2 * @brief Lightning memory-mapped database library
4 * A Btree-based database management library modeled loosely on the
5 * BerkeleyDB API, but much simplified.
8 * Copyright 2011-2014 Howard Chu, Symas Corp.
11 * Redistribution and use in source and binary forms, with or without
12 * modification, are permitted only as authorized by the OpenLDAP
15 * A copy of this license is available in the file LICENSE in the
16 * top-level directory of the distribution or, alternatively, at
17 * <http://www.OpenLDAP.org/license.html>.
19 * This code is derived from btree.c written by Martin Hedenfalk.
21 * Copyright (c) 2009, 2010 Martin Hedenfalk <martin@bzero.se>
23 * Permission to use, copy, modify, and distribute this software for any
24 * purpose with or without fee is hereby granted, provided that the above
25 * copyright notice and this permission notice appear in all copies.
27 * THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES
28 * WITH REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF
29 * MERCHANTABILITY AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR
30 * ANY SPECIAL, DIRECT, INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES
31 * WHATSOEVER RESULTING FROM LOSS OF USE, DATA OR PROFITS, WHETHER IN AN
32 * ACTION OF CONTRACT, NEGLIGENCE OR OTHER TORTIOUS ACTION, ARISING OUT OF
33 * OR IN CONNECTION WITH THE USE OR PERFORMANCE OF THIS SOFTWARE.
41 /** getpid() returns int; MinGW defines pid_t but MinGW64 typedefs it
42 * as int64 which is wrong. MSVC doesn't define it at all, so just
46 #define MDB_THR_T DWORD
47 #include <sys/types.h>
50 # include <sys/param.h>
52 # define LITTLE_ENDIAN 1234
53 # define BIG_ENDIAN 4321
54 # define BYTE_ORDER LITTLE_ENDIAN
56 # define SSIZE_MAX INT_MAX
60 #include <sys/types.h>
62 #define MDB_PID_T pid_t
63 #define MDB_THR_T pthread_t
64 #include <sys/param.h>
67 #ifdef HAVE_SYS_FILE_H
83 #if !(defined(BYTE_ORDER) || defined(__BYTE_ORDER))
84 #include <netinet/in.h>
85 #include <resolv.h> /* defines BYTE_ORDER on HPUX and Solaris */
88 #if defined(__APPLE__) || defined (BSD)
89 # define MDB_USE_POSIX_SEM 1
90 # define MDB_FDATASYNC fsync
91 #elif defined(ANDROID)
92 # define MDB_FDATASYNC fsync
97 #ifdef MDB_USE_POSIX_SEM
98 # define MDB_USE_HASH 1
99 #include <semaphore.h>
104 #include <valgrind/memcheck.h>
105 #define VGMEMP_CREATE(h,r,z) VALGRIND_CREATE_MEMPOOL(h,r,z)
106 #define VGMEMP_ALLOC(h,a,s) VALGRIND_MEMPOOL_ALLOC(h,a,s)
107 #define VGMEMP_FREE(h,a) VALGRIND_MEMPOOL_FREE(h,a)
108 #define VGMEMP_DESTROY(h) VALGRIND_DESTROY_MEMPOOL(h)
109 #define VGMEMP_DEFINED(a,s) VALGRIND_MAKE_MEM_DEFINED(a,s)
111 #define VGMEMP_CREATE(h,r,z)
112 #define VGMEMP_ALLOC(h,a,s)
113 #define VGMEMP_FREE(h,a)
114 #define VGMEMP_DESTROY(h)
115 #define VGMEMP_DEFINED(a,s)
119 # if (defined(_LITTLE_ENDIAN) || defined(_BIG_ENDIAN)) && !(defined(_LITTLE_ENDIAN) && defined(_BIG_ENDIAN))
120 /* Solaris just defines one or the other */
121 # define LITTLE_ENDIAN 1234
122 # define BIG_ENDIAN 4321
123 # ifdef _LITTLE_ENDIAN
124 # define BYTE_ORDER LITTLE_ENDIAN
126 # define BYTE_ORDER BIG_ENDIAN
129 # define BYTE_ORDER __BYTE_ORDER
133 #ifndef LITTLE_ENDIAN
134 #define LITTLE_ENDIAN __LITTLE_ENDIAN
137 #define BIG_ENDIAN __BIG_ENDIAN
140 #if defined(__i386) || defined(__x86_64) || defined(_M_IX86)
141 #define MISALIGNED_OK 1
147 #if (BYTE_ORDER == LITTLE_ENDIAN) == (BYTE_ORDER == BIG_ENDIAN)
148 # error "Unknown or unsupported endianness (BYTE_ORDER)"
149 #elif (-6 & 5) || CHAR_BIT != 8 || UINT_MAX < 0xffffffff || ULONG_MAX % 0xFFFF
150 # error "Two's complement, reasonably sized integer types, please"
154 /** Put infrequently used env functions in separate section */
155 #define ESECT __attribute__ ((section("text_env")))
160 /** @defgroup internal LMDB Internals
163 /** @defgroup compat Compatibility Macros
164 * A bunch of macros to minimize the amount of platform-specific ifdefs
165 * needed throughout the rest of the code. When the features this library
166 * needs are similar enough to POSIX to be hidden in a one-or-two line
167 * replacement, this macro approach is used.
171 /* Features under development */
176 /** Wrapper around __func__, which is a C99 feature */
177 #if __STDC_VERSION__ >= 199901L
178 # define mdb_func_ __func__
179 #elif __GNUC__ >= 2 || _MSC_VER >= 1300
180 # define mdb_func_ __FUNCTION__
182 /* If a debug message says <mdb_unknown>(), update the #if statements above */
183 # define mdb_func_ "<mdb_unknown>"
187 #define MDB_USE_HASH 1
188 #define MDB_PIDLOCK 0
189 #define THREAD_RET DWORD
190 #define pthread_t HANDLE
191 #define pthread_mutex_t HANDLE
192 #define pthread_cond_t HANDLE
193 #define pthread_key_t DWORD
194 #define pthread_self() GetCurrentThreadId()
195 #define pthread_key_create(x,y) \
196 ((*(x) = TlsAlloc()) == TLS_OUT_OF_INDEXES ? ErrCode() : 0)
197 #define pthread_key_delete(x) TlsFree(x)
198 #define pthread_getspecific(x) TlsGetValue(x)
199 #define pthread_setspecific(x,y) (TlsSetValue(x,y) ? 0 : ErrCode())
200 #define pthread_mutex_unlock(x) ReleaseMutex(*x)
201 #define pthread_mutex_lock(x) WaitForSingleObject(*x, INFINITE)
202 #define pthread_cond_signal(x) SetEvent(*x)
203 #define pthread_cond_wait(cond,mutex) do{SignalObjectAndWait(*mutex, *cond, INFINITE, FALSE); WaitForSingleObject(*mutex, INFINITE);}while(0)
204 #define THREAD_CREATE(thr,start,arg) thr=CreateThread(NULL,0,start,arg,0,NULL)
205 #define THREAD_FINISH(thr) WaitForSingleObject(thr, INFINITE)
206 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_rmutex)
207 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_rmutex)
208 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_wmutex)
209 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_wmutex)
210 #define getpid() GetCurrentProcessId()
211 #define MDB_FDATASYNC(fd) (!FlushFileBuffers(fd))
212 #define MDB_MSYNC(addr,len,flags) (!FlushViewOfFile(addr,len))
213 #define ErrCode() GetLastError()
214 #define GET_PAGESIZE(x) {SYSTEM_INFO si; GetSystemInfo(&si); (x) = si.dwPageSize;}
215 #define close(fd) (CloseHandle(fd) ? 0 : -1)
216 #define munmap(ptr,len) UnmapViewOfFile(ptr)
217 #ifdef PROCESS_QUERY_LIMITED_INFORMATION
218 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION PROCESS_QUERY_LIMITED_INFORMATION
220 #define MDB_PROCESS_QUERY_LIMITED_INFORMATION 0x1000
224 #define THREAD_RET void *
225 #define THREAD_CREATE(thr,start,arg) pthread_create(&thr,NULL,start,arg)
226 #define THREAD_FINISH(thr) pthread_join(thr,NULL)
227 #define Z "z" /**< printf format modifier for size_t */
229 /** For MDB_LOCK_FORMAT: True if readers take a pid lock in the lockfile */
230 #define MDB_PIDLOCK 1
232 #ifdef MDB_USE_POSIX_SEM
234 #define LOCK_MUTEX_R(env) mdb_sem_wait((env)->me_rmutex)
235 #define UNLOCK_MUTEX_R(env) sem_post((env)->me_rmutex)
236 #define LOCK_MUTEX_W(env) mdb_sem_wait((env)->me_wmutex)
237 #define UNLOCK_MUTEX_W(env) sem_post((env)->me_wmutex)
240 mdb_sem_wait(sem_t *sem)
243 while ((rc = sem_wait(sem)) && (rc = errno) == EINTR) ;
248 /** Lock the reader mutex.
250 #define LOCK_MUTEX_R(env) pthread_mutex_lock(&(env)->me_txns->mti_mutex)
251 /** Unlock the reader mutex.
253 #define UNLOCK_MUTEX_R(env) pthread_mutex_unlock(&(env)->me_txns->mti_mutex)
255 /** Lock the writer mutex.
256 * Only a single write transaction is allowed at a time. Other writers
257 * will block waiting for this mutex.
259 #define LOCK_MUTEX_W(env) pthread_mutex_lock(&(env)->me_txns->mti_wmutex)
260 /** Unlock the writer mutex.
262 #define UNLOCK_MUTEX_W(env) pthread_mutex_unlock(&(env)->me_txns->mti_wmutex)
263 #endif /* MDB_USE_POSIX_SEM */
265 /** Get the error code for the last failed system function.
267 #define ErrCode() errno
269 /** An abstraction for a file handle.
270 * On POSIX systems file handles are small integers. On Windows
271 * they're opaque pointers.
275 /** A value for an invalid file handle.
276 * Mainly used to initialize file variables and signify that they are
279 #define INVALID_HANDLE_VALUE (-1)
281 /** Get the size of a memory page for the system.
282 * This is the basic size that the platform's memory manager uses, and is
283 * fundamental to the use of memory-mapped files.
285 #define GET_PAGESIZE(x) ((x) = sysconf(_SC_PAGE_SIZE))
288 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
291 #define MNAME_LEN (sizeof(pthread_mutex_t))
297 /** A flag for opening a file and requesting synchronous data writes.
298 * This is only used when writing a meta page. It's not strictly needed;
299 * we could just do a normal write and then immediately perform a flush.
300 * But if this flag is available it saves us an extra system call.
302 * @note If O_DSYNC is undefined but exists in /usr/include,
303 * preferably set some compiler flag to get the definition.
304 * Otherwise compile with the less efficient -DMDB_DSYNC=O_SYNC.
307 # define MDB_DSYNC O_DSYNC
311 /** Function for flushing the data of a file. Define this to fsync
312 * if fdatasync() is not supported.
314 #ifndef MDB_FDATASYNC
315 # define MDB_FDATASYNC fdatasync
319 # define MDB_MSYNC(addr,len,flags) msync(addr,len,flags)
330 /** A page number in the database.
331 * Note that 64 bit page numbers are overkill, since pages themselves
332 * already represent 12-13 bits of addressable memory, and the OS will
333 * always limit applications to a maximum of 63 bits of address space.
335 * @note In the #MDB_node structure, we only store 48 bits of this value,
336 * which thus limits us to only 60 bits of addressable data.
338 typedef MDB_ID pgno_t;
340 /** A transaction ID.
341 * See struct MDB_txn.mt_txnid for details.
343 typedef MDB_ID txnid_t;
345 /** @defgroup debug Debug Macros
349 /** Enable debug output. Needs variable argument macros (a C99 feature).
350 * Set this to 1 for copious tracing. Set to 2 to add dumps of all IDLs
351 * read from and written to the database (used for free space management).
357 static int mdb_debug;
358 static txnid_t mdb_debug_start;
360 /** Print a debug message with printf formatting.
361 * Requires double parenthesis around 2 or more args.
363 # define DPRINTF(args) ((void) ((mdb_debug) && DPRINTF0 args))
364 # define DPRINTF0(fmt, ...) \
365 fprintf(stderr, "%s:%d " fmt "\n", mdb_func_, __LINE__, __VA_ARGS__)
367 # define DPRINTF(args) ((void) 0)
369 /** Print a debug string.
370 * The string is printed literally, with no format processing.
372 #define DPUTS(arg) DPRINTF(("%s", arg))
373 /** Debuging output value of a cursor DBI: Negative in a sub-cursor. */
375 (((mc)->mc_flags & C_SUB) ? -(int)(mc)->mc_dbi : (int)(mc)->mc_dbi)
378 /** @brief The maximum size of a database page.
380 * It is 32k or 64k, since value-PAGEBASE must fit in
381 * #MDB_page.%mp_upper.
383 * LMDB will use database pages < OS pages if needed.
384 * That causes more I/O in write transactions: The OS must
385 * know (read) the whole page before writing a partial page.
387 * Note that we don't currently support Huge pages. On Linux,
388 * regular data files cannot use Huge pages, and in general
389 * Huge pages aren't actually pageable. We rely on the OS
390 * demand-pager to read our data and page it out when memory
391 * pressure from other processes is high. So until OSs have
392 * actual paging support for Huge pages, they're not viable.
394 #define MAX_PAGESIZE (PAGEBASE ? 0x10000 : 0x8000)
396 /** The minimum number of keys required in a database page.
397 * Setting this to a larger value will place a smaller bound on the
398 * maximum size of a data item. Data items larger than this size will
399 * be pushed into overflow pages instead of being stored directly in
400 * the B-tree node. This value used to default to 4. With a page size
401 * of 4096 bytes that meant that any item larger than 1024 bytes would
402 * go into an overflow page. That also meant that on average 2-3KB of
403 * each overflow page was wasted space. The value cannot be lower than
404 * 2 because then there would no longer be a tree structure. With this
405 * value, items larger than 2KB will go into overflow pages, and on
406 * average only 1KB will be wasted.
408 #define MDB_MINKEYS 2
410 /** A stamp that identifies a file as an LMDB file.
411 * There's nothing special about this value other than that it is easily
412 * recognizable, and it will reflect any byte order mismatches.
414 #define MDB_MAGIC 0xBEEFC0DE
416 /** The version number for a database's datafile format. */
417 #define MDB_DATA_VERSION ((MDB_DEVEL) ? 999 : 1)
418 /** The version number for a database's lockfile format. */
419 #define MDB_LOCK_VERSION 1
421 /** @brief The max size of a key we can write, or 0 for dynamic max.
423 * Define this as 0 to compute the max from the page size. 511
424 * is default for backwards compat: liblmdb <= 0.9.10 can break
425 * when modifying a DB with keys/dupsort data bigger than its max.
427 * Data items in an #MDB_DUPSORT database are also limited to
428 * this size, since they're actually keys of a sub-DB. Keys and
429 * #MDB_DUPSORT data items must fit on a node in a regular page.
431 #ifndef MDB_MAXKEYSIZE
432 #define MDB_MAXKEYSIZE 511
435 /** The maximum size of a key we can write to the environment. */
437 #define ENV_MAXKEY(env) (MDB_MAXKEYSIZE)
439 #define ENV_MAXKEY(env) ((env)->me_maxkey)
442 /** @brief The maximum size of a data item.
444 * We only store a 32 bit value for node sizes.
446 #define MAXDATASIZE 0xffffffffUL
449 /** Key size which fits in a #DKBUF.
452 #define DKBUF_MAXKEYSIZE ((MDB_MAXKEYSIZE) > 0 ? (MDB_MAXKEYSIZE) : 511)
455 * This is used for printing a hex dump of a key's contents.
457 #define DKBUF char kbuf[DKBUF_MAXKEYSIZE*2+1]
458 /** Display a key in hex.
460 * Invoke a function to display a key in hex.
462 #define DKEY(x) mdb_dkey(x, kbuf)
468 /** An invalid page number.
469 * Mainly used to denote an empty tree.
471 #define P_INVALID (~(pgno_t)0)
473 /** Test if the flags \b f are set in a flag word \b w. */
474 #define F_ISSET(w, f) (((w) & (f)) == (f))
476 /** Round \b n up to an even number. */
477 #define EVEN(n) (((n) + 1U) & -2) /* sign-extending -2 to match n+1U */
479 /** Used for offsets within a single page.
480 * Since memory pages are typically 4 or 8KB in size, 12-13 bits,
483 typedef uint16_t indx_t;
485 /** Default size of memory map.
486 * This is certainly too small for any actual applications. Apps should always set
487 * the size explicitly using #mdb_env_set_mapsize().
489 #define DEFAULT_MAPSIZE 1048576
491 /** @defgroup readers Reader Lock Table
492 * Readers don't acquire any locks for their data access. Instead, they
493 * simply record their transaction ID in the reader table. The reader
494 * mutex is needed just to find an empty slot in the reader table. The
495 * slot's address is saved in thread-specific data so that subsequent read
496 * transactions started by the same thread need no further locking to proceed.
498 * If #MDB_NOTLS is set, the slot address is not saved in thread-specific data.
500 * No reader table is used if the database is on a read-only filesystem, or
501 * if #MDB_NOLOCK is set.
503 * Since the database uses multi-version concurrency control, readers don't
504 * actually need any locking. This table is used to keep track of which
505 * readers are using data from which old transactions, so that we'll know
506 * when a particular old transaction is no longer in use. Old transactions
507 * that have discarded any data pages can then have those pages reclaimed
508 * for use by a later write transaction.
510 * The lock table is constructed such that reader slots are aligned with the
511 * processor's cache line size. Any slot is only ever used by one thread.
512 * This alignment guarantees that there will be no contention or cache
513 * thrashing as threads update their own slot info, and also eliminates
514 * any need for locking when accessing a slot.
516 * A writer thread will scan every slot in the table to determine the oldest
517 * outstanding reader transaction. Any freed pages older than this will be
518 * reclaimed by the writer. The writer doesn't use any locks when scanning
519 * this table. This means that there's no guarantee that the writer will
520 * see the most up-to-date reader info, but that's not required for correct
521 * operation - all we need is to know the upper bound on the oldest reader,
522 * we don't care at all about the newest reader. So the only consequence of
523 * reading stale information here is that old pages might hang around a
524 * while longer before being reclaimed. That's actually good anyway, because
525 * the longer we delay reclaiming old pages, the more likely it is that a
526 * string of contiguous pages can be found after coalescing old pages from
527 * many old transactions together.
530 /** Number of slots in the reader table.
531 * This value was chosen somewhat arbitrarily. 126 readers plus a
532 * couple mutexes fit exactly into 8KB on my development machine.
533 * Applications should set the table size using #mdb_env_set_maxreaders().
535 #define DEFAULT_READERS 126
537 /** The size of a CPU cache line in bytes. We want our lock structures
538 * aligned to this size to avoid false cache line sharing in the
540 * This value works for most CPUs. For Itanium this should be 128.
546 /** The information we store in a single slot of the reader table.
547 * In addition to a transaction ID, we also record the process and
548 * thread ID that owns a slot, so that we can detect stale information,
549 * e.g. threads or processes that went away without cleaning up.
550 * @note We currently don't check for stale records. We simply re-init
551 * the table when we know that we're the only process opening the
554 typedef struct MDB_rxbody {
555 /** Current Transaction ID when this transaction began, or (txnid_t)-1.
556 * Multiple readers that start at the same time will probably have the
557 * same ID here. Again, it's not important to exclude them from
558 * anything; all we need to know is which version of the DB they
559 * started from so we can avoid overwriting any data used in that
560 * particular version.
563 /** The process ID of the process owning this reader txn. */
565 /** The thread ID of the thread owning this txn. */
569 /** The actual reader record, with cacheline padding. */
570 typedef struct MDB_reader {
573 /** shorthand for mrb_txnid */
574 #define mr_txnid mru.mrx.mrb_txnid
575 #define mr_pid mru.mrx.mrb_pid
576 #define mr_tid mru.mrx.mrb_tid
577 /** cache line alignment */
578 char pad[(sizeof(MDB_rxbody)+CACHELINE-1) & ~(CACHELINE-1)];
582 /** The header for the reader table.
583 * The table resides in a memory-mapped file. (This is a different file
584 * than is used for the main database.)
586 * For POSIX the actual mutexes reside in the shared memory of this
587 * mapped file. On Windows, mutexes are named objects allocated by the
588 * kernel; we store the mutex names in this mapped file so that other
589 * processes can grab them. This same approach is also used on
590 * MacOSX/Darwin (using named semaphores) since MacOSX doesn't support
591 * process-shared POSIX mutexes. For these cases where a named object
592 * is used, the object name is derived from a 64 bit FNV hash of the
593 * environment pathname. As such, naming collisions are extremely
594 * unlikely. If a collision occurs, the results are unpredictable.
596 typedef struct MDB_txbody {
597 /** Stamp identifying this as an LMDB file. It must be set
600 /** Format of this lock file. Must be set to #MDB_LOCK_FORMAT. */
602 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
603 char mtb_rmname[MNAME_LEN];
605 /** Mutex protecting access to this table.
606 * This is the reader lock that #LOCK_MUTEX_R acquires.
608 pthread_mutex_t mtb_mutex;
610 /** The ID of the last transaction committed to the database.
611 * This is recorded here only for convenience; the value can always
612 * be determined by reading the main database meta pages.
615 /** The number of slots that have been used in the reader table.
616 * This always records the maximum count, it is not decremented
617 * when readers release their slots.
619 unsigned mtb_numreaders;
622 /** The actual reader table definition. */
623 typedef struct MDB_txninfo {
626 #define mti_magic mt1.mtb.mtb_magic
627 #define mti_format mt1.mtb.mtb_format
628 #define mti_mutex mt1.mtb.mtb_mutex
629 #define mti_rmname mt1.mtb.mtb_rmname
630 #define mti_txnid mt1.mtb.mtb_txnid
631 #define mti_numreaders mt1.mtb.mtb_numreaders
632 char pad[(sizeof(MDB_txbody)+CACHELINE-1) & ~(CACHELINE-1)];
635 #if defined(_WIN32) || defined(MDB_USE_POSIX_SEM)
636 char mt2_wmname[MNAME_LEN];
637 #define mti_wmname mt2.mt2_wmname
639 pthread_mutex_t mt2_wmutex;
640 #define mti_wmutex mt2.mt2_wmutex
642 char pad[(MNAME_LEN+CACHELINE-1) & ~(CACHELINE-1)];
644 MDB_reader mti_readers[1];
647 /** Lockfile format signature: version, features and field layout */
648 #define MDB_LOCK_FORMAT \
650 ((MDB_LOCK_VERSION) \
651 /* Flags which describe functionality */ \
652 + (((MDB_PIDLOCK) != 0) << 16)))
655 /** Common header for all page types.
656 * Overflow records occupy a number of contiguous pages with no
657 * headers on any page after the first.
659 typedef struct MDB_page {
660 #define mp_pgno mp_p.p_pgno
661 #define mp_next mp_p.p_next
663 pgno_t p_pgno; /**< page number */
664 struct MDB_page *p_next; /**< for in-memory list of freed pages */
667 /** @defgroup mdb_page Page Flags
669 * Flags for the page headers.
672 #define P_BRANCH 0x01 /**< branch page */
673 #define P_LEAF 0x02 /**< leaf page */
674 #define P_OVERFLOW 0x04 /**< overflow page */
675 #define P_META 0x08 /**< meta page */
676 #define P_DIRTY 0x10 /**< dirty page, also set for #P_SUBP pages */
677 #define P_LEAF2 0x20 /**< for #MDB_DUPFIXED records */
678 #define P_SUBP 0x40 /**< for #MDB_DUPSORT sub-pages */
679 #define P_LOOSE 0x4000 /**< page was dirtied then freed, can be reused */
680 #define P_KEEP 0x8000 /**< leave this page alone during spill */
682 uint16_t mp_flags; /**< @ref mdb_page */
683 #define mp_lower mp_pb.pb.pb_lower
684 #define mp_upper mp_pb.pb.pb_upper
685 #define mp_pages mp_pb.pb_pages
688 indx_t pb_lower; /**< lower bound of free space */
689 indx_t pb_upper; /**< upper bound of free space */
691 uint32_t pb_pages; /**< number of overflow pages */
693 indx_t mp_ptrs[1]; /**< dynamic size */
696 /** Size of the page header, excluding dynamic data at the end */
697 #define PAGEHDRSZ ((unsigned) offsetof(MDB_page, mp_ptrs))
699 /** Address of first usable data byte in a page, after the header */
700 #define METADATA(p) ((void *)((char *)(p) + PAGEHDRSZ))
702 /** ITS#7713, change PAGEBASE to handle 65536 byte pages */
703 #define PAGEBASE ((MDB_DEVEL) ? PAGEHDRSZ : 0)
705 /** Number of nodes on a page */
706 #define NUMKEYS(p) (((p)->mp_lower - (PAGEHDRSZ-PAGEBASE)) >> 1)
708 /** The amount of space remaining in the page */
709 #define SIZELEFT(p) (indx_t)((p)->mp_upper - (p)->mp_lower)
711 /** The percentage of space used in the page, in tenths of a percent. */
712 #define PAGEFILL(env, p) (1000L * ((env)->me_psize - PAGEHDRSZ - SIZELEFT(p)) / \
713 ((env)->me_psize - PAGEHDRSZ))
714 /** The minimum page fill factor, in tenths of a percent.
715 * Pages emptier than this are candidates for merging.
717 #define FILL_THRESHOLD 250
719 /** Test if a page is a leaf page */
720 #define IS_LEAF(p) F_ISSET((p)->mp_flags, P_LEAF)
721 /** Test if a page is a LEAF2 page */
722 #define IS_LEAF2(p) F_ISSET((p)->mp_flags, P_LEAF2)
723 /** Test if a page is a branch page */
724 #define IS_BRANCH(p) F_ISSET((p)->mp_flags, P_BRANCH)
725 /** Test if a page is an overflow page */
726 #define IS_OVERFLOW(p) F_ISSET((p)->mp_flags, P_OVERFLOW)
727 /** Test if a page is a sub page */
728 #define IS_SUBP(p) F_ISSET((p)->mp_flags, P_SUBP)
730 /** The number of overflow pages needed to store the given size. */
731 #define OVPAGES(size, psize) ((PAGEHDRSZ-1 + (size)) / (psize) + 1)
733 /** Link in #MDB_txn.%mt_loose_pages list */
734 #define NEXT_LOOSE_PAGE(p) (*(MDB_page **)((p) + 2))
736 /** Header for a single key/data pair within a page.
737 * Used in pages of type #P_BRANCH and #P_LEAF without #P_LEAF2.
738 * We guarantee 2-byte alignment for 'MDB_node's.
740 typedef struct MDB_node {
741 /** lo and hi are used for data size on leaf nodes and for
742 * child pgno on branch nodes. On 64 bit platforms, flags
743 * is also used for pgno. (Branch nodes have no flags).
744 * They are in host byte order in case that lets some
745 * accesses be optimized into a 32-bit word access.
747 #if BYTE_ORDER == LITTLE_ENDIAN
748 unsigned short mn_lo, mn_hi; /**< part of data size or pgno */
750 unsigned short mn_hi, mn_lo;
752 /** @defgroup mdb_node Node Flags
754 * Flags for node headers.
757 #define F_BIGDATA 0x01 /**< data put on overflow page */
758 #define F_SUBDATA 0x02 /**< data is a sub-database */
759 #define F_DUPDATA 0x04 /**< data has duplicates */
761 /** valid flags for #mdb_node_add() */
762 #define NODE_ADD_FLAGS (F_DUPDATA|F_SUBDATA|MDB_RESERVE|MDB_APPEND)
765 unsigned short mn_flags; /**< @ref mdb_node */
766 unsigned short mn_ksize; /**< key size */
767 char mn_data[1]; /**< key and data are appended here */
770 /** Size of the node header, excluding dynamic data at the end */
771 #define NODESIZE offsetof(MDB_node, mn_data)
773 /** Bit position of top word in page number, for shifting mn_flags */
774 #define PGNO_TOPWORD ((pgno_t)-1 > 0xffffffffu ? 32 : 0)
776 /** Size of a node in a branch page with a given key.
777 * This is just the node header plus the key, there is no data.
779 #define INDXSIZE(k) (NODESIZE + ((k) == NULL ? 0 : (k)->mv_size))
781 /** Size of a node in a leaf page with a given key and data.
782 * This is node header plus key plus data size.
784 #define LEAFSIZE(k, d) (NODESIZE + (k)->mv_size + (d)->mv_size)
786 /** Address of node \b i in page \b p */
787 #define NODEPTR(p, i) ((MDB_node *)((char *)(p) + (p)->mp_ptrs[i] + PAGEBASE))
789 /** Address of the key for the node */
790 #define NODEKEY(node) (void *)((node)->mn_data)
792 /** Address of the data for a node */
793 #define NODEDATA(node) (void *)((char *)(node)->mn_data + (node)->mn_ksize)
795 /** Get the page number pointed to by a branch node */
796 #define NODEPGNO(node) \
797 ((node)->mn_lo | ((pgno_t) (node)->mn_hi << 16) | \
798 (PGNO_TOPWORD ? ((pgno_t) (node)->mn_flags << PGNO_TOPWORD) : 0))
799 /** Set the page number in a branch node */
800 #define SETPGNO(node,pgno) do { \
801 (node)->mn_lo = (pgno) & 0xffff; (node)->mn_hi = (pgno) >> 16; \
802 if (PGNO_TOPWORD) (node)->mn_flags = (pgno) >> PGNO_TOPWORD; } while(0)
804 /** Get the size of the data in a leaf node */
805 #define NODEDSZ(node) ((node)->mn_lo | ((unsigned)(node)->mn_hi << 16))
806 /** Set the size of the data for a leaf node */
807 #define SETDSZ(node,size) do { \
808 (node)->mn_lo = (size) & 0xffff; (node)->mn_hi = (size) >> 16;} while(0)
809 /** The size of a key in a node */
810 #define NODEKSZ(node) ((node)->mn_ksize)
812 /** Copy a page number from src to dst */
814 #define COPY_PGNO(dst,src) dst = src
816 #if SIZE_MAX > 4294967295UL
817 #define COPY_PGNO(dst,src) do { \
818 unsigned short *s, *d; \
819 s = (unsigned short *)&(src); \
820 d = (unsigned short *)&(dst); \
827 #define COPY_PGNO(dst,src) do { \
828 unsigned short *s, *d; \
829 s = (unsigned short *)&(src); \
830 d = (unsigned short *)&(dst); \
836 /** The address of a key in a LEAF2 page.
837 * LEAF2 pages are used for #MDB_DUPFIXED sorted-duplicate sub-DBs.
838 * There are no node headers, keys are stored contiguously.
840 #define LEAF2KEY(p, i, ks) ((char *)(p) + PAGEHDRSZ + ((i)*(ks)))
842 /** Set the \b node's key into \b keyptr, if requested. */
843 #define MDB_GET_KEY(node, keyptr) { if ((keyptr) != NULL) { \
844 (keyptr)->mv_size = NODEKSZ(node); (keyptr)->mv_data = NODEKEY(node); } }
846 /** Set the \b node's key into \b key. */
847 #define MDB_GET_KEY2(node, key) { key.mv_size = NODEKSZ(node); key.mv_data = NODEKEY(node); }
849 /** Information about a single database in the environment. */
850 typedef struct MDB_db {
851 uint32_t md_pad; /**< also ksize for LEAF2 pages */
852 uint16_t md_flags; /**< @ref mdb_dbi_open */
853 uint16_t md_depth; /**< depth of this tree */
854 pgno_t md_branch_pages; /**< number of internal pages */
855 pgno_t md_leaf_pages; /**< number of leaf pages */
856 pgno_t md_overflow_pages; /**< number of overflow pages */
857 size_t md_entries; /**< number of data items */
858 pgno_t md_root; /**< the root page of this tree */
861 /** mdb_dbi_open flags */
862 #define MDB_VALID 0x8000 /**< DB handle is valid, for me_dbflags */
863 #define PERSISTENT_FLAGS (0xffff & ~(MDB_VALID))
864 #define VALID_FLAGS (MDB_REVERSEKEY|MDB_DUPSORT|MDB_INTEGERKEY|MDB_DUPFIXED|\
865 MDB_INTEGERDUP|MDB_REVERSEDUP|MDB_CREATE)
867 /** Handle for the DB used to track free pages. */
869 /** Handle for the default DB. */
872 /** Meta page content.
873 * A meta page is the start point for accessing a database snapshot.
874 * Pages 0-1 are meta pages. Transaction N writes meta page #(N % 2).
876 typedef struct MDB_meta {
877 /** Stamp identifying this as an LMDB file. It must be set
880 /** Version number of this lock file. Must be set to #MDB_DATA_VERSION. */
882 void *mm_address; /**< address for fixed mapping */
883 size_t mm_mapsize; /**< size of mmap region */
884 MDB_db mm_dbs[2]; /**< first is free space, 2nd is main db */
885 /** The size of pages used in this DB */
886 #define mm_psize mm_dbs[0].md_pad
887 /** Any persistent environment flags. @ref mdb_env */
888 #define mm_flags mm_dbs[0].md_flags
889 pgno_t mm_last_pg; /**< last used page in file */
890 txnid_t mm_txnid; /**< txnid that committed this page */
893 /** Buffer for a stack-allocated meta page.
894 * The members define size and alignment, and silence type
895 * aliasing warnings. They are not used directly; that could
896 * mean incorrectly using several union members in parallel.
898 typedef union MDB_metabuf {
901 char mm_pad[PAGEHDRSZ];
906 /** Auxiliary DB info.
907 * The information here is mostly static/read-only. There is
908 * only a single copy of this record in the environment.
910 typedef struct MDB_dbx {
911 MDB_val md_name; /**< name of the database */
912 MDB_cmp_func *md_cmp; /**< function for comparing keys */
913 MDB_cmp_func *md_dcmp; /**< function for comparing data items */
914 MDB_rel_func *md_rel; /**< user relocate function */
915 void *md_relctx; /**< user-provided context for md_rel */
918 /** A database transaction.
919 * Every operation requires a transaction handle.
922 MDB_txn *mt_parent; /**< parent of a nested txn */
923 MDB_txn *mt_child; /**< nested txn under this txn */
924 pgno_t mt_next_pgno; /**< next unallocated page */
925 /** The ID of this transaction. IDs are integers incrementing from 1.
926 * Only committed write transactions increment the ID. If a transaction
927 * aborts, the ID may be re-used by the next writer.
930 MDB_env *mt_env; /**< the DB environment */
931 /** The list of pages that became unused during this transaction.
934 /** The list of loose pages that became unused and may be reused
935 * in this transaction, linked through #NEXT_LOOSE_PAGE(page).
937 MDB_page *mt_loose_pgs;
938 /** The sorted list of dirty pages we temporarily wrote to disk
939 * because the dirty list was full. page numbers in here are
940 * shifted left by 1, deleted slots have the LSB set.
942 MDB_IDL mt_spill_pgs;
944 /** For write txns: Modified pages. Sorted when not MDB_WRITEMAP. */
946 /** For read txns: This thread/txn's reader table slot, or NULL. */
949 /** Array of records for each DB known in the environment. */
951 /** Array of MDB_db records for each known DB */
953 /** @defgroup mt_dbflag Transaction DB Flags
957 #define DB_DIRTY 0x01 /**< DB was modified or is DUPSORT data */
958 #define DB_STALE 0x02 /**< Named-DB record is older than txnID */
959 #define DB_NEW 0x04 /**< Named-DB handle opened in this txn */
960 #define DB_VALID 0x08 /**< DB handle is valid, see also #MDB_VALID */
962 /** In write txns, array of cursors for each DB */
963 MDB_cursor **mt_cursors;
964 /** Array of flags for each DB */
965 unsigned char *mt_dbflags;
966 /** Number of DB records in use. This number only ever increments;
967 * we don't decrement it when individual DB handles are closed.
971 /** @defgroup mdb_txn Transaction Flags
975 #define MDB_TXN_RDONLY 0x01 /**< read-only transaction */
976 #define MDB_TXN_ERROR 0x02 /**< txn is unusable after an error */
977 #define MDB_TXN_DIRTY 0x04 /**< must write, even if dirty list is empty */
978 #define MDB_TXN_SPILLS 0x08 /**< txn or a parent has spilled pages */
980 unsigned int mt_flags; /**< @ref mdb_txn */
981 /** #dirty_list room: Array size - \#dirty pages visible to this txn.
