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")))
188 /** @defgroup internal LMDB Internals
191 /** @defgroup compat Compatibility Macros
192 * A bunch of macros to minimize the amount of platform-specific ifdefs
193 * needed throughout the rest of the code. When the features this library
194 * needs are similar enough to POSIX to be hidden in a one-or-two line
195 * replacement, this macro approach is used.
199 /** Features under development */
204 /** Wrapper around __func__, which is a C99 feature */
205 #if __STDC_VERSION__ >= 199901L
206 # define mdb_func_ __func__
207 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
208 # define mdb_func_ __FUNCTION__
210 /* If a debug message says <mdb_unknown>(), update the #if statements above */
211 # define mdb_func_ "<mdb_unknown>"
215 #define MDB_USE_HASH 1
216 #define MDB_PIDLOCK 0
217 #define THREAD_RET DWORD
218 #define pthread_t HANDLE
219 #define pthread_mutex_t HANDLE
220 #define pthread_cond_t HANDLE
221 #define pthread_key_t DWORD
222 #define pthread_self() GetCurrentThreadId()
223 #define pthread_key_create(x,y) \
224 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
225 #define pthread_key_delete(x) TlsFree(x)
226 #define pthread_getspecific(x) TlsGetValue(x)
227 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
228 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
229 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
230 #define pthread_cond_signal(x) SetEvent(*x)
231 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
232 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
233 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
234 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
235 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
236 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
237 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
238 #define getpid() GetCurrentProcessId()
239 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
240 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
241 #define ErrCode() GetLastError()
242 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
243 #define close(fd) (CloseHandle(fd) ? 0 : -1)
244 #define munmap(ptr,len) UnmapViewOfFile(ptr)
245 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
246 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
248 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
252 #define THREAD_RET void *
253 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
254 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
255 #define Z "z" /**< printf format modifier for size_t */
257 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
258 #define MDB_PIDLOCK 1
260 #ifdef MDB_USE_POSIX_SEM
262 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
263 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
264 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
265 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
268 mdb_sem_wait(sem_t *sem)
271 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
276 /** Lock the reader mutex.
278 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
279 /** Unlock the reader mutex.
281 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
283 /** Lock the writer mutex.
284 * Only a single write transaction is allowed at a time. Other writers
285 * will block waiting for this mutex.
287 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
288 /** Unlock the writer mutex.
290 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
291 #endif /* MDB_USE_POSIX_SEM */
293 /** Get the error code for the last failed system function.
295 #define ErrCode() errno
297 /** An abstraction for a file handle.
298 * On POSIX systems file handles are small integers. On Windows
299 * they're opaque pointers.
303 /** A value for an invalid file handle.
304 * Mainly used to initialize file variables and signify that they are
307 #define INVALID_HANDLE_VALUE (-1)
309 /** Get the size of a memory page for the system.
310 * This is the basic size that the platform's memory manager uses, and is
311 * fundamental to the use of memory-mapped files.
313 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
316 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
319 #define MNAME_LEN (sizeof(pthread_mutex_t))
325 /** A flag for opening a file and requesting synchronous data writes.
326 * This is only used when writing a meta page. It's not strictly needed;
327 * we could just do a normal write and then immediately perform a flush.
328 * But if this flag is available it saves us an extra system call.
330 * @note If O_DSYNC is undefined but exists in /usr/include,
331 * preferably set some compiler flag to get the definition.
332 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
335 # define MDB_DSYNC O_DSYNC
339 /** Function for flushing the data of a file. Define this to fsync
340 * if fdatasync() is not supported.
342 #ifndef MDB_FDATASYNC
343 # define MDB_FDATASYNC fdatasync
347 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
358 /** A page number in the database.
359 * Note that 64 bit page numbers are overkill, since pages themselves
360 * already represent 12-13 bits of addressable memory, and the OS will
361 * always limit applications to a maximum of 63 bits of address space.
363 * @note In the #MDB_node structure, we only store 48 bits of this value,
364 * which thus limits us to only 60 bits of addressable data.
366 typedef MDB_ID pgno_t;
368 /** A transaction ID.
369 * See struct MDB_txn.mt_txnid for details.
371 typedef MDB_ID txnid_t;
373 /** @defgroup debug Debug Macros
377 /** Enable debug output. Needs variable argument macros (a C99 feature).
378 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
379 * read from and written to the database (used for free space management).
385 static int mdb_debug;
386 static txnid_t mdb_debug_start;
388 /** Print a debug message with printf formatting.
389 * Requires double parenthesis around 2 or more args.
391 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
392 # define DPRINTF0(fmt, ...) \
393 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
395 # define DPRINTF(args) ((void) 0)
397 /** Print a debug string.
398 * The string is printed literally, with no format processing.
400 #define DPUTS(arg) DPRINTF(("%s", arg))
401 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
403 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
406 /** @brief The maximum size of a database page.
408 * It is 32k or 64k, since value-PAGEBASE must fit in
409 * #MDB_page.%mp_upper.
411 * LMDB will use database pages < OS pages if needed.
412 * That causes more I/O in write transactions: The OS must
413 * know (read) the whole page before writing a partial page.
415 * Note that we don't currently support Huge pages. On Linux,
416 * regular data files cannot use Huge pages, and in general
417 * Huge pages aren't actually pageable. We rely on the OS
418 * demand-pager to read our data and page it out when memory
419 * pressure from other processes is high. So until OSs have
420 * actual paging support for Huge pages, they're not viable.
422 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
424 /** The minimum number of keys required in a database page.
425 * Setting this to a larger value will place a smaller bound on the
426 * maximum size of a data item. Data items larger than this size will
427 * be pushed into overflow pages instead of being stored directly in
428 * the B-tree node. This value used to default to 4. With a page size
429 * of 4096 bytes that meant that any item larger than 1024 bytes would
430 * go into an overflow page. That also meant that on average 2-3KB of
431 * each overflow page was wasted space. The value cannot be lower than
432 * 2 because then there would no longer be a tree structure. With this
433 * value, items larger than 2KB will go into overflow pages, and on
434 * average only 1KB will be wasted.
436 #define MDB_MINKEYS 2
438 /** A stamp that identifies a file as an LMDB file.
439 * There's nothing special about this value other than that it is easily
440 * recognizable, and it will reflect any byte order mismatches.
442 #define MDB_MAGIC 0xBEEFC0DE
444 /** The version number for a database's datafile format. */
445 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
446 /** The version number for a database's lockfile format. */
447 #define MDB_LOCK_VERSION 1
449 /** @brief The max size of a key we can write, or 0 for dynamic max.
451 * Define this as 0 to compute the max from the page size. 511
452 * is default for backwards compat: liblmdb <= 0.9.10 can break
453 * when modifying a DB with keys/dupsort data bigger than its max.
454 * #MDB_DEVEL sets the default to 0.
456 * Data items in an #MDB_DUPSORT database are also limited to
457 * this size, since they're actually keys of a sub-DB. Keys and
458 * #MDB_DUPSORT data items must fit on a node in a regular page.
460 #ifndef MDB_MAXKEYSIZE
461 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
464 /** The maximum size of a key we can write to the environment. */
466 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
468 #define ENV_MAXKEY(env) ((env)->me_maxkey)
471 /** @brief The maximum size of a data item.
473 * We only store a 32 bit value for node sizes.
475 #define MAXDATASIZE 0xffffffffUL
478 /** Key size which fits in a #DKBUF.
481 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
484 * This is used for printing a hex dump of a key's contents.
486 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
487 /** Display a key in hex.
489 * Invoke a function to display a key in hex.
491 #define DKEY(x) mdb_dkey(x, kbuf)
497 /** An invalid page number.
498 * Mainly used to denote an empty tree.
500 #define P_INVALID (~(pgno_t)0)
502 /** Test if the flags \b f are set in a flag word \b w. */
503 #define F_ISSET(w, f) (((w) & (f)) == (f))
505 /** Round \b n up to an even number. */
506 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
508 /** Used for offsets within a single page.
509 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
512 typedef uint16_t indx_t;
514 /** Default size of memory map.
515 * This is certainly too small for any actual applications. Apps should always set
516 * the size explicitly using #mdb_env_set_mapsize().
518 #define DEFAULT_MAPSIZE 1048576
520 /** @defgroup readers Reader Lock Table
521 * Readers don't acquire any locks for their data access. Instead, they
522 * simply record their transaction ID in the reader table. The reader
523 * mutex is needed just to find an empty slot in the reader table. The
524 * slot's address is saved in thread-specific data so that subsequent read
525 * transactions started by the same thread need no further locking to proceed.
527 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
529 * No reader table is used if the database is on a read-only filesystem, or
530 * if #MDB_NOLOCK is set.
532 * Since the database uses multi-version concurrency control, readers don't
533 * actually need any locking. This table is used to keep track of which
534 * readers are using data from which old transactions, so that we'll know
535 * when a particular old transaction is no longer in use. Old transactions
536 * that have discarded any data pages can then have those pages reclaimed
537 * for use by a later write transaction.
539 * The lock table is constructed such that reader slots are aligned with the
540 * processor's cache line size. Any slot is only ever used by one thread.
541 * This alignment guarantees that there will be no contention or cache
542 * thrashing as threads update their own slot info, and also eliminates
543 * any need for locking when accessing a slot.
545 * A writer thread will scan every slot in the table to determine the oldest
546 * outstanding reader transaction. Any freed pages older than this will be
547 * reclaimed by the writer. The writer doesn't use any locks when scanning
548 * this table. This means that there's no guarantee that the writer will
549 * see the most up-to-date reader info, but that's not required for correct
550 * operation - all we need is to know the upper bound on the oldest reader,
551 * we don't care at all about the newest reader. So the only consequence of
552 * reading stale information here is that old pages might hang around a
553 * while longer before being reclaimed. That's actually good anyway, because
554 * the longer we delay reclaiming old pages, the more likely it is that a
555 * string of contiguous pages can be found after coalescing old pages from
556 * many old transactions together.
559 /** Number of slots in the reader table.
560 * This value was chosen somewhat arbitrarily. 126 readers plus a
561 * couple mutexes fit exactly into 8KB on my development machine.
562 * Applications should set the table size using #mdb_env_set_maxreaders().
564 #define DEFAULT_READERS 126
566 /** The size of a CPU cache line in bytes. We want our lock structures
567 * aligned to this size to avoid false cache line sharing in the
569 * This value works for most CPUs. For Itanium this should be 128.
575 /** The information we store in a single slot of the reader table.
576 * In addition to a transaction ID, we also record the process and
577 * thread ID that owns a slot, so that we can detect stale information,
578 * e.g. threads or processes that went away without cleaning up.
579 * @note We currently don't check for stale records. We simply re-init
580 * the table when we know that we're the only process opening the
583 typedef struct MDB_rxbody {
584 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
585 * Multiple readers that start at the same time will probably have the
586 * same ID here. Again, it's not important to exclude them from
587 * anything; all we need to know is which version of the DB they
588 * started from so we can avoid overwriting any data used in that
589 * particular version.
591 volatile txnid_t mrb_txnid;
592 /** The process ID of the process owning this reader txn. */
593 volatile MDB_PID_T mrb_pid;
594 /** The thread ID of the thread owning this txn. */
595 volatile MDB_THR_T mrb_tid;
598 /** The actual reader record, with cacheline padding. */
599 typedef struct MDB_reader {
602 /** shorthand for mrb_txnid */
603 #define mr_txnid mru.mrx.mrb_txnid
604 #define mr_pid mru.mrx.mrb_pid
605 #define mr_tid mru.mrx.mrb_tid
606 /** cache line alignment */
607 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
611 /** The header for the reader table.
612 * The table resides in a memory-mapped file. (This is a different file
613 * than is used for the main database.)
615 * For POSIX the actual mutexes reside in the shared memory of this
616 * mapped file. On Windows, mutexes are named objects allocated by the
617 * kernel; we store the mutex names in this mapped file so that other
618 * processes can grab them. This same approach is also used on
619 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
620 * process-shared POSIX mutexes. For these cases where a named object
621 * is used, the object name is derived from a 64 bit FNV hash of the
622 * environment pathname. As such, naming collisions are extremely
623 * unlikely. If a collision occurs, the results are unpredictable.
625 typedef struct MDB_txbody {
626 /** Stamp identifying this as an LMDB file. It must be set
629 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
631 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
632 char mtb_rmname[MNAME_LEN];
634 /** Mutex protecting access to this table.
635 * This is the reader lock that #LOCK_MUTEX_R acquires.
637 pthread_mutex_t mtb_mutex;
639 /** The ID of the last transaction committed to the database.
640 * This is recorded here only for convenience; the value can always
641 * be determined by reading the main database meta pages.
643 volatile txnid_t mtb_txnid;
644 /** The number of slots that have been used in the reader table.
645 * This always records the maximum count, it is not decremented
646 * when readers release their slots.
648 volatile unsigned mtb_numreaders;
651 /** The actual reader table definition. */
652 typedef struct MDB_txninfo {
655 #define mti_magic mt1.mtb.mtb_magic
656 #define mti_format mt1.mtb.mtb_format
657 #define mti_mutex mt1.mtb.mtb_mutex
658 #define mti_rmname mt1.mtb.mtb_rmname
659 #define mti_txnid mt1.mtb.mtb_txnid
660 #define mti_numreaders mt1.mtb.mtb_numreaders
661 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
664 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
665 char mt2_wmname[MNAME_LEN];
666 #define mti_wmname mt2.mt2_wmname
668 pthread_mutex_t mt2_wmutex;
669 #define mti_wmutex mt2.mt2_wmutex
671 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
673 MDB_reader mti_readers[1];
676 /** Lockfile format signature: version, features and field layout */
677 #define MDB_LOCK_FORMAT \
679 ((MDB_LOCK_VERSION) \
680 /* Flags which describe functionality */ \
681 + (((MDB_PIDLOCK) != 0) << 16)))
684 /** Common header for all page types.
685 * Overflow records occupy a number of contiguous pages with no
686 * headers on any page after the first.
688 typedef struct MDB_page {
689 #define mp_pgno mp_p.p_pgno
690 #define mp_next mp_p.p_next
692 pgno_t p_pgno; /**< page number */
693 struct MDB_page *p_next; /**< for in-memory list of freed pages */
696 /** @defgroup mdb_page Page Flags
698 * Flags for the page headers.
701 #define P_BRANCH 0x01 /**< branch page */
702 #define P_LEAF 0x02 /**< leaf page */
703 #define P_OVERFLOW 0x04 /**< overflow page */
704 #define P_META 0x08 /**< meta page */
705 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
706 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
707 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
708 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
709 #define P_KEEP 0x8000 /**< leave this page alone during spill */
711 uint16_t mp_flags; /**< @ref mdb_page */
712 #define mp_lower mp_pb.pb.pb_lower
713 #define mp_upper mp_pb.pb.pb_upper
714 #define mp_pages mp_pb.pb_pages
717 indx_t pb_lower; /**< lower bound of free space */
718 indx_t pb_upper; /**< upper bound of free space */
720 uint32_t pb_pages; /**< number of overflow pages */
722 indx_t mp_ptrs[1]; /**< dynamic size */
725 /** Size of the page header, excluding dynamic data at the end */
726 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
728 /** Address of first usable data byte in a page, after the header */
729 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
731 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
732 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
734 /** Number of nodes on a page */
735 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
737 /** The amount of space remaining in the page */
738 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
740 /** The percentage of space used in the page, in tenths of a percent. */
741 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
742 ((env)->me_psize - PAGEHDRSZ))
743 /** The minimum page fill factor, in tenths of a percent.
744 * Pages emptier than this are candidates for merging.
746 #define FILL_THRESHOLD 250
748 /** Test if a page is a leaf page */
749 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
750 /** Test if a page is a LEAF2 page */
751 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
752 /** Test if a page is a branch page */
753 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
754 /** Test if a page is an overflow page */
755 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
756 /** Test if a page is a sub page */
757 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
759 /** The number of overflow pages needed to store the given size. */
760 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
762 /** Link in #MDB_txn.%mt_loose_pgs list */
763 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
765 /** Header for a single key/data pair within a page.
766 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
767 * We guarantee 2-byte alignment for 'MDB_node's.
769 typedef struct MDB_node {
770 /** lo and hi are used for data size on leaf nodes and for
771 * child pgno on branch nodes. On 64 bit platforms, flags
772 * is also used for pgno. (Branch nodes have no flags).
773 * They are in host byte order in case that lets some
774 * accesses be optimized into a 32-bit word access.
776 #if BYTE_ORDER == LITTLE_ENDIAN
777 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
779 unsigned short mn_hi, mn_lo;
781 /** @defgroup mdb_node Node Flags
783 * Flags for node headers.
786 #define F_BIGDATA 0x01 /**< data put on overflow page */
787 #define F_SUBDATA 0x02 /**< data is a sub-database */
788 #define F_DUPDATA 0x04 /**< data has duplicates */
790 /** valid flags for #mdb_node_add() */
791 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
794 unsigned short mn_flags; /**< @ref mdb_node */
795 unsigned short mn_ksize; /**< key size */
796 char mn_data[1]; /**< key and data are appended here */
799 /** Size of the node header, excluding dynamic data at the end */
800 #define NODESIZE offsetof(MDB_node, mn_data)
802 /** Bit position of top word in page number, for shifting mn_flags */
803 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
805 /** Size of a node in a branch page with a given key.
806 * This is just the node header plus the key, there is no data.
808 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
810 /** Size of a node in a leaf page with a given key and data.
811 * This is node header plus key plus data size.
813 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
815 /** Address of node \b i in page \b p */
816 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
818 /** Address of the key for the node */
819 #define NODEKEY(node) (void *)((node)->mn_data)
821 /** Address of the data for a node */
822 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
824 /** Get the page number pointed to by a branch node */
825 #define NODEPGNO(node) \
826 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
827 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
828 /** Set the page number in a branch node */
829 #define SETPGNO(node,pgno) do { \
830 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
831 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
833 /** Get the size of the data in a leaf node */
834 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
835 /** Set the size of the data for a leaf node */
836 #define SETDSZ(node,size) do { \
837 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
838 /** The size of a key in a node */
839 #define NODEKSZ(node) ((node)->mn_ksize)
841 /** Copy a page number from src to dst */
843 #define COPY_PGNO(dst,src) dst = src
845 #if SIZE_MAX > 4294967295UL
846 #define COPY_PGNO(dst,src) do { \
847 unsigned short *s, *d; \
848 s = (unsigned short *)&(src); \
849 d = (unsigned short *)&(dst); \
856 #define COPY_PGNO(dst,src) do { \
857 unsigned short *s, *d; \
858 s = (unsigned short *)&(src); \
859 d = (unsigned short *)&(dst); \
865 /** The address of a key in a LEAF2 page.
866 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
867 * There are no node headers, keys are stored contiguously.
869 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
871 /** Set the \b node's key into \b keyptr, if requested. */
872 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
873 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
875 /** Set the \b node's key into \b key. */
876 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
878 /** Information about a single database in the environment. */
879 typedef struct MDB_db {
880 uint32_t md_pad; /**< also ksize for LEAF2 pages */
881 uint16_t md_flags; /**< @ref mdb_dbi_open */
882 uint16_t md_depth; /**< depth of this tree */
883 pgno_t md_branch_pages; /**< number of internal pages */
884 pgno_t md_leaf_pages; /**< number of leaf pages */
885 pgno_t md_overflow_pages; /**< number of overflow pages */
886 size_t md_entries; /**< number of data items */
887 pgno_t md_root; /**< the root page of this tree */
890 /** mdb_dbi_open flags */
891 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
892 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
893 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
894 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
896 /** Handle for the DB used to track free pages. */
898 /** Handle for the default DB. */
901 /** Meta page content.
902 * A meta page is the start point for accessing a database snapshot.
903 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
905 typedef struct MDB_meta {
906 /** Stamp identifying this as an LMDB file. It must be set
909 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
911 void *mm_address; /**< address for fixed mapping */
912 size_t mm_mapsize; /**< size of mmap region */
913 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
914 /** The size of pages used in this DB */
915 #define mm_psize mm_dbs[0].md_pad
916 /** Any persistent environment flags. @ref mdb_env */
917 #define mm_flags mm_dbs[0].md_flags
918 pgno_t mm_last_pg; /**< last used page in file */
919 volatile txnid_t mm_txnid; /**< txnid that committed this page */
922 /** Buffer for a stack-allocated meta page.
923 * The members define size and alignment, and silence type
924 * aliasing warnings. They are not used directly; that could
925 * mean incorrectly using several union members in parallel.
927 typedef union MDB_metabuf {
930 char mm_pad[PAGEHDRSZ];
935 /** Auxiliary DB info.
936 * The information here is mostly static/read-only. There is
937 * only a single copy of this record in the environment.
939 typedef struct MDB_dbx {
940 MDB_val md_name; /**< name of the database */
941 MDB_cmp_func *md_cmp; /**< function for comparing keys */
942 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
943 MDB_rel_func *md_rel; /**< user relocate function */
944 void *md_relctx; /**< user-provided context for md_rel */
947 /** A database transaction.
948 * Every operation requires a transaction handle.
951 MDB_txn *mt_parent; /**< parent of a nested txn */
952 MDB_txn *mt_child; /**< nested txn under this txn */
953 pgno_t mt_next_pgno; /**< next unallocated page */
954 /** The ID of this transaction. IDs are integers incrementing from 1.
955 * Only committed write transactions increment the ID. If a transaction
956 * aborts, the ID may be re-used by the next writer.
959 MDB_env *mt_env; /**< the DB environment */
960 /** The list of pages that became unused during this transaction.
963 /** The list of loose pages that became unused and may be reused
964 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
966 MDB_page *mt_loose_pgs;
967 /* #Number of loose pages (#mt_loose_pgs) */
969 /** The sorted list of dirty pages we temporarily wrote to disk
970 * because the dirty list was full. page numbers in here are
971 * shifted left by 1, deleted slots have the LSB set.
973 MDB_IDL mt_spill_pgs;
975 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
977 /** For read txns: This thread/txn's reader table slot, or NULL. */
980 /** Array of records for each DB known in the environment. */
982 /** Array of MDB_db records for each known DB */
984 /** Array of sequence numbers for each DB handle */
985 unsigned int *mt_dbiseqs;
986 /** @defgroup mt_dbflag Transaction DB Flags
990 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
991 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
992 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
993 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
995 /** In write txns, array of cursors for each DB */
996 MDB_cursor **mt_cursors;
997 /** Array of flags for each DB */
998 unsigned char *mt_dbflags;
999 /** Number of DB records in use. This number only ever increments;
1000 * we don't decrement it when individual DB handles are closed.
1004 /** @defgroup mdb_txn Transaction Flags
1008 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
1009 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1010 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1011 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1013 unsigned int mt_flags; /**< @ref mdb_txn */
1014 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1015 * Includes ancestor txns' dirty pages not hidden by other txns'
1016 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1017 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1019 unsigned int mt_dirty_room;
1022 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1023 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1024 * raise this on a 64 bit machine.
1026 #define CURSOR_STACK 32
1030 /** Cursors are used for all DB operations.
1031 * A cursor holds a path of (page pointer, key index) from the DB
1032 * root to a position in the DB, plus other state. #MDB_DUPSORT
1033 * cursors include an xcursor to the current data item. Write txns
1034 * track their cursors and keep them up to date when data moves.
1035 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1036 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1039 /** Next cursor on this DB in this txn */
1040 MDB_cursor *mc_next;
1041 /** Backup of the original cursor if this cursor is a shadow */
1042 MDB_cursor *mc_backup;
1043 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1044 struct MDB_xcursor *mc_xcursor;
1045 /** The transaction that owns this cursor */
1047 /** The database handle this cursor operates on */
1049 /** The database record for this cursor */
1051 /** The database auxiliary record for this cursor */
1053 /** The @ref mt_dbflag for this database */
1054 unsigned char *mc_dbflag;
1055 unsigned short mc_snum; /**< number of pushed pages */
1056 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1057 /** @defgroup mdb_cursor Cursor Flags
1059 * Cursor state flags.
1062 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1063 #define C_EOF 0x02 /**< No more data */
1064 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1065 #define C_DEL 0x08 /**< last op was a cursor_del */
1066 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1067 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1069 unsigned int mc_flags; /**< @ref mdb_cursor */
1070 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1071 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1074 /** Context for sorted-dup records.
