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 computed max.
451 * This macro should normally be left alone or set to 0.
452 * Note that a database with big keys or dupsort data cannot be
453 * reliably modified by a liblmdb which uses a smaller max.
454 * The default is 511 for backwards compat, or 0 when #MDB_DEVEL.
456 * Other values are allowed, for backwards compat. However:
457 * A value bigger than the computed max can break if you do not
458 * know what you are doing, and liblmdb <= 0.9.10 can break when
459 * modifying a DB with keys/dupsort data bigger than its max.
461 * Data items in an #MDB_DUPSORT database are also limited to
462 * this size, since they're actually keys of a sub-DB. Keys and
463 * #MDB_DUPSORT data items must fit on a node in a regular page.
465 #ifndef MDB_MAXKEYSIZE
466 #define MDB_MAXKEYSIZE ((MDB_DEVEL) ? 0 : 511)
469 /** The maximum size of a key we can write to the environment. */
471 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
473 #define ENV_MAXKEY(env) ((env)->me_maxkey)
476 /** @brief The maximum size of a data item.
478 * We only store a 32 bit value for node sizes.
480 #define MAXDATASIZE 0xffffffffUL
483 /** Key size which fits in a #DKBUF.
486 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
489 * This is used for printing a hex dump of a key's contents.
491 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
492 /** Display a key in hex.
494 * Invoke a function to display a key in hex.
496 #define DKEY(x) mdb_dkey(x, kbuf)
502 /** An invalid page number.
503 * Mainly used to denote an empty tree.
505 #define P_INVALID (~(pgno_t)0)
507 /** Test if the flags \b f are set in a flag word \b w. */
508 #define F_ISSET(w, f) (((w) & (f)) == (f))
510 /** Round \b n up to an even number. */
511 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
513 /** Used for offsets within a single page.
514 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
517 typedef uint16_t indx_t;
519 /** Default size of memory map.
520 * This is certainly too small for any actual applications. Apps should always set
521 * the size explicitly using #mdb_env_set_mapsize().
523 #define DEFAULT_MAPSIZE 1048576
525 /** @defgroup readers Reader Lock Table
526 * Readers don't acquire any locks for their data access. Instead, they
527 * simply record their transaction ID in the reader table. The reader
528 * mutex is needed just to find an empty slot in the reader table. The
529 * slot's address is saved in thread-specific data so that subsequent read
530 * transactions started by the same thread need no further locking to proceed.
532 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
534 * No reader table is used if the database is on a read-only filesystem, or
535 * if #MDB_NOLOCK is set.
537 * Since the database uses multi-version concurrency control, readers don't
538 * actually need any locking. This table is used to keep track of which
539 * readers are using data from which old transactions, so that we'll know
540 * when a particular old transaction is no longer in use. Old transactions
541 * that have discarded any data pages can then have those pages reclaimed
542 * for use by a later write transaction.
544 * The lock table is constructed such that reader slots are aligned with the
545 * processor's cache line size. Any slot is only ever used by one thread.
546 * This alignment guarantees that there will be no contention or cache
547 * thrashing as threads update their own slot info, and also eliminates
548 * any need for locking when accessing a slot.
550 * A writer thread will scan every slot in the table to determine the oldest
551 * outstanding reader transaction. Any freed pages older than this will be
552 * reclaimed by the writer. The writer doesn't use any locks when scanning
553 * this table. This means that there's no guarantee that the writer will
554 * see the most up-to-date reader info, but that's not required for correct
555 * operation - all we need is to know the upper bound on the oldest reader,
556 * we don't care at all about the newest reader. So the only consequence of
557 * reading stale information here is that old pages might hang around a
558 * while longer before being reclaimed. That's actually good anyway, because
559 * the longer we delay reclaiming old pages, the more likely it is that a
560 * string of contiguous pages can be found after coalescing old pages from
561 * many old transactions together.
564 /** Number of slots in the reader table.
565 * This value was chosen somewhat arbitrarily. 126 readers plus a
566 * couple mutexes fit exactly into 8KB on my development machine.
567 * Applications should set the table size using #mdb_env_set_maxreaders().
569 #define DEFAULT_READERS 126
571 /** The size of a CPU cache line in bytes. We want our lock structures
572 * aligned to this size to avoid false cache line sharing in the
574 * This value works for most CPUs. For Itanium this should be 128.
580 /** The information we store in a single slot of the reader table.
581 * In addition to a transaction ID, we also record the process and
582 * thread ID that owns a slot, so that we can detect stale information,
583 * e.g. threads or processes that went away without cleaning up.
584 * @note We currently don't check for stale records. We simply re-init
585 * the table when we know that we're the only process opening the
588 typedef struct MDB_rxbody {
589 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
590 * Multiple readers that start at the same time will probably have the
591 * same ID here. Again, it's not important to exclude them from
592 * anything; all we need to know is which version of the DB they
593 * started from so we can avoid overwriting any data used in that
594 * particular version.
596 volatile txnid_t mrb_txnid;
597 /** The process ID of the process owning this reader txn. */
598 volatile MDB_PID_T mrb_pid;
599 /** The thread ID of the thread owning this txn. */
600 volatile MDB_THR_T mrb_tid;
603 /** The actual reader record, with cacheline padding. */
604 typedef struct MDB_reader {
607 /** shorthand for mrb_txnid */
608 #define mr_txnid mru.mrx.mrb_txnid
609 #define mr_pid mru.mrx.mrb_pid
610 #define mr_tid mru.mrx.mrb_tid
611 /** cache line alignment */
612 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
616 /** The header for the reader table.
617 * The table resides in a memory-mapped file. (This is a different file
618 * than is used for the main database.)
620 * For POSIX the actual mutexes reside in the shared memory of this
621 * mapped file. On Windows, mutexes are named objects allocated by the
622 * kernel; we store the mutex names in this mapped file so that other
623 * processes can grab them. This same approach is also used on
624 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
625 * process-shared POSIX mutexes. For these cases where a named object
626 * is used, the object name is derived from a 64 bit FNV hash of the
627 * environment pathname. As such, naming collisions are extremely
628 * unlikely. If a collision occurs, the results are unpredictable.
630 typedef struct MDB_txbody {
631 /** Stamp identifying this as an LMDB file. It must be set
634 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
636 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
637 char mtb_rmname[MNAME_LEN];
639 /** Mutex protecting access to this table.
640 * This is the reader lock that #LOCK_MUTEX_R acquires.
642 pthread_mutex_t mtb_mutex;
644 /** The ID of the last transaction committed to the database.
645 * This is recorded here only for convenience; the value can always
646 * be determined by reading the main database meta pages.
648 volatile txnid_t mtb_txnid;
649 /** The number of slots that have been used in the reader table.
650 * This always records the maximum count, it is not decremented
651 * when readers release their slots.
653 volatile unsigned mtb_numreaders;
656 /** The actual reader table definition. */
657 typedef struct MDB_txninfo {
660 #define mti_magic mt1.mtb.mtb_magic
661 #define mti_format mt1.mtb.mtb_format
662 #define mti_mutex mt1.mtb.mtb_mutex
663 #define mti_rmname mt1.mtb.mtb_rmname
664 #define mti_txnid mt1.mtb.mtb_txnid
665 #define mti_numreaders mt1.mtb.mtb_numreaders
666 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
669 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
670 char mt2_wmname[MNAME_LEN];
671 #define mti_wmname mt2.mt2_wmname
673 pthread_mutex_t mt2_wmutex;
674 #define mti_wmutex mt2.mt2_wmutex
676 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
678 MDB_reader mti_readers[1];
681 /** Lockfile format signature: version, features and field layout */
682 #define MDB_LOCK_FORMAT \
684 ((MDB_LOCK_VERSION) \
685 /* Flags which describe functionality */ \
686 + (((MDB_PIDLOCK) != 0) << 16)))
689 /** Common header for all page types.
690 * Overflow records occupy a number of contiguous pages with no
691 * headers on any page after the first.
693 typedef struct MDB_page {
694 #define mp_pgno mp_p.p_pgno
695 #define mp_next mp_p.p_next
697 pgno_t p_pgno; /**< page number */
698 struct MDB_page *p_next; /**< for in-memory list of freed pages */
701 /** @defgroup mdb_page Page Flags
703 * Flags for the page headers.
706 #define P_BRANCH 0x01 /**< branch page */
707 #define P_LEAF 0x02 /**< leaf page */
708 #define P_OVERFLOW 0x04 /**< overflow page */
709 #define P_META 0x08 /**< meta page */
710 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
711 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
712 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
713 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
714 #define P_KEEP 0x8000 /**< leave this page alone during spill */
716 uint16_t mp_flags; /**< @ref mdb_page */
717 #define mp_lower mp_pb.pb.pb_lower
718 #define mp_upper mp_pb.pb.pb_upper
719 #define mp_pages mp_pb.pb_pages
722 indx_t pb_lower; /**< lower bound of free space */
723 indx_t pb_upper; /**< upper bound of free space */
725 uint32_t pb_pages; /**< number of overflow pages */
727 indx_t mp_ptrs[1]; /**< dynamic size */
730 /** Size of the page header, excluding dynamic data at the end */
731 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
733 /** Address of first usable data byte in a page, after the header */
734 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
736 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
737 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
739 /** Number of nodes on a page */
740 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
742 /** The amount of space remaining in the page */
743 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
745 /** The percentage of space used in the page, in tenths of a percent. */
746 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
747 ((env)->me_psize - PAGEHDRSZ))
748 /** The minimum page fill factor, in tenths of a percent.
749 * Pages emptier than this are candidates for merging.
751 #define FILL_THRESHOLD 250
753 /** Test if a page is a leaf page */
754 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
755 /** Test if a page is a LEAF2 page */
756 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
757 /** Test if a page is a branch page */
758 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
759 /** Test if a page is an overflow page */
760 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
761 /** Test if a page is a sub page */
762 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
764 /** The number of overflow pages needed to store the given size. */
765 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
767 /** Link in #MDB_txn.%mt_loose_pgs list */
768 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
770 /** Header for a single key/data pair within a page.
771 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
772 * We guarantee 2-byte alignment for 'MDB_node's.
774 typedef struct MDB_node {
775 /** lo and hi are used for data size on leaf nodes and for
776 * child pgno on branch nodes. On 64 bit platforms, flags
777 * is also used for pgno. (Branch nodes have no flags).
778 * They are in host byte order in case that lets some
779 * accesses be optimized into a 32-bit word access.
781 #if BYTE_ORDER == LITTLE_ENDIAN
782 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
784 unsigned short mn_hi, mn_lo;
786 /** @defgroup mdb_node Node Flags
788 * Flags for node headers.
791 #define F_BIGDATA 0x01 /**< data put on overflow page */
792 #define F_SUBDATA 0x02 /**< data is a sub-database */
793 #define F_DUPDATA 0x04 /**< data has duplicates */
795 /** valid flags for #mdb_node_add() */
796 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
799 unsigned short mn_flags; /**< @ref mdb_node */
800 unsigned short mn_ksize; /**< key size */
801 char mn_data[1]; /**< key and data are appended here */
804 /** Size of the node header, excluding dynamic data at the end */
805 #define NODESIZE offsetof(MDB_node, mn_data)
807 /** Bit position of top word in page number, for shifting mn_flags */
808 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
810 /** Size of a node in a branch page with a given key.
811 * This is just the node header plus the key, there is no data.
813 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
815 /** Size of a node in a leaf page with a given key and data.
816 * This is node header plus key plus data size.
818 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
820 /** Address of node \b i in page \b p */
821 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
823 /** Address of the key for the node */
824 #define NODEKEY(node) (void *)((node)->mn_data)
826 /** Address of the data for a node */
827 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
829 /** Get the page number pointed to by a branch node */
830 #define NODEPGNO(node) \
831 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
832 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
833 /** Set the page number in a branch node */
834 #define SETPGNO(node,pgno) do { \
835 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
836 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
838 /** Get the size of the data in a leaf node */
839 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
840 /** Set the size of the data for a leaf node */
841 #define SETDSZ(node,size) do { \
842 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
843 /** The size of a key in a node */
844 #define NODEKSZ(node) ((node)->mn_ksize)
846 /** Copy a page number from src to dst */
848 #define COPY_PGNO(dst,src) dst = src
850 #if SIZE_MAX > 4294967295UL
851 #define COPY_PGNO(dst,src) do { \
852 unsigned short *s, *d; \
853 s = (unsigned short *)&(src); \
854 d = (unsigned short *)&(dst); \
861 #define COPY_PGNO(dst,src) do { \
862 unsigned short *s, *d; \
863 s = (unsigned short *)&(src); \
864 d = (unsigned short *)&(dst); \
870 /** The address of a key in a LEAF2 page.
871 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
872 * There are no node headers, keys are stored contiguously.
874 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
876 /** Set the \b node's key into \b keyptr, if requested. */
877 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
878 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
880 /** Set the \b node's key into \b key. */
881 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
883 /** Information about a single database in the environment. */
884 typedef struct MDB_db {
885 uint32_t md_pad; /**< also ksize for LEAF2 pages */
886 uint16_t md_flags; /**< @ref mdb_dbi_open */
887 uint16_t md_depth; /**< depth of this tree */
888 pgno_t md_branch_pages; /**< number of internal pages */
889 pgno_t md_leaf_pages; /**< number of leaf pages */
890 pgno_t md_overflow_pages; /**< number of overflow pages */
891 size_t md_entries; /**< number of data items */
892 pgno_t md_root; /**< the root page of this tree */
895 /** mdb_dbi_open flags */
896 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
897 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
898 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
899 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
901 /** Handle for the DB used to track free pages. */
903 /** Handle for the default DB. */
906 /** Meta page content.
907 * A meta page is the start point for accessing a database snapshot.
908 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
910 typedef struct MDB_meta {
911 /** Stamp identifying this as an LMDB file. It must be set
914 /** Version number of this file. Must be set to #MDB_DATA_VERSION. */
916 void *mm_address; /**< address for fixed mapping */
917 size_t mm_mapsize; /**< size of mmap region */
918 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
919 /** The size of pages used in this DB */
920 #define mm_psize mm_dbs[0].md_pad
921 /** Any persistent environment flags. @ref mdb_env */
922 #define mm_flags mm_dbs[0].md_flags
923 pgno_t mm_last_pg; /**< last used page in file */
924 volatile txnid_t mm_txnid; /**< txnid that committed this page */
927 /** Buffer for a stack-allocated meta page.
928 * The members define size and alignment, and silence type
929 * aliasing warnings. They are not used directly; that could
930 * mean incorrectly using several union members in parallel.
932 typedef union MDB_metabuf {
935 char mm_pad[PAGEHDRSZ];
940 /** Auxiliary DB info.
941 * The information here is mostly static/read-only. There is
942 * only a single copy of this record in the environment.
944 typedef struct MDB_dbx {
945 MDB_val md_name; /**< name of the database */
946 MDB_cmp_func *md_cmp; /**< function for comparing keys */
947 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
948 MDB_rel_func *md_rel; /**< user relocate function */
949 void *md_relctx; /**< user-provided context for md_rel */
952 /** A database transaction.
953 * Every operation requires a transaction handle.
956 MDB_txn *mt_parent; /**< parent of a nested txn */
957 MDB_txn *mt_child; /**< nested txn under this txn */
958 pgno_t mt_next_pgno; /**< next unallocated page */
959 /** The ID of this transaction. IDs are integers incrementing from 1.
960 * Only committed write transactions increment the ID. If a transaction
961 * aborts, the ID may be re-used by the next writer.
964 MDB_env *mt_env; /**< the DB environment */
965 /** The list of pages that became unused during this transaction.
968 /** The list of loose pages that became unused and may be reused
969 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
971 MDB_page *mt_loose_pgs;
972 /* #Number of loose pages (#mt_loose_pgs) */
974 /** The sorted list of dirty pages we temporarily wrote to disk
975 * because the dirty list was full. page numbers in here are
976 * shifted left by 1, deleted slots have the LSB set.
978 MDB_IDL mt_spill_pgs;
980 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
982 /** For read txns: This thread/txn's reader table slot, or NULL. */
985 /** Array of records for each DB known in the environment. */
987 /** Array of MDB_db records for each known DB */
989 /** Array of sequence numbers for each DB handle */
990 unsigned int *mt_dbiseqs;
991 /** @defgroup mt_dbflag Transaction DB Flags
995 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
996 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
997 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
998 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
1000 /** In write txns, array of cursors for each DB */
1001 MDB_cursor **mt_cursors;
1002 /** Array of flags for each DB */
1003 unsigned char *mt_dbflags;
1004 /** Number of DB records in use. This number only ever increments;
1005 * we don't decrement it when individual DB handles are closed.
1009 /** @defgroup mdb_txn Transaction Flags
1013 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
1014 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
1015 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
1016 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
1018 unsigned int mt_flags; /**< @ref mdb_txn */
1019 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
1020 * Includes ancestor txns' dirty pages not hidden by other txns'
1021 * dirty/spilled pages. Thus commit(nested txn) has room to merge
1022 * dirty_list into mt_parent after freeing hidden mt_parent pages.
1024 unsigned int mt_dirty_room;
1027 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
1028 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
1029 * raise this on a 64 bit machine.
1031 #define CURSOR_STACK 32
1035 /** Cursors are used for all DB operations.
1036 * A cursor holds a path of (page pointer, key index) from the DB
1037 * root to a position in the DB, plus other state. #MDB_DUPSORT
1038 * cursors include an xcursor to the current data item. Write txns
1039 * track their cursors and keep them up to date when data moves.
1040 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1041 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1044 /** Next cursor on this DB in this txn */
1045 MDB_cursor *mc_next;
1046 /** Backup of the original cursor if this cursor is a shadow */
1047 MDB_cursor *mc_backup;
1048 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1049 struct MDB_xcursor *mc_xcursor;
1050 /** The transaction that owns this cursor */
1052 /** The database handle this cursor operates on */
1054 /** The database record for this cursor */
1056 /** The database auxiliary record for this cursor */
1058 /** The @ref mt_dbflag for this database */
1059 unsigned char *mc_dbflag;
1060 unsigned short mc_snum; /**< number of pushed pages */
1061 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1062 /** @defgroup mdb_cursor Cursor Flags
1064 * Cursor state flags.
1067 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1068 #define C_EOF 0x02 /**< No more data */
1069 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1070 #define C_DEL 0x08 /**< last op was a cursor_del */
1071 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1072 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1074 unsigned int mc_flags; /**< @ref mdb_cursor */
1075 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1076 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1079 /** Context for sorted-dup records.
1080 * We could have gone to a fully recursive design, with arbitrarily
1081 * deep nesting of sub-databases. But for now we only handle these
1082 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1084 typedef struct MDB_xcursor {
1085 /** A sub-cursor for traversing the Dup DB */
1086 MDB_cursor mx_cursor;
1087 /** The database record for this Dup DB */
1089 /** The auxiliary DB record for this Dup DB */
1091 /** The @ref mt_dbflag for this Dup DB */
1092 unsigned char mx_dbflag;
1095 /** State of FreeDB old pages, stored in the MDB_env */
1096 typedef struct MDB_pgstate {
1097 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1098 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1101 /** The database environment. */
1103 HANDLE me_fd; /**< The main data file */
1104 HANDLE me_lfd; /**< The lock file */
1105 HANDLE me_mfd; /**< just for writing the meta pages */
1106 /** Failed to update the meta page. Probably an I/O error. */
1107 #define MDB_FATAL_ERROR 0x80000000U
1108 /** Some fields are initialized. */
1109 #define MDB_ENV_ACTIVE 0x20000000U
1110 /** me_txkey is set */
1111 #define MDB_ENV_TXKEY 0x10000000U
1112 /** fdatasync is unreliable */
1113 #define MDB_FSYNCONLY 0x08000000U
1114 uint32_t me_flags; /**< @ref mdb_env */
1115 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1116 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1117 unsigned int me_maxreaders; /**< size of the reader table */
1118 unsigned int me_numreaders; /**< max numreaders set by this env */
1119 MDB_dbi me_numdbs; /**< number of DBs opened */
1120 MDB_dbi me_maxdbs; /**< size of the DB table */
1121 MDB_PID_T me_pid; /**< process ID of this env */
1122 char *me_path; /**< path to the DB files */
1123 char *me_map; /**< the memory map of the data file */
1124 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1125 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1126 void *me_pbuf; /**< scratch area for DUPSORT put() */
1127 MDB_txn *me_txn; /**< current write transaction */
1128 MDB_txn *me_txn0; /**< prealloc'd write transaction */
1129 size_t me_mapsize; /**< size of the data memory map */
1130 off_t me_size; /**< current file size */
1131 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1132 MDB_dbx *me_dbxs; /**< array of static DB info */
1133 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1134 unsigned int *me_dbiseqs; /**< array of dbi sequence numbers */
1135 pthread_key_t me_txkey; /**< thread-key for readers */
1136 txnid_t me_pgoldest; /**< ID of oldest reader last time we looked */
1137 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1138 # define me_pglast me_pgstate.mf_pglast
1139 # define me_pghead me_pgstate.mf_pghead
1140 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1141 /** IDL of pages that became unused in a write txn */
1142 MDB_IDL me_free_pgs;
1143 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1144 MDB_ID2L me_dirty_list;
1145 /** Max number of freelist items that can fit in a single overflow page */
1147 /** Max size of a node on a page */
1148 unsigned int me_nodemax;
1149 #if !(MDB_MAXKEYSIZE)
1150 unsigned int me_maxkey; /**< max size of a key */
1152 int me_live_reader; /**< have liveness lock in reader table */
1154 int me_pidquery; /**< Used in OpenProcess */
1155 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1157 #elif defined(MDB_USE_POSIX_SEM)
1158 sem_t *me_rmutex; /* Shared mutexes are not supported */
1161 void *me_userctx; /**< User-settable context */
1162 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1165 /** Nested transaction */
1166 typedef struct MDB_ntxn {
1167 MDB_txn mnt_txn; /**< the transaction */
1168 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1171 /** max number of pages to commit in one writev() call */
1172 #define MDB_COMMIT_PAGES 64
1173 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1174 #undef MDB_COMMIT_PAGES
1175 #define MDB_COMMIT_PAGES IOV_MAX
1178 /** max bytes to write in one call */
1179 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1181 /** Check \b txn and \b dbi arguments to a function */
1182 #define TXN_DBI_EXIST(txn, dbi) \
1183 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1185 /** Check for misused \b dbi handles */
1186 #define TXN_DBI_CHANGED(txn, dbi) \
1187 ((txn)->mt_dbiseqs[dbi] != (txn)->mt_env->me_dbiseqs[dbi])
1189 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1190 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1191 static int mdb_page_touch(MDB_cursor *mc);
1193 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1194 static int mdb_page_search_root(MDB_cursor *mc,
1195 MDB_val *key, int modify);
1196 #define MDB_PS_MODIFY 1
1197 #define MDB_PS_ROOTONLY 2
1198 #define MDB_PS_FIRST 4
1199 #define MDB_PS_LAST 8
1200 static int mdb_page_search(MDB_cursor *mc,
1201 MDB_val *key, int flags);
1202 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1204 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1205 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1206 pgno_t newpgno, unsigned int nflags);
1208 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1209 static int mdb_env_pick_meta(const MDB_env *env);
1210 static int mdb_env_write_meta(MDB_txn *txn);
1211 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1212 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1214 static void mdb_env_close0(MDB_env *env, int excl);
1216 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1217 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1218 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1219 static void mdb_node_del(MDB_cursor *mc, int ksize);
1220 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1221 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1222 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1223 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1224 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1226 static int mdb_rebalance(MDB_cursor *mc);
1227 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1229 static void mdb_cursor_pop(MDB_cursor *mc);
1230 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1232 static int mdb_cursor_del0(MDB_cursor *mc);
1233 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1234 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1235 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1236 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1237 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1239 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1240 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1242 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1243 static void mdb_xcursor_init0(MDB_cursor *mc);
1244 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1246 static int mdb_drop0(MDB_cursor *mc, int subs);
1247 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1250 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1253 /** Compare two items pointing at size_t's of unknown alignment. */
1254 #ifdef MISALIGNED_OK
1255 # define mdb_cmp_clong mdb_cmp_long
1257 # define mdb_cmp_clong mdb_cmp_cint
1261 static SECURITY_DESCRIPTOR mdb_null_sd;
1262 static SECURITY_ATTRIBUTES mdb_all_sa;
1263 static int mdb_sec_inited;
1266 /** Return the library version info. */
1268 mdb_version(int *major, int *minor, int *patch)
1270 if (major) *major = MDB_VERSION_MAJOR;
1271 if (minor) *minor = MDB_VERSION_MINOR;
1272 if (patch) *patch = MDB_VERSION_PATCH;
1273 return MDB_VERSION_STRING;
1276 /** Table of descriptions for LMDB @ref errors */
1277 static char *const mdb_errstr[] = {
1278 "MDB_KEYEXIST: Key/data pair already exists",
1279 "MDB_NOTFOUND: No matching key/data pair found",
1280 "MDB_PAGE_NOTFOUND: Requested page not found",
1281 "MDB_CORRUPTED: Located page was wrong type",
1282 "MDB_PANIC: Update of meta page failed",
1283 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1284 "MDB_INVALID: File is not an LMDB file",
1285 "MDB_MAP_FULL: Environment mapsize limit reached",
1286 "MDB_DBS_FULL: Environment maxdbs limit reached",
1287 "MDB_READERS_FULL: Environment maxreaders limit reached",
1288 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1289 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1290 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1291 "MDB_PAGE_FULL: Internal error - page has no more space",
1292 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1293 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1294 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1295 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1296 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1297 "MDB_BAD_DBI: The specified DBI handle was closed/changed unexpectedly",
1301 mdb_strerror(int err)
1304 /** HACK: pad 4KB on stack over the buf. Return system msgs in buf.