982 * Includes ancestor txns' dirty pages not hidden by other txns'
983 * dirty/spilled pages. Thus commit(nested txn) has room to merge
984 * dirty_list into mt_parent after freeing hidden mt_parent pages.
986 unsigned int mt_dirty_room;
989 /** Enough space for 2^32 nodes with minimum of 2 keys per node. I.e., plenty.
990 * At 4 keys per node, enough for 2^64 nodes, so there's probably no need to
991 * raise this on a 64 bit machine.
993 #define CURSOR_STACK 32
997 /** Cursors are used for all DB operations.
998 * A cursor holds a path of (page pointer, key index) from the DB
999 * root to a position in the DB, plus other state. #MDB_DUPSORT
1000 * cursors include an xcursor to the current data item. Write txns
1001 * track their cursors and keep them up to date when data moves.
1002 * Exception: An xcursor's pointer to a #P_SUBP page can be stale.
1003 * (A node with #F_DUPDATA but no #F_SUBDATA contains a subpage).
1006 /** Next cursor on this DB in this txn */
1007 MDB_cursor *mc_next;
1008 /** Backup of the original cursor if this cursor is a shadow */
1009 MDB_cursor *mc_backup;
1010 /** Context used for databases with #MDB_DUPSORT, otherwise NULL */
1011 struct MDB_xcursor *mc_xcursor;
1012 /** The transaction that owns this cursor */
1014 /** The database handle this cursor operates on */
1016 /** The database record for this cursor */
1018 /** The database auxiliary record for this cursor */
1020 /** The @ref mt_dbflag for this database */
1021 unsigned char *mc_dbflag;
1022 unsigned short mc_snum; /**< number of pushed pages */
1023 unsigned short mc_top; /**< index of top page, normally mc_snum-1 */
1024 /** @defgroup mdb_cursor Cursor Flags
1026 * Cursor state flags.
1029 #define C_INITIALIZED 0x01 /**< cursor has been initialized and is valid */
1030 #define C_EOF 0x02 /**< No more data */
1031 #define C_SUB 0x04 /**< Cursor is a sub-cursor */
1032 #define C_DEL 0x08 /**< last op was a cursor_del */
1033 #define C_SPLITTING 0x20 /**< Cursor is in page_split */
1034 #define C_UNTRACK 0x40 /**< Un-track cursor when closing */
1036 unsigned int mc_flags; /**< @ref mdb_cursor */
1037 MDB_page *mc_pg[CURSOR_STACK]; /**< stack of pushed pages */
1038 indx_t mc_ki[CURSOR_STACK]; /**< stack of page indices */
1041 /** Context for sorted-dup records.
1042 * We could have gone to a fully recursive design, with arbitrarily
1043 * deep nesting of sub-databases. But for now we only handle these
1044 * levels - main DB, optional sub-DB, sorted-duplicate DB.
1046 typedef struct MDB_xcursor {
1047 /** A sub-cursor for traversing the Dup DB */
1048 MDB_cursor mx_cursor;
1049 /** The database record for this Dup DB */
1051 /** The auxiliary DB record for this Dup DB */
1053 /** The @ref mt_dbflag for this Dup DB */
1054 unsigned char mx_dbflag;
1057 /** State of FreeDB old pages, stored in the MDB_env */
1058 typedef struct MDB_pgstate {
1059 pgno_t *mf_pghead; /**< Reclaimed freeDB pages, or NULL before use */
1060 txnid_t mf_pglast; /**< ID of last used record, or 0 if !mf_pghead */
1063 /** The database environment. */
1065 HANDLE me_fd; /**< The main data file */
1066 HANDLE me_lfd; /**< The lock file */
1067 HANDLE me_mfd; /**< just for writing the meta pages */
1068 /** Failed to update the meta page. Probably an I/O error. */
1069 #define MDB_FATAL_ERROR 0x80000000U
1070 /** Some fields are initialized. */
1071 #define MDB_ENV_ACTIVE 0x20000000U
1072 /** me_txkey is set */
1073 #define MDB_ENV_TXKEY 0x10000000U
1074 uint32_t me_flags; /**< @ref mdb_env */
1075 unsigned int me_psize; /**< DB page size, inited from me_os_psize */
1076 unsigned int me_os_psize; /**< OS page size, from #GET_PAGESIZE */
1077 unsigned int me_maxreaders; /**< size of the reader table */
1078 unsigned int me_numreaders; /**< max numreaders set by this env */
1079 MDB_dbi me_numdbs; /**< number of DBs opened */
1080 MDB_dbi me_maxdbs; /**< size of the DB table */
1081 MDB_PID_T me_pid; /**< process ID of this env */
1082 char *me_path; /**< path to the DB files */
1083 char *me_map; /**< the memory map of the data file */
1084 MDB_txninfo *me_txns; /**< the memory map of the lock file or NULL */
1085 MDB_meta *me_metas[2]; /**< pointers to the two meta pages */
1086 void *me_pbuf; /**< scratch area for DUPSORT put() */
1087 MDB_txn *me_txn; /**< current write transaction */
1088 size_t me_mapsize; /**< size of the data memory map */
1089 off_t me_size; /**< current file size */
1090 pgno_t me_maxpg; /**< me_mapsize / me_psize */
1091 MDB_dbx *me_dbxs; /**< array of static DB info */
1092 uint16_t *me_dbflags; /**< array of flags from MDB_db.md_flags */
1093 pthread_key_t me_txkey; /**< thread-key for readers */
1094 MDB_pgstate me_pgstate; /**< state of old pages from freeDB */
1095 # define me_pglast me_pgstate.mf_pglast
1096 # define me_pghead me_pgstate.mf_pghead
1097 MDB_page *me_dpages; /**< list of malloc'd blocks for re-use */
1098 /** IDL of pages that became unused in a write txn */
1099 MDB_IDL me_free_pgs;
1100 /** ID2L of pages written during a write txn. Length MDB_IDL_UM_SIZE. */
1101 MDB_ID2L me_dirty_list;
1102 /** Max number of freelist items that can fit in a single overflow page */
1104 /** Max size of a node on a page */
1105 unsigned int me_nodemax;
1106 #if !(MDB_MAXKEYSIZE)
1107 unsigned int me_maxkey; /**< max size of a key */
1109 int me_live_reader; /**< have liveness lock in reader table */
1111 int me_pidquery; /**< Used in OpenProcess */
1112 HANDLE me_rmutex; /* Windows mutexes don't reside in shared mem */
1114 #elif defined(MDB_USE_POSIX_SEM)
1115 sem_t *me_rmutex; /* Shared mutexes are not supported */
1118 void *me_userctx; /**< User-settable context */
1119 MDB_assert_func *me_assert_func; /**< Callback for assertion failures */
1122 /** Nested transaction */
1123 typedef struct MDB_ntxn {
1124 MDB_txn mnt_txn; /**< the transaction */
1125 MDB_pgstate mnt_pgstate; /**< parent transaction's saved freestate */
1128 /** max number of pages to commit in one writev() call */
1129 #define MDB_COMMIT_PAGES 64
1130 #if defined(IOV_MAX) && IOV_MAX < MDB_COMMIT_PAGES
1131 #undef MDB_COMMIT_PAGES
1132 #define MDB_COMMIT_PAGES IOV_MAX
1135 /** max bytes to write in one call */
1136 #define MAX_WRITE (0x80000000U >> (sizeof(ssize_t) == 4))
1138 /** Check \b txn and \b dbi arguments to a function */
1139 #define TXN_DBI_EXIST(txn, dbi) \
1140 ((txn) && (dbi) < (txn)->mt_numdbs && ((txn)->mt_dbflags[dbi] & DB_VALID))
1142 static int mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp);
1143 static int mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp);
1144 static int mdb_page_touch(MDB_cursor *mc);
1146 static int mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **mp, int *lvl);
1147 static int mdb_page_search_root(MDB_cursor *mc,
1148 MDB_val *key, int modify);
1149 #define MDB_PS_MODIFY 1
1150 #define MDB_PS_ROOTONLY 2
1151 #define MDB_PS_FIRST 4
1152 #define MDB_PS_LAST 8
1153 static int mdb_page_search(MDB_cursor *mc,
1154 MDB_val *key, int flags);
1155 static int mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst);
1157 #define MDB_SPLIT_REPLACE MDB_APPENDDUP /**< newkey is not new */
1158 static int mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata,
1159 pgno_t newpgno, unsigned int nflags);
1161 static int mdb_env_read_header(MDB_env *env, MDB_meta *meta);
1162 static int mdb_env_pick_meta(const MDB_env *env);
1163 static int mdb_env_write_meta(MDB_txn *txn);
1164 #if !(defined(_WIN32) || defined(MDB_USE_POSIX_SEM)) /* Drop unused excl arg */
1165 # define mdb_env_close0(env, excl) mdb_env_close1(env)
1167 static void mdb_env_close0(MDB_env *env, int excl);
1169 static MDB_node *mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp);
1170 static int mdb_node_add(MDB_cursor *mc, indx_t indx,
1171 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags);
1172 static void mdb_node_del(MDB_cursor *mc, int ksize);
1173 static void mdb_node_shrink(MDB_page *mp, indx_t indx);
1174 static int mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst);
1175 static int mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data);
1176 static size_t mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data);
1177 static size_t mdb_branch_size(MDB_env *env, MDB_val *key);
1179 static int mdb_rebalance(MDB_cursor *mc);
1180 static int mdb_update_key(MDB_cursor *mc, MDB_val *key);
1182 static void mdb_cursor_pop(MDB_cursor *mc);
1183 static int mdb_cursor_push(MDB_cursor *mc, MDB_page *mp);
1185 static int mdb_cursor_del0(MDB_cursor *mc);
1186 static int mdb_del0(MDB_txn *txn, MDB_dbi dbi, MDB_val *key, MDB_val *data, unsigned flags);
1187 static int mdb_cursor_sibling(MDB_cursor *mc, int move_right);
1188 static int mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1189 static int mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op);
1190 static int mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op,
1192 static int mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1193 static int mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data);
1195 static void mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx);
1196 static void mdb_xcursor_init0(MDB_cursor *mc);
1197 static void mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node);
1199 static int mdb_drop0(MDB_cursor *mc, int subs);
1200 static void mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi);
1203 static MDB_cmp_func mdb_cmp_memn, mdb_cmp_memnr, mdb_cmp_int, mdb_cmp_cint, mdb_cmp_long;
1207 static SECURITY_DESCRIPTOR mdb_null_sd;
1208 static SECURITY_ATTRIBUTES mdb_all_sa;
1209 static int mdb_sec_inited;
1212 /** Return the library version info. */
1214 mdb_version(int *major, int *minor, int *patch)
1216 if (major) *major = MDB_VERSION_MAJOR;
1217 if (minor) *minor = MDB_VERSION_MINOR;
1218 if (patch) *patch = MDB_VERSION_PATCH;
1219 return MDB_VERSION_STRING;
1222 /** Table of descriptions for LMDB @ref errors */
1223 static char *const mdb_errstr[] = {
1224 "MDB_KEYEXIST: Key/data pair already exists",
1225 "MDB_NOTFOUND: No matching key/data pair found",
1226 "MDB_PAGE_NOTFOUND: Requested page not found",
1227 "MDB_CORRUPTED: Located page was wrong type",
1228 "MDB_PANIC: Update of meta page failed",
1229 "MDB_VERSION_MISMATCH: Database environment version mismatch",
1230 "MDB_INVALID: File is not an LMDB file",
1231 "MDB_MAP_FULL: Environment mapsize limit reached",
1232 "MDB_DBS_FULL: Environment maxdbs limit reached",
1233 "MDB_READERS_FULL: Environment maxreaders limit reached",
1234 "MDB_TLS_FULL: Thread-local storage keys full - too many environments open",
1235 "MDB_TXN_FULL: Transaction has too many dirty pages - transaction too big",
1236 "MDB_CURSOR_FULL: Internal error - cursor stack limit reached",
1237 "MDB_PAGE_FULL: Internal error - page has no more space",
1238 "MDB_MAP_RESIZED: Database contents grew beyond environment mapsize",
1239 "MDB_INCOMPATIBLE: Operation and DB incompatible, or DB flags changed",
1240 "MDB_BAD_RSLOT: Invalid reuse of reader locktable slot",
1241 "MDB_BAD_TXN: Transaction cannot recover - it must be aborted",
1242 "MDB_BAD_VALSIZE: Unsupported size of key/DB name/data, or wrong DUPFIXED size",
1246 mdb_strerror(int err)
1250 return ("Successful return: 0");
1252 if (err >= MDB_KEYEXIST && err <= MDB_LAST_ERRCODE) {
1253 i = err - MDB_KEYEXIST;
1254 return mdb_errstr[i];
1257 return strerror(err);
1260 /** assert(3) variant in cursor context */
1261 #define mdb_cassert(mc, expr) mdb_assert0((mc)->mc_txn->mt_env, expr, #expr)
1262 /** assert(3) variant in transaction context */
1263 #define mdb_tassert(mc, expr) mdb_assert0((txn)->mt_env, expr, #expr)
1264 /** assert(3) variant in environment context */
1265 #define mdb_eassert(env, expr) mdb_assert0(env, expr, #expr)
1268 # define mdb_assert0(env, expr, expr_txt) ((expr) ? (void)0 : \
1269 mdb_assert_fail(env, expr_txt, mdb_func_, __FILE__, __LINE__))
1272 mdb_assert_fail(MDB_env *env, const char *expr_txt,
1273 const char *func, const char *file, int line)
1276 sprintf(buf, "%.100s:%d: Assertion '%.200s' failed in %.40s()",
1277 file, line, expr_txt, func);
1278 if (env->me_assert_func)
1279 env->me_assert_func(env, buf);
1280 fprintf(stderr, "%s\n", buf);
1284 # define mdb_assert0(env, expr, expr_txt) ((void) 0)
1288 /** Return the page number of \b mp which may be sub-page, for debug output */
1290 mdb_dbg_pgno(MDB_page *mp)
1293 COPY_PGNO(ret, mp->mp_pgno);
1297 /** Display a key in hexadecimal and return the address of the result.
1298 * @param[in] key the key to display
1299 * @param[in] buf the buffer to write into. Should always be #DKBUF.
1300 * @return The key in hexadecimal form.
1303 mdb_dkey(MDB_val *key, char *buf)
1306 unsigned char *c = key->mv_data;
1312 if (key->mv_size > DKBUF_MAXKEYSIZE)
1313 return "MDB_MAXKEYSIZE";
1314 /* may want to make this a dynamic check: if the key is mostly
1315 * printable characters, print it as-is instead of converting to hex.
1319 for (i=0; i<key->mv_size; i++)
1320 ptr += sprintf(ptr, "%02x", *c++);
1322 sprintf(buf, "%.*s", key->mv_size, key->mv_data);
1328 mdb_leafnode_type(MDB_node *n)
1330 static char *const tp[2][2] = {{"", ": DB"}, {": sub-page", ": sub-DB"}};
1331 return F_ISSET(n->mn_flags, F_BIGDATA) ? ": overflow page" :
1332 tp[F_ISSET(n->mn_flags, F_DUPDATA)][F_ISSET(n->mn_flags, F_SUBDATA)];
1335 /** Display all the keys in the page. */
1337 mdb_page_list(MDB_page *mp)
1339 pgno_t pgno = mdb_dbg_pgno(mp);
1340 const char *type, *state = (mp->mp_flags & P_DIRTY) ? ", dirty" : "";
1342 unsigned int i, nkeys, nsize, total = 0;
1346 switch (mp->mp_flags & (P_BRANCH|P_LEAF|P_LEAF2|P_META|P_OVERFLOW|P_SUBP)) {
1347 case P_BRANCH: type = "Branch page"; break;
1348 case P_LEAF: type = "Leaf page"; break;
1349 case P_LEAF|P_SUBP: type = "Sub-page"; break;
1350 case P_LEAF|P_LEAF2: type = "LEAF2 page"; break;
1351 case P_LEAF|P_LEAF2|P_SUBP: type = "LEAF2 sub-page"; break;
1353 fprintf(stderr, "Overflow page %"Z"u pages %u%s\n",
1354 pgno, mp->mp_pages, state);
1357 fprintf(stderr, "Meta-page %"Z"u txnid %"Z"u\n",
1358 pgno, ((MDB_meta *)METADATA(mp))->mm_txnid);
1361 fprintf(stderr, "Bad page %"Z"u flags 0x%u\n", pgno, mp->mp_flags);
1365 nkeys = NUMKEYS(mp);
1366 fprintf(stderr, "%s %"Z"u numkeys %d%s\n", type, pgno, nkeys, state);
1368 for (i=0; i<nkeys; i++) {
1369 if (IS_LEAF2(mp)) { /* LEAF2 pages have no mp_ptrs[] or node headers */
1370 key.mv_size = nsize = mp->mp_pad;
1371 key.mv_data = LEAF2KEY(mp, i, nsize);
1373 fprintf(stderr, "key %d: nsize %d, %s\n", i, nsize, DKEY(&key));
1376 node = NODEPTR(mp, i);
1377 key.mv_size = node->mn_ksize;
1378 key.mv_data = node->mn_data;
1379 nsize = NODESIZE + key.mv_size;
1380 if (IS_BRANCH(mp)) {
1381 fprintf(stderr, "key %d: page %"Z"u, %s\n", i, NODEPGNO(node),
1385 if (F_ISSET(node->mn_flags, F_BIGDATA))
1386 nsize += sizeof(pgno_t);
1388 nsize += NODEDSZ(node);
1390 nsize += sizeof(indx_t);
1391 fprintf(stderr, "key %d: nsize %d, %s%s\n",
1392 i, nsize, DKEY(&key), mdb_leafnode_type(node));
1394 total = EVEN(total);
1396 fprintf(stderr, "Total: header %d + contents %d + unused %d\n",
1397 IS_LEAF2(mp) ? PAGEHDRSZ : PAGEBASE + mp->mp_lower, total, SIZELEFT(mp));
1401 mdb_cursor_chk(MDB_cursor *mc)
1407 if (!mc->mc_snum && !(mc->mc_flags & C_INITIALIZED)) return;
1408 for (i=0; i<mc->mc_top; i++) {
1410 node = NODEPTR(mp, mc->mc_ki[i]);
1411 if (NODEPGNO(node) != mc->mc_pg[i+1]->mp_pgno)
1414 if (mc->mc_ki[i] >= NUMKEYS(mc->mc_pg[i]))
1420 /** Count all the pages in each DB and in the freelist
1421 * and make sure it matches the actual number of pages
1423 * All named DBs must be open for a correct count.
1425 static void mdb_audit(MDB_txn *txn)
1429 MDB_ID freecount, count;
1434 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
1435 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
1436 freecount += *(MDB_ID *)data.mv_data;
1437 mdb_tassert(txn, rc == MDB_NOTFOUND);
1440 for (i = 0; i<txn->mt_numdbs; i++) {
1442 if (!(txn->mt_dbflags[i] & DB_VALID))
1444 mdb_cursor_init(&mc, txn, i, &mx);
1445 if (txn->mt_dbs[i].md_root == P_INVALID)
1447 count += txn->mt_dbs[i].md_branch_pages +
1448 txn->mt_dbs[i].md_leaf_pages +
1449 txn->mt_dbs[i].md_overflow_pages;
1450 if (txn->mt_dbs[i].md_flags & MDB_DUPSORT) {
1451 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST);
1452 for (; rc == MDB_SUCCESS; rc = mdb_cursor_sibling(&mc, 1)) {
1455 mp = mc.mc_pg[mc.mc_top];
1456 for (j=0; j<NUMKEYS(mp); j++) {
1457 MDB_node *leaf = NODEPTR(mp, j);
1458 if (leaf->mn_flags & F_SUBDATA) {
1460 memcpy(&db, NODEDATA(leaf), sizeof(db));
1461 count += db.md_branch_pages + db.md_leaf_pages +
1462 db.md_overflow_pages;
1466 mdb_tassert(txn, rc == MDB_NOTFOUND);
1469 if (freecount + count + 2 /* metapages */ != txn->mt_next_pgno) {
1470 fprintf(stderr, "audit: %lu freecount: %lu count: %lu total: %lu next_pgno: %lu\n",
1471 txn->mt_txnid, freecount, count+2, freecount+count+2, txn->mt_next_pgno);
1477 mdb_cmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1479 return txn->mt_dbxs[dbi].md_cmp(a, b);
1483 mdb_dcmp(MDB_txn *txn, MDB_dbi dbi, const MDB_val *a, const MDB_val *b)
1485 return txn->mt_dbxs[dbi].md_dcmp(a, b);
1488 /** Allocate memory for a page.
1489 * Re-use old malloc'd pages first for singletons, otherwise just malloc.
1492 mdb_page_malloc(MDB_txn *txn, unsigned num)
1494 MDB_env *env = txn->mt_env;
1495 MDB_page *ret = env->me_dpages;
1496 size_t psize = env->me_psize, sz = psize, off;
1497 /* For ! #MDB_NOMEMINIT, psize counts how much to init.
1498 * For a single page alloc, we init everything after the page header.
1499 * For multi-page, we init the final page; if the caller needed that
1500 * many pages they will be filling in at least up to the last page.
1504 VGMEMP_ALLOC(env, ret, sz);
1505 VGMEMP_DEFINED(ret, sizeof(ret->mp_next));
1506 env->me_dpages = ret->mp_next;
1509 psize -= off = PAGEHDRSZ;
1514 if ((ret = malloc(sz)) != NULL) {
1515 VGMEMP_ALLOC(env, ret, sz);
1516 if (!(env->me_flags & MDB_NOMEMINIT)) {
1517 memset((char *)ret + off, 0, psize);
1521 txn->mt_flags |= MDB_TXN_ERROR;
1525 /** Free a single page.
1526 * Saves single pages to a list, for future reuse.
1527 * (This is not used for multi-page overflow pages.)
1530 mdb_page_free(MDB_env *env, MDB_page *mp)
1532 mp->mp_next = env->me_dpages;
1533 VGMEMP_FREE(env, mp);
1534 env->me_dpages = mp;
1537 /** Free a dirty page */
1539 mdb_dpage_free(MDB_env *env, MDB_page *dp)
1541 if (!IS_OVERFLOW(dp) || dp->mp_pages == 1) {
1542 mdb_page_free(env, dp);
1544 /* large pages just get freed directly */
1545 VGMEMP_FREE(env, dp);
1550 /** Return all dirty pages to dpage list */
1552 mdb_dlist_free(MDB_txn *txn)
1554 MDB_env *env = txn->mt_env;
1555 MDB_ID2L dl = txn->mt_u.dirty_list;
1556 unsigned i, n = dl[0].mid;
1558 for (i = 1; i <= n; i++) {
1559 mdb_dpage_free(env, dl[i].mptr);
1564 /** Loosen or free a single page.
1565 * Saves single pages to a list for future reuse
1566 * in this same txn. It has been pulled from the freeDB
1567 * and already resides on the dirty list, but has been
1568 * deleted. Use these pages first before pulling again
1571 * If the page wasn't dirtied in this txn, just add it
1572 * to this txn's free list.
1575 mdb_page_loose(MDB_cursor *mc, MDB_page *mp)
1578 pgno_t pgno = mp->mp_pgno;
1580 if ((mp->mp_flags & P_DIRTY) && mc->mc_dbi != FREE_DBI) {
1581 if (mc->mc_txn->mt_parent) {
1582 MDB_ID2 *dl = mc->mc_txn->mt_u.dirty_list;
1583 /* If txn has a parent, make sure the page is in our
1587 unsigned x = mdb_mid2l_search(dl, pgno);
1588 if (x <= dl[0].mid && dl[x].mid == pgno) {
1589 if (mp != dl[x].mptr) { /* bad cursor? */
1590 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
1591 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
1592 return MDB_CORRUPTED;
1599 /* no parent txn, so it's just ours */
1604 DPRINTF(("loosen db %d page %"Z"u", DDBI(mc),
1606 NEXT_LOOSE_PAGE(mp) = mc->mc_txn->mt_loose_pgs;
1607 mc->mc_txn->mt_loose_pgs = mp;
1608 mp->mp_flags |= P_LOOSE;
1610 int rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, pgno);
1618 /** Set or clear P_KEEP in dirty, non-overflow, non-sub pages watched by txn.
1619 * @param[in] mc A cursor handle for the current operation.
1620 * @param[in] pflags Flags of the pages to update:
1621 * P_DIRTY to set P_KEEP, P_DIRTY|P_KEEP to clear it.
1622 * @param[in] all No shortcuts. Needed except after a full #mdb_page_flush().
1623 * @return 0 on success, non-zero on failure.
1626 mdb_pages_xkeep(MDB_cursor *mc, unsigned pflags, int all)
1628 enum { Mask = P_SUBP|P_DIRTY|P_LOOSE|P_KEEP };
1629 MDB_txn *txn = mc->mc_txn;
1635 int rc = MDB_SUCCESS, level;
1637 /* Mark pages seen by cursors */
1638 if (mc->mc_flags & C_UNTRACK)
1639 mc = NULL; /* will find mc in mt_cursors */
1640 for (i = txn->mt_numdbs;; mc = txn->mt_cursors[--i]) {
1641 for (; mc; mc=mc->mc_next) {
1642 if (!(mc->mc_flags & C_INITIALIZED))
1644 for (m3 = mc;; m3 = &mx->mx_cursor) {
1646 for (j=0; j<m3->mc_snum; j++) {
1648 if ((mp->mp_flags & Mask) == pflags)
1649 mp->mp_flags ^= P_KEEP;
1651 mx = m3->mc_xcursor;
1652 /* Proceed to mx if it is at a sub-database */
1653 if (! (mx && (mx->mx_cursor.mc_flags & C_INITIALIZED)))
1655 if (! (mp && (mp->mp_flags & P_LEAF)))
1657 leaf = NODEPTR(mp, m3->mc_ki[j-1]);
1658 if (!(leaf->mn_flags & F_SUBDATA))
1667 /* Mark dirty root pages */
1668 for (i=0; i<txn->mt_numdbs; i++) {
1669 if (txn->mt_dbflags[i] & DB_DIRTY) {
1670 pgno_t pgno = txn->mt_dbs[i].md_root;
1671 if (pgno == P_INVALID)
1673 if ((rc = mdb_page_get(txn, pgno, &dp, &level)) != MDB_SUCCESS)
1675 if ((dp->mp_flags & Mask) == pflags && level <= 1)
1676 dp->mp_flags ^= P_KEEP;
1684 static int mdb_page_flush(MDB_txn *txn, int keep);
1686 /** Spill pages from the dirty list back to disk.
1687 * This is intended to prevent running into #MDB_TXN_FULL situations,
1688 * but note that they may still occur in a few cases:
1689 * 1) our estimate of the txn size could be too small. Currently this
1690 * seems unlikely, except with a large number of #MDB_MULTIPLE items.
1691 * 2) child txns may run out of space if their parents dirtied a
1692 * lot of pages and never spilled them. TODO: we probably should do
1693 * a preemptive spill during #mdb_txn_begin() of a child txn, if
1694 * the parent's dirty_room is below a given threshold.
1696 * Otherwise, if not using nested txns, it is expected that apps will
1697 * not run into #MDB_TXN_FULL any more. The pages are flushed to disk
1698 * the same way as for a txn commit, e.g. their P_DIRTY flag is cleared.
1699 * If the txn never references them again, they can be left alone.
1700 * If the txn only reads them, they can be used without any fuss.
1701 * If the txn writes them again, they can be dirtied immediately without
1702 * going thru all of the work of #mdb_page_touch(). Such references are
1703 * handled by #mdb_page_unspill().
1705 * Also note, we never spill DB root pages, nor pages of active cursors,
1706 * because we'll need these back again soon anyway. And in nested txns,
1707 * we can't spill a page in a child txn if it was already spilled in a
1708 * parent txn. That would alter the parent txns' data even though
1709 * the child hasn't committed yet, and we'd have no way to undo it if
1710 * the child aborted.
1712 * @param[in] m0 cursor A cursor handle identifying the transaction and
1713 * database for which we are checking space.
1714 * @param[in] key For a put operation, the key being stored.
1715 * @param[in] data For a put operation, the data being stored.
1716 * @return 0 on success, non-zero on failure.
1719 mdb_page_spill(MDB_cursor *m0, MDB_val *key, MDB_val *data)
1721 MDB_txn *txn = m0->mc_txn;
1723 MDB_ID2L dl = txn->mt_u.dirty_list;
1724 unsigned int i, j, need;
1727 if (m0->mc_flags & C_SUB)
1730 /* Estimate how much space this op will take */
1731 i = m0->mc_db->md_depth;
1732 /* Named DBs also dirty the main DB */
1733 if (m0->mc_dbi > MAIN_DBI)
1734 i += txn->mt_dbs[MAIN_DBI].md_depth;
1735 /* For puts, roughly factor in the key+data size */
1737 i += (LEAFSIZE(key, data) + txn->mt_env->me_psize) / txn->mt_env->me_psize;
1738 i += i; /* double it for good measure */
1741 if (txn->mt_dirty_room > i)
1744 if (!txn->mt_spill_pgs) {
1745 txn->mt_spill_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX);
1746 if (!txn->mt_spill_pgs)
1749 /* purge deleted slots */
1750 MDB_IDL sl = txn->mt_spill_pgs;
1751 unsigned int num = sl[0];
1753 for (i=1; i<=num; i++) {
1760 /* Preserve pages which may soon be dirtied again */
1761 if ((rc = mdb_pages_xkeep(m0, P_DIRTY, 1)) != MDB_SUCCESS)
1764 /* Less aggressive spill - we originally spilled the entire dirty list,
1765 * with a few exceptions for cursor pages and DB root pages. But this
1766 * turns out to be a lot of wasted effort because in a large txn many
1767 * of those pages will need to be used again. So now we spill only 1/8th
1768 * of the dirty pages. Testing revealed this to be a good tradeoff,
1769 * better than 1/2, 1/4, or 1/10.
1771 if (need < MDB_IDL_UM_MAX / 8)
1772 need = MDB_IDL_UM_MAX / 8;
1774 /* Save the page IDs of all the pages we're flushing */
1775 /* flush from the tail forward, this saves a lot of shifting later on. */
1776 for (i=dl[0].mid; i && need; i--) {
1777 MDB_ID pn = dl[i].mid << 1;
1779 if (dp->mp_flags & (P_LOOSE|P_KEEP))
1781 /* Can't spill twice, make sure it's not already in a parent's
1784 if (txn->mt_parent) {
1786 for (tx2 = txn->mt_parent; tx2; tx2 = tx2->mt_parent) {
1787 if (tx2->mt_spill_pgs) {
1788 j = mdb_midl_search(tx2->mt_spill_pgs, pn);
1789 if (j <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[j] == pn) {
1790 dp->mp_flags |= P_KEEP;
1798 if ((rc = mdb_midl_append(&txn->mt_spill_pgs, pn)))
1802 mdb_midl_sort(txn->mt_spill_pgs);
1804 /* Flush the spilled part of dirty list */
1805 if ((rc = mdb_page_flush(txn, i)) != MDB_SUCCESS)
1808 /* Reset any dirty pages we kept that page_flush didn't see */
1809 rc = mdb_pages_xkeep(m0, P_DIRTY|P_KEEP, i);
1812 txn->mt_flags |= rc ? MDB_TXN_ERROR : MDB_TXN_SPILLS;
1816 /** Find oldest txnid still referenced. Expects txn->mt_txnid > 0. */
1818 mdb_find_oldest(MDB_txn *txn)
1821 txnid_t mr, oldest = txn->mt_txnid - 1;
1822 if (txn->mt_env->me_txns) {
1823 MDB_reader *r = txn->mt_env->me_txns->mti_readers;
1824 for (i = txn->mt_env->me_txns->mti_numreaders; --i >= 0; ) {
1835 /** Add a page to the txn's dirty list */
1837 mdb_page_dirty(MDB_txn *txn, MDB_page *mp)
1840 int rc, (*insert)(MDB_ID2L, MDB_ID2 *);
1842 if (txn->mt_env->me_flags & MDB_WRITEMAP) {
1843 insert = mdb_mid2l_append;
1845 insert = mdb_mid2l_insert;
1847 mid.mid = mp->mp_pgno;
1849 rc = insert(txn->mt_u.dirty_list, &mid);
1850 mdb_tassert(txn, rc == 0);
1851 txn->mt_dirty_room--;
1854 /** Allocate page numbers and memory for writing. Maintain me_pglast,
1855 * me_pghead and mt_next_pgno.
1857 * If there are free pages available from older transactions, they
1858 * are re-used first. Otherwise allocate a new page at mt_next_pgno.
1859 * Do not modify the freedB, just merge freeDB records into me_pghead[]
1860 * and move me_pglast to say which records were consumed. Only this
1861 * function can create me_pghead and move me_pglast/mt_next_pgno.
1862 * @param[in] mc cursor A cursor handle identifying the transaction and
1863 * database for which we are allocating.
1864 * @param[in] num the number of pages to allocate.
1865 * @param[out] mp Address of the allocated page(s). Requests for multiple pages
1866 * will always be satisfied by a single contiguous chunk of memory.
1867 * @return 0 on success, non-zero on failure.
1870 mdb_page_alloc(MDB_cursor *mc, int num, MDB_page **mp)
1872 #ifdef MDB_PARANOID /* Seems like we can ignore this now */
1873 /* Get at most <Max_retries> more freeDB records once me_pghead
1874 * has enough pages. If not enough, use new pages from the map.
1875 * If <Paranoid> and mc is updating the freeDB, only get new
1876 * records if me_pghead is empty. Then the freelist cannot play
1877 * catch-up with itself by growing while trying to save it.
1879 enum { Paranoid = 1, Max_retries = 500 };
1881 enum { Paranoid = 0, Max_retries = INT_MAX /*infinite*/ };
1883 int rc, retry = num * 20;
1884 MDB_txn *txn = mc->mc_txn;
1885 MDB_env *env = txn->mt_env;
1886 pgno_t pgno, *mop = env->me_pghead;
1887 unsigned i, j, k, mop_len = mop ? mop[0] : 0, n2 = num-1;
1889 txnid_t oldest = 0, last;
1893 /* If there are any loose pages, just use them */
1894 if (num == 1 && txn->mt_loose_pgs) {
1895 np = txn->mt_loose_pgs;
1896 txn->mt_loose_pgs = NEXT_LOOSE_PAGE(np);
1897 DPRINTF(("db %d use loose page %"Z"u", DDBI(mc),
1905 /* If our dirty list is already full, we can't do anything */
1906 if (txn->mt_dirty_room == 0) {
1911 for (op = MDB_FIRST;; op = MDB_NEXT) {
1914 pgno_t *idl, old_id, new_id;
1916 /* Seek a big enough contiguous page range. Prefer
1917 * pages at the tail, just truncating the list.
1923 if (mop[i-n2] == pgno+n2)
1930 if (op == MDB_FIRST) { /* 1st iteration */
1931 /* Prepare to fetch more and coalesce */
1932 oldest = mdb_find_oldest(txn);
1933 last = env->me_pglast;
1934 mdb_cursor_init(&m2, txn, FREE_DBI, NULL);
1937 key.mv_data = &last; /* will look up last+1 */
1938 key.mv_size = sizeof(last);
1940 if (Paranoid && mc->mc_dbi == FREE_DBI)
1943 if (Paranoid && retry < 0 && mop_len)
1947 /* Do not fetch more if the record will be too recent */
1950 rc = mdb_cursor_get(&m2, &key, NULL, op);
1952 if (rc == MDB_NOTFOUND)
1956 last = *(txnid_t*)key.mv_data;
1959 np = m2.mc_pg[m2.mc_top];
1960 leaf = NODEPTR(np, m2.mc_ki[m2.mc_top]);
1961 if ((rc = mdb_node_read(txn, leaf, &data)) != MDB_SUCCESS)
1964 idl = (MDB_ID *) data.mv_data;
1967 if (!(env->me_pghead = mop = mdb_midl_alloc(i))) {
1972 if ((rc = mdb_midl_need(&env->me_pghead, i)) != 0)
1974 mop = env->me_pghead;
1976 env->me_pglast = last;
1978 DPRINTF(("IDL read txn %"Z"u root %"Z"u num %u",
1979 last, txn->mt_dbs[FREE_DBI].md_root, i));
1981 DPRINTF(("IDL %"Z"u", idl[k]));
1983 /* Merge in descending sorted order */
1986 mop[0] = (pgno_t)-1;
1990 for (; old_id < new_id; old_id = mop[--j])
1997 /* Use new pages from the map when nothing suitable in the freeDB */
1999 pgno = txn->mt_next_pgno;
2000 if (pgno + num >= env->me_maxpg) {
2001 DPUTS("DB size maxed out");
2007 if (env->me_flags & MDB_WRITEMAP) {
2008 np = (MDB_page *)(env->me_map + env->me_psize * pgno);
2010 if (!(np = mdb_page_malloc(txn, num))) {
2016 mop[0] = mop_len -= num;
2017 /* Move any stragglers down */
2018 for (j = i-num; j < mop_len; )
2019 mop[++j] = mop[++i];
2021 txn->mt_next_pgno = pgno + num;
2024 mdb_page_dirty(txn, np);
2030 txn->mt_flags |= MDB_TXN_ERROR;
2034 /** Copy the used portions of a non-overflow page.