1075 * We could have gone to a fully recursive design, with arbitrarily
1076 * deep nesting of sub-databases. But for now we only handle these
1077 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1079 typedef struct MDB_xcursor {
1080 /** A sub-cursor for traversing the Dup DB */
1081 MDB_cursor mx_cursor;
1082 /** The database record for this Dup DB */
1084 /** The auxiliary DB record for this Dup DB */
1086 /** The @ref mt_dbflag for this Dup DB */
1087 unsigned char mx_dbflag;
1090 /** State of FreeDB old pages, stored in the MDB_env */
1091 typedef struct MDB_pgstate {
1092 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1093 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1096 /** The database environment. */
1098 HANDLE me_fd; /**< The main data file */
1099 HANDLE me_lfd; /**< The lock file */
1100 HANDLE me_mfd; /**< just for writing the meta pages */
1101 /** Failed to update the meta page. Probably an I/O error. */
1102 #define MDB_FATAL_ERROR 0x80000000U
1103 /** Some fields are initialized. */
1104 #define MDB_ENV_ACTIVE 0x20000000U
1105 /** me_txkey is set */
1106 #define MDB_ENV_TXKEY 0x10000000U
1107 /** fdatasync is unreliable */
1108 #define MDB_FSYNCONLY 0x08000000U
1109 uint32_t me_flags; /**< @ref mdb_env */
1110 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1111 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1112 unsigned int me_maxreaders; /**< size of the reader table */
1113 unsigned int me_numreaders; /**< max numreaders set by this env */
1114 MDB_dbi me_numdbs; /**< number of DBs opened */
1115 MDB_dbi me_maxdbs; /**< size of the DB table */
1116 MDB_PID_T me_pid; /**< process ID of this env */
1117 char *me_path; /**< path to the DB files */
1118 char *me_map; /**< the memory map of the data file */
1119 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1120 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1121 void *me_pbuf; /**< scratch area for DUPSORT put() */
1122 MDB_txn *me_txn; /**< current write transaction */
1123 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1124 size_t me_mapsize; /**< size of the data memory map */
1125 off_t me_size; /**< current file size */
1126 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1127 MDB_dbx *me_dbxs; /**< array of static DB info */
1128 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1129 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1130 pthread_key_t me_txkey; /**< thread-key for readers */
1131 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1132 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1133 # define me_pglast me_pgstate.mf_pglast
1134 # define me_pghead me_pgstate.mf_pghead
1135 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1136 /** IDL of pages that became unused in a write txn */
1137 MDB_IDL me_free_pgs;
1138 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1139 MDB_ID2L me_dirty_list;
1140 /** Max number of freelist items that can fit in a single overflow page */
1142 /** Max size of a node on a page */
1143 unsigned int me_nodemax;
1144 #if !(MDB_MAXKEYSIZE)
1145 unsigned int me_maxkey; /**< max size of a key */
1147 int me_live_reader; /**< have liveness lock in reader table */
1149 int me_pidquery; /**< Used in OpenProcess */
1150 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1152 #elif defined(MDB_USE_POSIX_SEM)
1153 sem_t *me_rmutex; /* Shared mutexes are not supported */
1156 void *me_userctx; /**< User-settable context */
1157 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1160 /** Nested transaction */
1161 typedef struct MDB_ntxn {
1162 MDB_txn mnt_txn; /**< the transaction */
1163 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1166 /** max number of pages to commit in one writev() call */
1167 #define MDB_COMMIT_PAGES 64
1168 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1169 #undef MDB_COMMIT_PAGES
1170 #define MDB_COMMIT_PAGES IOV_MAX
1173 /** max bytes to write in one call */
1174 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1176 /** Check \b txn and \b dbi arguments to a function */
1177 #define TXN_DBI_EXIST(txn, dbi) \
1178 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1180 /** Check for misused \b dbi handles */
1181 #define TXN_DBI_CHANGED(txn, dbi) \
1182 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1184 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1185 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1186 static int mdb_page_touch(MDB_cursor *mc);
1188 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1189 static int mdb_page_search_root(MDB_cursor *mc,
1190 MDB_val *key, int modify);
1191 #define MDB_PS_MODIFY 1
1192 #define MDB_PS_ROOTONLY 2
1193 #define MDB_PS_FIRST 4
1194 #define MDB_PS_LAST 8
1195 static int mdb_page_search(MDB_cursor *mc,
1196 MDB_val *key, int flags);
1197 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1199 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1200 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1201 pgno_t newpgno, unsigned int nflags);
1203 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1204 static int mdb_env_pick_meta(const MDB_env *env);
1205 static int mdb_env_write_meta(MDB_txn *txn);
1206 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1207 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1209 static void mdb_env_close0(MDB_env *env, int excl);
1211 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1212 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1213 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1214 static void mdb_node_del(MDB_cursor *mc, int ksize);
1215 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1216 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1217 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1218 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1219 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1221 static int mdb_rebalance(MDB_cursor *mc);
1222 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1224 static void mdb_cursor_pop(MDB_cursor *mc);
1225 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1227 static int mdb_cursor_del0(MDB_cursor *mc);
1228 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1229 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1230 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1231 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1232 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1234 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1235 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1237 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1238 static void mdb_xcursor_init0(MDB_cursor *mc);
1239 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1241 static int mdb_drop0(MDB_cursor *mc, int subs);
1242 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1245 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1249 static SECURITY_DESCRIPTOR mdb_null_sd;
1250 static SECURITY_ATTRIBUTES mdb_all_sa;
1251 static int mdb_sec_inited;
1254 /** Return the library version info. */
1256 mdb_version(int *major, int *minor, int *patch)
1258 if (major) *major = MDB_VERSION_MAJOR;
1259 if (minor) *minor = MDB_VERSION_MINOR;
1260 if (patch) *patch = MDB_VERSION_PATCH;
1261 return MDB_VERSION_STRING;
1264 /** Table of descriptions for LMDB @ref errors */
1265 static char *const mdb_errstr[] = {
1266 "MDB_KEYEXIST: Key/data pair already exists",
1267 "MDB_NOTFOUND: No matching key/data pair found",
1268 "MDB_PAGE_NOTFOUND: Requested page not found",
1269 "MDB_CORRUPTED: Located page was wrong type",
1270 "MDB_PANIC: Update of meta page failed",
1271 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1272 "MDB_INVALID: File is not an LMDB file",
1273 "MDB_MAP_FULL: Environment mapsize limit reached",
1274 "MDB_DBS_FULL: Environment maxdbs limit reached",
1275 "MDB_READERS_FULL: Environment maxreaders limit reached",
1276 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1277 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1278 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1279 "MDB_PAGE_FULL: Internal error - page has no more space",
1280 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1281 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1282 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1283 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1284 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1285 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1289 mdb_strerror(int err)
1292 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1293 * This works as long as no function between the call to mdb_strerror
1294 * and the actual use of the message uses more than 4K of stack.
1297 char buf[1024], *ptr = buf;
1301 return ("Successful return: 0");
1303 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1304 i = err - MDB_KEYEXIST;
1305 return mdb_errstr[i];
1309 /* These are the C-runtime error codes we use. The comment indicates
1310 * their numeric value, and the Win32 error they would correspond to
1311 * if the error actually came from a Win32 API. A major mess, we should
1312 * have used LMDB-specific error codes for everything.
1315 case ENOENT: /* 2, FILE_NOT_FOUND */
1316 case EIO: /* 5, ACCESS_DENIED */
1317 case ENOMEM: /* 12, INVALID_ACCESS */
1318 case EACCES: /* 13, INVALID_DATA */
1319 case EBUSY: /* 16, CURRENT_DIRECTORY */
1320 case EINVAL: /* 22, BAD_COMMAND */
1321 case ENOSPC: /* 28, OUT_OF_PAPER */
1322 return strerror(err);
1327 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1328 FORMAT_MESSAGE_IGNORE_INSERTS,
1329 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1332 return strerror(err);
1336 /** assert(3) variant in cursor context */
1337 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1338 /** assert(3) variant in transaction context */
1339 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1340 /** assert(3) variant in environment context */
1341 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1344 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1345 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1348 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1349 const char *func, const char *file, int line)
1352 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1353 file, line, expr_txt, func);
1354 if (env->me_assert_func)
1355 env->me_assert_func(env, buf);
1356 fprintf(stderr, "%s\n", buf);
1360 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1364 /** Return the page number of \b mp which may be sub-page, for debug output */
1366 mdb_dbg_pgno(MDB_page *mp)
1369 COPY_PGNO(ret, mp->mp_pgno);
1373 /** Display a key in hexadecimal and return the address of the result.
1374 * @param[in] key the key to display
1375 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1376 * @return The key in hexadecimal form.
1379 mdb_dkey(MDB_val *key, char *buf)
1382 unsigned char *c = key->mv_data;
1388 if (key->mv_size > DKBUF_MAXKEYSIZE)
1389 return "MDB_MAXKEYSIZE";
1390 /* may want to make this a dynamic check: if the key is mostly
1391 * printable characters, print it as-is instead of converting to hex.
1395 for (i=0; i<key->mv_size; i++)
1396 ptr += sprintf(ptr, "%02x", *c++);
1398 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1404 mdb_leafnode_type(MDB_node *n)
1406 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1407 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1408 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1411 /** Display all the keys in the page. */
1413 mdb_page_list(MDB_page *mp)
1415 pgno_t pgno = mdb_dbg_pgno(mp);
1416 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1418 unsigned int i, nkeys, nsize, total = 0;
1422 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1423 case P_BRANCH: type = "Branch page"; break;
1424 case P_LEAF: type = "Leaf page"; break;
1425 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1426 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1427 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1429 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1430 pgno, mp->mp_pages, state);
1433 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1434 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1437 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1441 nkeys = NUMKEYS(mp);
1442 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1444 for (i=0; i<nkeys; i++) {
1445 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1446 key.mv_size = nsize = mp->mp_pad;
1447 key.mv_data = LEAF2KEY(mp, i, nsize);
1449 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1452 node = NODEPTR(mp, i);
1453 key.mv_size = node->mn_ksize;
1454 key.mv_data = node->mn_data;
1455 nsize = NODESIZE + key.mv_size;
1456 if (IS_BRANCH(mp)) {
1457 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1461 if (F_ISSET(node->mn_flags, F_BIGDATA))
1462 nsize += sizeof(pgno_t);
1464 nsize += NODEDSZ(node);
1466 nsize += sizeof(indx_t);
1467 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1468 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1470 total = EVEN(total);
1472 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1473 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1477 mdb_cursor_chk(MDB_cursor *mc)
1483 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1484 for (i=0; i<mc->mc_top; i++) {
1486 node = NODEPTR(mp, mc->mc_ki[i]);
1487 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1490 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1496 /** Count all the pages in each DB and in the freelist
1497 * and make sure it matches the actual number of pages
1499 * All named DBs must be open for a correct count.
1501 static void mdb_audit(MDB_txn *txn)
1505 MDB_ID freecount, count;
1510 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1511 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1512 freecount += *(MDB_ID *)data.mv_data;
1513 mdb_tassert(txn, rc == MDB_NOTFOUND);
1516 for (i = 0; i<txn->mt_numdbs; i++) {
1518 if (!(txn->mt_dbflags[i] & DB_VALID))
1520 mdb_cursor_init(&mc, txn, i, &mx);
1521 if (txn->mt_dbs[i].md_root == P_INVALID)
1523 count += txn->mt_dbs[i].md_branch_pages +
1524 txn->mt_dbs[i].md_leaf_pages +
1525 txn->mt_dbs[i].md_overflow_pages;
1526 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1527 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1528 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1531 mp = mc.mc_pg[mc.mc_top];
1532 for (j=0; j<NUMKEYS(mp); j++) {
1533 MDB_node *leaf = NODEPTR(mp, j);
1534 if (leaf->mn_flags & F_SUBDATA) {
1536 memcpy(&db, NODEDATA(leaf), sizeof(db));
1537 count += db.md_branch_pages + db.md_leaf_pages +
1538 db.md_overflow_pages;
1542 mdb_tassert(txn, rc == MDB_NOTFOUND);
1545 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1546 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1547 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1553 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1555 return txn->mt_dbxs[dbi].md_cmp(a, b);
1559 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1561 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1564 /** Allocate memory for a page.
1565 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1568 mdb_page_malloc(MDB_txn *txn, unsigned num)
1570 MDB_env *env = txn->mt_env;
1571 MDB_page *ret = env->me_dpages;
1572 size_t psize = env->me_psize, sz = psize, off;
1573 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1574 * For a single page alloc, we init everything after the page header.
1575 * For multi-page, we init the final page; if the caller needed that
1576 * many pages they will be filling in at least up to the last page.
1580 VGMEMP_ALLOC(env, ret, sz);
1581 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1582 env->me_dpages = ret->mp_next;
1585 psize -= off = PAGEHDRSZ;
1590 if ((ret = malloc(sz)) != NULL) {
1591 VGMEMP_ALLOC(env, ret, sz);
1592 if (!(env->me_flags & MDB_NOMEMINIT)) {
1593 memset((char *)ret + off, 0, psize);
1597 txn->mt_flags |= MDB_TXN_ERROR;
1601 /** Free a single page.
1602 * Saves single pages to a list, for future reuse.
1603 * (This is not used for multi-page overflow pages.)
1606 mdb_page_free(MDB_env *env, MDB_page *mp)
1608 mp->mp_next = env->me_dpages;
1609 VGMEMP_FREE(env, mp);
1610 env->me_dpages = mp;
1613 /** Free a dirty page */
1615 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1617 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1618 mdb_page_free(env, dp);
1620 /* large pages just get freed directly */
1621 VGMEMP_FREE(env, dp);
1626 /** Return all dirty pages to dpage list */
1628 mdb_dlist_free(MDB_txn *txn)
1630 MDB_env *env = txn->mt_env;
1631 MDB_ID2L dl = txn->mt_u.dirty_list;
1632 unsigned i, n = dl[0].mid;
1634 for (i = 1; i <= n; i++) {
1635 mdb_dpage_free(env, dl[i].mptr);
1640 /** Loosen or free a single page.
1641 * Saves single pages to a list for future reuse
1642 * in this same txn. It has been pulled from the freeDB
1643 * and already resides on the dirty list, but has been
1644 * deleted. Use these pages first before pulling again
1647 * If the page wasn't dirtied in this txn, just add it
1648 * to this txn's free list.
1651 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1654 pgno_t pgno = mp->mp_pgno;
1655 MDB_txn *txn = mc->mc_txn;
1657 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1658 if (txn->mt_parent) {
1659 MDB_ID2 *dl = txn->mt_u.dirty_list;
1660 /* If txn has a parent, make sure the page is in our
1664 unsigned x = mdb_mid2l_search(dl, pgno);
1665 if (x <= dl[0].mid && dl[x].mid == pgno) {
1666 if (mp != dl[x].mptr) { /* bad cursor? */
1667 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1668 txn->mt_flags |= MDB_TXN_ERROR;
1669 return MDB_CORRUPTED;
1676 /* no parent txn, so it's just ours */
1681 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1683 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1684 txn->mt_loose_pgs = mp;
1685 txn->mt_loose_count++;
1686 mp->mp_flags |= P_LOOSE;
1688 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1696 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1697 * @param[in] mc A cursor handle for the current operation.
1698 * @param[in] pflags Flags of the pages to update:
1699 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1700 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1701 * @return 0 on success, non-zero on failure.
1704 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1706 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1707 MDB_txn *txn = mc->mc_txn;
1713 int rc = MDB_SUCCESS, level;
1715 /* Mark pages seen by cursors */
1716 if (mc->mc_flags & C_UNTRACK)
1717 mc = NULL; /* will find mc in mt_cursors */
1718 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1719 for (; mc; mc=mc->mc_next) {
1720 if (!(mc->mc_flags & C_INITIALIZED))
1722 for (m3 = mc;; m3 = &mx->mx_cursor) {
1724 for (j=0; j<m3->mc_snum; j++) {
1726 if ((mp->mp_flags & Mask) == pflags)
1727 mp->mp_flags ^= P_KEEP;
1729 mx = m3->mc_xcursor;
1730 /* Proceed to mx if it is at a sub-database */
1731 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1733 if (! (mp && (mp->mp_flags & P_LEAF)))
1735 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1736 if (!(leaf->mn_flags & F_SUBDATA))
1745 /* Mark dirty root pages */
1746 for (i=0; i<txn->mt_numdbs; i++) {
1747 if (txn->mt_dbflags[i] & DB_DIRTY) {
1748 pgno_t pgno = txn->mt_dbs[i].md_root;
1749 if (pgno == P_INVALID)
1751 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1753 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1754 dp->mp_flags ^= P_KEEP;
1762 static int mdb_page_flush(MDB_txn *txn, int keep);
1764 /** Spill pages from the dirty list back to disk.
1765 * This is intended to prevent running into #MDB_TXN_FULL situations,
1766 * but note that they may still occur in a few cases:
1767 * 1) our estimate of the txn size could be too small. Currently this
1768 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1769 * 2) child txns may run out of space if their parents dirtied a
1770 * lot of pages and never spilled them. TODO: we probably should do
1771 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1772 * the parent's dirty_room is below a given threshold.
1774 * Otherwise, if not using nested txns, it is expected that apps will
1775 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1776 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1777 * If the txn never references them again, they can be left alone.
1778 * If the txn only reads them, they can be used without any fuss.
1779 * If the txn writes them again, they can be dirtied immediately without
1780 * going thru all of the work of #mdb_page_touch(). Such references are
1781 * handled by #mdb_page_unspill().
1783 * Also note, we never spill DB root pages, nor pages of active cursors,
1784 * because we'll need these back again soon anyway. And in nested txns,
1785 * we can't spill a page in a child txn if it was already spilled in a
1786 * parent txn. That would alter the parent txns' data even though
1787 * the child hasn't committed yet, and we'd have no way to undo it if
1788 * the child aborted.
1790 * @param[in] m0 cursor A cursor handle identifying the transaction and
1791 * database for which we are checking space.
1792 * @param[in] key For a put operation, the key being stored.
1793 * @param[in] data For a put operation, the data being stored.
1794 * @return 0 on success, non-zero on failure.
1797 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1799 MDB_txn *txn = m0->mc_txn;
1801 MDB_ID2L dl = txn->mt_u.dirty_list;
1802 unsigned int i, j, need;
1805 if (m0->mc_flags & C_SUB)
1808 /* Estimate how much space this op will take */
1809 i = m0->mc_db->md_depth;
1810 /* Named DBs also dirty the main DB */
1811 if (m0->mc_dbi > MAIN_DBI)
1812 i += txn->mt_dbs[MAIN_DBI].md_depth;
1813 /* For puts, roughly factor in the key+data size */
1815 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1816 i += i; /* double it for good measure */
1819 if (txn->mt_dirty_room > i)
1822 if (!txn->mt_spill_pgs) {
1823 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1824 if (!txn->mt_spill_pgs)
1827 /* purge deleted slots */
1828 MDB_IDL sl = txn->mt_spill_pgs;
1829 unsigned int num = sl[0];
1831 for (i=1; i<=num; i++) {
1838 /* Preserve pages which may soon be dirtied again */
1839 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1842 /* Less aggressive spill - we originally spilled the entire dirty list,
1843 * with a few exceptions for cursor pages and DB root pages. But this
1844 * turns out to be a lot of wasted effort because in a large txn many
1845 * of those pages will need to be used again. So now we spill only 1/8th
1846 * of the dirty pages. Testing revealed this to be a good tradeoff,
1847 * better than 1/2, 1/4, or 1/10.
1849 if (need < MDB_IDL_UM_MAX / 8)
1850 need = MDB_IDL_UM_MAX / 8;
1852 /* Save the page IDs of all the pages we're flushing */
1853 /* flush from the tail forward, this saves a lot of shifting later on. */
1854 for (i=dl[0].mid; i && need; i--) {
1855 MDB_ID pn = dl[i].mid << 1;
1857 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1859 /* Can't spill twice, make sure it's not already in a parent's
1862 if (txn->mt_parent) {
1864 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1865 if (tx2->mt_spill_pgs) {
1866 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1867 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1868 dp->mp_flags |= P_KEEP;
1876 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1880 mdb_midl_sort(txn->mt_spill_pgs);
1882 /* Flush the spilled part of dirty list */
1883 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1886 /* Reset any dirty pages we kept that page_flush didn't see */
1887 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1890 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1894 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1896 mdb_find_oldest(MDB_txn *txn)
1899 txnid_t mr, oldest = txn->mt_txnid - 1;
1900 if (txn->mt_env->me_txns) {
1901 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1902 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1913 /** Add a page to the txn's dirty list */
1915 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1918 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1920 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1921 insert = mdb_mid2l_append;
1923 insert = mdb_mid2l_insert;
1925 mid.mid = mp->mp_pgno;
1927 rc = insert(txn->mt_u.dirty_list, &mid);
1928 mdb_tassert(txn, rc == 0);
1929 txn->mt_dirty_room--;
1932 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1933 * me_pghead and mt_next_pgno.
1935 * If there are free pages available from older transactions, they
1936 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1937 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1938 * and move me_pglast to say which records were consumed. Only this
1939 * function can create me_pghead and move me_pglast/mt_next_pgno.
1940 * @param[in] mc cursor A cursor handle identifying the transaction and
1941 * database for which we are allocating.
1942 * @param[in] num the number of pages to allocate.
1943 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1944 * will always be satisfied by a single contiguous chunk of memory.
1945 * @return 0 on success, non-zero on failure.
1948 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1950 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1951 /* Get at most <Max_retries> more freeDB records once me_pghead
1952 * has enough pages. If not enough, use new pages from the map.
1953 * If <Paranoid> and mc is updating the freeDB, only get new
1954 * records if me_pghead is empty. Then the freelist cannot play
1955 * catch-up with itself by growing while trying to save it.
1957 enum { Paranoid = 1, Max_retries = 500 };
1959 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1961 int rc, retry = num * 60;
1962 MDB_txn *txn = mc->mc_txn;
1963 MDB_env *env = txn->mt_env;
1964 pgno_t pgno, *mop = env->me_pghead;
1965 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1967 txnid_t oldest = 0, last;
1972 /* If there are any loose pages, just use them */
1973 if (num == 1 && txn->mt_loose_pgs) {
1974 np = txn->mt_loose_pgs;
1975 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1976 txn->mt_loose_count--;
1977 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1985 /* If our dirty list is already full, we can't do anything */
1986 if (txn->mt_dirty_room == 0) {
1991 for (op = MDB_FIRST;; op = MDB_NEXT) {
1996 /* Seek a big enough contiguous page range. Prefer
1997 * pages at the tail, just truncating the list.
2003 if (mop[i-n2] == pgno+n2)
2010 if (op == MDB_FIRST) { /* 1st iteration */
2011 /* Prepare to fetch more and coalesce */
2012 last = env->me_pglast;
2013 oldest = env->me_pgoldest;
2014 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2017 key.mv_data = &last; /* will look up last+1 */
2018 key.mv_size = sizeof(last);
2020 if (Paranoid && mc->mc_dbi == FREE_DBI)
2023 if (Paranoid && retry < 0 && mop_len)
2027 /* Do not fetch more if the record will be too recent */
2028 if (oldest <= last) {
2030 oldest = mdb_find_oldest(txn);
2031 env->me_pgoldest = oldest;
2037 rc = mdb_cursor_get(&m2, &key, NULL, op);
2039 if (rc == MDB_NOTFOUND)
2043 last = *(txnid_t*)key.mv_data;
2044 if (oldest <= last) {
2046 oldest = mdb_find_oldest(txn);
2047 env->me_pgoldest = oldest;
2053 np = m2.mc_pg[m2.mc_top];
2054 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2055 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2058 idl = (MDB_ID *) data.mv_data;
2061 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2066 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2068 mop = env->me_pghead;
2070 env->me_pglast = last;
2072 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2073 last, txn->mt_dbs[FREE_DBI].md_root, i));
2075 DPRINTF(("IDL %"Z"u", idl[j]));
2077 /* Merge in descending sorted order */
2078 mdb_midl_xmerge(mop, idl);
2082 /* Use new pages from the map when nothing suitable in the freeDB */
2084 pgno = txn->mt_next_pgno;
2085 if (pgno + num >= env->me_maxpg) {
2086 DPUTS("DB size maxed out");
2092 if (env->me_flags & MDB_WRITEMAP) {
2093 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2095 if (!(np = mdb_page_malloc(txn, num))) {
2101 mop[0] = mop_len -= num;
2102 /* Move any stragglers down */
2103 for (j = i-num; j < mop_len; )
2104 mop[++j] = mop[++i];
2106 txn->mt_next_pgno = pgno + num;
2109 mdb_page_dirty(txn, np);
2115 txn->mt_flags |= MDB_TXN_ERROR;
2119 /** Copy the used portions of a non-overflow page.
2120 * @param[in] dst page to copy into
2121 * @param[in] src page to copy from
2122 * @param[in] psize size of a page
2125 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2127 enum { Align = sizeof(pgno_t) };
2128 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2130 /* If page isn't full, just copy the used portion. Adjust
2131 * alignment so memcpy may copy words instead of bytes.
2133 if ((unused &= -Align) && !IS_LEAF2(src)) {
2134 upper = (upper + PAGEBASE) & -Align;
2135 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2136 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2139 memcpy(dst, src, psize - unused);
2143 /** Pull a page off the txn's spill list, if present.
2144 * If a page being referenced was spilled to disk in this txn, bring
2145 * it back and make it dirty/writable again.
2146 * @param[in] txn the transaction handle.
2147 * @param[in] mp the page being referenced. It must not be dirty.
2148 * @param[out] ret the writable page, if any. ret is unchanged if
2149 * mp wasn't spilled.
2152 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2154 MDB_env *env = txn->mt_env;
2157 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2159 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2160 if (!tx2->mt_spill_pgs)
2162 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2163 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2166 if (txn->mt_dirty_room == 0)
2167 return MDB_TXN_FULL;
2168 if (IS_OVERFLOW(mp))
2172 if (env->me_flags & MDB_WRITEMAP) {
2175 np = mdb_page_malloc(txn, num);
2179 memcpy(np, mp, num * env->me_psize);
2181 mdb_page_copy(np, mp, env->me_psize);
2184 /* If in current txn, this page is no longer spilled.
2185 * If it happens to be the last page, truncate the spill list.
2186 * Otherwise mark it as deleted by setting the LSB.