1305 * This works as long as no function between the call to mdb_strerror
1306 * and the actual use of the message uses more than 4K of stack.
1309 char buf[1024], *ptr = buf;
1313 return ("Successful return: 0");
1315 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1316 i = err - MDB_KEYEXIST;
1317 return mdb_errstr[i];
1321 /* These are the C-runtime error codes we use. The comment indicates
1322 * their numeric value, and the Win32 error they would correspond to
1323 * if the error actually came from a Win32 API. A major mess, we should
1324 * have used LMDB-specific error codes for everything.
1327 case ENOENT: /* 2, FILE_NOT_FOUND */
1328 case EIO: /* 5, ACCESS_DENIED */
1329 case ENOMEM: /* 12, INVALID_ACCESS */
1330 case EACCES: /* 13, INVALID_DATA */
1331 case EBUSY: /* 16, CURRENT_DIRECTORY */
1332 case EINVAL: /* 22, BAD_COMMAND */
1333 case ENOSPC: /* 28, OUT_OF_PAPER */
1334 return strerror(err);
1339 FormatMessage(FORMAT_MESSAGE_FROM_SYSTEM |
1340 FORMAT_MESSAGE_IGNORE_INSERTS,
1341 NULL, err, 0, ptr, sizeof(buf), (va_list *)pad);
1344 return strerror(err);
1348 /** assert(3) variant in cursor context */
1349 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1350 /** assert(3) variant in transaction context */
1351 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1352 /** assert(3) variant in environment context */
1353 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1356 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1357 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1360 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1361 const char *func, const char *file, int line)
1364 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1365 file, line, expr_txt, func);
1366 if (env->me_assert_func)
1367 env->me_assert_func(env, buf);
1368 fprintf(stderr, "%s\n", buf);
1372 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1376 /** Return the page number of \b mp which may be sub-page, for debug output */
1378 mdb_dbg_pgno(MDB_page *mp)
1381 COPY_PGNO(ret, mp->mp_pgno);
1385 /** Display a key in hexadecimal and return the address of the result.
1386 * @param[in] key the key to display
1387 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1388 * @return The key in hexadecimal form.
1391 mdb_dkey(MDB_val *key, char *buf)
1394 unsigned char *c = key->mv_data;
1400 if (key->mv_size > DKBUF_MAXKEYSIZE)
1401 return "MDB_MAXKEYSIZE";
1402 /* may want to make this a dynamic check: if the key is mostly
1403 * printable characters, print it as-is instead of converting to hex.
1407 for (i=0; i<key->mv_size; i++)
1408 ptr += sprintf(ptr, "%02x", *c++);
1410 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1416 mdb_leafnode_type(MDB_node *n)
1418 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1419 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1420 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1423 /** Display all the keys in the page. */
1425 mdb_page_list(MDB_page *mp)
1427 pgno_t pgno = mdb_dbg_pgno(mp);
1428 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1430 unsigned int i, nkeys, nsize, total = 0;
1434 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1435 case P_BRANCH: type = "Branch page"; break;
1436 case P_LEAF: type = "Leaf page"; break;
1437 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1438 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1439 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1441 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1442 pgno, mp->mp_pages, state);
1445 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1446 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1449 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1453 nkeys = NUMKEYS(mp);
1454 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1456 for (i=0; i<nkeys; i++) {
1457 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1458 key.mv_size = nsize = mp->mp_pad;
1459 key.mv_data = LEAF2KEY(mp, i, nsize);
1461 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1464 node = NODEPTR(mp, i);
1465 key.mv_size = node->mn_ksize;
1466 key.mv_data = node->mn_data;
1467 nsize = NODESIZE + key.mv_size;
1468 if (IS_BRANCH(mp)) {
1469 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1473 if (F_ISSET(node->mn_flags, F_BIGDATA))
1474 nsize += sizeof(pgno_t);
1476 nsize += NODEDSZ(node);
1478 nsize += sizeof(indx_t);
1479 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1480 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1482 total = EVEN(total);
1484 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1485 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1489 mdb_cursor_chk(MDB_cursor *mc)
1495 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1496 for (i=0; i<mc->mc_top; i++) {
1498 node = NODEPTR(mp, mc->mc_ki[i]);
1499 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1502 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1508 /** Count all the pages in each DB and in the freelist
1509 * and make sure it matches the actual number of pages
1511 * All named DBs must be open for a correct count.
1513 static void mdb_audit(MDB_txn *txn)
1517 MDB_ID freecount, count;
1522 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1523 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1524 freecount += *(MDB_ID *)data.mv_data;
1525 mdb_tassert(txn, rc == MDB_NOTFOUND);
1528 for (i = 0; i<txn->mt_numdbs; i++) {
1530 if (!(txn->mt_dbflags[i] & DB_VALID))
1532 mdb_cursor_init(&mc, txn, i, &mx);
1533 if (txn->mt_dbs[i].md_root == P_INVALID)
1535 count += txn->mt_dbs[i].md_branch_pages +
1536 txn->mt_dbs[i].md_leaf_pages +
1537 txn->mt_dbs[i].md_overflow_pages;
1538 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1539 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1540 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1543 mp = mc.mc_pg[mc.mc_top];
1544 for (j=0; j<NUMKEYS(mp); j++) {
1545 MDB_node *leaf = NODEPTR(mp, j);
1546 if (leaf->mn_flags & F_SUBDATA) {
1548 memcpy(&db, NODEDATA(leaf), sizeof(db));
1549 count += db.md_branch_pages + db.md_leaf_pages +
1550 db.md_overflow_pages;
1554 mdb_tassert(txn, rc == MDB_NOTFOUND);
1557 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1558 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1559 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1565 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1567 return txn->mt_dbxs[dbi].md_cmp(a, b);
1571 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1573 MDB_cmp_func *dcmp = txn->mt_dbxs[dbi].md_dcmp;
1574 #if UINT_MAX < SIZE_MAX
1575 if (dcmp == mdb_cmp_int && a->mv_size == sizeof(size_t))
1576 dcmp = mdb_cmp_clong;
1581 /** Allocate memory for a page.
1582 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1585 mdb_page_malloc(MDB_txn *txn, unsigned num)
1587 MDB_env *env = txn->mt_env;
1588 MDB_page *ret = env->me_dpages;
1589 size_t psize = env->me_psize, sz = psize, off;
1590 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1591 * For a single page alloc, we init everything after the page header.
1592 * For multi-page, we init the final page; if the caller needed that
1593 * many pages they will be filling in at least up to the last page.
1597 VGMEMP_ALLOC(env, ret, sz);
1598 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1599 env->me_dpages = ret->mp_next;
1602 psize -= off = PAGEHDRSZ;
1607 if ((ret = malloc(sz)) != NULL) {
1608 VGMEMP_ALLOC(env, ret, sz);
1609 if (!(env->me_flags & MDB_NOMEMINIT)) {
1610 memset((char *)ret + off, 0, psize);
1614 txn->mt_flags |= MDB_TXN_ERROR;
1618 /** Free a single page.
1619 * Saves single pages to a list, for future reuse.
1620 * (This is not used for multi-page overflow pages.)
1623 mdb_page_free(MDB_env *env, MDB_page *mp)
1625 mp->mp_next = env->me_dpages;
1626 VGMEMP_FREE(env, mp);
1627 env->me_dpages = mp;
1630 /** Free a dirty page */
1632 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1634 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1635 mdb_page_free(env, dp);
1637 /* large pages just get freed directly */
1638 VGMEMP_FREE(env, dp);
1643 /** Return all dirty pages to dpage list */
1645 mdb_dlist_free(MDB_txn *txn)
1647 MDB_env *env = txn->mt_env;
1648 MDB_ID2L dl = txn->mt_u.dirty_list;
1649 unsigned i, n = dl[0].mid;
1651 for (i = 1; i <= n; i++) {
1652 mdb_dpage_free(env, dl[i].mptr);
1657 /** Loosen or free a single page.
1658 * Saves single pages to a list for future reuse
1659 * in this same txn. It has been pulled from the freeDB
1660 * and already resides on the dirty list, but has been
1661 * deleted. Use these pages first before pulling again
1664 * If the page wasn't dirtied in this txn, just add it
1665 * to this txn's free list.
1668 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1671 pgno_t pgno = mp->mp_pgno;
1672 MDB_txn *txn = mc->mc_txn;
1674 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1675 if (txn->mt_parent) {
1676 MDB_ID2 *dl = txn->mt_u.dirty_list;
1677 /* If txn has a parent, make sure the page is in our
1681 unsigned x = mdb_mid2l_search(dl, pgno);
1682 if (x <= dl[0].mid && dl[x].mid == pgno) {
1683 if (mp != dl[x].mptr) { /* bad cursor? */
1684 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1685 txn->mt_flags |= MDB_TXN_ERROR;
1686 return MDB_CORRUPTED;
1693 /* no parent txn, so it's just ours */
1698 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1700 NEXT_LOOSE_PAGE(mp) = txn->mt_loose_pgs;
1701 txn->mt_loose_pgs = mp;
1702 txn->mt_loose_count++;
1703 mp->mp_flags |= P_LOOSE;
1705 int rc = mdb_midl_append(&txn->mt_free_pgs, pgno);
1713 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1714 * @param[in] mc A cursor handle for the current operation.
1715 * @param[in] pflags Flags of the pages to update:
1716 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1717 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1718 * @return 0 on success, non-zero on failure.
1721 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1723 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1724 MDB_txn *txn = mc->mc_txn;
1730 int rc = MDB_SUCCESS, level;
1732 /* Mark pages seen by cursors */
1733 if (mc->mc_flags & C_UNTRACK)
1734 mc = NULL; /* will find mc in mt_cursors */
1735 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1736 for (; mc; mc=mc->mc_next) {
1737 if (!(mc->mc_flags & C_INITIALIZED))
1739 for (m3 = mc;; m3 = &mx->mx_cursor) {
1741 for (j=0; j<m3->mc_snum; j++) {
1743 if ((mp->mp_flags & Mask) == pflags)
1744 mp->mp_flags ^= P_KEEP;
1746 mx = m3->mc_xcursor;
1747 /* Proceed to mx if it is at a sub-database */
1748 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1750 if (! (mp && (mp->mp_flags & P_LEAF)))
1752 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1753 if (!(leaf->mn_flags & F_SUBDATA))
1762 /* Mark dirty root pages */
1763 for (i=0; i<txn->mt_numdbs; i++) {
1764 if (txn->mt_dbflags[i] & DB_DIRTY) {
1765 pgno_t pgno = txn->mt_dbs[i].md_root;
1766 if (pgno == P_INVALID)
1768 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1770 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1771 dp->mp_flags ^= P_KEEP;
1779 static int mdb_page_flush(MDB_txn *txn, int keep);
1781 /** Spill pages from the dirty list back to disk.
1782 * This is intended to prevent running into #MDB_TXN_FULL situations,
1783 * but note that they may still occur in a few cases:
1784 * 1) our estimate of the txn size could be too small. Currently this
1785 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1786 * 2) child txns may run out of space if their parents dirtied a
1787 * lot of pages and never spilled them. TODO: we probably should do
1788 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1789 * the parent's dirty_room is below a given threshold.
1791 * Otherwise, if not using nested txns, it is expected that apps will
1792 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1793 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1794 * If the txn never references them again, they can be left alone.
1795 * If the txn only reads them, they can be used without any fuss.
1796 * If the txn writes them again, they can be dirtied immediately without
1797 * going thru all of the work of #mdb_page_touch(). Such references are
1798 * handled by #mdb_page_unspill().
1800 * Also note, we never spill DB root pages, nor pages of active cursors,
1801 * because we'll need these back again soon anyway. And in nested txns,
1802 * we can't spill a page in a child txn if it was already spilled in a
1803 * parent txn. That would alter the parent txns' data even though
1804 * the child hasn't committed yet, and we'd have no way to undo it if
1805 * the child aborted.
1807 * @param[in] m0 cursor A cursor handle identifying the transaction and
1808 * database for which we are checking space.
1809 * @param[in] key For a put operation, the key being stored.
1810 * @param[in] data For a put operation, the data being stored.
1811 * @return 0 on success, non-zero on failure.
1814 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1816 MDB_txn *txn = m0->mc_txn;
1818 MDB_ID2L dl = txn->mt_u.dirty_list;
1819 unsigned int i, j, need;
1822 if (m0->mc_flags & C_SUB)
1825 /* Estimate how much space this op will take */
1826 i = m0->mc_db->md_depth;
1827 /* Named DBs also dirty the main DB */
1828 if (m0->mc_dbi > MAIN_DBI)
1829 i += txn->mt_dbs[MAIN_DBI].md_depth;
1830 /* For puts, roughly factor in the key+data size */
1832 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1833 i += i; /* double it for good measure */
1836 if (txn->mt_dirty_room > i)
1839 if (!txn->mt_spill_pgs) {
1840 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1841 if (!txn->mt_spill_pgs)
1844 /* purge deleted slots */
1845 MDB_IDL sl = txn->mt_spill_pgs;
1846 unsigned int num = sl[0];
1848 for (i=1; i<=num; i++) {
1855 /* Preserve pages which may soon be dirtied again */
1856 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1859 /* Less aggressive spill - we originally spilled the entire dirty list,
1860 * with a few exceptions for cursor pages and DB root pages. But this
1861 * turns out to be a lot of wasted effort because in a large txn many
1862 * of those pages will need to be used again. So now we spill only 1/8th
1863 * of the dirty pages. Testing revealed this to be a good tradeoff,
1864 * better than 1/2, 1/4, or 1/10.
1866 if (need < MDB_IDL_UM_MAX / 8)
1867 need = MDB_IDL_UM_MAX / 8;
1869 /* Save the page IDs of all the pages we're flushing */
1870 /* flush from the tail forward, this saves a lot of shifting later on. */
1871 for (i=dl[0].mid; i && need; i--) {
1872 MDB_ID pn = dl[i].mid << 1;
1874 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1876 /* Can't spill twice, make sure it's not already in a parent's
1879 if (txn->mt_parent) {
1881 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1882 if (tx2->mt_spill_pgs) {
1883 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1884 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1885 dp->mp_flags |= P_KEEP;
1893 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1897 mdb_midl_sort(txn->mt_spill_pgs);
1899 /* Flush the spilled part of dirty list */
1900 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1903 /* Reset any dirty pages we kept that page_flush didn't see */
1904 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1907 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1911 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1913 mdb_find_oldest(MDB_txn *txn)
1916 txnid_t mr, oldest = txn->mt_txnid - 1;
1917 if (txn->mt_env->me_txns) {
1918 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1919 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1930 /** Add a page to the txn's dirty list */
1932 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1935 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1937 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1938 insert = mdb_mid2l_append;
1940 insert = mdb_mid2l_insert;
1942 mid.mid = mp->mp_pgno;
1944 rc = insert(txn->mt_u.dirty_list, &mid);
1945 mdb_tassert(txn, rc == 0);
1946 txn->mt_dirty_room--;
1949 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1950 * me_pghead and mt_next_pgno.
1952 * If there are free pages available from older transactions, they
1953 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1954 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1955 * and move me_pglast to say which records were consumed. Only this
1956 * function can create me_pghead and move me_pglast/mt_next_pgno.
1957 * @param[in] mc cursor A cursor handle identifying the transaction and
1958 * database for which we are allocating.
1959 * @param[in] num the number of pages to allocate.
1960 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1961 * will always be satisfied by a single contiguous chunk of memory.
1962 * @return 0 on success, non-zero on failure.
1965 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1967 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1968 /* Get at most <Max_retries> more freeDB records once me_pghead
1969 * has enough pages. If not enough, use new pages from the map.
1970 * If <Paranoid> and mc is updating the freeDB, only get new
1971 * records if me_pghead is empty. Then the freelist cannot play
1972 * catch-up with itself by growing while trying to save it.
1974 enum { Paranoid = 1, Max_retries = 500 };
1976 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1978 int rc, retry = num * 60;
1979 MDB_txn *txn = mc->mc_txn;
1980 MDB_env *env = txn->mt_env;
1981 pgno_t pgno, *mop = env->me_pghead;
1982 unsigned i, j, mop_len = mop ? mop[0] : 0, n2 = num-1;
1984 txnid_t oldest = 0, last;
1989 /* If there are any loose pages, just use them */
1990 if (num == 1 && txn->mt_loose_pgs) {
1991 np = txn->mt_loose_pgs;
1992 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1993 txn->mt_loose_count--;
1994 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
2002 /* If our dirty list is already full, we can't do anything */
2003 if (txn->mt_dirty_room == 0) {
2008 for (op = MDB_FIRST;; op = MDB_NEXT) {
2013 /* Seek a big enough contiguous page range. Prefer
2014 * pages at the tail, just truncating the list.
2020 if (mop[i-n2] == pgno+n2)
2027 if (op == MDB_FIRST) { /* 1st iteration */
2028 /* Prepare to fetch more and coalesce */
2029 last = env->me_pglast;
2030 oldest = env->me_pgoldest;
2031 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
2034 key.mv_data = &last; /* will look up last+1 */
2035 key.mv_size = sizeof(last);
2037 if (Paranoid && mc->mc_dbi == FREE_DBI)
2040 if (Paranoid && retry < 0 && mop_len)
2044 /* Do not fetch more if the record will be too recent */
2045 if (oldest <= last) {
2047 oldest = mdb_find_oldest(txn);
2048 env->me_pgoldest = oldest;
2054 rc = mdb_cursor_get(&m2, &key, NULL, op);
2056 if (rc == MDB_NOTFOUND)
2060 last = *(txnid_t*)key.mv_data;
2061 if (oldest <= last) {
2063 oldest = mdb_find_oldest(txn);
2064 env->me_pgoldest = oldest;
2070 np = m2.mc_pg[m2.mc_top];
2071 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
2072 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
2075 idl = (MDB_ID *) data.mv_data;
2078 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
2083 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
2085 mop = env->me_pghead;
2087 env->me_pglast = last;
2089 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
2090 last, txn->mt_dbs[FREE_DBI].md_root, i));
2092 DPRINTF(("IDL %"Z"u", idl[j]));
2094 /* Merge in descending sorted order */
2095 mdb_midl_xmerge(mop, idl);
2099 /* Use new pages from the map when nothing suitable in the freeDB */
2101 pgno = txn->mt_next_pgno;
2102 if (pgno + num >= env->me_maxpg) {
2103 DPUTS("DB size maxed out");
2109 if (env->me_flags & MDB_WRITEMAP) {
2110 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2112 if (!(np = mdb_page_malloc(txn, num))) {
2118 mop[0] = mop_len -= num;
2119 /* Move any stragglers down */
2120 for (j = i-num; j < mop_len; )
2121 mop[++j] = mop[++i];
2123 txn->mt_next_pgno = pgno + num;
2126 mdb_page_dirty(txn, np);
2132 txn->mt_flags |= MDB_TXN_ERROR;
2136 /** Copy the used portions of a non-overflow page.
2137 * @param[in] dst page to copy into
2138 * @param[in] src page to copy from
2139 * @param[in] psize size of a page
2142 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2144 enum { Align = sizeof(pgno_t) };
2145 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2147 /* If page isn't full, just copy the used portion. Adjust
2148 * alignment so memcpy may copy words instead of bytes.
2150 if ((unused &= -Align) && !IS_LEAF2(src)) {
2151 upper = (upper + PAGEBASE) & -Align;
2152 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2153 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2156 memcpy(dst, src, psize - unused);
2160 /** Pull a page off the txn's spill list, if present.
2161 * If a page being referenced was spilled to disk in this txn, bring
2162 * it back and make it dirty/writable again.
2163 * @param[in] txn the transaction handle.
2164 * @param[in] mp the page being referenced. It must not be dirty.
2165 * @param[out] ret the writable page, if any. ret is unchanged if
2166 * mp wasn't spilled.
2169 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2171 MDB_env *env = txn->mt_env;
2174 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2176 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2177 if (!tx2->mt_spill_pgs)
2179 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2180 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2183 if (txn->mt_dirty_room == 0)
2184 return MDB_TXN_FULL;
2185 if (IS_OVERFLOW(mp))
2189 if (env->me_flags & MDB_WRITEMAP) {
2192 np = mdb_page_malloc(txn, num);
2196 memcpy(np, mp, num * env->me_psize);
2198 mdb_page_copy(np, mp, env->me_psize);
2201 /* If in current txn, this page is no longer spilled.
2202 * If it happens to be the last page, truncate the spill list.
2203 * Otherwise mark it as deleted by setting the LSB.
2205 if (x == txn->mt_spill_pgs[0])
2206 txn->mt_spill_pgs[0]--;
2208 txn->mt_spill_pgs[x] |= 1;
2209 } /* otherwise, if belonging to a parent txn, the
2210 * page remains spilled until child commits
2213 mdb_page_dirty(txn, np);
2214 np->mp_flags |= P_DIRTY;
2222 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2223 * @param[in] mc cursor pointing to the page to be touched
2224 * @return 0 on success, non-zero on failure.
2227 mdb_page_touch(MDB_cursor *mc)
2229 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2230 MDB_txn *txn = mc->mc_txn;
2231 MDB_cursor *m2, *m3;
2235 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2236 if (txn->mt_flags & MDB_TXN_SPILLS) {
2238 rc = mdb_page_unspill(txn, mp, &np);
2244 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2245 (rc = mdb_page_alloc(mc, 1, &np)))
2248 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2249 mp->mp_pgno, pgno));
2250 mdb_cassert(mc, mp->mp_pgno != pgno);
2251 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2252 /* Update the parent page, if any, to point to the new page */
2254 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2255 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2256 SETPGNO(node, pgno);
2258 mc->mc_db->md_root = pgno;
2260 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2261 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2263 /* If txn has a parent, make sure the page is in our
2267 unsigned x = mdb_mid2l_search(dl, pgno);
2268 if (x <= dl[0].mid && dl[x].mid == pgno) {
2269 if (mp != dl[x].mptr) { /* bad cursor? */
2270 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2271 txn->mt_flags |= MDB_TXN_ERROR;
2272 return MDB_CORRUPTED;
2277 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2279 np = mdb_page_malloc(txn, 1);
2284 rc = mdb_mid2l_insert(dl, &mid);
2285 mdb_cassert(mc, rc == 0);
2290 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2292 np->mp_flags |= P_DIRTY;
2295 /* Adjust cursors pointing to mp */
2296 mc->mc_pg[mc->mc_top] = np;
2297 m2 = txn->mt_cursors[mc->mc_dbi];
2298 if (mc->mc_flags & C_SUB) {
2299 for (; m2; m2=m2->mc_next) {
2300 m3 = &m2->mc_xcursor->mx_cursor;
2301 if (m3->mc_snum < mc->mc_snum) continue;
2302 if (m3->mc_pg[mc->mc_top] == mp)
2303 m3->mc_pg[mc->mc_top] = np;
2306 for (; m2; m2=m2->mc_next) {
2307 if (m2->mc_snum < mc->mc_snum) continue;
2308 if (m2->mc_pg[mc->mc_top] == mp) {
2309 m2->mc_pg[mc->mc_top] = np;
2310 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2312 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2314 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2315 if (!(leaf->mn_flags & F_SUBDATA))
2316 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2324 txn->mt_flags |= MDB_TXN_ERROR;
2329 mdb_env_sync(MDB_env *env, int force)
2332 if (env->me_flags & MDB_RDONLY)
2334 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2335 if (env->me_flags & MDB_WRITEMAP) {
2336 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2337 ? MS_ASYNC : MS_SYNC;
2338 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2341 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2345 #ifdef BROKEN_FDATASYNC
2346 if (env->me_flags & MDB_FSYNCONLY) {
2347 if (fsync(env->me_fd))
2351 if (MDB_FDATASYNC(env->me_fd))
2358 /** Back up parent txn's cursors, then grab the originals for tracking */
2360 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2362 MDB_cursor *mc, *bk;
2367 for (i = src->mt_numdbs; --i >= 0; ) {
2368 if ((mc = src->mt_cursors[i]) != NULL) {
2369 size = sizeof(MDB_cursor);
2371 size += sizeof(MDB_xcursor);
2372 for (; mc; mc = bk->mc_next) {
2378 mc->mc_db = &dst->mt_dbs[i];
2379 /* Kill pointers into src - and dst to reduce abuse: The
2380 * user may not use mc until dst ends. Otherwise we'd...
2382 mc->mc_txn = NULL; /* ...set this to dst */
2383 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2384 if ((mx = mc->mc_xcursor) != NULL) {
2385 *(MDB_xcursor *)(bk+1) = *mx;
2386 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2388 mc->mc_next = dst->mt_cursors[i];
2389 dst->mt_cursors[i] = mc;
2396 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2397 * @param[in] txn the transaction handle.
2398 * @param[in] merge true to keep changes to parent cursors, false to revert.
2399 * @return 0 on success, non-zero on failure.