2035 * @param[in] dst page to copy into
2036 * @param[in] src page to copy from
2037 * @param[in] psize size of a page
2040 mdb_page_copy(MDB_page *dst, MDB_page *src, unsigned int psize)
2042 enum { Align = sizeof(pgno_t) };
2043 indx_t upper = src->mp_upper, lower = src->mp_lower, unused = upper-lower;
2045 /* If page isn't full, just copy the used portion. Adjust
2046 * alignment so memcpy may copy words instead of bytes.
2048 if ((unused &= -Align) && !IS_LEAF2(src)) {
2049 upper = (upper + PAGEBASE) & -Align;
2050 memcpy(dst, src, (lower + PAGEBASE + (Align-1)) & -Align);
2051 memcpy((pgno_t *)((char *)dst+upper), (pgno_t *)((char *)src+upper),
2054 memcpy(dst, src, psize - unused);
2058 /** Pull a page off the txn's spill list, if present.
2059 * If a page being referenced was spilled to disk in this txn, bring
2060 * it back and make it dirty/writable again.
2061 * @param[in] txn the transaction handle.
2062 * @param[in] mp the page being referenced. It must not be dirty.
2063 * @param[out] ret the writable page, if any. ret is unchanged if
2064 * mp wasn't spilled.
2067 mdb_page_unspill(MDB_txn *txn, MDB_page *mp, MDB_page **ret)
2069 MDB_env *env = txn->mt_env;
2072 pgno_t pgno = mp->mp_pgno, pn = pgno << 1;
2074 for (tx2 = txn; tx2; tx2=tx2->mt_parent) {
2075 if (!tx2->mt_spill_pgs)
2077 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
2078 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
2081 if (txn->mt_dirty_room == 0)
2082 return MDB_TXN_FULL;
2083 if (IS_OVERFLOW(mp))
2087 if (env->me_flags & MDB_WRITEMAP) {
2090 np = mdb_page_malloc(txn, num);
2094 memcpy(np, mp, num * env->me_psize);
2096 mdb_page_copy(np, mp, env->me_psize);
2099 /* If in current txn, this page is no longer spilled.
2100 * If it happens to be the last page, truncate the spill list.
2101 * Otherwise mark it as deleted by setting the LSB.
2103 if (x == txn->mt_spill_pgs[0])
2104 txn->mt_spill_pgs[0]--;
2106 txn->mt_spill_pgs[x] |= 1;
2107 } /* otherwise, if belonging to a parent txn, the
2108 * page remains spilled until child commits
2111 mdb_page_dirty(txn, np);
2112 np->mp_flags |= P_DIRTY;
2120 /** Touch a page: make it dirty and re-insert into tree with updated pgno.
2121 * @param[in] mc cursor pointing to the page to be touched
2122 * @return 0 on success, non-zero on failure.
2125 mdb_page_touch(MDB_cursor *mc)
2127 MDB_page *mp = mc->mc_pg[mc->mc_top], *np;
2128 MDB_txn *txn = mc->mc_txn;
2129 MDB_cursor *m2, *m3;
2133 if (!F_ISSET(mp->mp_flags, P_DIRTY)) {
2134 if (txn->mt_flags & MDB_TXN_SPILLS) {
2136 rc = mdb_page_unspill(txn, mp, &np);
2142 if ((rc = mdb_midl_need(&txn->mt_free_pgs, 1)) ||
2143 (rc = mdb_page_alloc(mc, 1, &np)))
2146 DPRINTF(("touched db %d page %"Z"u -> %"Z"u", DDBI(mc),
2147 mp->mp_pgno, pgno));
2148 mdb_cassert(mc, mp->mp_pgno != pgno);
2149 mdb_midl_xappend(txn->mt_free_pgs, mp->mp_pgno);
2150 /* Update the parent page, if any, to point to the new page */
2152 MDB_page *parent = mc->mc_pg[mc->mc_top-1];
2153 MDB_node *node = NODEPTR(parent, mc->mc_ki[mc->mc_top-1]);
2154 SETPGNO(node, pgno);
2156 mc->mc_db->md_root = pgno;
2158 } else if (txn->mt_parent && !IS_SUBP(mp)) {
2159 MDB_ID2 mid, *dl = txn->mt_u.dirty_list;
2161 /* If txn has a parent, make sure the page is in our
2165 unsigned x = mdb_mid2l_search(dl, pgno);
2166 if (x <= dl[0].mid && dl[x].mid == pgno) {
2167 if (mp != dl[x].mptr) { /* bad cursor? */
2168 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
2169 txn->mt_flags |= MDB_TXN_ERROR;
2170 return MDB_CORRUPTED;
2175 mdb_cassert(mc, dl[0].mid < MDB_IDL_UM_MAX);
2177 np = mdb_page_malloc(txn, 1);
2182 rc = mdb_mid2l_insert(dl, &mid);
2183 mdb_cassert(mc, rc == 0);
2188 mdb_page_copy(np, mp, txn->mt_env->me_psize);
2190 np->mp_flags |= P_DIRTY;
2193 /* Adjust cursors pointing to mp */
2194 mc->mc_pg[mc->mc_top] = np;
2195 m2 = txn->mt_cursors[mc->mc_dbi];
2196 if (mc->mc_flags & C_SUB) {
2197 for (; m2; m2=m2->mc_next) {
2198 m3 = &m2->mc_xcursor->mx_cursor;
2199 if (m3->mc_snum < mc->mc_snum) continue;
2200 if (m3->mc_pg[mc->mc_top] == mp)
2201 m3->mc_pg[mc->mc_top] = np;
2204 for (; m2; m2=m2->mc_next) {
2205 if (m2->mc_snum < mc->mc_snum) continue;
2206 if (m2->mc_pg[mc->mc_top] == mp) {
2207 m2->mc_pg[mc->mc_top] = np;
2208 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
2210 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
2212 MDB_node *leaf = NODEPTR(np, mc->mc_ki[mc->mc_top]);
2213 if (!(leaf->mn_flags & F_SUBDATA))
2214 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
2222 txn->mt_flags |= MDB_TXN_ERROR;
2227 mdb_env_sync(MDB_env *env, int force)
2230 if (force || !F_ISSET(env->me_flags, MDB_NOSYNC)) {
2231 if (env->me_flags & MDB_WRITEMAP) {
2232 int flags = ((env->me_flags & MDB_MAPASYNC) && !force)
2233 ? MS_ASYNC : MS_SYNC;
2234 if (MDB_MSYNC(env->me_map, env->me_mapsize, flags))
2237 else if (flags == MS_SYNC && MDB_FDATASYNC(env->me_fd))
2241 if (MDB_FDATASYNC(env->me_fd))
2248 /** Back up parent txn's cursors, then grab the originals for tracking */
2250 mdb_cursor_shadow(MDB_txn *src, MDB_txn *dst)
2252 MDB_cursor *mc, *bk;
2257 for (i = src->mt_numdbs; --i >= 0; ) {
2258 if ((mc = src->mt_cursors[i]) != NULL) {
2259 size = sizeof(MDB_cursor);
2261 size += sizeof(MDB_xcursor);
2262 for (; mc; mc = bk->mc_next) {
2268 mc->mc_db = &dst->mt_dbs[i];
2269 /* Kill pointers into src - and dst to reduce abuse: The
2270 * user may not use mc until dst ends. Otherwise we'd...
2272 mc->mc_txn = NULL; /* ...set this to dst */
2273 mc->mc_dbflag = NULL; /* ...and &dst->mt_dbflags[i] */
2274 if ((mx = mc->mc_xcursor) != NULL) {
2275 *(MDB_xcursor *)(bk+1) = *mx;
2276 mx->mx_cursor.mc_txn = NULL; /* ...and dst. */
2278 mc->mc_next = dst->mt_cursors[i];
2279 dst->mt_cursors[i] = mc;
2286 /** Close this write txn's cursors, give parent txn's cursors back to parent.
2287 * @param[in] txn the transaction handle.
2288 * @param[in] merge true to keep changes to parent cursors, false to revert.
2289 * @return 0 on success, non-zero on failure.
2292 mdb_cursors_close(MDB_txn *txn, unsigned merge)
2294 MDB_cursor **cursors = txn->mt_cursors, *mc, *next, *bk;
2298 for (i = txn->mt_numdbs; --i >= 0; ) {
2299 for (mc = cursors[i]; mc; mc = next) {
2301 if ((bk = mc->mc_backup) != NULL) {
2303 /* Commit changes to parent txn */
2304 mc->mc_next = bk->mc_next;
2305 mc->mc_backup = bk->mc_backup;
2306 mc->mc_txn = bk->mc_txn;
2307 mc->mc_db = bk->mc_db;
2308 mc->mc_dbflag = bk->mc_dbflag;
2309 if ((mx = mc->mc_xcursor) != NULL)
2310 mx->mx_cursor.mc_txn = bk->mc_txn;
2312 /* Abort nested txn */
2314 if ((mx = mc->mc_xcursor) != NULL)
2315 *mx = *(MDB_xcursor *)(bk+1);
2319 /* Only malloced cursors are permanently tracked. */
2327 #define mdb_txn_reset0(txn, act) mdb_txn_reset0(txn)
2330 mdb_txn_reset0(MDB_txn *txn, const char *act);
2332 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2338 Pidset = F_SETLK, Pidcheck = F_GETLK
2342 /** Set or check a pid lock. Set returns 0 on success.
2343 * Check returns 0 if the process is certainly dead, nonzero if it may
2344 * be alive (the lock exists or an error happened so we do not know).
2346 * On Windows Pidset is a no-op, we merely check for the existence
2347 * of the process with the given pid. On POSIX we use a single byte
2348 * lock on the lockfile, set at an offset equal to the pid.
2351 mdb_reader_pid(MDB_env *env, enum Pidlock_op op, MDB_PID_T pid)
2353 #if !(MDB_PIDLOCK) /* Currently the same as defined(_WIN32) */
2356 if (op == Pidcheck) {
2357 h = OpenProcess(env->me_pidquery, FALSE, pid);
2358 /* No documented "no such process" code, but other program use this: */
2360 return ErrCode() != ERROR_INVALID_PARAMETER;
2361 /* A process exists until all handles to it close. Has it exited? */
2362 ret = WaitForSingleObject(h, 0) != 0;
2369 struct flock lock_info;
2370 memset(&lock_info, 0, sizeof(lock_info));
2371 lock_info.l_type = F_WRLCK;
2372 lock_info.l_whence = SEEK_SET;
2373 lock_info.l_start = pid;
2374 lock_info.l_len = 1;
2375 if ((rc = fcntl(env->me_lfd, op, &lock_info)) == 0) {
2376 if (op == F_GETLK && lock_info.l_type != F_UNLCK)
2378 } else if ((rc = ErrCode()) == EINTR) {
2386 /** Common code for #mdb_txn_begin() and #mdb_txn_renew().
2387 * @param[in] txn the transaction handle to initialize
2388 * @return 0 on success, non-zero on failure.
2391 mdb_txn_renew0(MDB_txn *txn)
2393 MDB_env *env = txn->mt_env;
2394 MDB_txninfo *ti = env->me_txns;
2398 int rc, new_notls = 0;
2401 txn->mt_numdbs = env->me_numdbs;
2402 txn->mt_dbxs = env->me_dbxs; /* mostly static anyway */
2404 if (txn->mt_flags & MDB_TXN_RDONLY) {
2406 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2407 txn->mt_txnid = meta->mm_txnid;
2408 txn->mt_u.reader = NULL;
2410 MDB_reader *r = (env->me_flags & MDB_NOTLS) ? txn->mt_u.reader :
2411 pthread_getspecific(env->me_txkey);
2413 if (r->mr_pid != env->me_pid || r->mr_txnid != (txnid_t)-1)
2414 return MDB_BAD_RSLOT;
2416 MDB_PID_T pid = env->me_pid;
2417 MDB_THR_T tid = pthread_self();
2419 if (!env->me_live_reader) {
2420 rc = mdb_reader_pid(env, Pidset, pid);
2423 env->me_live_reader = 1;
2427 nr = ti->mti_numreaders;
2428 for (i=0; i<nr; i++)
2429 if (ti->mti_readers[i].mr_pid == 0)
2431 if (i == env->me_maxreaders) {
2432 UNLOCK_MUTEX_R(env);
2433 return MDB_READERS_FULL;
2435 ti->mti_readers[i].mr_pid = pid;
2436 ti->mti_readers[i].mr_tid = tid;
2438 ti->mti_numreaders = ++nr;
2439 /* Save numreaders for un-mutexed mdb_env_close() */
2440 env->me_numreaders = nr;
2441 UNLOCK_MUTEX_R(env);
2443 r = &ti->mti_readers[i];
2444 new_notls = (env->me_flags & MDB_NOTLS);
2445 if (!new_notls && (rc=pthread_setspecific(env->me_txkey, r))) {
2450 txn->mt_txnid = r->mr_txnid = ti->mti_txnid;
2451 txn->mt_u.reader = r;
2452 meta = env->me_metas[txn->mt_txnid & 1];
2458 txn->mt_txnid = ti->mti_txnid;
2459 meta = env->me_metas[txn->mt_txnid & 1];
2461 meta = env->me_metas[ mdb_env_pick_meta(env) ];
2462 txn->mt_txnid = meta->mm_txnid;
2466 if (txn->mt_txnid == mdb_debug_start)
2469 txn->mt_dirty_room = MDB_IDL_UM_MAX;
2470 txn->mt_u.dirty_list = env->me_dirty_list;
2471 txn->mt_u.dirty_list[0].mid = 0;
2472 txn->mt_free_pgs = env->me_free_pgs;
2473 txn->mt_free_pgs[0] = 0;
2474 txn->mt_spill_pgs = NULL;
2478 /* Copy the DB info and flags */
2479 memcpy(txn->mt_dbs, meta->mm_dbs, 2 * sizeof(MDB_db));
2481 /* Moved to here to avoid a data race in read TXNs */
2482 txn->mt_next_pgno = meta->mm_last_pg+1;
2484 for (i=2; i<txn->mt_numdbs; i++) {
2485 x = env->me_dbflags[i];
2486 txn->mt_dbs[i].md_flags = x & PERSISTENT_FLAGS;
2487 txn->mt_dbflags[i] = (x & MDB_VALID) ? DB_VALID|DB_STALE : 0;
2489 txn->mt_dbflags[0] = txn->mt_dbflags[1] = DB_VALID;
2491 if (env->me_maxpg < txn->mt_next_pgno) {
2492 mdb_txn_reset0(txn, "renew0-mapfail");
2494 txn->mt_u.reader->mr_pid = 0;
2495 txn->mt_u.reader = NULL;
2497 return MDB_MAP_RESIZED;
2504 mdb_txn_renew(MDB_txn *txn)
2508 if (!txn || txn->mt_dbxs) /* A reset txn has mt_dbxs==NULL */
2511 if (txn->mt_env->me_flags & MDB_FATAL_ERROR) {
2512 DPUTS("environment had fatal error, must shutdown!");
2516 rc = mdb_txn_renew0(txn);
2517 if (rc == MDB_SUCCESS) {
2518 DPRINTF(("renew txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2519 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2520 (void *)txn, (void *)txn->mt_env, txn->mt_dbs[MAIN_DBI].md_root));
2526 mdb_txn_begin(MDB_env *env, MDB_txn *parent, unsigned int flags, MDB_txn **ret)
2530 int rc, size, tsize = sizeof(MDB_txn);
2532 if (env->me_flags & MDB_FATAL_ERROR) {
2533 DPUTS("environment had fatal error, must shutdown!");
2536 if ((env->me_flags & MDB_RDONLY) && !(flags & MDB_RDONLY))
2539 /* Nested transactions: Max 1 child, write txns only, no writemap */
2540 if (parent->mt_child ||
2541 (flags & MDB_RDONLY) ||
2542 (parent->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR)) ||
2543 (env->me_flags & MDB_WRITEMAP))
2545 return (parent->mt_flags & MDB_TXN_RDONLY) ? EINVAL : MDB_BAD_TXN;
2547 tsize = sizeof(MDB_ntxn);
2549 size = tsize + env->me_maxdbs * (sizeof(MDB_db)+1);
2550 if (!(flags & MDB_RDONLY))
2551 size += env->me_maxdbs * sizeof(MDB_cursor *);
2553 if ((txn = calloc(1, size)) == NULL) {
2554 DPRINTF(("calloc: %s", strerror(ErrCode())));
2557 txn->mt_dbs = (MDB_db *) ((char *)txn + tsize);
2558 if (flags & MDB_RDONLY) {
2559 txn->mt_flags |= MDB_TXN_RDONLY;
2560 txn->mt_dbflags = (unsigned char *)(txn->mt_dbs + env->me_maxdbs);
2562 txn->mt_cursors = (MDB_cursor **)(txn->mt_dbs + env->me_maxdbs);
2563 txn->mt_dbflags = (unsigned char *)(txn->mt_cursors + env->me_maxdbs);
2569 txn->mt_u.dirty_list = malloc(sizeof(MDB_ID2)*MDB_IDL_UM_SIZE);
2570 if (!txn->mt_u.dirty_list ||
2571 !(txn->mt_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)))
2573 free(txn->mt_u.dirty_list);
2577 txn->mt_txnid = parent->mt_txnid;
2578 txn->mt_dirty_room = parent->mt_dirty_room;
2579 txn->mt_u.dirty_list[0].mid = 0;
2580 txn->mt_spill_pgs = NULL;
2581 txn->mt_next_pgno = parent->mt_next_pgno;
2582 parent->mt_child = txn;
2583 txn->mt_parent = parent;
2584 txn->mt_numdbs = parent->mt_numdbs;
2585 txn->mt_flags = parent->mt_flags;
2586 txn->mt_dbxs = parent->mt_dbxs;
2587 memcpy(txn->mt_dbs, parent->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
2588 /* Copy parent's mt_dbflags, but clear DB_NEW */
2589 for (i=0; i<txn->mt_numdbs; i++)
2590 txn->mt_dbflags[i] = parent->mt_dbflags[i] & ~DB_NEW;
2592 ntxn = (MDB_ntxn *)txn;
2593 ntxn->mnt_pgstate = env->me_pgstate; /* save parent me_pghead & co */
2594 if (env->me_pghead) {
2595 size = MDB_IDL_SIZEOF(env->me_pghead);
2596 env->me_pghead = mdb_midl_alloc(env->me_pghead[0]);
2598 memcpy(env->me_pghead, ntxn->mnt_pgstate.mf_pghead, size);
2603 rc = mdb_cursor_shadow(parent, txn);
2605 mdb_txn_reset0(txn, "beginchild-fail");
2607 rc = mdb_txn_renew0(txn);
2613 DPRINTF(("begin txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2614 txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2615 (void *) txn, (void *) env, txn->mt_dbs[MAIN_DBI].md_root));
2622 mdb_txn_env(MDB_txn *txn)
2624 if(!txn) return NULL;
2628 /** Export or close DBI handles opened in this txn. */
2630 mdb_dbis_update(MDB_txn *txn, int keep)
2633 MDB_dbi n = txn->mt_numdbs;
2634 MDB_env *env = txn->mt_env;
2635 unsigned char *tdbflags = txn->mt_dbflags;
2637 for (i = n; --i >= 2;) {
2638 if (tdbflags[i] & DB_NEW) {
2640 env->me_dbflags[i] = txn->mt_dbs[i].md_flags | MDB_VALID;
2642 char *ptr = env->me_dbxs[i].md_name.mv_data;
2643 env->me_dbxs[i].md_name.mv_data = NULL;
2644 env->me_dbxs[i].md_name.mv_size = 0;
2645 env->me_dbflags[i] = 0;
2650 if (keep && env->me_numdbs < n)
2654 /** Common code for #mdb_txn_reset() and #mdb_txn_abort().
2655 * May be called twice for readonly txns: First reset it, then abort.
2656 * @param[in] txn the transaction handle to reset
2657 * @param[in] act why the transaction is being reset
2660 mdb_txn_reset0(MDB_txn *txn, const char *act)
2662 MDB_env *env = txn->mt_env;
2664 /* Close any DBI handles opened in this txn */
2665 mdb_dbis_update(txn, 0);
2667 DPRINTF(("%s txn %"Z"u%c %p on mdbenv %p, root page %"Z"u",
2668 act, txn->mt_txnid, (txn->mt_flags & MDB_TXN_RDONLY) ? 'r' : 'w',
2669 (void *) txn, (void *)env, txn->mt_dbs[MAIN_DBI].md_root));
2671 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
2672 if (txn->mt_u.reader) {
2673 txn->mt_u.reader->mr_txnid = (txnid_t)-1;
2674 if (!(env->me_flags & MDB_NOTLS))
2675 txn->mt_u.reader = NULL; /* txn does not own reader */
2677 txn->mt_numdbs = 0; /* close nothing if called again */
2678 txn->mt_dbxs = NULL; /* mark txn as reset */
2680 mdb_cursors_close(txn, 0);
2682 if (!(env->me_flags & MDB_WRITEMAP)) {
2683 mdb_dlist_free(txn);
2685 mdb_midl_free(env->me_pghead);
2687 if (txn->mt_parent) {
2688 txn->mt_parent->mt_child = NULL;
2689 env->me_pgstate = ((MDB_ntxn *)txn)->mnt_pgstate;
2690 mdb_midl_free(txn->mt_free_pgs);
2691 mdb_midl_free(txn->mt_spill_pgs);
2692 free(txn->mt_u.dirty_list);
2696 if (mdb_midl_shrink(&txn->mt_free_pgs))
2697 env->me_free_pgs = txn->mt_free_pgs;
2698 env->me_pghead = NULL;
2702 /* The writer mutex was locked in mdb_txn_begin. */
2704 UNLOCK_MUTEX_W(env);
2709 mdb_txn_reset(MDB_txn *txn)
2714 /* This call is only valid for read-only txns */
2715 if (!(txn->mt_flags & MDB_TXN_RDONLY))
2718 mdb_txn_reset0(txn, "reset");
2722 mdb_txn_abort(MDB_txn *txn)
2728 mdb_txn_abort(txn->mt_child);
2730 mdb_txn_reset0(txn, "abort");
2731 /* Free reader slot tied to this txn (if MDB_NOTLS && writable FS) */
2732 if ((txn->mt_flags & MDB_TXN_RDONLY) && txn->mt_u.reader)
2733 txn->mt_u.reader->mr_pid = 0;
2738 /** Save the freelist as of this transaction to the freeDB.
2739 * This changes the freelist. Keep trying until it stabilizes.
2742 mdb_freelist_save(MDB_txn *txn)
2744 /* env->me_pghead[] can grow and shrink during this call.
2745 * env->me_pglast and txn->mt_free_pgs[] can only grow.
2746 * Page numbers cannot disappear from txn->mt_free_pgs[].
2749 MDB_env *env = txn->mt_env;
2750 int rc, maxfree_1pg = env->me_maxfree_1pg, more = 1;
2751 txnid_t pglast = 0, head_id = 0;
2752 pgno_t freecnt = 0, *free_pgs, *mop;
2753 ssize_t head_room = 0, total_room = 0, mop_len, clean_limit;
2755 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
2757 if (env->me_pghead) {
2758 /* Make sure first page of freeDB is touched and on freelist */
2759 rc = mdb_page_search(&mc, NULL, MDB_PS_FIRST|MDB_PS_MODIFY);
2760 if (rc && rc != MDB_NOTFOUND)
2764 /* Dispose of loose pages. Usually they will have all
2765 * been used up by the time we get here.
2767 if (txn->mt_loose_pgs) {
2768 MDB_page *mp = txn->mt_loose_pgs;
2769 /* Just return them to freeDB */
2770 if (env->me_pghead) {
2772 mop = env->me_pghead;
2773 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2774 pgno_t pg = mp->mp_pgno;
2776 for (i = mop[0]; i && mop[i] < pg; i--)
2782 /* Oh well, they were wasted. Put on freelist */
2783 for (; mp; mp = NEXT_LOOSE_PAGE(mp)) {
2784 mdb_midl_append(&txn->mt_free_pgs, mp->mp_pgno);
2787 txn->mt_loose_pgs = NULL;
2790 /* MDB_RESERVE cancels meminit in ovpage malloc (when no WRITEMAP) */
2791 clean_limit = (env->me_flags & (MDB_NOMEMINIT|MDB_WRITEMAP))
2792 ? SSIZE_MAX : maxfree_1pg;
2795 /* Come back here after each Put() in case freelist changed */
2800 /* If using records from freeDB which we have not yet
2801 * deleted, delete them and any we reserved for me_pghead.
2803 while (pglast < env->me_pglast) {
2804 rc = mdb_cursor_first(&mc, &key, NULL);
2807 pglast = head_id = *(txnid_t *)key.mv_data;
2808 total_room = head_room = 0;
2809 mdb_tassert(txn, pglast <= env->me_pglast);
2810 rc = mdb_cursor_del(&mc, 0);
2815 /* Save the IDL of pages freed by this txn, to a single record */
2816 if (freecnt < txn->mt_free_pgs[0]) {
2818 /* Make sure last page of freeDB is touched and on freelist */
2819 rc = mdb_page_search(&mc, NULL, MDB_PS_LAST|MDB_PS_MODIFY);
2820 if (rc && rc != MDB_NOTFOUND)
2823 free_pgs = txn->mt_free_pgs;
2824 /* Write to last page of freeDB */
2825 key.mv_size = sizeof(txn->mt_txnid);
2826 key.mv_data = &txn->mt_txnid;
2828 freecnt = free_pgs[0];
2829 data.mv_size = MDB_IDL_SIZEOF(free_pgs);
2830 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2833 /* Retry if mt_free_pgs[] grew during the Put() */
2834 free_pgs = txn->mt_free_pgs;
2835 } while (freecnt < free_pgs[0]);
2836 mdb_midl_sort(free_pgs);
2837 memcpy(data.mv_data, free_pgs, data.mv_size);
2840 unsigned int i = free_pgs[0];
2841 DPRINTF(("IDL write txn %"Z"u root %"Z"u num %u",
2842 txn->mt_txnid, txn->mt_dbs[FREE_DBI].md_root, i));
2844 DPRINTF(("IDL %"Z"u", free_pgs[i]));
2850 mop = env->me_pghead;
2851 mop_len = mop ? mop[0] : 0;
2853 /* Reserve records for me_pghead[]. Split it if multi-page,
2854 * to avoid searching freeDB for a page range. Use keys in
2855 * range [1,me_pglast]: Smaller than txnid of oldest reader.
2857 if (total_room >= mop_len) {
2858 if (total_room == mop_len || --more < 0)
2860 } else if (head_room >= maxfree_1pg && head_id > 1) {
2861 /* Keep current record (overflow page), add a new one */
2865 /* (Re)write {key = head_id, IDL length = head_room} */
2866 total_room -= head_room;
2867 head_room = mop_len - total_room;
2868 if (head_room > maxfree_1pg && head_id > 1) {
2869 /* Overflow multi-page for part of me_pghead */
2870 head_room /= head_id; /* amortize page sizes */
2871 head_room += maxfree_1pg - head_room % (maxfree_1pg + 1);
2872 } else if (head_room < 0) {
2873 /* Rare case, not bothering to delete this record */
2876 key.mv_size = sizeof(head_id);
2877 key.mv_data = &head_id;
2878 data.mv_size = (head_room + 1) * sizeof(pgno_t);
2879 rc = mdb_cursor_put(&mc, &key, &data, MDB_RESERVE);
2882 /* IDL is initially empty, zero out at least the length */
2883 pgs = (pgno_t *)data.mv_data;
2884 j = head_room > clean_limit ? head_room : 0;
2888 total_room += head_room;
2891 /* Fill in the reserved me_pghead records */
2897 rc = mdb_cursor_first(&mc, &key, &data);
2898 for (; !rc; rc = mdb_cursor_next(&mc, &key, &data, MDB_NEXT)) {
2899 txnid_t id = *(txnid_t *)key.mv_data;
2900 ssize_t len = (ssize_t)(data.mv_size / sizeof(MDB_ID)) - 1;
2903 mdb_tassert(txn, len >= 0 && id <= env->me_pglast);
2905 if (len > mop_len) {
2907 data.mv_size = (len + 1) * sizeof(MDB_ID);
2909 data.mv_data = mop -= len;
2912 rc = mdb_cursor_put(&mc, &key, &data, MDB_CURRENT);
2914 if (rc || !(mop_len -= len))
2921 /** Flush (some) dirty pages to the map, after clearing their dirty flag.
2922 * @param[in] txn the transaction that's being committed
2923 * @param[in] keep number of initial pages in dirty_list to keep dirty.
2924 * @return 0 on success, non-zero on failure.
2927 mdb_page_flush(MDB_txn *txn, int keep)
2929 MDB_env *env = txn->mt_env;
2930 MDB_ID2L dl = txn->mt_u.dirty_list;
2931 unsigned psize = env->me_psize, j;
2932 int i, pagecount = dl[0].mid, rc;
2933 size_t size = 0, pos = 0;
2935 MDB_page *dp = NULL;
2939 struct iovec iov[MDB_COMMIT_PAGES];
2940 ssize_t wpos = 0, wsize = 0, wres;
2941 size_t next_pos = 1; /* impossible pos, so pos != next_pos */
2947 if (env->me_flags & MDB_WRITEMAP) {
2948 /* Clear dirty flags */
2949 while (++i <= pagecount) {
2951 /* Don't flush this page yet */
2952 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
2953 dp->mp_flags &= ~P_KEEP;
2957 dp->mp_flags &= ~P_DIRTY;
2962 /* Write the pages */
2964 if (++i <= pagecount) {
2966 /* Don't flush this page yet */
2967 if (dp->mp_flags & (P_LOOSE|P_KEEP)) {
2968 dp->mp_flags &= ~P_KEEP;
2973 /* clear dirty flag */
2974 dp->mp_flags &= ~P_DIRTY;
2977 if (IS_OVERFLOW(dp)) size *= dp->mp_pages;
2982 /* Windows actually supports scatter/gather I/O, but only on
2983 * unbuffered file handles. Since we're relying on the OS page
2984 * cache for all our data, that's self-defeating. So we just
2985 * write pages one at a time. We use the ov structure to set
2986 * the write offset, to at least save the overhead of a Seek
2989 DPRINTF(("committing page %"Z"u", pgno));
2990 memset(&ov, 0, sizeof(ov));
2991 ov.Offset = pos & 0xffffffff;
2992 ov.OffsetHigh = pos >> 16 >> 16;
2993 if (!WriteFile(env->me_fd, dp, size, NULL, &ov)) {
2995 DPRINTF(("WriteFile: %d", rc));
2999 /* Write up to MDB_COMMIT_PAGES dirty pages at a time. */
3000 if (pos!=next_pos || n==MDB_COMMIT_PAGES || wsize+size>MAX_WRITE) {
3002 /* Write previous page(s) */
3003 #ifdef MDB_USE_PWRITEV
3004 wres = pwritev(env->me_fd, iov, n, wpos);
3007 wres = pwrite(env->me_fd, iov[0].iov_base, wsize, wpos);
3009 if (lseek(env->me_fd, wpos, SEEK_SET) == -1) {
3011 DPRINTF(("lseek: %s", strerror(rc)));
3014 wres = writev(env->me_fd, iov, n);
3017 if (wres != wsize) {
3020 DPRINTF(("Write error: %s", strerror(rc)));
3022 rc = EIO; /* TODO: Use which error code? */
3023 DPUTS("short write, filesystem full?");
3034 DPRINTF(("committing page %"Z"u", pgno));
3035 next_pos = pos + size;
3036 iov[n].iov_len = size;
3037 iov[n].iov_base = (char *)dp;
3043 for (i = keep; ++i <= pagecount; ) {
3045 /* This is a page we skipped above */
3048 dl[j].mid = dp->mp_pgno;
3051 mdb_dpage_free(env, dp);
3056 txn->mt_dirty_room += i - j;
3062 mdb_txn_commit(MDB_txn *txn)
3068 if (txn == NULL || txn->mt_env == NULL)
3071 if (txn->mt_child) {
3072 rc = mdb_txn_commit(txn->mt_child);
3073 txn->mt_child = NULL;
3080 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY)) {
3081 mdb_dbis_update(txn, 1);
3082 txn->mt_numdbs = 2; /* so txn_abort() doesn't close any new handles */
3087 if (F_ISSET(txn->mt_flags, MDB_TXN_ERROR)) {
3088 DPUTS("error flag is set, can't commit");
3090 txn->mt_parent->mt_flags |= MDB_TXN_ERROR;
3095 if (txn->mt_parent) {
3096 MDB_txn *parent = txn->mt_parent;
3100 unsigned x, y, len, ps_len;
3102 /* Append our free list to parent's */
3103 rc = mdb_midl_append_list(&parent->mt_free_pgs, txn->mt_free_pgs);
3106 mdb_midl_free(txn->mt_free_pgs);
3107 /* Failures after this must either undo the changes
3108 * to the parent or set MDB_TXN_ERROR in the parent.
3111 parent->mt_next_pgno = txn->mt_next_pgno;
3112 parent->mt_flags = txn->mt_flags;
3114 /* Merge our cursors into parent's and close them */
3115 mdb_cursors_close(txn, 1);
3117 /* Update parent's DB table. */
3118 memcpy(parent->mt_dbs, txn->mt_dbs, txn->mt_numdbs * sizeof(MDB_db));
3119 parent->mt_numdbs = txn->mt_numdbs;
3120 parent->mt_dbflags[0] = txn->mt_dbflags[0];
3121 parent->mt_dbflags[1] = txn->mt_dbflags[1];
3122 for (i=2; i<txn->mt_numdbs; i++) {
3123 /* preserve parent's DB_NEW status */
3124 x = parent->mt_dbflags[i] & DB_NEW;
3125 parent->mt_dbflags[i] = txn->mt_dbflags[i] | x;
3128 dst = parent->mt_u.dirty_list;
3129 src = txn->mt_u.dirty_list;
3130 /* Remove anything in our dirty list from parent's spill list */
3131 if ((pspill = parent->mt_spill_pgs) && (ps_len = pspill[0])) {
3133 pspill[0] = (pgno_t)-1;
3134 /* Mark our dirty pages as deleted in parent spill list */
3135 for (i=0, len=src[0].mid; ++i <= len; ) {
3136 MDB_ID pn = src[i].mid << 1;
3137 while (pn > pspill[x])
3139 if (pn == pspill[x]) {
3144 /* Squash deleted pagenums if we deleted any */
3145 for (x=y; ++x <= ps_len; )
3146 if (!(pspill[x] & 1))
3147 pspill[++y] = pspill[x];
3151 /* Find len = length of merging our dirty list with parent's */
3153 dst[0].mid = 0; /* simplify loops */
3154 if (parent->mt_parent) {
3155 len = x + src[0].mid;
3156 y = mdb_mid2l_search(src, dst[x].mid + 1) - 1;
3157 for (i = x; y && i; y--) {
3158 pgno_t yp = src[y].mid;
3159 while (yp < dst[i].mid)
3161 if (yp == dst[i].mid) {
3166 } else { /* Simplify the above for single-ancestor case */
3167 len = MDB_IDL_UM_MAX - txn->mt_dirty_room;
3169 /* Merge our dirty list with parent's */
3171 for (i = len; y; dst[i--] = src[y--]) {
3172 pgno_t yp = src[y].mid;
3173 while (yp < dst[x].mid)
3174 dst[i--] = dst[x--];
3175 if (yp == dst[x].mid)
3176 free(dst[x--].mptr);
3178 mdb_tassert(txn, i == x);
3180 free(txn->mt_u.dirty_list);
3181 parent->mt_dirty_room = txn->mt_dirty_room;
3182 if (txn->mt_spill_pgs) {
3183 if (parent->mt_spill_pgs) {
3184 /* TODO: Prevent failure here, so parent does not fail */
3185 rc = mdb_midl_append_list(&parent->mt_spill_pgs, txn->mt_spill_pgs);
3187 parent->mt_flags |= MDB_TXN_ERROR;
3188 mdb_midl_free(txn->mt_spill_pgs);
3189 mdb_midl_sort(parent->mt_spill_pgs);
3191 parent->mt_spill_pgs = txn->mt_spill_pgs;
3195 /* Append our loose page list to parent's */
3196 for (lp = &parent->mt_loose_pgs; *lp; lp = &NEXT_LOOSE_PAGE(lp))
3198 *lp = txn->mt_loose_pgs;
3200 parent->mt_child = NULL;
3201 mdb_midl_free(((MDB_ntxn *)txn)->mnt_pgstate.mf_pghead);
3206 if (txn != env->me_txn) {
3207 DPUTS("attempt to commit unknown transaction");
3212 mdb_cursors_close(txn, 0);
3214 if (!txn->mt_u.dirty_list[0].mid &&
3215 !(txn->mt_flags & (MDB_TXN_DIRTY|MDB_TXN_SPILLS)))
3218 DPRINTF(("committing txn %"Z"u %p on mdbenv %p, root page %"Z"u",
3219 txn->mt_txnid, (void*)txn, (void*)env, txn->mt_dbs[MAIN_DBI].md_root));
3221 /* Update DB root pointers */
3222 if (txn->mt_numdbs > 2) {
3226 data.mv_size = sizeof(MDB_db);
3228 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
3229 for (i = 2; i < txn->mt_numdbs; i++) {
3230 if (txn->mt_dbflags[i] & DB_DIRTY) {
3231 data.mv_data = &txn->mt_dbs[i];
3232 rc = mdb_cursor_put(&mc, &txn->mt_dbxs[i].md_name, &data, 0);
3239 rc = mdb_freelist_save(txn);
3243 mdb_midl_free(env->me_pghead);
3244 env->me_pghead = NULL;
3245 if (mdb_midl_shrink(&txn->mt_free_pgs))
3246 env->me_free_pgs = txn->mt_free_pgs;
3252 if ((rc = mdb_page_flush(txn, 0)) ||
3253 (rc = mdb_env_sync(env, 0)) ||
3254 (rc = mdb_env_write_meta(txn)))
3260 mdb_dbis_update(txn, 1);
3263 UNLOCK_MUTEX_W(env);
3273 /** Read the environment parameters of a DB environment before
3274 * mapping it into memory.