2188 if (x == txn->mt_spill_pgs[0])
2189 txn->mt_spill_pgs[0]--;
2191 txn->mt_spill_pgs[x] |= 1;
2192 } /* otherwise, if belonging to a parent txn, the
2193 * page remains spilled until child commits
2196 mdb_page_dirty(txn, np);
2197 np->mp_flags |= P_DIRTY;
2205 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2206 * @param[in] mc cursor pointing to the page to be touched
2207 * @return 0 on success, non-zero on failure.
2210 mdb_page_touch(MDB_cursor *mc)
2212 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2213 MDB_txn *txn = mc->mc_txn;
2214 MDB_cursor *m2, *m3;
2218 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2219 if (txn->mt_flags & MDB_TXN_SPILLS) {
2221 rc = mdb_page_unspill(txn, mp, &np);
2227 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2228 (rc = mdb_page_alloc(mc, 1, &np)))
2231 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2232 mp->mp_pgno, pgno));
2233 mdb_cassert(mc, mp->mp_pgno != pgno);
2234 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2235 /* Update the parent page, if any, to point to the new page */
2237 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2238 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2239 SETPGNO(node, pgno);
2241 mc->mc_db->md_root = pgno;
2243 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2244 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2246 /* If txn has a parent, make sure the page is in our
2250 unsigned x = mdb_mid2l_search(dl, pgno);
2251 if (x <= dl[0].mid && dl[x].mid == pgno) {
2252 if (mp != dl[x].mptr) { /* bad cursor? */
2253 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2254 txn->mt_flags |= MDB_TXN_ERROR;
2255 return MDB_CORRUPTED;
2260 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2262 np = mdb_page_malloc(txn, 1);
2267 rc = mdb_mid2l_insert(dl, &mid);
2268 mdb_cassert(mc, rc == 0);
2273 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2275 np->mp_flags |= P_DIRTY;
2278 /* Adjust cursors pointing to mp */
2279 mc->mc_pg[mc->mc_top] = np;
2280 m2 = txn->mt_cursors[mc->mc_dbi];
2281 if (mc->mc_flags & C_SUB) {
2282 for (; m2; m2=m2->mc_next) {
2283 m3 = &m2->mc_xcursor->mx_cursor;
2284 if (m3->mc_snum < mc->mc_snum) continue;
2285 if (m3->mc_pg[mc->mc_top] == mp)
2286 m3->mc_pg[mc->mc_top] = np;
2289 for (; m2; m2=m2->mc_next) {
2290 if (m2->mc_snum < mc->mc_snum) continue;
2291 if (m2->mc_pg[mc->mc_top] == mp) {
2292 m2->mc_pg[mc->mc_top] = np;
2293 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2295 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2297 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2298 if (!(leaf->mn_flags & F_SUBDATA))
2299 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2307 txn->mt_flags |= MDB_TXN_ERROR;
2312 mdb_env_sync(MDB_env *env, int force)
2315 if (env->me_flags & MDB_RDONLY)
2317 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2318 if (env->me_flags & MDB_WRITEMAP) {
2319 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2320 ? MS_ASYNC : MS_SYNC;
2321 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2324 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2328 #ifdef BROKEN_FDATASYNC
2329 if (env->me_flags & MDB_FSYNCONLY) {
2330 if (fsync(env->me_fd))
2334 if (MDB_FDATASYNC(env->me_fd))
2341 /** Back up parent txn's cursors, then grab the originals for tracking */
2343 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2345 MDB_cursor *mc, *bk;
2350 for (i = src->mt_numdbs; --i >= 0; ) {
2351 if ((mc = src->mt_cursors[i]) != NULL) {
2352 size = sizeof(MDB_cursor);
2354 size += sizeof(MDB_xcursor);
2355 for (; mc; mc = bk->mc_next) {
2361 mc->mc_db = &dst->mt_dbs[i];
2362 /* Kill pointers into src - and dst to reduce abuse: The
2363 * user may not use mc until dst ends. Otherwise we'd...
2365 mc->mc_txn = NULL; /* ...set this to dst */
2366 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2367 if ((mx = mc->mc_xcursor) != NULL) {
2368 *(MDB_xcursor *)(bk+1) = *mx;
2369 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2371 mc->mc_next = dst->mt_cursors[i];
2372 dst->mt_cursors[i] = mc;
2379 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2380 * @param[in] txn the transaction handle.
2381 * @param[in] merge true to keep changes to parent cursors, false to revert.
2382 * @return 0 on success, non-zero on failure.
2385 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2387 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2391 for (i = txn->mt_numdbs; --i >= 0; ) {
2392 for (mc = cursors[i]; mc; mc = next) {
2394 if ((bk = mc->mc_backup) != NULL) {
2396 /* Commit changes to parent txn */
2397 mc->mc_next = bk->mc_next;
2398 mc->mc_backup = bk->mc_backup;
2399 mc->mc_txn = bk->mc_txn;
2400 mc->mc_db = bk->mc_db;
2401 mc->mc_dbflag = bk->mc_dbflag;
2402 if ((mx = mc->mc_xcursor) != NULL)
2403 mx->mx_cursor.mc_txn = bk->mc_txn;
2405 /* Abort nested txn */
2407 if ((mx = mc->mc_xcursor) != NULL)
2408 *mx = *(MDB_xcursor *)(bk+1);
2412 /* Only malloced cursors are permanently tracked. */
2420 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2423 mdb_txn_reset0(MDB_txn *txn, const char *act);
2425 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2431 Pidset = F_SETLK, Pidcheck = F_GETLK
2435 /** Set or check a pid lock. Set returns 0 on success.
2436 * Check returns 0 if the process is certainly dead, nonzero if it may
2437 * be alive (the lock exists or an error happened so we do not know).
2439 * On Windows Pidset is a no-op, we merely check for the existence
2440 * of the process with the given pid. On POSIX we use a single byte
2441 * lock on the lockfile, set at an offset equal to the pid.
2444 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2446 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2449 if (op == Pidcheck) {
2450 h = OpenProcess(env->me_pidquery, FALSE, pid);
2451 /* No documented "no such process" code, but other program use this: */
2453 return ErrCode() != ERROR_INVALID_PARAMETER;
2454 /* A process exists until all handles to it close. Has it exited? */
2455 ret = WaitForSingleObject(h, 0) != 0;
2462 struct flock lock_info;
2463 memset(&lock_info, 0, sizeof(lock_info));
2464 lock_info.l_type = F_WRLCK;
2465 lock_info.l_whence = SEEK_SET;
2466 lock_info.l_start = pid;
2467 lock_info.l_len = 1;
2468 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2469 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2471 } else if ((rc = ErrCode()) == EINTR) {
2479 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2480 * @param[in] txn the transaction handle to initialize
2481 * @return 0 on success, non-zero on failure.
2484 mdb_txn_renew0(MDB_txn *txn)
2486 MDB_env *env = txn->mt_env;
2487 MDB_txninfo *ti = env->me_txns;
2491 int rc, new_notls = 0;
2493 if (txn->mt_flags & MDB_TXN_RDONLY) {
2494 txn->mt_flags = MDB_TXN_RDONLY;
2496 txn->mt_numdbs = env->me_numdbs;
2497 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2499 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2500 txn->mt_txnid = meta->mm_txnid;
2501 txn->mt_u.reader = NULL;
2503 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2504 pthread_getspecific(env->me_txkey);
2506 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2507 return MDB_BAD_RSLOT;
2509 MDB_PID_T pid = env->me_pid;
2510 MDB_THR_T tid = pthread_self();
2512 if (!env->me_live_reader) {
2513 rc = mdb_reader_pid(env, Pidset, pid);
2516 env->me_live_reader = 1;
2520 nr = ti->mti_numreaders;
2521 for (i=0; i<nr; i++)
2522 if (ti->mti_readers[i].mr_pid == 0)
2524 if (i == env->me_maxreaders) {
2525 UNLOCK_MUTEX_R(env);
2526 return MDB_READERS_FULL;
2528 ti->mti_readers[i].mr_pid = pid;
2529 ti->mti_readers[i].mr_tid = tid;
2531 ti->mti_numreaders = ++nr;
2532 /* Save numreaders for un-mutexed mdb_env_close() */
2533 env->me_numreaders = nr;
2534 UNLOCK_MUTEX_R(env);
2536 r = &ti->mti_readers[i];
2537 new_notls = (env->me_flags & MDB_NOTLS);
2538 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2543 do /* LY: Retry on a race, ITS#7970. */
2544 r->mr_txnid = ti->mti_txnid;
2545 while(r->mr_txnid != ti->mti_txnid);
2546 txn->mt_txnid = r->mr_txnid;
2547 txn->mt_u.reader = r;
2548 meta = env->me_metas[txn->mt_txnid & 1];
2554 txn->mt_txnid = ti->mti_txnid;
2555 meta = env->me_metas[txn->mt_txnid & 1];
2557 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2558 txn->mt_txnid = meta->mm_txnid;
2561 txn->mt_numdbs = env->me_numdbs;
2564 if (txn->mt_txnid == mdb_debug_start)
2568 txn->mt_child = NULL;
2569 txn->mt_loose_pgs = NULL;
2570 txn->mt_loose_count = 0;
2571 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2572 txn->mt_u.dirty_list = env->me_dirty_list;
2573 txn->mt_u.dirty_list[0].mid = 0;
2574 txn->mt_free_pgs = env->me_free_pgs;
2575 txn->mt_free_pgs[0] = 0;
2576 txn->mt_spill_pgs = NULL;
2578 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2581 /* Copy the DB info and flags */
2582 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2584 /* Moved to here to avoid a data race in read TXNs */
2585 txn->mt_next_pgno = meta->mm_last_pg+1;
2587 for (i=2; i<txn->mt_numdbs; i++) {
2588 x = env->me_dbflags[i];
2589 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2590 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2592 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2594 if (env->me_maxpg < txn->mt_next_pgno) {
2595 mdb_txn_reset0(txn, "renew0-mapfail");
2597 txn->mt_u.reader->mr_pid = 0;
2598 txn->mt_u.reader = NULL;
2600 return MDB_MAP_RESIZED;
2607 mdb_txn_renew(MDB_txn *txn)
2611 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2614 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2615 DPUTS("environment had fatal error, must shutdown!");
2619 rc = mdb_txn_renew0(txn);
2620 if (rc == MDB_SUCCESS) {
2621 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2622 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2623 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2629 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2633 int rc, size, tsize = sizeof(MDB_txn);
2635 if (env->me_flags & MDB_FATAL_ERROR) {
2636 DPUTS("environment had fatal error, must shutdown!");
2639 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2642 /* Nested transactions: Max 1 child, write txns only, no writemap */
2643 if (parent->mt_child ||
2644 (flags & MDB_RDONLY) ||
2645 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2646 (env->me_flags & MDB_WRITEMAP))
2648 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2650 tsize = sizeof(MDB_ntxn);
2653 if (!(flags & MDB_RDONLY)) {
2655 txn = env->me_txn0; /* just reuse preallocated write txn */
2658 /* child txns use own copy of cursors */
2659 size += env->me_maxdbs * sizeof(MDB_cursor *);
2661 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2663 if ((txn = calloc(1, size)) == NULL) {
2664 DPRINTF(("calloc: %s", strerror(errno)));
2667 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2668 if (flags & MDB_RDONLY) {
2669 txn->mt_flags |= MDB_TXN_RDONLY;
2670 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2671 txn->mt_dbiseqs = env->me_dbiseqs;
2673 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2675 txn->mt_dbiseqs = parent->mt_dbiseqs;
2676 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2678 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2679 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2687 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2688 if (!txn->mt_u.dirty_list ||
2689 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2691 free(txn->mt_u.dirty_list);
2695 txn->mt_txnid = parent->mt_txnid;
2696 txn->mt_dirty_room = parent->mt_dirty_room;
2697 txn->mt_u.dirty_list[0].mid = 0;
2698 txn->mt_spill_pgs = NULL;
2699 txn->mt_next_pgno = parent->mt_next_pgno;
2700 parent->mt_child = txn;
2701 txn->mt_parent = parent;
2702 txn->mt_numdbs = parent->mt_numdbs;
2703 txn->mt_flags = parent->mt_flags;
2704 txn->mt_dbxs = parent->mt_dbxs;
2705 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2706 /* Copy parent's mt_dbflags, but clear DB_NEW */
2707 for (i=0; i<txn->mt_numdbs; i++)
2708 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2710 ntxn = (MDB_ntxn *)txn;
2711 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2712 if (env->me_pghead) {
2713 size = MDB_IDL_SIZEOF(env->me_pghead);
2714 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2716 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2721 rc = mdb_cursor_shadow(parent, txn);
2723 mdb_txn_reset0(txn, "beginchild-fail");
2725 rc = mdb_txn_renew0(txn);
2728 if (txn != env->me_txn0)
2732 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2733 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2734 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2741 mdb_txn_env(MDB_txn *txn)
2743 if(!txn) return NULL;
2747 /** Export or close DBI handles opened in this txn. */
2749 mdb_dbis_update(MDB_txn *txn, int keep)
2752 MDB_dbi n = txn->mt_numdbs;
2753 MDB_env *env = txn->mt_env;
2754 unsigned char *tdbflags = txn->mt_dbflags;
2756 for (i = n; --i >= 2;) {
2757 if (tdbflags[i] & DB_NEW) {
2759 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2761 char *ptr = env->me_dbxs[i].md_name.mv_data;
2763 env->me_dbxs[i].md_name.mv_data = NULL;
2764 env->me_dbxs[i].md_name.mv_size = 0;
2765 env->me_dbflags[i] = 0;
2766 env->me_dbiseqs[i]++;
2772 if (keep && env->me_numdbs < n)
2776 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2777 * May be called twice for readonly txns: First reset it, then abort.
2778 * @param[in] txn the transaction handle to reset
2779 * @param[in] act why the transaction is being reset
2782 mdb_txn_reset0(MDB_txn *txn, const char *act)
2784 MDB_env *env = txn->mt_env;
2786 /* Close any DBI handles opened in this txn */
2787 mdb_dbis_update(txn, 0);
2789 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2790 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2791 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2793 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2794 if (txn->mt_u.reader) {
2795 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2796 if (!(env->me_flags & MDB_NOTLS))
2797 txn->mt_u.reader = NULL; /* txn does not own reader */
2799 txn->mt_numdbs = 0; /* close nothing if called again */
2800 txn->mt_dbxs = NULL; /* mark txn as reset */
2802 pgno_t *pghead = env->me_pghead;
2804 mdb_cursors_close(txn, 0);
2805 if (!(env->me_flags & MDB_WRITEMAP)) {
2806 mdb_dlist_free(txn);
2809 if (!txn->mt_parent) {
2810 if (mdb_midl_shrink(&txn->mt_free_pgs))
2811 env->me_free_pgs = txn->mt_free_pgs;
2813 env->me_pghead = NULL;
2817 /* The writer mutex was locked in mdb_txn_begin. */
2819 UNLOCK_MUTEX_W(env);
2821 txn->mt_parent->mt_child = NULL;
2822 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2823 mdb_midl_free(txn->mt_free_pgs);
2824 mdb_midl_free(txn->mt_spill_pgs);
2825 free(txn->mt_u.dirty_list);
2828 mdb_midl_free(pghead);
2833 mdb_txn_reset(MDB_txn *txn)
2838 /* This call is only valid for read-only txns */
2839 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2842 mdb_txn_reset0(txn, "reset");
2846 mdb_txn_abort(MDB_txn *txn)
2852 mdb_txn_abort(txn->mt_child);
2854 mdb_txn_reset0(txn, "abort");
2855 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2856 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2857 txn->mt_u.reader->mr_pid = 0;
2859 if (txn != txn->mt_env->me_txn0)
2863 /** Save the freelist as of this transaction to the freeDB.
2864 * This changes the freelist. Keep trying until it stabilizes.
2867 mdb_freelist_save(MDB_txn *txn)
2869 /* env->me_pghead[] can grow and shrink during this call.
2870 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2871 * Page numbers cannot disappear from txn->mt_free_pgs[].
2874 MDB_env *env = txn->mt_env;
2875 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2876 txnid_t pglast = 0, head_id = 0;
2877 pgno_t freecnt = 0, *free_pgs, *mop;
2878 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2880 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2882 if (env->me_pghead) {
2883 /* Make sure first page of freeDB is touched and on freelist */
2884 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2885 if (rc && rc != MDB_NOTFOUND)
2889 if (!env->me_pghead && txn->mt_loose_pgs) {
2890 /* Put loose page numbers in mt_free_pgs, since
2891 * we may be unable to return them to me_pghead.
2893 MDB_page *mp = txn->mt_loose_pgs;
2894 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2896 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2897 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2898 txn->mt_loose_pgs = NULL;
2899 txn->mt_loose_count = 0;
2902 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2903 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2904 ? SSIZE_MAX : maxfree_1pg;
2907 /* Come back here after each Put() in case freelist changed */
2912 /* If using records from freeDB which we have not yet
2913 * deleted, delete them and any we reserved for me_pghead.
2915 while (pglast < env->me_pglast) {
2916 rc = mdb_cursor_first(&mc, &key, NULL);
2919 pglast = head_id = *(txnid_t *)key.mv_data;
2920 total_room = head_room = 0;
2921 mdb_tassert(txn, pglast <= env->me_pglast);
2922 rc = mdb_cursor_del(&mc, 0);
2927 /* Save the IDL of pages freed by this txn, to a single record */
2928 if (freecnt < txn->mt_free_pgs[0]) {
2930 /* Make sure last page of freeDB is touched and on freelist */
2931 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2932 if (rc && rc != MDB_NOTFOUND)
2935 free_pgs = txn->mt_free_pgs;
2936 /* Write to last page of freeDB */
2937 key.mv_size = sizeof(txn->mt_txnid);
2938 key.mv_data = &txn->mt_txnid;
2940 freecnt = free_pgs[0];
2941 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2942 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2945 /* Retry if mt_free_pgs[] grew during the Put() */
2946 free_pgs = txn->mt_free_pgs;
2947 } while (freecnt < free_pgs[0]);
2948 mdb_midl_sort(free_pgs);
2949 memcpy(data.mv_data, free_pgs, data.mv_size);
2952 unsigned int i = free_pgs[0];
2953 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2954 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2956 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2962 mop = env->me_pghead;
2963 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2965 /* Reserve records for me_pghead[]. Split it if multi-page,
2966 * to avoid searching freeDB for a page range. Use keys in
2967 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2969 if (total_room >= mop_len) {
2970 if (total_room == mop_len || --more < 0)
2972 } else if (head_room >= maxfree_1pg && head_id > 1) {
2973 /* Keep current record (overflow page), add a new one */
2977 /* (Re)write {key = head_id, IDL length = head_room} */
2978 total_room -= head_room;
2979 head_room = mop_len - total_room;
2980 if (head_room > maxfree_1pg && head_id > 1) {
2981 /* Overflow multi-page for part of me_pghead */
2982 head_room /= head_id; /* amortize page sizes */
2983 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2984 } else if (head_room < 0) {
2985 /* Rare case, not bothering to delete this record */
2988 key.mv_size = sizeof(head_id);
2989 key.mv_data = &head_id;
2990 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2991 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2994 /* IDL is initially empty, zero out at least the length */
2995 pgs = (pgno_t *)data.mv_data;
2996 j = head_room > clean_limit ? head_room : 0;
3000 total_room += head_room;
3003 /* Return loose page numbers to me_pghead, though usually none are
3004 * left at this point. The pages themselves remain in dirty_list.
3006 if (txn->mt_loose_pgs) {
3007 MDB_page *mp = txn->mt_loose_pgs;
3008 unsigned count = txn->mt_loose_count;
3010 /* Room for loose pages + temp IDL with same */
3011 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3013 mop = env->me_pghead;
3014 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3015 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3016 loose[ ++count ] = mp->mp_pgno;
3018 mdb_midl_sort(loose);
3019 mdb_midl_xmerge(mop, loose);
3020 txn->mt_loose_pgs = NULL;
3021 txn->mt_loose_count = 0;
3025 /* Fill in the reserved me_pghead records */
3031 rc = mdb_cursor_first(&mc, &key, &data);
3032 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3033 txnid_t id = *(txnid_t *)key.mv_data;
3034 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3037 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3039 if (len > mop_len) {
3041 data.mv_size = (len + 1) * sizeof(MDB_ID);
3043 data.mv_data = mop -= len;
3046 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3048 if (rc || !(mop_len -= len))
3055 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3056 * @param[in] txn the transaction that's being committed
3057 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3058 * @return 0 on success, non-zero on failure.
3061 mdb_page_flush(MDB_txn *txn, int keep)
3063 MDB_env *env = txn->mt_env;
3064 MDB_ID2L dl = txn->mt_u.dirty_list;
3065 unsigned psize = env->me_psize, j;
3066 int i, pagecount = dl[0].mid, rc;
3067 size_t size = 0, pos = 0;
3069 MDB_page *dp = NULL;
3073 struct iovec iov[MDB_COMMIT_PAGES];
3074 ssize_t wpos = 0, wsize = 0, wres;
3075 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3081 if (env->me_flags & MDB_WRITEMAP) {
3082 /* Clear dirty flags */
3083 while (++i <= pagecount) {
3085 /* Don't flush this page yet */
3086 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3087 dp->mp_flags &= ~P_KEEP;
3091 dp->mp_flags &= ~P_DIRTY;
3096 /* Write the pages */
3098 if (++i <= pagecount) {
3100 /* Don't flush this page yet */
3101 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3102 dp->mp_flags &= ~P_KEEP;
3107 /* clear dirty flag */
3108 dp->mp_flags &= ~P_DIRTY;
3111 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3116 /* Windows actually supports scatter/gather I/O, but only on
3117 * unbuffered file handles. Since we're relying on the OS page
3118 * cache for all our data, that's self-defeating. So we just
3119 * write pages one at a time. We use the ov structure to set
3120 * the write offset, to at least save the overhead of a Seek
3123 DPRINTF(("committing page %"Z"u", pgno));
3124 memset(&ov, 0, sizeof(ov));
3125 ov.Offset = pos & 0xffffffff;
3126 ov.OffsetHigh = pos >> 16 >> 16;
3127 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3129 DPRINTF(("WriteFile: %d", rc));
3133 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3134 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3137 /* Write previous page(s) */
3138 #ifdef MDB_USE_PWRITEV
3139 wres = pwritev(env->me_fd, iov, n, wpos);
3142 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3145 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3149 DPRINTF(("lseek: %s", strerror(rc)));
3152 wres = writev(env->me_fd, iov, n);
3155 if (wres != wsize) {
3160 DPRINTF(("Write error: %s", strerror(rc)));
3162 rc = EIO; /* TODO: Use which error code? */
3163 DPUTS("short write, filesystem full?");
3174 DPRINTF(("committing page %"Z"u", pgno));
3175 next_pos = pos + size;
3176 iov[n].iov_len = size;
3177 iov[n].iov_base = (char *)dp;
3183 /* MIPS has cache coherency issues, this is a no-op everywhere else
3184 * Note: for any size >= on-chip cache size, entire on-chip cache is
3187 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3189 for (i = keep; ++i <= pagecount; ) {
3191 /* This is a page we skipped above */
3194 dl[j].mid = dp->mp_pgno;
3197 mdb_dpage_free(env, dp);
3202 txn->mt_dirty_room += i - j;
3208 mdb_txn_commit(MDB_txn *txn)
3214 if (txn == NULL || txn->mt_env == NULL)
3217 if (txn->mt_child) {
3218 rc = mdb_txn_commit(txn->mt_child);
3219 txn->mt_child = NULL;
3226 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3227 mdb_dbis_update(txn, 1);
3228 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3233 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3234 DPUTS("error flag is set, can't commit");
3236 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3241 if (txn->mt_parent) {
3242 MDB_txn *parent = txn->mt_parent;
3246 unsigned x, y, len, ps_len;
3248 /* Append our free list to parent's */
3249 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3252 mdb_midl_free(txn->mt_free_pgs);
3253 /* Failures after this must either undo the changes
3254 * to the parent or set MDB_TXN_ERROR in the parent.