2402 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2404 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2408 for (i = txn->mt_numdbs; --i >= 0; ) {
2409 for (mc = cursors[i]; mc; mc = next) {
2411 if ((bk = mc->mc_backup) != NULL) {
2413 /* Commit changes to parent txn */
2414 mc->mc_next = bk->mc_next;
2415 mc->mc_backup = bk->mc_backup;
2416 mc->mc_txn = bk->mc_txn;
2417 mc->mc_db = bk->mc_db;
2418 mc->mc_dbflag = bk->mc_dbflag;
2419 if ((mx = mc->mc_xcursor) != NULL)
2420 mx->mx_cursor.mc_txn = bk->mc_txn;
2422 /* Abort nested txn */
2424 if ((mx = mc->mc_xcursor) != NULL)
2425 *mx = *(MDB_xcursor *)(bk+1);
2429 /* Only malloced cursors are permanently tracked. */
2437 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2440 mdb_txn_reset0(MDB_txn *txn, const char *act);
2442 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2448 Pidset = F_SETLK, Pidcheck = F_GETLK
2452 /** Set or check a pid lock. Set returns 0 on success.
2453 * Check returns 0 if the process is certainly dead, nonzero if it may
2454 * be alive (the lock exists or an error happened so we do not know).
2456 * On Windows Pidset is a no-op, we merely check for the existence
2457 * of the process with the given pid. On POSIX we use a single byte
2458 * lock on the lockfile, set at an offset equal to the pid.
2461 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2463 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2466 if (op == Pidcheck) {
2467 h = OpenProcess(env->me_pidquery, FALSE, pid);
2468 /* No documented "no such process" code, but other program use this: */
2470 return ErrCode() != ERROR_INVALID_PARAMETER;
2471 /* A process exists until all handles to it close. Has it exited? */
2472 ret = WaitForSingleObject(h, 0) != 0;
2479 struct flock lock_info;
2480 memset(&lock_info, 0, sizeof(lock_info));
2481 lock_info.l_type = F_WRLCK;
2482 lock_info.l_whence = SEEK_SET;
2483 lock_info.l_start = pid;
2484 lock_info.l_len = 1;
2485 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2486 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2488 } else if ((rc = ErrCode()) == EINTR) {
2496 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2497 * @param[in] txn the transaction handle to initialize
2498 * @return 0 on success, non-zero on failure.
2501 mdb_txn_renew0(MDB_txn *txn)
2503 MDB_env *env = txn->mt_env;
2504 MDB_txninfo *ti = env->me_txns;
2506 unsigned int i, nr, flags = txn->mt_flags;
2508 int rc, new_notls = 0;
2510 if ((flags &= MDB_TXN_RDONLY) != 0) {
2512 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2513 txn->mt_txnid = meta->mm_txnid;
2514 txn->mt_u.reader = NULL;
2516 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2517 pthread_getspecific(env->me_txkey);
2519 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2520 return MDB_BAD_RSLOT;
2522 MDB_PID_T pid = env->me_pid;
2523 MDB_THR_T tid = pthread_self();
2525 if (!env->me_live_reader) {
2526 rc = mdb_reader_pid(env, Pidset, pid);
2529 env->me_live_reader = 1;
2533 nr = ti->mti_numreaders;
2534 for (i=0; i<nr; i++)
2535 if (ti->mti_readers[i].mr_pid == 0)
2537 if (i == env->me_maxreaders) {
2538 UNLOCK_MUTEX_R(env);
2539 return MDB_READERS_FULL;
2541 ti->mti_readers[i].mr_pid = pid;
2542 ti->mti_readers[i].mr_tid = tid;
2544 ti->mti_numreaders = ++nr;
2545 /* Save numreaders for un-mutexed mdb_env_close() */
2546 env->me_numreaders = nr;
2547 UNLOCK_MUTEX_R(env);
2549 r = &ti->mti_readers[i];
2550 new_notls = (env->me_flags & MDB_NOTLS);
2551 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2556 do /* LY: Retry on a race, ITS#7970. */
2557 r->mr_txnid = ti->mti_txnid;
2558 while(r->mr_txnid != ti->mti_txnid);
2559 txn->mt_txnid = r->mr_txnid;
2560 txn->mt_u.reader = r;
2561 meta = env->me_metas[txn->mt_txnid & 1];
2563 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2568 txn->mt_txnid = ti->mti_txnid;
2569 meta = env->me_metas[txn->mt_txnid & 1];
2571 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2572 txn->mt_txnid = meta->mm_txnid;
2576 if (txn->mt_txnid == mdb_debug_start)
2579 txn->mt_child = NULL;
2580 txn->mt_loose_pgs = NULL;
2581 txn->mt_loose_count = 0;
2582 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2583 txn->mt_u.dirty_list = env->me_dirty_list;
2584 txn->mt_u.dirty_list[0].mid = 0;
2585 txn->mt_free_pgs = env->me_free_pgs;
2586 txn->mt_free_pgs[0] = 0;
2587 txn->mt_spill_pgs = NULL;
2589 memcpy(txn->mt_dbiseqs, env->me_dbiseqs, env->me_maxdbs * sizeof(unsigned int));
2592 /* Copy the DB info and flags */
2593 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2595 /* Moved to here to avoid a data race in read TXNs */
2596 txn->mt_next_pgno = meta->mm_last_pg+1;
2598 txn->mt_flags = flags;
2601 txn->mt_numdbs = env->me_numdbs;
2602 for (i=2; i<txn->mt_numdbs; i++) {
2603 x = env->me_dbflags[i];
2604 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2605 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2607 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2609 if (env->me_maxpg < txn->mt_next_pgno) {
2610 mdb_txn_reset0(txn, "renew0-mapfail");
2612 txn->mt_u.reader->mr_pid = 0;
2613 txn->mt_u.reader = NULL;
2615 return MDB_MAP_RESIZED;
2622 mdb_txn_renew(MDB_txn *txn)
2626 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2629 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2630 DPUTS("environment had fatal error, must shutdown!");
2634 rc = mdb_txn_renew0(txn);
2635 if (rc == MDB_SUCCESS) {
2636 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2637 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2638 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2644 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2648 int rc, size, tsize = sizeof(MDB_txn);
2650 if (env->me_flags & MDB_FATAL_ERROR) {
2651 DPUTS("environment had fatal error, must shutdown!");
2654 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2657 /* Nested transactions: Max 1 child, write txns only, no writemap */
2658 if (parent->mt_child ||
2659 (flags & MDB_RDONLY) ||
2660 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2661 (env->me_flags & MDB_WRITEMAP))
2663 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2665 tsize = sizeof(MDB_ntxn);
2668 if (!(flags & MDB_RDONLY)) {
2670 txn = env->me_txn0; /* just reuse preallocated write txn */
2673 /* child txns use own copy of cursors */
2674 size += env->me_maxdbs * sizeof(MDB_cursor *);
2676 size += env->me_maxdbs * (sizeof(MDB_db)+1);
2678 if ((txn = calloc(1, size)) == NULL) {
2679 DPRINTF(("calloc: %s", strerror(errno)));
2682 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2683 if (flags & MDB_RDONLY) {
2684 txn->mt_flags |= MDB_TXN_RDONLY;
2685 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2686 txn->mt_dbiseqs = env->me_dbiseqs;
2688 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2690 txn->mt_dbiseqs = parent->mt_dbiseqs;
2691 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2693 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
2694 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
2702 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2703 if (!txn->mt_u.dirty_list ||
2704 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2706 free(txn->mt_u.dirty_list);
2710 txn->mt_txnid = parent->mt_txnid;
2711 txn->mt_dirty_room = parent->mt_dirty_room;
2712 txn->mt_u.dirty_list[0].mid = 0;
2713 txn->mt_spill_pgs = NULL;
2714 txn->mt_next_pgno = parent->mt_next_pgno;
2715 parent->mt_child = txn;
2716 txn->mt_parent = parent;
2717 txn->mt_numdbs = parent->mt_numdbs;
2718 txn->mt_flags = parent->mt_flags;
2719 txn->mt_dbxs = parent->mt_dbxs;
2720 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2721 /* Copy parent's mt_dbflags, but clear DB_NEW */
2722 for (i=0; i<txn->mt_numdbs; i++)
2723 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2725 ntxn = (MDB_ntxn *)txn;
2726 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2727 if (env->me_pghead) {
2728 size = MDB_IDL_SIZEOF(env->me_pghead);
2729 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2731 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2736 rc = mdb_cursor_shadow(parent, txn);
2738 mdb_txn_reset0(txn, "beginchild-fail");
2740 rc = mdb_txn_renew0(txn);
2743 if (txn != env->me_txn0)
2747 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2748 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2749 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2756 mdb_txn_env(MDB_txn *txn)
2758 if(!txn) return NULL;
2762 /** Export or close DBI handles opened in this txn. */
2764 mdb_dbis_update(MDB_txn *txn, int keep)
2767 MDB_dbi n = txn->mt_numdbs;
2768 MDB_env *env = txn->mt_env;
2769 unsigned char *tdbflags = txn->mt_dbflags;
2771 for (i = n; --i >= 2;) {
2772 if (tdbflags[i] & DB_NEW) {
2774 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2776 char *ptr = env->me_dbxs[i].md_name.mv_data;
2778 env->me_dbxs[i].md_name.mv_data = NULL;
2779 env->me_dbxs[i].md_name.mv_size = 0;
2780 env->me_dbflags[i] = 0;
2781 env->me_dbiseqs[i]++;
2787 if (keep && env->me_numdbs < n)
2791 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2792 * May be called twice for readonly txns: First reset it, then abort.
2793 * @param[in] txn the transaction handle to reset
2794 * @param[in] act why the transaction is being reset
2797 mdb_txn_reset0(MDB_txn *txn, const char *act)
2799 MDB_env *env = txn->mt_env;
2801 /* Close any DBI handles opened in this txn */
2802 mdb_dbis_update(txn, 0);
2804 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2805 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2806 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2808 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2809 if (txn->mt_u.reader) {
2810 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2811 if (!(env->me_flags & MDB_NOTLS))
2812 txn->mt_u.reader = NULL; /* txn does not own reader */
2814 txn->mt_numdbs = 0; /* close nothing if called again */
2815 txn->mt_dbxs = NULL; /* mark txn as reset */
2817 pgno_t *pghead = env->me_pghead;
2819 mdb_cursors_close(txn, 0);
2820 if (!(env->me_flags & MDB_WRITEMAP)) {
2821 mdb_dlist_free(txn);
2824 if (!txn->mt_parent) {
2825 mdb_midl_shrink(&txn->mt_free_pgs);
2826 env->me_free_pgs = txn->mt_free_pgs;
2828 env->me_pghead = NULL;
2832 /* The writer mutex was locked in mdb_txn_begin. */
2834 UNLOCK_MUTEX_W(env);
2836 txn->mt_parent->mt_child = NULL;
2837 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2838 mdb_midl_free(txn->mt_free_pgs);
2839 mdb_midl_free(txn->mt_spill_pgs);
2840 free(txn->mt_u.dirty_list);
2843 mdb_midl_free(pghead);
2848 mdb_txn_reset(MDB_txn *txn)
2853 /* This call is only valid for read-only txns */
2854 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2857 mdb_txn_reset0(txn, "reset");
2861 mdb_txn_abort(MDB_txn *txn)
2867 mdb_txn_abort(txn->mt_child);
2869 mdb_txn_reset0(txn, "abort");
2870 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2871 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2872 txn->mt_u.reader->mr_pid = 0;
2874 if (txn != txn->mt_env->me_txn0)
2878 /** Save the freelist as of this transaction to the freeDB.
2879 * This changes the freelist. Keep trying until it stabilizes.
2882 mdb_freelist_save(MDB_txn *txn)
2884 /* env->me_pghead[] can grow and shrink during this call.
2885 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2886 * Page numbers cannot disappear from txn->mt_free_pgs[].
2889 MDB_env *env = txn->mt_env;
2890 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2891 txnid_t pglast = 0, head_id = 0;
2892 pgno_t freecnt = 0, *free_pgs, *mop;
2893 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2895 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2897 if (env->me_pghead) {
2898 /* Make sure first page of freeDB is touched and on freelist */
2899 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2900 if (rc && rc != MDB_NOTFOUND)
2904 if (!env->me_pghead && txn->mt_loose_pgs) {
2905 /* Put loose page numbers in mt_free_pgs, since
2906 * we may be unable to return them to me_pghead.
2908 MDB_page *mp = txn->mt_loose_pgs;
2909 if ((rc = mdb_midl_need(&txn->mt_free_pgs, txn->mt_loose_count)) != 0)
2911 for (; mp; mp = NEXT_LOOSE_PAGE(mp))
2912 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2913 txn->mt_loose_pgs = NULL;
2914 txn->mt_loose_count = 0;
2917 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2918 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2919 ? SSIZE_MAX : maxfree_1pg;
2922 /* Come back here after each Put() in case freelist changed */
2927 /* If using records from freeDB which we have not yet
2928 * deleted, delete them and any we reserved for me_pghead.
2930 while (pglast < env->me_pglast) {
2931 rc = mdb_cursor_first(&mc, &key, NULL);
2934 pglast = head_id = *(txnid_t *)key.mv_data;
2935 total_room = head_room = 0;
2936 mdb_tassert(txn, pglast <= env->me_pglast);
2937 rc = mdb_cursor_del(&mc, 0);
2942 /* Save the IDL of pages freed by this txn, to a single record */
2943 if (freecnt < txn->mt_free_pgs[0]) {
2945 /* Make sure last page of freeDB is touched and on freelist */
2946 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2947 if (rc && rc != MDB_NOTFOUND)
2950 free_pgs = txn->mt_free_pgs;
2951 /* Write to last page of freeDB */
2952 key.mv_size = sizeof(txn->mt_txnid);
2953 key.mv_data = &txn->mt_txnid;
2955 freecnt = free_pgs[0];
2956 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2957 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2960 /* Retry if mt_free_pgs[] grew during the Put() */
2961 free_pgs = txn->mt_free_pgs;
2962 } while (freecnt < free_pgs[0]);
2963 mdb_midl_sort(free_pgs);
2964 memcpy(data.mv_data, free_pgs, data.mv_size);
2967 unsigned int i = free_pgs[0];
2968 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2969 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2971 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2977 mop = env->me_pghead;
2978 mop_len = (mop ? mop[0] : 0) + txn->mt_loose_count;
2980 /* Reserve records for me_pghead[]. Split it if multi-page,
2981 * to avoid searching freeDB for a page range. Use keys in
2982 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2984 if (total_room >= mop_len) {
2985 if (total_room == mop_len || --more < 0)
2987 } else if (head_room >= maxfree_1pg && head_id > 1) {
2988 /* Keep current record (overflow page), add a new one */
2992 /* (Re)write {key = head_id, IDL length = head_room} */
2993 total_room -= head_room;
2994 head_room = mop_len - total_room;
2995 if (head_room > maxfree_1pg && head_id > 1) {
2996 /* Overflow multi-page for part of me_pghead */
2997 head_room /= head_id; /* amortize page sizes */
2998 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2999 } else if (head_room < 0) {
3000 /* Rare case, not bothering to delete this record */
3003 key.mv_size = sizeof(head_id);
3004 key.mv_data = &head_id;
3005 data.mv_size = (head_room + 1) * sizeof(pgno_t);
3006 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
3009 /* IDL is initially empty, zero out at least the length */
3010 pgs = (pgno_t *)data.mv_data;
3011 j = head_room > clean_limit ? head_room : 0;
3015 total_room += head_room;
3018 /* Return loose page numbers to me_pghead, though usually none are
3019 * left at this point. The pages themselves remain in dirty_list.
3021 if (txn->mt_loose_pgs) {
3022 MDB_page *mp = txn->mt_loose_pgs;
3023 unsigned count = txn->mt_loose_count;
3025 /* Room for loose pages + temp IDL with same */
3026 if ((rc = mdb_midl_need(&env->me_pghead, 2*count+1)) != 0)
3028 mop = env->me_pghead;
3029 loose = mop + MDB_IDL_ALLOCLEN(mop) - count;
3030 for (count = 0; mp; mp = NEXT_LOOSE_PAGE(mp))
3031 loose[ ++count ] = mp->mp_pgno;
3033 mdb_midl_sort(loose);
3034 mdb_midl_xmerge(mop, loose);
3035 txn->mt_loose_pgs = NULL;
3036 txn->mt_loose_count = 0;
3040 /* Fill in the reserved me_pghead records */
3046 rc = mdb_cursor_first(&mc, &key, &data);
3047 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
3048 txnid_t id = *(txnid_t *)key.mv_data;
3049 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
3052 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
3054 if (len > mop_len) {
3056 data.mv_size = (len + 1) * sizeof(MDB_ID);
3058 data.mv_data = mop -= len;
3061 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
3063 if (rc || !(mop_len -= len))
3070 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
3071 * @param[in] txn the transaction that's being committed
3072 * @param[in] keep number of initial pages in dirty_list to keep dirty.
3073 * @return 0 on success, non-zero on failure.
3076 mdb_page_flush(MDB_txn *txn, int keep)
3078 MDB_env *env = txn->mt_env;
3079 MDB_ID2L dl = txn->mt_u.dirty_list;
3080 unsigned psize = env->me_psize, j;
3081 int i, pagecount = dl[0].mid, rc;
3082 size_t size = 0, pos = 0;
3084 MDB_page *dp = NULL;
3088 struct iovec iov[MDB_COMMIT_PAGES];
3089 ssize_t wpos = 0, wsize = 0, wres;
3090 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
3096 if (env->me_flags & MDB_WRITEMAP) {
3097 /* Clear dirty flags */
3098 while (++i <= pagecount) {
3100 /* Don't flush this page yet */
3101 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3102 dp->mp_flags &= ~P_KEEP;
3106 dp->mp_flags &= ~P_DIRTY;
3111 /* Write the pages */
3113 if (++i <= pagecount) {
3115 /* Don't flush this page yet */
3116 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
3117 dp->mp_flags &= ~P_KEEP;
3122 /* clear dirty flag */
3123 dp->mp_flags &= ~P_DIRTY;
3126 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
3131 /* Windows actually supports scatter/gather I/O, but only on
3132 * unbuffered file handles. Since we're relying on the OS page
3133 * cache for all our data, that's self-defeating. So we just
3134 * write pages one at a time. We use the ov structure to set
3135 * the write offset, to at least save the overhead of a Seek
3138 DPRINTF(("committing page %"Z"u", pgno));
3139 memset(&ov, 0, sizeof(ov));
3140 ov.Offset = pos & 0xffffffff;
3141 ov.OffsetHigh = pos >> 16 >> 16;
3142 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
3144 DPRINTF(("WriteFile: %d", rc));
3148 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3149 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3152 /* Write previous page(s) */
3153 #ifdef MDB_USE_PWRITEV
3154 wres = pwritev(env->me_fd, iov, n, wpos);
3157 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3160 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3164 DPRINTF(("lseek: %s", strerror(rc)));
3167 wres = writev(env->me_fd, iov, n);
3170 if (wres != wsize) {
3175 DPRINTF(("Write error: %s", strerror(rc)));
3177 rc = EIO; /* TODO: Use which error code? */
3178 DPUTS("short write, filesystem full?");
3189 DPRINTF(("committing page %"Z"u", pgno));
3190 next_pos = pos + size;
3191 iov[n].iov_len = size;
3192 iov[n].iov_base = (char *)dp;
3198 /* MIPS has cache coherency issues, this is a no-op everywhere else
3199 * Note: for any size >= on-chip cache size, entire on-chip cache is
3202 CACHEFLUSH(env->me_map, txn->mt_next_pgno * env->me_psize, DCACHE);
3204 for (i = keep; ++i <= pagecount; ) {
3206 /* This is a page we skipped above */
3209 dl[j].mid = dp->mp_pgno;
3212 mdb_dpage_free(env, dp);
3217 txn->mt_dirty_room += i - j;
3223 mdb_txn_commit(MDB_txn *txn)
3229 if (txn == NULL || txn->mt_env == NULL)
3232 if (txn->mt_child) {
3233 rc = mdb_txn_commit(txn->mt_child);
3234 txn->mt_child = NULL;
3241 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3242 mdb_dbis_update(txn, 1);
3243 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3248 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3249 DPUTS("error flag is set, can't commit");
3251 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3256 if (txn->mt_parent) {
3257 MDB_txn *parent = txn->mt_parent;
3261 unsigned x, y, len, ps_len;
3263 /* Append our free list to parent's */
3264 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3267 mdb_midl_free(txn->mt_free_pgs);
3268 /* Failures after this must either undo the changes
3269 * to the parent or set MDB_TXN_ERROR in the parent.
3272 parent->mt_next_pgno = txn->mt_next_pgno;
3273 parent->mt_flags = txn->mt_flags;
3275 /* Merge our cursors into parent's and close them */
3276 mdb_cursors_close(txn, 1);
3278 /* Update parent's DB table. */
3279 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3280 parent->mt_numdbs = txn->mt_numdbs;
3281 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3282 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3283 for (i=2; i<txn->mt_numdbs; i++) {
3284 /* preserve parent's DB_NEW status */
3285 x = parent->mt_dbflags[i] & DB_NEW;
3286 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3289 dst = parent->mt_u.dirty_list;
3290 src = txn->mt_u.dirty_list;
3291 /* Remove anything in our dirty list from parent's spill list */
3292 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3294 pspill[0] = (pgno_t)-1;
3295 /* Mark our dirty pages as deleted in parent spill list */
3296 for (i=0, len=src[0].mid; ++i <= len; ) {
3297 MDB_ID pn = src[i].mid << 1;
3298 while (pn > pspill[x])
3300 if (pn == pspill[x]) {
3305 /* Squash deleted pagenums if we deleted any */
3306 for (x=y; ++x <= ps_len; )
3307 if (!(pspill[x] & 1))
3308 pspill[++y] = pspill[x];
3312 /* Find len = length of merging our dirty list with parent's */
3314 dst[0].mid = 0; /* simplify loops */
3315 if (parent->mt_parent) {
3316 len = x + src[0].mid;
3317 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3318 for (i = x; y && i; y--) {
3319 pgno_t yp = src[y].mid;
3320 while (yp < dst[i].mid)
3322 if (yp == dst[i].mid) {
3327 } else { /* Simplify the above for single-ancestor case */
3328 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3330 /* Merge our dirty list with parent's */
3332 for (i = len; y; dst[i--] = src[y--]) {
3333 pgno_t yp = src[y].mid;
3334 while (yp < dst[x].mid)
3335 dst[i--] = dst[x--];
3336 if (yp == dst[x].mid)
3337 free(dst[x--].mptr);
3339 mdb_tassert(txn, i == x);
3341 free(txn->mt_u.dirty_list);
3342 parent->mt_dirty_room = txn->mt_dirty_room;
3343 if (txn->mt_spill_pgs) {
3344 if (parent->mt_spill_pgs) {
3345 /* TODO: Prevent failure here, so parent does not fail */
3346 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3348 parent->mt_flags |= MDB_TXN_ERROR;
3349 mdb_midl_free(txn->mt_spill_pgs);
3350 mdb_midl_sort(parent->mt_spill_pgs);
3352 parent->mt_spill_pgs = txn->mt_spill_pgs;
3356 /* Append our loose page list to parent's */
3357 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3359 *lp = txn->mt_loose_pgs;
3360 parent->mt_loose_count += txn->mt_loose_count;
3362 parent->mt_child = NULL;
3363 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3368 if (txn != env->me_txn) {
3369 DPUTS("attempt to commit unknown transaction");
3374 mdb_cursors_close(txn, 0);
3376 if (!txn->mt_u.dirty_list[0].mid &&
3377 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3380 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3381 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3383 /* Update DB root pointers */
3384 if (txn->mt_numdbs > 2) {
3388 data.mv_size = sizeof(MDB_db);
3390 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3391 for (i = 2; i < txn->mt_numdbs; i++) {
3392 if (txn->mt_dbflags[i] & DB_DIRTY) {
3393 if (TXN_DBI_CHANGED(txn, i)) {
3397 data.mv_data = &txn->mt_dbs[i];
3398 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data,
3406 rc = mdb_freelist_save(txn);
3410 mdb_midl_free(env->me_pghead);
3411 env->me_pghead = NULL;
3412 mdb_midl_shrink(&txn->mt_free_pgs);
3413 env->me_free_pgs = txn->mt_free_pgs;
3419 if ((rc = mdb_page_flush(txn, 0)) ||
3420 (rc = mdb_env_sync(env, 0)) ||
3421 (rc = mdb_env_write_meta(txn)))
3424 /* Free P_LOOSE pages left behind in dirty_list */
3425 if (!(env->me_flags & MDB_WRITEMAP))
3426 mdb_dlist_free(txn);
3431 mdb_dbis_update(txn, 1);
3434 UNLOCK_MUTEX_W(env);
3435 if (txn != env->me_txn0)
3445 /** Read the environment parameters of a DB environment before
3446 * mapping it into memory.
3447 * @param[in] env the environment handle
3448 * @param[out] meta address of where to store the meta information
3449 * @return 0 on success, non-zero on failure.
3452 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3458 enum { Size = sizeof(pbuf) };
3460 /* We don't know the page size yet, so use a minimum value.
3461 * Read both meta pages so we can use the latest one.
3464 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3468 memset(&ov, 0, sizeof(ov));
3470 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3471 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3474 rc = pread(env->me_fd, &pbuf, Size, off);
3477 if (rc == 0 && off == 0)
3479 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3480 DPRINTF(("read: %s", mdb_strerror(rc)));
3484 p = (MDB_page *)&pbuf;
3486 if (!F_ISSET(p->mp_flags, P_META)) {
3487 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3492 if (m->mm_magic != MDB_MAGIC) {
3493 DPUTS("meta has invalid magic");
3497 if (m->mm_version != MDB_DATA_VERSION) {
3498 DPRINTF(("database is version %u, expected version %u",
3499 m->mm_version, MDB_DATA_VERSION));
3500 return MDB_VERSION_MISMATCH;
3503 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3510 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3512 meta->mm_magic = MDB_MAGIC;
3513 meta->mm_version = MDB_DATA_VERSION;
3514 meta->mm_mapsize = env->me_mapsize;
3515 meta->mm_psize = env->me_psize;
3516 meta->mm_last_pg = 1;
3517 meta->mm_flags = env->me_flags & 0xffff;
3518 meta->mm_flags |= MDB_INTEGERKEY;
3519 meta->mm_dbs[0].md_root = P_INVALID;
3520 meta->mm_dbs[1].md_root = P_INVALID;
3523 /** Write the environment parameters of a freshly created DB environment.