3275 * @param[in] env the environment handle
3276 * @param[out] meta address of where to store the meta information
3277 * @return 0 on success, non-zero on failure.
3280 mdb_env_read_header(MDB_env *env, MDB_meta *meta)
3286 enum { Size = sizeof(pbuf) };
3288 /* We don't know the page size yet, so use a minimum value.
3289 * Read both meta pages so we can use the latest one.
3292 for (i=off=0; i<2; i++, off = meta->mm_psize) {
3296 memset(&ov, 0, sizeof(ov));
3298 rc = ReadFile(env->me_fd, &pbuf, Size, &len, &ov) ? (int)len : -1;
3299 if (rc == -1 && ErrCode() == ERROR_HANDLE_EOF)
3302 rc = pread(env->me_fd, &pbuf, Size, off);
3305 if (rc == 0 && off == 0)
3307 rc = rc < 0 ? (int) ErrCode() : MDB_INVALID;
3308 DPRINTF(("read: %s", mdb_strerror(rc)));
3312 p = (MDB_page *)&pbuf;
3314 if (!F_ISSET(p->mp_flags, P_META)) {
3315 DPRINTF(("page %"Z"u not a meta page", p->mp_pgno));
3320 if (m->mm_magic != MDB_MAGIC) {
3321 DPUTS("meta has invalid magic");
3325 if (m->mm_version != MDB_DATA_VERSION) {
3326 DPRINTF(("database is version %u, expected version %u",
3327 m->mm_version, MDB_DATA_VERSION));
3328 return MDB_VERSION_MISMATCH;
3331 if (off == 0 || m->mm_txnid > meta->mm_txnid)
3338 mdb_env_init_meta0(MDB_env *env, MDB_meta *meta)
3340 meta->mm_magic = MDB_MAGIC;
3341 meta->mm_version = MDB_DATA_VERSION;
3342 meta->mm_mapsize = env->me_mapsize;
3343 meta->mm_psize = env->me_psize;
3344 meta->mm_last_pg = 1;
3345 meta->mm_flags = env->me_flags & 0xffff;
3346 meta->mm_flags |= MDB_INTEGERKEY;
3347 meta->mm_dbs[0].md_root = P_INVALID;
3348 meta->mm_dbs[1].md_root = P_INVALID;
3351 /** Write the environment parameters of a freshly created DB environment.
3352 * @param[in] env the environment handle
3353 * @param[out] meta address of where to store the meta information
3354 * @return 0 on success, non-zero on failure.
3357 mdb_env_init_meta(MDB_env *env, MDB_meta *meta)
3365 memset(&ov, 0, sizeof(ov));
3366 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3368 rc = WriteFile(fd, ptr, size, &len, &ov); } while(0)
3371 #define DO_PWRITE(rc, fd, ptr, size, len, pos) do { \
3372 len = pwrite(fd, ptr, size, pos); \
3373 rc = (len >= 0); } while(0)
3376 DPUTS("writing new meta page");
3378 psize = env->me_psize;
3380 mdb_env_init_meta0(env, meta);
3382 p = calloc(2, psize);
3384 p->mp_flags = P_META;
3385 *(MDB_meta *)METADATA(p) = *meta;
3387 q = (MDB_page *)((char *)p + psize);
3389 q->mp_flags = P_META;
3390 *(MDB_meta *)METADATA(q) = *meta;
3392 DO_PWRITE(rc, env->me_fd, p, psize * 2, len, 0);
3395 else if ((unsigned) len == psize * 2)
3403 /** Update the environment info to commit a transaction.
3404 * @param[in] txn the transaction that's being committed
3405 * @return 0 on success, non-zero on failure.
3408 mdb_env_write_meta(MDB_txn *txn)
3411 MDB_meta meta, metab, *mp;
3413 int rc, len, toggle;
3422 toggle = txn->mt_txnid & 1;
3423 DPRINTF(("writing meta page %d for root page %"Z"u",
3424 toggle, txn->mt_dbs[MAIN_DBI].md_root));
3427 mp = env->me_metas[toggle];
3429 if (env->me_flags & MDB_WRITEMAP) {
3430 /* Persist any increases of mapsize config */
3431 if (env->me_mapsize > mp->mm_mapsize)
3432 mp->mm_mapsize = env->me_mapsize;
3433 mp->mm_dbs[0] = txn->mt_dbs[0];
3434 mp->mm_dbs[1] = txn->mt_dbs[1];
3435 mp->mm_last_pg = txn->mt_next_pgno - 1;
3436 mp->mm_txnid = txn->mt_txnid;
3437 if (!(env->me_flags & (MDB_NOMETASYNC|MDB_NOSYNC))) {
3438 unsigned meta_size = env->me_psize;
3439 rc = (env->me_flags & MDB_MAPASYNC) ? MS_ASYNC : MS_SYNC;
3442 #ifndef _WIN32 /* POSIX msync() requires ptr = start of OS page */
3443 if (meta_size < env->me_os_psize)
3444 meta_size += meta_size;
3449 if (MDB_MSYNC(ptr, meta_size, rc)) {
3456 metab.mm_txnid = env->me_metas[toggle]->mm_txnid;
3457 metab.mm_last_pg = env->me_metas[toggle]->mm_last_pg;
3459 ptr = (char *)&meta;
3460 if (env->me_mapsize > mp->mm_mapsize) {
3461 /* Persist any increases of mapsize config */
3462 meta.mm_mapsize = env->me_mapsize;
3463 off = offsetof(MDB_meta, mm_mapsize);
3465 off = offsetof(MDB_meta, mm_dbs[0].md_depth);
3467 len = sizeof(MDB_meta) - off;
3470 meta.mm_dbs[0] = txn->mt_dbs[0];
3471 meta.mm_dbs[1] = txn->mt_dbs[1];
3472 meta.mm_last_pg = txn->mt_next_pgno - 1;
3473 meta.mm_txnid = txn->mt_txnid;
3476 off += env->me_psize;
3479 /* Write to the SYNC fd */
3480 mfd = env->me_flags & (MDB_NOSYNC|MDB_NOMETASYNC) ?
3481 env->me_fd : env->me_mfd;
3484 memset(&ov, 0, sizeof(ov));
3486 if (!WriteFile(mfd, ptr, len, (DWORD *)&rc, &ov))
3490 rc = pwrite(mfd, ptr, len, off);
3493 rc = rc < 0 ? ErrCode() : EIO;
3494 DPUTS("write failed, disk error?");
3495 /* On a failure, the pagecache still contains the new data.
3496 * Write some old data back, to prevent it from being used.
3497 * Use the non-SYNC fd; we know it will fail anyway.
3499 meta.mm_last_pg = metab.mm_last_pg;
3500 meta.mm_txnid = metab.mm_txnid;
3502 memset(&ov, 0, sizeof(ov));
3504 WriteFile(env->me_fd, ptr, len, NULL, &ov);
3506 r2 = pwrite(env->me_fd, ptr, len, off);
3507 (void)r2; /* Silence warnings. We don't care about pwrite's return value */
3510 env->me_flags |= MDB_FATAL_ERROR;
3514 /* Memory ordering issues are irrelevant; since the entire writer
3515 * is wrapped by wmutex, all of these changes will become visible
3516 * after the wmutex is unlocked. Since the DB is multi-version,
3517 * readers will get consistent data regardless of how fresh or
3518 * how stale their view of these values is.
3521 env->me_txns->mti_txnid = txn->mt_txnid;
3526 /** Check both meta pages to see which one is newer.
3527 * @param[in] env the environment handle
3528 * @return meta toggle (0 or 1).
3531 mdb_env_pick_meta(const MDB_env *env)
3533 return (env->me_metas[0]->mm_txnid < env->me_metas[1]->mm_txnid);
3537 mdb_env_create(MDB_env **env)
3541 e = calloc(1, sizeof(MDB_env));
3545 e->me_maxreaders = DEFAULT_READERS;
3546 e->me_maxdbs = e->me_numdbs = 2;
3547 e->me_fd = INVALID_HANDLE_VALUE;
3548 e->me_lfd = INVALID_HANDLE_VALUE;
3549 e->me_mfd = INVALID_HANDLE_VALUE;
3550 #ifdef MDB_USE_POSIX_SEM
3551 e->me_rmutex = SEM_FAILED;
3552 e->me_wmutex = SEM_FAILED;
3554 e->me_pid = getpid();
3555 GET_PAGESIZE(e->me_os_psize);
3556 VGMEMP_CREATE(e,0,0);
3562 mdb_env_map(MDB_env *env, void *addr, int newsize)
3565 unsigned int flags = env->me_flags;
3569 LONG sizelo, sizehi;
3572 if (flags & MDB_RDONLY) {
3577 msize = env->me_mapsize;
3578 sizelo = msize & 0xffffffff;
3579 sizehi = msize >> 16 >> 16; /* only needed on Win64 */
3582 /* Windows won't create mappings for zero length files.
3583 * Just allocate the maxsize right now.
3586 if (SetFilePointer(env->me_fd, sizelo, &sizehi, 0) != (DWORD)sizelo
3587 || !SetEndOfFile(env->me_fd)
3588 || SetFilePointer(env->me_fd, 0, NULL, 0) != 0)
3591 mh = CreateFileMapping(env->me_fd, NULL, flags & MDB_WRITEMAP ?
3592 PAGE_READWRITE : PAGE_READONLY,
3593 sizehi, sizelo, NULL);
3596 env->me_map = MapViewOfFileEx(mh, flags & MDB_WRITEMAP ?
3597 FILE_MAP_WRITE : FILE_MAP_READ,
3599 rc = env->me_map ? 0 : ErrCode();
3604 int prot = PROT_READ;
3605 if (flags & MDB_WRITEMAP) {
3607 if (ftruncate(env->me_fd, env->me_mapsize) < 0)
3610 env->me_map = mmap(addr, env->me_mapsize, prot, MAP_SHARED,
3612 if (env->me_map == MAP_FAILED) {
3617 if (flags & MDB_NORDAHEAD) {
3618 /* Turn off readahead. It's harmful when the DB is larger than RAM. */
3620 madvise(env->me_map, env->me_mapsize, MADV_RANDOM);
3622 #ifdef POSIX_MADV_RANDOM
3623 posix_madvise(env->me_map, env->me_mapsize, POSIX_MADV_RANDOM);
3624 #endif /* POSIX_MADV_RANDOM */
3625 #endif /* MADV_RANDOM */
3629 /* Can happen because the address argument to mmap() is just a
3630 * hint. mmap() can pick another, e.g. if the range is in use.
3631 * The MAP_FIXED flag would prevent that, but then mmap could
3632 * instead unmap existing pages to make room for the new map.
3634 if (addr && env->me_map != addr)
3635 return EBUSY; /* TODO: Make a new MDB_* error code? */
3637 p = (MDB_page *)env->me_map;
3638 env->me_metas[0] = METADATA(p);
3639 env->me_metas[1] = (MDB_meta *)((char *)env->me_metas[0] + env->me_psize);
3645 mdb_env_set_mapsize(MDB_env *env, size_t size)
3647 /* If env is already open, caller is responsible for making
3648 * sure there are no active txns.
3656 size = env->me_metas[mdb_env_pick_meta(env)]->mm_mapsize;
3657 else if (size < env->me_mapsize) {
3658 /* If the configured size is smaller, make sure it's
3659 * still big enough. Silently round up to minimum if not.
3661 size_t minsize = (env->me_metas[mdb_env_pick_meta(env)]->mm_last_pg + 1) * env->me_psize;
3665 munmap(env->me_map, env->me_mapsize);
3666 env->me_mapsize = size;
3667 old = (env->me_flags & MDB_FIXEDMAP) ? env->me_map : NULL;
3668 rc = mdb_env_map(env, old, 1);
3672 env->me_mapsize = size;
3674 env->me_maxpg = env->me_mapsize / env->me_psize;
3679 mdb_env_set_maxdbs(MDB_env *env, MDB_dbi dbs)
3683 env->me_maxdbs = dbs + 2; /* Named databases + main and free DB */
3688 mdb_env_set_maxreaders(MDB_env *env, unsigned int readers)
3690 if (env->me_map || readers < 1)
3692 env->me_maxreaders = readers;
3697 mdb_env_get_maxreaders(MDB_env *env, unsigned int *readers)
3699 if (!env || !readers)
3701 *readers = env->me_maxreaders;
3705 /** Further setup required for opening an LMDB environment
3708 mdb_env_open2(MDB_env *env)
3710 unsigned int flags = env->me_flags;
3711 int i, newenv = 0, rc;
3715 /* See if we should use QueryLimited */
3717 if ((rc & 0xff) > 5)
3718 env->me_pidquery = MDB_PROCESS_QUERY_LIMITED_INFORMATION;
3720 env->me_pidquery = PROCESS_QUERY_INFORMATION;
3723 memset(&meta, 0, sizeof(meta));
3725 if ((i = mdb_env_read_header(env, &meta)) != 0) {
3728 DPUTS("new mdbenv");
3730 env->me_psize = env->me_os_psize;
3731 if (env->me_psize > MAX_PAGESIZE)
3732 env->me_psize = MAX_PAGESIZE;
3734 env->me_psize = meta.mm_psize;
3737 /* Was a mapsize configured? */
3738 if (!env->me_mapsize) {
3739 /* If this is a new environment, take the default,
3740 * else use the size recorded in the existing env.
3742 env->me_mapsize = newenv ? DEFAULT_MAPSIZE : meta.mm_mapsize;
3743 } else if (env->me_mapsize < meta.mm_mapsize) {
3744 /* If the configured size is smaller, make sure it's
3745 * still big enough. Silently round up to minimum if not.
3747 size_t minsize = (meta.mm_last_pg + 1) * meta.mm_psize;
3748 if (env->me_mapsize < minsize)
3749 env->me_mapsize = minsize;
3752 rc = mdb_env_map(env, meta.mm_address, newenv || env->me_mapsize != meta.mm_mapsize);
3757 if (flags & MDB_FIXEDMAP)
3758 meta.mm_address = env->me_map;
3759 i = mdb_env_init_meta(env, &meta);
3760 if (i != MDB_SUCCESS) {
3765 env->me_maxfree_1pg = (env->me_psize - PAGEHDRSZ) / sizeof(pgno_t) - 1;
3766 env->me_nodemax = (((env->me_psize - PAGEHDRSZ) / MDB_MINKEYS) & -2)
3768 #if !(MDB_MAXKEYSIZE)
3769 env->me_maxkey = env->me_nodemax - (NODESIZE + sizeof(MDB_db));
3771 env->me_maxpg = env->me_mapsize / env->me_psize;
3775 int toggle = mdb_env_pick_meta(env);
3776 MDB_db *db = &env->me_metas[toggle]->mm_dbs[MAIN_DBI];
3778 DPRINTF(("opened database version %u, pagesize %u",
3779 env->me_metas[0]->mm_version, env->me_psize));
3780 DPRINTF(("using meta page %d", toggle));
3781 DPRINTF(("depth: %u", db->md_depth));
3782 DPRINTF(("entries: %"Z"u", db->md_entries));
3783 DPRINTF(("branch pages: %"Z"u", db->md_branch_pages));
3784 DPRINTF(("leaf pages: %"Z"u", db->md_leaf_pages));
3785 DPRINTF(("overflow pages: %"Z"u", db->md_overflow_pages));
3786 DPRINTF(("root: %"Z"u", db->md_root));
3794 /** Release a reader thread's slot in the reader lock table.
3795 * This function is called automatically when a thread exits.
3796 * @param[in] ptr This points to the slot in the reader lock table.
3799 mdb_env_reader_dest(void *ptr)
3801 MDB_reader *reader = ptr;
3807 /** Junk for arranging thread-specific callbacks on Windows. This is
3808 * necessarily platform and compiler-specific. Windows supports up
3809 * to 1088 keys. Let's assume nobody opens more than 64 environments
3810 * in a single process, for now. They can override this if needed.
3812 #ifndef MAX_TLS_KEYS
3813 #define MAX_TLS_KEYS 64
3815 static pthread_key_t mdb_tls_keys[MAX_TLS_KEYS];
3816 static int mdb_tls_nkeys;
3818 static void NTAPI mdb_tls_callback(PVOID module, DWORD reason, PVOID ptr)
3822 case DLL_PROCESS_ATTACH: break;
3823 case DLL_THREAD_ATTACH: break;
3824 case DLL_THREAD_DETACH:
3825 for (i=0; i<mdb_tls_nkeys; i++) {
3826 MDB_reader *r = pthread_getspecific(mdb_tls_keys[i]);
3828 mdb_env_reader_dest(r);
3832 case DLL_PROCESS_DETACH: break;
3837 const PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3839 PIMAGE_TLS_CALLBACK mdb_tls_cbp __attribute__((section (".CRT$XLB"))) = mdb_tls_callback;
3843 /* Force some symbol references.
3844 * _tls_used forces the linker to create the TLS directory if not already done
3845 * mdb_tls_cbp prevents whole-program-optimizer from dropping the symbol.
3847 #pragma comment(linker, "/INCLUDE:_tls_used")
3848 #pragma comment(linker, "/INCLUDE:mdb_tls_cbp")
3849 #pragma const_seg(".CRT$XLB")
3850 extern const PIMAGE_TLS_CALLBACK mdb_tls_cbp;
3851 const PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3854 #pragma comment(linker, "/INCLUDE:__tls_used")
3855 #pragma comment(linker, "/INCLUDE:_mdb_tls_cbp")
3856 #pragma data_seg(".CRT$XLB")
3857 PIMAGE_TLS_CALLBACK mdb_tls_cbp = mdb_tls_callback;
3859 #endif /* WIN 32/64 */
3860 #endif /* !__GNUC__ */
3863 /** Downgrade the exclusive lock on the region back to shared */
3865 mdb_env_share_locks(MDB_env *env, int *excl)
3867 int rc = 0, toggle = mdb_env_pick_meta(env);
3869 env->me_txns->mti_txnid = env->me_metas[toggle]->mm_txnid;
3874 /* First acquire a shared lock. The Unlock will
3875 * then release the existing exclusive lock.
3877 memset(&ov, 0, sizeof(ov));
3878 if (!LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3881 UnlockFile(env->me_lfd, 0, 0, 1, 0);
3887 struct flock lock_info;
3888 /* The shared lock replaces the existing lock */
3889 memset((void *)&lock_info, 0, sizeof(lock_info));
3890 lock_info.l_type = F_RDLCK;
3891 lock_info.l_whence = SEEK_SET;
3892 lock_info.l_start = 0;
3893 lock_info.l_len = 1;
3894 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3895 (rc = ErrCode()) == EINTR) ;
3896 *excl = rc ? -1 : 0; /* error may mean we lost the lock */
3903 /** Try to get exlusive lock, otherwise shared.
3904 * Maintain *excl = -1: no/unknown lock, 0: shared, 1: exclusive.
3907 mdb_env_excl_lock(MDB_env *env, int *excl)
3911 if (LockFile(env->me_lfd, 0, 0, 1, 0)) {
3915 memset(&ov, 0, sizeof(ov));
3916 if (LockFileEx(env->me_lfd, 0, 0, 1, 0, &ov)) {
3923 struct flock lock_info;
3924 memset((void *)&lock_info, 0, sizeof(lock_info));
3925 lock_info.l_type = F_WRLCK;
3926 lock_info.l_whence = SEEK_SET;
3927 lock_info.l_start = 0;
3928 lock_info.l_len = 1;
3929 while ((rc = fcntl(env->me_lfd, F_SETLK, &lock_info)) &&
3930 (rc = ErrCode()) == EINTR) ;
3934 # ifdef MDB_USE_POSIX_SEM
3935 if (*excl < 0) /* always true when !MDB_USE_POSIX_SEM */
3938 lock_info.l_type = F_RDLCK;
3939 while ((rc = fcntl(env->me_lfd, F_SETLKW, &lock_info)) &&
3940 (rc = ErrCode()) == EINTR) ;
3950 * hash_64 - 64 bit Fowler/Noll/Vo-0 FNV-1a hash code
3952 * @(#) $Revision: 5.1 $
3953 * @(#) $Id: hash_64a.c,v 5.1 2009/06/30 09:01:38 chongo Exp $
3954 * @(#) $Source: /usr/local/src/cmd/fnv/RCS/hash_64a.c,v $
3956 * http://www.isthe.com/chongo/tech/comp/fnv/index.html
3960 * Please do not copyright this code. This code is in the public domain.
3962 * LANDON CURT NOLL DISCLAIMS ALL WARRANTIES WITH REGARD TO THIS SOFTWARE,
3963 * INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS. IN NO
3964 * EVENT SHALL LANDON CURT NOLL BE LIABLE FOR ANY SPECIAL, INDIRECT OR
3965 * CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM LOSS OF
3966 * USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
3967 * OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
3968 * PERFORMANCE OF THIS SOFTWARE.
3971 * chongo <Landon Curt Noll> /\oo/\
3972 * http://www.isthe.com/chongo/
3974 * Share and Enjoy! :-)
3977 typedef unsigned long long mdb_hash_t;
3978 #define MDB_HASH_INIT ((mdb_hash_t)0xcbf29ce484222325ULL)
3980 /** perform a 64 bit Fowler/Noll/Vo FNV-1a hash on a buffer
3981 * @param[in] val value to hash
3982 * @param[in] hval initial value for hash
3983 * @return 64 bit hash
3985 * NOTE: To use the recommended 64 bit FNV-1a hash, use MDB_HASH_INIT as the
3986 * hval arg on the first call.
3989 mdb_hash_val(MDB_val *val, mdb_hash_t hval)
3991 unsigned char *s = (unsigned char *)val->mv_data; /* unsigned string */
3992 unsigned char *end = s + val->mv_size;
3994 * FNV-1a hash each octet of the string
3997 /* xor the bottom with the current octet */
3998 hval ^= (mdb_hash_t)*s++;
4000 /* multiply by the 64 bit FNV magic prime mod 2^64 */
4001 hval += (hval << 1) + (hval << 4) + (hval << 5) +
4002 (hval << 7) + (hval << 8) + (hval << 40);
4004 /* return our new hash value */
4008 /** Hash the string and output the encoded hash.
4009 * This uses modified RFC1924 Ascii85 encoding to accommodate systems with
4010 * very short name limits. We don't care about the encoding being reversible,
4011 * we just want to preserve as many bits of the input as possible in a
4012 * small printable string.
4013 * @param[in] str string to hash
4014 * @param[out] encbuf an array of 11 chars to hold the hash
4016 static const char mdb_a85[]= "0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz!#$%&()*+-;<=>?@^_`{|}~";
4019 mdb_pack85(unsigned long l, char *out)
4023 for (i=0; i<5; i++) {
4024 *out++ = mdb_a85[l % 85];
4030 mdb_hash_enc(MDB_val *val, char *encbuf)
4032 mdb_hash_t h = mdb_hash_val(val, MDB_HASH_INIT);
4034 mdb_pack85(h, encbuf);
4035 mdb_pack85(h>>32, encbuf+5);
4040 /** Open and/or initialize the lock region for the environment.
4041 * @param[in] env The LMDB environment.
4042 * @param[in] lpath The pathname of the file used for the lock region.
4043 * @param[in] mode The Unix permissions for the file, if we create it.
4044 * @param[out] excl Resulting file lock type: -1 none, 0 shared, 1 exclusive
4045 * @param[in,out] excl In -1, out lock type: -1 none, 0 shared, 1 exclusive
4046 * @return 0 on success, non-zero on failure.
4049 mdb_env_setup_locks(MDB_env *env, char *lpath, int mode, int *excl)
4052 # define MDB_ERRCODE_ROFS ERROR_WRITE_PROTECT
4054 # define MDB_ERRCODE_ROFS EROFS
4055 #ifdef O_CLOEXEC /* Linux: Open file and set FD_CLOEXEC atomically */
4056 # define MDB_CLOEXEC O_CLOEXEC
4059 # define MDB_CLOEXEC 0
4066 env->me_lfd = CreateFile(lpath, GENERIC_READ|GENERIC_WRITE,
4067 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, OPEN_ALWAYS,
4068 FILE_ATTRIBUTE_NORMAL, NULL);
4070 env->me_lfd = open(lpath, O_RDWR|O_CREAT|MDB_CLOEXEC, mode);
4072 if (env->me_lfd == INVALID_HANDLE_VALUE) {
4074 if (rc == MDB_ERRCODE_ROFS && (env->me_flags & MDB_RDONLY)) {
4079 #if ! ((MDB_CLOEXEC) || defined(_WIN32))
4080 /* Lose record locks when exec*() */
4081 if ((fdflags = fcntl(env->me_lfd, F_GETFD) | FD_CLOEXEC) >= 0)
4082 fcntl(env->me_lfd, F_SETFD, fdflags);
4085 if (!(env->me_flags & MDB_NOTLS)) {
4086 rc = pthread_key_create(&env->me_txkey, mdb_env_reader_dest);
4089 env->me_flags |= MDB_ENV_TXKEY;
4091 /* Windows TLS callbacks need help finding their TLS info. */
4092 if (mdb_tls_nkeys >= MAX_TLS_KEYS) {
4096 mdb_tls_keys[mdb_tls_nkeys++] = env->me_txkey;
4100 /* Try to get exclusive lock. If we succeed, then
4101 * nobody is using the lock region and we should initialize it.
4103 if ((rc = mdb_env_excl_lock(env, excl))) goto fail;
4106 size = GetFileSize(env->me_lfd, NULL);
4108 size = lseek(env->me_lfd, 0, SEEK_END);
4109 if (size == -1) goto fail_errno;
4111 rsize = (env->me_maxreaders-1) * sizeof(MDB_reader) + sizeof(MDB_txninfo);
4112 if (size < rsize && *excl > 0) {
4114 if (SetFilePointer(env->me_lfd, rsize, NULL, FILE_BEGIN) != (DWORD)rsize
4115 || !SetEndOfFile(env->me_lfd))
4118 if (ftruncate(env->me_lfd, rsize) != 0) goto fail_errno;
4122 size = rsize - sizeof(MDB_txninfo);
4123 env->me_maxreaders = size/sizeof(MDB_reader) + 1;
4128 mh = CreateFileMapping(env->me_lfd, NULL, PAGE_READWRITE,
4130 if (!mh) goto fail_errno;
4131 env->me_txns = MapViewOfFileEx(mh, FILE_MAP_WRITE, 0, 0, rsize, NULL);
4133 if (!env->me_txns) goto fail_errno;
4135 void *m = mmap(NULL, rsize, PROT_READ|PROT_WRITE, MAP_SHARED,
4137 if (m == MAP_FAILED) goto fail_errno;
4143 BY_HANDLE_FILE_INFORMATION stbuf;
4152 if (!mdb_sec_inited) {
4153 InitializeSecurityDescriptor(&mdb_null_sd,
4154 SECURITY_DESCRIPTOR_REVISION);
4155 SetSecurityDescriptorDacl(&mdb_null_sd, TRUE, 0, FALSE);
4156 mdb_all_sa.nLength = sizeof(SECURITY_ATTRIBUTES);
4157 mdb_all_sa.bInheritHandle = FALSE;
4158 mdb_all_sa.lpSecurityDescriptor = &mdb_null_sd;
4161 if (!GetFileInformationByHandle(env->me_lfd, &stbuf)) goto fail_errno;
4162 idbuf.volume = stbuf.dwVolumeSerialNumber;
4163 idbuf.nhigh = stbuf.nFileIndexHigh;
4164 idbuf.nlow = stbuf.nFileIndexLow;
4165 val.mv_data = &idbuf;
4166 val.mv_size = sizeof(idbuf);
4167 mdb_hash_enc(&val, encbuf);
4168 sprintf(env->me_txns->mti_rmname, "Global\\MDBr%s", encbuf);
4169 sprintf(env->me_txns->mti_wmname, "Global\\MDBw%s", encbuf);
4170 env->me_rmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_rmname);
4171 if (!env->me_rmutex) goto fail_errno;
4172 env->me_wmutex = CreateMutex(&mdb_all_sa, FALSE, env->me_txns->mti_wmname);
4173 if (!env->me_wmutex) goto fail_errno;
4174 #elif defined(MDB_USE_POSIX_SEM)
4183 #if defined(__NetBSD__)
4184 #define MDB_SHORT_SEMNAMES 1 /* limited to 14 chars */
4186 if (fstat(env->me_lfd, &stbuf)) goto fail_errno;
4187 idbuf.dev = stbuf.st_dev;
4188 idbuf.ino = stbuf.st_ino;
4189 val.mv_data = &idbuf;
4190 val.mv_size = sizeof(idbuf);
4191 mdb_hash_enc(&val, encbuf);
4192 #ifdef MDB_SHORT_SEMNAMES
4193 encbuf[9] = '\0'; /* drop name from 15 chars to 14 chars */
4195 sprintf(env->me_txns->mti_rmname, "/MDBr%s", encbuf);
4196 sprintf(env->me_txns->mti_wmname, "/MDBw%s", encbuf);
4197 /* Clean up after a previous run, if needed: Try to
4198 * remove both semaphores before doing anything else.
4200 sem_unlink(env->me_txns->mti_rmname);
4201 sem_unlink(env->me_txns->mti_wmname);
4202 env->me_rmutex = sem_open(env->me_txns->mti_rmname,
4203 O_CREAT|O_EXCL, mode, 1);
4204 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4205 env->me_wmutex = sem_open(env->me_txns->mti_wmname,
4206 O_CREAT|O_EXCL, mode, 1);
4207 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4208 #else /* MDB_USE_POSIX_SEM */
4209 pthread_mutexattr_t mattr;
4211 if ((rc = pthread_mutexattr_init(&mattr))
4212 || (rc = pthread_mutexattr_setpshared(&mattr, PTHREAD_PROCESS_SHARED))
4213 || (rc = pthread_mutex_init(&env->me_txns->mti_mutex, &mattr))
4214 || (rc = pthread_mutex_init(&env->me_txns->mti_wmutex, &mattr)))
4216 pthread_mutexattr_destroy(&mattr);
4217 #endif /* _WIN32 || MDB_USE_POSIX_SEM */
4219 env->me_txns->mti_magic = MDB_MAGIC;
4220 env->me_txns->mti_format = MDB_LOCK_FORMAT;
4221 env->me_txns->mti_txnid = 0;
4222 env->me_txns->mti_numreaders = 0;
4225 if (env->me_txns->mti_magic != MDB_MAGIC) {
4226 DPUTS("lock region has invalid magic");
4230 if (env->me_txns->mti_format != MDB_LOCK_FORMAT) {
4231 DPRINTF(("lock region has format+version 0x%x, expected 0x%x",
4232 env->me_txns->mti_format, MDB_LOCK_FORMAT));
4233 rc = MDB_VERSION_MISMATCH;
4237 if (rc && rc != EACCES && rc != EAGAIN) {
4241 env->me_rmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_rmname);
4242 if (!env->me_rmutex) goto fail_errno;
4243 env->me_wmutex = OpenMutex(SYNCHRONIZE, FALSE, env->me_txns->mti_wmname);
4244 if (!env->me_wmutex) goto fail_errno;
4245 #elif defined(MDB_USE_POSIX_SEM)
4246 env->me_rmutex = sem_open(env->me_txns->mti_rmname, 0);
4247 if (env->me_rmutex == SEM_FAILED) goto fail_errno;
4248 env->me_wmutex = sem_open(env->me_txns->mti_wmname, 0);
4249 if (env->me_wmutex == SEM_FAILED) goto fail_errno;
4260 /** The name of the lock file in the DB environment */
4261 #define LOCKNAME "/lock.mdb"
4262 /** The name of the data file in the DB environment */
4263 #define DATANAME "/data.mdb"
4264 /** The suffix of the lock file when no subdir is used */
4265 #define LOCKSUFF "-lock"
4266 /** Only a subset of the @ref mdb_env flags can be changed
4267 * at runtime. Changing other flags requires closing the
4268 * environment and re-opening it with the new flags.
4270 #define CHANGEABLE (MDB_NOSYNC|MDB_NOMETASYNC|MDB_MAPASYNC|MDB_NOMEMINIT)
4271 #define CHANGELESS (MDB_FIXEDMAP|MDB_NOSUBDIR|MDB_RDONLY|MDB_WRITEMAP| \
4272 MDB_NOTLS|MDB_NOLOCK|MDB_NORDAHEAD)
4274 #if VALID_FLAGS & PERSISTENT_FLAGS & (CHANGEABLE|CHANGELESS)
4275 # error "Persistent DB flags & env flags overlap, but both go in mm_flags"
4279 mdb_env_open(MDB_env *env, const char *path, unsigned int flags, mdb_mode_t mode)
4281 int oflags, rc, len, excl = -1;
4282 char *lpath, *dpath;
4284 if (env->me_fd!=INVALID_HANDLE_VALUE || (flags & ~(CHANGEABLE|CHANGELESS)))
4288 if (flags & MDB_NOSUBDIR) {
4289 rc = len + sizeof(LOCKSUFF) + len + 1;
4291 rc = len + sizeof(LOCKNAME) + len + sizeof(DATANAME);
4296 if (flags & MDB_NOSUBDIR) {
4297 dpath = lpath + len + sizeof(LOCKSUFF);
4298 sprintf(lpath, "%s" LOCKSUFF, path);
4299 strcpy(dpath, path);
4301 dpath = lpath + len + sizeof(LOCKNAME);
4302 sprintf(lpath, "%s" LOCKNAME, path);
4303 sprintf(dpath, "%s" DATANAME, path);
4307 flags |= env->me_flags;
4308 if (flags & MDB_RDONLY) {
4309 /* silently ignore WRITEMAP when we're only getting read access */
4310 flags &= ~MDB_WRITEMAP;
4312 if (!((env->me_free_pgs = mdb_midl_alloc(MDB_IDL_UM_MAX)) &&
4313 (env->me_dirty_list = calloc(MDB_IDL_UM_SIZE, sizeof(MDB_ID2)))))
4316 env->me_flags = flags |= MDB_ENV_ACTIVE;
4320 env->me_path = strdup(path);
4321 env->me_dbxs = calloc(env->me_maxdbs, sizeof(MDB_dbx));
4322 env->me_dbflags = calloc(env->me_maxdbs, sizeof(uint16_t));
4323 if (!(env->me_dbxs && env->me_path && env->me_dbflags)) {
4328 /* For RDONLY, get lockfile after we know datafile exists */
4329 if (!(flags & (MDB_RDONLY|MDB_NOLOCK))) {
4330 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4336 if (F_ISSET(flags, MDB_RDONLY)) {
4337 oflags = GENERIC_READ;
4338 len = OPEN_EXISTING;
4340 oflags = GENERIC_READ|GENERIC_WRITE;
4343 mode = FILE_ATTRIBUTE_NORMAL;
4344 env->me_fd = CreateFile(dpath, oflags, FILE_SHARE_READ|FILE_SHARE_WRITE,
4345 NULL, len, mode, NULL);
4347 if (F_ISSET(flags, MDB_RDONLY))
4350 oflags = O_RDWR | O_CREAT;
4352 env->me_fd = open(dpath, oflags, mode);
4354 if (env->me_fd == INVALID_HANDLE_VALUE) {
4359 if ((flags & (MDB_RDONLY|MDB_NOLOCK)) == MDB_RDONLY) {
4360 rc = mdb_env_setup_locks(env, lpath, mode, &excl);
4365 if ((rc = mdb_env_open2(env)) == MDB_SUCCESS) {
4366 if (flags & (MDB_RDONLY|MDB_WRITEMAP)) {
4367 env->me_mfd = env->me_fd;
4369 /* Synchronous fd for meta writes. Needed even with
4370 * MDB_NOSYNC/MDB_NOMETASYNC, in case these get reset.
