATA))
8308 psize += sizeof(pgno_t);
8310 psize += NODEDSZ(node);
8312 psize = EVEN(psize);
8314 if (psize > pmax || i == k-j) {
8315 split_indx = i + (j<0);
8320 if (split_indx == newindx) {
8321 sepkey.mv_size = newkey->mv_size;
8322 sepkey.mv_data = newkey->mv_data;
8324 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8325 sepkey.mv_size = node->mn_ksize;
8326 sepkey.mv_data = NODEKEY(node);
8331 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8333 /* Copy separator key to the parent.
8335 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8339 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8344 if (mn.mc_snum == mc->mc_snum) {
8345 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8346 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8347 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8348 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8353 /* Right page might now have changed parent.
8354 * Check if left page also changed parent.
8356 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8357 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8358 for (i=0; i<ptop; i++) {
8359 mc->mc_pg[i] = mn.mc_pg[i];
8360 mc->mc_ki[i] = mn.mc_ki[i];
8362 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8363 if (mn.mc_ki[ptop]) {
8364 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8366 /* find right page's left sibling */
8367 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8368 mdb_cursor_sibling(mc, 0);
8373 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8376 mc->mc_flags ^= C_SPLITTING;
8377 if (rc != MDB_SUCCESS) {
8380 if (nflags & MDB_APPEND) {
8381 mc->mc_pg[mc->mc_top] = rp;
8382 mc->mc_ki[mc->mc_top] = 0;
8383 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8386 for (i=0; i<mc->mc_top; i++)
8387 mc->mc_ki[i] = mn.mc_ki[i];
8388 } else if (!IS_LEAF2(mp)) {
8390 mc->mc_pg[mc->mc_top] = rp;
8395 rkey.mv_data = newkey->mv_data;
8396 rkey.mv_size = newkey->mv_size;
8402 /* Update index for the new key. */
8403 mc->mc_ki[mc->mc_top] = j;
8405 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8406 rkey.mv_data = NODEKEY(node);
8407 rkey.mv_size = node->mn_ksize;
8409 xdata.mv_data = NODEDATA(node);
8410 xdata.mv_size = NODEDSZ(node);
8413 pgno = NODEPGNO(node);
8414 flags = node->mn_flags;
8417 if (!IS_LEAF(mp) && j == 0) {
8418 /* First branch index doesn't need key data. */
8422 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8428 mc->mc_pg[mc->mc_top] = copy;
8433 } while (i != split_indx);
8435 nkeys = NUMKEYS(copy);
8436 for (i=0; i<nkeys; i++)
8437 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8438 mp->mp_lower = copy->mp_lower;
8439 mp->mp_upper = copy->mp_upper;
8440 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8441 env->me_psize - copy->mp_upper - PAGEBASE);
8443 /* reset back to original page */
8444 if (newindx < split_indx) {
8445 mc->mc_pg[mc->mc_top] = mp;
8446 if (nflags & MDB_RESERVE) {
8447 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8448 if (!(node->mn_flags & F_BIGDATA))
8449 newdata->mv_data = NODEDATA(node);
8452 mc->mc_pg[mc->mc_top] = rp;
8454 /* Make sure mc_ki is still valid.