3257 parent->mt_next_pgno = txn->mt_next_pgno;
3258 parent->mt_flags = txn->mt_flags;
3260 /* Merge our cursors into parent's and close them */
3261 mdb_cursors_close(txn, 1);
3263 /* Update parent's DB table. */
3264 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3265 parent->mt_numdbs = txn->mt_numdbs;
3266 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3267 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3268 for (i=2; i<txn->mt_numdbs; i++) {
3269 /* preserve parent's DB_NEW status */
3270 x = parent->mt_dbflags[i] & DB_NEW;
3271 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3274 dst = parent->mt_u.dirty_list;
3275 src = txn->mt_u.dirty_list;
3276 /* Remove anything in our dirty list from parent's spill list */
3277 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3279 pspill[0] = (pgno_t)-1;
3280 /* Mark our dirty pages as deleted in parent spill list */
3281 for (i=0, len=src[0].mid; ++i <= len; ) {
3282 MDB_ID pn = src[i].mid << 1;
3283 while (pn > pspill[x])
3285 if (pn == pspill[x]) {
3290 /* Squash deleted pagenums if we deleted any */
3291 for (x=y; ++x <= ps_len; )
3292 if (!(pspill[x] & 1))
3293 pspill[++y] = pspill[x];
3297 /* Find len = length of merging our dirty list with parent's */
3299 dst[0].mid = 0; /* simplify loops */
3300 if (parent->mt_parent) {
3301 len = x + src[0].mid;
3302 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3303 for (i = x; y && i; y--) {
3304 pgno_t yp = src[y].mid;
3305 while (yp < dst[i].mid)
3307 if (yp == dst[i].mid) {
3312 } else { /* Simplify the above for single-ancestor case */
3313 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3315 /* Merge our dirty list with parent's */
3317 for (i = len; y; dst[i--] = src[y--]) {
3318 pgno_t yp = src[y].mid;
3319 while (yp < dst[x].mid)
3320 dst[i--] = dst[x--];
3321 if (yp == dst[x].mid)
3322 free(dst[x--].mptr);
3324 mdb_tassert(txn, i == x);
3326 free(txn->mt_u.dirty_list);
3327 parent->mt_dirty_room = txn->mt_dirty_room;
3328 if (txn->mt_spill_pgs) {
3329 if (parent->mt_spill_pgs) {
3330 /* TODO: Prevent failure here, so parent does not fail */
3331 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3333 parent->mt_flags |= MDB_TXN_ERROR;
3334 mdb_midl_free(txn->mt_spill_pgs);
3335 mdb_midl_sort(parent->mt_spill_pgs);
3337 parent->mt_spill_pgs = txn->mt_spill_pgs;
3341 /* Append our loose page list to parent's */
3342 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3344 *lp = txn->mt_loose_pgs;
3345 parent->mt_loose_count += txn->mt_loose_count;
3347 parent->mt_child = NULL;
3348 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3353 if (txn != env->me_txn) {
3354 DPUTS("attempt to commit unknown transaction");
3359 mdb_cursors_close(txn, 0);
3361 if (!txn->mt_u.dirty_list[0].mid &&
3362 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3365 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3366 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3368 /* Update DB root pointers */
3369 if (txn->mt_numdbs > 2) {
3373 data.mv_size = sizeof(MDB_db);
3375 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3376 for (i = 2; i < txn->mt_numdbs; i++) {
3377 if (txn->mt_dbflags[i] & DB_DIRTY) {
3378 if (TXN_DBI_CHANGED(txn, i)) {
3382 data.mv_data = &txn->mt_dbs[i];
3383 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3390 rc = mdb_freelist_save(txn);
3394 mdb_midl_free(env->me_pghead);
3395 env->me_pghead = NULL;
3396 if (mdb_midl_shrink(&txn->mt_free_pgs))
3397 env->me_free_pgs = txn->mt_free_pgs;
3403 if ((rc = mdb_page_flush(txn, 0)) ||
3404 (rc = mdb_env_sync(env, 0)) ||
3405 (rc = mdb_env_write_meta(txn)))
3408 /* Free P_LOOSE pages left behind in dirty_list */
3409 if (!(env->me_flags & MDB_WRITEMAP))
3410 mdb_dlist_free(txn);
3415 mdb_dbis_update(txn, 1);
3418 UNLOCK_MUTEX_W(env);
3419 if (txn != env->me_txn0)
3429 /** Read the environment parameters of a DB environment before
3430 * mapping it into memory.
3431 * @param[in] env the environment handle
3432 * @param[out] meta address of where to store the meta information
3433 * @return 0 on success, non-zero on failure.
3436 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3442 enum { Size = sizeof(pbuf) };
3444 /* We don't know the page size yet, so use a minimum value.
3445 * Read both meta pages so we can use the latest one.
3448 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3452 memset(&ov, 0, sizeof(ov));
3454 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3455 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3458 rc = pread(env->me_fd, &pbuf, Size, off);
3461 if (rc == 0 && off == 0)
3463 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3464 DPRINTF(("read: %s", mdb_strerror(rc)));
3468 p = (MDB_page *)&pbuf;
3470 if (!F_ISSET(p->mp_flags, P_META)) {
3471 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3476 if (m->mm_magic != MDB_MAGIC) {
3477 DPUTS("meta has invalid magic");
3481 if (m->mm_version != MDB_DATA_VERSION) {
3482 DPRINTF(("database is version %u, expected version %u",
3483 m->mm_version, MDB_DATA_VERSION));
3484 return MDB_VERSION_MISMATCH;
3487 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3494 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3496 meta->mm_magic = MDB_MAGIC;
3497 meta->mm_version = MDB_DATA_VERSION;
3498 meta->mm_mapsize = env->me_mapsize;
3499 meta->mm_psize = env->me_psize;
3500 meta->mm_last_pg = 1;
3501 meta->mm_flags = env->me_flags & 0xffff;
3502 meta->mm_flags |= MDB_INTEGERKEY;
3503 meta->mm_dbs[0].md_root = P_INVALID;
3504 meta->mm_dbs[1].md_root = P_INVALID;
3507 /** Write the environment parameters of a freshly created DB environment.
3508 * @param[in] env the environment handle
3509 * @param[out] meta address of where to store the meta information
3510 * @return 0 on success, non-zero on failure.
3513 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3521 memset(&ov, 0, sizeof(ov));
3522 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3524 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3527 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3528 len = pwrite(fd, ptr, size, pos); \
3529 if (len == -1 && ErrCode() == EINTR) continue; \
3530 rc = (len >= 0); break; } while(1)
3533 DPUTS("writing new meta page");
3535 psize = env->me_psize;
3537 mdb_env_init_meta0(env, meta);
3539 p = calloc(2, psize);
3541 p->mp_flags = P_META;
3542 *(MDB_meta *)METADATA(p) = *meta;
3544 q = (MDB_page *)((char *)p + psize);
3546 q->mp_flags = P_META;
3547 *(MDB_meta *)METADATA(q) = *meta;
3549 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3552 else if ((unsigned) len == psize * 2)
3560 /** Update the environment info to commit a transaction.
3561 * @param[in] txn the transaction that's being committed
3562 * @return 0 on success, non-zero on failure.
3565 mdb_env_write_meta(MDB_txn *txn)
3568 MDB_meta meta, metab, *mp;
3571 int rc, len, toggle;
3580 toggle = txn->mt_txnid & 1;
3581 DPRINTF(("writing meta page %d for root page %"Z"u",
3582 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3585 mp = env->me_metas[toggle];
3586 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3587 /* Persist any increases of mapsize config */
3588 if (mapsize < env->me_mapsize)
3589 mapsize = env->me_mapsize;
3591 if (env->me_flags & MDB_WRITEMAP) {
3592 mp->mm_mapsize = mapsize;
3593 mp->mm_dbs[0] = txn->mt_dbs[0];
3594 mp->mm_dbs[1] = txn->mt_dbs[1];
3595 mp->mm_last_pg = txn->mt_next_pgno - 1;
3596 mp->mm_txnid = txn->mt_txnid;
3597 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3598 unsigned meta_size = env->me_psize;
3599 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3602 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3603 if (meta_size < env->me_os_psize)
3604 meta_size += meta_size;
3609 if (MDB_MSYNC(ptr, meta_size, rc)) {
3616 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3617 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3619 meta.mm_mapsize = mapsize;
3620 meta.mm_dbs[0] = txn->mt_dbs[0];
3621 meta.mm_dbs[1] = txn->mt_dbs[1];
3622 meta.mm_last_pg = txn->mt_next_pgno - 1;
3623 meta.mm_txnid = txn->mt_txnid;
3625 off = offsetof(MDB_meta, mm_mapsize);
3626 ptr = (char *)&meta + off;
3627 len = sizeof(MDB_meta) - off;
3629 off += env->me_psize;
3632 /* Write to the SYNC fd */
3633 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3634 env->me_fd : env->me_mfd;
3638 memset(&ov, 0, sizeof(ov));
3640 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3644 rc = pwrite(mfd, ptr, len, off);
3647 rc = rc < 0 ? ErrCode() : EIO;
3650 DPUTS("write failed, disk error?");
3651 /* On a failure, the pagecache still contains the new data.
3652 * Write some old data back, to prevent it from being used.
3653 * Use the non-SYNC fd; we know it will fail anyway.
3655 meta.mm_last_pg = metab.mm_last_pg;
3656 meta.mm_txnid = metab.mm_txnid;
3658 memset(&ov, 0, sizeof(ov));
3660 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3662 r2 = pwrite(env->me_fd, ptr, len, off);
3663 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3666 env->me_flags |= MDB_FATAL_ERROR;
3669 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3670 CACHEFLUSH(env->me_map + off, len, DCACHE);
3672 /* Memory ordering issues are irrelevant; since the entire writer
3673 * is wrapped by wmutex, all of these changes will become visible
3674 * after the wmutex is unlocked. Since the DB is multi-version,
3675 * readers will get consistent data regardless of how fresh or
3676 * how stale their view of these values is.
3679 env->me_txns->mti_txnid = txn->mt_txnid;
3684 /** Check both meta pages to see which one is newer.
3685 * @param[in] env the environment handle
3686 * @return meta toggle (0 or 1).
3689 mdb_env_pick_meta(const MDB_env *env)
3691 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3695 mdb_env_create(MDB_env **env)
3699 e = calloc(1, sizeof(MDB_env));
3703 e->me_maxreaders = DEFAULT_READERS;
3704 e->me_maxdbs = e->me_numdbs = 2;
3705 e->me_fd = INVALID_HANDLE_VALUE;
3706 e->me_lfd = INVALID_HANDLE_VALUE;
3707 e->me_mfd = INVALID_HANDLE_VALUE;
3708 #ifdef MDB_USE_POSIX_SEM
3709 e->me_rmutex = SEM_FAILED;
3710 e->me_wmutex = SEM_FAILED;
3712 e->me_pid = getpid();
3713 GET_PAGESIZE(e->me_os_psize);
3714 VGMEMP_CREATE(e,0,0);
3720 mdb_env_map(MDB_env *env, void *addr)
3723 unsigned int flags = env->me_flags;
3727 LONG sizelo, sizehi;
3730 if (flags & MDB_RDONLY) {
3731 /* Don't set explicit map size, use whatever exists */
3736 msize = env->me_mapsize;
3737 sizelo = msize & 0xffffffff;
3738 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3740 /* Windows won't create mappings for zero length files.
3741 * and won't map more than the file size.
3742 * Just set the maxsize right now.
3744 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3745 || !SetEndOfFile(env->me_fd)
3746 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3750 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3751 PAGE_READWRITE : PAGE_READONLY,
3752 sizehi, sizelo, NULL);
3755 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3756 FILE_MAP_WRITE : FILE_MAP_READ,
3758 rc = env->me_map ? 0 : ErrCode();
3763 int prot = PROT_READ;
3764 if (flags & MDB_WRITEMAP) {
3766 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3769 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3771 if (env->me_map == MAP_FAILED) {
3776 if (flags & MDB_NORDAHEAD) {
3777 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3779 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3781 #ifdef POSIX_MADV_RANDOM
3782 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3783 #endif /* POSIX_MADV_RANDOM */
3784 #endif /* MADV_RANDOM */
3788 /* Can happen because the address argument to mmap() is just a
3789 * hint. mmap() can pick another, e.g. if the range is in use.
3790 * The MAP_FIXED flag would prevent that, but then mmap could
3791 * instead unmap existing pages to make room for the new map.
3793 if (addr && env->me_map != addr)
3794 return EBUSY; /* TODO: Make a new MDB_* error code? */
3796 p = (MDB_page *)env->me_map;
3797 env->me_metas[0] = METADATA(p);
3798 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3804 mdb_env_set_mapsize(MDB_env *env, size_t size)
3806 /* If env is already open, caller is responsible for making
3807 * sure there are no active txns.
3815 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3816 else if (size < env->me_mapsize) {
3817 /* If the configured size is smaller, make sure it's
3818 * still big enough. Silently round up to minimum if not.
3820 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3824 munmap(env->me_map, env->me_mapsize);
3825 env->me_mapsize = size;
3826 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3827 rc = mdb_env_map(env, old);
3831 env->me_mapsize = size;
3833 env->me_maxpg = env->me_mapsize / env->me_psize;
3838 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3842 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3847 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3849 if (env->me_map || readers < 1)
3851 env->me_maxreaders = readers;
3856 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3858 if (!env || !readers)
3860 *readers = env->me_maxreaders;
3865 mdb_fsize(HANDLE fd, size_t *size)
3868 LARGE_INTEGER fsize;
3870 if (!GetFileSizeEx(fd, &fsize))
3873 *size = fsize.QuadPart;
3885 #ifdef BROKEN_FDATASYNC
3886 #include <sys/utsname.h>
3887 #include <sys/vfs.h>
3890 /** Further setup required for opening an LMDB environment
3893 mdb_env_open2(MDB_env *env)
3895 unsigned int flags = env->me_flags;
3896 int i, newenv = 0, rc;
3900 /* See if we should use QueryLimited */
3902 if ((rc & 0xff) > 5)
3903 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3905 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3907 #ifdef BROKEN_FDATASYNC
3908 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
3909 * https://lkml.org/lkml/2012/9/3/83
3910 * Kernels after 3.6-rc6 are known good.
3911 * https://lkml.org/lkml/2012/9/10/556
3912 * See if the DB is on ext3/ext4, then check for new enough kernel
3913 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
3918 fstatfs(env->me_fd, &st);
3919 while (st.f_type == 0xEF53) {
3923 if (uts.release[0] < '3') {
3924 if (!strncmp(uts.release, "2.6.32.", 7)) {
3925 i = atoi(uts.release+7);
3927 break; /* 2.6.32.60 and newer is OK */
3928 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
3929 i = atoi(uts.release+7);
3931 break; /* 2.6.34.15 and newer is OK */
3933 } else if (uts.release[0] == '3') {
3934 i = atoi(uts.release+2);
3936 break; /* 3.6 and newer is OK */
3938 i = atoi(uts.release+4);
3940 break; /* 3.5.4 and newer is OK */
3941 } else if (i == 2) {
3942 i = atoi(uts.release+4);
3944 break; /* 3.2.30 and newer is OK */
3946 } else { /* 4.x and newer is OK */
3949 env->me_flags |= MDB_FSYNCONLY;
3955 memset(&meta, 0, sizeof(meta));
3957 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3960 DPUTS("new mdbenv");
3962 env->me_psize = env->me_os_psize;
3963 if (env->me_psize > MAX_PAGESIZE)
3964 env->me_psize = MAX_PAGESIZE;
3966 env->me_psize = meta.mm_psize;
3969 /* Was a mapsize configured? */
3970 if (!env->me_mapsize) {
3971 /* If this is a new environment, take the default,
3972 * else use the size recorded in the existing env.
3974 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3975 } else if (env->me_mapsize < meta.mm_mapsize) {
3976 /* If the configured size is smaller, make sure it's
3977 * still big enough. Silently round up to minimum if not.
3979 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3980 if (env->me_mapsize < minsize)
3981 env->me_mapsize = minsize;
3984 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
3989 if (flags & MDB_FIXEDMAP)
3990 meta.mm_address = env->me_map;
3991 i = mdb_env_init_meta(env, &meta);
3992 if (i != MDB_SUCCESS) {
3997 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3998 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4000 #if !(MDB_MAXKEYSIZE)
4001 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4003 env->me_maxpg = env->me_mapsize / env->me_psize;
4007 int toggle = mdb_env_pick_meta(env);
4008 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4010 DPRINTF(("opened database version %u, pagesize %u",
4011 env->me_metas[0]->mm_version, env->me_psize));
4012 DPRINTF(("using meta page %d", toggle));
4013 DPRINTF(("depth: %u", db->md_depth));
4014 DPRINTF(("entries: %"Z"u", db->md_entries));
4015 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4016 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4017 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4018 DPRINTF(("root: %"Z"u", db->md_root));
4026 /** Release a reader thread's slot in the reader lock table.
4027 * This function is called automatically when a thread exits.
4028 * @param[in] ptr This points to the slot in the reader lock table.
4031 mdb_env_reader_dest(void *ptr)
4033 MDB_reader *reader = ptr;
4039 /** Junk for arranging thread-specific callbacks on Windows. This is
4040 * necessarily platform and compiler-specific. Windows supports up
4041 * to 1088 keys. Let's assume nobody opens more than 64 environments
4042 * in a single process, for now. They can override this if needed.
4044 #ifndef MAX_TLS_KEYS
4045 #define MAX_TLS_KEYS 64
4047 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4048 static int mdb_tls_nkeys;
4050 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4054 case DLL_PROCESS_ATTACH: break;
4055 case DLL_THREAD_ATTACH: break;
4056 case DLL_THREAD_DETACH:
4057 for (i=0; i<mdb_tls_nkeys; i++) {
4058 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4060 mdb_env_reader_dest(r);
4064 case DLL_PROCESS_DETACH: break;
4069 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4071 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4075 /* Force some symbol references.
4076 * _tls_used forces the linker to create the TLS directory if not already done
4077 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4079 #pragma comment(linker, "/INCLUDE:_tls_used")
4080 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4081 #pragma const_seg(".CRT$XLB")
4082 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4083 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4086 #pragma comment(linker, "/INCLUDE:__tls_used")
4087 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4088 #pragma data_seg(".CRT$XLB")
4089 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4091 #endif /* WIN 32/64 */
4092 #endif /* !__GNUC__ */
4095 /** Downgrade the exclusive lock on the region back to shared */
4097 mdb_env_share_locks(MDB_env *env, int *excl)
4099 int rc = 0, toggle = mdb_env_pick_meta(env);
4101 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4106 /* First acquire a shared lock. The Unlock will
4107 * then release the existing exclusive lock.
4109 memset(&ov, 0, sizeof(ov));
4110 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4113 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4119 struct flock lock_info;
4120 /* The shared lock replaces the existing lock */
4121 memset((void *)&lock_info, 0, sizeof(lock_info));
4122 lock_info.l_type = F_RDLCK;
4123 lock_info.l_whence = SEEK_SET;
4124 lock_info.l_start = 0;
4125 lock_info.l_len = 1;
4126 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4127 (rc = ErrCode()) == EINTR) ;
4128 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4135 /** Try to get exclusive lock, otherwise shared.
4136 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4139 mdb_env_excl_lock(MDB_env *env, int *excl)
4143 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4147 memset(&ov, 0, sizeof(ov));
4148 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4155 struct flock lock_info;
4156 memset((void *)&lock_info, 0, sizeof(lock_info));
4157 lock_info.l_type = F_WRLCK;
4158 lock_info.l_whence = SEEK_SET;
4159 lock_info.l_start = 0;
4160 lock_info.l_len = 1;
4161 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4162 (rc = ErrCode()) == EINTR) ;
4166 # ifdef MDB_USE_POSIX_SEM
4167 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4170 lock_info.l_type = F_RDLCK;
4171 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4172 (rc = ErrCode()) == EINTR) ;
4182 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4184 * @(#) $Revision: 5.1 $
4185 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4186 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4188 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4192 * Please do not copyright this code. This code is in the public domain.
4194 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4195 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4196 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4197 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4198 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4199 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4200 * PERFORMANCE OF THIS SOFTWARE.
4203 * chongo <Landon Curt Noll> /\oo/\
4204 * http://www.isthe.com/chongo/
4206 * Share and Enjoy! :-)
4209 typedef unsigned long long mdb_hash_t;
4210 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4212 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4213 * @param[in] val value to hash
4214 * @param[in] hval initial value for hash
4215 * @return 64 bit hash
4217 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4218 * hval arg on the first call.
4221 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4223 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4224 unsigned char *end = s + val->mv_size;
4226 * FNV-1a hash each octet of the string
4229 /* xor the bottom with the current octet */
4230 hval ^= (mdb_hash_t)*s++;
4232 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4233 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4234 (hval << 7) + (hval << 8) + (hval << 40);
4236 /* return our new hash value */
4240 /** Hash the string and output the encoded hash.
4241 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4242 * very short name limits. We don't care about the encoding being reversible,
4243 * we just want to preserve as many bits of the input as possible in a
4244 * small printable string.
4245 * @param[in] str string to hash
4246 * @param[out] encbuf an array of 11 chars to hold the hash
4248 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4251 mdb_pack85(unsigned long l, char *out)
4255 for (i=0; i<5; i++) {
4256 *out++ = mdb_a85[l % 85];
4262 mdb_hash_enc(MDB_val *val, char *encbuf)
4264 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4266 mdb_pack85(h, encbuf);
4267 mdb_pack85(h>>32, encbuf+5);
4272 /** Open and/or initialize the lock region for the environment.
4273 * @param[in] env The LMDB environment.
4274 * @param[in] lpath The pathname of the file used for the lock region.
4275 * @param[in] mode The Unix permissions for the file, if we create it.
4276 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4277 * @return 0 on success, non-zero on failure.
4280 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4283 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4285 # define MDB_ERRCODE_ROFS EROFS
4286 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4287 # define MDB_CLOEXEC O_CLOEXEC
4290 # define MDB_CLOEXEC 0
4297 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4298 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4299 FILE_ATTRIBUTE_NORMAL, NULL);
4301 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4303 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4305 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4310 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4311 /* Lose record locks when exec*() */
4312 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4313 fcntl(env->me_lfd, F_SETFD, fdflags);
4316 if (!(env->me_flags & MDB_NOTLS)) {
4317 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4320 env->me_flags |= MDB_ENV_TXKEY;
4322 /* Windows TLS callbacks need help finding their TLS info. */
4323 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4327 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4331 /* Try to get exclusive lock. If we succeed, then
4332 * nobody is using the lock region and we should initialize it.
4334 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4337 size = GetFileSize(env->me_lfd, NULL);
4339 size = lseek(env->me_lfd, 0, SEEK_END);
4340 if (size == -1) goto fail_errno;
4342 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4343 if (size < rsize && *excl > 0) {
4345 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4346 || !SetEndOfFile(env->me_lfd))
4349 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4353 size = rsize - sizeof(MDB_txninfo);
4354 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4359 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4361 if (!mh) goto fail_errno;
4362 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4364 if (!env->me_txns) goto fail_errno;
4366 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4368 if (m == MAP_FAILED) goto fail_errno;
4374 BY_HANDLE_FILE_INFORMATION stbuf;
4383 if (!mdb_sec_inited) {
4384 InitializeSecurityDescriptor(&mdb_null_sd,
4385 SECURITY_DESCRIPTOR_REVISION);
4386 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4387 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4388 mdb_all_sa.bInheritHandle = FALSE;
4389 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4392 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4393 idbuf.volume = stbuf.dwVolumeSerialNumber;
4394 idbuf.nhigh = stbuf.nFileIndexHigh;
4395 idbuf.nlow = stbuf.nFileIndexLow;
4396 val.mv_data = &idbuf;
4397 val.mv_size = sizeof(idbuf);
4398 mdb_hash_enc(&val, encbuf);
4399 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4400 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4401 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4402 if (!env->me_rmutex) goto fail_errno;
4403 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4404 if (!env->me_wmutex) goto fail_errno;
4405 #elif defined(MDB_USE_POSIX_SEM)
4414 #if defined(__NetBSD__)
4415 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4417 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4418 idbuf.dev = stbuf.st_dev;
4419 idbuf.ino = stbuf.st_ino;
4420 val.mv_data = &idbuf;
4421 val.mv_size = sizeof(idbuf);
4422 mdb_hash_enc(&val, encbuf);
4423 #ifdef MDB_SHORT_SEMNAMES
4424 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4426 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4427 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4428 /* Clean up after a previous run, if needed: Try to
4429 * remove both semaphores before doing anything else.
4431 sem_unlink(env->me_txns->mti_rmname);
4432 sem_unlink(env->me_txns->mti_wmname);
4433 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4434 O_CREAT|O_EXCL, mode, 1);
4435 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4436 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4437 O_CREAT|O_EXCL, mode, 1);
4438 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4439 #else /* MDB_USE_POSIX_SEM */
4440 pthread_mutexattr_t mattr;
4442 if ((rc = pthread_mutexattr_init(&mattr))
4443 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4444 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4445 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4447 pthread_mutexattr_destroy(&mattr);
4448 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4450 env->me_txns->mti_magic = MDB_MAGIC;
4451 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4452 env->me_txns->mti_txnid = 0;
4453 env->me_txns->mti_numreaders = 0;
4456 if (env->me_txns->mti_magic != MDB_MAGIC) {
4457 DPUTS("lock region has invalid magic");
4461 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4462 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4463 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4464 rc = MDB_VERSION_MISMATCH;
4468 if (rc && rc != EACCES && rc != EAGAIN) {
4472 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4473 if (!env->me_rmutex) goto fail_errno;
4474 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4475 if (!env->me_wmutex) goto fail_errno;
4476 #elif defined(MDB_USE_POSIX_SEM)
4477 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4478 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4479 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4480 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4491 /** The name of the lock file in the DB environment */
4492 #define LOCKNAME "/lock.mdb"
4493 /** The name of the data file in the DB environment */
4494 #define DATANAME "/data.mdb"
4495 /** The suffix of the lock file when no subdir is used */
4496 #define LOCKSUFF "-lock"
4497 /** Only a subset of the @ref mdb_env flags can be changed
4498 * at runtime. Changing other flags requires closing the
4499 * environment and re-opening it with the new flags.