3524 * @param[in] env the environment handle
3525 * @param[out] meta address of where to store the meta information
3526 * @return 0 on success, non-zero on failure.
3529 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3537 memset(&ov, 0, sizeof(ov));
3538 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3540 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3543 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3544 len = pwrite(fd, ptr, size, pos); \
3545 if (len == -1 && ErrCode() == EINTR) continue; \
3546 rc = (len >= 0); break; } while(1)
3549 DPUTS("writing new meta page");
3551 psize = env->me_psize;
3553 mdb_env_init_meta0(env, meta);
3555 p = calloc(2, psize);
3557 p->mp_flags = P_META;
3558 *(MDB_meta *)METADATA(p) = *meta;
3560 q = (MDB_page *)((char *)p + psize);
3562 q->mp_flags = P_META;
3563 *(MDB_meta *)METADATA(q) = *meta;
3565 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3568 else if ((unsigned) len == psize * 2)
3576 /** Update the environment info to commit a transaction.
3577 * @param[in] txn the transaction that's being committed
3578 * @return 0 on success, non-zero on failure.
3581 mdb_env_write_meta(MDB_txn *txn)
3584 MDB_meta meta, metab, *mp;
3587 int rc, len, toggle;
3596 toggle = txn->mt_txnid & 1;
3597 DPRINTF(("writing meta page %d for root page %"Z"u",
3598 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3601 mp = env->me_metas[toggle];
3602 mapsize = env->me_metas[toggle ^ 1]->mm_mapsize;
3603 /* Persist any increases of mapsize config */
3604 if (mapsize < env->me_mapsize)
3605 mapsize = env->me_mapsize;
3607 if (env->me_flags & MDB_WRITEMAP) {
3608 mp->mm_mapsize = mapsize;
3609 mp->mm_dbs[0] = txn->mt_dbs[0];
3610 mp->mm_dbs[1] = txn->mt_dbs[1];
3611 mp->mm_last_pg = txn->mt_next_pgno - 1;
3612 mp->mm_txnid = txn->mt_txnid;
3613 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3614 unsigned meta_size = env->me_psize;
3615 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3618 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3619 if (meta_size < env->me_os_psize)
3620 meta_size += meta_size;
3625 if (MDB_MSYNC(ptr, meta_size, rc)) {
3632 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3633 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3635 meta.mm_mapsize = mapsize;
3636 meta.mm_dbs[0] = txn->mt_dbs[0];
3637 meta.mm_dbs[1] = txn->mt_dbs[1];
3638 meta.mm_last_pg = txn->mt_next_pgno - 1;
3639 meta.mm_txnid = txn->mt_txnid;
3641 off = offsetof(MDB_meta, mm_mapsize);
3642 ptr = (char *)&meta + off;
3643 len = sizeof(MDB_meta) - off;
3645 off += env->me_psize;
3648 /* Write to the SYNC fd */
3649 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3650 env->me_fd : env->me_mfd;
3653 memset(&ov, 0, sizeof(ov));
3655 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3660 rc = pwrite(mfd, ptr, len, off);
3663 rc = rc < 0 ? ErrCode() : EIO;
3668 DPUTS("write failed, disk error?");
3669 /* On a failure, the pagecache still contains the new data.
3670 * Write some old data back, to prevent it from being used.
3671 * Use the non-SYNC fd; we know it will fail anyway.
3673 meta.mm_last_pg = metab.mm_last_pg;
3674 meta.mm_txnid = metab.mm_txnid;
3676 memset(&ov, 0, sizeof(ov));
3678 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3680 r2 = pwrite(env->me_fd, ptr, len, off);
3681 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3684 env->me_flags |= MDB_FATAL_ERROR;
3687 /* MIPS has cache coherency issues, this is a no-op everywhere else */
3688 CACHEFLUSH(env->me_map + off, len, DCACHE);
3690 /* Memory ordering issues are irrelevant; since the entire writer
3691 * is wrapped by wmutex, all of these changes will become visible
3692 * after the wmutex is unlocked. Since the DB is multi-version,
3693 * readers will get consistent data regardless of how fresh or
3694 * how stale their view of these values is.
3697 env->me_txns->mti_txnid = txn->mt_txnid;
3702 /** Check both meta pages to see which one is newer.
3703 * @param[in] env the environment handle
3704 * @return meta toggle (0 or 1).
3707 mdb_env_pick_meta(const MDB_env *env)
3709 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3713 mdb_env_create(MDB_env **env)
3717 e = calloc(1, sizeof(MDB_env));
3721 e->me_maxreaders = DEFAULT_READERS;
3722 e->me_maxdbs = e->me_numdbs = 2;
3723 e->me_fd = INVALID_HANDLE_VALUE;
3724 e->me_lfd = INVALID_HANDLE_VALUE;
3725 e->me_mfd = INVALID_HANDLE_VALUE;
3726 #ifdef MDB_USE_POSIX_SEM
3727 e->me_rmutex = SEM_FAILED;
3728 e->me_wmutex = SEM_FAILED;
3730 e->me_pid = getpid();
3731 GET_PAGESIZE(e->me_os_psize);
3732 VGMEMP_CREATE(e,0,0);
3738 mdb_env_map(MDB_env *env, void *addr)
3741 unsigned int flags = env->me_flags;
3745 LONG sizelo, sizehi;
3748 if (flags & MDB_RDONLY) {
3749 /* Don't set explicit map size, use whatever exists */
3754 msize = env->me_mapsize;
3755 sizelo = msize & 0xffffffff;
3756 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3758 /* Windows won't create mappings for zero length files.
3759 * and won't map more than the file size.
3760 * Just set the maxsize right now.
3762 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3763 || !SetEndOfFile(env->me_fd)
3764 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3768 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3769 PAGE_READWRITE : PAGE_READONLY,
3770 sizehi, sizelo, NULL);
3773 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3774 FILE_MAP_WRITE : FILE_MAP_READ,
3776 rc = env->me_map ? 0 : ErrCode();
3781 int prot = PROT_READ;
3782 if (flags & MDB_WRITEMAP) {
3784 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3787 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3789 if (env->me_map == MAP_FAILED) {
3794 if (flags & MDB_NORDAHEAD) {
3795 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3797 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3799 #ifdef POSIX_MADV_RANDOM
3800 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3801 #endif /* POSIX_MADV_RANDOM */
3802 #endif /* MADV_RANDOM */
3806 /* Can happen because the address argument to mmap() is just a
3807 * hint. mmap() can pick another, e.g. if the range is in use.
3808 * The MAP_FIXED flag would prevent that, but then mmap could
3809 * instead unmap existing pages to make room for the new map.
3811 if (addr && env->me_map != addr)
3812 return EBUSY; /* TODO: Make a new MDB_* error code? */
3814 p = (MDB_page *)env->me_map;
3815 env->me_metas[0] = METADATA(p);
3816 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3822 mdb_env_set_mapsize(MDB_env *env, size_t size)
3824 /* If env is already open, caller is responsible for making
3825 * sure there are no active txns.
3833 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3834 else if (size < env->me_mapsize) {
3835 /* If the configured size is smaller, make sure it's
3836 * still big enough. Silently round up to minimum if not.
3838 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3842 munmap(env->me_map, env->me_mapsize);
3843 env->me_mapsize = size;
3844 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3845 rc = mdb_env_map(env, old);
3849 env->me_mapsize = size;
3851 env->me_maxpg = env->me_mapsize / env->me_psize;
3856 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3860 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3865 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3867 if (env->me_map || readers < 1)
3869 env->me_maxreaders = readers;
3874 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3876 if (!env || !readers)
3878 *readers = env->me_maxreaders;
3883 mdb_fsize(HANDLE fd, size_t *size)
3886 LARGE_INTEGER fsize;
3888 if (!GetFileSizeEx(fd, &fsize))
3891 *size = fsize.QuadPart;
3903 #ifdef BROKEN_FDATASYNC
3904 #include <sys/utsname.h>
3905 #include <sys/vfs.h>
3908 /** Further setup required for opening an LMDB environment
3911 mdb_env_open2(MDB_env *env)
3913 unsigned int flags = env->me_flags;
3914 int i, newenv = 0, rc;
3918 /* See if we should use QueryLimited */
3920 if ((rc & 0xff) > 5)
3921 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3923 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3925 #ifdef BROKEN_FDATASYNC
3926 /* ext3/ext4 fdatasync is broken on some older Linux kernels.
3927 * https://lkml.org/lkml/2012/9/3/83
3928 * Kernels after 3.6-rc6 are known good.
3929 * https://lkml.org/lkml/2012/9/10/556
3930 * See if the DB is on ext3/ext4, then check for new enough kernel
3931 * Kernels 2.6.32.60, 2.6.34.15, 3.2.30, and 3.5.4 are also known
3936 fstatfs(env->me_fd, &st);
3937 while (st.f_type == 0xEF53) {
3941 if (uts.release[0] < '3') {
3942 if (!strncmp(uts.release, "2.6.32.", 7)) {
3943 i = atoi(uts.release+7);
3945 break; /* 2.6.32.60 and newer is OK */
3946 } else if (!strncmp(uts.release, "2.6.34.", 7)) {
3947 i = atoi(uts.release+7);
3949 break; /* 2.6.34.15 and newer is OK */
3951 } else if (uts.release[0] == '3') {
3952 i = atoi(uts.release+2);
3954 break; /* 3.6 and newer is OK */
3956 i = atoi(uts.release+4);
3958 break; /* 3.5.4 and newer is OK */
3959 } else if (i == 2) {
3960 i = atoi(uts.release+4);
3962 break; /* 3.2.30 and newer is OK */
3964 } else { /* 4.x and newer is OK */
3967 env->me_flags |= MDB_FSYNCONLY;
3973 memset(&meta, 0, sizeof(meta));
3975 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3978 DPUTS("new mdbenv");
3980 env->me_psize = env->me_os_psize;
3981 if (env->me_psize > MAX_PAGESIZE)
3982 env->me_psize = MAX_PAGESIZE;
3984 env->me_psize = meta.mm_psize;
3987 /* Was a mapsize configured? */
3988 if (!env->me_mapsize) {
3989 /* If this is a new environment, take the default,
3990 * else use the size recorded in the existing env.
3992 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3993 } else if (env->me_mapsize < meta.mm_mapsize) {
3994 /* If the configured size is smaller, make sure it's
3995 * still big enough. Silently round up to minimum if not.
3997 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3998 if (env->me_mapsize < minsize)
3999 env->me_mapsize = minsize;
4002 rc = mdb_env_map(env, (flags & MDB_FIXEDMAP) ? meta.mm_address : NULL);
4007 if (flags & MDB_FIXEDMAP)
4008 meta.mm_address = env->me_map;
4009 i = mdb_env_init_meta(env, &meta);
4010 if (i != MDB_SUCCESS) {
4015 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
4016 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
4018 #if !(MDB_MAXKEYSIZE)
4019 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
4021 env->me_maxpg = env->me_mapsize / env->me_psize;
4025 int toggle = mdb_env_pick_meta(env);
4026 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
4028 DPRINTF(("opened database version %u, pagesize %u",
4029 env->me_metas[0]->mm_version, env->me_psize));
4030 DPRINTF(("using meta page %d", toggle));
4031 DPRINTF(("depth: %u", db->md_depth));
4032 DPRINTF(("entries: %"Z"u", db->md_entries));
4033 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
4034 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
4035 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
4036 DPRINTF(("root: %"Z"u", db->md_root));
4044 /** Release a reader thread's slot in the reader lock table.
4045 * This function is called automatically when a thread exits.
4046 * @param[in] ptr This points to the slot in the reader lock table.
4049 mdb_env_reader_dest(void *ptr)
4051 MDB_reader *reader = ptr;
4057 /** Junk for arranging thread-specific callbacks on Windows. This is
4058 * necessarily platform and compiler-specific. Windows supports up
4059 * to 1088 keys. Let's assume nobody opens more than 64 environments
4060 * in a single process, for now. They can override this if needed.
4062 #ifndef MAX_TLS_KEYS
4063 #define MAX_TLS_KEYS 64
4065 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
4066 static int mdb_tls_nkeys;
4068 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
4072 case DLL_PROCESS_ATTACH: break;
4073 case DLL_THREAD_ATTACH: break;
4074 case DLL_THREAD_DETACH:
4075 for (i=0; i<mdb_tls_nkeys; i++) {
4076 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
4078 mdb_env_reader_dest(r);
4082 case DLL_PROCESS_DETACH: break;
4087 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4089 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
4093 /* Force some symbol references.
4094 * _tls_used forces the linker to create the TLS directory if not already done
4095 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
4097 #pragma comment(linker, "/INCLUDE:_tls_used")
4098 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
4099 #pragma const_seg(".CRT$XLB")
4100 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
4101 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4104 #pragma comment(linker, "/INCLUDE:__tls_used")
4105 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
4106 #pragma data_seg(".CRT$XLB")
4107 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
4109 #endif /* WIN 32/64 */
4110 #endif /* !__GNUC__ */
4113 /** Downgrade the exclusive lock on the region back to shared */
4115 mdb_env_share_locks(MDB_env *env, int *excl)
4117 int rc = 0, toggle = mdb_env_pick_meta(env);
4119 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
4124 /* First acquire a shared lock. The Unlock will
4125 * then release the existing exclusive lock.
4127 memset(&ov, 0, sizeof(ov));
4128 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4131 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4137 struct flock lock_info;
4138 /* The shared lock replaces the existing lock */
4139 memset((void *)&lock_info, 0, sizeof(lock_info));
4140 lock_info.l_type = F_RDLCK;
4141 lock_info.l_whence = SEEK_SET;
4142 lock_info.l_start = 0;
4143 lock_info.l_len = 1;
4144 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4145 (rc = ErrCode()) == EINTR) ;
4146 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
4153 /** Try to get exclusive lock, otherwise shared.
4154 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
4157 mdb_env_excl_lock(MDB_env *env, int *excl)
4161 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
4165 memset(&ov, 0, sizeof(ov));
4166 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
4173 struct flock lock_info;
4174 memset((void *)&lock_info, 0, sizeof(lock_info));
4175 lock_info.l_type = F_WRLCK;
4176 lock_info.l_whence = SEEK_SET;
4177 lock_info.l_start = 0;
4178 lock_info.l_len = 1;
4179 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
4180 (rc = ErrCode()) == EINTR) ;
4184 # ifdef MDB_USE_POSIX_SEM
4185 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
4188 lock_info.l_type = F_RDLCK;
4189 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
4190 (rc = ErrCode()) == EINTR) ;
4200 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
4202 * @(#) $Revision: 5.1 $
4203 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
4204 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
4206 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
4210 * Please do not copyright this code. This code is in the public domain.
4212 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
4213 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
4214 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
4215 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
4216 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
4217 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
4218 * PERFORMANCE OF THIS SOFTWARE.
4221 * chongo <Landon Curt Noll> /\oo/\
4222 * http://www.isthe.com/chongo/
4224 * Share and Enjoy! :-)
4227 typedef unsigned long long mdb_hash_t;
4228 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
4230 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
4231 * @param[in] val value to hash
4232 * @param[in] hval initial value for hash
4233 * @return 64 bit hash
4235 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
4236 * hval arg on the first call.
4239 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
4241 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
4242 unsigned char *end = s + val->mv_size;
4244 * FNV-1a hash each octet of the string
4247 /* xor the bottom with the current octet */
4248 hval ^= (mdb_hash_t)*s++;
4250 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4251 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4252 (hval << 7) + (hval << 8) + (hval << 40);
4254 /* return our new hash value */
4258 /** Hash the string and output the encoded hash.
4259 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4260 * very short name limits. We don't care about the encoding being reversible,
4261 * we just want to preserve as many bits of the input as possible in a
4262 * small printable string.
4263 * @param[in] str string to hash
4264 * @param[out] encbuf an array of 11 chars to hold the hash
4266 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4269 mdb_pack85(unsigned long l, char *out)
4273 for (i=0; i<5; i++) {
4274 *out++ = mdb_a85[l % 85];
4280 mdb_hash_enc(MDB_val *val, char *encbuf)
4282 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4284 mdb_pack85(h, encbuf);
4285 mdb_pack85(h>>32, encbuf+5);
4290 /** Open and/or initialize the lock region for the environment.
4291 * @param[in] env The LMDB environment.
4292 * @param[in] lpath The pathname of the file used for the lock region.
4293 * @param[in] mode The Unix permissions for the file, if we create it.
4294 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4295 * @return 0 on success, non-zero on failure.
4298 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4301 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4303 # define MDB_ERRCODE_ROFS EROFS
4304 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4305 # define MDB_CLOEXEC O_CLOEXEC
4308 # define MDB_CLOEXEC 0
4315 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4316 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4317 FILE_ATTRIBUTE_NORMAL, NULL);
4319 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4321 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4323 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4328 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4329 /* Lose record locks when exec*() */
4330 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4331 fcntl(env->me_lfd, F_SETFD, fdflags);
4334 if (!(env->me_flags & MDB_NOTLS)) {
4335 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4338 env->me_flags |= MDB_ENV_TXKEY;
4340 /* Windows TLS callbacks need help finding their TLS info. */
4341 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4345 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4349 /* Try to get exclusive lock. If we succeed, then
4350 * nobody is using the lock region and we should initialize it.
4352 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4355 size = GetFileSize(env->me_lfd, NULL);
4357 size = lseek(env->me_lfd, 0, SEEK_END);
4358 if (size == -1) goto fail_errno;
4360 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4361 if (size < rsize && *excl > 0) {
4363 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4364 || !SetEndOfFile(env->me_lfd))
4367 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4371 size = rsize - sizeof(MDB_txninfo);
4372 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4377 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4379 if (!mh) goto fail_errno;
4380 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4382 if (!env->me_txns) goto fail_errno;
4384 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4386 if (m == MAP_FAILED) goto fail_errno;
4392 BY_HANDLE_FILE_INFORMATION stbuf;
4401 if (!mdb_sec_inited) {
4402 InitializeSecurityDescriptor(&mdb_null_sd,
4403 SECURITY_DESCRIPTOR_REVISION);
4404 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4405 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4406 mdb_all_sa.bInheritHandle = FALSE;
4407 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4410 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4411 idbuf.volume = stbuf.dwVolumeSerialNumber;
4412 idbuf.nhigh = stbuf.nFileIndexHigh;
4413 idbuf.nlow = stbuf.nFileIndexLow;
4414 val.mv_data = &idbuf;
4415 val.mv_size = sizeof(idbuf);
4416 mdb_hash_enc(&val, encbuf);
4417 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4418 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4419 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4420 if (!env->me_rmutex) goto fail_errno;
4421 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4422 if (!env->me_wmutex) goto fail_errno;
4423 #elif defined(MDB_USE_POSIX_SEM)
4432 #if defined(__NetBSD__)
4433 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4435 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4436 idbuf.dev = stbuf.st_dev;
4437 idbuf.ino = stbuf.st_ino;
4438 val.mv_data = &idbuf;
4439 val.mv_size = sizeof(idbuf);
4440 mdb_hash_enc(&val, encbuf);
4441 #ifdef MDB_SHORT_SEMNAMES
4442 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4444 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4445 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4446 /* Clean up after a previous run, if needed: Try to
4447 * remove both semaphores before doing anything else.
4449 sem_unlink(env->me_txns->mti_rmname);
4450 sem_unlink(env->me_txns->mti_wmname);
4451 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4452 O_CREAT|O_EXCL, mode, 1);
4453 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4454 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4455 O_CREAT|O_EXCL, mode, 1);
4456 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4457 #else /* MDB_USE_POSIX_SEM */
4458 pthread_mutexattr_t mattr;
4460 if ((rc = pthread_mutexattr_init(&mattr))
4461 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4462 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4463 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4465 pthread_mutexattr_destroy(&mattr);
4466 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4468 env->me_txns->mti_magic = MDB_MAGIC;
4469 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4470 env->me_txns->mti_txnid = 0;
4471 env->me_txns->mti_numreaders = 0;
4474 if (env->me_txns->mti_magic != MDB_MAGIC) {
4475 DPUTS("lock region has invalid magic");
4479 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4480 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4481 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4482 rc = MDB_VERSION_MISMATCH;
4486 if (rc && rc != EACCES && rc != EAGAIN) {
4490 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4491 if (!env->me_rmutex) goto fail_errno;
4492 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4493 if (!env->me_wmutex) goto fail_errno;
4494 #elif defined(MDB_USE_POSIX_SEM)
4495 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4496 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4497 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4498 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4509 /** The name of the lock file in the DB environment */
4510 #define LOCKNAME "/lock.mdb"
4511 /** The name of the data file in the DB environment */
4512 #define DATANAME "/data.mdb"
4513 /** The suffix of the lock file when no subdir is used */
4514 #define LOCKSUFF "-lock"
4515 /** Only a subset of the @ref mdb_env flags can be changed
4516 * at runtime. Changing other flags requires closing the
4517 * environment and re-opening it with the new flags.
4519 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4520 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4521 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4523 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4524 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4528 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4530 int oflags, rc, len, excl = -1;
4531 char *lpath, *dpath;
4533 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4537 if (flags & MDB_NOSUBDIR) {
4538 rc = len + sizeof(LOCKSUFF) + len + 1;
4540 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4545 if (flags & MDB_NOSUBDIR) {
4546 dpath = lpath + len + sizeof(LOCKSUFF);
4547 sprintf(lpath, "%s" LOCKSUFF, path);
4548 strcpy(dpath, path);
4550 dpath = lpath + len + sizeof(LOCKNAME);
4551 sprintf(lpath, "%s" LOCKNAME, path);
4552 sprintf(dpath, "%s" DATANAME, path);
4556 flags |= env->me_flags;
4557 if (flags & MDB_RDONLY) {
4558 /* silently ignore WRITEMAP when we're only getting read access */
4559 flags &= ~MDB_WRITEMAP;
4561 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4562 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4565 env->me_flags = flags |= MDB_ENV_ACTIVE;
4569 env->me_path = strdup(path);
4570 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4571 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4572 env->me_dbiseqs = calloc(env->me_maxdbs, sizeof(unsigned int));
4573 if (!(env->me_dbxs && env->me_path && env->me_dbflags && env->me_dbiseqs)) {
4578 /* For RDONLY, get lockfile after we know datafile exists */
4579 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4580 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4586 if (F_ISSET(flags, MDB_RDONLY)) {
4587 oflags = GENERIC_READ;
4588 len = OPEN_EXISTING;
4590 oflags = GENERIC_READ|GENERIC_WRITE;
4593 mode = FILE_ATTRIBUTE_NORMAL;
4594 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4595 NULL, len, mode, NULL);
4597 if (F_ISSET(flags, MDB_RDONLY))
4600 oflags = O_RDWR | O_CREAT;
4602 env->me_fd = open(dpath, oflags, mode);
4604 if (env->me_fd == INVALID_HANDLE_VALUE) {
4609 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4610 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4615 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4616 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4617 env->me_mfd = env->me_fd;
4619 /* Synchronous fd for meta writes. Needed even with
4620 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4623 len = OPEN_EXISTING;
4624 env->me_mfd = CreateFile(dpath, oflags,
4625 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4626 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4629 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4631 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4636 DPRINTF(("opened dbenv %p", (void *) env));
4638 rc = mdb_env_share_locks(env, &excl);
4642 if (!((flags & MDB_RDONLY) ||
4643 (env->me_pbuf = calloc(1, env->me_psize))))
4645 if (!(flags & MDB_RDONLY)) {
4647 int tsize = sizeof(MDB_txn), size = tsize + env->me_maxdbs *
4648 (sizeof(MDB_db)+sizeof(MDB_cursor *)+sizeof(unsigned int)+1);
4649 txn = calloc(1, size);
4651 txn->mt_dbs = (MDB_db *)((char *)txn + tsize);
4652 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
4653 txn->mt_dbiseqs = (unsigned int *)(txn->mt_cursors + env->me_maxdbs);
4654 txn->mt_dbflags = (unsigned char *)(txn->mt_dbiseqs + env->me_maxdbs);
4656 txn->mt_dbxs = env->me_dbxs;
4666 mdb_env_close0(env, excl);
4672 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4674 mdb_env_close0(MDB_env *env, int excl)
4678 if (!(env->me_flags & MDB_ENV_ACTIVE))
4681 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4683 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4684 free(env->me_dbxs[i].md_name.mv_data);
4689 free(env->me_dbiseqs);
4690 free(env->me_dbflags);
4692 free(env->me_dirty_list);
4694 mdb_midl_free(env->me_free_pgs);
4696 if (env->me_flags & MDB_ENV_TXKEY) {
4697 pthread_key_delete(env->me_txkey);
4699 /* Delete our key from the global list */
4700 for (i=0; i<mdb_tls_nkeys; i++)
4701 if (mdb_tls_keys[i] == env->me_txkey) {
4702 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4710 munmap(env->me_map, env->me_mapsize);
4712 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4713 (void) close(env->me_mfd);
4714 if (env->me_fd != INVALID_HANDLE_VALUE)
4715 (void) close(env->me_fd);
4717 MDB_PID_T pid = env->me_pid;
4718 /* Clearing readers is done in this function because
4719 * me_txkey with its destructor must be disabled first.