8456 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8457 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8458 for (i=0; i<=ptop; i++) {
8459 mc->mc_pg[i] = mn.mc_pg[i];
8460 mc->mc_ki[i] = mn.mc_ki[i];
8467 /* Adjust other cursors pointing to mp */
8468 MDB_cursor *m2, *m3;
8469 MDB_dbi dbi = mc->mc_dbi;
8470 int fixup = NUMKEYS(mp);
8472 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8473 if (mc->mc_flags & C_SUB)
8474 m3 = &m2->mc_xcursor->mx_cursor;
8479 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8481 if (m3->mc_flags & C_SPLITTING)
8486 for (k=new_root; k>=0; k--) {
8487 m3->mc_ki[k+1] = m3->mc_ki[k];
8488 m3->mc_pg[k+1] = m3->mc_pg[k];
8490 if (m3->mc_ki[0] >= split_indx) {
8495 m3->mc_pg[0] = mc->mc_pg[0];
8499 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8500 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8501 m3->mc_ki[mc->mc_top]++;
8502 if (m3->mc_ki[mc->mc_top] >= fixup) {
8503 m3->mc_pg[mc->mc_top] = rp;
8504 m3->mc_ki[mc->mc_top] -= fixup;
8505 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8507 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8508 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8513 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8516 if (copy) /* tmp page */
8517 mdb_page_free(env, copy);
8519 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8524 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8525 MDB_val *key, MDB_val *data, unsigned int flags)
8530 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8533 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8536 mdb_cursor_init(&mc, txn, dbi, &mx);
8537 return mdb_cursor_put(&mc, key, data, flags);
8541 #define MDB_WBUF (1024*1024)
8544 /** State needed for a compacting copy. */
8545 typedef struct mdb_copy {
8546 pthread_mutex_t mc_mutex;
8547 pthread_cond_t mc_cond;
8554 pgno_t mc_next_pgno;
8557 volatile int mc_new;
8562 /** Dedicated writer thread for compacting copy. */
8563 static THREAD_RET ESECT CALL_CONV
8564 mdb_env_copythr(void *arg)
8568 int toggle = 0, wsize, rc;
8571 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8574 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8577 pthread_mutex_lock(&my->mc_mutex);
8579 pthread_cond_signal(&my->mc_cond);
8582 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8583 if (my->mc_new < 0) {
8588 wsize = my->mc_wlen[toggle];
8589 ptr = my->mc_wbuf[toggle];
8592 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8596 } else if (len > 0) {
8610 /* If there's an overflow page tail, write it too */
8611 if (my->mc_olen[toggle]) {
8612 wsize = my->mc_olen[toggle];
8613 ptr = my->mc_over[toggle];
8614 my->mc_olen[toggle] = 0;
8617 my->mc_wlen[toggle] = 0;
8619 pthread_cond_signal(&my->mc_cond);
8621 pthread_cond_signal(&my->mc_cond);
8622 pthread_mutex_unlock(&my->mc_mutex);
8623 return (THREAD_RET)0;
8627 /** Tell the writer thread there's a buffer ready to write */
8629 mdb_env_cthr_toggle(mdb_copy *my, int st)
8631 int toggle = my->mc_toggle ^ 1;
8632 pthread_mutex_lock(&my->mc_mutex);
8633 if (my->mc_status) {
8634 pthread_mutex_unlock(&my->mc_mutex);
8635 return my->mc_status;
8637 while (my->mc_new == 1)
8638 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8640 my->mc_toggle = toggle;
8641 pthread_cond_signal(&my->mc_cond);
8642 pthread_mutex_unlock(&my->mc_mutex);
8646 /** Depth-first tree traversal for compacting copy. */
8648 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8651 MDB_txn *txn = my->mc_txn;
8653 MDB_page *mo, *mp, *leaf;
8658 /* Empty DB, nothing to do */
8659 if (*pg == P_INVALID)
8666 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8669 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8673 /* Make cursor pages writable */
8674 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8678 for (i=0; i<mc.mc_top; i++) {
8679 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8680 mc.mc_pg[i] = (MDB_page *)ptr;
8681 ptr += my->mc_env->me_psize;
8684 /* This is writable space for a leaf page. Usually not needed. */
8685 leaf = (MDB_page *)ptr;
8687 toggle = my->mc_toggle;
8688 while (mc.mc_snum > 0) {
8690 mp = mc.mc_pg[mc.mc_top];
8694 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8695 for (i=0; i<n; i++) {
8696 ni = NODEPTR(mp, i);
8697 if (ni->mn_flags & F_BIGDATA) {
8701 /* Need writable leaf */
8703 mc.mc_pg[mc.mc_top] = leaf;
8704 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8706 ni = NODEPTR(mp, i);
8709 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8710 rc = mdb_page_get(txn, pg, &omp, NULL);
8713 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8714 rc = mdb_env_cthr_toggle(my, 1);
8717 toggle = my->mc_toggle;
8719 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8720 memcpy(mo, omp, my->mc_env->me_psize);
8721 mo->mp_pgno = my->mc_next_pgno;
8722 my->mc_next_pgno += omp->mp_pages;
8723 my->mc_wlen[toggle] += my->mc_env->me_psize;
8724 if (omp->mp_pages > 1) {
8725 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8726 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8727 rc = mdb_env_cthr_toggle(my, 1);
8730 toggle = my->mc_toggle;
8732 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8733 } else if (ni->mn_flags & F_SUBDATA) {
8736 /* Need writable leaf */
8738 mc.mc_pg[mc.mc_top] = leaf;
8739 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8741 ni = NODEPTR(mp, i);
8744 memcpy(&db, NODEDATA(ni), sizeof(db));
8745 my->mc_toggle = toggle;
8746 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8749 toggle = my->mc_toggle;
8750 memcpy(NODEDATA(ni), &db, sizeof(db));
8755 mc.mc_ki[mc.mc_top]++;
8756 if (mc.mc_ki[mc.mc_top] < n) {
8759 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8761 rc = mdb_page_get(txn, pg, &mp, NULL);
8766 mc.mc_ki[mc.mc_top] = 0;
8767 if (IS_BRANCH(mp)) {
8768 /* Whenever we advance to a sibling branch page,
8769 * we must proceed all the way down to its first leaf.