4501 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4502 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4503 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4505 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4506 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4510 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4512 int oflags, rc, len, excl = -1;
4513 char *lpath, *dpath;
4515 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4519 if (flags & MDB_NOSUBDIR) {
4520 rc = len + sizeof(LOCKSUFF) + len + 1;
4522 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4527 if (flags & MDB_NOSUBDIR) {
4528 dpath = lpath + len + sizeof(LOCKSUFF);
4529 sprintf(lpath, "%s" LOCKSUFF, path);
4530 strcpy(dpath, path);
4532 dpath = lpath + len + sizeof(LOCKNAME);
4533 sprintf(lpath, "%s" LOCKNAME, path);
4534 sprintf(dpath, "%s" DATANAME, path);
4538 flags |= env->me_flags;
4539 if (flags & MDB_RDONLY) {
4540 /* silently ignore WRITEMAP when we're only getting read access */
4541 flags &= ~MDB_WRITEMAP;
4543 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4544 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4547 env->me_flags = flags |= MDB_ENV_ACTIVE;
4551 env->me_path = strdup(path);
4552 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4553 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4554 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4555 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4560 /* For RDONLY, get lockfile after we know datafile exists */
4561 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4562 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4568 if (F_ISSET(flags, MDB_RDONLY)) {
4569 oflags = GENERIC_READ;
4570 len = OPEN_EXISTING;
4572 oflags = GENERIC_READ|GENERIC_WRITE;
4575 mode = FILE_ATTRIBUTE_NORMAL;
4576 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4577 NULL, len, mode, NULL);
4579 if (F_ISSET(flags, MDB_RDONLY))
4582 oflags = O_RDWR | O_CREAT;
4584 env->me_fd = open(dpath, oflags, mode);
4586 if (env->me_fd == INVALID_HANDLE_VALUE) {
4591 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4592 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4597 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4598 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4599 env->me_mfd = env->me_fd;
4601 /* Synchronous fd for meta writes. Needed even with
4602 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4605 len = OPEN_EXISTING;
4606 env->me_mfd = CreateFile(dpath, oflags,
4607 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4608 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4611 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4613 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4618 DPRINTF(("opened dbenv %p", (void *) env));
4620 rc = mdb_env_share_locks(env, &excl);
4624 if (!((flags & MDB_RDONLY) ||
4625 (env->me_pbuf = calloc(1, env->me_psize))))
4627 if (!(flags & MDB_RDONLY)) {
4629 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4630 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4631 txn = calloc(1, size);
4633 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4634 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4635 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4636 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4638 txn->mt_dbxs = env->me_dbxs;
4648 mdb_env_close0(env, excl);
4654 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4656 mdb_env_close0(MDB_env *env, int excl)
4660 if (!(env->me_flags & MDB_ENV_ACTIVE))
4663 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4665 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4666 free(env->me_dbxs[i].md_name.mv_data);
4671 free(env->me_dbiseqs);
4672 free(env->me_dbflags);
4674 free(env->me_dirty_list);
4676 mdb_midl_free(env->me_free_pgs);
4678 if (env->me_flags & MDB_ENV_TXKEY) {
4679 pthread_key_delete(env->me_txkey);
4681 /* Delete our key from the global list */
4682 for (i=0; i<mdb_tls_nkeys; i++)
4683 if (mdb_tls_keys[i] == env->me_txkey) {
4684 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4692 munmap(env->me_map, env->me_mapsize);
4694 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4695 (void) close(env->me_mfd);
4696 if (env->me_fd != INVALID_HANDLE_VALUE)
4697 (void) close(env->me_fd);
4699 MDB_PID_T pid = env->me_pid;
4700 /* Clearing readers is done in this function because
4701 * me_txkey with its destructor must be disabled first.
4703 for (i = env->me_numreaders; --i >= 0; )
4704 if (env->me_txns->mti_readers[i].mr_pid == pid)
4705 env->me_txns->mti_readers[i].mr_pid = 0;
4707 if (env->me_rmutex) {
4708 CloseHandle(env->me_rmutex);
4709 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4711 /* Windows automatically destroys the mutexes when
4712 * the last handle closes.
4714 #elif defined(MDB_USE_POSIX_SEM)
4715 if (env->me_rmutex != SEM_FAILED) {
4716 sem_close(env->me_rmutex);
4717 if (env->me_wmutex != SEM_FAILED)
4718 sem_close(env->me_wmutex);
4719 /* If we have the filelock: If we are the
4720 * only remaining user, clean up semaphores.
4723 mdb_env_excl_lock(env, &excl);
4725 sem_unlink(env->me_txns->mti_rmname);
4726 sem_unlink(env->me_txns->mti_wmname);
4730 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4732 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4735 /* Unlock the lockfile. Windows would have unlocked it
4736 * after closing anyway, but not necessarily at once.
4738 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4741 (void) close(env->me_lfd);
4744 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4749 mdb_env_close(MDB_env *env)
4756 VGMEMP_DESTROY(env);
4757 while ((dp = env->me_dpages) != NULL) {
4758 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4759 env->me_dpages = dp->mp_next;
4763 mdb_env_close0(env, 0);
4767 /** Compare two items pointing at aligned size_t's */
4769 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4771 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4772 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4775 /** Compare two items pointing at aligned unsigned int's */
4777 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4779 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4780 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4783 /** Compare two items pointing at unsigned ints of unknown alignment.
4784 * Nodes and keys are guaranteed to be 2-byte aligned.
4787 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4789 #if BYTE_ORDER == LITTLE_ENDIAN
4790 unsigned short *u, *c;
4793 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4794 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4797 } while(!x && u > (unsigned short *)a->mv_data);
4800 unsigned short *u, *c, *end;
4803 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4804 u = (unsigned short *)a->mv_data;
4805 c = (unsigned short *)b->mv_data;
4808 } while(!x && u < end);
4813 /** Compare two items pointing at size_t's of unknown alignment. */
4814 #ifdef MISALIGNED_OK
4815 # define mdb_cmp_clong mdb_cmp_long
4817 # define mdb_cmp_clong mdb_cmp_cint
4820 /** Compare two items lexically */
4822 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4829 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4835 diff = memcmp(a->mv_data, b->mv_data, len);
4836 return diff ? diff : len_diff<0 ? -1 : len_diff;
4839 /** Compare two items in reverse byte order */
4841 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4843 const unsigned char *p1, *p2, *p1_lim;
4847 p1_lim = (const unsigned char *)a->mv_data;
4848 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4849 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4851 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4857 while (p1 > p1_lim) {
4858 diff = *--p1 - *--p2;
4862 return len_diff<0 ? -1 : len_diff;
4865 /** Search for key within a page, using binary search.
4866 * Returns the smallest entry larger or equal to the key.
4867 * If exactp is non-null, stores whether the found entry was an exact match
4868 * in *exactp (1 or 0).
4869 * Updates the cursor index with the index of the found entry.
4870 * If no entry larger or equal to the key is found, returns NULL.
4873 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4875 unsigned int i = 0, nkeys;
4878 MDB_page *mp = mc->mc_pg[mc->mc_top];
4879 MDB_node *node = NULL;
4884 nkeys = NUMKEYS(mp);
4886 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4887 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4890 low = IS_LEAF(mp) ? 0 : 1;
4892 cmp = mc->mc_dbx->md_cmp;
4894 /* Branch pages have no data, so if using integer keys,
4895 * alignment is guaranteed. Use faster mdb_cmp_int.
4897 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4898 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4905 nodekey.mv_size = mc->mc_db->md_pad;
4906 node = NODEPTR(mp, 0); /* fake */
4907 while (low <= high) {
4908 i = (low + high) >> 1;
4909 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4910 rc = cmp(key, &nodekey);
4911 DPRINTF(("found leaf index %u [%s], rc = %i",
4912 i, DKEY(&nodekey), rc));
4921 while (low <= high) {
4922 i = (low + high) >> 1;
4924 node = NODEPTR(mp, i);
4925 nodekey.mv_size = NODEKSZ(node);
4926 nodekey.mv_data = NODEKEY(node);
4928 rc = cmp(key, &nodekey);
4931 DPRINTF(("found leaf index %u [%s], rc = %i",
4932 i, DKEY(&nodekey), rc));
4934 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4935 i, DKEY(&nodekey), NODEPGNO(node), rc));
4946 if (rc > 0) { /* Found entry is less than the key. */
4947 i++; /* Skip to get the smallest entry larger than key. */
4949 node = NODEPTR(mp, i);
4952 *exactp = (rc == 0 && nkeys > 0);
4953 /* store the key index */
4954 mc->mc_ki[mc->mc_top] = i;
4956 /* There is no entry larger or equal to the key. */
4959 /* nodeptr is fake for LEAF2 */
4965 mdb_cursor_adjust(MDB_cursor *mc, func)
4969 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4970 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4977 /** Pop a page off the top of the cursor's stack. */
4979 mdb_cursor_pop(MDB_cursor *mc)
4983 MDB_page *top = mc->mc_pg[mc->mc_top];
4989 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4990 DDBI(mc), (void *) mc));
4994 /** Push a page onto the top of the cursor's stack. */
4996 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4998 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4999 DDBI(mc), (void *) mc));
5001 if (mc->mc_snum >= CURSOR_STACK) {
5002 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5003 return MDB_CURSOR_FULL;
5006 mc->mc_top = mc->mc_snum++;
5007 mc->mc_pg[mc->mc_top] = mp;
5008 mc->mc_ki[mc->mc_top] = 0;
5013 /** Find the address of the page corresponding to a given page number.
5014 * @param[in] txn the transaction for this access.
5015 * @param[in] pgno the page number for the page to retrieve.
5016 * @param[out] ret address of a pointer where the page's address will be stored.
5017 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5018 * @return 0 on success, non-zero on failure.
5021 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5023 MDB_env *env = txn->mt_env;
5027 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
5031 MDB_ID2L dl = tx2->mt_u.dirty_list;
5033 /* Spilled pages were dirtied in this txn and flushed
5034 * because the dirty list got full. Bring this page
5035 * back in from the map (but don't unspill it here,
5036 * leave that unless page_touch happens again).
5038 if (tx2->mt_spill_pgs) {
5039 MDB_ID pn = pgno << 1;
5040 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5041 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5042 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5047 unsigned x = mdb_mid2l_search(dl, pgno);
5048 if (x <= dl[0].mid && dl[x].mid == pgno) {
5054 } while ((tx2 = tx2->mt_parent) != NULL);
5057 if (pgno < txn->mt_next_pgno) {
5059 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5061 DPRINTF(("page %"Z"u not found", pgno));
5062 txn->mt_flags |= MDB_TXN_ERROR;
5063 return MDB_PAGE_NOTFOUND;
5073 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5074 * The cursor is at the root page, set up the rest of it.
5077 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5079 MDB_page *mp = mc->mc_pg[mc->mc_top];
5083 while (IS_BRANCH(mp)) {
5087 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5088 mdb_cassert(mc, NUMKEYS(mp) > 1);
5089 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5091 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5093 if (flags & MDB_PS_LAST)
5094 i = NUMKEYS(mp) - 1;
5097 node = mdb_node_search(mc, key, &exact);
5099 i = NUMKEYS(mp) - 1;
5101 i = mc->mc_ki[mc->mc_top];
5103 mdb_cassert(mc, i > 0);
5107 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5110 mdb_cassert(mc, i < NUMKEYS(mp));
5111 node = NODEPTR(mp, i);
5113 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5116 mc->mc_ki[mc->mc_top] = i;
5117 if ((rc = mdb_cursor_push(mc, mp)))
5120 if (flags & MDB_PS_MODIFY) {
5121 if ((rc = mdb_page_touch(mc)) != 0)
5123 mp = mc->mc_pg[mc->mc_top];
5128 DPRINTF(("internal error, index points to a %02X page!?",
5130 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5131 return MDB_CORRUPTED;
5134 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5135 key ? DKEY(key) : "null"));
5136 mc->mc_flags |= C_INITIALIZED;
5137 mc->mc_flags &= ~C_EOF;
5142 /** Search for the lowest key under the current branch page.
5143 * This just bypasses a NUMKEYS check in the current page
5144 * before calling mdb_page_search_root(), because the callers
5145 * are all in situations where the current page is known to
5149 mdb_page_search_lowest(MDB_cursor *mc)
5151 MDB_page *mp = mc->mc_pg[mc->mc_top];
5152 MDB_node *node = NODEPTR(mp, 0);
5155 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5158 mc->mc_ki[mc->mc_top] = 0;
5159 if ((rc = mdb_cursor_push(mc, mp)))
5161 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5164 /** Search for the page a given key should be in.
5165 * Push it and its parent pages on the cursor stack.
5166 * @param[in,out] mc the cursor for this operation.
5167 * @param[in] key the key to search for, or NULL for first/last page.
5168 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5169 * are touched (updated with new page numbers).
5170 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5171 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5172 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5173 * @return 0 on success, non-zero on failure.
5176 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5181 /* Make sure the txn is still viable, then find the root from
5182 * the txn's db table and set it as the root of the cursor's stack.
5184 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5185 DPUTS("transaction has failed, must abort");
5188 /* Make sure we're using an up-to-date root */
5189 if (*mc->mc_dbflag & DB_STALE) {
5191 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5193 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5194 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5201 MDB_node *leaf = mdb_node_search(&mc2,
5202 &mc->mc_dbx->md_name, &exact);
5204 return MDB_NOTFOUND;
5205 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5208 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5210 /* The txn may not know this DBI, or another process may
5211 * have dropped and recreated the DB with other flags.
5213 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5214 return MDB_INCOMPATIBLE;
5215 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5217 *mc->mc_dbflag &= ~DB_STALE;
5219 root = mc->mc_db->md_root;
5221 if (root == P_INVALID) { /* Tree is empty. */
5222 DPUTS("tree is empty");
5223 return MDB_NOTFOUND;
5227 mdb_cassert(mc, root > 1);
5228 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5229 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5235 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5236 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5238 if (flags & MDB_PS_MODIFY) {
5239 if ((rc = mdb_page_touch(mc)))
5243 if (flags & MDB_PS_ROOTONLY)
5246 return mdb_page_search_root(mc, key, flags);
5250 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5252 MDB_txn *txn = mc->mc_txn;
5253 pgno_t pg = mp->mp_pgno;
5254 unsigned x = 0, ovpages = mp->mp_pages;
5255 MDB_env *env = txn->mt_env;
5256 MDB_IDL sl = txn->mt_spill_pgs;
5257 MDB_ID pn = pg << 1;
5260 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5261 /* If the page is dirty or on the spill list we just acquired it,
5262 * so we should give it back to our current free list, if any.
5263 * Otherwise put it onto the list of pages we freed in this txn.
5265 * Won't create me_pghead: me_pglast must be inited along with it.
5266 * Unsupported in nested txns: They would need to hide the page
5267 * range in ancestor txns' dirty and spilled lists.
5269 if (env->me_pghead &&
5271 ((mp->mp_flags & P_DIRTY) ||
5272 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5276 MDB_ID2 *dl, ix, iy;
5277 rc = mdb_midl_need(&env->me_pghead, ovpages);
5280 if (!(mp->mp_flags & P_DIRTY)) {
5281 /* This page is no longer spilled */
5288 /* Remove from dirty list */
5289 dl = txn->mt_u.dirty_list;
5291 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5297 mdb_cassert(mc, x > 1);
5299 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5300 txn->mt_flags |= MDB_TXN_ERROR;
5301 return MDB_CORRUPTED;
5304 if (!(env->me_flags & MDB_WRITEMAP))
5305 mdb_dpage_free(env, mp);
5307 /* Insert in me_pghead */
5308 mop = env->me_pghead;
5309 j = mop[0] + ovpages;
5310 for (i = mop[0]; i && mop[i] < pg; i--)
5316 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5320 mc->mc_db->md_overflow_pages -= ovpages;
5324 /** Return the data associated with a given node.
5325 * @param[in] txn The transaction for this operation.
5326 * @param[in] leaf The node being read.
5327 * @param[out] data Updated to point to the node's data.
5328 * @return 0 on success, non-zero on failure.
5331 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5333 MDB_page *omp; /* overflow page */
5337 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5338 data->mv_size = NODEDSZ(leaf);
5339 data->mv_data = NODEDATA(leaf);
5343 /* Read overflow data.
5345 data->mv_size = NODEDSZ(leaf);
5346 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5347 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5348 DPRINTF(("read overflow page %"Z"u failed", pgno));
5351 data->mv_data = METADATA(omp);
5357 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5358 MDB_val *key, MDB_val *data)
5365 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5367 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5370 if (txn->mt_flags & MDB_TXN_ERROR)
5373 mdb_cursor_init(&mc, txn, dbi, &mx);
5374 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5377 /** Find a sibling for a page.
5378 * Replaces the page at the top of the cursor's stack with the
5379 * specified sibling, if one exists.
5380 * @param[in] mc The cursor for this operation.
5381 * @param[in] move_right Non-zero if the right sibling is requested,
5382 * otherwise the left sibling.
5383 * @return 0 on success, non-zero on failure.
5386 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5392 if (mc->mc_snum < 2) {
5393 return MDB_NOTFOUND; /* root has no siblings */
5397 DPRINTF(("parent page is page %"Z"u, index %u",
5398 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5400 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5401 : (mc->mc_ki[mc->mc_top] == 0)) {
5402 DPRINTF(("no more keys left, moving to %s sibling",
5403 move_right ? "right" : "left"));
5404 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5405 /* undo cursor_pop before returning */
5412 mc->mc_ki[mc->mc_top]++;
5414 mc->mc_ki[mc->mc_top]--;
5415 DPRINTF(("just moving to %s index key %u",
5416 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5418 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5420 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5421 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5422 /* mc will be inconsistent if caller does mc_snum++ as above */
5423 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5427 mdb_cursor_push(mc, mp);
5429 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5434 /** Move the cursor to the next data item. */
5436 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5442 if (mc->mc_flags & C_EOF) {
5443 return MDB_NOTFOUND;
5446 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5448 mp = mc->mc_pg[mc->mc_top];
5450 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5451 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5452 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5453 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5454 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5455 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5456 if (rc == MDB_SUCCESS)
5457 MDB_GET_KEY(leaf, key);
5462 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5463 if (op == MDB_NEXT_DUP)
5464 return MDB_NOTFOUND;
5468 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5469 mdb_dbg_pgno(mp), (void *) mc));
5470 if (mc->mc_flags & C_DEL)
5473 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5474 DPUTS("=====> move to next sibling page");
5475 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5476 mc->mc_flags |= C_EOF;
5479 mp = mc->mc_pg[mc->mc_top];
5480 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5482 mc->mc_ki[mc->mc_top]++;
5485 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5486 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5489 key->mv_size = mc->mc_db->md_pad;
5490 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5494 mdb_cassert(mc, IS_LEAF(mp));
5495 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5497 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5498 mdb_xcursor_init1(mc, leaf);
5501 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5504 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5505 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5506 if (rc != MDB_SUCCESS)
5511 MDB_GET_KEY(leaf, key);
5515 /** Move the cursor to the previous data item. */
5517 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5523 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5525 mp = mc->mc_pg[mc->mc_top];
5527 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5528 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5529 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5530 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5531 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5532 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5533 if (rc == MDB_SUCCESS) {
5534 MDB_GET_KEY(leaf, key);
5535 mc->mc_flags &= ~C_EOF;
5541 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5542 if (op == MDB_PREV_DUP)
5543 return MDB_NOTFOUND;
5547 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5548 mdb_dbg_pgno(mp), (void *) mc));
5550 if (mc->mc_ki[mc->mc_top] == 0) {
5551 DPUTS("=====> move to prev sibling page");
5552 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5555 mp = mc->mc_pg[mc->mc_top];
5556 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5557 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5559 mc->mc_ki[mc->mc_top]--;
5561 mc->mc_flags &= ~C_EOF;
5563 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5564 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5567 key->mv_size = mc->mc_db->md_pad;
5568 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5572 mdb_cassert(mc, IS_LEAF(mp));
5573 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5575 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5576 mdb_xcursor_init1(mc, leaf);
5579 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5582 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5583 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5584 if (rc != MDB_SUCCESS)
5589 MDB_GET_KEY(leaf, key);
5593 /** Set the cursor on a specific data item. */
5595 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5596 MDB_cursor_op op, int *exactp)
5600 MDB_node *leaf = NULL;
5603 if (key->mv_size == 0)
5604 return MDB_BAD_VALSIZE;
5607 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5609 /* See if we're already on the right page */
5610 if (mc->mc_flags & C_INITIALIZED) {
5613 mp = mc->mc_pg[mc->mc_top];
5615 mc->mc_ki[mc->mc_top] = 0;
5616 return MDB_NOTFOUND;
5618 if (mp->mp_flags & P_LEAF2) {
5619 nodekey.mv_size = mc->mc_db->md_pad;
5620 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5622 leaf = NODEPTR(mp, 0);
5623 MDB_GET_KEY2(leaf, nodekey);
5625 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5627 /* Probably happens rarely, but first node on the page
5628 * was the one we wanted.
5630 mc->mc_ki[mc->mc_top] = 0;
5637 unsigned int nkeys = NUMKEYS(mp);
5639 if (mp->mp_flags & P_LEAF2) {
5640 nodekey.mv_data = LEAF2KEY(mp,
5641 nkeys-1, nodekey.mv_size);
5643 leaf = NODEPTR(mp, nkeys-1);
5644 MDB_GET_KEY2(leaf, nodekey);
5646 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5648 /* last node was the one we wanted */
5649 mc->mc_ki[mc->mc_top] = nkeys-1;
5655 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5656 /* This is definitely the right page, skip search_page */
5657 if (mp->mp_flags & P_LEAF2) {
5658 nodekey.mv_data = LEAF2KEY(mp,
5659 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5661 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5662 MDB_GET_KEY2(leaf, nodekey);
5664 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5666 /* current node was the one we wanted */
5676 /* If any parents have right-sibs, search.
5677 * Otherwise, there's nothing further.
5679 for (i=0; i<mc->mc_top; i++)
5681 NUMKEYS(mc->mc_pg[i])-1)
5683 if (i == mc->mc_top) {
5684 /* There are no other pages */
5685 mc->mc_ki[mc->mc_top] = nkeys;
5686 return MDB_NOTFOUND;
5690 /* There are no other pages */
5691 mc->mc_ki[mc->mc_top] = 0;
5692 if (op == MDB_SET_RANGE && !exactp) {
5696 return MDB_NOTFOUND;
5700 rc = mdb_page_search(mc, key, 0);
5701 if (rc != MDB_SUCCESS)
5704 mp = mc->mc_pg[mc->mc_top];
5705 mdb_cassert(mc, IS_LEAF(mp));
5708 leaf = mdb_node_search(mc, key, exactp);
5709 if (exactp != NULL && !*exactp) {
5710 /* MDB_SET specified and not an exact match. */
5711 return MDB_NOTFOUND;
5715 DPUTS("===> inexact leaf not found, goto sibling");
5716 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5717 return rc; /* no entries matched */
5718 mp = mc->mc_pg[mc->mc_top];
5719 mdb_cassert(mc, IS_LEAF(mp));
5720 leaf = NODEPTR(mp, 0);
5724 mc->mc_flags |= C_INITIALIZED;
5725 mc->mc_flags &= ~C_EOF;
5728 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5729 key->mv_size = mc->mc_db->md_pad;
5730 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5735 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5736 mdb_xcursor_init1(mc, leaf);
5739 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5740 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5741 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5744 if (op == MDB_GET_BOTH) {
5750 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5751 if (rc != MDB_SUCCESS)
5754 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5756 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5758 rc = mc->mc_dbx->md_dcmp(data, &d2);
5760 if (op == MDB_GET_BOTH || rc > 0)
5761 return MDB_NOTFOUND;
5768 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5769 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5774 /* The key already matches in all other cases */
5775 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5776 MDB_GET_KEY(leaf, key);
5777 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5782 /** Move the cursor to the first item in the database. */
5784 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5790 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5792 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5793 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5794 if (rc != MDB_SUCCESS)
5797 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5799 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5800 mc->mc_flags |= C_INITIALIZED;
5801 mc->mc_flags &= ~C_EOF;
5803 mc->mc_ki[mc->mc_top] = 0;
5805 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5806 key->mv_size = mc->mc_db->md_pad;
5807 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5812 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5813 mdb_xcursor_init1(mc, leaf);
5814 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5818 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5822 MDB_GET_KEY(leaf, key);
5826 /** Move the cursor to the last item in the database. */
5828 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5834 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5836 if (!(mc->mc_flags & C_EOF)) {
5838 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5839 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5840 if (rc != MDB_SUCCESS)
5843 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5846 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5847 mc->mc_flags |= C_INITIALIZED|C_EOF;
5848 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5850 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5851 key->mv_size = mc->mc_db->md_pad;
5852 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5857 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5858 mdb_xcursor_init1(mc, leaf);
5859 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5863 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5868 MDB_GET_KEY(leaf, key);
5873 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5878 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5883 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5887 case MDB_GET_CURRENT:
5888 if (!(mc->mc_flags & C_INITIALIZED)) {
5891 MDB_page *mp = mc->mc_pg[mc->mc_top];
5892 int nkeys = NUMKEYS(mp);
5893 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5894 mc->mc_ki[mc->mc_top] = nkeys;
5900 key->mv_size = mc->mc_db->md_pad;
5901 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5903 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5904 MDB_GET_KEY(leaf, key);
5906 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5907 if (mc->mc_flags & C_DEL)
5908 mdb_xcursor_init1(mc, leaf);
5909 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5911 rc = mdb_node_read(mc->mc_txn, leaf, data);
5918 case MDB_GET_BOTH_RANGE:
5923 if (mc->mc_xcursor == NULL) {
5924 rc = MDB_INCOMPATIBLE;
5934 rc = mdb_cursor_set(mc, key, data, op,
5935 op == MDB_SET_RANGE ? NULL : &exact);
5938 case MDB_GET_MULTIPLE:
5939 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5943 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5944 rc = MDB_INCOMPATIBLE;
5948 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5949 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5952 case MDB_NEXT_MULTIPLE:
5957 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5958 rc = MDB_INCOMPATIBLE;
5961 if (!(mc->mc_flags & C_INITIALIZED))
5962 rc = mdb_cursor_first(mc, key, data);
5964 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5965 if (rc == MDB_SUCCESS) {
5966 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5969 mx = &mc->mc_xcursor->mx_cursor;
5970 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5972 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5973 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5981 case MDB_NEXT_NODUP:
5982 if (!(mc->mc_flags & C_INITIALIZED))
5983 rc = mdb_cursor_first(mc, key, data);
5985 rc = mdb_cursor_next(mc, key, data, op);
5989 case MDB_PREV_NODUP:
5990 if (!(mc->mc_flags & C_INITIALIZED)) {
5991 rc = mdb_cursor_last(mc, key, data);
5994 mc->mc_flags |= C_INITIALIZED;
5995 mc->mc_ki[mc->mc_top]++;
5997 rc = mdb_cursor_prev(mc, key, data, op);
6000 rc = mdb_cursor_first(mc, key, data);
6003 mfunc = mdb_cursor_first;
6005 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6009 if (mc->mc_xcursor == NULL) {
6010 rc = MDB_INCOMPATIBLE;
6014 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6015 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6016 MDB_GET_KEY(leaf, key);
6017 rc = mdb_node_read(mc->mc_txn, leaf, data);
6021 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6025 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6028 rc = mdb_cursor_last(mc, key, data);
6031 mfunc = mdb_cursor_last;
6034 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6039 if (mc->mc_flags & C_DEL)
6040 mc->mc_flags ^= C_DEL;
6045 /** Touch all the pages in the cursor stack. Set mc_top.