4721 for (i = env->me_numreaders; --i >= 0; )
4722 if (env->me_txns->mti_readers[i].mr_pid == pid)
4723 env->me_txns->mti_readers[i].mr_pid = 0;
4725 if (env->me_rmutex) {
4726 CloseHandle(env->me_rmutex);
4727 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4729 /* Windows automatically destroys the mutexes when
4730 * the last handle closes.
4732 #elif defined(MDB_USE_POSIX_SEM)
4733 if (env->me_rmutex != SEM_FAILED) {
4734 sem_close(env->me_rmutex);
4735 if (env->me_wmutex != SEM_FAILED)
4736 sem_close(env->me_wmutex);
4737 /* If we have the filelock: If we are the
4738 * only remaining user, clean up semaphores.
4741 mdb_env_excl_lock(env, &excl);
4743 sem_unlink(env->me_txns->mti_rmname);
4744 sem_unlink(env->me_txns->mti_wmname);
4748 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4750 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4753 /* Unlock the lockfile. Windows would have unlocked it
4754 * after closing anyway, but not necessarily at once.
4756 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4759 (void) close(env->me_lfd);
4762 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4767 mdb_env_close(MDB_env *env)
4774 VGMEMP_DESTROY(env);
4775 while ((dp = env->me_dpages) != NULL) {
4776 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4777 env->me_dpages = dp->mp_next;
4781 mdb_env_close0(env, 0);
4785 /** Compare two items pointing at aligned size_t's */
4787 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4789 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4790 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4793 /** Compare two items pointing at aligned unsigned int's.
4795 * This is also set as #MDB_INTEGERDUP|#MDB_DUPFIXED's #MDB_dbx.%md_dcmp,
4796 * but #mdb_cmp_clong() is called instead if the data type is size_t.
4799 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4801 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4802 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4805 /** Compare two items pointing at unsigned ints of unknown alignment.
4806 * Nodes and keys are guaranteed to be 2-byte aligned.
4809 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4811 #if BYTE_ORDER == LITTLE_ENDIAN
4812 unsigned short *u, *c;
4815 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4816 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4819 } while(!x && u > (unsigned short *)a->mv_data);
4822 unsigned short *u, *c, *end;
4825 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4826 u = (unsigned short *)a->mv_data;
4827 c = (unsigned short *)b->mv_data;
4830 } while(!x && u < end);
4835 /** Compare two items lexically */
4837 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4844 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4850 diff = memcmp(a->mv_data, b->mv_data, len);
4851 return diff ? diff : len_diff<0 ? -1 : len_diff;
4854 /** Compare two items in reverse byte order */
4856 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4858 const unsigned char *p1, *p2, *p1_lim;
4862 p1_lim = (const unsigned char *)a->mv_data;
4863 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4864 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4866 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4872 while (p1 > p1_lim) {
4873 diff = *--p1 - *--p2;
4877 return len_diff<0 ? -1 : len_diff;
4880 /** Search for key within a page, using binary search.
4881 * Returns the smallest entry larger or equal to the key.
4882 * If exactp is non-null, stores whether the found entry was an exact match
4883 * in *exactp (1 or 0).
4884 * Updates the cursor index with the index of the found entry.
4885 * If no entry larger or equal to the key is found, returns NULL.
4888 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4890 unsigned int i = 0, nkeys;
4893 MDB_page *mp = mc->mc_pg[mc->mc_top];
4894 MDB_node *node = NULL;
4899 nkeys = NUMKEYS(mp);
4901 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4902 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4905 low = IS_LEAF(mp) ? 0 : 1;
4907 cmp = mc->mc_dbx->md_cmp;
4909 /* Branch pages have no data, so if using integer keys,
4910 * alignment is guaranteed. Use faster mdb_cmp_int.
4912 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4913 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4920 nodekey.mv_size = mc->mc_db->md_pad;
4921 node = NODEPTR(mp, 0); /* fake */
4922 while (low <= high) {
4923 i = (low + high) >> 1;
4924 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4925 rc = cmp(key, &nodekey);
4926 DPRINTF(("found leaf index %u [%s], rc = %i",
4927 i, DKEY(&nodekey), rc));
4936 while (low <= high) {
4937 i = (low + high) >> 1;
4939 node = NODEPTR(mp, i);
4940 nodekey.mv_size = NODEKSZ(node);
4941 nodekey.mv_data = NODEKEY(node);
4943 rc = cmp(key, &nodekey);
4946 DPRINTF(("found leaf index %u [%s], rc = %i",
4947 i, DKEY(&nodekey), rc));
4949 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4950 i, DKEY(&nodekey), NODEPGNO(node), rc));
4961 if (rc > 0) { /* Found entry is less than the key. */
4962 i++; /* Skip to get the smallest entry larger than key. */
4964 node = NODEPTR(mp, i);
4967 *exactp = (rc == 0 && nkeys > 0);
4968 /* store the key index */
4969 mc->mc_ki[mc->mc_top] = i;
4971 /* There is no entry larger or equal to the key. */
4974 /* nodeptr is fake for LEAF2 */
4980 mdb_cursor_adjust(MDB_cursor *mc, func)
4984 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4985 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4992 /** Pop a page off the top of the cursor's stack. */
4994 mdb_cursor_pop(MDB_cursor *mc)
4998 MDB_page *top = mc->mc_pg[mc->mc_top];
5004 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
5005 DDBI(mc), (void *) mc));
5009 /** Push a page onto the top of the cursor's stack. */
5011 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
5013 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
5014 DDBI(mc), (void *) mc));
5016 if (mc->mc_snum >= CURSOR_STACK) {
5017 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5018 return MDB_CURSOR_FULL;
5021 mc->mc_top = mc->mc_snum++;
5022 mc->mc_pg[mc->mc_top] = mp;
5023 mc->mc_ki[mc->mc_top] = 0;
5028 /** Find the address of the page corresponding to a given page number.
5029 * @param[in] txn the transaction for this access.
5030 * @param[in] pgno the page number for the page to retrieve.
5031 * @param[out] ret address of a pointer where the page's address will be stored.
5032 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
5033 * @return 0 on success, non-zero on failure.
5036 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
5038 MDB_env *env = txn->mt_env;
5042 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
5046 MDB_ID2L dl = tx2->mt_u.dirty_list;
5048 /* Spilled pages were dirtied in this txn and flushed
5049 * because the dirty list got full. Bring this page
5050 * back in from the map (but don't unspill it here,
5051 * leave that unless page_touch happens again).
5053 if (tx2->mt_spill_pgs) {
5054 MDB_ID pn = pgno << 1;
5055 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
5056 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
5057 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5062 unsigned x = mdb_mid2l_search(dl, pgno);
5063 if (x <= dl[0].mid && dl[x].mid == pgno) {
5069 } while ((tx2 = tx2->mt_parent) != NULL);
5072 if (pgno < txn->mt_next_pgno) {
5074 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
5076 DPRINTF(("page %"Z"u not found", pgno));
5077 txn->mt_flags |= MDB_TXN_ERROR;
5078 return MDB_PAGE_NOTFOUND;
5088 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
5089 * The cursor is at the root page, set up the rest of it.
5092 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
5094 MDB_page *mp = mc->mc_pg[mc->mc_top];
5098 while (IS_BRANCH(mp)) {
5102 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
5103 mdb_cassert(mc, NUMKEYS(mp) > 1);
5104 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
5106 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
5108 if (flags & MDB_PS_LAST)
5109 i = NUMKEYS(mp) - 1;
5112 node = mdb_node_search(mc, key, &exact);
5114 i = NUMKEYS(mp) - 1;
5116 i = mc->mc_ki[mc->mc_top];
5118 mdb_cassert(mc, i > 0);
5122 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
5125 mdb_cassert(mc, i < NUMKEYS(mp));
5126 node = NODEPTR(mp, i);
5128 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5131 mc->mc_ki[mc->mc_top] = i;
5132 if ((rc = mdb_cursor_push(mc, mp)))
5135 if (flags & MDB_PS_MODIFY) {
5136 if ((rc = mdb_page_touch(mc)) != 0)
5138 mp = mc->mc_pg[mc->mc_top];
5143 DPRINTF(("internal error, index points to a %02X page!?",
5145 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
5146 return MDB_CORRUPTED;
5149 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
5150 key ? DKEY(key) : "null"));
5151 mc->mc_flags |= C_INITIALIZED;
5152 mc->mc_flags &= ~C_EOF;
5157 /** Search for the lowest key under the current branch page.
5158 * This just bypasses a NUMKEYS check in the current page
5159 * before calling mdb_page_search_root(), because the callers
5160 * are all in situations where the current page is known to
5164 mdb_page_search_lowest(MDB_cursor *mc)
5166 MDB_page *mp = mc->mc_pg[mc->mc_top];
5167 MDB_node *node = NODEPTR(mp, 0);
5170 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
5173 mc->mc_ki[mc->mc_top] = 0;
5174 if ((rc = mdb_cursor_push(mc, mp)))
5176 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
5179 /** Search for the page a given key should be in.
5180 * Push it and its parent pages on the cursor stack.
5181 * @param[in,out] mc the cursor for this operation.
5182 * @param[in] key the key to search for, or NULL for first/last page.
5183 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
5184 * are touched (updated with new page numbers).
5185 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
5186 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
5187 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
5188 * @return 0 on success, non-zero on failure.
5191 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
5196 /* Make sure the txn is still viable, then find the root from
5197 * the txn's db table and set it as the root of the cursor's stack.
5199 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
5200 DPUTS("transaction has failed, must abort");
5203 /* Make sure we're using an up-to-date root */
5204 if (*mc->mc_dbflag & DB_STALE) {
5206 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
5208 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
5209 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
5216 MDB_node *leaf = mdb_node_search(&mc2,
5217 &mc->mc_dbx->md_name, &exact);
5219 return MDB_NOTFOUND;
5220 if ((leaf->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
5221 return MDB_INCOMPATIBLE; /* not a named DB */
5222 rc = mdb_node_read(mc->mc_txn, leaf, &data);
5225 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
5227 /* The txn may not know this DBI, or another process may
5228 * have dropped and recreated the DB with other flags.
5230 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
5231 return MDB_INCOMPATIBLE;
5232 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
5234 *mc->mc_dbflag &= ~DB_STALE;
5236 root = mc->mc_db->md_root;
5238 if (root == P_INVALID) { /* Tree is empty. */
5239 DPUTS("tree is empty");
5240 return MDB_NOTFOUND;
5244 mdb_cassert(mc, root > 1);
5245 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
5246 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
5252 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
5253 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
5255 if (flags & MDB_PS_MODIFY) {
5256 if ((rc = mdb_page_touch(mc)))
5260 if (flags & MDB_PS_ROOTONLY)
5263 return mdb_page_search_root(mc, key, flags);
5267 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
5269 MDB_txn *txn = mc->mc_txn;
5270 pgno_t pg = mp->mp_pgno;
5271 unsigned x = 0, ovpages = mp->mp_pages;
5272 MDB_env *env = txn->mt_env;
5273 MDB_IDL sl = txn->mt_spill_pgs;
5274 MDB_ID pn = pg << 1;
5277 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5278 /* If the page is dirty or on the spill list we just acquired it,
5279 * so we should give it back to our current free list, if any.
5280 * Otherwise put it onto the list of pages we freed in this txn.
5282 * Won't create me_pghead: me_pglast must be inited along with it.
5283 * Unsupported in nested txns: They would need to hide the page
5284 * range in ancestor txns' dirty and spilled lists.
5286 if (env->me_pghead &&
5288 ((mp->mp_flags & P_DIRTY) ||
5289 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5293 MDB_ID2 *dl, ix, iy;
5294 rc = mdb_midl_need(&env->me_pghead, ovpages);
5297 if (!(mp->mp_flags & P_DIRTY)) {
5298 /* This page is no longer spilled */
5305 /* Remove from dirty list */
5306 dl = txn->mt_u.dirty_list;
5308 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5314 mdb_cassert(mc, x > 1);
5316 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5317 txn->mt_flags |= MDB_TXN_ERROR;
5318 return MDB_CORRUPTED;
5321 if (!(env->me_flags & MDB_WRITEMAP))
5322 mdb_dpage_free(env, mp);
5324 /* Insert in me_pghead */
5325 mop = env->me_pghead;
5326 j = mop[0] + ovpages;
5327 for (i = mop[0]; i && mop[i] < pg; i--)
5333 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5337 mc->mc_db->md_overflow_pages -= ovpages;
5341 /** Return the data associated with a given node.
5342 * @param[in] txn The transaction for this operation.
5343 * @param[in] leaf The node being read.
5344 * @param[out] data Updated to point to the node's data.
5345 * @return 0 on success, non-zero on failure.
5348 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5350 MDB_page *omp; /* overflow page */
5354 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5355 data->mv_size = NODEDSZ(leaf);
5356 data->mv_data = NODEDATA(leaf);
5360 /* Read overflow data.
5362 data->mv_size = NODEDSZ(leaf);
5363 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5364 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5365 DPRINTF(("read overflow page %"Z"u failed", pgno));
5368 data->mv_data = METADATA(omp);
5374 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5375 MDB_val *key, MDB_val *data)
5382 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5384 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5387 if (txn->mt_flags & MDB_TXN_ERROR)
5390 mdb_cursor_init(&mc, txn, dbi, &mx);
5391 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5394 /** Find a sibling for a page.
5395 * Replaces the page at the top of the cursor's stack with the
5396 * specified sibling, if one exists.
5397 * @param[in] mc The cursor for this operation.
5398 * @param[in] move_right Non-zero if the right sibling is requested,
5399 * otherwise the left sibling.
5400 * @return 0 on success, non-zero on failure.
5403 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5409 if (mc->mc_snum < 2) {
5410 return MDB_NOTFOUND; /* root has no siblings */
5414 DPRINTF(("parent page is page %"Z"u, index %u",
5415 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5417 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5418 : (mc->mc_ki[mc->mc_top] == 0)) {
5419 DPRINTF(("no more keys left, moving to %s sibling",
5420 move_right ? "right" : "left"));
5421 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5422 /* undo cursor_pop before returning */
5429 mc->mc_ki[mc->mc_top]++;
5431 mc->mc_ki[mc->mc_top]--;
5432 DPRINTF(("just moving to %s index key %u",
5433 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5435 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5437 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5438 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5439 /* mc will be inconsistent if caller does mc_snum++ as above */
5440 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5444 mdb_cursor_push(mc, mp);
5446 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5451 /** Move the cursor to the next data item. */
5453 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5459 if (mc->mc_flags & C_EOF) {
5460 return MDB_NOTFOUND;
5463 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5465 mp = mc->mc_pg[mc->mc_top];
5467 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5468 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5469 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5470 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5471 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5472 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5473 if (rc == MDB_SUCCESS)
5474 MDB_GET_KEY(leaf, key);
5479 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5480 if (op == MDB_NEXT_DUP)
5481 return MDB_NOTFOUND;
5485 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5486 mdb_dbg_pgno(mp), (void *) mc));
5487 if (mc->mc_flags & C_DEL)
5490 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5491 DPUTS("=====> move to next sibling page");
5492 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5493 mc->mc_flags |= C_EOF;
5496 mp = mc->mc_pg[mc->mc_top];
5497 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5499 mc->mc_ki[mc->mc_top]++;
5502 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5503 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5506 key->mv_size = mc->mc_db->md_pad;
5507 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5511 mdb_cassert(mc, IS_LEAF(mp));
5512 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5514 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5515 mdb_xcursor_init1(mc, leaf);
5518 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5521 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5522 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5523 if (rc != MDB_SUCCESS)
5528 MDB_GET_KEY(leaf, key);
5532 /** Move the cursor to the previous data item. */
5534 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5540 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5542 mp = mc->mc_pg[mc->mc_top];
5544 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5545 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5546 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5547 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5548 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5549 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5550 if (rc == MDB_SUCCESS) {
5551 MDB_GET_KEY(leaf, key);
5552 mc->mc_flags &= ~C_EOF;
5558 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5559 if (op == MDB_PREV_DUP)
5560 return MDB_NOTFOUND;
5564 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5565 mdb_dbg_pgno(mp), (void *) mc));
5567 if (mc->mc_ki[mc->mc_top] == 0) {
5568 DPUTS("=====> move to prev sibling page");
5569 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5572 mp = mc->mc_pg[mc->mc_top];
5573 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5574 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5576 mc->mc_ki[mc->mc_top]--;
5578 mc->mc_flags &= ~C_EOF;
5580 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5581 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5584 key->mv_size = mc->mc_db->md_pad;
5585 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5589 mdb_cassert(mc, IS_LEAF(mp));
5590 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5592 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5593 mdb_xcursor_init1(mc, leaf);
5596 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5599 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5600 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5601 if (rc != MDB_SUCCESS)
5606 MDB_GET_KEY(leaf, key);
5610 /** Set the cursor on a specific data item. */
5612 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5613 MDB_cursor_op op, int *exactp)
5617 MDB_node *leaf = NULL;
5620 if (key->mv_size == 0)
5621 return MDB_BAD_VALSIZE;
5624 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5626 /* See if we're already on the right page */
5627 if (mc->mc_flags & C_INITIALIZED) {
5630 mp = mc->mc_pg[mc->mc_top];
5632 mc->mc_ki[mc->mc_top] = 0;
5633 return MDB_NOTFOUND;
5635 if (mp->mp_flags & P_LEAF2) {
5636 nodekey.mv_size = mc->mc_db->md_pad;
5637 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5639 leaf = NODEPTR(mp, 0);
5640 MDB_GET_KEY2(leaf, nodekey);
5642 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5644 /* Probably happens rarely, but first node on the page
5645 * was the one we wanted.
5647 mc->mc_ki[mc->mc_top] = 0;
5654 unsigned int nkeys = NUMKEYS(mp);
5656 if (mp->mp_flags & P_LEAF2) {
5657 nodekey.mv_data = LEAF2KEY(mp,
5658 nkeys-1, nodekey.mv_size);
5660 leaf = NODEPTR(mp, nkeys-1);
5661 MDB_GET_KEY2(leaf, nodekey);
5663 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5665 /* last node was the one we wanted */
5666 mc->mc_ki[mc->mc_top] = nkeys-1;
5672 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5673 /* This is definitely the right page, skip search_page */
5674 if (mp->mp_flags & P_LEAF2) {
5675 nodekey.mv_data = LEAF2KEY(mp,
5676 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5678 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5679 MDB_GET_KEY2(leaf, nodekey);
5681 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5683 /* current node was the one we wanted */
5693 /* If any parents have right-sibs, search.
5694 * Otherwise, there's nothing further.
5696 for (i=0; i<mc->mc_top; i++)
5698 NUMKEYS(mc->mc_pg[i])-1)
5700 if (i == mc->mc_top) {
5701 /* There are no other pages */
5702 mc->mc_ki[mc->mc_top] = nkeys;
5703 return MDB_NOTFOUND;
5707 /* There are no other pages */
5708 mc->mc_ki[mc->mc_top] = 0;
5709 if (op == MDB_SET_RANGE && !exactp) {
5713 return MDB_NOTFOUND;
5717 rc = mdb_page_search(mc, key, 0);
5718 if (rc != MDB_SUCCESS)
5721 mp = mc->mc_pg[mc->mc_top];
5722 mdb_cassert(mc, IS_LEAF(mp));
5725 leaf = mdb_node_search(mc, key, exactp);
5726 if (exactp != NULL && !*exactp) {
5727 /* MDB_SET specified and not an exact match. */
5728 return MDB_NOTFOUND;
5732 DPUTS("===> inexact leaf not found, goto sibling");
5733 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5734 mc->mc_flags |= C_EOF;
5735 return rc; /* no entries matched */
5737 mp = mc->mc_pg[mc->mc_top];
5738 mdb_cassert(mc, IS_LEAF(mp));
5739 leaf = NODEPTR(mp, 0);
5743 mc->mc_flags |= C_INITIALIZED;
5744 mc->mc_flags &= ~C_EOF;
5747 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5748 key->mv_size = mc->mc_db->md_pad;
5749 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5754 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5755 mdb_xcursor_init1(mc, leaf);
5758 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5759 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5760 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5763 if (op == MDB_GET_BOTH) {
5769 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5770 if (rc != MDB_SUCCESS)
5773 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5776 if ((rc = mdb_node_read(mc->mc_txn, leaf, &olddata)) != MDB_SUCCESS)
5778 dcmp = mc->mc_dbx->md_dcmp;
5779 #if UINT_MAX < SIZE_MAX
5780 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5781 dcmp = mdb_cmp_clong;
5783 rc = dcmp(data, &olddata);
5785 if (op == MDB_GET_BOTH || rc > 0)
5786 return MDB_NOTFOUND;
5793 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5794 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5799 /* The key already matches in all other cases */
5800 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5801 MDB_GET_KEY(leaf, key);
5802 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5807 /** Move the cursor to the first item in the database. */
5809 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5815 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5817 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5818 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5819 if (rc != MDB_SUCCESS)
5822 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5824 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5825 mc->mc_flags |= C_INITIALIZED;
5826 mc->mc_flags &= ~C_EOF;
5828 mc->mc_ki[mc->mc_top] = 0;
5830 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5831 key->mv_size = mc->mc_db->md_pad;
5832 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5837 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5838 mdb_xcursor_init1(mc, leaf);
5839 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5843 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5847 MDB_GET_KEY(leaf, key);
5851 /** Move the cursor to the last item in the database. */
5853 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5859 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5861 if (!(mc->mc_flags & C_EOF)) {
5863 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5864 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5865 if (rc != MDB_SUCCESS)
5868 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5871 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5872 mc->mc_flags |= C_INITIALIZED|C_EOF;
5873 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5875 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5876 key->mv_size = mc->mc_db->md_pad;
5877 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5882 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5883 mdb_xcursor_init1(mc, leaf);
5884 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5888 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5893 MDB_GET_KEY(leaf, key);
5898 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5903 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5908 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5912 case MDB_GET_CURRENT:
5913 if (!(mc->mc_flags & C_INITIALIZED)) {
5916 MDB_page *mp = mc->mc_pg[mc->mc_top];
5917 int nkeys = NUMKEYS(mp);
5918 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5919 mc->mc_ki[mc->mc_top] = nkeys;
5925 key->mv_size = mc->mc_db->md_pad;
5926 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5928 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5929 MDB_GET_KEY(leaf, key);
5931 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5932 if (mc->mc_flags & C_DEL)
5933 mdb_xcursor_init1(mc, leaf);
5934 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5936 rc = mdb_node_read(mc->mc_txn, leaf, data);
5943 case MDB_GET_BOTH_RANGE:
5948 if (mc->mc_xcursor == NULL) {
5949 rc = MDB_INCOMPATIBLE;
5959 rc = mdb_cursor_set(mc, key, data, op,
5960 op == MDB_SET_RANGE ? NULL : &exact);
5963 case MDB_GET_MULTIPLE:
5964 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5968 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5969 rc = MDB_INCOMPATIBLE;
5973 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5974 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5977 case MDB_NEXT_MULTIPLE:
5982 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5983 rc = MDB_INCOMPATIBLE;
5986 if (!(mc->mc_flags & C_INITIALIZED))
5987 rc = mdb_cursor_first(mc, key, data);
5989 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5990 if (rc == MDB_SUCCESS) {
5991 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5994 mx = &mc->mc_xcursor->mx_cursor;
5995 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5997 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5998 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
6006 case MDB_NEXT_NODUP:
6007 if (!(mc->mc_flags & C_INITIALIZED))
6008 rc = mdb_cursor_first(mc, key, data);
6010 rc = mdb_cursor_next(mc, key, data, op);
6014 case MDB_PREV_NODUP:
6015 if (!(mc->mc_flags & C_INITIALIZED)) {
6016 rc = mdb_cursor_last(mc, key, data);
6019 mc->mc_flags |= C_INITIALIZED;
6020 mc->mc_ki[mc->mc_top]++;
6022 rc = mdb_cursor_prev(mc, key, data, op);
6025 rc = mdb_cursor_first(mc, key, data);
6028 mfunc = mdb_cursor_first;
6030 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
6034 if (mc->mc_xcursor == NULL) {
6035 rc = MDB_INCOMPATIBLE;
6039 MDB_node *leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6040 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6041 MDB_GET_KEY(leaf, key);
6042 rc = mdb_node_read(mc->mc_txn, leaf, data);
6046 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
6050 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
6053 rc = mdb_cursor_last(mc, key, data);
6056 mfunc = mdb_cursor_last;
6059 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
6064 if (mc->mc_flags & C_DEL)
6065 mc->mc_flags ^= C_DEL;
6070 /** Touch all the pages in the cursor stack. Set mc_top.
6071 * Makes sure all the pages are writable, before attempting a write operation.
6072 * @param[in] mc The cursor to operate on.