8771 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8774 mc.mc_pg[mc.mc_top] = mp;
8778 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8779 rc = mdb_env_cthr_toggle(my, 1);
8782 toggle = my->mc_toggle;
8784 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8785 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8786 mo->mp_pgno = my->mc_next_pgno++;
8787 my->mc_wlen[toggle] += my->mc_env->me_psize;
8789 /* Update parent if there is one */
8790 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8791 SETPGNO(ni, mo->mp_pgno);
8792 mdb_cursor_pop(&mc);
8794 /* Otherwise we're done */
8804 /** Copy environment with compaction. */
8806 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8811 MDB_txn *txn = NULL;
8816 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8817 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8818 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8819 if (my.mc_wbuf[0] == NULL)
8822 pthread_mutex_init(&my.mc_mutex, NULL);
8823 pthread_cond_init(&my.mc_cond, NULL);
8824 #ifdef HAVE_MEMALIGN
8825 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8826 if (my.mc_wbuf[0] == NULL)
8829 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8834 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8835 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8840 my.mc_next_pgno = 2;
8846 THREAD_CREATE(thr, mdb_env_copythr, &my);
8848 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8852 mp = (MDB_page *)my.mc_wbuf[0];
8853 memset(mp, 0, 2*env->me_psize);
8855 mp->mp_flags = P_META;
8856 mm = (MDB_meta *)METADATA(mp);
8857 mdb_env_init_meta0(env, mm);
8858 mm->mm_address = env->me_metas[0]->mm_address;
8860 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8862 mp->mp_flags = P_META;
8863 *(MDB_meta *)METADATA(mp) = *mm;
8864 mm = (MDB_meta *)METADATA(mp);
8866 /* Count the number of free pages, subtract from lastpg to find
8867 * number of active pages
8870 MDB_ID freecount = 0;
8873 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8874 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8875 freecount += *(MDB_ID *)data.mv_data;
8876 freecount += txn->mt_dbs[0].md_branch_pages +
8877 txn->mt_dbs[0].md_leaf_pages +
8878 txn->mt_dbs[0].md_overflow_pages;
8880 /* Set metapage 1 */
8881 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8882 mm->mm_dbs[1] = txn->mt_dbs[1];
8883 if (mm->mm_last_pg > 1) {
8884 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8887 mm->mm_dbs[1].md_root = P_INVALID;
8890 my.mc_wlen[0] = env->me_psize * 2;
8892 pthread_mutex_lock(&my.mc_mutex);
8894 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8895 pthread_mutex_unlock(&my.mc_mutex);
8896 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8897 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8898 rc = mdb_env_cthr_toggle(&my, 1);
8899 mdb_env_cthr_toggle(&my, -1);
8900 pthread_mutex_lock(&my.mc_mutex);
8902 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8903 pthread_mutex_unlock(&my.mc_mutex);
8908 CloseHandle(my.mc_cond);
8909 CloseHandle(my.mc_mutex);
8910 _aligned_free(my.mc_wbuf[0]);
8912 pthread_cond_destroy(&my.mc_cond);
8913 pthread_mutex_destroy(&my.mc_mutex);
8914 free(my.mc_wbuf[0]);
8919 /** Copy environment as-is. */
8921 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8923 MDB_txn *txn = NULL;
8929 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8933 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8936 /* Do the lock/unlock of the reader mutex before starting the
8937 * write txn. Otherwise other read txns could block writers.