4373 len = OPEN_EXISTING;
4374 env->me_mfd = CreateFile(dpath, oflags,
4375 FILE_SHARE_READ|FILE_SHARE_WRITE, NULL, len,
4376 mode | FILE_FLAG_WRITE_THROUGH, NULL);
4379 env->me_mfd = open(dpath, oflags | MDB_DSYNC, mode);
4381 if (env->me_mfd == INVALID_HANDLE_VALUE) {
4386 DPRINTF(("opened dbenv %p", (void *) env));
4388 rc = mdb_env_share_locks(env, &excl);
4392 if (!((flags & MDB_RDONLY) ||
4393 (env->me_pbuf = calloc(1, env->me_psize))))
4399 mdb_env_close0(env, excl);
4405 /** Destroy resources from mdb_env_open(), clear our readers & DBIs */
4407 mdb_env_close0(MDB_env *env, int excl)
4411 if (!(env->me_flags & MDB_ENV_ACTIVE))
4414 /* Doing this here since me_dbxs may not exist during mdb_env_close */
4415 for (i = env->me_maxdbs; --i > MAIN_DBI; )
4416 free(env->me_dbxs[i].md_name.mv_data);
4419 free(env->me_dbflags);
4422 free(env->me_dirty_list);
4423 mdb_midl_free(env->me_free_pgs);
4425 if (env->me_flags & MDB_ENV_TXKEY) {
4426 pthread_key_delete(env->me_txkey);
4428 /* Delete our key from the global list */
4429 for (i=0; i<mdb_tls_nkeys; i++)
4430 if (mdb_tls_keys[i] == env->me_txkey) {
4431 mdb_tls_keys[i] = mdb_tls_keys[mdb_tls_nkeys-1];
4439 munmap(env->me_map, env->me_mapsize);
4441 if (env->me_mfd != env->me_fd && env->me_mfd != INVALID_HANDLE_VALUE)
4442 (void) close(env->me_mfd);
4443 if (env->me_fd != INVALID_HANDLE_VALUE)
4444 (void) close(env->me_fd);
4446 MDB_PID_T pid = env->me_pid;
4447 /* Clearing readers is done in this function because
4448 * me_txkey with its destructor must be disabled first.
4450 for (i = env->me_numreaders; --i >= 0; )
4451 if (env->me_txns->mti_readers[i].mr_pid == pid)
4452 env->me_txns->mti_readers[i].mr_pid = 0;
4454 if (env->me_rmutex) {
4455 CloseHandle(env->me_rmutex);
4456 if (env->me_wmutex) CloseHandle(env->me_wmutex);
4458 /* Windows automatically destroys the mutexes when
4459 * the last handle closes.
4461 #elif defined(MDB_USE_POSIX_SEM)
4462 if (env->me_rmutex != SEM_FAILED) {
4463 sem_close(env->me_rmutex);
4464 if (env->me_wmutex != SEM_FAILED)
4465 sem_close(env->me_wmutex);
4466 /* If we have the filelock: If we are the
4467 * only remaining user, clean up semaphores.
4470 mdb_env_excl_lock(env, &excl);
4472 sem_unlink(env->me_txns->mti_rmname);
4473 sem_unlink(env->me_txns->mti_wmname);
4477 munmap((void *)env->me_txns, (env->me_maxreaders-1)*sizeof(MDB_reader)+sizeof(MDB_txninfo));
4479 if (env->me_lfd != INVALID_HANDLE_VALUE) {
4482 /* Unlock the lockfile. Windows would have unlocked it
4483 * after closing anyway, but not necessarily at once.
4485 UnlockFile(env->me_lfd, 0, 0, 1, 0);
4488 (void) close(env->me_lfd);
4491 env->me_flags &= ~(MDB_ENV_ACTIVE|MDB_ENV_TXKEY);
4496 mdb_env_close(MDB_env *env)
4503 VGMEMP_DESTROY(env);
4504 while ((dp = env->me_dpages) != NULL) {
4505 VGMEMP_DEFINED(&dp->mp_next, sizeof(dp->mp_next));
4506 env->me_dpages = dp->mp_next;
4510 mdb_env_close0(env, 0);
4514 /** Compare two items pointing at aligned size_t's */
4516 mdb_cmp_long(const MDB_val *a, const MDB_val *b)
4518 return (*(size_t *)a->mv_data < *(size_t *)b->mv_data) ? -1 :
4519 *(size_t *)a->mv_data > *(size_t *)b->mv_data;
4522 /** Compare two items pointing at aligned unsigned int's */
4524 mdb_cmp_int(const MDB_val *a, const MDB_val *b)
4526 return (*(unsigned int *)a->mv_data < *(unsigned int *)b->mv_data) ? -1 :
4527 *(unsigned int *)a->mv_data > *(unsigned int *)b->mv_data;
4530 /** Compare two items pointing at unsigned ints of unknown alignment.
4531 * Nodes and keys are guaranteed to be 2-byte aligned.
4534 mdb_cmp_cint(const MDB_val *a, const MDB_val *b)
4536 #if BYTE_ORDER == LITTLE_ENDIAN
4537 unsigned short *u, *c;
4540 u = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4541 c = (unsigned short *) ((char *) b->mv_data + a->mv_size);
4544 } while(!x && u > (unsigned short *)a->mv_data);
4547 unsigned short *u, *c, *end;
4550 end = (unsigned short *) ((char *) a->mv_data + a->mv_size);
4551 u = (unsigned short *)a->mv_data;
4552 c = (unsigned short *)b->mv_data;
4555 } while(!x && u < end);
4560 /** Compare two items pointing at size_t's of unknown alignment. */
4561 #ifdef MISALIGNED_OK
4562 # define mdb_cmp_clong mdb_cmp_long
4564 # define mdb_cmp_clong mdb_cmp_cint
4567 /** Compare two items lexically */
4569 mdb_cmp_memn(const MDB_val *a, const MDB_val *b)
4576 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4582 diff = memcmp(a->mv_data, b->mv_data, len);
4583 return diff ? diff : len_diff<0 ? -1 : len_diff;
4586 /** Compare two items in reverse byte order */
4588 mdb_cmp_memnr(const MDB_val *a, const MDB_val *b)
4590 const unsigned char *p1, *p2, *p1_lim;
4594 p1_lim = (const unsigned char *)a->mv_data;
4595 p1 = (const unsigned char *)a->mv_data + a->mv_size;
4596 p2 = (const unsigned char *)b->mv_data + b->mv_size;
4598 len_diff = (ssize_t) a->mv_size - (ssize_t) b->mv_size;
4604 while (p1 > p1_lim) {
4605 diff = *--p1 - *--p2;
4609 return len_diff<0 ? -1 : len_diff;
4612 /** Search for key within a page, using binary search.
4613 * Returns the smallest entry larger or equal to the key.
4614 * If exactp is non-null, stores whether the found entry was an exact match
4615 * in *exactp (1 or 0).
4616 * Updates the cursor index with the index of the found entry.
4617 * If no entry larger or equal to the key is found, returns NULL.
4620 mdb_node_search(MDB_cursor *mc, MDB_val *key, int *exactp)
4622 unsigned int i = 0, nkeys;
4625 MDB_page *mp = mc->mc_pg[mc->mc_top];
4626 MDB_node *node = NULL;
4631 nkeys = NUMKEYS(mp);
4633 DPRINTF(("searching %u keys in %s %spage %"Z"u",
4634 nkeys, IS_LEAF(mp) ? "leaf" : "branch", IS_SUBP(mp) ? "sub-" : "",
4637 low = IS_LEAF(mp) ? 0 : 1;
4639 cmp = mc->mc_dbx->md_cmp;
4641 /* Branch pages have no data, so if using integer keys,
4642 * alignment is guaranteed. Use faster mdb_cmp_int.
4644 if (cmp == mdb_cmp_cint && IS_BRANCH(mp)) {
4645 if (NODEPTR(mp, 1)->mn_ksize == sizeof(size_t))
4652 nodekey.mv_size = mc->mc_db->md_pad;
4653 node = NODEPTR(mp, 0); /* fake */
4654 while (low <= high) {
4655 i = (low + high) >> 1;
4656 nodekey.mv_data = LEAF2KEY(mp, i, nodekey.mv_size);
4657 rc = cmp(key, &nodekey);
4658 DPRINTF(("found leaf index %u [%s], rc = %i",
4659 i, DKEY(&nodekey), rc));
4668 while (low <= high) {
4669 i = (low + high) >> 1;
4671 node = NODEPTR(mp, i);
4672 nodekey.mv_size = NODEKSZ(node);
4673 nodekey.mv_data = NODEKEY(node);
4675 rc = cmp(key, &nodekey);
4678 DPRINTF(("found leaf index %u [%s], rc = %i",
4679 i, DKEY(&nodekey), rc));
4681 DPRINTF(("found branch index %u [%s -> %"Z"u], rc = %i",
4682 i, DKEY(&nodekey), NODEPGNO(node), rc));
4693 if (rc > 0) { /* Found entry is less than the key. */
4694 i++; /* Skip to get the smallest entry larger than key. */
4696 node = NODEPTR(mp, i);
4699 *exactp = (rc == 0 && nkeys > 0);
4700 /* store the key index */
4701 mc->mc_ki[mc->mc_top] = i;
4703 /* There is no entry larger or equal to the key. */
4706 /* nodeptr is fake for LEAF2 */
4712 mdb_cursor_adjust(MDB_cursor *mc, func)
4716 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
4717 if (m2->mc_pg[m2->mc_top] == mc->mc_pg[mc->mc_top]) {
4724 /** Pop a page off the top of the cursor's stack. */
4726 mdb_cursor_pop(MDB_cursor *mc)
4730 MDB_page *top = mc->mc_pg[mc->mc_top];
4736 DPRINTF(("popped page %"Z"u off db %d cursor %p", top->mp_pgno,
4737 DDBI(mc), (void *) mc));
4741 /** Push a page onto the top of the cursor's stack. */
4743 mdb_cursor_push(MDB_cursor *mc, MDB_page *mp)
4745 DPRINTF(("pushing page %"Z"u on db %d cursor %p", mp->mp_pgno,
4746 DDBI(mc), (void *) mc));
4748 if (mc->mc_snum >= CURSOR_STACK) {
4749 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4750 return MDB_CURSOR_FULL;
4753 mc->mc_top = mc->mc_snum++;
4754 mc->mc_pg[mc->mc_top] = mp;
4755 mc->mc_ki[mc->mc_top] = 0;
4760 /** Find the address of the page corresponding to a given page number.
4761 * @param[in] txn the transaction for this access.
4762 * @param[in] pgno the page number for the page to retrieve.
4763 * @param[out] ret address of a pointer where the page's address will be stored.
4764 * @param[out] lvl dirty_list inheritance level of found page. 1=current txn, 0=mapped page.
4765 * @return 0 on success, non-zero on failure.
4768 mdb_page_get(MDB_txn *txn, pgno_t pgno, MDB_page **ret, int *lvl)
4770 MDB_env *env = txn->mt_env;
4774 if (!((txn->mt_flags & MDB_TXN_RDONLY) | (env->me_flags & MDB_WRITEMAP))) {
4778 MDB_ID2L dl = tx2->mt_u.dirty_list;
4780 /* Spilled pages were dirtied in this txn and flushed
4781 * because the dirty list got full. Bring this page
4782 * back in from the map (but don't unspill it here,
4783 * leave that unless page_touch happens again).
4785 if (tx2->mt_spill_pgs) {
4786 MDB_ID pn = pgno << 1;
4787 x = mdb_midl_search(tx2->mt_spill_pgs, pn);
4788 if (x <= tx2->mt_spill_pgs[0] && tx2->mt_spill_pgs[x] == pn) {
4789 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4794 unsigned x = mdb_mid2l_search(dl, pgno);
4795 if (x <= dl[0].mid && dl[x].mid == pgno) {
4801 } while ((tx2 = tx2->mt_parent) != NULL);
4804 if (pgno < txn->mt_next_pgno) {
4806 p = (MDB_page *)(env->me_map + env->me_psize * pgno);
4808 DPRINTF(("page %"Z"u not found", pgno));
4809 txn->mt_flags |= MDB_TXN_ERROR;
4810 return MDB_PAGE_NOTFOUND;
4820 /** Finish #mdb_page_search() / #mdb_page_search_lowest().
4821 * The cursor is at the root page, set up the rest of it.
4824 mdb_page_search_root(MDB_cursor *mc, MDB_val *key, int flags)
4826 MDB_page *mp = mc->mc_pg[mc->mc_top];
4830 while (IS_BRANCH(mp)) {
4834 DPRINTF(("branch page %"Z"u has %u keys", mp->mp_pgno, NUMKEYS(mp)));
4835 mdb_cassert(mc, NUMKEYS(mp) > 1);
4836 DPRINTF(("found index 0 to page %"Z"u", NODEPGNO(NODEPTR(mp, 0))));
4838 if (flags & (MDB_PS_FIRST|MDB_PS_LAST)) {
4840 if (flags & MDB_PS_LAST)
4841 i = NUMKEYS(mp) - 1;
4844 node = mdb_node_search(mc, key, &exact);
4846 i = NUMKEYS(mp) - 1;
4848 i = mc->mc_ki[mc->mc_top];
4850 mdb_cassert(mc, i > 0);
4854 DPRINTF(("following index %u for key [%s]", i, DKEY(key)));
4857 mdb_cassert(mc, i < NUMKEYS(mp));
4858 node = NODEPTR(mp, i);
4860 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4863 mc->mc_ki[mc->mc_top] = i;
4864 if ((rc = mdb_cursor_push(mc, mp)))
4867 if (flags & MDB_PS_MODIFY) {
4868 if ((rc = mdb_page_touch(mc)) != 0)
4870 mp = mc->mc_pg[mc->mc_top];
4875 DPRINTF(("internal error, index points to a %02X page!?",
4877 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
4878 return MDB_CORRUPTED;
4881 DPRINTF(("found leaf page %"Z"u for key [%s]", mp->mp_pgno,
4882 key ? DKEY(key) : "null"));
4883 mc->mc_flags |= C_INITIALIZED;
4884 mc->mc_flags &= ~C_EOF;
4889 /** Search for the lowest key under the current branch page.
4890 * This just bypasses a NUMKEYS check in the current page
4891 * before calling mdb_page_search_root(), because the callers
4892 * are all in situations where the current page is known to
4896 mdb_page_search_lowest(MDB_cursor *mc)
4898 MDB_page *mp = mc->mc_pg[mc->mc_top];
4899 MDB_node *node = NODEPTR(mp, 0);
4902 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(node), &mp, NULL)) != 0)
4905 mc->mc_ki[mc->mc_top] = 0;
4906 if ((rc = mdb_cursor_push(mc, mp)))
4908 return mdb_page_search_root(mc, NULL, MDB_PS_FIRST);
4911 /** Search for the page a given key should be in.
4912 * Push it and its parent pages on the cursor stack.
4913 * @param[in,out] mc the cursor for this operation.
4914 * @param[in] key the key to search for, or NULL for first/last page.
4915 * @param[in] flags If MDB_PS_MODIFY is set, visited pages in the DB
4916 * are touched (updated with new page numbers).
4917 * If MDB_PS_FIRST or MDB_PS_LAST is set, find first or last leaf.
4918 * This is used by #mdb_cursor_first() and #mdb_cursor_last().
4919 * If MDB_PS_ROOTONLY set, just fetch root node, no further lookups.
4920 * @return 0 on success, non-zero on failure.
4923 mdb_page_search(MDB_cursor *mc, MDB_val *key, int flags)
4928 /* Make sure the txn is still viable, then find the root from
4929 * the txn's db table and set it as the root of the cursor's stack.
4931 if (F_ISSET(mc->mc_txn->mt_flags, MDB_TXN_ERROR)) {
4932 DPUTS("transaction has failed, must abort");
4935 /* Make sure we're using an up-to-date root */
4936 if (*mc->mc_dbflag & DB_STALE) {
4938 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, NULL);
4939 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, 0);
4946 MDB_node *leaf = mdb_node_search(&mc2,
4947 &mc->mc_dbx->md_name, &exact);
4949 return MDB_NOTFOUND;
4950 rc = mdb_node_read(mc->mc_txn, leaf, &data);
4953 memcpy(&flags, ((char *) data.mv_data + offsetof(MDB_db, md_flags)),
4955 /* The txn may not know this DBI, or another process may
4956 * have dropped and recreated the DB with other flags.
4958 if ((mc->mc_db->md_flags & PERSISTENT_FLAGS) != flags)
4959 return MDB_INCOMPATIBLE;
4960 memcpy(mc->mc_db, data.mv_data, sizeof(MDB_db));
4962 *mc->mc_dbflag &= ~DB_STALE;
4964 root = mc->mc_db->md_root;
4966 if (root == P_INVALID) { /* Tree is empty. */
4967 DPUTS("tree is empty");
4968 return MDB_NOTFOUND;
4972 mdb_cassert(mc, root > 1);
4973 if (!mc->mc_pg[0] || mc->mc_pg[0]->mp_pgno != root)
4974 if ((rc = mdb_page_get(mc->mc_txn, root, &mc->mc_pg[0], NULL)) != 0)
4980 DPRINTF(("db %d root page %"Z"u has flags 0x%X",
4981 DDBI(mc), root, mc->mc_pg[0]->mp_flags));
4983 if (flags & MDB_PS_MODIFY) {
4984 if ((rc = mdb_page_touch(mc)))
4988 if (flags & MDB_PS_ROOTONLY)
4991 return mdb_page_search_root(mc, key, flags);
4995 mdb_ovpage_free(MDB_cursor *mc, MDB_page *mp)
4997 MDB_txn *txn = mc->mc_txn;
4998 pgno_t pg = mp->mp_pgno;
4999 unsigned x = 0, ovpages = mp->mp_pages;
5000 MDB_env *env = txn->mt_env;
5001 MDB_IDL sl = txn->mt_spill_pgs;
5002 MDB_ID pn = pg << 1;
5005 DPRINTF(("free ov page %"Z"u (%d)", pg, ovpages));
5006 /* If the page is dirty or on the spill list we just acquired it,
5007 * so we should give it back to our current free list, if any.
5008 * Otherwise put it onto the list of pages we freed in this txn.
5010 * Won't create me_pghead: me_pglast must be inited along with it.
5011 * Unsupported in nested txns: They would need to hide the page
5012 * range in ancestor txns' dirty and spilled lists.
5014 if (env->me_pghead &&
5016 ((mp->mp_flags & P_DIRTY) ||
5017 (sl && (x = mdb_midl_search(sl, pn)) <= sl[0] && sl[x] == pn)))
5021 MDB_ID2 *dl, ix, iy;
5022 rc = mdb_midl_need(&env->me_pghead, ovpages);
5025 if (!(mp->mp_flags & P_DIRTY)) {
5026 /* This page is no longer spilled */
5033 /* Remove from dirty list */
5034 dl = txn->mt_u.dirty_list;
5036 for (ix = dl[x]; ix.mptr != mp; ix = iy) {
5042 mdb_cassert(mc, x > 1);
5044 dl[j] = ix; /* Unsorted. OK when MDB_TXN_ERROR. */
5045 txn->mt_flags |= MDB_TXN_ERROR;
5046 return MDB_CORRUPTED;
5049 if (!(env->me_flags & MDB_WRITEMAP))
5050 mdb_dpage_free(env, mp);
5052 /* Insert in me_pghead */
5053 mop = env->me_pghead;
5054 j = mop[0] + ovpages;
5055 for (i = mop[0]; i && mop[i] < pg; i--)
5061 rc = mdb_midl_append_range(&txn->mt_free_pgs, pg, ovpages);
5065 mc->mc_db->md_overflow_pages -= ovpages;
5069 /** Return the data associated with a given node.
5070 * @param[in] txn The transaction for this operation.
5071 * @param[in] leaf The node being read.
5072 * @param[out] data Updated to point to the node's data.
5073 * @return 0 on success, non-zero on failure.
5076 mdb_node_read(MDB_txn *txn, MDB_node *leaf, MDB_val *data)
5078 MDB_page *omp; /* overflow page */
5082 if (!F_ISSET(leaf->mn_flags, F_BIGDATA)) {
5083 data->mv_size = NODEDSZ(leaf);
5084 data->mv_data = NODEDATA(leaf);
5088 /* Read overflow data.
5090 data->mv_size = NODEDSZ(leaf);
5091 memcpy(&pgno, NODEDATA(leaf), sizeof(pgno));
5092 if ((rc = mdb_page_get(txn, pgno, &omp, NULL)) != 0) {
5093 DPRINTF(("read overflow page %"Z"u failed", pgno));
5096 data->mv_data = METADATA(omp);
5102 mdb_get(MDB_txn *txn, MDB_dbi dbi,
5103 MDB_val *key, MDB_val *data)
5110 DPRINTF(("===> get db %u key [%s]", dbi, DKEY(key)));
5112 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
5115 if (txn->mt_flags & MDB_TXN_ERROR)
5118 mdb_cursor_init(&mc, txn, dbi, &mx);
5119 return mdb_cursor_set(&mc, key, data, MDB_SET, &exact);
5122 /** Find a sibling for a page.
5123 * Replaces the page at the top of the cursor's stack with the
5124 * specified sibling, if one exists.
5125 * @param[in] mc The cursor for this operation.
5126 * @param[in] move_right Non-zero if the right sibling is requested,
5127 * otherwise the left sibling.
5128 * @return 0 on success, non-zero on failure.
5131 mdb_cursor_sibling(MDB_cursor *mc, int move_right)
5137 if (mc->mc_snum < 2) {
5138 return MDB_NOTFOUND; /* root has no siblings */
5142 DPRINTF(("parent page is page %"Z"u, index %u",
5143 mc->mc_pg[mc->mc_top]->mp_pgno, mc->mc_ki[mc->mc_top]));
5145 if (move_right ? (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mc->mc_pg[mc->mc_top]))
5146 : (mc->mc_ki[mc->mc_top] == 0)) {
5147 DPRINTF(("no more keys left, moving to %s sibling",
5148 move_right ? "right" : "left"));
5149 if ((rc = mdb_cursor_sibling(mc, move_right)) != MDB_SUCCESS) {
5150 /* undo cursor_pop before returning */
5157 mc->mc_ki[mc->mc_top]++;
5159 mc->mc_ki[mc->mc_top]--;
5160 DPRINTF(("just moving to %s index key %u",
5161 move_right ? "right" : "left", mc->mc_ki[mc->mc_top]));
5163 mdb_cassert(mc, IS_BRANCH(mc->mc_pg[mc->mc_top]));
5165 indx = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5166 if ((rc = mdb_page_get(mc->mc_txn, NODEPGNO(indx), &mp, NULL)) != 0) {
5167 /* mc will be inconsistent if caller does mc_snum++ as above */
5168 mc->mc_flags &= ~(C_INITIALIZED|C_EOF);
5172 mdb_cursor_push(mc, mp);
5174 mc->mc_ki[mc->mc_top] = NUMKEYS(mp)-1;
5179 /** Move the cursor to the next data item. */
5181 mdb_cursor_next(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5187 if (mc->mc_flags & C_EOF) {
5188 return MDB_NOTFOUND;
5191 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5193 mp = mc->mc_pg[mc->mc_top];
5195 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5196 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5197 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5198 if (op == MDB_NEXT || op == MDB_NEXT_DUP) {
5199 rc = mdb_cursor_next(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_NEXT);
5200 if (op != MDB_NEXT || rc != MDB_NOTFOUND) {
5201 if (rc == MDB_SUCCESS)
5202 MDB_GET_KEY(leaf, key);
5207 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5208 if (op == MDB_NEXT_DUP)
5209 return MDB_NOTFOUND;
5213 DPRINTF(("cursor_next: top page is %"Z"u in cursor %p",
5214 mdb_dbg_pgno(mp), (void *) mc));
5215 if (mc->mc_flags & C_DEL)
5218 if (mc->mc_ki[mc->mc_top] + 1u >= NUMKEYS(mp)) {
5219 DPUTS("=====> move to next sibling page");
5220 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS) {
5221 mc->mc_flags |= C_EOF;
5224 mp = mc->mc_pg[mc->mc_top];
5225 DPRINTF(("next page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5227 mc->mc_ki[mc->mc_top]++;
5230 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5231 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5234 key->mv_size = mc->mc_db->md_pad;
5235 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5239 mdb_cassert(mc, IS_LEAF(mp));
5240 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5242 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5243 mdb_xcursor_init1(mc, leaf);
5246 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5249 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5250 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5251 if (rc != MDB_SUCCESS)
5256 MDB_GET_KEY(leaf, key);
5260 /** Move the cursor to the previous data item. */
5262 mdb_cursor_prev(MDB_cursor *mc, MDB_val *key, MDB_val *data, MDB_cursor_op op)
5268 mdb_cassert(mc, mc->mc_flags & C_INITIALIZED);
5270 mp = mc->mc_pg[mc->mc_top];
5272 if (mc->mc_db->md_flags & MDB_DUPSORT) {
5273 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5274 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5275 if (op == MDB_PREV || op == MDB_PREV_DUP) {
5276 rc = mdb_cursor_prev(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_PREV);
5277 if (op != MDB_PREV || rc != MDB_NOTFOUND) {
5278 if (rc == MDB_SUCCESS)
5279 MDB_GET_KEY(leaf, key);
5283 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5284 if (op == MDB_PREV_DUP)
5285 return MDB_NOTFOUND;
5290 DPRINTF(("cursor_prev: top page is %"Z"u in cursor %p",
5291 mdb_dbg_pgno(mp), (void *) mc));
5293 if (mc->mc_ki[mc->mc_top] == 0) {
5294 DPUTS("=====> move to prev sibling page");
5295 if ((rc = mdb_cursor_sibling(mc, 0)) != MDB_SUCCESS) {
5298 mp = mc->mc_pg[mc->mc_top];
5299 mc->mc_ki[mc->mc_top] = NUMKEYS(mp) - 1;
5300 DPRINTF(("prev page is %"Z"u, key index %u", mp->mp_pgno, mc->mc_ki[mc->mc_top]));
5302 mc->mc_ki[mc->mc_top]--;
5304 mc->mc_flags &= ~C_EOF;
5306 DPRINTF(("==> cursor points to page %"Z"u with %u keys, key index %u",
5307 mdb_dbg_pgno(mp), NUMKEYS(mp), mc->mc_ki[mc->mc_top]));
5310 key->mv_size = mc->mc_db->md_pad;
5311 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5315 mdb_cassert(mc, IS_LEAF(mp));
5316 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5318 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5319 mdb_xcursor_init1(mc, leaf);
5322 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5325 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5326 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5327 if (rc != MDB_SUCCESS)
5332 MDB_GET_KEY(leaf, key);
5336 /** Set the cursor on a specific data item. */
5338 mdb_cursor_set(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5339 MDB_cursor_op op, int *exactp)
5343 MDB_node *leaf = NULL;
5346 if (key->mv_size == 0)
5347 return MDB_BAD_VALSIZE;
5350 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5352 /* See if we're already on the right page */
5353 if (mc->mc_flags & C_INITIALIZED) {
5356 mp = mc->mc_pg[mc->mc_top];
5358 mc->mc_ki[mc->mc_top] = 0;
5359 return MDB_NOTFOUND;
5361 if (mp->mp_flags & P_LEAF2) {
5362 nodekey.mv_size = mc->mc_db->md_pad;
5363 nodekey.mv_data = LEAF2KEY(mp, 0, nodekey.mv_size);
5365 leaf = NODEPTR(mp, 0);
5366 MDB_GET_KEY2(leaf, nodekey);
5368 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5370 /* Probably happens rarely, but first node on the page
5371 * was the one we wanted.
5373 mc->mc_ki[mc->mc_top] = 0;
5380 unsigned int nkeys = NUMKEYS(mp);
5382 if (mp->mp_flags & P_LEAF2) {
5383 nodekey.mv_data = LEAF2KEY(mp,
5384 nkeys-1, nodekey.mv_size);
5386 leaf = NODEPTR(mp, nkeys-1);
5387 MDB_GET_KEY2(leaf, nodekey);
5389 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5391 /* last node was the one we wanted */
5392 mc->mc_ki[mc->mc_top] = nkeys-1;
5398 if (mc->mc_ki[mc->mc_top] < NUMKEYS(mp)) {
5399 /* This is definitely the right page, skip search_page */
5400 if (mp->mp_flags & P_LEAF2) {
5401 nodekey.mv_data = LEAF2KEY(mp,
5402 mc->mc_ki[mc->mc_top], nodekey.mv_size);
5404 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5405 MDB_GET_KEY2(leaf, nodekey);
5407 rc = mc->mc_dbx->md_cmp(key, &nodekey);
5409 /* current node was the one we wanted */
5419 /* If any parents have right-sibs, search.
5420 * Otherwise, there's nothing further.
5422 for (i=0; i<mc->mc_top; i++)
5424 NUMKEYS(mc->mc_pg[i])-1)
5426 if (i == mc->mc_top) {
5427 /* There are no other pages */
5428 mc->mc_ki[mc->mc_top] = nkeys;
5429 return MDB_NOTFOUND;
5433 /* There are no other pages */
5434 mc->mc_ki[mc->mc_top] = 0;
5435 if (op == MDB_SET_RANGE && !exactp) {
5439 return MDB_NOTFOUND;
5443 rc = mdb_page_search(mc, key, 0);
5444 if (rc != MDB_SUCCESS)
5447 mp = mc->mc_pg[mc->mc_top];
5448 mdb_cassert(mc, IS_LEAF(mp));
5451 leaf = mdb_node_search(mc, key, exactp);
5452 if (exactp != NULL && !*exactp) {
5453 /* MDB_SET specified and not an exact match. */
5454 return MDB_NOTFOUND;
5458 DPUTS("===> inexact leaf not found, goto sibling");
5459 if ((rc = mdb_cursor_sibling(mc, 1)) != MDB_SUCCESS)
5460 return rc; /* no entries matched */
5461 mp = mc->mc_pg[mc->mc_top];
5462 mdb_cassert(mc, IS_LEAF(mp));
5463 leaf = NODEPTR(mp, 0);
5467 mc->mc_flags |= C_INITIALIZED;
5468 mc->mc_flags &= ~C_EOF;
5471 if (op == MDB_SET_RANGE || op == MDB_SET_KEY) {
5472 key->mv_size = mc->mc_db->md_pad;
5473 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5478 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5479 mdb_xcursor_init1(mc, leaf);
5482 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5483 if (op == MDB_SET || op == MDB_SET_KEY || op == MDB_SET_RANGE) {
5484 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5487 if (op == MDB_GET_BOTH) {
5493 rc = mdb_cursor_set(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_SET_RANGE, ex2p);
5494 if (rc != MDB_SUCCESS)
5497 } else if (op == MDB_GET_BOTH || op == MDB_GET_BOTH_RANGE) {
5499 if ((rc = mdb_node_read(mc->mc_txn, leaf, &d2)) != MDB_SUCCESS)
5501 rc = mc->mc_dbx->md_dcmp(data, &d2);
5503 if (op == MDB_GET_BOTH || rc > 0)
5504 return MDB_NOTFOUND;
5511 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5512 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5517 /* The key already matches in all other cases */
5518 if (op == MDB_SET_RANGE || op == MDB_SET_KEY)
5519 MDB_GET_KEY(leaf, key);
5520 DPRINTF(("==> cursor placed on key [%s]", DKEY(key)));
5525 /** Move the cursor to the first item in the database. */
5527 mdb_cursor_first(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5533 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5535 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5536 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
5537 if (rc != MDB_SUCCESS)
5540 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5542 leaf = NODEPTR(mc->mc_pg[mc->mc_top], 0);
5543 mc->mc_flags |= C_INITIALIZED;
5544 mc->mc_flags &= ~C_EOF;
5546 mc->mc_ki[mc->mc_top] = 0;
5548 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5549 key->mv_size = mc->mc_db->md_pad;
5550 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], 0, key->mv_size);
5555 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5556 mdb_xcursor_init1(mc, leaf);
5557 rc = mdb_cursor_first(&mc->mc_xcursor->mx_cursor, data, NULL);
5561 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5565 MDB_GET_KEY(leaf, key);
5569 /** Move the cursor to the last item in the database. */
5571 mdb_cursor_last(MDB_cursor *mc, MDB_val *key, MDB_val *data)
5577 mc->mc_xcursor->mx_cursor.mc_flags &= ~(C_INITIALIZED|C_EOF);
5579 if (!(mc->mc_flags & C_EOF)) {
5581 if (!(mc->mc_flags & C_INITIALIZED) || mc->mc_top) {
5582 rc = mdb_page_search(mc, NULL, MDB_PS_LAST);
5583 if (rc != MDB_SUCCESS)
5586 mdb_cassert(mc, IS_LEAF(mc->mc_pg[mc->mc_top]));
5589 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]) - 1;
5590 mc->mc_flags |= C_INITIALIZED|C_EOF;
5591 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5593 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5594 key->mv_size = mc->mc_db->md_pad;
5595 key->mv_data = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], key->mv_size);
5600 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5601 mdb_xcursor_init1(mc, leaf);
5602 rc = mdb_cursor_last(&mc->mc_xcursor->mx_cursor, data, NULL);
5606 if ((rc = mdb_node_read(mc->mc_txn, leaf, data)) != MDB_SUCCESS)
5611 MDB_GET_KEY(leaf, key);
5616 mdb_cursor_get(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5621 int (*mfunc)(MDB_cursor *mc, MDB_val *key, MDB_val *data);
5626 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
5630 case MDB_GET_CURRENT:
5631 if (!(mc->mc_flags & C_INITIALIZED)) {
5634 MDB_page *mp = mc->mc_pg[mc->mc_top];
5635 int nkeys = NUMKEYS(mp);
5636 if (!nkeys || mc->mc_ki[mc->mc_top] >= nkeys) {
5637 mc->mc_ki[mc->mc_top] = nkeys;
5643 key->mv_size = mc->mc_db->md_pad;
5644 key->mv_data = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], key->mv_size);
5646 MDB_node *leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
5647 MDB_GET_KEY(leaf, key);
5649 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5650 if (mc->mc_flags & C_DEL)
5651 mdb_xcursor_init1(mc, leaf);
5652 rc = mdb_cursor_get(&mc->mc_xcursor->mx_cursor, data, NULL, MDB_GET_CURRENT);
5654 rc = mdb_node_read(mc->mc_txn, leaf, data);
5661 case MDB_GET_BOTH_RANGE:
5666 if (mc->mc_xcursor == NULL) {
5667 rc = MDB_INCOMPATIBLE;
5677 rc = mdb_cursor_set(mc, key, data, op,
5678 op == MDB_SET_RANGE ? NULL : &exact);
5681 case MDB_GET_MULTIPLE:
5682 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5686 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5687 rc = MDB_INCOMPATIBLE;
5691 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) ||
5692 (mc->mc_xcursor->mx_cursor.mc_flags & C_EOF))
5695 case MDB_NEXT_MULTIPLE:
5700 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
5701 rc = MDB_INCOMPATIBLE;
5704 if (!(mc->mc_flags & C_INITIALIZED))
5705 rc = mdb_cursor_first(mc, key, data);
5707 rc = mdb_cursor_next(mc, key, data, MDB_NEXT_DUP);
5708 if (rc == MDB_SUCCESS) {
5709 if (mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED) {
5712 mx = &mc->mc_xcursor->mx_cursor;
5713 data->mv_size = NUMKEYS(mx->mc_pg[mx->mc_top]) *
5715 data->mv_data = METADATA(mx->mc_pg[mx->mc_top]);
5716 mx->mc_ki[mx->mc_top] = NUMKEYS(mx->mc_pg[mx->mc_top])-1;
5724 case MDB_NEXT_NODUP:
5725 if (!(mc->mc_flags & C_INITIALIZED))
5726 rc = mdb_cursor_first(mc, key, data);
5728 rc = mdb_cursor_next(mc, key, data, op);
5732 case MDB_PREV_NODUP:
5733 if (!(mc->mc_flags & C_INITIALIZED)) {
5734 rc = mdb_cursor_last(mc, key, data);
5737 mc->mc_flags |= C_INITIALIZED;
5738 mc->mc_ki[mc->mc_top]++;
5740 rc = mdb_cursor_prev(mc, key, data, op);
5743 rc = mdb_cursor_first(mc, key, data);
5746 mfunc = mdb_cursor_first;
5748 if (data == NULL || !(mc->mc_flags & C_INITIALIZED)) {
5752 if (mc->mc_xcursor == NULL) {
5753 rc = MDB_INCOMPATIBLE;
5756 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED)) {
5760 rc = mfunc(&mc->mc_xcursor->mx_cursor, data, NULL);
5763 rc = mdb_cursor_last(mc, key, data);
5766 mfunc = mdb_cursor_last;
5769 DPRINTF(("unhandled/unimplemented cursor operation %u", op));
5774 if (mc->mc_flags & C_DEL)
5775 mc->mc_flags ^= C_DEL;
5780 /** Touch all the pages in the cursor stack. Set mc_top.