6075 mdb_cursor_touch(MDB_cursor *mc)
6077 int rc = MDB_SUCCESS;
6079 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
6082 if (TXN_DBI_CHANGED(mc->mc_txn, mc->mc_dbi))
6084 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
6085 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
6088 *mc->mc_dbflag |= DB_DIRTY;
6093 rc = mdb_page_touch(mc);
6094 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
6095 mc->mc_top = mc->mc_snum-1;
6100 /** Do not spill pages to disk if txn is getting full, may fail instead */
6101 #define MDB_NOSPILL 0x8000
6104 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
6107 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
6109 MDB_node *leaf = NULL;
6112 MDB_val xdata, *rdata, dkey, olddata;
6114 int do_sub = 0, insert_key, insert_data;
6115 unsigned int mcount = 0, dcount = 0, nospill;
6118 unsigned int nflags;
6121 if (mc == NULL || key == NULL)
6124 env = mc->mc_txn->mt_env;
6126 /* Check this first so counter will always be zero on any
6129 if (flags & MDB_MULTIPLE) {
6130 dcount = data[1].mv_size;
6131 data[1].mv_size = 0;
6132 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
6133 return MDB_INCOMPATIBLE;
6136 nospill = flags & MDB_NOSPILL;
6137 flags &= ~MDB_NOSPILL;
6139 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6140 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6142 if (key->mv_size-1 >= ENV_MAXKEY(env))
6143 return MDB_BAD_VALSIZE;
6145 #if SIZE_MAX > MAXDATASIZE
6146 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
6147 return MDB_BAD_VALSIZE;
6149 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
6150 return MDB_BAD_VALSIZE;
6153 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
6154 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
6158 if (flags == MDB_CURRENT) {
6159 if (!(mc->mc_flags & C_INITIALIZED))
6162 } else if (mc->mc_db->md_root == P_INVALID) {
6163 /* new database, cursor has nothing to point to */
6166 mc->mc_flags &= ~C_INITIALIZED;
6171 if (flags & MDB_APPEND) {
6173 rc = mdb_cursor_last(mc, &k2, &d2);
6175 rc = mc->mc_dbx->md_cmp(key, &k2);
6178 mc->mc_ki[mc->mc_top]++;
6180 /* new key is <= last key */
6185 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
6187 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
6188 DPRINTF(("duplicate key [%s]", DKEY(key)));
6190 return MDB_KEYEXIST;
6192 if (rc && rc != MDB_NOTFOUND)
6196 if (mc->mc_flags & C_DEL)
6197 mc->mc_flags ^= C_DEL;
6199 /* Cursor is positioned, check for room in the dirty list */
6201 if (flags & MDB_MULTIPLE) {
6203 xdata.mv_size = data->mv_size * dcount;
6207 if ((rc2 = mdb_page_spill(mc, key, rdata)))
6211 if (rc == MDB_NO_ROOT) {
6213 /* new database, write a root leaf page */
6214 DPUTS("allocating new root leaf page");
6215 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
6218 mdb_cursor_push(mc, np);
6219 mc->mc_db->md_root = np->mp_pgno;
6220 mc->mc_db->md_depth++;
6221 *mc->mc_dbflag |= DB_DIRTY;
6222 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
6224 np->mp_flags |= P_LEAF2;
6225 mc->mc_flags |= C_INITIALIZED;
6227 /* make sure all cursor pages are writable */
6228 rc2 = mdb_cursor_touch(mc);
6233 insert_key = insert_data = rc;
6235 /* The key does not exist */
6236 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
6237 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
6238 LEAFSIZE(key, data) > env->me_nodemax)
6240 /* Too big for a node, insert in sub-DB. Set up an empty
6241 * "old sub-page" for prep_subDB to expand to a full page.
6243 fp_flags = P_LEAF|P_DIRTY;
6245 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
6246 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
6247 olddata.mv_size = PAGEHDRSZ;
6251 /* there's only a key anyway, so this is a no-op */
6252 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
6254 unsigned int ksize = mc->mc_db->md_pad;
6255 if (key->mv_size != ksize)
6256 return MDB_BAD_VALSIZE;
6257 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
6258 memcpy(ptr, key->mv_data, ksize);
6260 /* if overwriting slot 0 of leaf, need to
6261 * update branch key if there is a parent page
6263 if (mc->mc_top && !mc->mc_ki[mc->mc_top]) {
6264 unsigned short top = mc->mc_top;
6266 /* slot 0 is always an empty key, find real slot */
6267 while (mc->mc_top && !mc->mc_ki[mc->mc_top])
6269 if (mc->mc_ki[mc->mc_top])
6270 rc2 = mdb_update_key(mc, key);
6281 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6282 olddata.mv_size = NODEDSZ(leaf);
6283 olddata.mv_data = NODEDATA(leaf);
6286 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
6287 /* Prepare (sub-)page/sub-DB to accept the new item,
6288 * if needed. fp: old sub-page or a header faking
6289 * it. mp: new (sub-)page. offset: growth in page
6290 * size. xdata: node data with new page or DB.
6292 unsigned i, offset = 0;
6293 mp = fp = xdata.mv_data = env->me_pbuf;
6294 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
6296 /* Was a single item before, must convert now */
6297 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6299 /* Just overwrite the current item */
6300 if (flags == MDB_CURRENT)
6302 dcmp = mc->mc_dbx->md_dcmp;
6303 #if UINT_MAX < SIZE_MAX
6304 if (dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
6305 dcmp = mdb_cmp_clong;
6307 /* does data match? */
6308 if (!dcmp(data, &olddata)) {
6309 if (flags & MDB_NODUPDATA)
6310 return MDB_KEYEXIST;
6315 /* Back up original data item */
6316 dkey.mv_size = olddata.mv_size;
6317 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6319 /* Make sub-page header for the dup items, with dummy body */
6320 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6321 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6322 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6323 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6324 fp->mp_flags |= P_LEAF2;
6325 fp->mp_pad = data->mv_size;
6326 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6328 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6329 (dkey.mv_size & 1) + (data->mv_size & 1);
6331 fp->mp_upper = xdata.mv_size - PAGEBASE;
6332 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6333 } else if (leaf->mn_flags & F_SUBDATA) {
6334 /* Data is on sub-DB, just store it */
6335 flags |= F_DUPDATA|F_SUBDATA;
6338 /* Data is on sub-page */
6339 fp = olddata.mv_data;
6342 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6343 offset = EVEN(NODESIZE + sizeof(indx_t) +
6347 offset = fp->mp_pad;
6348 if (SIZELEFT(fp) < offset) {
6349 offset *= 4; /* space for 4 more */
6352 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6354 fp->mp_flags |= P_DIRTY;
6355 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6356 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6360 xdata.mv_size = olddata.mv_size + offset;
6363 fp_flags = fp->mp_flags;
6364 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6365 /* Too big for a sub-page, convert to sub-DB */
6366 fp_flags &= ~P_SUBP;
6368 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6369 fp_flags |= P_LEAF2;
6370 dummy.md_pad = fp->mp_pad;
6371 dummy.md_flags = MDB_DUPFIXED;
6372 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6373 dummy.md_flags |= MDB_INTEGERKEY;
6379 dummy.md_branch_pages = 0;
6380 dummy.md_leaf_pages = 1;
6381 dummy.md_overflow_pages = 0;
6382 dummy.md_entries = NUMKEYS(fp);
6383 xdata.mv_size = sizeof(MDB_db);
6384 xdata.mv_data = &dummy;
6385 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6387 offset = env->me_psize - olddata.mv_size;
6388 flags |= F_DUPDATA|F_SUBDATA;
6389 dummy.md_root = mp->mp_pgno;
6392 mp->mp_flags = fp_flags | P_DIRTY;
6393 mp->mp_pad = fp->mp_pad;
6394 mp->mp_lower = fp->mp_lower;
6395 mp->mp_upper = fp->mp_upper + offset;
6396 if (fp_flags & P_LEAF2) {
6397 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6399 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6400 olddata.mv_size - fp->mp_upper - PAGEBASE);
6401 for (i=0; i<NUMKEYS(fp); i++)
6402 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6410 mdb_node_del(mc, 0);
6414 /* LMDB passes F_SUBDATA in 'flags' to write a DB record */
6415 if ((leaf->mn_flags ^ flags) & F_SUBDATA)
6416 return MDB_INCOMPATIBLE;
6417 /* overflow page overwrites need special handling */
6418 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6421 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6423 memcpy(&pg, olddata.mv_data, sizeof(pg));
6424 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6426 ovpages = omp->mp_pages;
6428 /* Is the ov page large enough? */
6429 if (ovpages >= dpages) {
6430 if (!(omp->mp_flags & P_DIRTY) &&
6431 (level || (env->me_flags & MDB_WRITEMAP)))
6433 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6436 level = 0; /* dirty in this txn or clean */
6439 if (omp->mp_flags & P_DIRTY) {
6440 /* yes, overwrite it. Note in this case we don't
6441 * bother to try shrinking the page if the new data
6442 * is smaller than the overflow threshold.
6445 /* It is writable only in a parent txn */
6446 size_t sz = (size_t) env->me_psize * ovpages, off;
6447 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6453 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6454 mdb_cassert(mc, rc2 == 0);
6455 if (!(flags & MDB_RESERVE)) {
6456 /* Copy end of page, adjusting alignment so
6457 * compiler may copy words instead of bytes.
6459 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6460 memcpy((size_t *)((char *)np + off),
6461 (size_t *)((char *)omp + off), sz - off);
6464 memcpy(np, omp, sz); /* Copy beginning of page */
6467 SETDSZ(leaf, data->mv_size);
6468 if (F_ISSET(flags, MDB_RESERVE))
6469 data->mv_data = METADATA(omp);
6471 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6475 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6477 } else if (data->mv_size == olddata.mv_size) {
6478 /* same size, just replace it. Note that we could
6479 * also reuse this node if the new data is smaller,
6480 * but instead we opt to shrink the node in that case.
6482 if (F_ISSET(flags, MDB_RESERVE))
6483 data->mv_data = olddata.mv_data;
6484 else if (!(mc->mc_flags & C_SUB))
6485 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6487 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6492 mdb_node_del(mc, 0);
6498 nflags = flags & NODE_ADD_FLAGS;
6499 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6500 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6501 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6502 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6504 nflags |= MDB_SPLIT_REPLACE;
6505 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6507 /* There is room already in this leaf page. */
6508 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6509 if (rc == 0 && insert_key) {
6510 /* Adjust other cursors pointing to mp */
6511 MDB_cursor *m2, *m3;
6512 MDB_dbi dbi = mc->mc_dbi;
6513 unsigned i = mc->mc_top;
6514 MDB_page *mp = mc->mc_pg[i];
6516 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6517 if (mc->mc_flags & C_SUB)
6518 m3 = &m2->mc_xcursor->mx_cursor;
6521 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6522 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6529 if (rc == MDB_SUCCESS) {
6530 /* Now store the actual data in the child DB. Note that we're
6531 * storing the user data in the keys field, so there are strict
6532 * size limits on dupdata. The actual data fields of the child
6533 * DB are all zero size.
6541 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6542 if (flags & MDB_CURRENT) {
6543 xflags = MDB_CURRENT|MDB_NOSPILL;
6545 mdb_xcursor_init1(mc, leaf);
6546 xflags = (flags & MDB_NODUPDATA) ?
6547 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6549 /* converted, write the original data first */
6551 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6555 /* Adjust other cursors pointing to mp */
6557 unsigned i = mc->mc_top;
6558 MDB_page *mp = mc->mc_pg[i];
6560 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6561 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6562 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6563 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6564 mdb_xcursor_init1(m2, leaf);
6568 /* we've done our job */
6571 ecount = mc->mc_xcursor->mx_db.md_entries;
6572 if (flags & MDB_APPENDDUP)
6573 xflags |= MDB_APPEND;
6574 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6575 if (flags & F_SUBDATA) {
6576 void *db = NODEDATA(leaf);
6577 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6579 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6581 /* Increment count unless we just replaced an existing item. */
6583 mc->mc_db->md_entries++;
6585 /* Invalidate txn if we created an empty sub-DB */
6588 /* If we succeeded and the key didn't exist before,
6589 * make sure the cursor is marked valid.
6591 mc->mc_flags |= C_INITIALIZED;
6593 if (flags & MDB_MULTIPLE) {
6596 /* let caller know how many succeeded, if any */
6597 data[1].mv_size = mcount;
6598 if (mcount < dcount) {
6599 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6600 insert_key = insert_data = 0;
6607 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6610 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6615 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6621 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6622 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6624 if (!(mc->mc_flags & C_INITIALIZED))
6627 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6628 return MDB_NOTFOUND;
6630 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6633 rc = mdb_cursor_touch(mc);
6637 mp = mc->mc_pg[mc->mc_top];
6640 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6642 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6643 if (flags & MDB_NODUPDATA) {
6644 /* mdb_cursor_del0() will subtract the final entry */
6645 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6647 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6648 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6650 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6653 /* If sub-DB still has entries, we're done */
6654 if (mc->mc_xcursor->mx_db.md_entries) {
6655 if (leaf->mn_flags & F_SUBDATA) {
6656 /* update subDB info */
6657 void *db = NODEDATA(leaf);
6658 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6661 /* shrink fake page */
6662 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6663 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6664 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6665 /* fix other sub-DB cursors pointed at this fake page */
6666 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6667 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6668 if (m2->mc_pg[mc->mc_top] == mp &&
6669 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6670 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6673 mc->mc_db->md_entries--;
6674 mc->mc_flags |= C_DEL;
6677 /* otherwise fall thru and delete the sub-DB */
6680 if (leaf->mn_flags & F_SUBDATA) {
6681 /* add all the child DB's pages to the free list */
6682 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6687 /* LMDB passes F_SUBDATA in 'flags' to delete a DB record */
6688 else if ((leaf->mn_flags ^ flags) & F_SUBDATA) {
6689 rc = MDB_INCOMPATIBLE;
6693 /* add overflow pages to free list */
6694 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6698 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6699 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6700 (rc = mdb_ovpage_free(mc, omp)))
6705 return mdb_cursor_del0(mc);
6708 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6712 /** Allocate and initialize new pages for a database.
6713 * @param[in] mc a cursor on the database being added to.
6714 * @param[in] flags flags defining what type of page is being allocated.
6715 * @param[in] num the number of pages to allocate. This is usually 1,
6716 * unless allocating overflow pages for a large record.
6717 * @param[out] mp Address of a page, or NULL on failure.
6718 * @return 0 on success, non-zero on failure.
6721 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6726 if ((rc = mdb_page_alloc(mc, num, &np)))
6728 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6729 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6730 np->mp_flags = flags | P_DIRTY;
6731 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6732 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6735 mc->mc_db->md_branch_pages++;
6736 else if (IS_LEAF(np))
6737 mc->mc_db->md_leaf_pages++;
6738 else if (IS_OVERFLOW(np)) {
6739 mc->mc_db->md_overflow_pages += num;
6747 /** Calculate the size of a leaf node.
6748 * The size depends on the environment's page size; if a data item
6749 * is too large it will be put onto an overflow page and the node
6750 * size will only include the key and not the data. Sizes are always
6751 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6752 * of the #MDB_node headers.
6753 * @param[in] env The environment handle.
6754 * @param[in] key The key for the node.
6755 * @param[in] data The data for the node.
6756 * @return The number of bytes needed to store the node.
6759 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6763 sz = LEAFSIZE(key, data);
6764 if (sz > env->me_nodemax) {
6765 /* put on overflow page */
6766 sz -= data->mv_size - sizeof(pgno_t);
6769 return EVEN(sz + sizeof(indx_t));
6772 /** Calculate the size of a branch node.
6773 * The size should depend on the environment's page size but since
6774 * we currently don't support spilling large keys onto overflow
6775 * pages, it's simply the size of the #MDB_node header plus the
6776 * size of the key. Sizes are always rounded up to an even number
6777 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6778 * @param[in] env The environment handle.
6779 * @param[in] key The key for the node.
6780 * @return The number of bytes needed to store the node.
6783 mdb_branch_size(MDB_env *env, MDB_val *key)
6788 if (sz > env->me_nodemax) {
6789 /* put on overflow page */
6790 /* not implemented */
6791 /* sz -= key->size - sizeof(pgno_t); */
6794 return sz + sizeof(indx_t);
6797 /** Add a node to the page pointed to by the cursor.
6798 * @param[in] mc The cursor for this operation.
6799 * @param[in] indx The index on the page where the new node should be added.
6800 * @param[in] key The key for the new node.
6801 * @param[in] data The data for the new node, if any.
6802 * @param[in] pgno The page number, if adding a branch node.
6803 * @param[in] flags Flags for the node.
6804 * @return 0 on success, non-zero on failure. Possible errors are:
6806 * <li>ENOMEM - failed to allocate overflow pages for the node.
6807 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6808 * should never happen since all callers already calculate the
6809 * page's free space before calling this function.
6813 mdb_node_add(MDB_cursor *mc, indx_t indx,
6814 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6817 size_t node_size = NODESIZE;
6821 MDB_page *mp = mc->mc_pg[mc->mc_top];
6822 MDB_page *ofp = NULL; /* overflow page */
6825 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6827 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6828 IS_LEAF(mp) ? "leaf" : "branch",
6829 IS_SUBP(mp) ? "sub-" : "",
6830 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6831 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6834 /* Move higher keys up one slot. */
6835 int ksize = mc->mc_db->md_pad, dif;
6836 char *ptr = LEAF2KEY(mp, indx, ksize);
6837 dif = NUMKEYS(mp) - indx;
6839 memmove(ptr+ksize, ptr, dif*ksize);
6840 /* insert new key */
6841 memcpy(ptr, key->mv_data, ksize);
6843 /* Just using these for counting */
6844 mp->mp_lower += sizeof(indx_t);
6845 mp->mp_upper -= ksize - sizeof(indx_t);
6849 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6851 node_size += key->mv_size;
6853 mdb_cassert(mc, data);
6854 if (F_ISSET(flags, F_BIGDATA)) {
6855 /* Data already on overflow page. */
6856 node_size += sizeof(pgno_t);
6857 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6858 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6860 /* Put data on overflow page. */
6861 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6862 data->mv_size, node_size+data->mv_size));
6863 node_size = EVEN(node_size + sizeof(pgno_t));
6864 if ((ssize_t)node_size > room)
6866 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6868 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6872 node_size += data->mv_size;
6875 node_size = EVEN(node_size);
6876 if ((ssize_t)node_size > room)
6880 /* Move higher pointers up one slot. */
6881 for (i = NUMKEYS(mp); i > indx; i--)
6882 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6884 /* Adjust free space offsets. */
6885 ofs = mp->mp_upper - node_size;
6886 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6887 mp->mp_ptrs[indx] = ofs;
6889 mp->mp_lower += sizeof(indx_t);
6891 /* Write the node data. */
6892 node = NODEPTR(mp, indx);
6893 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6894 node->mn_flags = flags;
6896 SETDSZ(node,data->mv_size);
6901 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6904 mdb_cassert(mc, key);
6906 if (F_ISSET(flags, F_BIGDATA))
6907 memcpy(node->mn_data + key->mv_size, data->mv_data,
6909 else if (F_ISSET(flags, MDB_RESERVE))
6910 data->mv_data = node->mn_data + key->mv_size;
6912 memcpy(node->mn_data + key->mv_size, data->mv_data,
6915 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6917 if (F_ISSET(flags, MDB_RESERVE))
6918 data->mv_data = METADATA(ofp);
6920 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6927 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6928 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6929 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6930 DPRINTF(("node size = %"Z"u", node_size));
6931 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6932 return MDB_PAGE_FULL;
6935 /** Delete the specified node from a page.
6936 * @param[in] mc Cursor pointing to the node to delete.
6937 * @param[in] ksize The size of a node. Only used if the page is
6938 * part of a #MDB_DUPFIXED database.
6941 mdb_node_del(MDB_cursor *mc, int ksize)
6943 MDB_page *mp = mc->mc_pg[mc->mc_top];
6944 indx_t indx = mc->mc_ki[mc->mc_top];
6946 indx_t i, j, numkeys, ptr;
6950 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6951 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6952 numkeys = NUMKEYS(mp);
6953 mdb_cassert(mc, indx < numkeys);
6956 int x = numkeys - 1 - indx;
6957 base = LEAF2KEY(mp, indx, ksize);
6959 memmove(base, base + ksize, x * ksize);
6960 mp->mp_lower -= sizeof(indx_t);
6961 mp->mp_upper += ksize - sizeof(indx_t);
6965 node = NODEPTR(mp, indx);
6966 sz = NODESIZE + node->mn_ksize;
6968 if (F_ISSET(node->mn_flags, F_BIGDATA))
6969 sz += sizeof(pgno_t);
6971 sz += NODEDSZ(node);
6975 ptr = mp->mp_ptrs[indx];
6976 for (i = j = 0; i < numkeys; i++) {
6978 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6979 if (mp->mp_ptrs[i] < ptr)
6980 mp->mp_ptrs[j] += sz;
6985 base = (char *)mp + mp->mp_upper + PAGEBASE;
6986 memmove(base + sz, base, ptr - mp->mp_upper);
6988 mp->mp_lower -= sizeof(indx_t);
6992 /** Compact the main page after deleting a node on a subpage.
6993 * @param[in] mp The main page to operate on.
6994 * @param[in] indx The index of the subpage on the main page.
6997 mdb_node_shrink(MDB_page *mp, indx_t indx)
7003 indx_t i, numkeys, ptr;
7005 node = NODEPTR(mp, indx);
7006 sp = (MDB_page *)NODEDATA(node);
7007 delta = SIZELEFT(sp);
7008 xp = (MDB_page *)((char *)sp + delta);
7010 /* shift subpage upward */
7012 nsize = NUMKEYS(sp) * sp->mp_pad;
7014 return; /* do not make the node uneven-sized */
7015 memmove(METADATA(xp), METADATA(sp), nsize);
7018 numkeys = NUMKEYS(sp);
7019 for (i=numkeys-1; i>=0; i--)
7020 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
7022 xp->mp_upper = sp->mp_lower;
7023 xp->mp_lower = sp->mp_lower;
7024 xp->mp_flags = sp->mp_flags;
7025 xp->mp_pad = sp->mp_pad;
7026 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
7028 nsize = NODEDSZ(node) - delta;
7029 SETDSZ(node, nsize);
7031 /* shift lower nodes upward */
7032 ptr = mp->mp_ptrs[indx];
7033 numkeys = NUMKEYS(mp);
7034 for (i = 0; i < numkeys; i++) {
7035 if (mp->mp_ptrs[i] <= ptr)
7036 mp->mp_ptrs[i] += delta;
7039 base = (char *)mp + mp->mp_upper + PAGEBASE;
7040 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
7041 mp->mp_upper += delta;
7044 /** Initial setup of a sorted-dups cursor.
7045 * Sorted duplicates are implemented as a sub-database for the given key.
7046 * The duplicate data items are actually keys of the sub-database.
7047 * Operations on the duplicate data items are performed using a sub-cursor
7048 * initialized when the sub-database is first accessed. This function does
7049 * the preliminary setup of the sub-cursor, filling in the fields that
7050 * depend only on the parent DB.
7051 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7054 mdb_xcursor_init0(MDB_cursor *mc)
7056 MDB_xcursor *mx = mc->mc_xcursor;
7058 mx->mx_cursor.mc_xcursor = NULL;
7059 mx->mx_cursor.mc_txn = mc->mc_txn;
7060 mx->mx_cursor.mc_db = &mx->mx_db;
7061 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
7062 mx->mx_cursor.mc_dbi = mc->mc_dbi;
7063 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
7064 mx->mx_cursor.mc_snum = 0;
7065 mx->mx_cursor.mc_top = 0;
7066 mx->mx_cursor.mc_flags = C_SUB;
7067 mx->mx_dbx.md_name.mv_size = 0;
7068 mx->mx_dbx.md_name.mv_data = NULL;
7069 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
7070 mx->mx_dbx.md_dcmp = NULL;
7071 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
7074 /** Final setup of a sorted-dups cursor.
7075 * Sets up the fields that depend on the data from the main cursor.
7076 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
7077 * @param[in] node The data containing the #MDB_db record for the
7078 * sorted-dup database.
7081 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
7083 MDB_xcursor *mx = mc->mc_xcursor;
7085 if (node->mn_flags & F_SUBDATA) {
7086 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
7087 mx->mx_cursor.mc_pg[0] = 0;
7088 mx->mx_cursor.mc_snum = 0;
7089 mx->mx_cursor.mc_top = 0;
7090 mx->mx_cursor.mc_flags = C_SUB;
7092 MDB_page *fp = NODEDATA(node);
7093 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
7094 mx->mx_db.md_flags = 0;
7095 mx->mx_db.md_depth = 1;
7096 mx->mx_db.md_branch_pages = 0;
7097 mx->mx_db.md_leaf_pages = 1;
7098 mx->mx_db.md_overflow_pages = 0;
7099 mx->mx_db.md_entries = NUMKEYS(fp);
7100 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
7101 mx->mx_cursor.mc_snum = 1;
7102 mx->mx_cursor.mc_top = 0;
7103 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
7104 mx->mx_cursor.mc_pg[0] = fp;
7105 mx->mx_cursor.mc_ki[0] = 0;
7106 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
7107 mx->mx_db.md_flags = MDB_DUPFIXED;
7108 mx->mx_db.md_pad = fp->mp_pad;
7109 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
7110 mx->mx_db.md_flags |= MDB_INTEGERKEY;
7113 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
7114 mx->mx_db.md_root));
7115 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
7116 #if UINT_MAX < SIZE_MAX
7117 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
7118 mx->mx_dbx.md_cmp = mdb_cmp_clong;
7122 /** Initialize a cursor for a given transaction and database. */
7124 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
7127 mc->mc_backup = NULL;
7130 mc->mc_db = &txn->mt_dbs[dbi];
7131 mc->mc_dbx = &txn->mt_dbxs[dbi];
7132 mc->mc_dbflag = &txn->mt_dbflags[dbi];
7138 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
7139 mdb_tassert(txn, mx != NULL);
7140 mc->mc_xcursor = mx;
7141 mdb_xcursor_init0(mc);
7143 mc->mc_xcursor = NULL;
7145 if (*mc->mc_dbflag & DB_STALE) {
7146 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
7151 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
7154 size_t size = sizeof(MDB_cursor);
7156 if (!ret || !TXN_DBI_EXIST(txn, dbi))
7159 if (txn->mt_flags & MDB_TXN_ERROR)
7162 /* Allow read access to the freelist */
7163 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
7166 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
7167 size += sizeof(MDB_xcursor);
7169 if ((mc = malloc(size)) != NULL) {
7170 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
7171 if (txn->mt_cursors) {
7172 mc->mc_next = txn->mt_cursors[dbi];
7173 txn->mt_cursors[dbi] = mc;
7174 mc->mc_flags |= C_UNTRACK;
7186 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
7188 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
7191 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
7194 if (txn->mt_flags & MDB_TXN_ERROR)
7197 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
7201 /* Return the count of duplicate data items for the current key */
7203 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
7207 if (mc == NULL || countp == NULL)
7210 if (mc->mc_xcursor == NULL)
7211 return MDB_INCOMPATIBLE;
7213 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
7216 if (!(mc->mc_flags & C_INITIALIZED))
7219 if (!mc->mc_snum || (mc->mc_flags & C_EOF))
7220 return MDB_NOTFOUND;
7222 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
7223 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
7226 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
7229 *countp = mc->mc_xcursor->mx_db.md_entries;
7235 mdb_cursor_close(MDB_cursor *mc)
7237 if (mc && !mc->mc_backup) {
7238 /* remove from txn, if tracked */
7239 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
7240 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
7241 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
7243 *prev = mc->mc_next;
7250 mdb_cursor_txn(MDB_cursor *mc)
7252 if (!mc) return NULL;
7257 mdb_cursor_dbi(MDB_cursor *mc)
7262 /** Replace the key for a branch node with a new key.