6046 * Makes sure all the pages are writable, before attempting a write operation.
6047 * @param[in] mc The cursor to operate on.
6050 mdb_cursor_touch(MDB_cursor *mc)
6052 int rc = MDB_SUCCESS;
6054 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6057 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6059 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6060 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6063 *mc->mc_dbflag |= DB_DIRTY;
6068 rc = mdb_page_touch(mc);
6069 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6070 mc->mc_top = mc->mc_snum-1;
6075 /** Do not spill pages to disk if txn is getting full, may fail instead */
6076 #define MDB_NOSPILL 0x8000
6079 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6082 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
6084 MDB_node *leaf = NULL;
6087 MDB_val xdata, *rdata, dkey, olddata;
6089 int do_sub = 0, insert_key, insert_data;
6090 unsigned int mcount = 0, dcount = 0, nospill;
6093 unsigned int nflags;
6096 if (mc == NULL || key == NULL)
6099 env = mc->mc_txn->mt_env;
6101 /* Check this first so counter will always be zero on any
6104 if (flags & MDB_MULTIPLE) {
6105 dcount = data[1].mv_size;
6106 data[1].mv_size = 0;
6107 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6108 return MDB_INCOMPATIBLE;
6111 nospill = flags & MDB_NOSPILL;
6112 flags &= ~MDB_NOSPILL;
6114 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6115 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6117 if (key->mv_size-1 >= ENV_MAXKEY(env))
6118 return MDB_BAD_VALSIZE;
6120 #if SIZE_MAX > MAXDATASIZE
6121 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6122 return MDB_BAD_VALSIZE;
6124 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6125 return MDB_BAD_VALSIZE;
6128 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6129 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6133 if (flags == MDB_CURRENT) {
6134 if (!(mc->mc_flags & C_INITIALIZED))
6137 } else if (mc->mc_db->md_root == P_INVALID) {
6138 /* new database, cursor has nothing to point to */
6141 mc->mc_flags &= ~C_INITIALIZED;
6146 if (flags & MDB_APPEND) {
6148 rc = mdb_cursor_last(mc, &k2, &d2);
6150 rc = mc->mc_dbx->md_cmp(key, &k2);
6153 mc->mc_ki[mc->mc_top]++;
6155 /* new key is <= last key */
6160 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6162 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6163 DPRINTF(("duplicate key [%s]", DKEY(key)));
6165 return MDB_KEYEXIST;
6167 if (rc && rc != MDB_NOTFOUND)
6171 if (mc->mc_flags & C_DEL)
6172 mc->mc_flags ^= C_DEL;
6174 /* Cursor is positioned, check for room in the dirty list */
6176 if (flags & MDB_MULTIPLE) {
6178 xdata.mv_size = data->mv_size * dcount;
6182 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6186 if (rc == MDB_NO_ROOT) {
6188 /* new database, write a root leaf page */
6189 DPUTS("allocating new root leaf page");
6190 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6193 mdb_cursor_push(mc, np);
6194 mc->mc_db->md_root = np->mp_pgno;
6195 mc->mc_db->md_depth++;
6196 *mc->mc_dbflag |= DB_DIRTY;
6197 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6199 np->mp_flags |= P_LEAF2;
6200 mc->mc_flags |= C_INITIALIZED;
6202 /* make sure all cursor pages are writable */
6203 rc2 = mdb_cursor_touch(mc);
6208 insert_key = insert_data = rc;
6210 /* The key does not exist */
6211 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6212 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6213 LEAFSIZE(key, data) > env->me_nodemax)
6215 /* Too big for a node, insert in sub-DB. Set up an empty
6216 * "old sub-page" for prep_subDB to expand to a full page.
6218 fp_flags = P_LEAF|P_DIRTY;
6220 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6221 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6222 olddata.mv_size = PAGEHDRSZ;
6226 /* there's only a key anyway, so this is a no-op */
6227 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6229 unsigned int ksize = mc->mc_db->md_pad;
6230 if (key->mv_size != ksize)
6231 return MDB_BAD_VALSIZE;
6232 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6233 memcpy(ptr, key->mv_data, ksize);
6235 /* if overwriting slot 0 of leaf, need to
6236 * update branch key if there is a parent page
6238 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6239 unsigned short top = mc->mc_top;
6241 /* slot 0 is always an empty key, find real slot */
6242 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6244 if (mc->mc_ki[mc->mc_top])
6245 rc2 = mdb_update_key(mc, key);
6256 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6257 olddata.mv_size = NODEDSZ(leaf);
6258 olddata.mv_data = NODEDATA(leaf);
6261 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6262 /* Prepare (sub-)page/sub-DB to accept the new item,
6263 * if needed. fp: old sub-page or a header faking
6264 * it. mp: new (sub-)page. offset: growth in page
6265 * size. xdata: node data with new page or DB.
6267 unsigned i, offset = 0;
6268 mp = fp = xdata.mv_data = env->me_pbuf;
6269 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6271 /* Was a single item before, must convert now */
6272 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6273 /* Just overwrite the current item */
6274 if (flags == MDB_CURRENT)
6277 #if UINT_MAX < SIZE_MAX
6278 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6279 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
6281 /* does data match? */
6282 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
6283 if (flags & MDB_NODUPDATA)
6284 return MDB_KEYEXIST;
6289 /* Back up original data item */
6290 dkey.mv_size = olddata.mv_size;
6291 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6293 /* Make sub-page header for the dup items, with dummy body */
6294 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6295 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6296 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6297 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6298 fp->mp_flags |= P_LEAF2;
6299 fp->mp_pad = data->mv_size;
6300 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6302 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6303 (dkey.mv_size & 1) + (data->mv_size & 1);
6305 fp->mp_upper = xdata.mv_size - PAGEBASE;
6306 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6307 } else if (leaf->mn_flags & F_SUBDATA) {
6308 /* Data is on sub-DB, just store it */
6309 flags |= F_DUPDATA|F_SUBDATA;
6312 /* Data is on sub-page */
6313 fp = olddata.mv_data;
6316 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6317 offset = EVEN(NODESIZE + sizeof(indx_t) +
6321 offset = fp->mp_pad;
6322 if (SIZELEFT(fp) < offset) {
6323 offset *= 4; /* space for 4 more */
6326 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6328 fp->mp_flags |= P_DIRTY;
6329 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6330 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6334 xdata.mv_size = olddata.mv_size + offset;
6337 fp_flags = fp->mp_flags;
6338 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6339 /* Too big for a sub-page, convert to sub-DB */
6340 fp_flags &= ~P_SUBP;
6342 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6343 fp_flags |= P_LEAF2;
6344 dummy.md_pad = fp->mp_pad;
6345 dummy.md_flags = MDB_DUPFIXED;
6346 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6347 dummy.md_flags |= MDB_INTEGERKEY;
6353 dummy.md_branch_pages = 0;
6354 dummy.md_leaf_pages = 1;
6355 dummy.md_overflow_pages = 0;
6356 dummy.md_entries = NUMKEYS(fp);
6357 xdata.mv_size = sizeof(MDB_db);
6358 xdata.mv_data = &dummy;
6359 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6361 offset = env->me_psize - olddata.mv_size;
6362 flags |= F_DUPDATA|F_SUBDATA;
6363 dummy.md_root = mp->mp_pgno;
6366 mp->mp_flags = fp_flags | P_DIRTY;
6367 mp->mp_pad = fp->mp_pad;
6368 mp->mp_lower = fp->mp_lower;
6369 mp->mp_upper = fp->mp_upper + offset;
6370 if (fp_flags & P_LEAF2) {
6371 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6373 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6374 olddata.mv_size - fp->mp_upper - PAGEBASE);
6375 for (i=0; i<NUMKEYS(fp); i++)
6376 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6384 mdb_node_del(mc, 0);
6388 /* overflow page overwrites need special handling */
6389 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6392 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6394 memcpy(&pg, olddata.mv_data, sizeof(pg));
6395 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6397 ovpages = omp->mp_pages;
6399 /* Is the ov page large enough? */
6400 if (ovpages >= dpages) {
6401 if (!(omp->mp_flags & P_DIRTY) &&
6402 (level || (env->me_flags & MDB_WRITEMAP)))
6404 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6407 level = 0; /* dirty in this txn or clean */
6410 if (omp->mp_flags & P_DIRTY) {
6411 /* yes, overwrite it. Note in this case we don't
6412 * bother to try shrinking the page if the new data
6413 * is smaller than the overflow threshold.
6416 /* It is writable only in a parent txn */
6417 size_t sz = (size_t) env->me_psize * ovpages, off;
6418 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6424 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6425 mdb_cassert(mc, rc2 == 0);
6426 if (!(flags & MDB_RESERVE)) {
6427 /* Copy end of page, adjusting alignment so
6428 * compiler may copy words instead of bytes.
6430 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6431 memcpy((size_t *)((char *)np + off),
6432 (size_t *)((char *)omp + off), sz - off);
6435 memcpy(np, omp, sz); /* Copy beginning of page */
6438 SETDSZ(leaf, data->mv_size);
6439 if (F_ISSET(flags, MDB_RESERVE))
6440 data->mv_data = METADATA(omp);
6442 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6446 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6448 } else if (data->mv_size == olddata.mv_size) {
6449 /* same size, just replace it. Note that we could
6450 * also reuse this node if the new data is smaller,
6451 * but instead we opt to shrink the node in that case.
6453 if (F_ISSET(flags, MDB_RESERVE))
6454 data->mv_data = olddata.mv_data;
6455 else if (!(mc->mc_flags & C_SUB))
6456 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6458 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6463 mdb_node_del(mc, 0);
6469 nflags = flags & NODE_ADD_FLAGS;
6470 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6471 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6472 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6473 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6475 nflags |= MDB_SPLIT_REPLACE;
6476 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6478 /* There is room already in this leaf page. */
6479 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6480 if (rc == 0 && insert_key) {
6481 /* Adjust other cursors pointing to mp */
6482 MDB_cursor *m2, *m3;
6483 MDB_dbi dbi = mc->mc_dbi;
6484 unsigned i = mc->mc_top;
6485 MDB_page *mp = mc->mc_pg[i];
6487 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6488 if (mc->mc_flags & C_SUB)
6489 m3 = &m2->mc_xcursor->mx_cursor;
6492 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6493 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6500 if (rc == MDB_SUCCESS) {
6501 /* Now store the actual data in the child DB. Note that we're
6502 * storing the user data in the keys field, so there are strict
6503 * size limits on dupdata. The actual data fields of the child
6504 * DB are all zero size.
6512 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6513 if (flags & MDB_CURRENT) {
6514 xflags = MDB_CURRENT|MDB_NOSPILL;
6516 mdb_xcursor_init1(mc, leaf);
6517 xflags = (flags & MDB_NODUPDATA) ?
6518 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6520 /* converted, write the original data first */
6522 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6526 /* Adjust other cursors pointing to mp */
6528 unsigned i = mc->mc_top;
6529 MDB_page *mp = mc->mc_pg[i];
6531 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6532 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6533 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6534 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6535 mdb_xcursor_init1(m2, leaf);
6539 /* we've done our job */
6542 ecount = mc->mc_xcursor->mx_db.md_entries;
6543 if (flags & MDB_APPENDDUP)
6544 xflags |= MDB_APPEND;
6545 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6546 if (flags & F_SUBDATA) {
6547 void *db = NODEDATA(leaf);
6548 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6550 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6552 /* Increment count unless we just replaced an existing item. */
6554 mc->mc_db->md_entries++;
6556 /* Invalidate txn if we created an empty sub-DB */
6559 /* If we succeeded and the key didn't exist before,
6560 * make sure the cursor is marked valid.
6562 mc->mc_flags |= C_INITIALIZED;
6564 if (flags & MDB_MULTIPLE) {
6567 /* let caller know how many succeeded, if any */
6568 data[1].mv_size = mcount;
6569 if (mcount < dcount) {
6570 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6571 insert_key = insert_data = 0;
6578 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6581 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6586 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6592 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6593 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6595 if (!(mc->mc_flags & C_INITIALIZED))
6598 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6599 return MDB_NOTFOUND;
6601 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6604 rc = mdb_cursor_touch(mc);
6608 mp = mc->mc_pg[mc->mc_top];
6611 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6613 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6614 if (flags & MDB_NODUPDATA) {
6615 /* mdb_cursor_del0() will subtract the final entry */
6616 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6618 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6619 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6621 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6624 /* If sub-DB still has entries, we're done */
6625 if (mc->mc_xcursor->mx_db.md_entries) {
6626 if (leaf->mn_flags & F_SUBDATA) {
6627 /* update subDB info */
6628 void *db = NODEDATA(leaf);
6629 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6632 /* shrink fake page */
6633 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6634 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6635 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6636 /* fix other sub-DB cursors pointed at this fake page */
6637 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6638 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6639 if (m2->mc_pg[mc->mc_top] == mp &&
6640 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6641 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6644 mc->mc_db->md_entries--;
6645 mc->mc_flags |= C_DEL;
6648 /* otherwise fall thru and delete the sub-DB */
6651 if (leaf->mn_flags & F_SUBDATA) {
6652 /* add all the child DB's pages to the free list */
6653 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6659 /* add overflow pages to free list */
6660 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6664 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6665 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6666 (rc = mdb_ovpage_free(mc, omp)))
6671 return mdb_cursor_del0(mc);
6674 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6678 /** Allocate and initialize new pages for a database.
6679 * @param[in] mc a cursor on the database being added to.
6680 * @param[in] flags flags defining what type of page is being allocated.
6681 * @param[in] num the number of pages to allocate. This is usually 1,
6682 * unless allocating overflow pages for a large record.
6683 * @param[out] mp Address of a page, or NULL on failure.
6684 * @return 0 on success, non-zero on failure.
6687 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6692 if ((rc = mdb_page_alloc(mc, num, &np)))
6694 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6695 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6696 np->mp_flags = flags | P_DIRTY;
6697 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6698 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6701 mc->mc_db->md_branch_pages++;
6702 else if (IS_LEAF(np))
6703 mc->mc_db->md_leaf_pages++;
6704 else if (IS_OVERFLOW(np)) {
6705 mc->mc_db->md_overflow_pages += num;
6713 /** Calculate the size of a leaf node.
6714 * The size depends on the environment's page size; if a data item
6715 * is too large it will be put onto an overflow page and the node
6716 * size will only include the key and not the data. Sizes are always
6717 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6718 * of the #MDB_node headers.
6719 * @param[in] env The environment handle.
6720 * @param[in] key The key for the node.
6721 * @param[in] data The data for the node.
6722 * @return The number of bytes needed to store the node.
6725 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6729 sz = LEAFSIZE(key, data);
6730 if (sz > env->me_nodemax) {
6731 /* put on overflow page */
6732 sz -= data->mv_size - sizeof(pgno_t);
6735 return EVEN(sz + sizeof(indx_t));
6738 /** Calculate the size of a branch node.
6739 * The size should depend on the environment's page size but since
6740 * we currently don't support spilling large keys onto overflow
6741 * pages, it's simply the size of the #MDB_node header plus the
6742 * size of the key. Sizes are always rounded up to an even number
6743 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6744 * @param[in] env The environment handle.
6745 * @param[in] key The key for the node.
6746 * @return The number of bytes needed to store the node.
6749 mdb_branch_size(MDB_env *env, MDB_val *key)
6754 if (sz > env->me_nodemax) {
6755 /* put on overflow page */
6756 /* not implemented */
6757 /* sz -= key->size - sizeof(pgno_t); */
6760 return sz + sizeof(indx_t);
6763 /** Add a node to the page pointed to by the cursor.
6764 * @param[in] mc The cursor for this operation.
6765 * @param[in] indx The index on the page where the new node should be added.
6766 * @param[in] key The key for the new node.
6767 * @param[in] data The data for the new node, if any.
6768 * @param[in] pgno The page number, if adding a branch node.
6769 * @param[in] flags Flags for the node.
6770 * @return 0 on success, non-zero on failure. Possible errors are:
6772 * <li>ENOMEM - failed to allocate overflow pages for the node.
6773 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6774 * should never happen since all callers already calculate the
6775 * page's free space before calling this function.
6779 mdb_node_add(MDB_cursor *mc, indx_t indx,
6780 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6783 size_t node_size = NODESIZE;
6787 MDB_page *mp = mc->mc_pg[mc->mc_top];
6788 MDB_page *ofp = NULL; /* overflow page */
6791 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6793 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6794 IS_LEAF(mp) ? "leaf" : "branch",
6795 IS_SUBP(mp) ? "sub-" : "",
6796 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6797 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6800 /* Move higher keys up one slot. */
6801 int ksize = mc->mc_db->md_pad, dif;
6802 char *ptr = LEAF2KEY(mp, indx, ksize);
6803 dif = NUMKEYS(mp) - indx;
6805 memmove(ptr+ksize, ptr, dif*ksize);
6806 /* insert new key */
6807 memcpy(ptr, key->mv_data, ksize);
6809 /* Just using these for counting */
6810 mp->mp_lower += sizeof(indx_t);
6811 mp->mp_upper -= ksize - sizeof(indx_t);
6815 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6817 node_size += key->mv_size;
6819 mdb_cassert(mc, data);
6820 if (F_ISSET(flags, F_BIGDATA)) {
6821 /* Data already on overflow page. */
6822 node_size += sizeof(pgno_t);
6823 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6824 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6826 /* Put data on overflow page. */
6827 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6828 data->mv_size, node_size+data->mv_size));
6829 node_size = EVEN(node_size + sizeof(pgno_t));
6830 if ((ssize_t)node_size > room)
6832 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6834 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6838 node_size += data->mv_size;
6841 node_size = EVEN(node_size);
6842 if ((ssize_t)node_size > room)
6846 /* Move higher pointers up one slot. */
6847 for (i = NUMKEYS(mp); i > indx; i--)
6848 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6850 /* Adjust free space offsets. */
6851 ofs = mp->mp_upper - node_size;
6852 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6853 mp->mp_ptrs[indx] = ofs;
6855 mp->mp_lower += sizeof(indx_t);
6857 /* Write the node data. */
6858 node = NODEPTR(mp, indx);
6859 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6860 node->mn_flags = flags;
6862 SETDSZ(node,data->mv_size);
6867 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6870 mdb_cassert(mc, key);
6872 if (F_ISSET(flags, F_BIGDATA))
6873 memcpy(node->mn_data + key->mv_size, data->mv_data,
6875 else if (F_ISSET(flags, MDB_RESERVE))
6876 data->mv_data = node->mn_data + key->mv_size;
6878 memcpy(node->mn_data + key->mv_size, data->mv_data,
6881 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6883 if (F_ISSET(flags, MDB_RESERVE))
6884 data->mv_data = METADATA(ofp);
6886 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6893 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6894 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6895 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6896 DPRINTF(("node size = %"Z"u", node_size));
6897 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6898 return MDB_PAGE_FULL;
6901 /** Delete the specified node from a page.
6902 * @param[in] mc Cursor pointing to the node to delete.
6903 * @param[in] ksize The size of a node. Only used if the page is
6904 * part of a #MDB_DUPFIXED database.
6907 mdb_node_del(MDB_cursor *mc, int ksize)
6909 MDB_page *mp = mc->mc_pg[mc->mc_top];
6910 indx_t indx = mc->mc_ki[mc->mc_top];
6912 indx_t i, j, numkeys, ptr;
6916 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6917 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6918 numkeys = NUMKEYS(mp);
6919 mdb_cassert(mc, indx < numkeys);
6922 int x = numkeys - 1 - indx;
6923 base = LEAF2KEY(mp, indx, ksize);
6925 memmove(base, base + ksize, x * ksize);
6926 mp->mp_lower -= sizeof(indx_t);
6927 mp->mp_upper += ksize - sizeof(indx_t);
6931 node = NODEPTR(mp, indx);
6932 sz = NODESIZE + node->mn_ksize;
6934 if (F_ISSET(node->mn_flags, F_BIGDATA))
6935 sz += sizeof(pgno_t);
6937 sz += NODEDSZ(node);
6941 ptr = mp->mp_ptrs[indx];
6942 for (i = j = 0; i < numkeys; i++) {
6944 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6945 if (mp->mp_ptrs[i] < ptr)
6946 mp->mp_ptrs[j] += sz;
6951 base = (char *)mp + mp->mp_upper + PAGEBASE;
6952 memmove(base + sz, base, ptr - mp->mp_upper);
6954 mp->mp_lower -= sizeof(indx_t);
6958 /** Compact the main page after deleting a node on a subpage.
6959 * @param[in] mp The main page to operate on.
6960 * @param[in] indx The index of the subpage on the main page.
6963 mdb_node_shrink(MDB_page *mp, indx_t indx)
6969 indx_t i, numkeys, ptr;
6971 node = NODEPTR(mp, indx);
6972 sp = (MDB_page *)NODEDATA(node);
6973 delta = SIZELEFT(sp);
6974 xp = (MDB_page *)((char *)sp + delta);
6976 /* shift subpage upward */
6978 nsize = NUMKEYS(sp) * sp->mp_pad;
6980 return; /* do not make the node uneven-sized */
6981 memmove(METADATA(xp), METADATA(sp), nsize);
6984 numkeys = NUMKEYS(sp);
6985 for (i=numkeys-1; i>=0; i--)
6986 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6988 xp->mp_upper = sp->mp_lower;
6989 xp->mp_lower = sp->mp_lower;
6990 xp->mp_flags = sp->mp_flags;
6991 xp->mp_pad = sp->mp_pad;
6992 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6994 nsize = NODEDSZ(node) - delta;
6995 SETDSZ(node, nsize);
6997 /* shift lower nodes upward */
6998 ptr = mp->mp_ptrs[indx];
6999 numkeys = NUMKEYS(mp);
7000 for (i = 0; i < numkeys; i++) {
7001 if (mp->mp_ptrs[i] <= ptr)
7002 mp->mp_ptrs[i] += delta;
7005 base = (char *)mp + mp->mp_upper + PAGEBASE;
7006 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7007 mp->mp_upper += delta;
7010 /** Initial setup of a sorted-dups cursor.
7011 * Sorted duplicates are implemented as a sub-database for the given key.
7012 * The duplicate data items are actually keys of the sub-database.
7013 * Operations on the duplicate data items are performed using a sub-cursor
7014 * initialized when the sub-database is first accessed. This function does
7015 * the preliminary setup of the sub-cursor, filling in the fields that
7016 * depend only on the parent DB.
7017 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7020 mdb_xcursor_init0(MDB_cursor *mc)
7022 MDB_xcursor *mx = mc->mc_xcursor;
7024 mx->mx_cursor.mc_xcursor = NULL;
7025 mx->mx_cursor.mc_txn = mc->mc_txn;
7026 mx->mx_cursor.mc_db = &mx->mx_db;
7027 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7028 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7029 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7030 mx->mx_cursor.mc_snum = 0;
7031 mx->mx_cursor.mc_top = 0;
7032 mx->mx_cursor.mc_flags = C_SUB;
7033 mx->mx_dbx.md_name.mv_size = 0;
7034 mx->mx_dbx.md_name.mv_data = NULL;
7035 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7036 mx->mx_dbx.md_dcmp = NULL;
7037 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7040 /** Final setup of a sorted-dups cursor.
7041 * Sets up the fields that depend on the data from the main cursor.
7042 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7043 * @param[in] node The data containing the #MDB_db record for the
7044 * sorted-dup database.