8939 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8944 /* We must start the actual read txn after blocking writers */
8945 mdb_txn_reset0(txn, "reset-stage1");
8947 /* Temporarily block writers until we snapshot the meta pages */
8950 rc = mdb_txn_renew0(txn);
8952 UNLOCK_MUTEX_W(env);
8957 wsize = env->me_psize * 2;
8961 DO_WRITE(rc, fd, ptr, w2, len);
8965 } else if (len > 0) {
8971 /* Non-blocking or async handles are not supported */
8977 UNLOCK_MUTEX_W(env);
8982 w2 = txn->mt_next_pgno * env->me_psize;
8985 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8992 if (wsize > MAX_WRITE)
8996 DO_WRITE(rc, fd, ptr, w2, len);
9000 } else if (len > 0) {
9017 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
9019 if (flags & MDB_CP_COMPACT)
9020 return mdb_env_copyfd1(env, fd);
9022 return mdb_env_copyfd0(env, fd);
9026 mdb_env_copyfd(MDB_env *env, HANDLE fd)
9028 return mdb_env_copyfd2(env, fd, 0);
9032 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
9036 HANDLE newfd = INVALID_HANDLE_VALUE;
9038 if (env->me_flags & MDB_NOSUBDIR) {
9039 lpath = (char *)path;
9042 len += sizeof(DATANAME);
9043 lpath = malloc(len);
9046 sprintf(lpath, "%s" DATANAME, path);
9049 /* The destination path must exist, but the destination file must not.
9050 * We don't want the OS to cache the writes, since the source data is
9051 * already in the OS cache.
9054 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
9055 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
9057 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
9059 if (newfd == INVALID_HANDLE_VALUE) {
9064 if (env->me_psize >= env->me_os_psize) {
9066 /* Set O_DIRECT if the file system supports it */
9067 if ((rc = fcntl(newfd, F_GETFL)) != -1)
9068 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
9070 #ifdef F_NOCACHE /* __APPLE__ */
9071 rc = fcntl(newfd, F_NOCACHE, 1);
9079 rc = mdb_env_copyfd2(env, newfd, flags);
9082 if (!(env->me_flags & MDB_NOSUBDIR))
9084 if (newfd != INVALID_HANDLE_VALUE)
9085 if (close(newfd) < 0 && rc == MDB_SUCCESS)
9092 mdb_env_copy(MDB_env *env, const char *path)
9094 return mdb_env_copy2(env, path, 0);
9098 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
9100 if ((flag & CHANGEABLE) != flag)
9103 env->me_flags |= flag;
9105 env->me_flags &= ~flag;
9110 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
9115 *arg = env->me_flags;
9120 mdb_env_set_userctx(MDB_env *env, void *ctx)
9124 env->me_userctx = ctx;
9129 mdb_env_get_userctx(MDB_env *env)
9131 return env ? env->me_userctx : NULL;
9135 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9140 env->me_assert_func = func;
9146 mdb_env_get_path(MDB_env *env, const char **arg)
9151 *arg = env->me_path;
9156 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9165 /** Common code for #mdb_stat() and #mdb_env_stat().
9166 * @param[in] env the environment to operate in.
9167 * @param[in] db the #MDB_db record containing the stats to return.
9168 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9169 * @return 0, this function always succeeds.
9172 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9174 arg->ms_psize = env->me_psize;
9175 arg->ms_depth = db->md_depth;
9176 arg->ms_branch_pages = db->md_branch_pages;
9177 arg->ms_leaf_pages = db->md_leaf_pages;
9178 arg->ms_overflow_pages = db->md_overflow_pages;
9179 arg->ms_entries = db->md_entries;
9185 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9189 if (env == NULL || arg == NULL)
9192 toggle = mdb_env_pick_meta(env);
9194 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9198 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9202 if (env == NULL || arg == NULL)
9205 toggle = mdb_env_pick_meta(env);
9206 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9207 arg->me_mapsize = env->me_mapsize;
9208 arg->me_maxreaders = env->me_maxreaders;
9209 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : 0;
9211 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9212 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9216 /** Set the default comparison functions for a database.
9217 * Called immediately after a database is opened to set the defaults.
9218 * The user can then override them with #mdb_set_compare() or
9219 * #mdb_set_dupsort().