7263 * @param[in] mc Cursor pointing to the node to operate on.
7264 * @param[in] key The new key to use.
7265 * @return 0 on success, non-zero on failure.
7268 mdb_update_key(MDB_cursor *mc, MDB_val *key)
7274 int delta, ksize, oksize;
7275 indx_t ptr, i, numkeys, indx;
7278 indx = mc->mc_ki[mc->mc_top];
7279 mp = mc->mc_pg[mc->mc_top];
7280 node = NODEPTR(mp, indx);
7281 ptr = mp->mp_ptrs[indx];
7285 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
7286 k2.mv_data = NODEKEY(node);
7287 k2.mv_size = node->mn_ksize;
7288 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
7290 mdb_dkey(&k2, kbuf2),
7296 /* Sizes must be 2-byte aligned. */
7297 ksize = EVEN(key->mv_size);
7298 oksize = EVEN(node->mn_ksize);
7299 delta = ksize - oksize;
7301 /* Shift node contents if EVEN(key length) changed. */
7303 if (delta > 0 && SIZELEFT(mp) < delta) {
7305 /* not enough space left, do a delete and split */
7306 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
7307 pgno = NODEPGNO(node);
7308 mdb_node_del(mc, 0);
7309 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
7312 numkeys = NUMKEYS(mp);
7313 for (i = 0; i < numkeys; i++) {
7314 if (mp->mp_ptrs[i] <= ptr)
7315 mp->mp_ptrs[i] -= delta;
7318 base = (char *)mp + mp->mp_upper + PAGEBASE;
7319 len = ptr - mp->mp_upper + NODESIZE;
7320 memmove(base - delta, base, len);
7321 mp->mp_upper -= delta;
7323 node = NODEPTR(mp, indx);
7326 /* But even if no shift was needed, update ksize */
7327 if (node->mn_ksize != key->mv_size)
7328 node->mn_ksize = key->mv_size;
7331 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7337 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7339 /** Move a node from csrc to cdst.
7342 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7349 unsigned short flags;
7353 /* Mark src and dst as dirty. */
7354 if ((rc = mdb_page_touch(csrc)) ||
7355 (rc = mdb_page_touch(cdst)))
7358 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7359 key.mv_size = csrc->mc_db->md_pad;
7360 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7362 data.mv_data = NULL;
7366 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7367 mdb_cassert(csrc, !((size_t)srcnode & 1));
7368 srcpg = NODEPGNO(srcnode);
7369 flags = srcnode->mn_flags;
7370 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7371 unsigned int snum = csrc->mc_snum;
7373 /* must find the lowest key below src */
7374 rc = mdb_page_search_lowest(csrc);
7377 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7378 key.mv_size = csrc->mc_db->md_pad;
7379 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7381 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7382 key.mv_size = NODEKSZ(s2);
7383 key.mv_data = NODEKEY(s2);
7385 csrc->mc_snum = snum--;
7386 csrc->mc_top = snum;
7388 key.mv_size = NODEKSZ(srcnode);
7389 key.mv_data = NODEKEY(srcnode);
7391 data.mv_size = NODEDSZ(srcnode);
7392 data.mv_data = NODEDATA(srcnode);
7394 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7395 unsigned int snum = cdst->mc_snum;
7398 /* must find the lowest key below dst */
7399 mdb_cursor_copy(cdst, &mn);
7400 rc = mdb_page_search_lowest(&mn);
7403 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7404 bkey.mv_size = mn.mc_db->md_pad;
7405 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7407 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7408 bkey.mv_size = NODEKSZ(s2);
7409 bkey.mv_data = NODEKEY(s2);
7411 mn.mc_snum = snum--;
7414 rc = mdb_update_key(&mn, &bkey);
7419 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7420 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7421 csrc->mc_ki[csrc->mc_top],
7423 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7424 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7426 /* Add the node to the destination page.
7428 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7429 if (rc != MDB_SUCCESS)
7432 /* Delete the node from the source page.
7434 mdb_node_del(csrc, key.mv_size);
7437 /* Adjust other cursors pointing to mp */
7438 MDB_cursor *m2, *m3;
7439 MDB_dbi dbi = csrc->mc_dbi;
7440 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7442 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7443 if (csrc->mc_flags & C_SUB)
7444 m3 = &m2->mc_xcursor->mx_cursor;
7447 if (m3 == csrc) continue;
7448 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7449 csrc->mc_ki[csrc->mc_top]) {
7450 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7451 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7456 /* Update the parent separators.
7458 if (csrc->mc_ki[csrc->mc_top] == 0) {
7459 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7460 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7461 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7463 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7464 key.mv_size = NODEKSZ(srcnode);
7465 key.mv_data = NODEKEY(srcnode);
7467 DPRINTF(("update separator for source page %"Z"u to [%s]",
7468 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7469 mdb_cursor_copy(csrc, &mn);
7472 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7475 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7477 indx_t ix = csrc->mc_ki[csrc->mc_top];
7478 nullkey.mv_size = 0;
7479 csrc->mc_ki[csrc->mc_top] = 0;
7480 rc = mdb_update_key(csrc, &nullkey);
7481 csrc->mc_ki[csrc->mc_top] = ix;
7482 mdb_cassert(csrc, rc == MDB_SUCCESS);
7486 if (cdst->mc_ki[cdst->mc_top] == 0) {
7487 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7488 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7489 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7491 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7492 key.mv_size = NODEKSZ(srcnode);
7493 key.mv_data = NODEKEY(srcnode);
7495 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7496 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7497 mdb_cursor_copy(cdst, &mn);
7500 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7503 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7505 indx_t ix = cdst->mc_ki[cdst->mc_top];
7506 nullkey.mv_size = 0;
7507 cdst->mc_ki[cdst->mc_top] = 0;
7508 rc = mdb_update_key(cdst, &nullkey);
7509 cdst->mc_ki[cdst->mc_top] = ix;
7510 mdb_cassert(cdst, rc == MDB_SUCCESS);
7517 /** Merge one page into another.
7518 * The nodes from the page pointed to by \b csrc will
7519 * be copied to the page pointed to by \b cdst and then
7520 * the \b csrc page will be freed.
7521 * @param[in] csrc Cursor pointing to the source page.
7522 * @param[in] cdst Cursor pointing to the destination page.
7523 * @return 0 on success, non-zero on failure.
7526 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7528 MDB_page *psrc, *pdst;
7535 psrc = csrc->mc_pg[csrc->mc_top];
7536 pdst = cdst->mc_pg[cdst->mc_top];
7538 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7540 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7541 mdb_cassert(csrc, cdst->mc_snum > 1);
7543 /* Mark dst as dirty. */
7544 if ((rc = mdb_page_touch(cdst)))
7547 /* Move all nodes from src to dst.
7549 j = nkeys = NUMKEYS(pdst);
7550 if (IS_LEAF2(psrc)) {
7551 key.mv_size = csrc->mc_db->md_pad;
7552 key.mv_data = METADATA(psrc);
7553 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7554 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7555 if (rc != MDB_SUCCESS)
7557 key.mv_data = (char *)key.mv_data + key.mv_size;
7560 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7561 srcnode = NODEPTR(psrc, i);
7562 if (i == 0 && IS_BRANCH(psrc)) {
7565 mdb_cursor_copy(csrc, &mn);
7566 /* must find the lowest key below src */
7567 rc = mdb_page_search_lowest(&mn);
7570 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7571 key.mv_size = mn.mc_db->md_pad;
7572 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7574 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7575 key.mv_size = NODEKSZ(s2);
7576 key.mv_data = NODEKEY(s2);
7579 key.mv_size = srcnode->mn_ksize;
7580 key.mv_data = NODEKEY(srcnode);
7583 data.mv_size = NODEDSZ(srcnode);
7584 data.mv_data = NODEDATA(srcnode);
7585 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7586 if (rc != MDB_SUCCESS)
7591 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7592 pdst->mp_pgno, NUMKEYS(pdst),
7593 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7595 /* Unlink the src page from parent and add to free list.
7598 mdb_node_del(csrc, 0);
7599 if (csrc->mc_ki[csrc->mc_top] == 0) {
7601 rc = mdb_update_key(csrc, &key);
7609 psrc = csrc->mc_pg[csrc->mc_top];
7610 /* If not operating on FreeDB, allow this page to be reused
7611 * in this txn. Otherwise just add to free list.
7613 rc = mdb_page_loose(csrc, psrc);
7617 csrc->mc_db->md_leaf_pages--;
7619 csrc->mc_db->md_branch_pages--;
7621 /* Adjust other cursors pointing to mp */
7622 MDB_cursor *m2, *m3;
7623 MDB_dbi dbi = csrc->mc_dbi;
7625 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7626 if (csrc->mc_flags & C_SUB)
7627 m3 = &m2->mc_xcursor->mx_cursor;
7630 if (m3 == csrc) continue;
7631 if (m3->mc_snum < csrc->mc_snum) continue;
7632 if (m3->mc_pg[csrc->mc_top] == psrc) {
7633 m3->mc_pg[csrc->mc_top] = pdst;
7634 m3->mc_ki[csrc->mc_top] += nkeys;
7639 unsigned int snum = cdst->mc_snum;
7640 uint16_t depth = cdst->mc_db->md_depth;
7641 mdb_cursor_pop(cdst);
7642 rc = mdb_rebalance(cdst);
7643 /* Did the tree shrink? */
7644 if (depth > cdst->mc_db->md_depth)
7646 cdst->mc_snum = snum;
7647 cdst->mc_top = snum-1;
7652 /** Copy the contents of a cursor.
7653 * @param[in] csrc The cursor to copy from.
7654 * @param[out] cdst The cursor to copy to.
7657 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7661 cdst->mc_txn = csrc->mc_txn;
7662 cdst->mc_dbi = csrc->mc_dbi;
7663 cdst->mc_db = csrc->mc_db;
7664 cdst->mc_dbx = csrc->mc_dbx;
7665 cdst->mc_snum = csrc->mc_snum;
7666 cdst->mc_top = csrc->mc_top;
7667 cdst->mc_flags = csrc->mc_flags;
7669 for (i=0; i<csrc->mc_snum; i++) {
7670 cdst->mc_pg[i] = csrc->mc_pg[i];
7671 cdst->mc_ki[i] = csrc->mc_ki[i];
7675 /** Rebalance the tree after a delete operation.
7676 * @param[in] mc Cursor pointing to the page where rebalancing
7678 * @return 0 on success, non-zero on failure.
7681 mdb_rebalance(MDB_cursor *mc)
7685 unsigned int ptop, minkeys;
7689 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7690 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7691 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7692 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7693 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7695 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7696 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7697 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7698 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7702 if (mc->mc_snum < 2) {
7703 MDB_page *mp = mc->mc_pg[0];
7705 DPUTS("Can't rebalance a subpage, ignoring");
7708 if (NUMKEYS(mp) == 0) {
7709 DPUTS("tree is completely empty");
7710 mc->mc_db->md_root = P_INVALID;
7711 mc->mc_db->md_depth = 0;
7712 mc->mc_db->md_leaf_pages = 0;
7713 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7716 /* Adjust cursors pointing to mp */
7719 mc->mc_flags &= ~C_INITIALIZED;
7721 MDB_cursor *m2, *m3;
7722 MDB_dbi dbi = mc->mc_dbi;
7724 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7725 if (mc->mc_flags & C_SUB)
7726 m3 = &m2->mc_xcursor->mx_cursor;
7729 if (m3->mc_snum < mc->mc_snum) continue;
7730 if (m3->mc_pg[0] == mp) {
7733 m3->mc_flags &= ~C_INITIALIZED;
7737 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7739 DPUTS("collapsing root page!");
7740 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7743 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7744 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7747 mc->mc_db->md_depth--;
7748 mc->mc_db->md_branch_pages--;
7749 mc->mc_ki[0] = mc->mc_ki[1];
7750 for (i = 1; i<mc->mc_db->md_depth; i++) {
7751 mc->mc_pg[i] = mc->mc_pg[i+1];
7752 mc->mc_ki[i] = mc->mc_ki[i+1];
7755 /* Adjust other cursors pointing to mp */
7756 MDB_cursor *m2, *m3;
7757 MDB_dbi dbi = mc->mc_dbi;
7759 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7760 if (mc->mc_flags & C_SUB)
7761 m3 = &m2->mc_xcursor->mx_cursor;
7764 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7765 if (m3->mc_pg[0] == mp) {
7766 for (i=0; i<m3->mc_snum; i++) {
7767 m3->mc_pg[i] = m3->mc_pg[i+1];
7768 m3->mc_ki[i] = m3->mc_ki[i+1];
7776 DPUTS("root page doesn't need rebalancing");
7780 /* The parent (branch page) must have at least 2 pointers,
7781 * otherwise the tree is invalid.
7783 ptop = mc->mc_top-1;
7784 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7786 /* Leaf page fill factor is below the threshold.
7787 * Try to move keys from left or right neighbor, or
7788 * merge with a neighbor page.
7793 mdb_cursor_copy(mc, &mn);
7794 mn.mc_xcursor = NULL;
7796 oldki = mc->mc_ki[mc->mc_top];
7797 if (mc->mc_ki[ptop] == 0) {
7798 /* We're the leftmost leaf in our parent.
7800 DPUTS("reading right neighbor");
7802 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7803 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7806 mn.mc_ki[mn.mc_top] = 0;
7807 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7809 /* There is at least one neighbor to the left.
7811 DPUTS("reading left neighbor");
7813 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7814 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7817 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7818 mc->mc_ki[mc->mc_top] = 0;
7821 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7822 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7823 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7825 /* If the neighbor page is above threshold and has enough keys,
7826 * move one key from it. Otherwise we should try to merge them.
7827 * (A branch page must never have less than 2 keys.)
7829 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7830 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7831 rc = mdb_node_move(&mn, mc);
7832 if (mc->mc_ki[ptop]) {
7836 if (mc->mc_ki[ptop] == 0) {
7837 rc = mdb_page_merge(&mn, mc);
7840 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7841 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7842 /* We want mdb_rebalance to find mn when doing fixups */
7843 if (mc->mc_flags & C_SUB) {
7844 dummy.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7845 mc->mc_txn->mt_cursors[mc->mc_dbi] = &dummy;
7846 dummy.mc_xcursor = (MDB_xcursor *)&mn;
7848 mn.mc_next = mc->mc_txn->mt_cursors[mc->mc_dbi];
7849 mc->mc_txn->mt_cursors[mc->mc_dbi] = &mn;
7851 rc = mdb_page_merge(mc, &mn);
7852 if (mc->mc_flags & C_SUB)
7853 mc->mc_txn->mt_cursors[mc->mc_dbi] = dummy.mc_next;
7855 mc->mc_txn->mt_cursors[mc->mc_dbi] = mn.mc_next;
7856 mdb_cursor_copy(&mn, mc);
7858 mc->mc_flags &= ~C_EOF;
7860 mc->mc_ki[mc->mc_top] = oldki;
7864 /** Complete a delete operation started by #mdb_cursor_del(). */
7866 mdb_cursor_del0(MDB_cursor *mc)
7873 ki = mc->mc_ki[mc->mc_top];
7874 mdb_node_del(mc, mc->mc_db->md_pad);
7875 mc->mc_db->md_entries--;
7876 rc = mdb_rebalance(mc);
7878 if (rc == MDB_SUCCESS) {
7879 MDB_cursor *m2, *m3;
7880 MDB_dbi dbi = mc->mc_dbi;
7882 /* DB is totally empty now, just bail out.
7883 * Other cursors adjustments were already done
7884 * by mdb_rebalance and aren't needed here.
7889 mp = mc->mc_pg[mc->mc_top];
7890 nkeys = NUMKEYS(mp);
7892 /* if mc points past last node in page, find next sibling */
7893 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7894 rc = mdb_cursor_sibling(mc, 1);
7895 if (rc == MDB_NOTFOUND) {
7896 mc->mc_flags |= C_EOF;
7901 /* Adjust other cursors pointing to mp */
7902 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7903 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7904 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7906 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7908 if (m3->mc_pg[mc->mc_top] == mp) {
7909 if (m3->mc_ki[mc->mc_top] >= ki) {
7910 m3->mc_flags |= C_DEL;
7911 if (m3->mc_ki[mc->mc_top] > ki)
7912 m3->mc_ki[mc->mc_top]--;
7913 else if (mc->mc_db->md_flags & MDB_DUPSORT)
7914 m3->mc_xcursor->mx_cursor.mc_flags |= C_EOF;
7916 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7917 rc = mdb_cursor_sibling(m3, 1);
7918 if (rc == MDB_NOTFOUND) {
7919 m3->mc_flags |= C_EOF;
7925 mc->mc_flags |= C_DEL;
7929 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7934 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7935 MDB_val *key, MDB_val *data)
7937 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7940 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7941 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7943 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7944 /* must ignore any data */
7948 return mdb_del0(txn, dbi, key, data, 0);
7952 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7953 MDB_val *key, MDB_val *data, unsigned flags)
7958 MDB_val rdata, *xdata;
7962 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7964 mdb_cursor_init(&mc, txn, dbi, &mx);
7973 flags |= MDB_NODUPDATA;
7975 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7977 /* let mdb_page_split know about this cursor if needed:
7978 * delete will trigger a rebalance; if it needs to move
7979 * a node from one page to another, it will have to
7980 * update the parent's separator key(s). If the new sepkey
7981 * is larger than the current one, the parent page may
7982 * run out of space, triggering a split. We need this
7983 * cursor to be consistent until the end of the rebalance.
7985 mc.mc_flags |= C_UNTRACK;
7986 mc.mc_next = txn->mt_cursors[dbi];
7987 txn->mt_cursors[dbi] = &mc;
7988 rc = mdb_cursor_del(&mc, flags);
7989 txn->mt_cursors[dbi] = mc.mc_next;
7994 /** Split a page and insert a new node.
7995 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7996 * The cursor will be updated to point to the actual page and index where
7997 * the node got inserted after the split.
7998 * @param[in] newkey The key for the newly inserted node.
7999 * @param[in] newdata The data for the newly inserted node.
8000 * @param[in] newpgno The page number, if the new node is a branch node.
8001 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
8002 * @return 0 on success, non-zero on failure.
8005 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
8006 unsigned int nflags)
8009 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
8012 int i, j, split_indx, nkeys, pmax;
8013 MDB_env *env = mc->mc_txn->mt_env;
8015 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
8016 MDB_page *copy = NULL;
8017 MDB_page *mp, *rp, *pp;
8022 mp = mc->mc_pg[mc->mc_top];
8023 newindx = mc->mc_ki[mc->mc_top];
8024 nkeys = NUMKEYS(mp);
8026 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
8027 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
8028 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
8030 /* Create a right sibling. */
8031 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
8033 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
8035 if (mc->mc_snum < 2) {
8036 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
8038 /* shift current top to make room for new parent */
8039 mc->mc_pg[1] = mc->mc_pg[0];
8040 mc->mc_ki[1] = mc->mc_ki[0];
8043 mc->mc_db->md_root = pp->mp_pgno;
8044 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
8045 mc->mc_db->md_depth++;
8048 /* Add left (implicit) pointer. */
8049 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
8050 /* undo the pre-push */
8051 mc->mc_pg[0] = mc->mc_pg[1];
8052 mc->mc_ki[0] = mc->mc_ki[1];
8053 mc->mc_db->md_root = mp->mp_pgno;
8054 mc->mc_db->md_depth--;
8061 ptop = mc->mc_top-1;
8062 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
8065 mc->mc_flags |= C_SPLITTING;
8066 mdb_cursor_copy(mc, &mn);
8067 mn.mc_pg[mn.mc_top] = rp;
8068 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
8070 if (nflags & MDB_APPEND) {
8071 mn.mc_ki[mn.mc_top] = 0;
8073 split_indx = newindx;
8077 split_indx = (nkeys+1) / 2;
8082 unsigned int lsize, rsize, ksize;
8083 /* Move half of the keys to the right sibling */
8084 x = mc->mc_ki[mc->mc_top] - split_indx;
8085 ksize = mc->mc_db->md_pad;
8086 split = LEAF2KEY(mp, split_indx, ksize);
8087 rsize = (nkeys - split_indx) * ksize;
8088 lsize = (nkeys - split_indx) * sizeof(indx_t);
8089 mp->mp_lower -= lsize;
8090 rp->mp_lower += lsize;
8091 mp->mp_upper += rsize - lsize;
8092 rp->mp_upper -= rsize - lsize;
8093 sepkey.mv_size = ksize;
8094 if (newindx == split_indx) {
8095 sepkey.mv_data = newkey->mv_data;
8097 sepkey.mv_data = split;
8100 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
8101 memcpy(rp->mp_ptrs, split, rsize);
8102 sepkey.mv_data = rp->mp_ptrs;
8103 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
8104 memcpy(ins, newkey->mv_data, ksize);
8105 mp->mp_lower += sizeof(indx_t);
8106 mp->mp_upper -= ksize - sizeof(indx_t);
8109 memcpy(rp->mp_ptrs, split, x * ksize);
8110 ins = LEAF2KEY(rp, x, ksize);
8111 memcpy(ins, newkey->mv_data, ksize);
8112 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
8113 rp->mp_lower += sizeof(indx_t);
8114 rp->mp_upper -= ksize - sizeof(indx_t);
8115 mc->mc_ki[mc->mc_top] = x;
8116 mc->mc_pg[mc->mc_top] = rp;
8119 int psize, nsize, k;
8120 /* Maximum free space in an empty page */
8121 pmax = env->me_psize - PAGEHDRSZ;
8123 nsize = mdb_leaf_size(env, newkey, newdata);
8125 nsize = mdb_branch_size(env, newkey);
8126 nsize = EVEN(nsize);
8128 /* grab a page to hold a temporary copy */
8129 copy = mdb_page_malloc(mc->mc_txn, 1);
8134 copy->mp_pgno = mp->mp_pgno;
8135 copy->mp_flags = mp->mp_flags;
8136 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
8137 copy->mp_upper = env->me_psize - PAGEBASE;
8139 /* prepare to insert */
8140 for (i=0, j=0; i<nkeys; i++) {
8142 copy->mp_ptrs[j++] = 0;
8144 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
8147 /* When items are relatively large the split point needs
8148 * to be checked, because being off-by-one will make the
8149 * difference between success or failure in mdb_node_add.
8151 * It's also relevant if a page happens to be laid out
8152 * such that one half of its nodes are all "small" and
8153 * the other half of its nodes are "large." If the new
8154 * item is also "large" and falls on the half with
8155 * "large" nodes, it also may not fit.
8157 * As a final tweak, if the new item goes on the last
8158 * spot on the page (and thus, onto the new page), bias
8159 * the split so the new page is emptier than the old page.
8160 * This yields better packing during sequential inserts.
8162 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
8163 /* Find split point */
8165 if (newindx <= split_indx || newindx >= nkeys) {
8167 k = newindx >= nkeys ? nkeys : split_indx+2;
8172 for (; i!=k; i+=j) {
8177 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8178 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
8180 if (F_ISSET(node->mn_flags, F_BIGDATA))
8181 psize += sizeof(pgno_t);
8183 psize += NODEDSZ(node);
8185 psize = EVEN(psize);
8187 if (psize > pmax || i == k-j) {
8188 split_indx = i + (j<0);
8193 if (split_indx == newindx) {
8194 sepkey.mv_size = newkey->mv_size;
8195 sepkey.mv_data = newkey->mv_data;
8197 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
8198 sepkey.mv_size = node->mn_ksize;
8199 sepkey.mv_data = NODEKEY(node);
8204 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
8206 /* Copy separator key to the parent.
8208 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
8212 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
8217 if (mn.mc_snum == mc->mc_snum) {
8218 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
8219 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
8220 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
8221 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
8226 /* Right page might now have changed parent.
8227 * Check if left page also changed parent.