5781 * Makes sure all the pages are writable, before attempting a write operation.
5782 * @param[in] mc The cursor to operate on.
5785 mdb_cursor_touch(MDB_cursor *mc)
5787 int rc = MDB_SUCCESS;
5789 if (mc->mc_dbi > MAIN_DBI && !(*mc->mc_dbflag & DB_DIRTY)) {
5792 mdb_cursor_init(&mc2, mc->mc_txn, MAIN_DBI, &mcx);
5793 rc = mdb_page_search(&mc2, &mc->mc_dbx->md_name, MDB_PS_MODIFY);
5796 *mc->mc_dbflag |= DB_DIRTY;
5801 rc = mdb_page_touch(mc);
5802 } while (!rc && ++(mc->mc_top) < mc->mc_snum);
5803 mc->mc_top = mc->mc_snum-1;
5808 /** Do not spill pages to disk if txn is getting full, may fail instead */
5809 #define MDB_NOSPILL 0x8000
5812 mdb_cursor_put(MDB_cursor *mc, MDB_val *key, MDB_val *data,
5815 enum { MDB_NO_ROOT = MDB_LAST_ERRCODE+10 }; /* internal code */
5817 MDB_node *leaf = NULL;
5820 MDB_val xdata, *rdata, dkey, olddata;
5822 int do_sub = 0, insert_key, insert_data;
5823 unsigned int mcount = 0, dcount = 0, nospill;
5826 unsigned int nflags;
5829 if (mc == NULL || key == NULL)
5832 env = mc->mc_txn->mt_env;
5834 /* Check this first so counter will always be zero on any
5837 if (flags & MDB_MULTIPLE) {
5838 dcount = data[1].mv_size;
5839 data[1].mv_size = 0;
5840 if (!F_ISSET(mc->mc_db->md_flags, MDB_DUPFIXED))
5841 return MDB_INCOMPATIBLE;
5844 nospill = flags & MDB_NOSPILL;
5845 flags &= ~MDB_NOSPILL;
5847 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
5848 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
5850 if (key->mv_size-1 >= ENV_MAXKEY(env))
5851 return MDB_BAD_VALSIZE;
5853 #if SIZE_MAX > MAXDATASIZE
5854 if (data->mv_size > ((mc->mc_db->md_flags & MDB_DUPSORT) ? ENV_MAXKEY(env) : MAXDATASIZE))
5855 return MDB_BAD_VALSIZE;
5857 if ((mc->mc_db->md_flags & MDB_DUPSORT) && data->mv_size > ENV_MAXKEY(env))
5858 return MDB_BAD_VALSIZE;
5861 DPRINTF(("==> put db %d key [%s], size %"Z"u, data size %"Z"u",
5862 DDBI(mc), DKEY(key), key ? key->mv_size : 0, data->mv_size));
5866 if (flags == MDB_CURRENT) {
5867 if (!(mc->mc_flags & C_INITIALIZED))
5870 } else if (mc->mc_db->md_root == P_INVALID) {
5871 /* new database, cursor has nothing to point to */
5874 mc->mc_flags &= ~C_INITIALIZED;
5879 if (flags & MDB_APPEND) {
5881 rc = mdb_cursor_last(mc, &k2, &d2);
5883 rc = mc->mc_dbx->md_cmp(key, &k2);
5886 mc->mc_ki[mc->mc_top]++;
5888 /* new key is <= last key */
5893 rc = mdb_cursor_set(mc, key, &d2, MDB_SET, &exact);
5895 if ((flags & MDB_NOOVERWRITE) && rc == 0) {
5896 DPRINTF(("duplicate key [%s]", DKEY(key)));
5898 return MDB_KEYEXIST;
5900 if (rc && rc != MDB_NOTFOUND)
5904 if (mc->mc_flags & C_DEL)
5905 mc->mc_flags ^= C_DEL;
5907 /* Cursor is positioned, check for room in the dirty list */
5909 if (flags & MDB_MULTIPLE) {
5911 xdata.mv_size = data->mv_size * dcount;
5915 if ((rc2 = mdb_page_spill(mc, key, rdata)))
5919 if (rc == MDB_NO_ROOT) {
5921 /* new database, write a root leaf page */
5922 DPUTS("allocating new root leaf page");
5923 if ((rc2 = mdb_page_new(mc, P_LEAF, 1, &np))) {
5926 mdb_cursor_push(mc, np);
5927 mc->mc_db->md_root = np->mp_pgno;
5928 mc->mc_db->md_depth++;
5929 *mc->mc_dbflag |= DB_DIRTY;
5930 if ((mc->mc_db->md_flags & (MDB_DUPSORT|MDB_DUPFIXED))
5932 np->mp_flags |= P_LEAF2;
5933 mc->mc_flags |= C_INITIALIZED;
5935 /* make sure all cursor pages are writable */
5936 rc2 = mdb_cursor_touch(mc);
5941 insert_key = insert_data = rc;
5943 /* The key does not exist */
5944 DPRINTF(("inserting key at index %i", mc->mc_ki[mc->mc_top]));
5945 if ((mc->mc_db->md_flags & MDB_DUPSORT) &&
5946 LEAFSIZE(key, data) > env->me_nodemax)
5948 /* Too big for a node, insert in sub-DB. Set up an empty
5949 * "old sub-page" for prep_subDB to expand to a full page.
5951 fp_flags = P_LEAF|P_DIRTY;
5953 fp->mp_pad = data->mv_size; /* used if MDB_DUPFIXED */
5954 fp->mp_lower = fp->mp_upper = (PAGEHDRSZ-PAGEBASE);
5955 olddata.mv_size = PAGEHDRSZ;
5959 /* there's only a key anyway, so this is a no-op */
5960 if (IS_LEAF2(mc->mc_pg[mc->mc_top])) {
5962 unsigned int ksize = mc->mc_db->md_pad;
5963 if (key->mv_size != ksize)
5964 return MDB_BAD_VALSIZE;
5965 ptr = LEAF2KEY(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top], ksize);
5966 memcpy(ptr, key->mv_data, ksize);
5971 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
5972 olddata.mv_size = NODEDSZ(leaf);
5973 olddata.mv_data = NODEDATA(leaf);
5976 if (F_ISSET(mc->mc_db->md_flags, MDB_DUPSORT)) {
5977 /* Prepare (sub-)page/sub-DB to accept the new item,
5978 * if needed. fp: old sub-page or a header faking
5979 * it. mp: new (sub-)page. offset: growth in page
5980 * size. xdata: node data with new page or DB.
5982 unsigned i, offset = 0;
5983 mp = fp = xdata.mv_data = env->me_pbuf;
5984 mp->mp_pgno = mc->mc_pg[mc->mc_top]->mp_pgno;
5986 /* Was a single item before, must convert now */
5987 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
5988 /* Just overwrite the current item */
5989 if (flags == MDB_CURRENT)
5992 #if UINT_MAX < SIZE_MAX
5993 if (mc->mc_dbx->md_dcmp == mdb_cmp_int && olddata.mv_size == sizeof(size_t))
5994 mc->mc_dbx->md_dcmp = mdb_cmp_clong;
5996 /* does data match? */
5997 if (!mc->mc_dbx->md_dcmp(data, &olddata)) {
5998 if (flags & MDB_NODUPDATA)
5999 return MDB_KEYEXIST;
6004 /* Back up original data item */
6005 dkey.mv_size = olddata.mv_size;
6006 dkey.mv_data = memcpy(fp+1, olddata.mv_data, olddata.mv_size);
6008 /* Make sub-page header for the dup items, with dummy body */
6009 fp->mp_flags = P_LEAF|P_DIRTY|P_SUBP;
6010 fp->mp_lower = (PAGEHDRSZ-PAGEBASE);
6011 xdata.mv_size = PAGEHDRSZ + dkey.mv_size + data->mv_size;
6012 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6013 fp->mp_flags |= P_LEAF2;
6014 fp->mp_pad = data->mv_size;
6015 xdata.mv_size += 2 * data->mv_size; /* leave space for 2 more */
6017 xdata.mv_size += 2 * (sizeof(indx_t) + NODESIZE) +
6018 (dkey.mv_size & 1) + (data->mv_size & 1);
6020 fp->mp_upper = xdata.mv_size - PAGEBASE;
6021 olddata.mv_size = xdata.mv_size; /* pretend olddata is fp */
6022 } else if (leaf->mn_flags & F_SUBDATA) {
6023 /* Data is on sub-DB, just store it */
6024 flags |= F_DUPDATA|F_SUBDATA;
6027 /* Data is on sub-page */
6028 fp = olddata.mv_data;
6031 if (!(mc->mc_db->md_flags & MDB_DUPFIXED)) {
6032 offset = EVEN(NODESIZE + sizeof(indx_t) +
6036 offset = fp->mp_pad;
6037 if (SIZELEFT(fp) < offset) {
6038 offset *= 4; /* space for 4 more */
6041 /* FALLTHRU: Big enough MDB_DUPFIXED sub-page */
6043 fp->mp_flags |= P_DIRTY;
6044 COPY_PGNO(fp->mp_pgno, mp->mp_pgno);
6045 mc->mc_xcursor->mx_cursor.mc_pg[0] = fp;
6049 xdata.mv_size = olddata.mv_size + offset;
6052 fp_flags = fp->mp_flags;
6053 if (NODESIZE + NODEKSZ(leaf) + xdata.mv_size > env->me_nodemax) {
6054 /* Too big for a sub-page, convert to sub-DB */
6055 fp_flags &= ~P_SUBP;
6057 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6058 fp_flags |= P_LEAF2;
6059 dummy.md_pad = fp->mp_pad;
6060 dummy.md_flags = MDB_DUPFIXED;
6061 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6062 dummy.md_flags |= MDB_INTEGERKEY;
6068 dummy.md_branch_pages = 0;
6069 dummy.md_leaf_pages = 1;
6070 dummy.md_overflow_pages = 0;
6071 dummy.md_entries = NUMKEYS(fp);
6072 xdata.mv_size = sizeof(MDB_db);
6073 xdata.mv_data = &dummy;
6074 if ((rc = mdb_page_alloc(mc, 1, &mp)))
6076 offset = env->me_psize - olddata.mv_size;
6077 flags |= F_DUPDATA|F_SUBDATA;
6078 dummy.md_root = mp->mp_pgno;
6081 mp->mp_flags = fp_flags | P_DIRTY;
6082 mp->mp_pad = fp->mp_pad;
6083 mp->mp_lower = fp->mp_lower;
6084 mp->mp_upper = fp->mp_upper + offset;
6085 if (fp_flags & P_LEAF2) {
6086 memcpy(METADATA(mp), METADATA(fp), NUMKEYS(fp) * fp->mp_pad);
6088 memcpy((char *)mp + mp->mp_upper + PAGEBASE, (char *)fp + fp->mp_upper + PAGEBASE,
6089 olddata.mv_size - fp->mp_upper - PAGEBASE);
6090 for (i=0; i<NUMKEYS(fp); i++)
6091 mp->mp_ptrs[i] = fp->mp_ptrs[i] + offset;
6099 mdb_node_del(mc, 0);
6103 /* overflow page overwrites need special handling */
6104 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6107 int level, ovpages, dpages = OVPAGES(data->mv_size, env->me_psize);
6109 memcpy(&pg, olddata.mv_data, sizeof(pg));
6110 if ((rc2 = mdb_page_get(mc->mc_txn, pg, &omp, &level)) != 0)
6112 ovpages = omp->mp_pages;
6114 /* Is the ov page large enough? */
6115 if (ovpages >= dpages) {
6116 if (!(omp->mp_flags & P_DIRTY) &&
6117 (level || (env->me_flags & MDB_WRITEMAP)))
6119 rc = mdb_page_unspill(mc->mc_txn, omp, &omp);
6122 level = 0; /* dirty in this txn or clean */
6125 if (omp->mp_flags & P_DIRTY) {
6126 /* yes, overwrite it. Note in this case we don't
6127 * bother to try shrinking the page if the new data
6128 * is smaller than the overflow threshold.
6131 /* It is writable only in a parent txn */
6132 size_t sz = (size_t) env->me_psize * ovpages, off;
6133 MDB_page *np = mdb_page_malloc(mc->mc_txn, ovpages);
6139 rc2 = mdb_mid2l_insert(mc->mc_txn->mt_u.dirty_list, &id2);
6140 mdb_cassert(mc, rc2 == 0);
6141 if (!(flags & MDB_RESERVE)) {
6142 /* Copy end of page, adjusting alignment so
6143 * compiler may copy words instead of bytes.
6145 off = (PAGEHDRSZ + data->mv_size) & -sizeof(size_t);
6146 memcpy((size_t *)((char *)np + off),
6147 (size_t *)((char *)omp + off), sz - off);
6150 memcpy(np, omp, sz); /* Copy beginning of page */
6153 SETDSZ(leaf, data->mv_size);
6154 if (F_ISSET(flags, MDB_RESERVE))
6155 data->mv_data = METADATA(omp);
6157 memcpy(METADATA(omp), data->mv_data, data->mv_size);
6161 if ((rc2 = mdb_ovpage_free(mc, omp)) != MDB_SUCCESS)
6163 } else if (data->mv_size == olddata.mv_size) {
6164 /* same size, just replace it. Note that we could
6165 * also reuse this node if the new data is smaller,
6166 * but instead we opt to shrink the node in that case.
6168 if (F_ISSET(flags, MDB_RESERVE))
6169 data->mv_data = olddata.mv_data;
6170 else if (!(mc->mc_flags & C_SUB))
6171 memcpy(olddata.mv_data, data->mv_data, data->mv_size);
6173 memcpy(NODEKEY(leaf), key->mv_data, key->mv_size);
6176 mdb_node_del(mc, 0);
6182 nflags = flags & NODE_ADD_FLAGS;
6183 nsize = IS_LEAF2(mc->mc_pg[mc->mc_top]) ? key->mv_size : mdb_leaf_size(env, key, rdata);
6184 if (SIZELEFT(mc->mc_pg[mc->mc_top]) < nsize) {
6185 if (( flags & (F_DUPDATA|F_SUBDATA)) == F_DUPDATA )
6186 nflags &= ~MDB_APPEND; /* sub-page may need room to grow */
6188 nflags |= MDB_SPLIT_REPLACE;
6189 rc = mdb_page_split(mc, key, rdata, P_INVALID, nflags);
6191 /* There is room already in this leaf page. */
6192 rc = mdb_node_add(mc, mc->mc_ki[mc->mc_top], key, rdata, 0, nflags);
6193 if (rc == 0 && insert_key) {
6194 /* Adjust other cursors pointing to mp */
6195 MDB_cursor *m2, *m3;
6196 MDB_dbi dbi = mc->mc_dbi;
6197 unsigned i = mc->mc_top;
6198 MDB_page *mp = mc->mc_pg[i];
6200 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
6201 if (mc->mc_flags & C_SUB)
6202 m3 = &m2->mc_xcursor->mx_cursor;
6205 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
6206 if (m3->mc_pg[i] == mp && m3->mc_ki[i] >= mc->mc_ki[i]) {
6213 if (rc == MDB_SUCCESS) {
6214 /* Now store the actual data in the child DB. Note that we're
6215 * storing the user data in the keys field, so there are strict
6216 * size limits on dupdata. The actual data fields of the child
6217 * DB are all zero size.
6225 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6226 if (flags & MDB_CURRENT) {
6227 xflags = MDB_CURRENT|MDB_NOSPILL;
6229 mdb_xcursor_init1(mc, leaf);
6230 xflags = (flags & MDB_NODUPDATA) ?
6231 MDB_NOOVERWRITE|MDB_NOSPILL : MDB_NOSPILL;
6233 /* converted, write the original data first */
6235 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, &dkey, &xdata, xflags);
6239 /* Adjust other cursors pointing to mp */
6241 unsigned i = mc->mc_top;
6242 MDB_page *mp = mc->mc_pg[i];
6244 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6245 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6246 if (!(m2->mc_flags & C_INITIALIZED)) continue;
6247 if (m2->mc_pg[i] == mp && m2->mc_ki[i] == mc->mc_ki[i]) {
6248 mdb_xcursor_init1(m2, leaf);
6252 /* we've done our job */
6255 ecount = mc->mc_xcursor->mx_db.md_entries;
6256 if (flags & MDB_APPENDDUP)
6257 xflags |= MDB_APPEND;
6258 rc = mdb_cursor_put(&mc->mc_xcursor->mx_cursor, data, &xdata, xflags);
6259 if (flags & F_SUBDATA) {
6260 void *db = NODEDATA(leaf);
6261 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6263 insert_data = mc->mc_xcursor->mx_db.md_entries - ecount;
6265 /* Increment count unless we just replaced an existing item. */
6267 mc->mc_db->md_entries++;
6269 /* Invalidate txn if we created an empty sub-DB */
6272 /* If we succeeded and the key didn't exist before,
6273 * make sure the cursor is marked valid.
6275 mc->mc_flags |= C_INITIALIZED;
6277 if (flags & MDB_MULTIPLE) {
6280 /* let caller know how many succeeded, if any */
6281 data[1].mv_size = mcount;
6282 if (mcount < dcount) {
6283 data[0].mv_data = (char *)data[0].mv_data + data[0].mv_size;
6284 insert_key = insert_data = 0;
6291 if (rc == MDB_KEYEXIST) /* should not happen, we deleted that item */
6294 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6299 mdb_cursor_del(MDB_cursor *mc, unsigned int flags)
6305 if (mc->mc_txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
6306 return (mc->mc_txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
6308 if (!(mc->mc_flags & C_INITIALIZED))
6311 if (mc->mc_ki[mc->mc_top] >= NUMKEYS(mc->mc_pg[mc->mc_top]))
6312 return MDB_NOTFOUND;
6314 if (!(flags & MDB_NOSPILL) && (rc = mdb_page_spill(mc, NULL, NULL)))
6317 rc = mdb_cursor_touch(mc);
6321 mp = mc->mc_pg[mc->mc_top];
6324 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6326 if (F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6327 if (flags & MDB_NODUPDATA) {
6328 /* mdb_cursor_del0() will subtract the final entry */
6329 mc->mc_db->md_entries -= mc->mc_xcursor->mx_db.md_entries - 1;
6331 if (!F_ISSET(leaf->mn_flags, F_SUBDATA)) {
6332 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6334 rc = mdb_cursor_del(&mc->mc_xcursor->mx_cursor, MDB_NOSPILL);
6337 /* If sub-DB still has entries, we're done */
6338 if (mc->mc_xcursor->mx_db.md_entries) {
6339 if (leaf->mn_flags & F_SUBDATA) {
6340 /* update subDB info */
6341 void *db = NODEDATA(leaf);
6342 memcpy(db, &mc->mc_xcursor->mx_db, sizeof(MDB_db));
6345 /* shrink fake page */
6346 mdb_node_shrink(mp, mc->mc_ki[mc->mc_top]);
6347 leaf = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
6348 mc->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6349 /* fix other sub-DB cursors pointed at this fake page */
6350 for (m2 = mc->mc_txn->mt_cursors[mc->mc_dbi]; m2; m2=m2->mc_next) {
6351 if (m2 == mc || m2->mc_snum < mc->mc_snum) continue;
6352 if (m2->mc_pg[mc->mc_top] == mp &&
6353 m2->mc_ki[mc->mc_top] == mc->mc_ki[mc->mc_top])
6354 m2->mc_xcursor->mx_cursor.mc_pg[0] = NODEDATA(leaf);
6357 mc->mc_db->md_entries--;
6358 mc->mc_flags |= C_DEL;
6361 /* otherwise fall thru and delete the sub-DB */
6364 if (leaf->mn_flags & F_SUBDATA) {
6365 /* add all the child DB's pages to the free list */
6366 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
6372 /* add overflow pages to free list */
6373 if (F_ISSET(leaf->mn_flags, F_BIGDATA)) {
6377 memcpy(&pg, NODEDATA(leaf), sizeof(pg));
6378 if ((rc = mdb_page_get(mc->mc_txn, pg, &omp, NULL)) ||
6379 (rc = mdb_ovpage_free(mc, omp)))
6384 return mdb_cursor_del0(mc);
6387 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6391 /** Allocate and initialize new pages for a database.
6392 * @param[in] mc a cursor on the database being added to.
6393 * @param[in] flags flags defining what type of page is being allocated.
6394 * @param[in] num the number of pages to allocate. This is usually 1,
6395 * unless allocating overflow pages for a large record.
6396 * @param[out] mp Address of a page, or NULL on failure.
6397 * @return 0 on success, non-zero on failure.
6400 mdb_page_new(MDB_cursor *mc, uint32_t flags, int num, MDB_page **mp)
6405 if ((rc = mdb_page_alloc(mc, num, &np)))
6407 DPRINTF(("allocated new mpage %"Z"u, page size %u",
6408 np->mp_pgno, mc->mc_txn->mt_env->me_psize));
6409 np->mp_flags = flags | P_DIRTY;
6410 np->mp_lower = (PAGEHDRSZ-PAGEBASE);
6411 np->mp_upper = mc->mc_txn->mt_env->me_psize - PAGEBASE;
6414 mc->mc_db->md_branch_pages++;
6415 else if (IS_LEAF(np))
6416 mc->mc_db->md_leaf_pages++;
6417 else if (IS_OVERFLOW(np)) {
6418 mc->mc_db->md_overflow_pages += num;
6426 /** Calculate the size of a leaf node.
6427 * The size depends on the environment's page size; if a data item
6428 * is too large it will be put onto an overflow page and the node
6429 * size will only include the key and not the data. Sizes are always
6430 * rounded up to an even number of bytes, to guarantee 2-byte alignment
6431 * of the #MDB_node headers.
6432 * @param[in] env The environment handle.
6433 * @param[in] key The key for the node.
6434 * @param[in] data The data for the node.
6435 * @return The number of bytes needed to store the node.
6438 mdb_leaf_size(MDB_env *env, MDB_val *key, MDB_val *data)
6442 sz = LEAFSIZE(key, data);
6443 if (sz > env->me_nodemax) {
6444 /* put on overflow page */
6445 sz -= data->mv_size - sizeof(pgno_t);
6448 return EVEN(sz + sizeof(indx_t));
6451 /** Calculate the size of a branch node.
6452 * The size should depend on the environment's page size but since
6453 * we currently don't support spilling large keys onto overflow
6454 * pages, it's simply the size of the #MDB_node header plus the
6455 * size of the key. Sizes are always rounded up to an even number
6456 * of bytes, to guarantee 2-byte alignment of the #MDB_node headers.
6457 * @param[in] env The environment handle.
6458 * @param[in] key The key for the node.
6459 * @return The number of bytes needed to store the node.
6462 mdb_branch_size(MDB_env *env, MDB_val *key)
6467 if (sz > env->me_nodemax) {
6468 /* put on overflow page */
6469 /* not implemented */
6470 /* sz -= key->size - sizeof(pgno_t); */
6473 return sz + sizeof(indx_t);
6476 /** Add a node to the page pointed to by the cursor.
6477 * @param[in] mc The cursor for this operation.
6478 * @param[in] indx The index on the page where the new node should be added.
6479 * @param[in] key The key for the new node.
6480 * @param[in] data The data for the new node, if any.
6481 * @param[in] pgno The page number, if adding a branch node.
6482 * @param[in] flags Flags for the node.
6483 * @return 0 on success, non-zero on failure. Possible errors are:
6485 * <li>ENOMEM - failed to allocate overflow pages for the node.
6486 * <li>MDB_PAGE_FULL - there is insufficient room in the page. This error
6487 * should never happen since all callers already calculate the
6488 * page's free space before calling this function.
6492 mdb_node_add(MDB_cursor *mc, indx_t indx,
6493 MDB_val *key, MDB_val *data, pgno_t pgno, unsigned int flags)
6496 size_t node_size = NODESIZE;
6500 MDB_page *mp = mc->mc_pg[mc->mc_top];
6501 MDB_page *ofp = NULL; /* overflow page */
6504 mdb_cassert(mc, mp->mp_upper >= mp->mp_lower);
6506 DPRINTF(("add to %s %spage %"Z"u index %i, data size %"Z"u key size %"Z"u [%s]",
6507 IS_LEAF(mp) ? "leaf" : "branch",
6508 IS_SUBP(mp) ? "sub-" : "",
6509 mdb_dbg_pgno(mp), indx, data ? data->mv_size : 0,
6510 key ? key->mv_size : 0, key ? DKEY(key) : "null"));
6513 /* Move higher keys up one slot. */
6514 int ksize = mc->mc_db->md_pad, dif;
6515 char *ptr = LEAF2KEY(mp, indx, ksize);
6516 dif = NUMKEYS(mp) - indx;
6518 memmove(ptr+ksize, ptr, dif*ksize);
6519 /* insert new key */
6520 memcpy(ptr, key->mv_data, ksize);
6522 /* Just using these for counting */
6523 mp->mp_lower += sizeof(indx_t);
6524 mp->mp_upper -= ksize - sizeof(indx_t);
6528 room = (ssize_t)SIZELEFT(mp) - (ssize_t)sizeof(indx_t);
6530 node_size += key->mv_size;
6532 mdb_cassert(mc, data);
6533 if (F_ISSET(flags, F_BIGDATA)) {
6534 /* Data already on overflow page. */
6535 node_size += sizeof(pgno_t);
6536 } else if (node_size + data->mv_size > mc->mc_txn->mt_env->me_nodemax) {
6537 int ovpages = OVPAGES(data->mv_size, mc->mc_txn->mt_env->me_psize);
6539 /* Put data on overflow page. */
6540 DPRINTF(("data size is %"Z"u, node would be %"Z"u, put data on overflow page",
6541 data->mv_size, node_size+data->mv_size));
6542 node_size = EVEN(node_size + sizeof(pgno_t));
6543 if ((ssize_t)node_size > room)
6545 if ((rc = mdb_page_new(mc, P_OVERFLOW, ovpages, &ofp)))
6547 DPRINTF(("allocated overflow page %"Z"u", ofp->mp_pgno));
6551 node_size += data->mv_size;
6554 node_size = EVEN(node_size);
6555 if ((ssize_t)node_size > room)
6559 /* Move higher pointers up one slot. */
6560 for (i = NUMKEYS(mp); i > indx; i--)
6561 mp->mp_ptrs[i] = mp->mp_ptrs[i - 1];
6563 /* Adjust free space offsets. */
6564 ofs = mp->mp_upper - node_size;
6565 mdb_cassert(mc, ofs >= mp->mp_lower + sizeof(indx_t));
6566 mp->mp_ptrs[indx] = ofs;
6568 mp->mp_lower += sizeof(indx_t);
6570 /* Write the node data. */
6571 node = NODEPTR(mp, indx);
6572 node->mn_ksize = (key == NULL) ? 0 : key->mv_size;
6573 node->mn_flags = flags;
6575 SETDSZ(node,data->mv_size);
6580 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
6583 mdb_cassert(mc, key);
6585 if (F_ISSET(flags, F_BIGDATA))
6586 memcpy(node->mn_data + key->mv_size, data->mv_data,
6588 else if (F_ISSET(flags, MDB_RESERVE))
6589 data->mv_data = node->mn_data + key->mv_size;
6591 memcpy(node->mn_data + key->mv_size, data->mv_data,
6594 memcpy(node->mn_data + key->mv_size, &ofp->mp_pgno,
6596 if (F_ISSET(flags, MDB_RESERVE))
6597 data->mv_data = METADATA(ofp);
6599 memcpy(METADATA(ofp), data->mv_data, data->mv_size);
6606 DPRINTF(("not enough room in page %"Z"u, got %u ptrs",
6607 mdb_dbg_pgno(mp), NUMKEYS(mp)));
6608 DPRINTF(("upper-lower = %u - %u = %"Z"d", mp->mp_upper,mp->mp_lower,room));
6609 DPRINTF(("node size = %"Z"u", node_size));
6610 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
6611 return MDB_PAGE_FULL;
6614 /** Delete the specified node from a page.
6615 * @param[in] mc Cursor pointing to the node to delete.
6616 * @param[in] ksize The size of a node. Only used if the page is
6617 * part of a #MDB_DUPFIXED database.
6620 mdb_node_del(MDB_cursor *mc, int ksize)
6622 MDB_page *mp = mc->mc_pg[mc->mc_top];
6623 indx_t indx = mc->mc_ki[mc->mc_top];
6625 indx_t i, j, numkeys, ptr;
6629 DPRINTF(("delete node %u on %s page %"Z"u", indx,
6630 IS_LEAF(mp) ? "leaf" : "branch", mdb_dbg_pgno(mp)));
6631 numkeys = NUMKEYS(mp);
6632 mdb_cassert(mc, indx < numkeys);
6635 int x = numkeys - 1 - indx;
6636 base = LEAF2KEY(mp, indx, ksize);
6638 memmove(base, base + ksize, x * ksize);
6639 mp->mp_lower -= sizeof(indx_t);
6640 mp->mp_upper += ksize - sizeof(indx_t);
6644 node = NODEPTR(mp, indx);
6645 sz = NODESIZE + node->mn_ksize;
6647 if (F_ISSET(node->mn_flags, F_BIGDATA))
6648 sz += sizeof(pgno_t);
6650 sz += NODEDSZ(node);
6654 ptr = mp->mp_ptrs[indx];
6655 for (i = j = 0; i < numkeys; i++) {
6657 mp->mp_ptrs[j] = mp->mp_ptrs[i];
6658 if (mp->mp_ptrs[i] < ptr)
6659 mp->mp_ptrs[j] += sz;
6664 base = (char *)mp + mp->mp_upper + PAGEBASE;
6665 memmove(base + sz, base, ptr - mp->mp_upper);
6667 mp->mp_lower -= sizeof(indx_t);
6671 /** Compact the main page after deleting a node on a subpage.
6672 * @param[in] mp The main page to operate on.
6673 * @param[in] indx The index of the subpage on the main page.
6676 mdb_node_shrink(MDB_page *mp, indx_t indx)
6682 indx_t i, numkeys, ptr;
6684 node = NODEPTR(mp, indx);
6685 sp = (MDB_page *)NODEDATA(node);
6686 delta = SIZELEFT(sp);
6687 xp = (MDB_page *)((char *)sp + delta);
6689 /* shift subpage upward */
6691 nsize = NUMKEYS(sp) * sp->mp_pad;
6693 return; /* do not make the node uneven-sized */
6694 memmove(METADATA(xp), METADATA(sp), nsize);
6697 numkeys = NUMKEYS(sp);
6698 for (i=numkeys-1; i>=0; i--)
6699 xp->mp_ptrs[i] = sp->mp_ptrs[i] - delta;
6701 xp->mp_upper = sp->mp_lower;
6702 xp->mp_lower = sp->mp_lower;
6703 xp->mp_flags = sp->mp_flags;
6704 xp->mp_pad = sp->mp_pad;
6705 COPY_PGNO(xp->mp_pgno, mp->mp_pgno);
6707 nsize = NODEDSZ(node) - delta;
6708 SETDSZ(node, nsize);
6710 /* shift lower nodes upward */
6711 ptr = mp->mp_ptrs[indx];
6712 numkeys = NUMKEYS(mp);
6713 for (i = 0; i < numkeys; i++) {
6714 if (mp->mp_ptrs[i] <= ptr)
6715 mp->mp_ptrs[i] += delta;
6718 base = (char *)mp + mp->mp_upper + PAGEBASE;
6719 memmove(base + delta, base, ptr - mp->mp_upper + NODESIZE + NODEKSZ(node));
6720 mp->mp_upper += delta;
6723 /** Initial setup of a sorted-dups cursor.
6724 * Sorted duplicates are implemented as a sub-database for the given key.
6725 * The duplicate data items are actually keys of the sub-database.
6726 * Operations on the duplicate data items are performed using a sub-cursor
6727 * initialized when the sub-database is first accessed. This function does
6728 * the preliminary setup of the sub-cursor, filling in the fields that
6729 * depend only on the parent DB.
6730 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6733 mdb_xcursor_init0(MDB_cursor *mc)
6735 MDB_xcursor *mx = mc->mc_xcursor;
6737 mx->mx_cursor.mc_xcursor = NULL;
6738 mx->mx_cursor.mc_txn = mc->mc_txn;
6739 mx->mx_cursor.mc_db = &mx->mx_db;
6740 mx->mx_cursor.mc_dbx = &mx->mx_dbx;
6741 mx->mx_cursor.mc_dbi = mc->mc_dbi;
6742 mx->mx_cursor.mc_dbflag = &mx->mx_dbflag;
6743 mx->mx_cursor.mc_snum = 0;
6744 mx->mx_cursor.mc_top = 0;
6745 mx->mx_cursor.mc_flags = C_SUB;
6746 mx->mx_dbx.md_name.mv_size = 0;
6747 mx->mx_dbx.md_name.mv_data = NULL;
6748 mx->mx_dbx.md_cmp = mc->mc_dbx->md_dcmp;
6749 mx->mx_dbx.md_dcmp = NULL;
6750 mx->mx_dbx.md_rel = mc->mc_dbx->md_rel;
6753 /** Final setup of a sorted-dups cursor.
6754 * Sets up the fields that depend on the data from the main cursor.
6755 * @param[in] mc The main cursor whose sorted-dups cursor is to be initialized.
6756 * @param[in] node The data containing the #MDB_db record for the
6757 * sorted-dup database.