9220 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9221 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9224 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9226 uint16_t f = txn->mt_dbs[dbi].md_flags;
9228 txn->mt_dbxs[dbi].md_cmp =
9229 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9230 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9232 txn->mt_dbxs[dbi].md_dcmp =
9233 !(f & MDB_DUPSORT) ? 0 :
9234 ((f & MDB_INTEGERDUP)
9235 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9236 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9239 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9245 int rc, dbflag, exact;
9246 unsigned int unused = 0, seq;
9249 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9250 mdb_default_cmp(txn, FREE_DBI);
9253 if ((flags & VALID_FLAGS) != flags)
9255 if (txn->mt_flags & MDB_TXN_ERROR)
9261 if (flags & PERSISTENT_FLAGS) {
9262 uint16_t f2 = flags & PERSISTENT_FLAGS;
9263 /* make sure flag changes get committed */
9264 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9265 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9266 txn->mt_flags |= MDB_TXN_DIRTY;
9269 mdb_default_cmp(txn, MAIN_DBI);
9273 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9274 mdb_default_cmp(txn, MAIN_DBI);
9277 /* Is the DB already open? */
9279 for (i=2; i<txn->mt_numdbs; i++) {
9280 if (!txn->mt_dbxs[i].md_name.mv_size) {
9281 /* Remember this free slot */
9282 if (!unused) unused = i;
9285 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9286 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9292 /* If no free slot and max hit, fail */
9293 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9294 return MDB_DBS_FULL;
9296 /* Cannot mix named databases with some mainDB flags */
9297 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9298 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9300 /* Find the DB info */
9301 dbflag = DB_NEW|DB_VALID;
9304 key.mv_data = (void *)name;
9305 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9306 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9307 if (rc == MDB_SUCCESS) {
9308 /* make sure this is actually a DB */
9309 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9310 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9311 return MDB_INCOMPATIBLE;
9312 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9313 /* Create if requested */
9314 data.mv_size = sizeof(MDB_db);
9315 data.mv_data = &dummy;
9316 memset(&dummy, 0, sizeof(dummy));
9317 dummy.md_root = P_INVALID;
9318 dummy.md_flags = flags & PERSISTENT_FLAGS;
9319 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9323 /* OK, got info, add to table */
9324 if (rc == MDB_SUCCESS) {
9325 unsigned int slot = unused ? unused : txn->mt_numdbs;
9326 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9327 txn->mt_dbxs[slot].md_name.mv_size = len;
9328 txn->mt_dbxs[slot].md_rel = NULL;
9329 txn->mt_dbflags[slot] = dbflag;
9330 /* txn-> and env-> are the same in read txns, use
9331 * tmp variable to avoid undefined assignment
9333 seq = ++txn->mt_env->me_dbiseqs[slot];
9334 txn->mt_dbiseqs[slot] = seq;
9336 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9338 mdb_default_cmp(txn, slot);
9347 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9349 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9352 if (txn->mt_flags & MDB_TXN_ERROR)
9355 if (txn->mt_dbflags[dbi] & DB_STALE) {
9358 /* Stale, must read the DB's root. cursor_init does it for us. */
9359 mdb_cursor_init(&mc, txn, dbi, &mx);
9361 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9364 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9367 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9369 ptr = env->me_dbxs[dbi].md_name.mv_data;
9370 /* If there was no name, this was already closed */
9372 env->me_dbxs[dbi].md_name.mv_data = NULL;
9373 env->me_dbxs[dbi].md_name.mv_size = 0;
9374 env->me_dbflags[dbi] = 0;
9375 env->me_dbiseqs[dbi]++;
9380 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9382 /* We could return the flags for the FREE_DBI too but what's the point? */
9383 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9385 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9389 /** Add all the DB's pages to the free list.
9390 * @param[in] mc Cursor on the DB to free.
9391 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9392 * @return 0 on success, non-zero on failure.
9395 mdb_drop0(MDB_cursor *mc, int subs)
9399 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9400 if (rc == MDB_SUCCESS) {
9401 MDB_txn *txn = mc->mc_txn;
9406 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9407 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9410 mdb_cursor_copy(mc, &mx);
9411 while (mc->mc_snum > 0) {
9412 MDB_page *mp = mc->mc_pg[mc->mc_top];
9413 unsigned n = NUMKEYS(mp);
9415 for (i=0; i<n; i++) {
9416 ni = NODEPTR(mp, i);
9417 if (ni->mn_flags & F_BIGDATA) {
9420 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9421 rc = mdb_page_get(txn, pg, &omp, NULL);
9424 mdb_cassert(mc, IS_OVERFLOW(omp));
9425 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9429 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9430 mdb_xcursor_init1(mc, ni);
9431 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9437 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9439 for (i=0; i<n; i++) {
9441 ni = NODEPTR(mp, i);
9444 mdb_midl_xappend(txn->mt_free_pgs, pg);
9449 mc->mc_ki[mc->mc_top] = i;
9450 rc = mdb_cursor_sibling(mc, 1);
9452 if (rc != MDB_NOTFOUND)
9454 /* no more siblings, go back to beginning
9455 * of previous level.