8229 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8230 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8231 for (i=0; i<ptop; i++) {
8232 mc->mc_pg[i] = mn.mc_pg[i];
8233 mc->mc_ki[i] = mn.mc_ki[i];
8235 mc->mc_pg[ptop] = mn.mc_pg[ptop];
8236 if (mn.mc_ki[ptop]) {
8237 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
8239 /* find right page's left sibling */
8240 mc->mc_ki[ptop] = mn.mc_ki[ptop];
8241 mdb_cursor_sibling(mc, 0);
8246 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
8249 mc->mc_flags ^= C_SPLITTING;
8250 if (rc != MDB_SUCCESS) {
8253 if (nflags & MDB_APPEND) {
8254 mc->mc_pg[mc->mc_top] = rp;
8255 mc->mc_ki[mc->mc_top] = 0;
8256 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
8259 for (i=0; i<mc->mc_top; i++)
8260 mc->mc_ki[i] = mn.mc_ki[i];
8261 } else if (!IS_LEAF2(mp)) {
8263 mc->mc_pg[mc->mc_top] = rp;
8268 rkey.mv_data = newkey->mv_data;
8269 rkey.mv_size = newkey->mv_size;
8275 /* Update index for the new key. */
8276 mc->mc_ki[mc->mc_top] = j;
8278 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
8279 rkey.mv_data = NODEKEY(node);
8280 rkey.mv_size = node->mn_ksize;
8282 xdata.mv_data = NODEDATA(node);
8283 xdata.mv_size = NODEDSZ(node);
8286 pgno = NODEPGNO(node);
8287 flags = node->mn_flags;
8290 if (!IS_LEAF(mp) && j == 0) {
8291 /* First branch index doesn't need key data. */
8295 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
8301 mc->mc_pg[mc->mc_top] = copy;
8306 } while (i != split_indx);
8308 nkeys = NUMKEYS(copy);
8309 for (i=0; i<nkeys; i++)
8310 mp->mp_ptrs[i] = copy->mp_ptrs[i];
8311 mp->mp_lower = copy->mp_lower;
8312 mp->mp_upper = copy->mp_upper;
8313 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
8314 env->me_psize - copy->mp_upper - PAGEBASE);
8316 /* reset back to original page */
8317 if (newindx < split_indx) {
8318 mc->mc_pg[mc->mc_top] = mp;
8319 if (nflags & MDB_RESERVE) {
8320 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
8321 if (!(node->mn_flags & F_BIGDATA))
8322 newdata->mv_data = NODEDATA(node);
8325 mc->mc_pg[mc->mc_top] = rp;
8327 /* Make sure mc_ki is still valid.
8329 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
8330 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
8331 for (i=0; i<=ptop; i++) {
8332 mc->mc_pg[i] = mn.mc_pg[i];
8333 mc->mc_ki[i] = mn.mc_ki[i];
8340 /* Adjust other cursors pointing to mp */
8341 MDB_cursor *m2, *m3;
8342 MDB_dbi dbi = mc->mc_dbi;
8343 int fixup = NUMKEYS(mp);
8345 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
8346 if (mc->mc_flags & C_SUB)
8347 m3 = &m2->mc_xcursor->mx_cursor;
8352 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8354 if (m3->mc_flags & C_SPLITTING)
8359 for (k=m3->mc_top; k>=0; k--) {
8360 m3->mc_ki[k+1] = m3->mc_ki[k];
8361 m3->mc_pg[k+1] = m3->mc_pg[k];
8363 if (m3->mc_ki[0] >= split_indx) {
8368 m3->mc_pg[0] = mc->mc_pg[0];
8372 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8373 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8374 m3->mc_ki[mc->mc_top]++;
8375 if (m3->mc_ki[mc->mc_top] >= fixup) {
8376 m3->mc_pg[mc->mc_top] = rp;
8377 m3->mc_ki[mc->mc_top] -= fixup;
8378 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8380 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8381 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8386 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8389 if (copy) /* tmp page */
8390 mdb_page_free(env, copy);
8392 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8397 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8398 MDB_val *key, MDB_val *data, unsigned int flags)
8403 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8406 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8409 mdb_cursor_init(&mc, txn, dbi, &mx);
8410 return mdb_cursor_put(&mc, key, data, flags);
8414 #define MDB_WBUF (1024*1024)
8417 /** State needed for a compacting copy. */
8418 typedef struct mdb_copy {
8419 pthread_mutex_t mc_mutex;
8420 pthread_cond_t mc_cond;
8427 pgno_t mc_next_pgno;
8430 volatile int mc_new;
8435 /** Dedicated writer thread for compacting copy. */
8436 static THREAD_RET ESECT
8437 mdb_env_copythr(void *arg)
8441 int toggle = 0, wsize, rc;
8444 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8447 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8450 pthread_mutex_lock(&my->mc_mutex);
8452 pthread_cond_signal(&my->mc_cond);
8455 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8456 if (my->mc_new < 0) {
8461 wsize = my->mc_wlen[toggle];
8462 ptr = my->mc_wbuf[toggle];
8465 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8469 } else if (len > 0) {
8483 /* If there's an overflow page tail, write it too */
8484 if (my->mc_olen[toggle]) {
8485 wsize = my->mc_olen[toggle];
8486 ptr = my->mc_over[toggle];
8487 my->mc_olen[toggle] = 0;
8490 my->mc_wlen[toggle] = 0;
8492 pthread_cond_signal(&my->mc_cond);
8494 pthread_cond_signal(&my->mc_cond);
8495 pthread_mutex_unlock(&my->mc_mutex);
8496 return (THREAD_RET)0;
8500 /** Tell the writer thread there's a buffer ready to write */
8502 mdb_env_cthr_toggle(mdb_copy *my, int st)
8504 int toggle = my->mc_toggle ^ 1;
8505 pthread_mutex_lock(&my->mc_mutex);
8506 if (my->mc_status) {
8507 pthread_mutex_unlock(&my->mc_mutex);
8508 return my->mc_status;
8510 while (my->mc_new == 1)
8511 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8513 my->mc_toggle = toggle;
8514 pthread_cond_signal(&my->mc_cond);
8515 pthread_mutex_unlock(&my->mc_mutex);
8519 /** Depth-first tree traversal for compacting copy. */
8521 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8524 MDB_txn *txn = my->mc_txn;
8526 MDB_page *mo, *mp, *leaf;
8531 /* Empty DB, nothing to do */
8532 if (*pg == P_INVALID)
8539 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8542 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8546 /* Make cursor pages writable */
8547 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8551 for (i=0; i<mc.mc_top; i++) {
8552 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8553 mc.mc_pg[i] = (MDB_page *)ptr;
8554 ptr += my->mc_env->me_psize;
8557 /* This is writable space for a leaf page. Usually not needed. */
8558 leaf = (MDB_page *)ptr;
8560 toggle = my->mc_toggle;
8561 while (mc.mc_snum > 0) {
8563 mp = mc.mc_pg[mc.mc_top];
8567 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8568 for (i=0; i<n; i++) {
8569 ni = NODEPTR(mp, i);
8570 if (ni->mn_flags & F_BIGDATA) {
8574 /* Need writable leaf */
8576 mc.mc_pg[mc.mc_top] = leaf;
8577 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8579 ni = NODEPTR(mp, i);
8582 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8583 rc = mdb_page_get(txn, pg, &omp, NULL);
8586 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8587 rc = mdb_env_cthr_toggle(my, 1);
8590 toggle = my->mc_toggle;
8592 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8593 memcpy(mo, omp, my->mc_env->me_psize);
8594 mo->mp_pgno = my->mc_next_pgno;
8595 my->mc_next_pgno += omp->mp_pages;
8596 my->mc_wlen[toggle] += my->mc_env->me_psize;
8597 if (omp->mp_pages > 1) {
8598 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8599 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8600 rc = mdb_env_cthr_toggle(my, 1);
8603 toggle = my->mc_toggle;
8605 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8606 } else if (ni->mn_flags & F_SUBDATA) {
8609 /* Need writable leaf */
8611 mc.mc_pg[mc.mc_top] = leaf;
8612 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8614 ni = NODEPTR(mp, i);
8617 memcpy(&db, NODEDATA(ni), sizeof(db));
8618 my->mc_toggle = toggle;
8619 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8622 toggle = my->mc_toggle;
8623 memcpy(NODEDATA(ni), &db, sizeof(db));
8628 mc.mc_ki[mc.mc_top]++;
8629 if (mc.mc_ki[mc.mc_top] < n) {
8632 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8634 rc = mdb_page_get(txn, pg, &mp, NULL);
8639 mc.mc_ki[mc.mc_top] = 0;
8640 if (IS_BRANCH(mp)) {
8641 /* Whenever we advance to a sibling branch page,
8642 * we must proceed all the way down to its first leaf.
8644 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8647 mc.mc_pg[mc.mc_top] = mp;
8651 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8652 rc = mdb_env_cthr_toggle(my, 1);
8655 toggle = my->mc_toggle;
8657 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8658 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8659 mo->mp_pgno = my->mc_next_pgno++;
8660 my->mc_wlen[toggle] += my->mc_env->me_psize;
8662 /* Update parent if there is one */
8663 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8664 SETPGNO(ni, mo->mp_pgno);
8665 mdb_cursor_pop(&mc);
8667 /* Otherwise we're done */
8677 /** Copy environment with compaction. */
8679 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8684 MDB_txn *txn = NULL;
8689 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8690 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8691 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_os_psize);
8692 if (my.mc_wbuf[0] == NULL)
8695 pthread_mutex_init(&my.mc_mutex, NULL);
8696 pthread_cond_init(&my.mc_cond, NULL);
8697 #ifdef HAVE_MEMALIGN
8698 my.mc_wbuf[0] = memalign(env->me_os_psize, MDB_WBUF*2);
8699 if (my.mc_wbuf[0] == NULL)
8702 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_os_psize, MDB_WBUF*2);
8707 memset(my.mc_wbuf[0], 0, MDB_WBUF*2);
8708 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8713 my.mc_next_pgno = 2;
8719 THREAD_CREATE(thr, mdb_env_copythr, &my);
8721 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8725 mp = (MDB_page *)my.mc_wbuf[0];
8726 memset(mp, 0, 2*env->me_psize);
8728 mp->mp_flags = P_META;
8729 mm = (MDB_meta *)METADATA(mp);
8730 mdb_env_init_meta0(env, mm);
8731 mm->mm_address = env->me_metas[0]->mm_address;
8733 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8735 mp->mp_flags = P_META;
8736 *(MDB_meta *)METADATA(mp) = *mm;
8737 mm = (MDB_meta *)METADATA(mp);
8739 /* Count the number of free pages, subtract from lastpg to find
8740 * number of active pages
8743 MDB_ID freecount = 0;
8746 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8747 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8748 freecount += *(MDB_ID *)data.mv_data;
8749 freecount += txn->mt_dbs[0].md_branch_pages +
8750 txn->mt_dbs[0].md_leaf_pages +
8751 txn->mt_dbs[0].md_overflow_pages;
8753 /* Set metapage 1 */
8754 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8755 mm->mm_dbs[1] = txn->mt_dbs[1];
8756 if (mm->mm_last_pg > 1) {
8757 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8760 mm->mm_dbs[1].md_root = P_INVALID;
8763 my.mc_wlen[0] = env->me_psize * 2;
8765 pthread_mutex_lock(&my.mc_mutex);
8767 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8768 pthread_mutex_unlock(&my.mc_mutex);
8769 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8770 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8771 rc = mdb_env_cthr_toggle(&my, 1);
8772 mdb_env_cthr_toggle(&my, -1);
8773 pthread_mutex_lock(&my.mc_mutex);
8775 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8776 pthread_mutex_unlock(&my.mc_mutex);
8781 CloseHandle(my.mc_cond);
8782 CloseHandle(my.mc_mutex);
8783 _aligned_free(my.mc_wbuf[0]);
8785 pthread_cond_destroy(&my.mc_cond);
8786 pthread_mutex_destroy(&my.mc_mutex);
8787 free(my.mc_wbuf[0]);
8792 /** Copy environment as-is. */
8794 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8796 MDB_txn *txn = NULL;
8802 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8806 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8809 /* Do the lock/unlock of the reader mutex before starting the
8810 * write txn. Otherwise other read txns could block writers.
8812 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8817 /* We must start the actual read txn after blocking writers */
8818 mdb_txn_reset0(txn, "reset-stage1");
8820 /* Temporarily block writers until we snapshot the meta pages */
8823 rc = mdb_txn_renew0(txn);
8825 UNLOCK_MUTEX_W(env);
8830 wsize = env->me_psize * 2;
8834 DO_WRITE(rc, fd, ptr, w2, len);
8838 } else if (len > 0) {
8844 /* Non-blocking or async handles are not supported */
8850 UNLOCK_MUTEX_W(env);
8855 w2 = txn->mt_next_pgno * env->me_psize;
8858 if ((rc = mdb_fsize(env->me_fd, &fsize)))
8865 if (wsize > MAX_WRITE)
8869 DO_WRITE(rc, fd, ptr, w2, len);
8873 } else if (len > 0) {
8890 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8892 if (flags & MDB_CP_COMPACT)
8893 return mdb_env_copyfd1(env, fd);
8895 return mdb_env_copyfd0(env, fd);
8899 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8901 return mdb_env_copyfd2(env, fd, 0);
8905 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8909 HANDLE newfd = INVALID_HANDLE_VALUE;
8911 if (env->me_flags & MDB_NOSUBDIR) {
8912 lpath = (char *)path;
8915 len += sizeof(DATANAME);
8916 lpath = malloc(len);
8919 sprintf(lpath, "%s" DATANAME, path);
8922 /* The destination path must exist, but the destination file must not.
8923 * We don't want the OS to cache the writes, since the source data is
8924 * already in the OS cache.
8927 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8928 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8930 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8932 if (newfd == INVALID_HANDLE_VALUE) {
8937 if (env->me_psize >= env->me_os_psize) {
8939 /* Set O_DIRECT if the file system supports it */
8940 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8941 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8943 #ifdef F_NOCACHE /* __APPLE__ */
8944 rc = fcntl(newfd, F_NOCACHE, 1);
8952 rc = mdb_env_copyfd2(env, newfd, flags);
8955 if (!(env->me_flags & MDB_NOSUBDIR))
8957 if (newfd != INVALID_HANDLE_VALUE)
8958 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8965 mdb_env_copy(MDB_env *env, const char *path)
8967 return mdb_env_copy2(env, path, 0);
8971 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8973 if ((flag & CHANGEABLE) != flag)
8976 env->me_flags |= flag;
8978 env->me_flags &= ~flag;
8983 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8988 *arg = env->me_flags;
8993 mdb_env_set_userctx(MDB_env *env, void *ctx)
8997 env->me_userctx = ctx;
9002 mdb_env_get_userctx(MDB_env *env)
9004 return env ? env->me_userctx : NULL;
9008 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
9013 env->me_assert_func = func;
9019 mdb_env_get_path(MDB_env *env, const char **arg)
9024 *arg = env->me_path;
9029 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
9038 /** Common code for #mdb_stat() and #mdb_env_stat().
9039 * @param[in] env the environment to operate in.
9040 * @param[in] db the #MDB_db record containing the stats to return.
9041 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
9042 * @return 0, this function always succeeds.
9045 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
9047 arg->ms_psize = env->me_psize;
9048 arg->ms_depth = db->md_depth;
9049 arg->ms_branch_pages = db->md_branch_pages;
9050 arg->ms_leaf_pages = db->md_leaf_pages;
9051 arg->ms_overflow_pages = db->md_overflow_pages;
9052 arg->ms_entries = db->md_entries;
9058 mdb_env_stat(MDB_env *env, MDB_stat *arg)
9062 if (env == NULL || arg == NULL)
9065 toggle = mdb_env_pick_meta(env);
9067 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
9071 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
9075 if (env == NULL || arg == NULL)
9078 toggle = mdb_env_pick_meta(env);
9079 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
9080 arg->me_mapsize = env->me_mapsize;
9081 arg->me_maxreaders = env->me_maxreaders;
9083 /* me_numreaders may be zero if this process never used any readers. Use
9084 * the shared numreader count if it exists.
9086 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
9088 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
9089 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
9093 /** Set the default comparison functions for a database.
9094 * Called immediately after a database is opened to set the defaults.
9095 * The user can then override them with #mdb_set_compare() or
9096 * #mdb_set_dupsort().
9097 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
9098 * @param[in] dbi A database handle returned by #mdb_dbi_open()
9101 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
9103 uint16_t f = txn->mt_dbs[dbi].md_flags;
9105 txn->mt_dbxs[dbi].md_cmp =
9106 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
9107 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
9109 txn->mt_dbxs[dbi].md_dcmp =
9110 !(f & MDB_DUPSORT) ? 0 :
9111 ((f & MDB_INTEGERDUP)
9112 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
9113 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
9116 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
9122 int rc, dbflag, exact;
9123 unsigned int unused = 0, seq;
9126 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
9127 mdb_default_cmp(txn, FREE_DBI);
9130 if ((flags & VALID_FLAGS) != flags)
9132 if (txn->mt_flags & MDB_TXN_ERROR)
9138 if (flags & PERSISTENT_FLAGS) {
9139 uint16_t f2 = flags & PERSISTENT_FLAGS;
9140 /* make sure flag changes get committed */
9141 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
9142 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
9143 txn->mt_flags |= MDB_TXN_DIRTY;
9146 mdb_default_cmp(txn, MAIN_DBI);
9150 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
9151 mdb_default_cmp(txn, MAIN_DBI);
9154 /* Is the DB already open? */
9156 for (i=2; i<txn->mt_numdbs; i++) {
9157 if (!txn->mt_dbxs[i].md_name.mv_size) {
9158 /* Remember this free slot */
9159 if (!unused) unused = i;
9162 if (len == txn->mt_dbxs[i].md_name.mv_size &&
9163 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
9169 /* If no free slot and max hit, fail */
9170 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
9171 return MDB_DBS_FULL;
9173 /* Cannot mix named databases with some mainDB flags */
9174 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
9175 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
9177 /* Find the DB info */
9178 dbflag = DB_NEW|DB_VALID;
9181 key.mv_data = (void *)name;
9182 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
9183 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
9184 if (rc == MDB_SUCCESS) {
9185 /* make sure this is actually a DB */
9186 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
9187 if ((node->mn_flags & (F_DUPDATA|F_SUBDATA)) != F_SUBDATA)
9188 return MDB_INCOMPATIBLE;
9189 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
9190 /* Create if requested */
9191 data.mv_size = sizeof(MDB_db);
9192 data.mv_data = &dummy;
9193 memset(&dummy, 0, sizeof(dummy));
9194 dummy.md_root = P_INVALID;
9195 dummy.md_flags = flags & PERSISTENT_FLAGS;
9196 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
9200 /* OK, got info, add to table */
9201 if (rc == MDB_SUCCESS) {
9202 unsigned int slot = unused ? unused : txn->mt_numdbs;
9203 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
9204 txn->mt_dbxs[slot].md_name.mv_size = len;
9205 txn->mt_dbxs[slot].md_rel = NULL;
9206 txn->mt_dbflags[slot] = dbflag;
9207 /* txn-> and env-> are the same in read txns, use
9208 * tmp variable to avoid undefined assignment
9210 seq = ++txn->mt_env->me_dbiseqs[slot];
9211 txn->mt_dbiseqs[slot] = seq;
9213 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
9215 mdb_default_cmp(txn, slot);
9224 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
9226 if (!arg || !TXN_DBI_EXIST(txn, dbi))
9229 if (txn->mt_flags & MDB_TXN_ERROR)
9232 if (txn->mt_dbflags[dbi] & DB_STALE) {
9235 /* Stale, must read the DB's root. cursor_init does it for us. */
9236 mdb_cursor_init(&mc, txn, dbi, &mx);
9238 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
9241 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
9244 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
9246 ptr = env->me_dbxs[dbi].md_name.mv_data;
9247 /* If there was no name, this was already closed */
9249 env->me_dbxs[dbi].md_name.mv_data = NULL;
9250 env->me_dbxs[dbi].md_name.mv_size = 0;
9251 env->me_dbflags[dbi] = 0;
9252 env->me_dbiseqs[dbi]++;
9257 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
9259 /* We could return the flags for the FREE_DBI too but what's the point? */
9260 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9262 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
9266 /** Add all the DB's pages to the free list.
9267 * @param[in] mc Cursor on the DB to free.
9268 * @param[in] subs non-Zero to check for sub-DBs in this DB.
9269 * @return 0 on success, non-zero on failure.
9272 mdb_drop0(MDB_cursor *mc, int subs)
9276 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
9277 if (rc == MDB_SUCCESS) {
9278 MDB_txn *txn = mc->mc_txn;
9283 /* LEAF2 pages have no nodes, cannot have sub-DBs */
9284 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
9287 mdb_cursor_copy(mc, &mx);
9288 while (mc->mc_snum > 0) {
9289 MDB_page *mp = mc->mc_pg[mc->mc_top];
9290 unsigned n = NUMKEYS(mp);
9292 for (i=0; i<n; i++) {
9293 ni = NODEPTR(mp, i);
9294 if (ni->mn_flags & F_BIGDATA) {
9297 memcpy(&pg, NODEDATA(ni), sizeof(pg));
9298 rc = mdb_page_get(txn, pg, &omp, NULL);
9301 mdb_cassert(mc, IS_OVERFLOW(omp));
9302 rc = mdb_midl_append_range(&txn->mt_free_pgs,
9306 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
9307 mdb_xcursor_init1(mc, ni);
9308 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
9314 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
9316 for (i=0; i<n; i++) {
9318 ni = NODEPTR(mp, i);
9321 mdb_midl_xappend(txn->mt_free_pgs, pg);
9326 mc->mc_ki[mc->mc_top] = i;
9327 rc = mdb_cursor_sibling(mc, 1);
9329 if (rc != MDB_NOTFOUND)
9331 /* no more siblings, go back to beginning
9332 * of previous level.
9336 for (i=1; i<mc->mc_snum; i++) {
9338 mc->mc_pg[i] = mx.mc_pg[i];
9343 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
9346 txn->mt_flags |= MDB_TXN_ERROR;
9347 } else if (rc == MDB_NOTFOUND) {
9353 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
9355 MDB_cursor *mc, *m2;
9358 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9361 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
9364 if (dbi > MAIN_DBI && TXN_DBI_CHANGED(txn, dbi))
9367 rc = mdb_cursor_open(txn, dbi, &mc);
9371 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9372 /* Invalidate the dropped DB's cursors */
9373 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9374 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9378 /* Can't delete the main DB */
9379 if (del && dbi > MAIN_DBI) {
9380 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, F_SUBDATA);
9382 txn->mt_dbflags[dbi] = DB_STALE;
9383 mdb_dbi_close(txn->mt_env, dbi);
9385 txn->mt_flags |= MDB_TXN_ERROR;
9388 /* reset the DB record, mark it dirty */
9389 txn->mt_dbflags[dbi] |= DB_DIRTY;
9390 txn->mt_dbs[dbi].md_depth = 0;
9391 txn->mt_dbs[dbi].md_branch_pages = 0;
9392 txn->mt_dbs[dbi].md_leaf_pages = 0;
9393 txn->mt_dbs[dbi].md_overflow_pages = 0;
9394 txn->mt_dbs[dbi].md_entries = 0;
9395 txn->mt_dbs[dbi].md_root = P_INVALID;
9397 txn->mt_flags |= MDB_TXN_DIRTY;
9400 mdb_cursor_close(mc);
9404 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9406 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9409 txn->mt_dbxs[dbi].md_cmp = cmp;
9413 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9415 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9418 txn->mt_dbxs[dbi].md_dcmp = cmp;
9422 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9424 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9427 txn->mt_dbxs[dbi].md_rel = rel;
9431 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9433 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9436 txn->mt_dbxs[dbi].md_relctx = ctx;
9441 mdb_env_get_maxkeysize(MDB_env *env)
9443 return ENV_MAXKEY(env);
9447 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9449 unsigned int i, rdrs;
9452 int rc = 0, first = 1;
9456 if (!env->me_txns) {
9457 return func("(no reader locks)\n", ctx);
9459 rdrs = env->me_txns->mti_numreaders;
9460 mr = env->me_txns->mti_readers;
9461 for (i=0; i<rdrs; i++) {
9463 txnid_t txnid = mr[i].mr_txnid;
9464 sprintf(buf, txnid == (txnid_t)-1 ?
9465 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9466 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9469 rc = func(" pid thread txnid\n", ctx);
9473 rc = func(buf, ctx);
9479 rc = func("(no active readers)\n", ctx);
9484 /** Insert pid into list if not already present.
9485 * return -1 if already present.
9488 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9490 /* binary search of pid in list */
9492 unsigned cursor = 1;
9494 unsigned n = ids[0];
9497 unsigned pivot = n >> 1;
9498 cursor = base + pivot + 1;
9499 val = pid - ids[cursor];
9504 } else if ( val > 0 ) {
9509 /* found, so it's a duplicate */
9518 for (n = ids[0]; n > cursor; n--)
9525 mdb_reader_check(MDB_env *env, int *dead)
9527 unsigned int i, j, rdrs;
9529 MDB_PID_T *pids, pid;
9538 rdrs = env->me_txns->mti_numreaders;
9539 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9543 mr = env->me_txns->mti_readers;
9544 for (i=0; i<rdrs; i++) {
9545 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9547 if (mdb_pid_insert(pids, pid) == 0) {
9548 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9550 /* Recheck, a new process may have reused pid */
9551 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9552 for (j=i; j<rdrs; j++)
9553 if (mr[j].mr_pid == pid) {
9554 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9555 (unsigned) pid, mr[j].mr_txnid));
9560 UNLOCK_MUTEX_R(env);