6760 mdb_xcursor_init1(MDB_cursor *mc, MDB_node *node)
6762 MDB_xcursor *mx = mc->mc_xcursor;
6764 if (node->mn_flags & F_SUBDATA) {
6765 memcpy(&mx->mx_db, NODEDATA(node), sizeof(MDB_db));
6766 mx->mx_cursor.mc_pg[0] = 0;
6767 mx->mx_cursor.mc_snum = 0;
6768 mx->mx_cursor.mc_top = 0;
6769 mx->mx_cursor.mc_flags = C_SUB;
6771 MDB_page *fp = NODEDATA(node);
6772 mx->mx_db.md_pad = mc->mc_pg[mc->mc_top]->mp_pad;
6773 mx->mx_db.md_flags = 0;
6774 mx->mx_db.md_depth = 1;
6775 mx->mx_db.md_branch_pages = 0;
6776 mx->mx_db.md_leaf_pages = 1;
6777 mx->mx_db.md_overflow_pages = 0;
6778 mx->mx_db.md_entries = NUMKEYS(fp);
6779 COPY_PGNO(mx->mx_db.md_root, fp->mp_pgno);
6780 mx->mx_cursor.mc_snum = 1;
6781 mx->mx_cursor.mc_top = 0;
6782 mx->mx_cursor.mc_flags = C_INITIALIZED|C_SUB;
6783 mx->mx_cursor.mc_pg[0] = fp;
6784 mx->mx_cursor.mc_ki[0] = 0;
6785 if (mc->mc_db->md_flags & MDB_DUPFIXED) {
6786 mx->mx_db.md_flags = MDB_DUPFIXED;
6787 mx->mx_db.md_pad = fp->mp_pad;
6788 if (mc->mc_db->md_flags & MDB_INTEGERDUP)
6789 mx->mx_db.md_flags |= MDB_INTEGERKEY;
6792 DPRINTF(("Sub-db -%u root page %"Z"u", mx->mx_cursor.mc_dbi,
6793 mx->mx_db.md_root));
6794 mx->mx_dbflag = DB_VALID|DB_DIRTY; /* DB_DIRTY guides mdb_cursor_touch */
6795 #if UINT_MAX < SIZE_MAX
6796 if (mx->mx_dbx.md_cmp == mdb_cmp_int && mx->mx_db.md_pad == sizeof(size_t))
6797 mx->mx_dbx.md_cmp = mdb_cmp_clong;
6801 /** Initialize a cursor for a given transaction and database. */
6803 mdb_cursor_init(MDB_cursor *mc, MDB_txn *txn, MDB_dbi dbi, MDB_xcursor *mx)
6806 mc->mc_backup = NULL;
6809 mc->mc_db = &txn->mt_dbs[dbi];
6810 mc->mc_dbx = &txn->mt_dbxs[dbi];
6811 mc->mc_dbflag = &txn->mt_dbflags[dbi];
6816 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT) {
6817 mdb_tassert(txn, mx != NULL);
6818 mc->mc_xcursor = mx;
6819 mdb_xcursor_init0(mc);
6821 mc->mc_xcursor = NULL;
6823 if (*mc->mc_dbflag & DB_STALE) {
6824 mdb_page_search(mc, NULL, MDB_PS_ROOTONLY);
6829 mdb_cursor_open(MDB_txn *txn, MDB_dbi dbi, MDB_cursor **ret)
6832 size_t size = sizeof(MDB_cursor);
6834 if (!ret || !TXN_DBI_EXIST(txn, dbi))
6837 if (txn->mt_flags & MDB_TXN_ERROR)
6840 /* Allow read access to the freelist */
6841 if (!dbi && !F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
6844 if (txn->mt_dbs[dbi].md_flags & MDB_DUPSORT)
6845 size += sizeof(MDB_xcursor);
6847 if ((mc = malloc(size)) != NULL) {
6848 mdb_cursor_init(mc, txn, dbi, (MDB_xcursor *)(mc + 1));
6849 if (txn->mt_cursors) {
6850 mc->mc_next = txn->mt_cursors[dbi];
6851 txn->mt_cursors[dbi] = mc;
6852 mc->mc_flags |= C_UNTRACK;
6864 mdb_cursor_renew(MDB_txn *txn, MDB_cursor *mc)
6866 if (!mc || !TXN_DBI_EXIST(txn, mc->mc_dbi))
6869 if ((mc->mc_flags & C_UNTRACK) || txn->mt_cursors)
6872 if (txn->mt_flags & MDB_TXN_ERROR)
6875 mdb_cursor_init(mc, txn, mc->mc_dbi, mc->mc_xcursor);
6879 /* Return the count of duplicate data items for the current key */
6881 mdb_cursor_count(MDB_cursor *mc, size_t *countp)
6885 if (mc == NULL || countp == NULL)
6888 if (mc->mc_xcursor == NULL)
6889 return MDB_INCOMPATIBLE;
6891 if (mc->mc_txn->mt_flags & MDB_TXN_ERROR)
6894 leaf = NODEPTR(mc->mc_pg[mc->mc_top], mc->mc_ki[mc->mc_top]);
6895 if (!F_ISSET(leaf->mn_flags, F_DUPDATA)) {
6898 if (!(mc->mc_xcursor->mx_cursor.mc_flags & C_INITIALIZED))
6901 *countp = mc->mc_xcursor->mx_db.md_entries;
6907 mdb_cursor_close(MDB_cursor *mc)
6909 if (mc && !mc->mc_backup) {
6910 /* remove from txn, if tracked */
6911 if ((mc->mc_flags & C_UNTRACK) && mc->mc_txn->mt_cursors) {
6912 MDB_cursor **prev = &mc->mc_txn->mt_cursors[mc->mc_dbi];
6913 while (*prev && *prev != mc) prev = &(*prev)->mc_next;
6915 *prev = mc->mc_next;
6922 mdb_cursor_txn(MDB_cursor *mc)
6924 if (!mc) return NULL;
6929 mdb_cursor_dbi(MDB_cursor *mc)
6934 /** Replace the key for a branch node with a new key.
6935 * @param[in] mc Cursor pointing to the node to operate on.
6936 * @param[in] key The new key to use.
6937 * @return 0 on success, non-zero on failure.
6940 mdb_update_key(MDB_cursor *mc, MDB_val *key)
6946 int delta, ksize, oksize;
6947 indx_t ptr, i, numkeys, indx;
6950 indx = mc->mc_ki[mc->mc_top];
6951 mp = mc->mc_pg[mc->mc_top];
6952 node = NODEPTR(mp, indx);
6953 ptr = mp->mp_ptrs[indx];
6957 char kbuf2[DKBUF_MAXKEYSIZE*2+1];
6958 k2.mv_data = NODEKEY(node);
6959 k2.mv_size = node->mn_ksize;
6960 DPRINTF(("update key %u (ofs %u) [%s] to [%s] on page %"Z"u",
6962 mdb_dkey(&k2, kbuf2),
6968 /* Sizes must be 2-byte aligned. */
6969 ksize = EVEN(key->mv_size);
6970 oksize = EVEN(node->mn_ksize);
6971 delta = ksize - oksize;
6973 /* Shift node contents if EVEN(key length) changed. */
6975 if (delta > 0 && SIZELEFT(mp) < delta) {
6977 /* not enough space left, do a delete and split */
6978 DPRINTF(("Not enough room, delta = %d, splitting...", delta));
6979 pgno = NODEPGNO(node);
6980 mdb_node_del(mc, 0);
6981 return mdb_page_split(mc, key, NULL, pgno, MDB_SPLIT_REPLACE);
6984 numkeys = NUMKEYS(mp);
6985 for (i = 0; i < numkeys; i++) {
6986 if (mp->mp_ptrs[i] <= ptr)
6987 mp->mp_ptrs[i] -= delta;
6990 base = (char *)mp + mp->mp_upper + PAGEBASE;
6991 len = ptr - mp->mp_upper + NODESIZE;
6992 memmove(base - delta, base, len);
6993 mp->mp_upper -= delta;
6995 node = NODEPTR(mp, indx);
6998 /* But even if no shift was needed, update ksize */
6999 if (node->mn_ksize != key->mv_size)
7000 node->mn_ksize = key->mv_size;
7003 memcpy(NODEKEY(node), key->mv_data, key->mv_size);
7009 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst);
7011 /** Move a node from csrc to cdst.
7014 mdb_node_move(MDB_cursor *csrc, MDB_cursor *cdst)
7021 unsigned short flags;
7025 /* Mark src and dst as dirty. */
7026 if ((rc = mdb_page_touch(csrc)) ||
7027 (rc = mdb_page_touch(cdst)))
7030 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7031 key.mv_size = csrc->mc_db->md_pad;
7032 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top], key.mv_size);
7034 data.mv_data = NULL;
7038 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], csrc->mc_ki[csrc->mc_top]);
7039 mdb_cassert(csrc, !((size_t)srcnode & 1));
7040 srcpg = NODEPGNO(srcnode);
7041 flags = srcnode->mn_flags;
7042 if (csrc->mc_ki[csrc->mc_top] == 0 && IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7043 unsigned int snum = csrc->mc_snum;
7045 /* must find the lowest key below src */
7046 rc = mdb_page_search_lowest(csrc);
7049 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7050 key.mv_size = csrc->mc_db->md_pad;
7051 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7053 s2 = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7054 key.mv_size = NODEKSZ(s2);
7055 key.mv_data = NODEKEY(s2);
7057 csrc->mc_snum = snum--;
7058 csrc->mc_top = snum;
7060 key.mv_size = NODEKSZ(srcnode);
7061 key.mv_data = NODEKEY(srcnode);
7063 data.mv_size = NODEDSZ(srcnode);
7064 data.mv_data = NODEDATA(srcnode);
7066 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top]) && cdst->mc_ki[cdst->mc_top] == 0) {
7067 unsigned int snum = cdst->mc_snum;
7070 /* must find the lowest key below dst */
7071 mdb_cursor_copy(cdst, &mn);
7072 rc = mdb_page_search_lowest(&mn);
7075 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7076 bkey.mv_size = mn.mc_db->md_pad;
7077 bkey.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, bkey.mv_size);
7079 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7080 bkey.mv_size = NODEKSZ(s2);
7081 bkey.mv_data = NODEKEY(s2);
7083 mn.mc_snum = snum--;
7086 rc = mdb_update_key(&mn, &bkey);
7091 DPRINTF(("moving %s node %u [%s] on page %"Z"u to node %u on page %"Z"u",
7092 IS_LEAF(csrc->mc_pg[csrc->mc_top]) ? "leaf" : "branch",
7093 csrc->mc_ki[csrc->mc_top],
7095 csrc->mc_pg[csrc->mc_top]->mp_pgno,
7096 cdst->mc_ki[cdst->mc_top], cdst->mc_pg[cdst->mc_top]->mp_pgno));
7098 /* Add the node to the destination page.
7100 rc = mdb_node_add(cdst, cdst->mc_ki[cdst->mc_top], &key, &data, srcpg, flags);
7101 if (rc != MDB_SUCCESS)
7104 /* Delete the node from the source page.
7106 mdb_node_del(csrc, key.mv_size);
7109 /* Adjust other cursors pointing to mp */
7110 MDB_cursor *m2, *m3;
7111 MDB_dbi dbi = csrc->mc_dbi;
7112 MDB_page *mp = csrc->mc_pg[csrc->mc_top];
7114 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7115 if (csrc->mc_flags & C_SUB)
7116 m3 = &m2->mc_xcursor->mx_cursor;
7119 if (m3 == csrc) continue;
7120 if (m3->mc_pg[csrc->mc_top] == mp && m3->mc_ki[csrc->mc_top] ==
7121 csrc->mc_ki[csrc->mc_top]) {
7122 m3->mc_pg[csrc->mc_top] = cdst->mc_pg[cdst->mc_top];
7123 m3->mc_ki[csrc->mc_top] = cdst->mc_ki[cdst->mc_top];
7128 /* Update the parent separators.
7130 if (csrc->mc_ki[csrc->mc_top] == 0) {
7131 if (csrc->mc_ki[csrc->mc_top-1] != 0) {
7132 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7133 key.mv_data = LEAF2KEY(csrc->mc_pg[csrc->mc_top], 0, key.mv_size);
7135 srcnode = NODEPTR(csrc->mc_pg[csrc->mc_top], 0);
7136 key.mv_size = NODEKSZ(srcnode);
7137 key.mv_data = NODEKEY(srcnode);
7139 DPRINTF(("update separator for source page %"Z"u to [%s]",
7140 csrc->mc_pg[csrc->mc_top]->mp_pgno, DKEY(&key)));
7141 mdb_cursor_copy(csrc, &mn);
7144 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7147 if (IS_BRANCH(csrc->mc_pg[csrc->mc_top])) {
7149 indx_t ix = csrc->mc_ki[csrc->mc_top];
7150 nullkey.mv_size = 0;
7151 csrc->mc_ki[csrc->mc_top] = 0;
7152 rc = mdb_update_key(csrc, &nullkey);
7153 csrc->mc_ki[csrc->mc_top] = ix;
7154 mdb_cassert(csrc, rc == MDB_SUCCESS);
7158 if (cdst->mc_ki[cdst->mc_top] == 0) {
7159 if (cdst->mc_ki[cdst->mc_top-1] != 0) {
7160 if (IS_LEAF2(csrc->mc_pg[csrc->mc_top])) {
7161 key.mv_data = LEAF2KEY(cdst->mc_pg[cdst->mc_top], 0, key.mv_size);
7163 srcnode = NODEPTR(cdst->mc_pg[cdst->mc_top], 0);
7164 key.mv_size = NODEKSZ(srcnode);
7165 key.mv_data = NODEKEY(srcnode);
7167 DPRINTF(("update separator for destination page %"Z"u to [%s]",
7168 cdst->mc_pg[cdst->mc_top]->mp_pgno, DKEY(&key)));
7169 mdb_cursor_copy(cdst, &mn);
7172 if ((rc = mdb_update_key(&mn, &key)) != MDB_SUCCESS)
7175 if (IS_BRANCH(cdst->mc_pg[cdst->mc_top])) {
7177 indx_t ix = cdst->mc_ki[cdst->mc_top];
7178 nullkey.mv_size = 0;
7179 cdst->mc_ki[cdst->mc_top] = 0;
7180 rc = mdb_update_key(cdst, &nullkey);
7181 cdst->mc_ki[cdst->mc_top] = ix;
7182 mdb_cassert(csrc, rc == MDB_SUCCESS);
7189 /** Merge one page into another.
7190 * The nodes from the page pointed to by \b csrc will
7191 * be copied to the page pointed to by \b cdst and then
7192 * the \b csrc page will be freed.
7193 * @param[in] csrc Cursor pointing to the source page.
7194 * @param[in] cdst Cursor pointing to the destination page.
7195 * @return 0 on success, non-zero on failure.
7198 mdb_page_merge(MDB_cursor *csrc, MDB_cursor *cdst)
7200 MDB_page *psrc, *pdst;
7207 psrc = csrc->mc_pg[csrc->mc_top];
7208 pdst = cdst->mc_pg[cdst->mc_top];
7210 DPRINTF(("merging page %"Z"u into %"Z"u", psrc->mp_pgno, pdst->mp_pgno));
7212 mdb_cassert(csrc, csrc->mc_snum > 1); /* can't merge root page */
7213 mdb_cassert(csrc, cdst->mc_snum > 1);
7215 /* Mark dst as dirty. */
7216 if ((rc = mdb_page_touch(cdst)))
7219 /* Move all nodes from src to dst.
7221 j = nkeys = NUMKEYS(pdst);
7222 if (IS_LEAF2(psrc)) {
7223 key.mv_size = csrc->mc_db->md_pad;
7224 key.mv_data = METADATA(psrc);
7225 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7226 rc = mdb_node_add(cdst, j, &key, NULL, 0, 0);
7227 if (rc != MDB_SUCCESS)
7229 key.mv_data = (char *)key.mv_data + key.mv_size;
7232 for (i = 0; i < NUMKEYS(psrc); i++, j++) {
7233 srcnode = NODEPTR(psrc, i);
7234 if (i == 0 && IS_BRANCH(psrc)) {
7237 mdb_cursor_copy(csrc, &mn);
7238 /* must find the lowest key below src */
7239 rc = mdb_page_search_lowest(&mn);
7242 if (IS_LEAF2(mn.mc_pg[mn.mc_top])) {
7243 key.mv_size = mn.mc_db->md_pad;
7244 key.mv_data = LEAF2KEY(mn.mc_pg[mn.mc_top], 0, key.mv_size);
7246 s2 = NODEPTR(mn.mc_pg[mn.mc_top], 0);
7247 key.mv_size = NODEKSZ(s2);
7248 key.mv_data = NODEKEY(s2);
7251 key.mv_size = srcnode->mn_ksize;
7252 key.mv_data = NODEKEY(srcnode);
7255 data.mv_size = NODEDSZ(srcnode);
7256 data.mv_data = NODEDATA(srcnode);
7257 rc = mdb_node_add(cdst, j, &key, &data, NODEPGNO(srcnode), srcnode->mn_flags);
7258 if (rc != MDB_SUCCESS)
7263 DPRINTF(("dst page %"Z"u now has %u keys (%.1f%% filled)",
7264 pdst->mp_pgno, NUMKEYS(pdst),
7265 (float)PAGEFILL(cdst->mc_txn->mt_env, pdst) / 10));
7267 /* Unlink the src page from parent and add to free list.
7270 mdb_node_del(csrc, 0);
7271 if (csrc->mc_ki[csrc->mc_top] == 0) {
7273 rc = mdb_update_key(csrc, &key);
7281 psrc = csrc->mc_pg[csrc->mc_top];
7282 /* If not operating on FreeDB, allow this page to be reused
7283 * in this txn. Otherwise just add to free list.
7285 rc = mdb_page_loose(csrc, psrc);
7289 csrc->mc_db->md_leaf_pages--;
7291 csrc->mc_db->md_branch_pages--;
7293 /* Adjust other cursors pointing to mp */
7294 MDB_cursor *m2, *m3;
7295 MDB_dbi dbi = csrc->mc_dbi;
7297 for (m2 = csrc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7298 if (csrc->mc_flags & C_SUB)
7299 m3 = &m2->mc_xcursor->mx_cursor;
7302 if (m3 == csrc) continue;
7303 if (m3->mc_snum < csrc->mc_snum) continue;
7304 if (m3->mc_pg[csrc->mc_top] == psrc) {
7305 m3->mc_pg[csrc->mc_top] = pdst;
7306 m3->mc_ki[csrc->mc_top] += nkeys;
7311 unsigned int snum = cdst->mc_snum;
7312 uint16_t depth = cdst->mc_db->md_depth;
7313 mdb_cursor_pop(cdst);
7314 rc = mdb_rebalance(cdst);
7315 /* Did the tree shrink? */
7316 if (depth > cdst->mc_db->md_depth)
7318 cdst->mc_snum = snum;
7319 cdst->mc_top = snum-1;
7324 /** Copy the contents of a cursor.
7325 * @param[in] csrc The cursor to copy from.
7326 * @param[out] cdst The cursor to copy to.
7329 mdb_cursor_copy(const MDB_cursor *csrc, MDB_cursor *cdst)
7333 cdst->mc_txn = csrc->mc_txn;
7334 cdst->mc_dbi = csrc->mc_dbi;
7335 cdst->mc_db = csrc->mc_db;
7336 cdst->mc_dbx = csrc->mc_dbx;
7337 cdst->mc_snum = csrc->mc_snum;
7338 cdst->mc_top = csrc->mc_top;
7339 cdst->mc_flags = csrc->mc_flags;
7341 for (i=0; i<csrc->mc_snum; i++) {
7342 cdst->mc_pg[i] = csrc->mc_pg[i];
7343 cdst->mc_ki[i] = csrc->mc_ki[i];
7347 /** Rebalance the tree after a delete operation.
7348 * @param[in] mc Cursor pointing to the page where rebalancing
7350 * @return 0 on success, non-zero on failure.
7353 mdb_rebalance(MDB_cursor *mc)
7357 unsigned int ptop, minkeys;
7361 minkeys = 1 + (IS_BRANCH(mc->mc_pg[mc->mc_top]));
7362 DPRINTF(("rebalancing %s page %"Z"u (has %u keys, %.1f%% full)",
7363 IS_LEAF(mc->mc_pg[mc->mc_top]) ? "leaf" : "branch",
7364 mdb_dbg_pgno(mc->mc_pg[mc->mc_top]), NUMKEYS(mc->mc_pg[mc->mc_top]),
7365 (float)PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) / 10));
7367 if (PAGEFILL(mc->mc_txn->mt_env, mc->mc_pg[mc->mc_top]) >= FILL_THRESHOLD &&
7368 NUMKEYS(mc->mc_pg[mc->mc_top]) >= minkeys) {
7369 DPRINTF(("no need to rebalance page %"Z"u, above fill threshold",
7370 mdb_dbg_pgno(mc->mc_pg[mc->mc_top])));
7374 if (mc->mc_snum < 2) {
7375 MDB_page *mp = mc->mc_pg[0];
7377 DPUTS("Can't rebalance a subpage, ignoring");
7380 if (NUMKEYS(mp) == 0) {
7381 DPUTS("tree is completely empty");
7382 mc->mc_db->md_root = P_INVALID;
7383 mc->mc_db->md_depth = 0;
7384 mc->mc_db->md_leaf_pages = 0;
7385 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7388 /* Adjust cursors pointing to mp */
7391 mc->mc_flags &= ~C_INITIALIZED;
7393 MDB_cursor *m2, *m3;
7394 MDB_dbi dbi = mc->mc_dbi;
7396 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7397 if (mc->mc_flags & C_SUB)
7398 m3 = &m2->mc_xcursor->mx_cursor;
7401 if (m3->mc_snum < mc->mc_snum) continue;
7402 if (m3->mc_pg[0] == mp) {
7405 m3->mc_flags &= ~C_INITIALIZED;
7409 } else if (IS_BRANCH(mp) && NUMKEYS(mp) == 1) {
7411 DPUTS("collapsing root page!");
7412 rc = mdb_midl_append(&mc->mc_txn->mt_free_pgs, mp->mp_pgno);
7415 mc->mc_db->md_root = NODEPGNO(NODEPTR(mp, 0));
7416 rc = mdb_page_get(mc->mc_txn,mc->mc_db->md_root,&mc->mc_pg[0],NULL);
7419 mc->mc_db->md_depth--;
7420 mc->mc_db->md_branch_pages--;
7421 mc->mc_ki[0] = mc->mc_ki[1];
7422 for (i = 1; i<mc->mc_db->md_depth; i++) {
7423 mc->mc_pg[i] = mc->mc_pg[i+1];
7424 mc->mc_ki[i] = mc->mc_ki[i+1];
7427 /* Adjust other cursors pointing to mp */
7428 MDB_cursor *m2, *m3;
7429 MDB_dbi dbi = mc->mc_dbi;
7431 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7432 if (mc->mc_flags & C_SUB)
7433 m3 = &m2->mc_xcursor->mx_cursor;
7436 if (m3 == mc || m3->mc_snum < mc->mc_snum) continue;
7437 if (m3->mc_pg[0] == mp) {
7440 for (i=0; i<m3->mc_snum; i++) {
7441 m3->mc_pg[i] = m3->mc_pg[i+1];
7442 m3->mc_ki[i] = m3->mc_ki[i+1];
7448 DPUTS("root page doesn't need rebalancing");
7452 /* The parent (branch page) must have at least 2 pointers,
7453 * otherwise the tree is invalid.
7455 ptop = mc->mc_top-1;
7456 mdb_cassert(mc, NUMKEYS(mc->mc_pg[ptop]) > 1);
7458 /* Leaf page fill factor is below the threshold.
7459 * Try to move keys from left or right neighbor, or
7460 * merge with a neighbor page.
7465 mdb_cursor_copy(mc, &mn);
7466 mn.mc_xcursor = NULL;
7468 oldki = mc->mc_ki[mc->mc_top];
7469 if (mc->mc_ki[ptop] == 0) {
7470 /* We're the leftmost leaf in our parent.
7472 DPUTS("reading right neighbor");
7474 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7475 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7478 mn.mc_ki[mn.mc_top] = 0;
7479 mc->mc_ki[mc->mc_top] = NUMKEYS(mc->mc_pg[mc->mc_top]);
7481 /* There is at least one neighbor to the left.
7483 DPUTS("reading left neighbor");
7485 node = NODEPTR(mc->mc_pg[ptop], mn.mc_ki[ptop]);
7486 rc = mdb_page_get(mc->mc_txn,NODEPGNO(node),&mn.mc_pg[mn.mc_top],NULL);
7489 mn.mc_ki[mn.mc_top] = NUMKEYS(mn.mc_pg[mn.mc_top]) - 1;
7490 mc->mc_ki[mc->mc_top] = 0;
7493 DPRINTF(("found neighbor page %"Z"u (%u keys, %.1f%% full)",
7494 mn.mc_pg[mn.mc_top]->mp_pgno, NUMKEYS(mn.mc_pg[mn.mc_top]),
7495 (float)PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) / 10));
7497 /* If the neighbor page is above threshold and has enough keys,
7498 * move one key from it. Otherwise we should try to merge them.
7499 * (A branch page must never have less than 2 keys.)
7501 minkeys = 1 + (IS_BRANCH(mn.mc_pg[mn.mc_top]));
7502 if (PAGEFILL(mc->mc_txn->mt_env, mn.mc_pg[mn.mc_top]) >= FILL_THRESHOLD && NUMKEYS(mn.mc_pg[mn.mc_top]) > minkeys) {
7503 rc = mdb_node_move(&mn, mc);
7504 if (mc->mc_ki[ptop]) {
7508 if (mc->mc_ki[ptop] == 0) {
7509 rc = mdb_page_merge(&mn, mc);
7511 oldki += NUMKEYS(mn.mc_pg[mn.mc_top]);
7512 mn.mc_ki[mn.mc_top] += mc->mc_ki[mn.mc_top] + 1;
7513 rc = mdb_page_merge(mc, &mn);
7514 mdb_cursor_copy(&mn, mc);
7516 mc->mc_flags &= ~C_EOF;
7518 mc->mc_ki[mc->mc_top] = oldki;
7522 /** Complete a delete operation started by #mdb_cursor_del(). */
7524 mdb_cursor_del0(MDB_cursor *mc)
7531 ki = mc->mc_ki[mc->mc_top];
7532 mdb_node_del(mc, mc->mc_db->md_pad);
7533 mc->mc_db->md_entries--;
7534 rc = mdb_rebalance(mc);
7536 if (rc == MDB_SUCCESS) {
7537 MDB_cursor *m2, *m3;
7538 MDB_dbi dbi = mc->mc_dbi;
7540 mp = mc->mc_pg[mc->mc_top];
7541 nkeys = NUMKEYS(mp);
7543 /* if mc points past last node in page, find next sibling */
7544 if (mc->mc_ki[mc->mc_top] >= nkeys) {
7545 rc = mdb_cursor_sibling(mc, 1);
7546 if (rc == MDB_NOTFOUND) {
7547 mc->mc_flags |= C_EOF;
7552 /* Adjust other cursors pointing to mp */
7553 for (m2 = mc->mc_txn->mt_cursors[dbi]; !rc && m2; m2=m2->mc_next) {
7554 m3 = (mc->mc_flags & C_SUB) ? &m2->mc_xcursor->mx_cursor : m2;
7555 if (! (m2->mc_flags & m3->mc_flags & C_INITIALIZED))
7557 if (m3 == mc || m3->mc_snum < mc->mc_snum)
7559 if (m3->mc_pg[mc->mc_top] == mp) {
7560 if (m3->mc_ki[mc->mc_top] >= ki) {
7561 m3->mc_flags |= C_DEL;
7562 if (m3->mc_ki[mc->mc_top] > ki)
7563 m3->mc_ki[mc->mc_top]--;
7565 if (m3->mc_ki[mc->mc_top] >= nkeys) {
7566 rc = mdb_cursor_sibling(m3, 1);
7567 if (rc == MDB_NOTFOUND) {
7568 m3->mc_flags |= C_EOF;
7574 mc->mc_flags |= C_DEL;
7578 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
7583 mdb_del(MDB_txn *txn, MDB_dbi dbi,
7584 MDB_val *key, MDB_val *data)
7586 if (!key || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
7589 if (txn->mt_flags & (MDB_TXN_RDONLY|MDB_TXN_ERROR))
7590 return (txn->mt_flags & MDB_TXN_RDONLY) ? EACCES : MDB_BAD_TXN;
7592 if (!F_ISSET(txn->mt_dbs[dbi].md_flags, MDB_DUPSORT)) {
7593 /* must ignore any data */
7597 return mdb_del0(txn, dbi, key, data, 0);
7601 mdb_del0(MDB_txn *txn, MDB_dbi dbi,
7602 MDB_val *key, MDB_val *data, unsigned flags)
7607 MDB_val rdata, *xdata;
7611 DPRINTF(("====> delete db %u key [%s]", dbi, DKEY(key)));
7613 mdb_cursor_init(&mc, txn, dbi, &mx);
7622 flags |= MDB_NODUPDATA;
7624 rc = mdb_cursor_set(&mc, key, xdata, op, &exact);
7626 /* let mdb_page_split know about this cursor if needed:
7627 * delete will trigger a rebalance; if it needs to move
7628 * a node from one page to another, it will have to
7629 * update the parent's separator key(s). If the new sepkey
7630 * is larger than the current one, the parent page may
7631 * run out of space, triggering a split. We need this
7632 * cursor to be consistent until the end of the rebalance.
7634 mc.mc_flags |= C_UNTRACK;
7635 mc.mc_next = txn->mt_cursors[dbi];
7636 txn->mt_cursors[dbi] = &mc;
7637 rc = mdb_cursor_del(&mc, flags);
7638 txn->mt_cursors[dbi] = mc.mc_next;
7643 /** Split a page and insert a new node.
7644 * @param[in,out] mc Cursor pointing to the page and desired insertion index.
7645 * The cursor will be updated to point to the actual page and index where
7646 * the node got inserted after the split.
7647 * @param[in] newkey The key for the newly inserted node.
7648 * @param[in] newdata The data for the newly inserted node.
7649 * @param[in] newpgno The page number, if the new node is a branch node.
7650 * @param[in] nflags The #NODE_ADD_FLAGS for the new node.
7651 * @return 0 on success, non-zero on failure.
7654 mdb_page_split(MDB_cursor *mc, MDB_val *newkey, MDB_val *newdata, pgno_t newpgno,
7655 unsigned int nflags)
7658 int rc = MDB_SUCCESS, new_root = 0, did_split = 0;
7661 int i, j, split_indx, nkeys, pmax;
7662 MDB_env *env = mc->mc_txn->mt_env;
7664 MDB_val sepkey, rkey, xdata, *rdata = &xdata;
7665 MDB_page *copy = NULL;
7666 MDB_page *mp, *rp, *pp;
7671 mp = mc->mc_pg[mc->mc_top];
7672 newindx = mc->mc_ki[mc->mc_top];
7673 nkeys = NUMKEYS(mp);
7675 DPRINTF(("-----> splitting %s page %"Z"u and adding [%s] at index %i/%i",
7676 IS_LEAF(mp) ? "leaf" : "branch", mp->mp_pgno,
7677 DKEY(newkey), mc->mc_ki[mc->mc_top], nkeys));
7679 /* Create a right sibling. */
7680 if ((rc = mdb_page_new(mc, mp->mp_flags, 1, &rp)))
7682 DPRINTF(("new right sibling: page %"Z"u", rp->mp_pgno));
7684 if (mc->mc_snum < 2) {
7685 if ((rc = mdb_page_new(mc, P_BRANCH, 1, &pp)))
7687 /* shift current top to make room for new parent */
7688 mc->mc_pg[1] = mc->mc_pg[0];
7689 mc->mc_ki[1] = mc->mc_ki[0];
7692 mc->mc_db->md_root = pp->mp_pgno;
7693 DPRINTF(("root split! new root = %"Z"u", pp->mp_pgno));
7694 mc->mc_db->md_depth++;
7697 /* Add left (implicit) pointer. */
7698 if ((rc = mdb_node_add(mc, 0, NULL, NULL, mp->mp_pgno, 0)) != MDB_SUCCESS) {
7699 /* undo the pre-push */
7700 mc->mc_pg[0] = mc->mc_pg[1];
7701 mc->mc_ki[0] = mc->mc_ki[1];
7702 mc->mc_db->md_root = mp->mp_pgno;
7703 mc->mc_db->md_depth--;
7710 ptop = mc->mc_top-1;
7711 DPRINTF(("parent branch page is %"Z"u", mc->mc_pg[ptop]->mp_pgno));
7714 mc->mc_flags |= C_SPLITTING;
7715 mdb_cursor_copy(mc, &mn);
7716 mn.mc_pg[mn.mc_top] = rp;
7717 mn.mc_ki[ptop] = mc->mc_ki[ptop]+1;
7719 if (nflags & MDB_APPEND) {
7720 mn.mc_ki[mn.mc_top] = 0;
7722 split_indx = newindx;
7726 split_indx = (nkeys+1) / 2;
7731 unsigned int lsize, rsize, ksize;
7732 /* Move half of the keys to the right sibling */
7733 x = mc->mc_ki[mc->mc_top] - split_indx;
7734 ksize = mc->mc_db->md_pad;
7735 split = LEAF2KEY(mp, split_indx, ksize);
7736 rsize = (nkeys - split_indx) * ksize;
7737 lsize = (nkeys - split_indx) * sizeof(indx_t);
7738 mp->mp_lower -= lsize;
7739 rp->mp_lower += lsize;
7740 mp->mp_upper += rsize - lsize;
7741 rp->mp_upper -= rsize - lsize;
7742 sepkey.mv_size = ksize;
7743 if (newindx == split_indx) {
7744 sepkey.mv_data = newkey->mv_data;
7746 sepkey.mv_data = split;
7749 ins = LEAF2KEY(mp, mc->mc_ki[mc->mc_top], ksize);
7750 memcpy(rp->mp_ptrs, split, rsize);
7751 sepkey.mv_data = rp->mp_ptrs;
7752 memmove(ins+ksize, ins, (split_indx - mc->mc_ki[mc->mc_top]) * ksize);
7753 memcpy(ins, newkey->mv_data, ksize);
7754 mp->mp_lower += sizeof(indx_t);
7755 mp->mp_upper -= ksize - sizeof(indx_t);
7758 memcpy(rp->mp_ptrs, split, x * ksize);
7759 ins = LEAF2KEY(rp, x, ksize);
7760 memcpy(ins, newkey->mv_data, ksize);
7761 memcpy(ins+ksize, split + x * ksize, rsize - x * ksize);
7762 rp->mp_lower += sizeof(indx_t);
7763 rp->mp_upper -= ksize - sizeof(indx_t);
7764 mc->mc_ki[mc->mc_top] = x;
7765 mc->mc_pg[mc->mc_top] = rp;
7768 int psize, nsize, k;
7769 /* Maximum free space in an empty page */
7770 pmax = env->me_psize - PAGEHDRSZ;
7772 nsize = mdb_leaf_size(env, newkey, newdata);
7774 nsize = mdb_branch_size(env, newkey);
7775 nsize = EVEN(nsize);
7777 /* grab a page to hold a temporary copy */
7778 copy = mdb_page_malloc(mc->mc_txn, 1);
7783 copy->mp_pgno = mp->mp_pgno;
7784 copy->mp_flags = mp->mp_flags;
7785 copy->mp_lower = (PAGEHDRSZ-PAGEBASE);
7786 copy->mp_upper = env->me_psize - PAGEBASE;
7788 /* prepare to insert */
7789 for (i=0, j=0; i<nkeys; i++) {
7791 copy->mp_ptrs[j++] = 0;
7793 copy->mp_ptrs[j++] = mp->mp_ptrs[i];
7796 /* When items are relatively large the split point needs
7797 * to be checked, because being off-by-one will make the
7798 * difference between success or failure in mdb_node_add.
7800 * It's also relevant if a page happens to be laid out
7801 * such that one half of its nodes are all "small" and
7802 * the other half of its nodes are "large." If the new
7803 * item is also "large" and falls on the half with
7804 * "large" nodes, it also may not fit.
7806 * As a final tweak, if the new item goes on the last
7807 * spot on the page (and thus, onto the new page), bias
7808 * the split so the new page is emptier than the old page.
7809 * This yields better packing during sequential inserts.
7811 if (nkeys < 20 || nsize > pmax/16 || newindx >= nkeys) {
7812 /* Find split point */
7814 if (newindx <= split_indx || newindx >= nkeys) {
7816 k = newindx >= nkeys ? nkeys : split_indx+2;
7821 for (; i!=k; i+=j) {
7826 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
7827 psize += NODESIZE + NODEKSZ(node) + sizeof(indx_t);
7829 if (F_ISSET(node->mn_flags, F_BIGDATA))
7830 psize += sizeof(pgno_t);
7832 psize += NODEDSZ(node);
7834 psize = EVEN(psize);
7836 if (psize > pmax || i == k-j) {
7837 split_indx = i + (j<0);
7842 if (split_indx == newindx) {
7843 sepkey.mv_size = newkey->mv_size;
7844 sepkey.mv_data = newkey->mv_data;
7846 node = (MDB_node *)((char *)mp + copy->mp_ptrs[split_indx] + PAGEBASE);
7847 sepkey.mv_size = node->mn_ksize;
7848 sepkey.mv_data = NODEKEY(node);
7853 DPRINTF(("separator is %d [%s]", split_indx, DKEY(&sepkey)));
7855 /* Copy separator key to the parent.