9459 for (i=1; i<mc->mc_snum; i++) {
9461 mc->mc_pg[i] = mx.mc_pg[i];
9466 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9469 txn->mt_flags |= MDB_TXN_ERROR;
9470 } else if (rc == MDB_NOTFOUND) {
9476 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9478 MDB_cursor *mc, *m2;
9481 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9484 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9487 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9490 rc = mdb_cursor_open(txn, dbi, &mc);
9494 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9495 /* Invalidate the dropped DB's cursors */
9496 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9497 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9501 /* Can't delete the main DB */
9502 if (del && dbi > MAIN_DBI) {
9503 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9505 txn->mt_dbflags[dbi] = DB_STALE;
9506 mdb_dbi_close(txn->mt_env, dbi);
9508 txn->mt_flags |= MDB_TXN_ERROR;
9511 /* reset the DB record, mark it dirty */
9512 txn->mt_dbflags[dbi] |= DB_DIRTY;
9513 txn->mt_dbs[dbi].md_depth = 0;
9514 txn->mt_dbs[dbi].md_branch_pages = 0;
9515 txn->mt_dbs[dbi].md_leaf_pages = 0;
9516 txn->mt_dbs[dbi].md_overflow_pages = 0;
9517 txn->mt_dbs[dbi].md_entries = 0;
9518 txn->mt_dbs[dbi].md_root = P_INVALID;
9520 txn->mt_flags |= MDB_TXN_DIRTY;
9523 mdb_cursor_close(mc);
9527 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9529 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9532 txn->mt_dbxs[dbi].md_cmp = cmp;
9536 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9538 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9541 txn->mt_dbxs[dbi].md_dcmp = cmp;
9545 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9547 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9550 txn->mt_dbxs[dbi].md_rel = rel;
9554 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9556 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9559 txn->mt_dbxs[dbi].md_relctx = ctx;
9564 mdb_env_get_maxkeysize(MDB_env *env)
9566 return ENV_MAXKEY(env);
9570 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9572 unsigned int i, rdrs;
9575 int rc = 0, first = 1;
9579 if (!env->me_txns) {
9580 return func("(no reader locks)\n", ctx);
9582 rdrs = env->me_txns->mti_numreaders;
9583 mr = env->me_txns->mti_readers;
9584 for (i=0; i<rdrs; i++) {
9586 txnid_t txnid = mr[i].mr_txnid;
9587 sprintf(buf, txnid == (txnid_t)-1 ?
9588 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9589 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9592 rc = func(" pid thread txnid\n", ctx);
9596 rc = func(buf, ctx);
9602 rc = func("(no active readers)\n", ctx);
9607 /** Insert pid into list if not already present.
9608 * return -1 if already present.
9611 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9613 /* binary search of pid in list */
9615 unsigned cursor = 1;
9617 unsigned n = ids[0];
9620 unsigned pivot = n >> 1;
9621 cursor = base + pivot + 1;
9622 val = pid - ids[cursor];
9627 } else if ( val > 0 ) {
9632 /* found, so it's a duplicate */
9641 for (n = ids[0]; n > cursor; n--)
9648 mdb_reader_check(MDB_env *env, int *dead)
9650 unsigned int i, j, rdrs;
9652 MDB_PID_T *pids, pid;
9661 rdrs = env->me_txns->mti_numreaders;
9662 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9666 mr = env->me_txns->mti_readers;
9667 for (i=0; i<rdrs; i++) {
9668 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9670 if (mdb_pid_insert(pids, pid) == 0) {
9671 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9673 /* Recheck, a new process may have reused pid */
9674 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9675 for (j=i; j<rdrs; j++)
9676 if (mr[j].mr_pid == pid) {
9677 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9678 (unsigned) pid, mr[j].mr_txnid));
9683 UNLOCK_MUTEX_R(env);