7047 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7049 MDB_xcursor *mx = mc->mc_xcursor;
7051 if (node->mn_flags & F_SUBDATA) {
7052 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7053 mx->mx_cursor.mc_pg[0] = 0;
7054 mx->mx_cursor.mc_snum = 0;
7055 mx->mx_cursor.mc_top = 0;
7056 mx->mx_cursor.mc_flags = C_SUB;
7058 MDB_page *fp = NODEDATA(node);
7059 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7060 mx->mx_db.md_flags = 0;
7061 mx->mx_db.md_depth = 1;
7062 mx->mx_db.md_branch_pages = 0;
7063 mx->mx_db.md_leaf_pages = 1;
7064 mx->mx_db.md_overflow_pages = 0;
7065 mx->mx_db.md_entries = NUMKEYS(fp);
7066 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7067 mx->mx_cursor.mc_snum = 1;
7068 mx->mx_cursor.mc_top = 0;
7069 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7070 mx->mx_cursor.mc_pg[0] = fp;
7071 mx->mx_cursor.mc_ki[0] = 0;
7072 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7073 mx->mx_db.md_flags = MDB_DUPFIXED;
7074 mx->mx_db.md_pad = fp->mp_pad;
7075 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7076 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7079 #if UINT_MAX < SIZE_MAX
7080 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t)) {
7081 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
7082 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7085 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7086 mx->mx_db.md_root));
7087 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7088 #if UINT_MAX < SIZE_MAX
7089 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7090 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7094 /** Initialize a cursor for a given transaction and database. */
7096 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7099 mc->mc_backup = NULL;
7102 mc->mc_db = &txn->mt_dbs[dbi];
7103 mc->mc_dbx = &txn->mt_dbxs[dbi];
7104 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7110 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7111 mdb_tassert(txn, mx != NULL);
7112 mc->mc_xcursor = mx;
7113 mdb_xcursor_init0(mc);
7115 mc->mc_xcursor = NULL;
7117 if (*mc->mc_dbflag & DB_STALE) {
7118 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7123 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7126 size_t size = sizeof(MDB_cursor);
7128 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7131 if (txn->mt_flags & MDB_TXN_ERROR)
7134 /* Allow read access to the freelist */
7135 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7138 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7139 size += sizeof(MDB_xcursor);
7141 if ((mc = malloc(size)) != NULL) {
7142 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7143 if (txn->mt_cursors) {
7144 mc->mc_next = txn->mt_cursors[dbi];
7145 txn->mt_cursors[dbi] = mc;
7146 mc->mc_flags |= C_UNTRACK;
7158 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7160 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7163 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7166 if (txn->mt_flags & MDB_TXN_ERROR)
7169 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7173 /* Return the count of duplicate data items for the current key */
7175 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7179 if (mc == NULL || countp == NULL)
7182 if (mc->mc_xcursor == NULL)
7183 return MDB_INCOMPATIBLE;
7185 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7188 if (!(mc->mc_flags & C_INITIALIZED))
7191 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7192 return MDB_NOTFOUND;
7194 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7195 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7198 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7201 *countp = mc->mc_xcursor->mx_db.md_entries;
7207 mdb_cursor_close(MDB_cursor *mc)
7209 if (mc && !mc->mc_backup) {
7210 /* remove from txn, if tracked */
7211 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7212 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7213 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7215 *prev = mc->mc_next;
7222 mdb_cursor_txn(MDB_cursor *mc)
7224 if (!mc) return NULL;
7229 mdb_cursor_dbi(MDB_cursor *mc)
7234 /** Replace the key for a branch node with a new key.
7235 * @param[in] mc Cursor pointing to the node to operate on.
7236 * @param[in] key The new key to use.
7237 * @return 0 on success, non-zero on failure.
7240 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7246 int delta, ksize, oksize;
7247 indx_t ptr, i, numkeys, indx;
7250 indx = mc->mc_ki[mc->mc_top];
7251 mp = mc->mc_pg[mc->mc_top];
7252 node = NODEPTR(mp, indx);
7253 ptr = mp->mp_ptrs[indx];
7257 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7258 k2.mv_data = NODEKEY(node);
7259 k2.mv_size = node->mn_ksize;
7260 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7262 mdb_dkey(&k2, kbuf2),
7268 /* Sizes must be 2-byte aligned. */
7269 ksize = EVEN(key->mv_size);
7270 oksize = EVEN(node->mn_ksize);
7271 delta = ksize - oksize;
7273 /* Shift node contents if EVEN(key length) changed. */
7275 if (delta > 0 && SIZELEFT(mp) < delta) {
7277 /* not enough space left, do a delete and split */
7278 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7279 pgno = NODEPGNO(node);
7280 mdb_node_del(mc, 0);
7281 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7284 numkeys = NUMKEYS(mp);
7285 for (i = 0; i < numkeys; i++) {
7286 if (mp->mp_ptrs[i] <= ptr)
7287 mp->mp_ptrs[i] -= delta;
7290 base = (char *)mp + mp->mp_upper + PAGEBASE;
7291 len = ptr - mp->mp_upper + NODESIZE;
7292 memmove(base - delta, base, len);
7293 mp->mp_upper -= delta;
7295 node = NODEPTR(mp, indx);
7298 /* But even if no shift was needed, update ksize */
7299 if (node->mn_ksize != key->mv_size)
7300 node->mn_ksize = key->mv_size;
7303 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7309 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7311 /** Move a node from csrc to cdst.
7314 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7321 unsigned short flags;
7325 /* Mark src and dst as dirty. */
7326 if ((rc = mdb_page_touch(csrc)) ||
7327 (rc = mdb_page_touch(cdst)))
7330 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7331 key.mv_size = csrc->mc_db->md_pad;
7332 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7334 data.mv_data = NULL;
7338 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7339 mdb_cassert(csrc, !((size_t)srcnode & 1));
7340 srcpg = NODEPGNO(srcnode);
7341 flags = srcnode->mn_flags;
7342 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7343 unsigned int snum = csrc->mc_snum;
7345 /* must find the lowest key below src */
7346 rc = mdb_page_search_lowest(csrc);
7349 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7350 key.mv_size = csrc->mc_db->md_pad;
7351 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7353 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7354 key.mv_size = NODEKSZ(s2);
7355 key.mv_data = NODEKEY(s2);
7357 csrc->mc_snum = snum--;
7358 csrc->mc_top = snum;
7360 key.mv_size = NODEKSZ(srcnode);
7361 key.mv_data = NODEKEY(srcnode);
7363 data.mv_size = NODEDSZ(srcnode);
7364 data.mv_data = NODEDATA(srcnode);
7366 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7367 unsigned int snum = cdst->mc_snum;
7370 /* must find the lowest key below dst */
7371 mdb_cursor_copy(cdst, &mn);
7372 rc = mdb_page_search_lowest(&mn);
7375 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7376 bkey.mv_size = mn.mc_db->md_pad;
7377 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7379 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7380 bkey.mv_size = NODEKSZ(s2);
7381 bkey.mv_data = NODEKEY(s2);
7383 mn.mc_snum = snum--;
7386 rc = mdb_update_key(&mn, &bkey);
7391 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7392 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7393 csrc->mc_ki[csrc->mc_top],
7395 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7396 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7398 /* Add the node to the destination page.
7400 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7401 if (rc != MDB_SUCCESS)
7404 /* Delete the node from the source page.
7406 mdb_node_del(csrc, key.mv_size);
7409 /* Adjust other cursors pointing to mp */
7410 MDB_cursor *m2, *m3;
7411 MDB_dbi dbi = csrc->mc_dbi;
7412 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7414 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7415 if (csrc->mc_flags & C_SUB)
7416 m3 = &m2->mc_xcursor->mx_cursor;
7419 if (m3 == csrc) continue;
7420 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7421 csrc->mc_ki[csrc->mc_top]) {
7422 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7423 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7428 /* Update the parent separators.
7430 if (csrc->mc_ki[csrc->mc_top] == 0) {
7431 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7432 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7433 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7435 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7436 key.mv_size = NODEKSZ(srcnode);
7437 key.mv_data = NODEKEY(srcnode);
7439 DPRINTF(("update separator for source page %"Z"u to [%s]",
7440 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7441 mdb_cursor_copy(csrc, &mn);
7444 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7447 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7449 indx_t ix = csrc->mc_ki[csrc->mc_top];
7450 nullkey.mv_size = 0;
7451 csrc->mc_ki[csrc->mc_top] = 0;
7452 rc = mdb_update_key(csrc, &nullkey);
7453 csrc->mc_ki[csrc->mc_top] = ix;
7454 mdb_cassert(csrc, rc == MDB_SUCCESS);
7458 if (cdst->mc_ki[cdst->mc_top] == 0) {
7459 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7460 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7461 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7463 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7464 key.mv_size = NODEKSZ(srcnode);
7465 key.mv_data = NODEKEY(srcnode);
7467 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7468 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7469 mdb_cursor_copy(cdst, &mn);
7472 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7475 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7477 indx_t ix = cdst->mc_ki[cdst->mc_top];
7478 nullkey.mv_size = 0;
7479 cdst->mc_ki[cdst->mc_top] = 0;
7480 rc = mdb_update_key(cdst, &nullkey);
7481 cdst->mc_ki[cdst->mc_top] = ix;
7482 mdb_cassert(cdst, rc == MDB_SUCCESS);
7489 /** Merge one page into another.
7490 * The nodes from the page pointed to by \b csrc will
7491 * be copied to the page pointed to by \b cdst and then
7492 * the \b csrc page will be freed.
7493 * @param[in] csrc Cursor pointing to the source page.
7494 * @param[in] cdst Cursor pointing to the destination page.
7495 * @return 0 on success, non-zero on failure.
7498 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7500 MDB_page *psrc, *pdst;
7507 psrc = csrc->mc_pg[csrc->mc_top];
7508 pdst = cdst->mc_pg[cdst->mc_top];
7510 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7512 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7513 mdb_cassert(csrc, cdst->mc_snum > 1);
7515 /* Mark dst as dirty. */
7516 if ((rc = mdb_page_touch(cdst)))
7519 /* Move all nodes from src to dst.
7521 j = nkeys = NUMKEYS(pdst);
7522 if (IS_LEAF2(psrc)) {
7523 key.mv_size = csrc->mc_db->md_pad;
7524 key.mv_data = METADATA(psrc);
7525 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7526 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7527 if (rc != MDB_SUCCESS)
7529 key.mv_data = (char *)key.mv_data + key.mv_size;
7532 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7533 srcnode = NODEPTR(psrc, i);
7534 if (i == 0 && IS_BRANCH(psrc)) {
7537 mdb_cursor_copy(csrc, &mn);
7538 /* must find the lowest key below src */
7539 rc = mdb_page_search_lowest(&mn);
7542 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7543 key.mv_size = mn.mc_db->md_pad;
7544 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7546 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7547 key.mv_size = NODEKSZ(s2);
7548 key.mv_data = NODEKEY(s2);
7551 key.mv_size = srcnode->mn_ksize;
7552 key.mv_data = NODEKEY(srcnode);
7555 data.mv_size = NODEDSZ(srcnode);
7556 data.mv_data = NODEDATA(srcnode);
7557 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7558 if (rc != MDB_SUCCESS)
7563 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7564 pdst->mp_pgno, NUMKEYS(pdst),
7565 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7567 /* Unlink the src page from parent and add to free list.
7570 mdb_node_del(csrc, 0);
7571 if (csrc->mc_ki[csrc->mc_top] == 0) {
7573 rc = mdb_update_key(csrc, &key);
7581 psrc = csrc->mc_pg[csrc->mc_top];
7582 /* If not operating on FreeDB, allow this page to be reused
7583 * in this txn. Otherwise just add to free list.
7585 rc = mdb_page_loose(csrc, psrc);
7589 csrc->mc_db->md_leaf_pages--;
7591 csrc->mc_db->md_branch_pages--;
7593 /* Adjust other cursors pointing to mp */
7594 MDB_cursor *m2, *m3;
7595 MDB_dbi dbi = csrc->mc_dbi;
7597 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7598 if (csrc->mc_flags & C_SUB)
7599 m3 = &m2->mc_xcursor->mx_cursor;
7602 if (m3 == csrc) continue;
7603 if (m3->mc_snum < csrc->mc_snum) continue;
7604 if (m3->mc_pg[csrc->mc_top] == psrc) {
7605 m3->mc_pg[csrc->mc_top] = pdst;
7606 m3->mc_ki[csrc->mc_top] += nkeys;
7611 unsigned int snum = cdst->mc_snum;
7612 uint16_t depth = cdst->mc_db->md_depth;
7613 mdb_cursor_pop(cdst);
7614 rc = mdb_rebalance(cdst);
7615 /* Did the tree shrink? */
7616 if (depth > cdst->mc_db->md_depth)
7618 cdst->mc_snum = snum;
7619 cdst->mc_top = snum-1;
7624 /** Copy the contents of a cursor.
7625 * @param[in] csrc The cursor to copy from.
7626 * @param[out] cdst The cursor to copy to.
7629 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7633 cdst->mc_txn = csrc->mc_txn;
7634 cdst->mc_dbi = csrc->mc_dbi;
7635 cdst->mc_db = csrc->mc_db;
7636 cdst->mc_dbx = csrc->mc_dbx;
7637 cdst->mc_snum = csrc->mc_snum;
7638 cdst->mc_top = csrc->mc_top;
7639 cdst->mc_flags = csrc->mc_flags;
7641 for (i=0; i<csrc->mc_snum; i++) {
7642 cdst->mc_pg[i] = csrc->mc_pg[i];
7643 cdst->mc_ki[i] = csrc->mc_ki[i];
7647 /** Rebalance the tree after a delete operation.
7648 * @param[in] mc Cursor pointing to the page where rebalancing
7650 * @return 0 on success, non-zero on failure.
7653 mdb_rebalance(MDB_cursor *mc)
7657 unsigned int ptop, minkeys;
7661 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7662 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7663 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7664 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7665 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7667 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7668 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7669 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7670 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7674 if (mc->mc_snum < 2) {
7675 MDB_page *mp = mc->mc_pg[0];
7677 DPUTS("Can't rebalance a subpage, ignoring");
7680 if (NUMKEYS(mp) == 0) {
7681 DPUTS("tree is completely empty");
7682 mc->mc_db->md_root = P_INVALID;
7683 mc->mc_db->md_depth = 0;
7684 mc->mc_db->md_leaf_pages = 0;
7685 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7688 /* Adjust cursors pointing to mp */
7691 mc->mc_flags &= ~C_INITIALIZED;
7693 MDB_cursor *m2, *m3;
7694 MDB_dbi dbi = mc->mc_dbi;
7696 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7697 if (mc->mc_flags & C_SUB)
7698 m3 = &m2->mc_xcursor->mx_cursor;
7701 if (m3->mc_snum < mc->mc_snum) continue;
7702 if (m3->mc_pg[0] == mp) {
7705 m3->mc_flags &= ~C_INITIALIZED;
7709 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7711 DPUTS("collapsing root page!");
7712 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7715 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7716 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7719 mc->mc_db->md_depth--;
7720 mc->mc_db->md_branch_pages--;
7721 mc->mc_ki[0] = mc->mc_ki[1];
7722 for (i = 1; i<mc->mc_db->md_depth; i++) {
7723 mc->mc_pg[i] = mc->mc_pg[i+1];
7724 mc->mc_ki[i] = mc->mc_ki[i+1];
7727 /* Adjust other cursors pointing to mp */
7728 MDB_cursor *m2, *m3;
7729 MDB_dbi dbi = mc->mc_dbi;
7731 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7732 if (mc->mc_flags & C_SUB)
7733 m3 = &m2->mc_xcursor->mx_cursor;
7736 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7737 if (m3->mc_pg[0] == mp) {
7738 for (i=0; i<m3->mc_snum; i++) {
7739 m3->mc_pg[i] = m3->mc_pg[i+1];
7740 m3->mc_ki[i] = m3->mc_ki[i+1];
7748 DPUTS("root page doesn't need rebalancing");
7752 /* The parent (branch page) must have at least 2 pointers,
7753 * otherwise the tree is invalid.
7755 ptop = mc->mc_top-1;
7756 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7758 /* Leaf page fill factor is below the threshold.
7759 * Try to move keys from left or right neighbor, or
7760 * merge with a neighbor page.
7765 mdb_cursor_copy(mc, &mn);
7766 mn.mc_xcursor = NULL;
7768 oldki = mc->mc_ki[mc->mc_top];
7769 if (mc->mc_ki[ptop] == 0) {
7770 /* We're the leftmost leaf in our parent.
7772 DPUTS("reading right neighbor");
7774 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7775 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7778 mn.mc_ki[mn.mc_top] = 0;
7779 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7781 /* There is at least one neighbor to the left.
7783 DPUTS("reading left neighbor");
7785 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7786 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7789 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7790 mc->mc_ki[mc->mc_top] = 0;
7793 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7794 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7795 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7797 /* If the neighbor page is above threshold and has enough keys,
7798 * move one key from it. Otherwise we should try to merge them.
7799 * (A branch page must never have less than 2 keys.)
7801 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7802 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7803 rc = mdb_node_move(&mn, mc);
7804 if (mc->mc_ki[ptop]) {
7808 if (mc->mc_ki[ptop] == 0) {
7809 rc = mdb_page_merge(&mn, mc);
7812 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7813 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7814 /* We want mdb_rebalance to find mn when doing fixups */
7815 if (mc->mc_flags & C_SUB) {
7816 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7817 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
7818 dummy.mc_xcursor = (MDB_xcursor *)&mn;
7820 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7821 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
7823 rc = mdb_page_merge(mc, &mn);
7824 if (mc->mc_flags & C_SUB)
7825 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
7827 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
7828 mdb_cursor_copy(&mn, mc);
7830 mc->mc_flags &= ~C_EOF;
7832 mc->mc_ki[mc->mc_top] = oldki;
7836 /** Complete a delete operation started by #mdb_cursor_del(). */
7838 mdb_cursor_del0(MDB_cursor *mc)
7845 ki = mc->mc_ki[mc->mc_top];
7846 mdb_node_del(mc, mc->mc_db->md_pad);
7847 mc->mc_db->md_entries--;
7848 rc = mdb_rebalance(mc);
7850 if (rc == MDB_SUCCESS) {
7851 MDB_cursor *m2, *m3;
7852 MDB_dbi dbi = mc->mc_dbi;
7854 /* DB is totally empty now, just bail out.
7855 * Other cursors adjustments were already done
7856 * by mdb_rebalance and aren't needed here.
7861 mp = mc->mc_pg[mc->mc_top];
7862 nkeys = NUMKEYS(mp);
7864 /* if mc points past last node in page, find next sibling */
7865 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7866 rc = mdb_cursor_sibling(mc, 1);
7867 if (rc == MDB_NOTFOUND) {
7868 mc->mc_flags |= C_EOF;
7873 /* Adjust other cursors pointing to mp */
7874 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7875 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7876 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7878 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7880 if (m3->mc_pg[mc->mc_top] == mp) {
7881 if (m3->mc_ki[mc->mc_top] >= ki) {
7882 m3->mc_flags |= C_DEL;
7883 if (m3->mc_ki[mc->mc_top] > ki)
7884 m3->mc_ki[mc->mc_top]--;
7885 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7886 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7888 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7889 rc = mdb_cursor_sibling(m3, 1);
7890 if (rc == MDB_NOTFOUND) {
7891 m3->mc_flags |= C_EOF;
7897 mc->mc_flags |= C_DEL;
7901 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7906 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7907 MDB_val *key, MDB_val *data)
7909 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7912 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7913 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7915 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7916 /* must ignore any data */
7920 return mdb_del0(txn, dbi, key, data, 0);
7924 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7925 MDB_val *key, MDB_val *data, unsigned flags)
7930 MDB_val rdata, *xdata;
7934 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7936 mdb_cursor_init(&mc, txn, dbi, &mx);
7945 flags |= MDB_NODUPDATA;
7947 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7949 /* let mdb_page_split know about this cursor if needed:
7950 * delete will trigger a rebalance; if it needs to move
7951 * a node from one page to another, it will have to
7952 * update the parent's separator key(s). If the new sepkey
7953 * is larger than the current one, the parent page may
7954 * run out of space, triggering a split. We need this
7955 * cursor to be consistent until the end of the rebalance.
7957 mc.mc_flags |= C_UNTRACK;
7958 mc.mc_next = txn->mt_cursors[dbi];
7959 txn->mt_cursors[dbi] = &mc;
7960 rc = mdb_cursor_del(&mc, flags);
7961 txn->mt_cursors[dbi] = mc.mc_next;
7966 /** Split a page and insert a new node.
7967 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7968 * The cursor will be updated to point to the actual page and index where
7969 * the node got inserted after the split.
7970 * @param[in] newkey The key for the newly inserted node.
7971 * @param[in] newdata The data for the newly inserted node.
7972 * @param[in] newpgno The page number, if the new node is a branch node.
7973 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7974 * @return 0 on success, non-zero on failure.
7977 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7978 unsigned int nflags)
7981 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7984 int i, j, split_indx, nkeys, pmax;
7985 MDB_env *env = mc->mc_txn->mt_env;
7987 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7988 MDB_page *copy = NULL;
7989 MDB_page *mp, *rp, *pp;
7994 mp = mc->mc_pg[mc->mc_top];
7995 newindx = mc->mc_ki[mc->mc_top];
7996 nkeys = NUMKEYS(mp);
7998 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7999 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8000 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8002 /* Create a right sibling. */
8003 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8005 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8007 if (mc->mc_snum < 2) {
8008 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8010 /* shift current top to make room for new parent */
8011 mc->mc_pg[1] = mc->mc_pg[0];
8012 mc->mc_ki[1] = mc->mc_ki[0];
8015 mc->mc_db->md_root = pp->mp_pgno;
8016 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8017 mc->mc_db->md_depth++;
8020 /* Add left (implicit) pointer. */
8021 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8022 /* undo the pre-push */
8023 mc->mc_pg[0] = mc->mc_pg[1];
8024 mc->mc_ki[0] = mc->mc_ki[1];
8025 mc->mc_db->md_root = mp->mp_pgno;
8026 mc->mc_db->md_depth--;
8033 ptop = mc->mc_top-1;
8034 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8037 mc->mc_flags |= C_SPLITTING;
8038 mdb_cursor_copy(mc, &mn);
8039 mn.mc_pg[mn.mc_top] = rp;
8040 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8042 if (nflags & MDB_APPEND) {
8043 mn.mc_ki[mn.mc_top] = 0;
8045 split_indx = newindx;
8049 split_indx = (nkeys+1) / 2;
8054 unsigned int lsize, rsize, ksize;
8055 /* Move half of the keys to the right sibling */
8056 x = mc->mc_ki[mc->mc_top] - split_indx;
8057 ksize = mc->mc_db->md_pad;
8058 split = LEAF2KEY(mp, split_indx, ksize);
8059 rsize = (nkeys - split_indx) * ksize;
8060 lsize = (nkeys - split_indx) * sizeof(indx_t);
8061 mp->mp_lower -= lsize;
8062 rp->mp_lower += lsize;
8063 mp->mp_upper += rsize - lsize;
8064 rp->mp_upper -= rsize - lsize;
8065 sepkey.mv_size = ksize;
8066 if (newindx == split_indx) {
8067 sepkey.mv_data = newkey->mv_data;
8069 sepkey.mv_data = split;
8072 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8073 memcpy(rp->mp_ptrs, split, rsize);
8074 sepkey.mv_data = rp->mp_ptrs;
8075 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8076 memcpy(ins, newkey->mv_data, ksize);
8077 mp->mp_lower += sizeof(indx_t);
8078 mp->mp_upper -= ksize - sizeof(indx_t);
8081 memcpy(rp->mp_ptrs, split, x * ksize);
8082 ins = LEAF2KEY(rp, x, ksize);
8083 memcpy(ins, newkey->mv_data, ksize);
8084 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8085 rp->mp_lower += sizeof(indx_t);
8086 rp->mp_upper -= ksize - sizeof(indx_t);
8087 mc->mc_ki[mc->mc_top] = x;
8088 mc->mc_pg[mc->mc_top] = rp;
8091 int psize, nsize, k;
8092 /* Maximum free space in an empty page */
8093 pmax = env->me_psize - PAGEHDRSZ;
8095 nsize = mdb_leaf_size(env, newkey, newdata);
8097 nsize = mdb_branch_size(env, newkey);
8098 nsize = EVEN(nsize);
8100 /* grab a page to hold a temporary copy */
8101 copy = mdb_page_malloc(mc->mc_txn, 1);
8106 copy->mp_pgno = mp->mp_pgno;
8107 copy->mp_flags = mp->mp_flags;
8108 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8109 copy->mp_upper = env->me_psize - PAGEBASE;
8111 /* prepare to insert */
8112 for (i=0, j=0; i<nkeys; i++) {
8114 copy->mp_ptrs[j++] = 0;
8116 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8119 /* When items are relatively large the split point needs
8120 * to be checked, because being off-by-one will make the
8121 * difference between success or failure in mdb_node_add.
8123 * It's also relevant if a page happens to be laid out
8124 * such that one half of its nodes are all "small" and
8125 * the other half of its nodes are "large." If the new
8126 * item is also "large" and falls on the half with
8127 * "large" nodes, it also may not fit.
8129 * As a final tweak, if the new item goes on the last
8130 * spot on the page (and thus, onto the new page), bias
8131 * the split so the new page is emptier than the old page.
8132 * This yields better packing during sequential inserts.