7857 if (SIZELEFT(mn.mc_pg[ptop]) < mdb_branch_size(env, &sepkey)) {
7861 rc = mdb_page_split(&mn, &sepkey, NULL, rp->mp_pgno, 0);
7866 if (mn.mc_snum == mc->mc_snum) {
7867 mc->mc_pg[mc->mc_snum] = mc->mc_pg[mc->mc_top];
7868 mc->mc_ki[mc->mc_snum] = mc->mc_ki[mc->mc_top];
7869 mc->mc_pg[mc->mc_top] = mc->mc_pg[ptop];
7870 mc->mc_ki[mc->mc_top] = mc->mc_ki[ptop];
7875 /* Right page might now have changed parent.
7876 * Check if left page also changed parent.
7878 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7879 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7880 for (i=0; i<ptop; i++) {
7881 mc->mc_pg[i] = mn.mc_pg[i];
7882 mc->mc_ki[i] = mn.mc_ki[i];
7884 mc->mc_pg[ptop] = mn.mc_pg[ptop];
7885 if (mn.mc_ki[ptop]) {
7886 mc->mc_ki[ptop] = mn.mc_ki[ptop] - 1;
7888 /* find right page's left sibling */
7889 mc->mc_ki[ptop] = mn.mc_ki[ptop];
7890 mdb_cursor_sibling(mc, 0);
7895 rc = mdb_node_add(&mn, mn.mc_ki[ptop], &sepkey, NULL, rp->mp_pgno, 0);
7898 mc->mc_flags ^= C_SPLITTING;
7899 if (rc != MDB_SUCCESS) {
7902 if (nflags & MDB_APPEND) {
7903 mc->mc_pg[mc->mc_top] = rp;
7904 mc->mc_ki[mc->mc_top] = 0;
7905 rc = mdb_node_add(mc, 0, newkey, newdata, newpgno, nflags);
7908 for (i=0; i<mc->mc_top; i++)
7909 mc->mc_ki[i] = mn.mc_ki[i];
7910 } else if (!IS_LEAF2(mp)) {
7912 mc->mc_pg[mc->mc_top] = rp;
7917 rkey.mv_data = newkey->mv_data;
7918 rkey.mv_size = newkey->mv_size;
7924 /* Update index for the new key. */
7925 mc->mc_ki[mc->mc_top] = j;
7927 node = (MDB_node *)((char *)mp + copy->mp_ptrs[i] + PAGEBASE);
7928 rkey.mv_data = NODEKEY(node);
7929 rkey.mv_size = node->mn_ksize;
7931 xdata.mv_data = NODEDATA(node);
7932 xdata.mv_size = NODEDSZ(node);
7935 pgno = NODEPGNO(node);
7936 flags = node->mn_flags;
7939 if (!IS_LEAF(mp) && j == 0) {
7940 /* First branch index doesn't need key data. */
7944 rc = mdb_node_add(mc, j, &rkey, rdata, pgno, flags);
7950 mc->mc_pg[mc->mc_top] = copy;
7955 } while (i != split_indx);
7957 nkeys = NUMKEYS(copy);
7958 for (i=0; i<nkeys; i++)
7959 mp->mp_ptrs[i] = copy->mp_ptrs[i];
7960 mp->mp_lower = copy->mp_lower;
7961 mp->mp_upper = copy->mp_upper;
7962 memcpy(NODEPTR(mp, nkeys-1), NODEPTR(copy, nkeys-1),
7963 env->me_psize - copy->mp_upper - PAGEBASE);
7965 /* reset back to original page */
7966 if (newindx < split_indx) {
7967 mc->mc_pg[mc->mc_top] = mp;
7968 if (nflags & MDB_RESERVE) {
7969 node = NODEPTR(mp, mc->mc_ki[mc->mc_top]);
7970 if (!(node->mn_flags & F_BIGDATA))
7971 newdata->mv_data = NODEDATA(node);
7974 mc->mc_pg[mc->mc_top] = rp;
7976 /* Make sure mc_ki is still valid.
7978 if (mn.mc_pg[ptop] != mc->mc_pg[ptop] &&
7979 mc->mc_ki[ptop] >= NUMKEYS(mc->mc_pg[ptop])) {
7980 for (i=0; i<=ptop; i++) {
7981 mc->mc_pg[i] = mn.mc_pg[i];
7982 mc->mc_ki[i] = mn.mc_ki[i];
7989 /* Adjust other cursors pointing to mp */
7990 MDB_cursor *m2, *m3;
7991 MDB_dbi dbi = mc->mc_dbi;
7992 int fixup = NUMKEYS(mp);
7994 for (m2 = mc->mc_txn->mt_cursors[dbi]; m2; m2=m2->mc_next) {
7995 if (mc->mc_flags & C_SUB)
7996 m3 = &m2->mc_xcursor->mx_cursor;
8001 if (!(m2->mc_flags & m3->mc_flags & C_INITIALIZED))
8003 if (m3->mc_flags & C_SPLITTING)
8008 for (k=m3->mc_top; k>=0; k--) {
8009 m3->mc_ki[k+1] = m3->mc_ki[k];
8010 m3->mc_pg[k+1] = m3->mc_pg[k];
8012 if (m3->mc_ki[0] >= split_indx) {
8017 m3->mc_pg[0] = mc->mc_pg[0];
8021 if (m3->mc_top >= mc->mc_top && m3->mc_pg[mc->mc_top] == mp) {
8022 if (m3->mc_ki[mc->mc_top] >= newindx && !(nflags & MDB_SPLIT_REPLACE))
8023 m3->mc_ki[mc->mc_top]++;
8024 if (m3->mc_ki[mc->mc_top] >= fixup) {
8025 m3->mc_pg[mc->mc_top] = rp;
8026 m3->mc_ki[mc->mc_top] -= fixup;
8027 m3->mc_ki[ptop] = mn.mc_ki[ptop];
8029 } else if (!did_split && m3->mc_top >= ptop && m3->mc_pg[ptop] == mc->mc_pg[ptop] &&
8030 m3->mc_ki[ptop] >= mc->mc_ki[ptop]) {
8035 DPRINTF(("mp left: %d, rp left: %d", SIZELEFT(mp), SIZELEFT(rp)));
8038 if (copy) /* tmp page */
8039 mdb_page_free(env, copy);
8041 mc->mc_txn->mt_flags |= MDB_TXN_ERROR;
8046 mdb_put(MDB_txn *txn, MDB_dbi dbi,
8047 MDB_val *key, MDB_val *data, unsigned int flags)
8052 if (!key || !data || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8055 if ((flags & (MDB_NOOVERWRITE|MDB_NODUPDATA|MDB_RESERVE|MDB_APPEND|MDB_APPENDDUP)) != flags)
8058 mdb_cursor_init(&mc, txn, dbi, &mx);
8059 return mdb_cursor_put(&mc, key, data, flags);
8063 #define MDB_WBUF (1024*1024)
8066 /** State needed for a compacting copy. */
8067 typedef struct mdb_copy {
8068 pthread_mutex_t mc_mutex;
8069 pthread_cond_t mc_cond;
8076 pgno_t mc_next_pgno;
8079 volatile int mc_new;
8084 /** Dedicated writer thread for compacting copy. */
8085 static THREAD_RET ESECT
8086 mdb_env_copythr(void *arg)
8090 int toggle = 0, wsize, rc;
8093 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8096 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8099 pthread_mutex_lock(&my->mc_mutex);
8101 pthread_cond_signal(&my->mc_cond);
8104 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8105 if (my->mc_new < 0) {
8110 wsize = my->mc_wlen[toggle];
8111 ptr = my->mc_wbuf[toggle];
8114 DO_WRITE(rc, my->mc_fd, ptr, wsize, len);
8118 } else if (len > 0) {
8132 /* If there's an overflow page tail, write it too */
8133 if (my->mc_olen[toggle]) {
8134 wsize = my->mc_olen[toggle];
8135 ptr = my->mc_over[toggle];
8136 my->mc_olen[toggle] = 0;
8139 my->mc_wlen[toggle] = 0;
8141 pthread_cond_signal(&my->mc_cond);
8143 pthread_cond_signal(&my->mc_cond);
8144 pthread_mutex_unlock(&my->mc_mutex);
8145 return (THREAD_RET)0;
8149 /** Tell the writer thread there's a buffer ready to write */
8151 mdb_env_cthr_toggle(mdb_copy *my, int st)
8153 int toggle = my->mc_toggle ^ 1;
8154 pthread_mutex_lock(&my->mc_mutex);
8155 if (my->mc_status) {
8156 pthread_mutex_unlock(&my->mc_mutex);
8157 return my->mc_status;
8159 while (my->mc_new == 1)
8160 pthread_cond_wait(&my->mc_cond, &my->mc_mutex);
8162 my->mc_toggle = toggle;
8163 pthread_cond_signal(&my->mc_cond);
8164 pthread_mutex_unlock(&my->mc_mutex);
8168 /** Depth-first tree traversal for compacting copy. */
8170 mdb_env_cwalk(mdb_copy *my, pgno_t *pg, int flags)
8173 MDB_txn *txn = my->mc_txn;
8175 MDB_page *mo, *mp, *leaf;
8180 /* Empty DB, nothing to do */
8181 if (*pg == P_INVALID)
8188 rc = mdb_page_get(my->mc_txn, *pg, &mc.mc_pg[0], NULL);
8191 rc = mdb_page_search_root(&mc, NULL, MDB_PS_FIRST);
8195 /* Make cursor pages writable */
8196 buf = ptr = malloc(my->mc_env->me_psize * mc.mc_snum);
8200 for (i=0; i<mc.mc_top; i++) {
8201 mdb_page_copy((MDB_page *)ptr, mc.mc_pg[i], my->mc_env->me_psize);
8202 mc.mc_pg[i] = (MDB_page *)ptr;
8203 ptr += my->mc_env->me_psize;
8206 /* This is writable space for a leaf page. Usually not needed. */
8207 leaf = (MDB_page *)ptr;
8209 toggle = my->mc_toggle;
8210 while (mc.mc_snum > 0) {
8212 mp = mc.mc_pg[mc.mc_top];
8216 if (!IS_LEAF2(mp) && !(flags & F_DUPDATA)) {
8217 for (i=0; i<n; i++) {
8218 ni = NODEPTR(mp, i);
8219 if (ni->mn_flags & F_BIGDATA) {
8223 /* Need writable leaf */
8225 mc.mc_pg[mc.mc_top] = leaf;
8226 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8228 ni = NODEPTR(mp, i);
8231 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8232 rc = mdb_page_get(txn, pg, &omp, NULL);
8235 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8236 rc = mdb_env_cthr_toggle(my, 1);
8239 toggle = my->mc_toggle;
8241 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8242 memcpy(mo, omp, my->mc_env->me_psize);
8243 mo->mp_pgno = my->mc_next_pgno;
8244 my->mc_next_pgno += omp->mp_pages;
8245 my->mc_wlen[toggle] += my->mc_env->me_psize;
8246 if (omp->mp_pages > 1) {
8247 my->mc_olen[toggle] = my->mc_env->me_psize * (omp->mp_pages - 1);
8248 my->mc_over[toggle] = (char *)omp + my->mc_env->me_psize;
8249 rc = mdb_env_cthr_toggle(my, 1);
8252 toggle = my->mc_toggle;
8254 memcpy(NODEDATA(ni), &mo->mp_pgno, sizeof(pgno_t));
8255 } else if (ni->mn_flags & F_SUBDATA) {
8258 /* Need writable leaf */
8260 mc.mc_pg[mc.mc_top] = leaf;
8261 mdb_page_copy(leaf, mp, my->mc_env->me_psize);
8263 ni = NODEPTR(mp, i);
8266 memcpy(&db, NODEDATA(ni), sizeof(db));
8267 my->mc_toggle = toggle;
8268 rc = mdb_env_cwalk(my, &db.md_root, ni->mn_flags & F_DUPDATA);
8271 toggle = my->mc_toggle;
8272 memcpy(NODEDATA(ni), &db, sizeof(db));
8277 mc.mc_ki[mc.mc_top]++;
8278 if (mc.mc_ki[mc.mc_top] < n) {
8281 ni = NODEPTR(mp, mc.mc_ki[mc.mc_top]);
8283 rc = mdb_page_get(txn, pg, &mp, NULL);
8288 mc.mc_ki[mc.mc_top] = 0;
8289 if (IS_BRANCH(mp)) {
8290 /* Whenever we advance to a sibling branch page,
8291 * we must proceed all the way down to its first leaf.
8293 mdb_page_copy(mc.mc_pg[mc.mc_top], mp, my->mc_env->me_psize);
8296 mc.mc_pg[mc.mc_top] = mp;
8300 if (my->mc_wlen[toggle] >= MDB_WBUF) {
8301 rc = mdb_env_cthr_toggle(my, 1);
8304 toggle = my->mc_toggle;
8306 mo = (MDB_page *)(my->mc_wbuf[toggle] + my->mc_wlen[toggle]);
8307 mdb_page_copy(mo, mp, my->mc_env->me_psize);
8308 mo->mp_pgno = my->mc_next_pgno++;
8309 my->mc_wlen[toggle] += my->mc_env->me_psize;
8311 /* Update parent if there is one */
8312 ni = NODEPTR(mc.mc_pg[mc.mc_top-1], mc.mc_ki[mc.mc_top-1]);
8313 SETPGNO(ni, mo->mp_pgno);
8314 mdb_cursor_pop(&mc);
8316 /* Otherwise we're done */
8326 /** Copy environment with compaction. */
8328 mdb_env_copyfd1(MDB_env *env, HANDLE fd)
8333 MDB_txn *txn = NULL;
8338 my.mc_mutex = CreateMutex(NULL, FALSE, NULL);
8339 my.mc_cond = CreateEvent(NULL, FALSE, FALSE, NULL);
8340 my.mc_wbuf[0] = _aligned_malloc(MDB_WBUF*2, env->me_psize);
8341 if (my.mc_wbuf[0] == NULL)
8344 pthread_mutex_init(&my.mc_mutex, NULL);
8345 pthread_cond_init(&my.mc_cond, NULL);
8346 rc = posix_memalign((void **)&my.mc_wbuf[0], env->me_psize, MDB_WBUF*2);
8350 my.mc_wbuf[1] = my.mc_wbuf[0] + MDB_WBUF;
8355 my.mc_next_pgno = 2;
8361 THREAD_CREATE(thr, mdb_env_copythr, &my);
8363 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8367 mp = (MDB_page *)my.mc_wbuf[0];
8368 memset(mp, 0, 2*env->me_psize);
8370 mp->mp_flags = P_META;
8371 mm = (MDB_meta *)METADATA(mp);
8372 mdb_env_init_meta0(env, mm);
8373 mm->mm_address = env->me_metas[0]->mm_address;
8375 mp = (MDB_page *)(my.mc_wbuf[0] + env->me_psize);
8377 mp->mp_flags = P_META;
8378 *(MDB_meta *)METADATA(mp) = *mm;
8379 mm = (MDB_meta *)METADATA(mp);
8381 /* Count the number of free pages, subtract from lastpg to find
8382 * number of active pages
8385 MDB_ID freecount = 0;
8388 mdb_cursor_init(&mc, txn, FREE_DBI, NULL);
8389 while ((rc = mdb_cursor_get(&mc, &key, &data, MDB_NEXT)) == 0)
8390 freecount += *(MDB_ID *)data.mv_data;
8391 freecount += txn->mt_dbs[0].md_branch_pages +
8392 txn->mt_dbs[0].md_leaf_pages +
8393 txn->mt_dbs[0].md_overflow_pages;
8395 /* Set metapage 1 */
8396 mm->mm_last_pg = txn->mt_next_pgno - freecount - 1;
8397 mm->mm_dbs[1] = txn->mt_dbs[1];
8398 mm->mm_dbs[1].md_root = mm->mm_last_pg;
8401 my.mc_wlen[0] = env->me_psize * 2;
8403 pthread_mutex_lock(&my.mc_mutex);
8405 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8406 pthread_mutex_unlock(&my.mc_mutex);
8407 rc = mdb_env_cwalk(&my, &txn->mt_dbs[1].md_root, 0);
8408 if (rc == MDB_SUCCESS && my.mc_wlen[my.mc_toggle])
8409 rc = mdb_env_cthr_toggle(&my, 1);
8410 mdb_env_cthr_toggle(&my, -1);
8411 pthread_mutex_lock(&my.mc_mutex);
8413 pthread_cond_wait(&my.mc_cond, &my.mc_mutex);
8414 pthread_mutex_unlock(&my.mc_mutex);
8419 CloseHandle(my.mc_cond);
8420 CloseHandle(my.mc_mutex);
8421 _aligned_free(my.mc_wbuf[0]);
8423 pthread_cond_destroy(&my.mc_cond);
8424 pthread_mutex_destroy(&my.mc_mutex);
8425 free(my.mc_wbuf[0]);
8430 /** Copy environment as-is. */
8432 mdb_env_copyfd0(MDB_env *env, HANDLE fd)
8434 MDB_txn *txn = NULL;
8440 #define DO_WRITE(rc, fd, ptr, w2, len) rc = WriteFile(fd, ptr, w2, &len, NULL)
8444 #define DO_WRITE(rc, fd, ptr, w2, len) len = write(fd, ptr, w2); rc = (len >= 0)
8447 /* Do the lock/unlock of the reader mutex before starting the
8448 * write txn. Otherwise other read txns could block writers.
8450 rc = mdb_txn_begin(env, NULL, MDB_RDONLY, &txn);
8455 /* We must start the actual read txn after blocking writers */
8456 mdb_txn_reset0(txn, "reset-stage1");
8458 /* Temporarily block writers until we snapshot the meta pages */
8461 rc = mdb_txn_renew0(txn);
8463 UNLOCK_MUTEX_W(env);
8468 wsize = env->me_psize * 2;
8472 DO_WRITE(rc, fd, ptr, w2, len);
8476 } else if (len > 0) {
8482 /* Non-blocking or async handles are not supported */
8488 UNLOCK_MUTEX_W(env);
8493 w2 = txn->mt_next_pgno * env->me_psize;
8496 LARGE_INTEGER fsize;
8497 GetFileSizeEx(env->me_fd, &fsize);
8498 if (w2 > fsize.QuadPart)
8499 w2 = fsize.QuadPart;
8504 fstat(env->me_fd, &st);
8505 if (w2 > (size_t)st.st_size)
8511 if (wsize > MAX_WRITE)
8515 DO_WRITE(rc, fd, ptr, w2, len);
8519 } else if (len > 0) {
8536 mdb_env_copyfd2(MDB_env *env, HANDLE fd, unsigned int flags)
8538 if (flags & MDB_CP_COMPACT)
8539 return mdb_env_copyfd1(env, fd);
8541 return mdb_env_copyfd0(env, fd);
8545 mdb_env_copyfd(MDB_env *env, HANDLE fd)
8547 return mdb_env_copyfd2(env, fd, 0);
8551 mdb_env_copy2(MDB_env *env, const char *path, unsigned int flags)
8555 HANDLE newfd = INVALID_HANDLE_VALUE;
8557 if (env->me_flags & MDB_NOSUBDIR) {
8558 lpath = (char *)path;
8561 len += sizeof(DATANAME);
8562 lpath = malloc(len);
8565 sprintf(lpath, "%s" DATANAME, path);
8568 /* The destination path must exist, but the destination file must not.
8569 * We don't want the OS to cache the writes, since the source data is
8570 * already in the OS cache.
8573 newfd = CreateFile(lpath, GENERIC_WRITE, 0, NULL, CREATE_NEW,
8574 FILE_FLAG_NO_BUFFERING|FILE_FLAG_WRITE_THROUGH, NULL);
8576 newfd = open(lpath, O_WRONLY|O_CREAT|O_EXCL, 0666);
8578 if (newfd == INVALID_HANDLE_VALUE) {
8584 /* Set O_DIRECT if the file system supports it */
8585 if ((rc = fcntl(newfd, F_GETFL)) != -1)
8586 (void) fcntl(newfd, F_SETFL, rc | O_DIRECT);
8588 #ifdef F_NOCACHE /* __APPLE__ */
8589 rc = fcntl(newfd, F_NOCACHE, 1);
8596 rc = mdb_env_copyfd2(env, newfd, flags);
8599 if (!(env->me_flags & MDB_NOSUBDIR))
8601 if (newfd != INVALID_HANDLE_VALUE)
8602 if (close(newfd) < 0 && rc == MDB_SUCCESS)
8609 mdb_env_copy(MDB_env *env, const char *path)
8611 return mdb_env_copy2(env, path, 0);
8615 mdb_env_set_flags(MDB_env *env, unsigned int flag, int onoff)
8617 if ((flag & CHANGEABLE) != flag)
8620 env->me_flags |= flag;
8622 env->me_flags &= ~flag;
8627 mdb_env_get_flags(MDB_env *env, unsigned int *arg)
8632 *arg = env->me_flags;
8637 mdb_env_set_userctx(MDB_env *env, void *ctx)
8641 env->me_userctx = ctx;
8646 mdb_env_get_userctx(MDB_env *env)
8648 return env ? env->me_userctx : NULL;
8652 mdb_env_set_assert(MDB_env *env, MDB_assert_func *func)
8657 env->me_assert_func = func;
8663 mdb_env_get_path(MDB_env *env, const char **arg)
8668 *arg = env->me_path;
8673 mdb_env_get_fd(MDB_env *env, mdb_filehandle_t *arg)
8682 /** Common code for #mdb_stat() and #mdb_env_stat().
8683 * @param[in] env the environment to operate in.
8684 * @param[in] db the #MDB_db record containing the stats to return.
8685 * @param[out] arg the address of an #MDB_stat structure to receive the stats.
8686 * @return 0, this function always succeeds.
8689 mdb_stat0(MDB_env *env, MDB_db *db, MDB_stat *arg)
8691 arg->ms_psize = env->me_psize;
8692 arg->ms_depth = db->md_depth;
8693 arg->ms_branch_pages = db->md_branch_pages;
8694 arg->ms_leaf_pages = db->md_leaf_pages;
8695 arg->ms_overflow_pages = db->md_overflow_pages;
8696 arg->ms_entries = db->md_entries;
8702 mdb_env_stat(MDB_env *env, MDB_stat *arg)
8706 if (env == NULL || arg == NULL)
8709 toggle = mdb_env_pick_meta(env);
8711 return mdb_stat0(env, &env->me_metas[toggle]->mm_dbs[MAIN_DBI], arg);
8715 mdb_env_info(MDB_env *env, MDB_envinfo *arg)
8719 if (env == NULL || arg == NULL)
8722 toggle = mdb_env_pick_meta(env);
8723 arg->me_mapaddr = env->me_metas[toggle]->mm_address;
8724 arg->me_mapsize = env->me_mapsize;
8725 arg->me_maxreaders = env->me_maxreaders;
8727 /* me_numreaders may be zero if this process never used any readers. Use
8728 * the shared numreader count if it exists.
8730 arg->me_numreaders = env->me_txns ? env->me_txns->mti_numreaders : env->me_numreaders;
8732 arg->me_last_pgno = env->me_metas[toggle]->mm_last_pg;
8733 arg->me_last_txnid = env->me_metas[toggle]->mm_txnid;
8737 /** Set the default comparison functions for a database.
8738 * Called immediately after a database is opened to set the defaults.
8739 * The user can then override them with #mdb_set_compare() or
8740 * #mdb_set_dupsort().
8741 * @param[in] txn A transaction handle returned by #mdb_txn_begin()
8742 * @param[in] dbi A database handle returned by #mdb_dbi_open()
8745 mdb_default_cmp(MDB_txn *txn, MDB_dbi dbi)
8747 uint16_t f = txn->mt_dbs[dbi].md_flags;
8749 txn->mt_dbxs[dbi].md_cmp =
8750 (f & MDB_REVERSEKEY) ? mdb_cmp_memnr :
8751 (f & MDB_INTEGERKEY) ? mdb_cmp_cint : mdb_cmp_memn;
8753 txn->mt_dbxs[dbi].md_dcmp =
8754 !(f & MDB_DUPSORT) ? 0 :
8755 ((f & MDB_INTEGERDUP)
8756 ? ((f & MDB_DUPFIXED) ? mdb_cmp_int : mdb_cmp_cint)
8757 : ((f & MDB_REVERSEDUP) ? mdb_cmp_memnr : mdb_cmp_memn));
8760 int mdb_dbi_open(MDB_txn *txn, const char *name, unsigned int flags, MDB_dbi *dbi)
8765 int rc, dbflag, exact;
8766 unsigned int unused = 0;
8769 if (txn->mt_dbxs[FREE_DBI].md_cmp == NULL) {
8770 mdb_default_cmp(txn, FREE_DBI);
8773 if ((flags & VALID_FLAGS) != flags)
8775 if (txn->mt_flags & MDB_TXN_ERROR)
8781 if (flags & PERSISTENT_FLAGS) {
8782 uint16_t f2 = flags & PERSISTENT_FLAGS;
8783 /* make sure flag changes get committed */
8784 if ((txn->mt_dbs[MAIN_DBI].md_flags | f2) != txn->mt_dbs[MAIN_DBI].md_flags) {
8785 txn->mt_dbs[MAIN_DBI].md_flags |= f2;
8786 txn->mt_flags |= MDB_TXN_DIRTY;
8789 mdb_default_cmp(txn, MAIN_DBI);
8793 if (txn->mt_dbxs[MAIN_DBI].md_cmp == NULL) {
8794 mdb_default_cmp(txn, MAIN_DBI);
8797 /* Is the DB already open? */
8799 for (i=2; i<txn->mt_numdbs; i++) {
8800 if (!txn->mt_dbxs[i].md_name.mv_size) {
8801 /* Remember this free slot */
8802 if (!unused) unused = i;
8805 if (len == txn->mt_dbxs[i].md_name.mv_size &&
8806 !strncmp(name, txn->mt_dbxs[i].md_name.mv_data, len)) {
8812 /* If no free slot and max hit, fail */
8813 if (!unused && txn->mt_numdbs >= txn->mt_env->me_maxdbs)
8814 return MDB_DBS_FULL;
8816 /* Cannot mix named databases with some mainDB flags */
8817 if (txn->mt_dbs[MAIN_DBI].md_flags & (MDB_DUPSORT|MDB_INTEGERKEY))
8818 return (flags & MDB_CREATE) ? MDB_INCOMPATIBLE : MDB_NOTFOUND;
8820 /* Find the DB info */
8821 dbflag = DB_NEW|DB_VALID;
8824 key.mv_data = (void *)name;
8825 mdb_cursor_init(&mc, txn, MAIN_DBI, NULL);
8826 rc = mdb_cursor_set(&mc, &key, &data, MDB_SET, &exact);
8827 if (rc == MDB_SUCCESS) {
8828 /* make sure this is actually a DB */
8829 MDB_node *node = NODEPTR(mc.mc_pg[mc.mc_top], mc.mc_ki[mc.mc_top]);
8830 if (!(node->mn_flags & F_SUBDATA))
8831 return MDB_INCOMPATIBLE;
8832 } else if (rc == MDB_NOTFOUND && (flags & MDB_CREATE)) {
8833 /* Create if requested */
8835 data.mv_size = sizeof(MDB_db);
8836 data.mv_data = &dummy;
8837 memset(&dummy, 0, sizeof(dummy));
8838 dummy.md_root = P_INVALID;
8839 dummy.md_flags = flags & PERSISTENT_FLAGS;
8840 rc = mdb_cursor_put(&mc, &key, &data, F_SUBDATA);
8844 /* OK, got info, add to table */
8845 if (rc == MDB_SUCCESS) {
8846 unsigned int slot = unused ? unused : txn->mt_numdbs;
8847 txn->mt_dbxs[slot].md_name.mv_data = strdup(name);
8848 txn->mt_dbxs[slot].md_name.mv_size = len;
8849 txn->mt_dbxs[slot].md_rel = NULL;
8850 txn->mt_dbflags[slot] = dbflag;
8851 memcpy(&txn->mt_dbs[slot], data.mv_data, sizeof(MDB_db));
8853 mdb_default_cmp(txn, slot);
8862 int mdb_stat(MDB_txn *txn, MDB_dbi dbi, MDB_stat *arg)
8864 if (!arg || !TXN_DBI_EXIST(txn, dbi))
8867 if (txn->mt_flags & MDB_TXN_ERROR)
8870 if (txn->mt_dbflags[dbi] & DB_STALE) {
8873 /* Stale, must read the DB's root. cursor_init does it for us. */
8874 mdb_cursor_init(&mc, txn, dbi, &mx);
8876 return mdb_stat0(txn->mt_env, &txn->mt_dbs[dbi], arg);
8879 void mdb_dbi_close(MDB_env *env, MDB_dbi dbi)
8882 if (dbi <= MAIN_DBI || dbi >= env->me_maxdbs)
8884 ptr = env->me_dbxs[dbi].md_name.mv_data;
8885 env->me_dbxs[dbi].md_name.mv_data = NULL;
8886 env->me_dbxs[dbi].md_name.mv_size = 0;
8887 env->me_dbflags[dbi] = 0;
8891 int mdb_dbi_flags(MDB_txn *txn, MDB_dbi dbi, unsigned int *flags)
8893 /* We could return the flags for the FREE_DBI too but what's the point? */
8894 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8896 *flags = txn->mt_dbs[dbi].md_flags & PERSISTENT_FLAGS;
8900 /** Add all the DB's pages to the free list.
8901 * @param[in] mc Cursor on the DB to free.
8902 * @param[in] subs non-Zero to check for sub-DBs in this DB.
8903 * @return 0 on success, non-zero on failure.
8906 mdb_drop0(MDB_cursor *mc, int subs)
8910 rc = mdb_page_search(mc, NULL, MDB_PS_FIRST);
8911 if (rc == MDB_SUCCESS) {
8912 MDB_txn *txn = mc->mc_txn;
8917 /* LEAF2 pages have no nodes, cannot have sub-DBs */
8918 if (IS_LEAF2(mc->mc_pg[mc->mc_top]))
8921 mdb_cursor_copy(mc, &mx);
8922 while (mc->mc_snum > 0) {
8923 MDB_page *mp = mc->mc_pg[mc->mc_top];
8924 unsigned n = NUMKEYS(mp);
8926 for (i=0; i<n; i++) {
8927 ni = NODEPTR(mp, i);
8928 if (ni->mn_flags & F_BIGDATA) {
8931 memcpy(&pg, NODEDATA(ni), sizeof(pg));
8932 rc = mdb_page_get(txn, pg, &omp, NULL);
8935 mdb_cassert(mc, IS_OVERFLOW(omp));
8936 rc = mdb_midl_append_range(&txn->mt_free_pgs,
8940 } else if (subs && (ni->mn_flags & F_SUBDATA)) {
8941 mdb_xcursor_init1(mc, ni);
8942 rc = mdb_drop0(&mc->mc_xcursor->mx_cursor, 0);
8948 if ((rc = mdb_midl_need(&txn->mt_free_pgs, n)) != 0)
8950 for (i=0; i<n; i++) {
8952 ni = NODEPTR(mp, i);
8955 mdb_midl_xappend(txn->mt_free_pgs, pg);
8960 mc->mc_ki[mc->mc_top] = i;
8961 rc = mdb_cursor_sibling(mc, 1);
8963 if (rc != MDB_NOTFOUND)
8965 /* no more siblings, go back to beginning
8966 * of previous level.
8970 for (i=1; i<mc->mc_snum; i++) {
8972 mc->mc_pg[i] = mx.mc_pg[i];
8977 rc = mdb_midl_append(&txn->mt_free_pgs, mc->mc_db->md_root);
8980 txn->mt_flags |= MDB_TXN_ERROR;
8981 } else if (rc == MDB_NOTFOUND) {
8987 int mdb_drop(MDB_txn *txn, MDB_dbi dbi, int del)
8989 MDB_cursor *mc, *m2;
8992 if ((unsigned)del > 1 || dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
8995 if (F_ISSET(txn->mt_flags, MDB_TXN_RDONLY))
8998 rc = mdb_cursor_open(txn, dbi, &mc);
9002 rc = mdb_drop0(mc, mc->mc_db->md_flags & MDB_DUPSORT);
9003 /* Invalidate the dropped DB's cursors */
9004 for (m2 = txn->mt_cursors[dbi]; m2; m2 = m2->mc_next)
9005 m2->mc_flags &= ~(C_INITIALIZED|C_EOF);
9009 /* Can't delete the main DB */
9010 if (del && dbi > MAIN_DBI) {
9011 rc = mdb_del0(txn, MAIN_DBI, &mc->mc_dbx->md_name, NULL, 0);
9013 txn->mt_dbflags[dbi] = DB_STALE;
9014 mdb_dbi_close(txn->mt_env, dbi);
9016 txn->mt_flags |= MDB_TXN_ERROR;
9019 /* reset the DB record, mark it dirty */
9020 txn->mt_dbflags[dbi] |= DB_DIRTY;
9021 txn->mt_dbs[dbi].md_depth = 0;
9022 txn->mt_dbs[dbi].md_branch_pages = 0;
9023 txn->mt_dbs[dbi].md_leaf_pages = 0;
9024 txn->mt_dbs[dbi].md_overflow_pages = 0;
9025 txn->mt_dbs[dbi].md_entries = 0;
9026 txn->mt_dbs[dbi].md_root = P_INVALID;
9028 txn->mt_flags |= MDB_TXN_DIRTY;
9031 mdb_cursor_close(mc);
9035 int mdb_set_compare(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9037 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9040 txn->mt_dbxs[dbi].md_cmp = cmp;
9044 int mdb_set_dupsort(MDB_txn *txn, MDB_dbi dbi, MDB_cmp_func *cmp)
9046 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9049 txn->mt_dbxs[dbi].md_dcmp = cmp;
9053 int mdb_set_relfunc(MDB_txn *txn, MDB_dbi dbi, MDB_rel_func *rel)
9055 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9058 txn->mt_dbxs[dbi].md_rel = rel;
9062 int mdb_set_relctx(MDB_txn *txn, MDB_dbi dbi, void *ctx)
9064 if (dbi == FREE_DBI || !TXN_DBI_EXIST(txn, dbi))
9067 txn->mt_dbxs[dbi].md_relctx = ctx;
9072 mdb_env_get_maxkeysize(MDB_env *env)
9074 return ENV_MAXKEY(env);
9078 mdb_reader_list(MDB_env *env, MDB_msg_func *func, void *ctx)
9080 unsigned int i, rdrs;
9083 int rc = 0, first = 1;
9087 if (!env->me_txns) {
9088 return func("(no reader locks)\n", ctx);
9090 rdrs = env->me_txns->mti_numreaders;
9091 mr = env->me_txns->mti_readers;
9092 for (i=0; i<rdrs; i++) {
9094 txnid_t txnid = mr[i].mr_txnid;
9095 sprintf(buf, txnid == (txnid_t)-1 ?
9096 "%10d %"Z"x -\n" : "%10d %"Z"x %"Z"u\n",
9097 (int)mr[i].mr_pid, (size_t)mr[i].mr_tid, txnid);
9100 rc = func(" pid thread txnid\n", ctx);
9104 rc = func(buf, ctx);
9110 rc = func("(no active readers)\n", ctx);
9115 /** Insert pid into list if not already present.
9116 * return -1 if already present.
9119 mdb_pid_insert(MDB_PID_T *ids, MDB_PID_T pid)
9121 /* binary search of pid in list */
9123 unsigned cursor = 1;
9125 unsigned n = ids[0];
9128 unsigned pivot = n >> 1;
9129 cursor = base + pivot + 1;
9130 val = pid - ids[cursor];
9135 } else if ( val > 0 ) {
9140 /* found, so it's a duplicate */
9149 for (n = ids[0]; n > cursor; n--)
9156 mdb_reader_check(MDB_env *env, int *dead)
9158 unsigned int i, j, rdrs;
9160 MDB_PID_T *pids, pid;
9169 rdrs = env->me_txns->mti_numreaders;
9170 pids = malloc((rdrs+1) * sizeof(MDB_PID_T));
9174 mr = env->me_txns->mti_readers;
9175 for (i=0; i<rdrs; i++) {
9176 if (mr[i].mr_pid && mr[i].mr_pid != env->me_pid) {
9178 if (mdb_pid_insert(pids, pid) == 0) {
9179 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9181 /* Recheck, a new process may have reused pid */
9182 if (!mdb_reader_pid(env, Pidcheck, pid)) {
9183 for (j=i; j<rdrs; j++)
9184 if (mr[j].mr_pid == pid) {
9185 DPRINTF(("clear stale reader pid %u txn %"Z"d",
9186 (unsigned) pid, mr[j].mr_txnid));
9191 UNLOCK_MUTEX_R(env);