8134 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8135 /* Find split point */
8137 if (newindx <= split_indx || newindx >= nkeys) {
8139 k = newindx >= nkeys ? nkeys : split_indx+2;
8144 for (; i!=k; i+=j) {
8149 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8150 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8152 if (F_ISSET(node->mn_flags, F_BIGDATA))
8153 psize += sizeof(pgno_t);
8155 psize += NODEDSZ(node);
8157 psize = EVEN(psize);
8159 if (psize > pmax || i == k-j) {
8160 split_indx = i + (j<0);
8165 if (split_indx == newindx) {
8166 sepkey.mv_size = newkey->mv_size;
8167 sepkey.mv_data = newkey->mv_data;
8169 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8170 sepkey.mv_size = node->mn_ksize;
8171 sepkey.mv_data = NODEKEY(node);
8176 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8178 /* Copy separator key to the parent.
8180 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8184 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8189 if (mn.mc_snum == mc->mc_snum) {
8190 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8191 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8192 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8193 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8198 /* Right page might now have changed parent.
8199 * Check if left page also changed parent.
8201 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8202 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8203 for (i=0; i<ptop; i++) {
8204 mc->mc_pg[i] = mn.mc_pg[i];
8205 mc->mc_ki[i] = mn.mc_ki[i];
8207 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8208 if (mn.mc_ki[ptop]) {
8209 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8211 /* find right page's left sibling */
8212 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8213 mdb_cursor_sibling(mc, 0);
8218 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8221 mc->mc_flags ^= C_SPLITTING;
8222 if (rc != MDB_SUCCESS) {
8225 if (nflags & MDB_APPEND) {
8226 mc->mc_pg[mc->mc_top] = rp;
8227 mc->mc_ki[mc->mc_top] = 0;
8228 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8231 for (i=0; i<mc->mc_top; i++)
8232 mc->mc_ki[i] = mn.mc_ki[i];
8233 } else if (!IS_LEAF2(mp)) {
8235 mc->mc_pg[mc->mc_top] = rp;
8240 rkey.mv_data = newkey->mv_data;
8241 rkey.mv_size = newkey->mv_size;
8247 /* Update index for the new key. */
8248 mc->mc_ki[mc->mc_top] = j;
8250 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8251 rkey.mv_data = NODEKEY(node);
8252 rkey.mv_size = node->mn_ksize;
8254 xdata.mv_data = NODEDATA(node);
8255 xdata.mv_size = NODEDSZ(node);
8258 pgno = NODEPGNO(node);
8259 flags = node->mn_flags;
8262 if (!IS_LEAF(mp) && j == 0) {
8263 /* First branch index doesn't need key data. */
8267 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8273 mc->mc_pg[mc->mc_top] = copy;
8278 } while (i != split_indx);
8280 nkeys = NUMKEYS(copy);
8281 for (i=0; i<nkeys; i++)
8282 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8283 mp->mp_lower = copy->mp_lower;
8284 mp->mp_upper = copy->mp_upper;
8285 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8286 env->me_psize - copy->mp_upper - PAGEBASE);
8288 /* reset back to original page */
8289 if (newindx < split_indx) {
8290 mc->mc_pg[mc->mc_top] = mp;
8291 if (nflags & MDB_RESERVE) {
8292 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8293 if (!(node->mn_flags & F_BIGDATA))
8294 newdata->mv_data = NODEDATA(node);
8297 mc->mc_pg[mc->mc_top] = rp;
8299 /* Make sure mc_ki is still valid.
8301 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8302 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8303 for (i=0; i<=ptop; i++) {
8304 mc->mc_pg[i] = mn.mc_pg[i];
8305 mc->mc_ki[i] = mn.mc_ki[i];
8312 /* Adjust other cursors pointing to mp */
8313 MDB_cursor *m2, *m3;
8314 MDB_dbi dbi = mc->mc_dbi;
8315 int fixup = NUMKEYS(mp);
8317 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8318 if (mc->mc_flags & C_SUB)
8319 m3 = &m2->mc_xcursor->mx_cursor;
8324 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8326 if (m3->mc_flags & C_SPLITTING)
8331 for (k=m3->mc_top; k>=0; k--) {
8332 m3->mc_ki[k+1] = m3->mc_ki[k];
8333 m3->mc_pg[k+1] = m3->mc_pg[k];
8335 if (m3->mc_ki[0] >= split_indx) {
8340 m3->mc_pg[0] = mc->mc_pg[0];
8344 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8345 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8346 m3->mc_ki[mc->mc_top]++;
8347 if (m3->mc_ki[mc->mc_top] >= fixup) {
8348 m3->mc_pg[mc->mc_top] = rp;
8349 m3->mc_ki[mc->mc_top] -= fixup;
8350 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8352 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8353 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8358 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8361 if (copy) /* tmp page */
8362 mdb_page_free(env, copy);
8364 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8369 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8370 MDB_val *key, MDB_val *data, unsigned int flags)
8375 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8378 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8381 mdb_cursor_init(&mc, txn, dbi, &mx);
8382 return mdb_cursor_put(&mc, key, data, flags);
8386 #define MDB_WBUF (1024*1024)
8389 /** State needed for a compacting copy. */
8390 typedef struct mdb_copy {
8391 pthread_mutex_t mc_mutex;
8392 pthread_cond_t mc_cond;
8399 pgno_t mc_next_pgno;
8402 volatile int mc_new;
8407 /** Dedicated writer thread for compacting copy. */
8408 static THREAD_RET ESECT
8409 mdb_env_copythr(void *arg)
8413 int toggle = 0, wsize, rc;
8416 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8419 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8422 pthread_mutex_lock(&my->mc_mutex);
8424 pthread_cond_signal(&my->mc_cond);
8427 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8428 if (my->mc_new < 0) {
8433 wsize = my->mc_wlen[toggle];
8434 ptr = my->mc_wbuf[toggle];
8437 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8441 } else if (len > 0) {
8455 /* If there's an overflow page tail, write it too */
8456 if (my->mc_olen[toggle]) {
8457 wsize = my->mc_olen[toggle];
8458 ptr = my->mc_over[toggle];
8459 my->mc_olen[toggle] = 0;
8462 my->mc_wlen[toggle] = 0;
8464 pthread_cond_signal(&my->mc_cond);
8466 pthread_cond_signal(&my->mc_cond);
8467 pthread_mutex_unlock(&my->mc_mutex);
8468 return (THREAD_RET)0;
8472 /** Tell the writer thread there's a buffer ready to write */
8474 mdb_env_cthr_toggle(mdb_copy *my, int st)
8476 int toggle = my->mc_toggle ^ 1;
8477 pthread_mutex_lock(&my->mc_mutex);
8478 if (my->mc_status) {
8479 pthread_mutex_unlock(&my->mc_mutex);
8480 return my->mc_status;
8482 while (my->mc_new == 1)
8483 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8485 my->mc_toggle = toggle;
8486 pthread_cond_signal(&my->mc_cond);
8487 pthread_mutex_unlock(&my->mc_mutex);
8491 /** Depth-first tree traversal for compacting copy. */
8493 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8496 MDB_txn *txn = my->mc_txn;
8498 MDB_page *mo, *mp, *leaf;
8503 /* Empty DB, nothing to do */
8504 if (*pg == P_INVALID)
8511 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8514 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8518 /* Make cursor pages writable */
8519 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8523 for (i=0; i<mc.mc_top; i++) {
8524 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8525 mc.mc_pg[i] = (MDB_page *)ptr;
8526 ptr += my->mc_env->me_psize;
8529 /* This is writable space for a leaf page. Usually not needed. */
8530 leaf = (MDB_page *)ptr;
8532 toggle = my->mc_toggle;
8533 while (mc.mc_snum > 0) {
8535 mp = mc.mc_pg[mc.mc_top];
8539 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8540 for (i=0; i<n; i++) {
8541 ni = NODEPTR(mp, i);
8542 if (ni->mn_flags & F_BIGDATA) {
8546 /* Need writable leaf */
8548 mc.mc_pg[mc.mc_top] = leaf;
8549 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8551 ni = NODEPTR(mp, i);
8554 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8555 rc = mdb_page_get(txn, pg, &omp, NULL);
8558 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8559 rc = mdb_env_cthr_toggle(my, 1);
8562 toggle = my->mc_toggle;
8564 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8565 memcpy(mo, omp, my->mc_env->me_psize);
8566 mo->mp_pgno = my->mc_next_pgno;
8567 my->mc_next_pgno += omp->mp_pages;
8568 my->mc_wlen[toggle] += my->mc_env->me_psize;
8569 if (omp->mp_pages > 1) {
8570 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8571 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8572 rc = mdb_env_cthr_toggle(my, 1);
8575 toggle = my->mc_toggle;
8577 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8578 } else if (ni->mn_flags & F_SUBDATA) {
8581 /* Need writable leaf */
8583 mc.mc_pg[mc.mc_top] = leaf;
8584 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8586 ni = NODEPTR(mp, i);
8589 memcpy(&db, NODEDATA(ni), sizeof(db));
8590 my->mc_toggle = toggle;
8591 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8594 toggle = my->mc_toggle;
8595 memcpy(NODEDATA(ni), &db, sizeof(db));
8600 mc.mc_ki[mc.mc_top]++;
8601 if (mc.mc_ki[mc.mc_top] < n) {
8604 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8606 rc = mdb_page_get(txn, pg, &mp, NULL);
8611 mc.mc_ki[mc.mc_top] = 0;
8612 if (IS_BRANCH(mp)) {
8613 /* Whenever we advance to a sibling branch page,
8614 * we must proceed all the way down to its first leaf.
8616 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8619 mc.mc_pg[mc.mc_top] = mp;
8623 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8624 rc = mdb_env_cthr_toggle(my, 1);
8627 toggle = my->mc_toggle;
8629 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8630 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8631 mo->mp_pgno = my->mc_next_pgno++;
8632 my->mc_wlen[toggle] += my->mc_env->me_psize;
8634 /* Update parent if there is one */
8635 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8636 SETPGNO(ni, mo->mp_pgno);
8637 mdb_cursor_pop(&mc);
8639 /* Otherwise we're done */
8649 /** Copy environment with compaction. */
8651 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8656 MDB_txn *txn = NULL;
8661 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8662 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8663 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8664 if (my.mc_wbuf[0] == NULL)
8667 pthread_mutex_init(&my.mc_mutex, NULL);
8668 pthread_cond_init(&my.mc_cond, NULL);
8669 #ifdef HAVE_MEMALIGN
8670 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8671 if (my.mc_wbuf[0] == NULL)
8674 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8679 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8680 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8685 my.mc_next_pgno = 2;
8691 THREAD_CREATE(thr, mdb_env_copythr, &my);
8693 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8697 mp = (MDB_page *)my.mc_wbuf[0];
8698 memset(mp, 0, 2*env->me_psize);
8700 mp->mp_flags = P_META;
8701 mm = (MDB_meta *)METADATA(mp);
8702 mdb_env_init_meta0(env, mm);
8703 mm->mm_address = env->me_metas[0]->mm_address;
8705 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8707 mp->mp_flags = P_META;
8708 *(MDB_meta *)METADATA(mp) = *mm;
8709 mm = (MDB_meta *)METADATA(mp);
8711 /* Count the number of free pages, subtract from lastpg to find
8712 * number of active pages
8715 MDB_ID freecount = 0;
8718 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8719 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8720 freecount += *(MDB_ID *)data.mv_data;
8721 freecount += txn->mt_dbs[0].md_branch_pages +
8722 txn->mt_dbs[0].md_leaf_pages +
8723 txn->mt_dbs[0].md_overflow_pages;
8725 /* Set metapage 1 */
8726 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8727 mm->mm_dbs[1] = txn->mt_dbs[1];
8728 if (mm->mm_last_pg > 1) {
8729 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8732 mm->mm_dbs[1].md_root = P_INVALID;
8735 my.mc_wlen[0] = env->me_psize * 2;
8737 pthread_mutex_lock(&my.mc_mutex);
8739 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8740 pthread_mutex_unlock(&my.mc_mutex);
8741 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8742 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8743 rc = mdb_env_cthr_toggle(&my, 1);
8744 mdb_env_cthr_toggle(&my, -1);
8745 pthread_mutex_lock(&my.mc_mutex);
8747 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8748 pthread_mutex_unlock(&my.mc_mutex);
8753 CloseHandle(my.mc_cond);
8754 CloseHandle(my.mc_mutex);
8755 _aligned_free(my.mc_wbuf[0]);
8757 pthread_cond_destroy(&my.mc_cond);
8758 pthread_mutex_destroy(&my.mc_mutex);
8759 free(my.mc_wbuf[0]);
8764 /** Copy environment as-is. */
8766 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8768 MDB_txn *txn = NULL;
8774 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8778 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8781 /* Do the lock/unlock of the reader mutex before starting the
8782 * write txn. Otherwise other read txns could block writers.
8784 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8789 /* We must start the actual read txn after blocking writers */
8790 mdb_txn_reset0(txn, "reset-stage1");
8792 /* Temporarily block writers until we snapshot the meta pages */
8795 rc = mdb_txn_renew0(txn);
8797 UNLOCK_MUTEX_W(env);
8802 wsize = env->me_psize * 2;
8806 DO_WRITE(rc, fd, ptr, w2, len);
8810 } else if (len > 0) {
8816 /* Non-blocking or async handles are not supported */
8822 UNLOCK_MUTEX_W(env);
8827 w2 = txn->mt_next_pgno * env->me_psize;
8830 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8837 if (wsize > MAX_WRITE)
8841 DO_WRITE(rc, fd, ptr, w2, len);
8845 } else if (len > 0) {
8862 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8864 if (flags & MDB_CP_COMPACT)
8865 return mdb_env_copyfd1(env, fd);
8867 return mdb_env_copyfd0(env, fd);
8871 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8873 return mdb_env_copyfd2(env, fd, 0);
8877 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8881 HANDLE newfd = INVALID_HANDLE_VALUE;
8883 if (env->me_flags & MDB_NOSUBDIR) {
8884 lpath = (char *)path;
8887 len += sizeof(DATANAME);
8888 lpath = malloc(len);
8891 sprintf(lpath, "%s" DATANAME, path);
8894 /* The destination path must exist, but the destination file must not.
8895 * We don't want the OS to cache the writes, since the source data is
8896 * already in the OS cache.
8899 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8900 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8902 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8904 if (newfd == INVALID_HANDLE_VALUE) {
8909 if (env->me_psize >= env->me_os_psize) {
8911 /* Set O_DIRECT if the file system supports it */
8912 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8913 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8915 #ifdef F_NOCACHE /* __APPLE__ */
8916 rc = fcntl(newfd, F_NOCACHE, 1);
8924 rc = mdb_env_copyfd2(env, newfd, flags);
8927 if (!(env->me_flags & MDB_NOSUBDIR))
8929 if (newfd != INVALID_HANDLE_VALUE)
8930 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8937 mdb_env_copy(MDB_env *env, const char *path)
8939 return mdb_env_copy2(env, path, 0);
8943 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8945 if ((flag & CHANGEABLE) != flag)
8948 env->me_flags |= flag;
8950 env->me_flags &= ~flag;
8955 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8960 *arg = env->me_flags;
8965 mdb_env_set_userctx(MDB_env *env, void *ctx)
8969 env->me_userctx = ctx;
8974 mdb_env_get_userctx(MDB_env *env)
8976 return env ? env->me_userctx : NULL;
8980 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8985 env->me_assert_func = func;
8991 mdb_env_get_path(MDB_env *env, const char **arg)
8996 *arg = env->me_path;
9001 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9010 /** Common code for #mdb_stat() and #mdb_env_stat().
9011 * @param[in] env the environment to operate in.
9012 * @param[in] db the #MDB_db record containing the stats to return.
9013 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9014 * @return 0, this function always succeeds.
9017 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9019 arg->ms_psize = env->me_psize;
9020 arg->ms_depth = db->md_depth;
9021 arg->ms_branch_pages = db->md_branch_pages;
9022 arg->ms_leaf_pages = db->md_leaf_pages;
9023 arg->ms_overflow_pages = db->md_overflow_pages;
9024 arg->ms_entries = db->md_entries;
9030 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9034 if (env == NULL || arg == NULL)
9037 toggle = mdb_env_pick_meta(env);
9039 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9043 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9047 if (env == NULL || arg == NULL)
9050 toggle = mdb_env_pick_meta(env);
9051 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9052 arg->me_mapsize = env->me_mapsize;
9053 arg->me_maxreaders = env->me_maxreaders;
9055 /* me_numreaders may be zero if this process never used any readers. Use
9056 * the shared numreader count if it exists.
9058 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
9060 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9061 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9065 /** Set the default comparison functions for a database.
9066 * Called immediately after a database is opened to set the defaults.
9067 * The user can then override them with #mdb_set_compare() or
9068 * #mdb_set_dupsort().
9069 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9070 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9073 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9075 uint16_t f = txn->mt_dbs[dbi].md_flags;
9077 txn->mt_dbxs[dbi].md_cmp =
9078 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9079 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9081 txn->mt_dbxs[dbi].md_dcmp =
9082 !(f & MDB_DUPSORT) ? 0 :
9083 ((f & MDB_INTEGERDUP)
9084 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9085 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9088 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9094 int rc, dbflag, exact;
9095 unsigned int unused = 0, seq;
9098 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9099 mdb_default_cmp(txn, FREE_DBI);
9102 if ((flags & VALID_FLAGS) != flags)
9104 if (txn->mt_flags & MDB_TXN_ERROR)
9110 if (flags & PERSISTENT_FLAGS) {
9111 uint16_t f2 = flags & PERSISTENT_FLAGS;
9112 /* make sure flag changes get committed */
9113 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9114 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9115 txn->mt_flags |= MDB_TXN_DIRTY;
9118 mdb_default_cmp(txn, MAIN_DBI);
9122 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9123 mdb_default_cmp(txn, MAIN_DBI);
9126 /* Is the DB already open? */
9128 for (i=2; i<txn->mt_numdbs; i++) {
9129 if (!txn->mt_dbxs[i].md_name.mv_size) {
9130 /* Remember this free slot */
9131 if (!unused) unused = i;
9134 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9135 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9141 /* If no free slot and max hit, fail */
9142 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9143 return MDB_DBS_FULL;
9145 /* Cannot mix named databases with some mainDB flags */
9146 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9147 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9149 /* Find the DB info */
9150 dbflag = DB_NEW|DB_VALID;
9153 key.mv_data = (void *)name;
9154 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9155 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9156 if (rc == MDB_SUCCESS) {
9157 /* make sure this is actually a DB */
9158 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9159 if (!(node->mn_flags & F_SUBDATA))
9160 return MDB_INCOMPATIBLE;
9161 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9162 /* Create if requested */
9163 data.mv_size = sizeof(MDB_db);
9164 data.mv_data = &dummy;
9165 memset(&dummy, 0, sizeof(dummy));
9166 dummy.md_root = P_INVALID;
9167 dummy.md_flags = flags & PERSISTENT_FLAGS;
9168 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9172 /* OK, got info, add to table */
9173 if (rc == MDB_SUCCESS) {
9174 unsigned int slot = unused ? unused : txn->mt_numdbs;
9175 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9176 txn->mt_dbxs[slot].md_name.mv_size = len;
9177 txn->mt_dbxs[slot].md_rel = NULL;
9178 txn->mt_dbflags[slot] = dbflag;
9179 /* txn-> and env-> are the same in read txns, use
9180 * tmp variable to avoid undefined assignment
9182 seq = ++txn->mt_env->me_dbiseqs[slot];
9183 txn->mt_dbiseqs[slot] = seq;
9185 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9187 mdb_default_cmp(txn, slot);
9196 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9198 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9201 if (txn->mt_flags & MDB_TXN_ERROR)
9204 if (txn->mt_dbflags[dbi] & DB_STALE) {
9207 /* Stale, must read the DB's root. cursor_init does it for us. */
9208 mdb_cursor_init(&mc, txn, dbi, &mx);
9210 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9213 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9216 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9218 ptr = env->me_dbxs[dbi].md_name.mv_data;
9219 /* If there was no name, this was already closed */
9221 env->me_dbxs[dbi].md_name.mv_data = NULL;
9222 env->me_dbxs[dbi].md_name.mv_size = 0;
9223 env->me_dbflags[dbi] = 0;
9224 env->me_dbiseqs[dbi]++;
9229 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9231 /* We could return the flags for the FREE_DBI too but what's the point? */
9232 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9234 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9238 /** Add all the DB's pages to the free list.
9239 * @param[in] mc Cursor on the DB to free.
9240 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9241 * @return 0 on success, non-zero on failure.
9244 mdb_drop0(MDB_cursor *mc, int subs)
9248 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9249 if (rc == MDB_SUCCESS) {
9250 MDB_txn *txn = mc->mc_txn;
9255 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9256 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9259 mdb_cursor_copy(mc, &mx);
9260 while (mc->mc_snum > 0) {
9261 MDB_page *mp = mc->mc_pg[mc->mc_top];
9262 unsigned n = NUMKEYS(mp);
9264 for (i=0; i<n; i++) {
9265 ni = NODEPTR(mp, i);
9266 if (ni->mn_flags & F_BIGDATA) {
9269 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9270 rc = mdb_page_get(txn, pg, &omp, NULL);
9273 mdb_cassert(mc, IS_OVERFLOW(omp));
9274 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9278 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9279 mdb_xcursor_init1(mc, ni);
9280 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9286 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9288 for (i=0; i<n; i++) {
9290 ni = NODEPTR(mp, i);
9293 mdb_midl_xappend(txn->mt_free_pgs, pg);
9298 mc->mc_ki[mc->mc_top] = i;
9299 rc = mdb_cursor_sibling(mc, 1);
9301 if (rc != MDB_NOTFOUND)
9303 /* no more siblings, go back to beginning
9304 * of previous level.
9308 for (i=1; i<mc->mc_snum; i++) {
9310 mc->mc_pg[i] = mx.mc_pg[i];
9315 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9318 txn->mt_flags |= MDB_TXN_ERROR;
9319 } else if (rc == MDB_NOTFOUND) {
9325 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9327 MDB_cursor *mc, *m2;
9330 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9333 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9336 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9339 rc = mdb_cursor_open(txn, dbi, &mc);
9343 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9344 /* Invalidate the dropped DB's cursors */
9345 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9346 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9350 /* Can't delete the main DB */
9351 if (del && dbi > MAIN_DBI) {
9352 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9354 txn->mt_dbflags[dbi] = DB_STALE;
9355 mdb_dbi_close(txn->mt_env, dbi);
9357 txn->mt_flags |= MDB_TXN_ERROR;
9360 /* reset the DB record, mark it dirty */
9361 txn->mt_dbflags[dbi] |= DB_DIRTY;
9362 txn->mt_dbs[dbi].md_depth = 0;
9363 txn->mt_dbs[dbi].md_branch_pages = 0;
9364 txn->mt_dbs[dbi].md_leaf_pages = 0;
9365 txn->mt_dbs[dbi].md_overflow_pages = 0;
9366 txn->mt_dbs[dbi].md_entries = 0;
9367 txn->mt_dbs[dbi].md_root = P_INVALID;
9369 txn->mt_flags |= MDB_TXN_DIRTY;
9372 mdb_cursor_close(mc);
9376 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9378 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9381 txn->mt_dbxs[dbi].md_cmp = cmp;
9385 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9387 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9390 txn->mt_dbxs[dbi].md_dcmp = cmp;
9394 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9396 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9399 txn->mt_dbxs[dbi].md_rel = rel;
9403 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9405 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9408 txn->mt_dbxs[dbi].md_relctx = ctx;
9413 mdb_env_get_maxkeysize(MDB_env *env)
9415 return ENV_MAXKEY(env);
9419 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9421 unsigned int i, rdrs;
9424 int rc = 0, first = 1;
9428 if (!env->me_txns) {
9429 return func("(no reader locks)\n", ctx);
9431 rdrs = env->me_txns->mti_numreaders;
9432 mr = env->me_txns->mti_readers;
9433 for (i=0; i<rdrs; i++) {
9435 txnid_t txnid = mr[i].mr_txnid;
9436 sprintf(buf, txnid == (txnid_t)-1 ?
9437 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9438 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9441 rc = func(" pid thread txnid\n", ctx);
9445 rc = func(buf, ctx);
9451 rc = func("(no active readers)\n", ctx);
9456 /** Insert pid into list if not already present.
9457 * return -1 if already present.
9460 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9462 /* binary search of pid in list */
9464 unsigned cursor = 1;
9466 unsigned n = ids[0];
9469 unsigned pivot = n >> 1;
9470 cursor = base + pivot + 1;
9471 val = pid - ids[cursor];
9476 } else if ( val > 0 ) {
9481 /* found, so it's a duplicate */
9490 for (n = ids[0]; n > cursor; n--)
9497 mdb_reader_check(MDB_env *env, int *dead)
9499 unsigned int i, j, rdrs;
9501 MDB_PID_T *pids, pid;
9510 rdrs = env->me_txns->mti_numreaders;
9511 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9515 mr = env->me_txns->mti_readers;
9516 for (i=0; i<rdrs; i++) {
9517 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9519 if (mdb_pid_insert(pids, pid) == 0) {
9520 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9522 /* Recheck, a new process may have reused pid */
9523 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9524 for (j=i; j<rdrs; j++)
9525 if (mr[j].mr_pid == pid) {
9526 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9527 (unsigned) pid, mr[j].mr_txnid));
9532 UNLOCK_MUTEX_R(env);
projects / openldap / blob